The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Welcome Robert H. Rasche W elcome to the Thirty-Second Annual Policy Conference sponsored by the Federal Reserve Bank of St. Louis. Our theme this year is “Monetary Policy Under Uncertainty.” We chose this topic to coordinate with Bill Poole’s imminent completion of 10 years of service at the Bank and his contributions over the years to the policy debate. Now that Bill’s forthcoming retirement as president of the Bank is official, we can plainly say that this conference is being held in his honor. We have tried to keep our motivation below Bill’s radar screen, though I suspect that we have not been completely successful. He has been gracious enough to limit his inquiries and not spoil our fun. Monetary policy under uncertainty has been one of Bill’s professional interests throughout his career. His 1970 Quarterly Journal of Economics paper, “Optimal Choice of Monetary Policy Instruments in a Simple Stochastic Macro Model,”1 is well-known and oft cited. (We have found 364 citations in the Social Sciences Citation Index to this publication over the years, and citations still continue 37 years later!) His interest in this subject has been clear during his service on the Federal Open Market Committee (FOMC) and in his speeches and publications on topics such as “A Policymaker Confronts Uncertainty,”2 “Perfecting the Market’s Knowledge of Monetary Policy,”3 “The Impact of Changes in FOMC Disclosure Practices on the Transparency of Monetary Policy: Are Markets and the FOMC Better ‘Synched’?,”4 “Fed Transparency: How, Not Whether,”5 and “How Predictable Is Fed Policy?”6 We are not allowing Bill to sit back completely and consume during this conference—we have included him in our panel discussion. We are very pleased with the distinguished authors and discussants who have agreed to contribute to this program in honor of Bill, as well as those of you who have set aside time to attend. We look forward to an active and stimulating discussion that will provide ideas for future research on this topic and possibly even provoke another speech from Bill before he retires. REFERENCES Poole, William. “Optimal Choice of Monetary Policy Instruments in a Simple Stochastic Macro Model.” Quarterly Journal of Economics, May 1970, 84(2), pp. 197-216. Poole, William. “A Policymaker Confronts Uncertainty.” Federal Reserve Bank of St. Louis Review, September/October 1998, 80(5), pp. 3-9. 1 Poole (1970). 2 Poole (1998). 3 Poole and Rasche (2000). 4 Poole and Rasche (2003). 5 Poole (2003). 6 Poole (2005). Robert H. Rasche is a senior vice president and director of research at the Federal Reserve Bank of St. Louis. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 269-70. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 269 Rasche Poole, William. “Fed Transparency: How, Not Whether.” Federal Reserve Bank of St. Louis Review, November/December 2003, 85(1), pp. 1-8. Poole, William. “How Predictable Is Fed Policy?” Federal Reserve Bank of St. Louis Review, November/December 2005, 87(6), pp. 659-68. Poole, William and Rasche, Robert H. “Perfecting the Market’s Knowledge of Monetary Policy.” Journal of Financial Services Research, December 2000, 18(2/3), pp. 255-98. Poole, William and Rasche, Robert H. “The Impact of Changes in FOMC Disclosure Practices on the Transparency of Monetary Policy: Are Markets and the FOMC Better ‘Synched’?” Federal Reserve Bank of St. Louis Review, January/February 2003, 85(1), pp. 1-10. 270 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Optimal Monetary Policy Under Uncertainty: A Markov Jump-Linear-Quadratic Approach Lars E.O. Svensson and Noah Williams This paper studies the design of optimal monetary policy under uncertainty using a Markov jumplinear-quadratic (MJLQ) approach. To approximate the uncertainty that policymakers face, the authors use different discrete modes in a Markov chain and take mode-dependent linear-quadratic approximations of the underlying model. This allows the authors to apply a powerful methodology with convenient solution algorithms that they have developed. They apply their methods to analyze the effects of uncertainty and potential gains from experimentation for two sources of uncertainty in the New Keynesian Phillips curve. The examples highlight that learning may have sizable effects on losses and, although it is generally beneficial, it need not always be so. The experimentation component typically has little effect and in some cases it can lead to attenuation of policy. (JEL E42, E52, E58) Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 275-93. I have long been interested in the analysis of monetary policy under uncertainty. The problems arise from what we do not know; we must deal with the uncertainty from the base of what we do know… The Fed faces many uncertainties, and must adjust its one policy instrument to navigate as best it can this sea of uncertainty. Our fundamental principle is that we must use that one policy instrument to achieve long-run price stability… My bottom line is that market participants should concentrate on the fundamentals. If the bond traders can get it right, they’ll do most of the stabilization work for us, and we at the Fed can sit back and enjoy life. —William Poole (1998), President of the Federal Reserve Bank of St. Louis (1998-2008) E arly in his tenure as president of the Federal Reserve Bank of St. Louis, William Poole laid out some of the issues that policymakers face when deciding on policy, as reflected in the quotations here. In this paper we take up some of these issues, applying a framework to help policymakers navigate the “sea of uncertainty.” We focus particularly on the issue of the knowledge and beliefs of the policymakers and the private sector—showing how both groups of agents learn from their observations and how this may or may not lead to enhanced economic stability. We also address the extent to which policymakers should “sit back” or, instead, actively intervene in markets in order to gain knowledge to help mitigate future uncertainty. In previous work, Svensson and Williams (2007a,b), we have developed methods to study optimal policy in Markov jump-linear-quadratic Lars E.O. Svensson is deputy governor of the Sveriges Riksbank and a professor of economics at Princeton University. Noah Williams is an assistant professor of economics at Princeton University. The authors thank James Bullard, Timothy Cogley, Andrew Levin, and William Poole for comments on this paper. Financial support from the National Science Foundation is gratefully acknowledged. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks, or the Executive Board of Sveriges Riksbank. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 275 Svensson and Williams (MJLQ) models with forward-looking variables: models with conditionally linear dynamics and conditionally quadratic preferences, where the matrices in both preferences and dynamics are random. In particular, each model has multiple modes, a finite collection of different possible values for the matrices, whose evolution is governed by a finite-state Markov chain. In our previous work, we have discussed how these modes could be structured to capture many different types of uncertainty relevant for policymakers. Here we put those suggestions into practice in a simple benchmark policy model. In a first paper, Svensson and Williams (2007a), we studied optimal policy design in MJLQ models when policymakers can or cannot observe the current mode, but we abstracted from any learning and inference about the current mode. Although in many cases the optimal policy under no learning (NL) is not a normatively desirable policy, it serves as a useful benchmark for our later policy analyses. In a second paper, Svensson and Williams (2007b), we focused on learning and inference in the more relevant situation, particularly for the model-uncertainty applications which interest us, in which the modes are not directly observable. Thus, decisionmakers must filter their observations to make inferences about the current mode. As in most Bayesian learning problems, the optimal policy thus typically includes an experimentation component reflecting the endogeneity of information. This class of problems has a long history in economics, and it is well-known that solutions are difficult to obtain. We developed algorithms to solve numerically for the optimal policy.1 Due to the 1 In addition to the classic literature (on such problems as a monopolist learning its demand curve), Wieland (2000 and 2006) and Beck and Wieland (2002) have recently examined Bayesian optimal policy and optimal experimentation in a context similar to ours but without forward-looking variables. Tesfaselassie, Schaling, and Eijffinger (2006) examine passive and active learning in a simple model with a forward-looking element in the form of a long interest rate in the aggregate-demand equation. Ellison and Valla (2001) and Cogley, Colacito, and Sargent (2007) study situations like ours but where the expectational component is as in the Lucas-supply curve (Et –1πt, for example) rather than our forwardlooking case (Et πt +1, for example). More closely related to our present paper, Ellison (2006) analyzes active and passive learning in a New Keynesian model with uncertainty about the slope of the Phillips curve. 276 J U LY / A U G U S T 2008 curse of dimensionality, the Bayesian optimal policy (BOP) is feasible only in relatively small models. Confronted with these difficulties, we also considered adaptive optimal policy (AOP).2 In this case, in each period the policymaker does update the probability distribution of the current mode in a Bayesian way, but the optimal policy is computed each period under the assumption that the policymaker will not learn in the future from observations. In our setting, the AOP is significantly easier to compute, and in many cases provides a good approximation to the BOP. Moreover, the AOP analysis is of some interest in its own right because it is closely related to specifications of adaptive learning that have been widely studied in macroeconomics (see Evans and Honkapohja, 2001, for an overview). Further, the AOP specification rules out the experimentation that some may view as objectionable in a policy context.3 In this paper, we apply our methodology to study optimal monetary policy design under uncertainty in dynamic stochastic general equilibrium (DSGE) models. We begin by summarizing the main findings from our previous work, leading to implementable algorithms for analyzing policy in MJLQ models. We then turn to examples that highlight the effects of learning and experimentation for two sources of uncertainty in the benchmark New Keynesian Phillips curve. In this model we compare and contrast optimal policies under NL, AOP, and BOP. We analyze whether learning is beneficial—it is not always so, a fact which at least partially reflects our assumption of symmetric information between the policymakers and the public—and then quantify the additional gains from experimentation.4 We find 2 What we call optimal policy under no learning, adaptive optimal policy, and Bayesian optimal policy have in the literature also been referred to as myopia, passive learning, and active learning, respectively. 3 In addition, AOP is useful for technical reasons because it gives us a good starting point for our more intensive numerical calculations in the BOP case. 4 In addition to our own previous work, MJLQ models have been widely studied in the control-theory literature for the special case when the model modes are observable and there are no forwardlooking variables (see Costa, Fragoso, and Marques, 2005, and the references therein); do Val and Başar (1999) provide an application of an adaptive-control MJLQ problem in economics. More recently, Zampolli (2006) has used such an MJLQ model to examine mone- F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams that the experimentation component is typically small. Recognizing the informational component of policy actions often leads policy to be slightly more aggressive, but, somewhat surprisingly, in one example here it leads to a less aggressive optimal policy. The paper is organized as follows: The next section presents the MJLQ framework and summarizes our earlier work. The third section presents our analysis of learning and experimentation in a simple benchmark New Keynesian model. The fourth section presents some conclusions and suggestions for further work. MJLQ ANALYSIS OF OPTIMAL POLICY This section summarizes our earlier work, Svensson and Williams (2007a,b). An MJLQ Model We consider an MJLQ model of an economy with forward-looking variables. The economy has a private sector and a policymaker. We let Xt denote an nx vector of predetermined variables in period t, xt an nx vector of forward-looking variables, and it an nx vector of (policymaker) instruments (control variables).5 We let model uncertainty be represented by nj possible modes and let jt 僆 Nj ⬅ {1,2,…,nj } denote the mode in tary policy under shifts between regimes with and without an asset-market bubble. Blake and Zampolli (2006) provide an extension of the MJLQ model with observable modes to include forwardlooking variables and present an algorithm for the solution of an equilibrium resulting from optimization under discretion. Svensson and Williams (2007a) provide a more general extension of the MJLQ framework with forward-looking variables and present algorithms for the solution of an equilibrium resulting from optimization under commitment in a timeless perspective as well as arbitrary time-varying or time-invariant policy rules, using the recursive saddlepoint method of Marcet and Marimon (1998). They also provide two concrete examples: an estimated backward-looking model (a three-mode variant of Rudebusch and Svensson, 1999) and an estimated forward-looking model (a three-mode variant of Lindé, 2005). Svensson and Williams (2007a) also extend the MJLQ framework to the more realistic case of unobservable modes, although without introducing learning and inference about the probability distribution of modes. Svensson and Williams (2007b) focus on learning and experimentation in the MJLQ framework. 5 The first component of Xt may be unity, in order to allow for mode-dependent intercepts in the model equations. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W period t. The model of the economy can then be written (1) X t +1 = A11 jt +1 X t + A12 jt +1 x t + B1 jt +1 it + C1 jt +1 ε t +1, (2) Ε t H jt +1 xt +1 = A21 jt X t + A22 jt x t + B2 jt it + C2 jt εt , where εt is a multivariate normally distributed random i.i.d. nε vector of shocks with mean zero and contemporaneous covariance matrix Inε . The matrices A11j ,A12j ,…,C2j have the appropriate dimensions and depend on the mode j. Because a structural model here is simply a collection of matrices, each mode can represent a different model of the economy. Thus, uncertainty about the prevailing mode is model uncertainty.6 Note that the matrices on the right side of (1) depend on the mode jt +1 in period t +1, whereas the matrices on the right side of (2) depend on the mode jt in period t. Equation (1) then determines the predetermined variables in period t +1 as a function of the mode and shocks in period t +1 and the predetermined variables, forward-looking variables, and instruments in period t. Equation (2) determines the forward-looking variables in period t as a function of the mode and shocks in period t, the expectations in period t of next period’s mode and forward-looking variables, and the predetermined variables and instruments in period t. The matrix A22j is nonsingular for each j 僆 Nj . The mode jt follows a Markov process with the transition matrix P ; Pjk .7 The shocks εt are mean zero and i.i.d. with probability density ϕ ; and without loss of generality we assume that εt is independent of jt .8 We also assume that c1j εt and c2k εt are independent for 6 See also Svensson and Williams (2007a), where we show how many different types of uncertainty can be mapped into our MJLQ framework. 7 Obvious special cases are P = In , when the modes are completely j –′ ( j 僆 N ), when persistent, and Pj = p the modes are serially i.i.d. j – with probability distribution p. 8 Because mode-dependent intercepts (as well as mode-dependent standard deviations) are allowed in the model, we can still incorporate additive mode-dependent shocks. J U LY / A U G U S T 2008 277 Svensson and Williams all j,k 僆 Nj . These shocks, along with the modes, are the driving forces in the model. They are not directly observed. For technical reasons, it is convenient but not necessary that they are independent. We let pt = 共p1t ,…,pnjt 兲′ denote the true probability distribution of jt in period t. We let pt+τ|t denote the policymaker and private sector estimate in the beginning of period t of the probability distribution in period t +τ. The prediction equation for the probability distribution is pt +1|t = P ′pt|t . (3) We let the operator Et[.] in the expression Et Hjt +1xt +1 on the left side of (2) denote expectations in period t conditional on policymaker and private sector information in the beginning of period t, including Xt , it , and pt|t but excluding jt and εt . Thus, the maintained assumption is symmetric information between the policymaker and the (aggregate) private sector. Because forwardlooking variables will be allowed to depend on jt , parts of the private sector, but not the aggregate private sector, may be able to observe jt and parts of εt . Note that although we focus on the determination of the optimal policy instrument, it , our results also show how private sector choices as embodied in xt are affected by uncertainty and learning. The precise informational assumptions and the determination of pt|t will be specified below. We let the policymaker intertemporal loss function in period t be Ε t ∑δ τ L ( X t +τ , x t +τ , it + τ , j t +τ ), ` (4) τ =0 where δ is a discount factor satisfying 0 < δ < 1, and the period loss, L共Xt , xt , it , jt 兲, satisfies (5) X t ′ L ( X t , x t , it , jt ); x t W jt it Xt , xt it where the matrix Wj (j 僆 Nj ) is positive semidefinite. We assume that the policymaker optimizes under commitment in a timeless perspective. As explained below, we will then add the term 278 J U LY / A U G U S T 2008 (6) 1 Ξt −1 Ε t H jt x t δ to the intertemporal loss function in period t. As we shall see below, the nx vector Ξt –1 is the vector of Lagrange multipliers for equation (2) from the optimization problem in period t –1. For the special case when there are no forward-looking variables (nx = 0), the model consists of (1) only, without the term A12jt+1xt ; the period loss function depends on Xt , it , and jt only; and there is no role for the Lagrange multipliers, Ξt –1, or the term (6). Approximate MJLQ models Although in this paper we start with an MJLQ model, it is natural to ask where such a model comes from, as usual formulations of economic models are not of this type. However, the same type of approximation methods that are widely used to convert nonlinear models into their linear counterparts can also convert nonlinear models into MJLQ models. We analyze this issue in Svensson and Williams (2007a) and present an illustration as well. Here we briefly discuss the main ideas. Rather than analyze local deviations from a single steady state as in conventional linearizations, for an MJLQ approximation we analyze the local deviations from (potentially) separate, mode-dependent steady states. Standard linearizations are justified as asymptotically valid for small shocks, as an increasing time is spent in the vicinity of the steady state. Our MJLQ approximations are asymptotically valid for small shocks and persistent modes, as an increasing time is spent in the vicinity of each mode-dependent steady state. Thus, for slowly varying Markov chains, our MJLQ model provides accurate approximations of nonlinear models with Markov switching. Types of Optimal Policies We will distinguish three cases: (i) optimal policy when there is no learning (NL), (ii) adaptive optimal policy (AOP), and (iii) Bayesian optimal policy (BOP). By NL, we refer to a situation when the policymaker and the aggregate private sector have a probability distribution pt|t over the modes in period t and update the probability F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams distribution in future periods using the transition matrix only, so the updating equation is (7) pt +1|t +1 = P ′pt|t . That is, the policymaker and the private sector do not use observations of the variables in the economy to update the probability distribution. The policymaker then determines optimal policy in period t conditional on pt|t and (7). This is a variant of a case examined in Svensson and Williams (2007a). By AOP, we refer to a situation when the policymaker in period t determines optimal policy as in the NL case, but then uses observations of the realization of the variables in the economy to update the probability distribution according to Bayes’s theorem. In this case, the instruments will generally have an effect on the updating of future probability distributions and through this channel separately affect the intertemporal loss. However, the policymaker does not exploit this channel in determining optimal policy. That is, the policymaker does not do any conscious experimentation. By BOP, we refer to a situation when the policymaker acknowledges that the current instruments will affect future inference and updating of the probability distribution and calculates optimal policy taking this separate channel into account. Therefore, BOP includes optimal experimentation, where for instance the policymaker may pursue policy that increases losses in the short run but improves the inference of the probability distribution and therefore lowers losses in the longer run. Optimal Policy with No Learning We first consider the NL case. Svensson and Williams (2007a) derive the equilibrium under commitment in a timeless perspective for the case when Xt, xt, and it are observable in period t, jt is unobservable, and the updating equation for pt|t is given by (7). Observations of Xt, xt, and it are then not used to update pt|t . It will be useful to replace equation (2) with the two equivalent equations, (8) Ε t H jt+1 x t +1 = zt , F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W (9) 0 = A21 jt X t + A22 jt x t − zt + B2 jt it + C2 jt ε t , where we introduce the nx vector of additional forward-looking variables, zt . Introducing this vector is a practical way of keeping track of the expectations term on the left side of (2). Furthermore, it will be practical to use (9) and solve xt as a function of Xt , zt , it , jt , and εt : (10) x t = x% ( X t , zt , it , jt , εt ) ( ) −1 ; A22 jt zt − A21 jt X t − B2 jt it − C 2 jt ε t . We note that, for given jt , this function is linear in Xt , zt , it , and εt . In order to solve for the optimal decisions, we use the recursive saddlepoint method (see Marcet and Marimon, 1998, Svensson and Williams, 2007a, and Svensson, 2007, for details of the recursive saddlepoint method). Thus, we introduce Lagrange multipliers for each forward-looking equation, the lagged values of which become state variables and reflect costs of commitment, while the current values become control variables. The dual period loss function can be written ( Ε t L% X% t , zt , it , γ t , j t , ε t ( ) ) ; ∑p jt|t ∫L% X% t , zt , it , γ t , j , ε t ϕ ( εt )dε t , j where X̃t ⬅ 共Xt′, Ξ′t –1兲′ is the 共nx + nx 兲 vector of extended predetermined variables (that is, including the nx vector Ξt–1), γt is an nx vector of Lagrange multipliers, ϕ 共.兲 denotes a generic probability density function (for εt , the standard normal density function), and ( ) L% X% t , zt , it , γ t , j t , ε t ; L X t , x% ( X t , zt , it , jt , ε t ), it , j t 1 −γ t′zt + Ξt′−1 H j t x% ( X t , zt , it , j t , ε t ) . δ (11) As discussed in Svensson and Williams (2007a), the failure of the law of iterated expectations leads us to introduce the collection of value functions, V̂共st ,j 兲, that condition on the mode, whereas the value function Ṽ共st 兲 averages over these and represents the solution of the dual optimization problem. The somewhat unusual J U LY / A U G U S T 2008 279 Svensson and Williams X t +1 st +1 ; Ξt = g ( st , zt , it , γ t , jt , εt , j t +1, ε t +1 ) pt +1|t +1 Bellman equation for the dual problem can be written (12) V% ( st ) ; ΕtVˆ ( st , jt ) ; ∑ j p jt|tVˆ ( st , j ) ( ) L% X% t , zt , it , γ t , jt , ε t = max min Εt st , zt , it , γ t , jt , γ t ( zt , it ) , j t +1 +δVˆ g ε , j , ε t t +1 t +1 L% X% t , zt ,iit , γ t , j , ε t ; max min ∑ p jt|t ∫ st , zt , it , γ t , j ˆ z , i γt ( t t) + δ ∑ kPjkV g , k j , ε t , k , εt +1 ( ) ϕ ( ε t )ϕ ( ε t +1 )dε tdε t +1, where st ⬅ 共X̃t′, p′t|t 兲′ denotes the perceived state of the economy (it includes the perceived probability distribution, pt|t, but not the true mode) and 共st, jt 兲 denotes the true state of the economy (it includes the true mode of the economy). As we discuss in more detail below, it is necessary to include the mode jt in the state vector because the beliefs do not satisfy the law of iterated expectations. In the BOP case, beliefs do satisfy this property, so the state vector is simply st . Also note that, in the Bellman equation, we require that all the choice variables respect the information constraints and thus depend on the perceived state, st , but not the mode jt directly. The optimization is subject to the transition equation for Xt , (13) X t +1 = A11 jt +1 X t + A12 jt +1 x% ( X t , zt , it , jt , εt ) + B1 jt +1 it + C1 jt +1 ε t +1, where we have substituted x̃ 共Xt, zt, it, jt, εt 兲 for xt ; the new dual transition equation for Ξt , Ξt = γ t ; (14) (15) and the transition equation (7) for pt|t . Combining equations, we have the transition for st : A11 jt +1 X t + A12 jt +1 x% ( X t , zt , it , j , ε t ) + B1 jt +1 it + C1 jt +1 ε t +1 γt ; . P ′pt|t It is straightforward to see that the solution of the dual optimization problem (12) is linear in X̃t for given pt|t, jt : Fz ( pt|t ) zt z (st ) = % (16) it i ( st ) = F ( pt|t ) X t ; Fi ( pt|t ) X% t , Fγ ( pt|t ) γ t γ (st ) (17) xt = x (st , jt , εt ) ; x% ( X t , z ( st ), i ( st ), jt , εt ) ( ) J U LY / A U G U S T 2008 ) This solution is also the solution to the original primal optimization problem. We note that xt is linear in εt for given pt|t and jt . The equilibrium transition equation is then given by (18) st +1 = gˆ ( st , j t , ε t , jt +1 , ε t +1 ) ; g st , z ( st ), i ( st ), γ (st ), j t , ε t , jt +1 , ε t +1 . As can be easily verified, the (unconditional) dual value function Ṽ共st 兲 is quadratic in X̃t for given pt|t, taking the form % V% ( st ) ; X% t′V%XX % % ( pt|t ) X t + w ( pt|t ) . The conditional dual value function, V̂共st ,jt 兲, gives the dual intertemporal loss conditional on the true state of the economy, 共st ,jt 兲. It follows that this function satisfies ( ) L% X% t , z (st ), i (st ), γ ( st ), j , ε t Vˆ ( st , j ) ; ∫ + δ ∑ k PjkVˆ gˆ ( st , j , ε t , k , ε t +1 ), k ϕ ( εt )ϕ (ε t +1 )dε tdε t +1 280 ( ; FxX% pt|t , jt X% t + Fxε pt|t , j t εt . ( j ∈ N ). j F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams The function V̂共st ,jt 兲 is also quadratic in X̃t for given pt|t and jt : % ˆ ( pt|t , j t ). Vˆ (st , jt ) ; X% t′VˆXX % % ( pt|t , jt ) X t + w ˆ % % ( p , j ), V%XX % % ( pt|t ) ; ∑ j p jt|tVXX t|t It follows that we have w ( pt|t ) ; ∑ j p jt|t wˆ ( pt|t , j ) . 1 V ( st ) ; V% ( st ) − Ξt′− 1 ∑p jt|t H j ∫x ( st , j , ε t )ϕ ( ε t )dεt δ j (19) 1 ∑pjt|t H j x (st , j , 0), δ j where the second equality follows because x共st,jt, εt 兲 is linear in εt for given st and jt. It is quadratic in X̃t for given pt|t: % V ( st ) ; X% t′VXX % % ( pt|t ) X t + w ( pt|t ) (the scalar w共pt|t 兲 in the primal value function is obviously identical to that in the dual value function). This is the value function conditional on X̃t and pt|t after Xt has been observed but before xt has been observed, taking into account that jt and εt are not observed. Hence, the second term on the right side of (19) contains the expectation of Hjtxt conditional on that information.9 Svensson and Williams (2007a,b) present algorithms to compute the solution and the primal and dual value functions for the NL case. For future reference, we note that the value function for the primal problem also satisfies 9 To be precise, the observation of Xt, which depends on C1j εt, allows some inference of εt, εt|t. xt will depend on jt and on εt, but on εt only through C2j εt. By assumption C1j εt and C2k εt are independent. Hence, any observation of Xt and C1j εt does not convey any information about C2j εt, so EtC2j εt = 0. t t t F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W where the conditional value function, V̆ 共st ,jt 兲, satisfies L X t , x ( st , j , εt ), i ( st ), j ( V ( st , j ) = ∫ ( (20) + δ ∑ k PjkV gˆ ( st , j , εt , k , εt +1 ), k ϕ ( εt )ϕ (ε t +1 )dε tdε t +1 Although we find the optimal policies from the dual problem, in order to measure true expected losses, we are interested in the value function for the primal problem (with the original, unmodified loss function). This value function, with the period loss function Et L共Xt, xt, it, jt 兲, rather than EtL̃ 共X̃t, zt, it, γt, jt , εt 兲, satisfies = V% ( st ) − Ξt′− 1 ( V ( st ) ; ∑ j p jt|tV (st , j ), ( j ∈ N ). j Adaptive Optimal Policy Consider now the case of AOP, where the policymaker uses the same policy function as in the NL case but each period updates the probabilities that this policy is conditioned on. This case is thus simple to implement recursively, as we have already discussed how to solve for the optimal decisions and below we show how to update probabilities. However, the ex ante evaluation of expected loss is more complex, as we show below. In particular, we assume that C2jt ⬅ / 0 and that both εt and jt are unobservable. The estimate pt|t is the result of Bayesian updating, using all information available, but the optimal policy in period t is computed under the perceived updating equation (7). That is, the fact that the policy choice will affect future pt+τ|t+τ and that future expected loss will change when pt+τ|t+τ changes is disregarded. Under the assumption that the expectations on the left side of (2) are conditional on (7), the variables zt, it, γt , and xt in period t are still determined by (16) and (17). In order to determine the updating equation for pt|t, we specify an explicit sequence of information revelation as follows, in no less than nine steps. The timing assumptions are necessary in order to spell out the appropriate conditioning for decisions and updating of beliefs. (i) The policymaker and the private sector enters period t with the prior pt|t–1. They know Xt–1, xt–1 = x共st–1,jt–1, εt–1兲, zt–1 = z共st–1兲, it–1 = i共st–1兲, and Ξt–1 = γ 共st–1兲 from the previous period. (ii) In the beginning of period t, the mode jt and the vector of shocks εt are realized. Then the vector of predetermined variables Xt is realized according to (1). J U LY / A U G U S T 2008 281 Svensson and Williams (iii) The policymaker and the private sector observe Xt . They then know X̃t ⬅ 共Xt ′, Ξ′t–1兲′. They do not observe jt or εt . (iv) The policymaker and the private sector update the prior pt|t–1 to the posterior pt|t according to Bayes’s theorem and the updating equation (21) p jt|t = ϕ ( X t | j t = j , X t −1, x t −1 , it −1, pt|t −1 ) p jt|t −1 ϕ ( X t | X t −1 , xt −1 , it −1, pt|t −1 ) ( j ∈ N ), The private sector and the policymaker can also infer Ξt from Ξt = γ ( st ). (23) This allows the private sector and the policymaker to form the expectations zt = z (st ) = Ε t H jt +1 x t +1|st = ∑ j , kPjk p jt|t H k x k , t +1|jt , (24) j where again ϕ 共.兲 denotes a generic density function.10 Then the policymaker and the private sector know st ⬅ 共X̃t′, p′t|t 兲′. (v) The policymaker solves the dual optimization problem, determines it = i共st 兲, and implements/announces the instrument setting, it. (vi) The private sector (and policymaker) expectations, where x k , t +1|jt ϕ (ε t )ϕ ( εt +1 )dε tdε t +1 ) ( A11k X t + A12k x ( st , j, 0 ) + B1k i ( st ) Ξt = x , k , 0 , P ′pt|t zt = Ε t H jt +1 x t +1 ; Ε H jt +1 x t +1|st , are formed. In equilibrium, these expectations will be determined by (16). In order to understand their determination better, we look at this in some detail. These expectations are by assumption formed before xt is observed. The private sector and the policymaker know that xt will in equilibrium be determined in the next step according to (17). Hence, they can form expectations of the soon-tobe determined xt conditional on jt = j,11 (22) 10 x jt|t = x ( st , j , 0) . The policymaker and private sector can also estimate the shocks εt|t as ε t t = ∑ j pjt t ε jt t , where ε jt t ; X t − A11 j X t −1 − A12 j x t −1 − B1j it −1 ( j ∈ N ). j However, because of the assumed independence of C1j εt and C2kεt, j,k 僆 Nj , we do not need to keep track of εjt|t. 11 Note that 0 instead of εjt|t enters above. This is because the inference εjt|t above is inference about C1j εt, whereas xt depends on εt through C2j εt. Because we assume that C1j εt and C2j εt are independent, there is no inference of C2j εt from observing Xt. Hence, EtC2j εt = 0. Because of the linearity of xt in εt, the integration of xt over εt results in x共st, jt, 0t 兲. where we have exploited the linearity of xt = x共st, jt, εt 兲 and xt +1 = x共st +1, jt +1, εt +1兲 in εt and εt +1. Note that zt is, under AOP, formed conditional on the belief that the probability distribution in period t +1 will be given by pt+1|t+1 = P ′pt|t , not by the true updating equation that we are about to specify. (vii) After the expectations zt have been formed, xt is determined as a function of Xt, zt, it, jt, and εt by (10). (viii) The policymaker and the private sector then use the observed xt to update pt|t to the new + posterior pt|t according to Bayes’s theorem, via the updating equation (25) p+jt|t = J U LY / A U G U S T 2008 ϕ ( xt | jt = j , X t , zt , it , pt|t ) p jt|t ϕ ( xt | X t , zt , it , pt|t ) ( j ∈ N ). j (ix) The policymaker and the private sector then leave period t and enter period t +1, with the prior pt+1|t given by the prediction equation t 282 A11k X t + A12 k x (st , j , εt ) + B1k i ( st ) , k, ε = ∫ x Ξt t +1 P ′pt|t (26) pt +1|t = P ′pt+|t . F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams In the beginning of period t +1, the mode jt +1 and the vector of shocks εt +1 are realized, and Xt +1 is determined by (1) and observed by the policymaker and private sector. The sequence of the nine steps above then repeats itself. For more detail on the explicit densities in the updating equations (21) and (25), see Svensson and Williams (2007b). The transition equation for pt+1|t+1 can be written (27) pt +1|t +1 = Q ( st , zt , it , jt , εt , j t +1, ε t +1 ), where Q共st, zt, it, jt, εt , jt +1, εt +1兲 is defined by the combination of (21) for period t +1 with (13) and (26). The equilibrium transition equation for the full state vector is then given by X t +1 st +1 ; Ξt = g (st , jt , εt , j t +1, ε t +1 ) pt +1|t +1 (28) A11 jt +1 X t + A12 jt +1 x (st , jt , εt ) + B1 jt +1 i (st ) + C1 jt +1 ε t +1 ; γ ( st ) , Q s , z s , i s , j , ε , j , ε ( t ( t ) ( t ) t t t +1 t +1 ) where the bottom block is given by the true updating equation (27) together with the policy function (16). Thus, we note that, in this AOP case, there is a distinction between the perceived transition and equilibrium transition equations, (15) and (18), which in the bottom block include the perceived updating equation, (7), and the true equilibrium transition equation, (28), which replaces the perceived updating equation, (7) with the true updating equation, (27). Note that V共st 兲 in (19), which is subject to the perceived transition equation, (15), does not give the true (unconditional) value function for the AOP case. This is instead given by ( V ( st ) ; ∑ j p jt|tV (st , j ), where the true conditional value function, V̆ 共st ,jt 兲, satisfies F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W L X t , x ( st , j , εt ), i ( st ), j ( V ( st , j ) = ∫ ( st , j , ε t , (29) + δ ∑ k PjkV g , k k , εt +1 ϕ ( εt )ϕ (ε t +1 )dε tdε t +1 ( j ∈ N ). j – That is, the true value function, V 共st 兲, takes into account the true updating equation for pt|t , (27), whereas the optimal policy and the perceived value function, V共st 兲, in (19), are conditional on the perceived updating equation, (7), and thereby the perceived transition equation, (15). Note also – that V 共st 兲 is the value function after X̃t has been observed but before xt is observed, so it is condi+ . Because the full tional on pt|t rather than pt|t transition equation, (28), is no longer linear due to the belief-updating equation, (27), the true – value function, V 共st 兲, is no longer quadratic in X̃t for given pt|t . Thus, more-complex numerical methods are required to evaluate losses in the AOP case, although policy is still determined simply as in the NL case. As we discuss in Svensson and Williams (2007b), the difference between the true updating equation for pt+1|t+1, (27), and the perceived updating equation, (7), is that, in the true updating equation, pt+1|t+1 becomes a random variable from the point of view of period t, with mean equal to pt+1|t . This is because pt+1|t+1 depends on the realization of jt +1 and εt +1. Thus Bayesian updating induces a mean-preserving spread over beliefs, which in turn sheds light on the gains from learning. If the conditional value function, V̆ 共st ,jt 兲, under NL is concave in pt|t for given X̃t and jt , then by Jensen’s inequality the true expected future loss under AOP will be lower than the true expected future loss under NL. That is, the concavity of the value function for beliefs means that learning leads to lower losses. Although it is likely that V̆ is indeed concave, as we show in the applications, it need not be globally so and thus learning need not always reduce losses. In some cases, the losses incurred by increased variability of beliefs may offset the expected precision gains. Furthermore, under BOP, it may be possible to adjust policy to further increase the variance of pt|t , that is, achieve a mean-preserving spread that might J U LY / A U G U S T 2008 283 Svensson and Williams further reduce the expected future loss.12 This amounts to optimal experimentation. Bayesian Optimal Policy Finally, we consider the BOP case, when optimal policy is determined while taking the updating equation, (27), into account. That is, we now allow the policymaker to choose it taking into account that his actions will affect pt+1|t+1, which in turn will affect future expected losses. In particular, experimentation is allowed and is optimally chosen. For the BOP case, there is hence no distinction between the perceived and true transition equation. The transition equation for the BOP case is X t +1 st +1 ; Ξt = g ( st , zt , it , γ t , jt , εt , j t +1, ε t +1 ) pt +1|t +1 (30) A11 jt +1 X t + A12 jt +1 x% ( st , zt , it , jt , εt ) + B1 jt +1 it + C1 jt +1 ε t +1 γt ; . Q ( st , zt , it , jt , ε t , j t +1, ε t +1 ) Then the dual optimization problem can be written as (12) subject to the above transition equation (30). However, in the Bayesian case, matters simplify somewhat, as we do not need to compute the conditional value functions, V̂共st ,jt 兲, which we recall were required because of the failure of the law of iterated expectations in the AOP case. We note now that the second term on the right side of (12) can be written as Ε tVˆ ( st +1, jt +1 ) ; Ε Vˆ (st +1 , jt +1 ) st . Because, in the Bayesian case, the beliefs do satisfy the law of iterated expectations, this is then the same as Ε Vˆ ( st +1 , j t +1 ) st = Ε V% (st +1 ) st . 12 Kiefer (1989) examines the properties of a value function, including concavity, under Bayesian learning for a simpler model without forward-looking variables. 284 J U LY / A U G U S T 2008 See Svensson and Williams (2007b) for a proof. Thus, the dual Bellman equation for the Bayesian optimal policy is ( (31) ) % % L X t , zt , it , γ t , jt , ε t V% ( st ) = max min Εt γt % ( zt , it ) +δV g ( st , zt , it , γ t , jt , ε t , jt +1, ε t +1 ) L% X% t , zt , it , γ t , j , εt ; max min ∑ p jt|t ∫ st , zt , it , γ t , % γt ( zt , it ) j + δ ∑ k PjkV g j , ε t , k , ε t +1 ϕ ( εt )ϕ ( εt +1 )dεtdεt +1, ( ) where the transition equation is given by (30). The solution to the optimization problem can be written (32) (33) z ( st ) zt ı%t ; it = ı% (st ) ≡ i (st ) γ ( st ) γ t Fz X% t , pt|t = F X% t , pt|t ; Fi X% t , pt|t Fγ X% t , pt|t ( ( ( ( ) ) ) , ) xt = x (st , jt , εt ) ; x% ( X t , z ( st ), i ( st ), jt , εt ) ( ) ; Fx X% t , pt|t , jt , ε t . Because of the nonlinearity of (27) and (30), the solution is no longer linear in X̃t for given pt|t . The dual value function, Ṽ共st 兲, is no longer quadratic in X̃t for given pt|t . The value function of the primal problem, V共st 兲, is given by, equivalently, (19), (29) (with the equilibrium transition equation (28) with the solution (32)), or L X t , x ( st , j , ε t ), i ( st ), j V ( st ) = ∑p jt|t ∫ (34) j + δ ∑ k PjkV g (st , j , ε t , k , ε t +1 ) ϕ (ε t )ϕ ( ε t +1 )dεtdεt +1. It is also no longer quadratic in X̃t for given pt|t . Thus, more complex and detailed numerical methods are necessary in this case to find the optimal policy and the value function. Therefore, little can be said in general about the solution of F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams the problem. Nonetheless, in numerical analysis it is very useful to have a good starting guess at a solution, which in our case comes from the AOP case. In our examples below we explain in more detail how the BOP and AOP cases differ and what drives the differences. Observable Modes In this paper we largely focus on the cases where the policymakers do not observe the current mode, which is certainly the more relevant case when analyzing model uncertainty. However, some situations may arguably be better modeled by observable shifts in modes, as in most of the econometric literature on regime-switching models. Moreover, one way to gauge the effects of uncertainty in a model is to move from a constantcoefficient specification to one in which the parameters are observable but may vary. (That is, the current values of parameters are known, but future values are uncertain.) For this reason, we use the observable mode case, to analyze implications of uncertainty on policy. In Svensson and Williams (2007a), we develop simple algorithms for observable changes in modes, which play off the fact that conditional on the mode the evolution of the economy is linear and preferences are quadratic. Thus, the optimal policy consists of a mode-dependent collection of linear policy rules and can be written it = Fijt X% t (35) for jt 僆 Nj . LEARNING AND EXPERIMENTATION IN A SIMPLE NEW KEYNESIAN MODEL The Model For our policy exercises, we consider a benchmark hybrid New Keynesian Phillips curve (see Woodford, 2003, for an exposition): ( ) (36) π t = 1 − ω jt π t −1 + ω jt Ε t π t +1 + γ jt y t + cε t . F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Here πt is the inflation rate, yt is the output gap, ωjt is a parameter reflecting the degree of forwardlooking behavior in price setting, and γjt is a composite parameter reflecting the elasticity of demand and frequency of price adjustment. For simplicity, we assume that policymakers can directly control the output gap, yt . In another paper, Svensson and Williams (2008), we consider optimal policy in the standard two-equation New Keynesian model that also includes a loglinearized consumption Euler equation. Many of the same issues that we focus on here arise there as well, but the simpler setting in the present paper allows us to focus more directly on the effects of uncertainty on policy. We focus on two key sources of uncertainty in the New Keynesian Phillips curve. Our first example considers the degree of forward-looking behavior in inflation. In the model, this translates to uncertainty about ωj . If this parameter is large, inflation is largely determined by current shocks and expectations of the future, whereas if ωj is small, then there is a substantial exogenous inertia in the inflation process. Our second example analyzes uncertainty about the slope of the Phillips curve, as reflected in the parameter γj . This could reflect changes in the degree of monopolistic competition (which also lead to varying markups) and/or changes in the degree of price stickiness. In each example, we look first at the effect of uncertainty, going from a constant-coefficient model to a model with random coefficients. Then, we analyze the effects of learning and experimentation on policy and losses. In both examples, we use the following loss function: (37) Lt = π t2 + λ y t2 . We set the loss-function parameters as δ = 0.98, λ = 0.1, and set the shock standard deviation to c = 0.5. Even though different structural parameters vary in the two examples, both examples use two modes and set the transition matrix to 0.98 0.02 P= . 0.02 0.98 J U LY / A U G U S T 2008 285 Svensson and Williams Figure 1 Policies and Losses from Observable and Constant Modes Policy: Observable and Constant Modes Loss: Observable and Constant Modes yt Loss 8 35 6 30 4 25 2 20 0 −2 Obs 1 15 −4 Obs 2 10 −6 E(Obs) −8 Constant −5 5 0 5 0 −5 0 5 πt−1 πt−1 NOTE: Obs 1 (2) is observable mode 1 (2); E(Obs) is the unconditional average policy. In both examples, we examine the value functions and optimal policies for this simple New Keynesian model under NL, AOP, and BOP. We have one forward-looking variable (xt ⬅ πt ) and consequently one Lagrange multiplier (Ξt–1 ⬅ Ξπ,t–1). We have one predetermined variable (Xt ⬅ πt–1) and the estimated mode probabilities (pt|t ⬅ 共p1t|t , p2t|t 兲′, of which we only need keep track of one, p1t|t ). Thus, the value and policy functions, V共st 兲 and i共st 兲, are all three dimensional (st = 共 πt–1, Ξπ,t–1, p1|t 兲′). For computational reasons, we are forced to restrict attention to relatively sparse grids with few points. The following plots show two-dimensional slices of the value and policy functions, focusing on the dependence on πt–1 and p1t|t (which we for simplicity denote by p1|t in the figures). In particular, all of the plots are for Ξπ,t–1 = 0. Example 1: How Forward-Looking Is Inflation? This example analyzes one of the main sources of uncertainty in the New Keynesian framework— the degree to which inflation is a forward-looking variable responding to expectations of future developments. Specifications that suggest that 286 J U LY / A U G U S T 2008 inflation has substantial exogenous persistence have tended to fit better empirically, while perhaps being less rigorous in their micro-foundations. In this example, we see how uncertainty about the degree of forward-looking behavior, as indexed by ωj, affects policy. Thus, we assume that there are two modes, one more forward looking, with ω1 = 0.8, and the other more backward looking, with ω2 = 0.2. Note that, with the transition matrix P as specified above, this means E共ωj 兲 = 0.5. For this example, we fix the slope parameter at γ = 0.1. In Figure 1, we illustrate the effects of uncertainty on policy and losses. In the left panel, we plot the two mode-dependent optimal policy functions for the MJLQ model with observable modes, labeled “Obs 1” for mode 1 and “Obs 2” for mode 2. Here, we see that the optimal policy is more aggressive in the more backward-looking mode 2, because in response to a higher inflation the optimal policy involves larger negative output gaps. The unconditional average policy is labeled “E(Obs)” and shown with a gray line. For comparison, the constant-coefficient case, where we set ω1 = ω2 = E共ωj 兲 = 0.5, is plotted with a black dashed line. Here, we see that optimal policy under uncertainty is more aggressive in respondF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams Figure 2 Losses and Differences in Losses from NL, AOP, and BOP Loss: BOP Loss: NL Loss Loss πt−1 = 0 πt−1 = −5 πt−1 = 3.33 80 70 80 70 60 60 50 50 40 40 0.2 30 0.2 0.4 0.6 0.4 0.8 0.6 0.8 p1t p1t Loss Differences: AOP−NL Loss Loss Differences: BOP−AOP Loss 1.5 −2 −3 1.0 −4 −5 0.5 −6 0 0.2 0.4 0.6 0.8 p1t ing to inflation movements than optimal policy in the absence of uncertainty. A common starting point for thinking about the effects of uncertainty on policy is Brainard’s (1967) classic analysis, which suggested that uncertainty should make policy more cautious. However, Brainard worked in a static framework and the source of uncertainty he analyzed was a slope coefficient on how policy affects the economy. Our second example below is closer to Brainard’s and comes to similar conclusions. But, in this example, our results suggest, at least for this parameterization, that uncertainty about the dynamics of inflation leads to more-aggressive policy. This is similar to what Söderström (2002) found in a backward-looking model. The right panel of Figure 1 plots the losses associated with the optimal policies in the different cases. When inflation is more forward looking, it is easier to control and so overall losses are F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W 0.2 0.4 0.6 0.8 p1t lower even with less-aggressive policies. However, uncertainty about the dynamics of inflation can have significant effects on losses for moderate to high inflation levels. This is evident by comparing the constant-coefficient and average observable curves, where we see that the loss nearly doubles at the edges of the plot. Now we keep the same specification, but make the more realistic assumption that the current mode is not observed. Thus, we analyze the effects of learning and experimentation on policy and losses. The top-two panels of Figure 2 show losses under NL and BOP as functions of p1t . The bottomtwo panels of the figure show the differences between losses under NL, AOP, and BOP. Figure 3 shows the corresponding policy functions and their differences. The top-two panels plot the policy functions under AOP and BOP as a function of inflation. The AOP policy is linear in πt , and clearly the BOP policy is nearly so. The botJ U LY / A U G U S T 2008 287 Svensson and Williams Figure 3 Optimal Policies and Their Differences Under AOP and BOP Policy: AOP Policy: BOP yt yt 5 5 p1t = 0.89 p1t = 0.5 p1t = 0.11 0 0 −5 −5 −5 −5 0 5 0 5 πt−1 πt−1 Policy Differences: BOP−AOP Policy: BOP yt yt πt−1 = 0 πt−1 = −5 πt−1 = 3.33 5 3 2 1 0 0 −1 −2 −5 0.2 0.4 0.6 0.8 p1t tom-left panel plots the BOP policy as a function of p1t , showing that policy is less aggressive (that is, has a smaller magnitude of response) the greater is the probability of being in the more forwardlooking mode 1. The bottom-right panel shows that the policy differences between AOP and BOP, the experimentation component of policy, are incredibly small. In Svensson and Williams (2007b), we show that learning implies a mean-preserving spread of the random variable pt +1|t +1 (which under learning is a random variable from the vantage point of period t). Hence, concavity of the value function under NL in p1t implies that learning is beneficial, because then a mean-preserving spread reduces the expected future loss. However, we see in Figure 2 that the value function is actually slightly convex in p1t , so learning is not beneficial here. Consequently, we see in Figure 2 that AOP gives higher losses than NL. In contrast, for a backward-looking example in Svensson and 288 J U LY / A U G U S T 2008 −3 −5 0 5 πt−1 Williams (2007b), the value function is concave and learning is beneficial. Experimentation is beneficial here, as BOP does give lower losses than AOP, but the difference is minuscule. So, for this example, learning has sizable effects on losses and is detrimental, whereas experimentation is beneficial but has negligible effects. Why would learning not be beneficial with forward-looking variables? It may at least partially be a remnant of our assumption of symmetric beliefs and information between the private sector and the policymaker. With backward-looking models, we have generally found that learning is beneficial. However, under our assumption of symmetric information and beliefs between the private sector and the policymaker, both the private sector and the policymaker learn. The difference between backward- and forward-looking models then comes from the way that private sector beliefs also respond to learning. Having more reactive private sector beliefs may add volatility F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams Figure 4 Policies and Losses from Observable and Constant Modes Policy: Observable and Constant Modes yt Loss: Observable and Constant Modes Loss 4 3 15 2 1 10 0 −1 Obs 1 −2 Obs 2 −3 −4 −5 5 E(Obs) Constant 0 5 0 −5 0 5 πt−1 πt−1 NOTE: Obs 1 (2) is observable mode 1 (2); E(Obs) is the unconditional average policy. and make it more difficult for the policymaker to stabilize the economy. Example 2: What Is the Slope of the Phillips Curve? This example analyzes the other main source of uncertainty in the New Keynesian Phillips curve—the extent to which inflation responds to fluctuations in the output gap. Once again, we assume that there are two modes: Now one has a Phillips curve that is flatter, with γ1 = 0.05, and the other has a steeper curve, with γ2 = 0.25. Note that with the transition matrix P as specified above, this means E共γj 兲 = 0.15. For this example, we fix the forward-looking expectations parameter at ω = 0.5. Because policymakers once again directly control the output gap, this example is a forward-looking counterpart to the classic Brainard (1967) analysis of uncertainty about the effectiveness of the control. In Figure 4 we illustrate the effects of uncertainty on policy and losses. As in the previous example, the left panel plots the two modedependent optimal policy functions for the MJLQ model with observable modes. Here, we see that the MJLQ optimal policies in both modes are less F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W aggressive than the constant-coefficient case. Thus, our results here are in accord with Brainard’s— uncertainty about the slope of the Phillips curve leads to more cautious policy. The right panel of Figure 4 plots the losses associated with the optimal policies in the different cases. When the Phillips curve is steeper, inflation responds more to the output gap, making inflation easier to control. Thus, overall losses are lower in mode 2, even with less-aggressive policies. However, once again uncertainty about this key parameter can have significant effects on losses for high inflation levels. This is evident by comparing the constant-coefficient and average observable curves, where we see that the loss nearly doubles at the edges of the plot. Now we again keep the same specification, but make the more realistic assumption that the current mode is not observed. The top-two panels of Figure 5 show losses under NL and BOP as functions of p1t . The bottom-two panels of the figure show the differences between losses under NL, AOP, and BOP. We see in Figure 2 that the value function is once again slightly convex in p1t , so learning is not beneficial here. Consequently, we see in the bottom-right panel of Figure 2 that AOP gives higher losses than NL. Thus, once J U LY / A U G U S T 2008 289 Svensson and Williams Figure 5 Losses and Differences in Losses from NL, AOP, and BOP Loss: BOP Loss: NL Loss Loss πt−1 = 0 πt−1 = −2 πt−1 = 3 28 26 28 26 24 24 22 22 20 20 0.2 0.2 0.4 0.6 0.4 0.8 0.6 0.8 p1t p1t Loss Differences: AOP−NL Loss Loss Differences: BOP−AOP Loss 0.7 −0.015 0.6 0.5 −0.02 0.4 0.3 −0.025 0.2 0.4 0.6 0.8 p1t again, the additional volatility outweighs the improved inference and makes learning detrimental in this example. Experimentation is once again beneficial, as BOP gives lower losses than AOP. And, while the effects of experimentation are an order of magnitude smaller than the effects of learning, the gains from recognizing the endogeneity of information are nonnegligible here. Thus, for uncertainty about the slope of the Phillips curve, policymakers may have an incentive to experiment—that is, to take actions to mitigate future uncertainty. Figure 6 shows the corresponding policy functions and their differences. The top-two panels plot the policy functions under AOP and BOP as a function of inflation. The AOP policy is linear in πt –1, and clearly the BOP policy is nearly so, although some differences are evident at the edge of the plot. The bottom-left panel plots the BOP policy as a function of p1t , showing that 290 J U LY / A U G U S T 2008 0.2 0.4 0.6 0.8 p1t the policy function is relatively flat in this dimension. The bottom-right panel plots the difference between the AOP and BOP policy functions, which shows that here the experimentation motive leads toward less-aggressive policy. This is counter to an example in Svensson and Williams (2007b), where we show that in a backward-looking model experimentation may lead to more-aggressive policy. There, policy makes outcomes more dispersed in order to sharpen inference over the modes. However, here, because learning is detrimental, the experimentation component of policy seeks to slow the effects of learning by making outcomes less dispersed. This serves to illustrate that the experimentation component of policy need not be associated with wild or aggressive policy action, but rather it optimally takes into account how information influences the targets of policy. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams Figure 6 Optimal Policies and Their Differences Under AOP and BOP Policy: AOP Policy: BOP yt yt 4 4 2 2 0 0 −2 −2 p1t = 0.92 p1t = 0.5 p1t = 0.08 −4 −4 −5 −5 0 5 0 5 πt−1 πt−1 Policy Differences: BOP−AOP Policy: BOP yt yt 0.4 1 0.2 0 0 πt−1 = 0 πt−1 = −2 πt−1 = 3 −1 −2 −0.2 −0.4 0.2 0.4 0.6 0.8 p1t CONCLUSION In this paper, we have presented a relatively general framework for analyzing model uncertainty and the interactions between learning and optimization. Although this is a classic issue, very little to date has been done for systems with forward-looking variables, which are essential elements of modern models for policy analysis. Our specification is general enough to cover many practical cases of interest, yet remains relatively tractable in implementation. This is definitely true for cases when decisionmakers do not learn from the data they observe (our case of no learning, NL) or when they do learn but do not account for learning in optimization (our case of adaptive optimal policy, AOP). In both of these cases, we have developed efficient algorithms for solving for the optimal policy, which can handle relatively large models with multiple modes and many state variables. However, in the case of the F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W −5 0 5 πt−1 Bayesian optimal policy (BOP), where the experimentation motive is taken into account, we must solve more-complex numerical dynamic programming problems. Thus, to fully examine optimal experimentation, we are haunted by the curse of dimensionality, forcing us to study relatively small and simple models. Thus, an issue of much practical importance is the size of the experimentation component of policy and the losses entailed by abstracting from it. Although our results in this paper are far from comprehensive, they suggest that in practical settings the experimentation motive may not be a concern. The above and similar examples that we have considered indicate that the benefits of learning (moving from NL to AOP) may be substantial, whereas the benefits from experimentation (moving from AOP to BOP) are modest or even insignificant. If this preliminary finding stands up to scrutiny, experimentation in economic policy in general and monetary policy in J U LY / A U G U S T 2008 291 Svensson and Williams particular may not be very beneficial, in which case there is little need to face the difficult ethical and other issues involved in conscious experimentation in economic policy. Furthermore, the AOP is much easier to compute and implement than the BOP. To have this truly be a robust implication, more simulations and cases need to be examined. Beck, Günter W. and Wieland, Volker. “Learning and Control in a Changing Economic Environment.” Journal of Economic Dynamics and Control, August 2002, 25(9-10), pp. 1359-77. Blake, Andrew P. and Zampolli, Fabrizio. “Time Consistent Policy in Markov Switching Models with Rational Expectations.” Working Paper No. 298, Bank of England, 2006. Brainard, William C. “Uncertainty and the Effectiveness of Policy.” American Economic Review, May 1967, 57(2), pp. 411-25. Cogley, Timothy; Colacito, Riccardo and Sargent, Thomas J. “The Benefits from U.S. Monetary Policy Experimentation in the Days of Samuelson and Solow and Lucas.” Journal of Money, Credit, and Banking, February 2007, 39(2), pp. 67-99. Costa, Oswaldo L.V.; Fragoso, Marecelo D. and Marques, Ricardo P. Discrete-Time Markov Jump Linear Systems. London: Springer, 2005. do Val, João B.R. and Başar, Tamer. “Receding Horizon Control of Jump Linear Systems and a Macroeconomic Policy Problem.” Journal of Economic Dynamics and Control, August 1999, 23(8), pp. 1099-31. Ellison, Martin. “The Learning Cost of Interest Rate Reversals.” Journal of Monetary Economics, November 2006, 53(8), pp. 1895-907. Ellison, Martin and Valla, Natacha. “Learning, Uncertainty and Central Bank Activism in an Economy with Strategic Interactions.” Journal of Monetary Economics, August 2001, 48(1), pp. 153-71. J U LY / A U G U S T Kiefer, Nicholas M. “A Value Function Arising in the Economics of Information.” Journal of Economic Dynamics and Control, April 1989, 13(2), pp. 201-23. Lindé, Jesper. “Estimating New-Keynesian Phillips Curves: A Full Information Maximum Likelihood Approach.” Journal of Monetary Economics, September 2005, 52(6), pp. 1135-49. REFERENCES 292 Evans, George and Honkapohja, Seppo. Learning and Expectations in Macroeconomics. Princeton, NJ: Princeton University Press, 2001. 2008 Marcet, Albert and Marimon, Ramon. “Recursive Contracts.” Working paper, Universitat Pompeu Fabra, Department of Economics and Business, 1998; www.econ.upf.edu. Poole, William. “A Policymaker Confronts Uncertainty.” Federal Reserve Bank of St. Louis Review, September 1998, 80(5), pp. 3-8. Rudebusch, Glenn D. and Svensson, Lars E.O. “Policy Rules for Inflation Targeting,” in John B. Taylor, ed., Monetary Policy Rules. Chicago: University of Chicago Press, 1999. Söderström, Ulf. “Monetary Policy with Uncertain Parameters.” Scandinavian Journal of Economics, 2002, 104(1), pp. 125-45. Svensson, Lars E.O. “Optimization under Commitment and Discretion, the Recursive Saddlepoint Method, and Targeting Rules and Instrument Rules.” Lecture notes, Princeton University, 2007; www.princeton.edu/svensson. Svensson, Lars E.O. and Williams, Noah. “Monetary Policy with Model Uncertainty: Distribution Forecast Targeting.” Working paper, Princeton University, May 2007a; www.princeton.edu/svensson/. Svensson, Lars E.O. and Williams, Noah. “Bayesian and Adaptive Optimal Policy Under Model Uncertainty.” NBER Working Paper No. 13414, National Bureau of Economic Research, 2007b. Svensson, Lars E.O. and Williams, Noah. “Optimal Monetary Policy in DSGE Models: A Markov JumpLinear-Quadratic Approach.” NBER Working Paper F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Svensson and Williams No. W13892, National Bureau of Economic Research, 2008. Tesfaselassie, Mewael F.; Schaling, Eric and Eijffinger, Sylvester C.W. “Learning about the Term Structure and Optimal Rules for Inflation Targeting.” CEPR Discussion Paper No. 5896, Centre for Economic Policy Research, 2006. Wieland, Volker. “Learning by Doing and the Value of Optimal Experimentation.” Journal of Economic Dynamics and Control, March 2000, 24(4), pp. 501-34. Wieland, Volker. “Monetary Policy and Uncertainty about the Natural Unemployment Rate: BrainardStyle Conservatism versus Experimental Activism.” Advances in Macroeconomics, March 2006, 6(1), pp. 1-34. Woodford, Michael. Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton, NJ: Princeton University Press, 2003. Zampolli, Fabrizio. “Optimal Monetary Policy in a Regime-Switching Economy: The Response to Abrupt Shifts in Exchange-Rate Dynamics.” Working Paper No. 297, Bank of England, 2006. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 293 294 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Timothy W. Cogley W illiam Poole has made a number of fundamental contributions to the theory and practice of monetary policy during his long and productive career. Among other things, Poole has long emphasized the importance of uncertainty in shaping monetary policy. Uncertainty takes many forms. The central bank must act in anticipation of future conditions, which are currently unknown. Because economists have not formed a consensus about the best way to model the monetary transmission mechanism, policymakers must also contemplate alternative theories with distinctive operating characteristics. Finally, even economists who agree on a modeling strategy sometimes disagree about the values of key parameters. Thus, central bankers must also confront parameter uncertainty within macroeconomic models. Addressing all these sources of uncertainty is a tall order, but economists have made considerable progress. Lars Svensson and Noah Williams are in the vanguard. In a series of important papers, they adapt and extend Markov jumplinear-quadratic (MJLQ) control algorithms so that they are suitable for application to monetary policy.1 Among other things, they extend MJLQ algorithms to handle forward-looking models and show how to design optimal policies under commitment. Their contribution to this volume (Svensson and Williams, 2008) provides a concise 1 See Svensson and Williams (2007a,b and 2008). technical introduction to their work and also describes a pair of thoughtful and well-designed examples that illustrate how uncertainty about the monetary transmission mechanism influences optimal policy. One lesson that emerges from their examples is that the benefits of learning are often substantial but that the gains from deliberate experimentation are slight. In their parlance, an “adaptive optimal policy” is almost as good as the fully optimal Bayesian policy. ATTITUDES ABOUT POLICY EXPERIMENTATION My comment focuses on the role of experimentation. A natural way to address parameter and/or model uncertainty is to cast an optimal policy problem as a Bayesian decision problem. The decisionmaker’s posterior distribution over unknown parameters and/or model probabilities becomes part of the state vector, and Bayes’s law becomes part of the transition equation. Because Bayes’s law is nonlinear, this breaks certainty equivalence,2 making the decision rule nonlinear. A Bellman equation instructs the decisionmaker to vary the policy instrument in order to generate information about unknown parameters and model probabilities. Hence, policymakers have an 2 Certainty equivalence would hold if the central bank’s objective function were quadratic and the transition equation were linear. The presence of Bayes’s law as a component of the transition equation makes it nonlinear and hence breaks certainty equivalence. Timothy W. Cogley is a professor of economics at the University of California, Davis. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 295-300. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 295 Cogley incentive to experiment to tighten that posterior in the future. Although experimentation causes near-term outcomes to deteriorate, it speeds learning and improves outcomes in the longer run. Whether the decisionmaker should experiment a little or a lot is unclear, but it is clear that a Bayesian policy should include some deliberate experimentation. Yet there is much aversion to deliberate experimentation among macroeconomists and policymakers. For instance, Robert Lucas (1981, p. 288) writes, Social experiments on the grand scale may be instructive and admirable, but they are best admired at a distance. The idea...must be to gain some confidence that the component parts of the program are in some sense reliable prior to running it at the expense of our neighbors. Alan Blinder (1998, p. 11) concurs, asserting that while there are some fairly sophisticated techniques for dealing with parameter uncertainty in optimal control models with learning, those methods have not attracted the attention of... policymakers. There is a good reason for this inattention, I think: You don’t conduct policy experiments on a real economy solely to sharpen your econometric estimates. One way to make sense of these conflicting attitudes is to invoke Milton Friedman’s precept that the best should not be an enemy of the good. According to Svensson and Williams, a good baseline policy involves learning but not deliberate experimentation. In principle, optimal experiments can improve on this baseline policy, but optimal experiments are hard to design because the policymaker’s Bellman equation is difficult to solve, the chief obstacle being the curse of dimensionality. Because Bellman equations for policy-relevant models are hard to solve, actual policy experiments are unlikely to be optimal. And although optimal experiments are guaranteed to be no worse than the “learn but don’t experiment” benchmark, suboptimal experiments are not. Indeed, they might be much worse. Perhaps this is what Lucas had in mind when he deprecated “grand” policy experiments.3 296 J U LY / A U G U S T 2008 Svensson and Williams have made substantial progress improving algorithms for solving Bayesian optimal policy problems. Without disparaging this contribution, my sense is that the curse of dimensionality will continue to be a significant barrier in practice. In view of this, their finding that the maximum benefit of experimentation is slight takes on greater importance, for it strengthens the case in favor of adaptive optimal policies. Their findings are example specific, but they are consistent with other examples in the literature. More examples would help clinch the argument. ANOTHER EXAMPLE Cogley, Colacito, and Sargent (2007; CCS) examine a central bank’s incentive to experiment in the context of two models of the Phillips curve. One model follows Samuelson and Solow (1960) and assumes an exploitable inflation-unemployment tradeoff. The other is inspired by Lucas (1972 and 1973) and Sargent (1973) and represents a rational expectations version of the natural rate hypothesis. Based on data through the mid-1960s, CCS estimate the following two specifications: Samuelson and Solow: U t = .0023 + .7971U t −1 − .2761πt + .0054η1,t π t = vt −1 + .0055η3t . U t = .0007 + .8468U t −1 − .2489 (πt − v t −1 ) + .0055η2,t Lucas and Sargent: πt = v t −1 + .0055η4t . The variable Ut represents the unemployment gap (i.e., the difference between actual unemployment and the natural rate), πt is inflation, vt –1 is programmed or expected inflation for period t con3 One of the initial objectives of Cogley, Colacito, and Sargent (2007) was to assess whether the Great Inflation could be interpreted as an optimal experiment. We found that it could not. At least in our model, optimal experiments did not generate a decade-long surge in inflation. On the contrary, they generated small, cyclically opportunistic perturbations of inflation relative to an adaptive, non-experimental policy. Whether the Great Inflation was initiated by a suboptimal policy experiment remains an open question. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Cogley Figure 1 Two Decision Rules α≈0 α = 0.2 Programmed Inflation 0.06 Programmed Inflation 0.06 0.04 0.04 0.02 0.02 0 0 –0.02 –0.02 –0.01 0 0.01 0.02 0.03 –0.01 Lagged Unemployment 0 0.01 0.02 0.03 Lagged Unemployment α = 0.4 α = 0.6 Programmed Inflation Programmed Inflation 0.06 0.06 0.04 0.04 0.02 0.02 0 0 –0.02 –0.02 –0.01 0 0.01 0.02 0.03 –0.01 Lagged Unemployment 0 0.01 0.02 0.03 Lagged Unemployment α = 0.8 α≈1 Programmed Inflation Programmed Inflation 0.06 0.06 0.04 0.04 0.02 0.02 0 0 –0.02 –0.02 –0.01 0 0.01 0.02 0.03 Lagged Unemployment ditioned on t –1 information, and ηit , i = 1,…,4, are standard normal innovations. CCS assume that one of these specifications is true but that the central bank does not know which one. As in Svensson and Williams (2008), the central bank formulates policy by solving Bayesian and adaptive optimal control problems. The key unknown parameter is the posterior probability, α, on the Samuelson and Solow model. This probability is updated every period in accordance with Bayes’s law. The cental bank minimizes a discounted quadratic loss function subject to the “natural” transition equations for F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W –0.01 0 0.01 0.02 0.03 Lagged Unemployment the two models and also a transition equation for α. The state vector consists of lagged unemployment and the posterior model probability, α, and the control variable is programmed inflation. For the adaptive policy, the central bank updates α every period, but then treats the current estimate as if it would remain constant forever. Thus, for adaptive control problem, the α transition equation is αt + j = αt ∀j ≥ 0. Because the other transition equations are also linear and the loss function is quadratic, it follows J U LY / A U G U S T 2008 297 Cogley Figure 2 Two Hard-to-Distinguish Value Functions Value Function –0.005 –0.01 –0.015 –0.02 –0.025 –0.03 –0.035 0.03 0.02 0.01 Unemployment (U) 0 –0.01 –0.02 0 that certainty equivalence holds and that the policy rule is linear. The thin gray lines in Figure 1 illustrate how programmed inflation is set as a function of α and lagged unemployment. Each panel refers to a different value for α , model uncertainty being most pronounced for α ⬇ 0.4. Lagged unemployment is shown on the x-axis in each panel, and programmed inflation is on the y-axis. Except when α is close to zero (the central bank puts high probability on the Lucas and Sargent model), programmed inflation is countercyclical. For the Bayesian optimal policy, the central bank recognizes that actions taken today influence future beliefs about the two models. Hence the α -transition equation is governed by Bayes’s law, αt = B (αt −1 , st ), where st represents the “natural” state variables for the two models. The thick blue lines in Figure 1 depict the Bayesian decision rule. For the most part, they differ only slightly from the adaptive optimal policy. The chief difference is 298 J U LY / A U G U S T 2008 0.2 0.4 0.6 0.8 1 Prior on Samuelson and Solow (α ) that the Bayesian policy is cyclically opportunistic when there is a lot of model uncertainty. When α ⬇ 0.4, the Bayesian policy calls for higher (lower) programmed inflation relative to the adaptive optimal policy when unemployment is high (low). In other words, a recession is the best time to experiment with Keynesian stimulus and a boom is the best time to experiment with disinflation. Because the two policy functions are so alike, it is not surprising that the benefits of deliberate experimentation are small. Figure 2 portrays the value functions associated with the adaptive and Bayesian policy rules. Because the adaptive policy is not optimal, it follows that VB 共s,α 兲 ≥ VB 共s;α 兲, with the discrepancy measuring the benefits of deliberate experimentation. However, the differences are so slight that they cannot be detected in the figure. Thus, the results of CCS agree with those of Svensson and Williams.4 4 Other aspects of monetary policy experimentation are examined by Wieland (2000a,b) and Beck and Wieland (2002). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Cogley WHY THE BENEFITS OF EXPERIMENTATION ARE SLIGHT Deliberate experiments are substitutes for natural experiments. Hence, the incidence of natural experiments arising from exogenous shocks influences the value of intentional experiments. In the CCS example, one reason why the adaptive policy well approximates the Bayesian policy is that enough natural experimentation occurs for the central bank eventually to learn the true model under the adaptive policy.5 Deliberate experimentation would speed learning, but not alter the limit point. In other models, such as Kasa (1999), there isn’t enough natural experimentation to learn the truth in the absence of intentional experimentation. In those environments, deliberate experimentation would alter not only the transition path but also the limit point of the learning process. Presumably that would enhance the value of deliberate experimentation, for in that case the central bank would collect dividends on experimentation forever. Another reason why the benefits of experimentation are small is that Bayesian updating makes posterior model probabilities a martingale (Doob, 1948), implying Et共αt+j 兲 = αt . Thus, the adaptive transition equation well approximates the center of the Bayesian predictive density for αt . The adaptive model poorly approximates its tails, however, because it disregards uncertainty about future model probabilities. Nevertheless, when precautionary motives are weak, decisions depend mostly on the mean, and errors in approximating the tails don’t matter much. In these examples, the central bank’s loss function is quadratic, so precautionary motives do not enter through preferences. Precautionary behavior comes in only because of nonlinearity in the transition equation. Accordingly, motives for experimentation might be strengthened by altering the central bank’s objective function. In principle, one way to reinforce precautionary motives is by introducing a concern for robustness. Building on work by Hansen and Sargent 5 El-Gamal and Sundaram (1993) highlight the importance of natural experiments. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W (2007), Cogley et al. (2008) replace the expectations operators that appear in a Bayesian value function with a pair of risk-sensitivity operators. One risk-sensitivity operator guards against misspecification of each of the submodels, and the other guards against misspecification of the central bank’s prior. The two risk-sensitivity operators can be interpreted as ways of seeking robustness with respect to forward- and backward-looking features of the model, respectively. Applying these operators to the Phillips curve models examined in CCS, Cogley et al. find that the forwardlooking risk-sensitivity operator strengthens experimental motives, whereas the backwardlooking operator mutes them. The combined effect is ambiguous and depends on the relative weight placed on the two operators. CONCLUSION Designing an optimal policy is substantially more complex when experimental motives are active. That easy-to-compute, nonexperimental policies well approximate hard-to-compute, fullyoptimal policies is an important result. If this conclusion holds up to further scrutiny, the analysis of monetary policy under model uncertainty will be greatly simplified. In this instance, it seems that “the good” is an excellent substitute for “the best.” REFERENCES Beck, Günter and Wieland, Volker. “Learning and Control in a Changing Environment.” Journal of Economic Dynamics and Control, November 2002, 26(9/10), pp. 1359-77. Blinder, Alan S. Central Banking in Theory and Practice. Cambridge, MA: MIT Press, 1998. Cogley, Timothy; Colacito, Riccardo and Sargent, Thomas J. “Benefits from U.S. Monetary Policy Experimentation in the Days of Samuelson and Solow and Lucas.” Journal of Money, Credit, and Banking, February 2007(Suppl.), 39, pp. 67-99. J U LY / A U G U S T 2008 299 Cogley Cogley, Timothy; Colacito, Riccardo; Hansen, Lars P. and Sargent, Thomas J. “Robustness and U.S. Monetary Policy Experimentation.” Journal of Money, Credit, and Banking, 2008 (forthcoming). Doob, Joseph L. “Application of the Theory of Martingales.” Colloques Internationaux du Centre National de la Recherché Scientifique, 1948, 36, pp. 23-27. El-Gamal, Mahmoud A. and Sundaram, Rangarajan K. “Bayesian Economists...Bayesian Agents: An Alternative Approach to Optimal Learning.” Journal of Economic Dynamics and Control, May 1993, 17(3), pp. 355-83. Hansen, Lars P. and Sargent, Thomas J. “Robust Estimation and Control without Commitment.” Journal of Economic Theory, September 2007, 136(1), pp. 1-27. Kasa, Kenneth. “Will the Fed Ever Learn?” Journal of Macroeconomics, Spring 1999, 21(2), pp. 279-92. Lucas, Robert E. Jr. “Expectations and the Neutrality of Money.” Journal of Economic Theory, April 1972, 4(2), pp. 103-24. Lucas, Robert E. Jr. “Some International Evidence on Output-Inflation Trade-Offs.” American Economic Review, June 1973, 63(3), pp. 326-34. Lucas, Robert E. Jr. “Methods and Problems in Business Cycle Theory,” in Robert E. Lucas Jr., ed., Studies in Business-Cycle Theory. Cambridge, MA: MIT Press, 1981. 300 J U LY / A U G U S T 2008 Samuelson, Paul A. and Solow, Robert M. “Analytical Aspects of Anti-Inflation Policy.” American Economic Review, May 1960, 50(2), pp. 177-84. Sargent, Thomas J. “Rational Expectations, the Real Rate of Interest, and the Natural Rate of Unemployment.” Brookings Papers on Economic Activity, 1973, Issue 2, pp. 429-72. Svensson, Lars E.O. and Williams, Noah. “Monetary Policy with Model Uncertainty: Distribution Forecast Targeting.” Working paper, Princeton University, 2007a; www.princeton.edu/~svensson. Svensson, Lars E.O. and Williams, Noah. “Bayesian and Adaptive Optimal Policy Under Model Uncertainty.” Working paper, Princeton University, 2007b; www.princeton.edu/~svensson. Svensson, Lars E.O. and Williams, Noah. “Optimal Monetary Under Uncertainty: A Markov JumpLinear-Quadratic Approach.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 275-93. Wieland, Volker. “Monetary Policy, Parameter Uncertainty, and Optimal Learning.” Journal of Monetary Economics, August 2000a, 46(1), pp. 199-228. Wieland, Volker. “Learning by Doing and the Value of Optimal Experimentation.” Journal of Economic Dynamics and Control, April 2000b, 24(4), pp. 501-34. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Andrew T. Levin O ver the past decade or so, researchers at academic institutions and central banks have been active in specifying and estimating dynamic stochastic general equilibrium (DSGE) models that can be used to analyze monetary policy.1 Although the first-generation models were relatively small and stylized, more recent models typically embed a much more elaborate dynamic structure aimed at capturing key aspects of the aggregate data.2 Indeed, several central banks now use DSGE models in the forecasting process and in formulating and communicating policy strategies.3 In following that approach, however, it is crucial to investigate the sensitivity of the optimal policy prescriptions of a given model—that is, comparing the policy implications of alternative specifications of the behavioral mechanisms or exogenous shocks—and to identify policy strategies that provide robust performance under model uncertainty. The authors’ paper (Svensson and Williams, 2008) makes an important contribution in analyzing Bayesian optimal monetary policy in an envi1 Pioneering early studies include King and Wolman (1996, 1999), Goodfriend and King (1997), Rotemberg and Woodford (1997, 1999), Clarida, Galí, and Gertler (1999), and McCallum and Nelson (1999). 2 See Christiano, Eichenbaum, and Evans (2005), Smets and Wouters (2003), Levin et al. (2006), and Schmitt-Gröhé and Uribe (2006). 3 Examples include the Bank of Canada, the Bank of England, the European Central Bank, and the Sveriges Riksbank. Recent DSGE model development at the Federal Reserve Board is described in Erceg, Guerrieri, and Gust (2006) and Edge, Kiley, and Laforte (2007). ronment in which the central bank faces a set of competing models and uses incoming information to update its probability assessments regarding which model is the best representation of the actual economy. Moreover, because private sector expectations play a key role in determining economic outcomes, the optimal policy not only characterizes the central bank’s current actions but also involves a complete set of commitments regarding which future actions will be taken under every possible contingency. Given this approach, the analysis is made tractable—and very elegant— by the use of Markov jump-linear-quadratic methods. In this environment, the Bayesian optimal policy is influenced by an “experimentation” motive, because the central bank recognizes that its current policy actions can influence the flow of incoming information and thereby affect the degree of model uncertainty in subsequent periods. In effect, experimentation is a form of public investment that incurs a short-run cost (in terms of greater macro volatility) in exchange for the medium-run benefit of a more precise estimate of the structure of the economy that will thereby facilitate better stabilization policies. Thus, the paper also makes a valuable contribution by comparing the Bayesian optimal policy with an “adaptive optimal control” strategy (in which the central bank updates its probability assessments of the competing models but does not engage in experimentation) and with the case of Andrew Levin is a deputy associate director in the Division of Monetary Affairs, Board of Governors of the Federal Reserve System. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 301-305. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 301 Levin “no learning” (in which the central bank never changes its probability assessments). Interestingly, this analysis reaches conclusions regarding the role of experimentation that are broadly similar to those obtained in earlier studies such as Wieland (2000, 2006). In particular, the experimentation motive has relatively modest effects on the characteristics of the Bayesian optimal policy, and welfare comparisons indicate fairly minimal costs of using adaptive optimal control. Indeed, as John Taylor described in a recent interview (Leeson, 2007), he arrived at essentially the same conclusions several decades ago when he applied Bayesian optimal control to a small structural macro model: “My Ph.D. thesis...problem was to find a good policy rule in a model where one does not know the parameters and therefore had to estimate them and control the dynamic system simultaneously. My main conclusion...was that in many models, simply following a rule without special experimentation features was a good approximation [to the optimal policy].” In the remainder of this commentary, I discuss a few conceptual issues regarding the formulation of model uncertainty, the characterization of optimal policy under commitment, and the specification of how the private sector’s information set differs from that of the central bank. CHARACTERIZING MODEL UNCERTAINTY In analyzing the monetary policy implications of model uncertainty, it seems reasonable to assume that there will never be any single “true” model, because every macro model is merely a stylized approximation of reality. Moreover, ongoing progress in economic theory and empirical analysis not only shifts policymakers’ probability assessments about which existing model is the best approximation, but it also inevitably generates a winnowing process whereby new modeling mechanisms are developed while obsolete models are completely discarded. Over the past few decades, for example, many central banks have undergone a sequence of transitions from 302 J U LY / A U G U S T 2008 traditional Phillips curve models (which implied a positive long-run relationship between output and inflation) to structural macro models embedding rational expectations—most recently to DSGE models with formal microeconomic foundations. Furthermore, it seems reasonable to anticipate that this process of model development and refinement will continue at a similar pace in the years ahead. From this perspective, a stationary Markov process does not seem to be the ideal approach to represent the sort of model uncertainty that is relevant for monetary policymaking. In the present analysis, each competing model corresponds to a specific node or “state” of the Markov process; hence, model uncertainty is represented by the policymaker’s assessments of the probability that each of these nodes is the correct model of the economy, and the learning process is represented by how these probability assessments are updated in response to incoming information. Thus, if the economy switches from one node to another, this implies that the “true” model of the economy has suddenly shifted. Such shifts may well occur, but it seems doubtful that the process is stationary: that is, the true economy does not shift back and forth among the members of the set of competing models. For example, a recent study of an empirical DSGE model of the U.S. economy found that two alternative specifications of the structure of nominal wage contracts—namely, Calvo-style contracts with random duration versus Taylor-style contracts with fixed duration—have markedly different implications for optimal monetary policy and welfare (Levin, Onatski, J. Williams, and N. Williams, 2006). The analytical framework of this paper can easily be used to characterize the Bayesian optimal policy for this specification uncertainty: One node would correspond to the Calvo-style contract structure, and the other node would correspond to the Taylor-style contract structure. But it does not seem plausible to specify this uncertainty as a stationary Markov process— after all, that would imply that the economy occasionally shifts back and forth between Calvostyle contracts and Taylor-style contracts! F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Levin Of course, a stationary Markov regime– switching specification may well be useful for representing occasional shifts in the state of the economy, such as stochastic transitions between low growth and high growth. But in the case of model uncertainty, it seems reasonable to specify a diagonal structure for the Markov transition matrix: that is, the true economy never shifts between competing models. In that case, the policymaker has prior beliefs that assign some positive weight to each of these models; these priors are then updated in response to incoming information. Alternatively, one might consider a triangular Markov transition matrix with very small off-diagonal elements, representing the notion that the true structure of the economy might experience very rare shifts but would never revert to its original structure. CHARACTERIZING OPTIMAL POLICY UNDER COMMITMENT The “timeless perspective” is an appealing approach to characterizing optimal policy under commitment in a stationary environment (Woodford, 2003). This approach is equivalent to assuming that the government agency established a complete set of state-contingent policy commitments at some point in the distant past (that is, time t = –⬁), and that the economy has converged to its stationary steady state under that regime by now (t = 0). Moreover, in the general case in which this steady state is not Pareto optimal, the quadratic approximation of household welfare depends on the steady-state values of the Lagrange multipliers of the original policymaking problem (Benigno and Woodford, 2005). In contrast, for the reasons described here previously, an environment of model uncertainty may be viewed as implying that the economy has not yet reached any stationary steady state, and hence that policy should not be characterized from a timeless perspective. Indeed, in this context it might be more natural to characterize optimal policy from the Ramsey perspective, that is, assuming that the policymaker is prepared to establish a complete set of state-contingent comF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W mitments starting in the present period (that is, as of time t = 0), where these commitments would reflect the anticipation that incoming information in future periods will gradually enable the policymaker to learn which model correctly represents the economy. Of course, that specification would raise further computational issues: Under the Ramsey policy (as opposed to the timeless perspective), the Lagrange multipliers corresponding to the implementation constraints cannot be substituted out of the problem but remain as essential state variables of the linear-quadratic approximation. CHARACTERIZING THE PRIVATE SECTOR’S INFORMATION Finally, it is worthwhile to consider the assumptions used in this analysis regarding the information available to private agents: 1. In the benchmark case of Bayesian optimal control, the analysis of this paper proceeds under the assumption that neither private agents nor the policymaker know which model is true. Unfortunately, this assumption is somewhat problematic in the context of DSGE models with explicit microeconomic foundations, because those models are formulated under the assumption that each household is aware of its own preferences and that each firm is aware of its own production technology and the characteristics of consumer demand. For example, in New Keynesian DSGE models with monopolistic competition and staggered price contracts, it is assumed that each firm sets the price of its product with full knowledge of its own production function and the elasticity of demand for its product. Nevertheless, econometricians may be unable to make precise distinctions regarding the extent to which aggregate price-setting behavior is influenced by factors such as firm-specific inputs and quasi-kinked demand; hence, there may be a strong motive for designing a monetary policy strategy that is robust to this source J U LY / A U G U S T 2008 303 Levin of model uncertainty (Levin, Lopez-Salido, and Yun, 2007). Similarly, DSGE models typically involve a consumption Euler equation that is derived from a particular specification of household preferences for consumption and leisure—and of course, each individual household is assumed to have full knowledge of its own preferences in making decisions about spending, labor supply, etc. Nevertheless, the available data may be insufficient to enable econometricians to resolve uncertainty regarding several competing specifications of household preferences. Therefore, the central bank may wish to follow a policy strategy that is robust to this source of model uncertainty (Levin et al., 2008). 2. In the case of adaptive optimal control, the analysis proceeds under the more restrictive assumption that neither private agents nor the policymaker observe the current vector of shocks—an assumption that precludes consideration of most (if not all) existing DSGE models. In many such models, for example, shocks to total factor productivity play a key role as a source of aggregate volatility in output and employment. But it is by no means clear how an individual firm could determine its own production if the firm did not have contemporaneous knowledge of its own productivity. 3. The case of “no learning” assumes that neither private agents nor the policymaker can recall any of the data that were observed in previous periods. In many DSGE models, however, these data do enter explicitly into agents’ decision rules. For example, in specifications with habit persistence in consumption, the household’s current spending decision partly reflects its spending in previous periods. Similarly, when investment in physical capital is subject to adjustment costs, each individual firm’s decision regarding its current level of investment depends explicitly on its prior investment decisions. 304 J U LY / A U G U S T 2008 Evidently, in analyzing optimal policy under model uncertainty in the context of DSGE models with explicit micro foundations, further progress is needed to distinguish between the information available to the central bank and the information that is available to individual households and firms. REFERENCES Benigno, Pierpaolo and Woodford, Michael. “Inflation Stabilization and Welfare: The Case of a Distorted Steady State.” Journal of the European Economics Association, 2005, 3, pp. 1185-236. Christiano, Lawrence; Eichenbaum, Martin and Evans, Charles. “Nominal Rigidities and the Dynamic Effects of a Shock to Monetary Policy.” Journal of Political Economy, 2005, 113, pp. 1-45. Clarida, Richard; Galí, Jordi and Gertler, Mark. “The Science of Monetary Policy: A New Keynesian Perspective.” Journal of Economic Literature, 1999, 37, pp. 1661-707. Edge, Rochelle; Kiley, Michael and Laforte, JeanPhillipe. “Natural Rate Measures in an Estimated DSGE Model of the U.S. Economy.” Finance and Economics Discussion Series, No. 2007-08, Board of Governors of the Federal Reserve System, 2007. Erceg, Christopher; Guerrieri, Luca and Gust, Christopher. “SIGMA: A New Open Economy Model for Policy Analysis.” International Journal of Central Banking, 2006, 2, pp. 1-50. Goodfriend, Marvin and King, Robert G. “The New Neoclassical Synthesis and the Role of Monetary Policy.” NBER Macroeconomics Annual 1997. Cambridge, MA: MIT Press, 1997. King, Robert G. and Wolman, Alexander L.“Inflation Targeting in a St. Louis Model of the 21st Century,” Federal Reserve Bank of St. Louis Review, May/June 1996, 78(3), pp. 83-107. King, Robert G. and Wolman, Alexander L. “What Should the Monetary Authority Do When Prices Are Sticky?” in John B. Taylor, ed., Monetary F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Levin Policy Rules. Chicago: University of Chicago Press, 1999, pp. 349-98. Leeson, Robert. “An Interview with John B. Taylor.” Unpublished manuscript, Murdoch University, 2007. Levin, Andrew; Onatski, Alexei; Williams, John C. and Williams, Noah. “Monetary Policy under Uncertainty in Micro-Founded Macroeconometric Models,” in Mark Gertler and Kenneth Rogoff, eds., NBER Macroeconomics Annual 2005. Cambridge, MA: MIT Press, 2006. Levin, Andrew; Lopez-Salido, J. David and Yun, Tack. “Strategic Complementarities and Optimal Monetary Policy.” Discussion Paper No. 6423, Centre for Economic Policy Research, 2007. Levin, Andrew; Lopez-Salido, J. David; Nelson, Edward and Yun, Tack. “Macroeconometric Equivalence, Microeconomic Dissonance, and the Design of Monetary Policy.” Journal of Monetary Economics, 2008 (forthcoming). McCallum, Bennett T. and Nelson, Edward. “Performance of Operational Policy Rules in an Estimated Semi-Classical Structural Model,” in John B. Taylor, ed., Monetary Policy Rules. Chicago: University of Chicago Press, 1999, pp. 15-45. Rotemberg, Julio J. and Woodford, Michael. “An Optimization-Based Econometric Framework for the Evaluation of Monetary Policy.” NBER Macroeconomics Annual 1997. Cambridge, MA: MIT Press, 1997, pp. 297-346. Schmitt-Gröhé, Stephanie and Uribe, Martin. “Optimal Fiscal and Monetary Policy in a Medium-Scale Macroeconomic Model,” in Mark Gertler and Kenneth Rogoff, eds., NBER Macroeconomics Annual 2005. Cambridge, MA: MIT Press, 2006. Smets, Frank and Wouters, Raf. “An Estimated Dynamic Stochastic General Equilibrium Model of the Euro Area.” Journal of the European Economic Association, 2003, 1, pp. 1123-75. Svensson, Lars E.O. and Williams, Noah. “Optimal Monetary Policy Under uncertainty: A Markov Jump-Linear-Quadratic Approach.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 275-93. Wieland, Volker. “Monetary Policy, Parameter Uncertainty, and Optimal Learning.” Journal of Monetary Economics, 2000, 46, pp. 199-228. Wieland, Volker. “Monetary Policy and Uncertainty about the Natural Unemployment Rate: BrainardStyle Conservatism versus Experimental Activism.” Berkeley Electronic Journal of Macroeconomics: Advances in Macroeconomics, 2006, 6(1), Article 1. Woodford, Michael. Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton, NJ: Princeton University Press, 2003. Rotemberg, Julio J. and Woodford, Michael. “Interest Rate Rules in an Estimated Sticky-Price Model,” in John B. Taylor, ed., Monetary Policy Rules. Chicago: University of Chicago Press, 1999. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 305 306 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Economic Projections and Rules of Thumb for Monetary Policy Athanasios Orphanides and Volker Wieland Monetary policy analysts often rely on rules of thumb, such as the Taylor rule, to describe historical monetary policy decisions and to compare current policy with historical norms. Analysis along these lines also permits evaluation of episodes where policy may have deviated from a simple rule and examination of the reasons behind such deviations. One interesting question is whether such rules of thumb should draw on policymakers’ forecasts of key variables, such as inflation and unemployment, or on observed outcomes. Importantly, deviations of the policy from the prescriptions of a Taylor rule that relies on outcomes may be the result of systematic responses to information captured in policymakers’ own projections. This paper investigates this proposition in the context of Federal Open Market Committee (FOMC) policy decisions over the past 20 years, using publicly available FOMC projections from the semiannual monetary policy reports to Congress (Humphrey-Hawkins reports). The results indicate that FOMC decisions can indeed be predominantly explained in terms of the FOMC’s own projections rather than observed outcomes. Thus, a forecast-based rule of thumb better characterizes FOMC decisionmaking. This paper also confirms that many of the apparent deviations of the federal funds rate from an outcome-based Taylor-style rule may be considered systematic responses to information contained in FOMC projections. (JEL E52) Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 307-24. W illiam Poole has been a long-time proponent of rules of thumb for monetary policy. Nearly four decades ago, as staff economist at the Board of Governors of the Federal Reserve System (BOG), Poole presented a reactive rule of thumb that he argued could serve as a robust guide to policy decisions (Poole, 1971). More recently, as president of the Federal Reserve Bank of St. Louis and a member of the Federal Open Market Committee (FOMC), he has highlighted how a simple Taylor rule that systematically responds to economic activity and inflation can serve as a useful tool for understanding historical monetary policy decisions (Poole, 2007). In both his recent and earlier work, Poole highlighted the usefulness of rules of thumb in the context of the complexity of the macroeconomy and our limited knowledge regarding it. In this light, a policy adviser cannot offer precise guidance about how the monetary authority should respond to every conceivable contingency to best achieve its goals. What a policy adviser can do is identify useful rules of thumb that can serve as appropriate guides to policy under most circumstances. To the extent policymakers rely on Athanasios Orphanides is the Governor of the Central Bank of Cyprus, and Volker Wieland is a professor at the Goethe University Frankfurt, director at the Center for Financial Studies, and fellow at the Centre for Economic Policy Research. Volker Wieland thanks the Stanford Center for International Development, where he was a visiting professor while writing this paper. The authors are grateful for excellent research assistance by Sebastian Schmidt from Goethe University Frankfurt. Helpful comments were provided by Greg Hess, Jim Hamilton, participants at the St. Louis conference, and the paper’s discussants, Charles Plosser and Patrick Minford. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, the regional Federal Reserve Banks, the Central Bank of Cyprus, or the Governing Council of the European Central Bank. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 307 Orphanides and Wieland a simple rule of thumb as an approximate policy guide, it should be possible to identify this rule and use it to understand historical policy decisions and to improve future policy. One of the difficulties in identifying a simple rule that can serve as a useful description of policy is that the policy prescriptions relevant for policy advice at any point in time reflect the information available to policymakers at that time. To the extent policy is based on observable macroeconomic variables, a simple rule could be estimated using real-time historical data. However, to the extent policymakers view projections of key macroeconomic variables as more useful summary descriptions of the current state of the economy, estimation of a simple rule based on those same policymaker projections would provide a more promising avenue. Poole (2007) examines FOMC policy decisions over the past 20 years using the simple outcome-based rule proposed by Taylor (1993). This rule uses the current inflation rate and output gap as inputs for federal funds rate decisions. Poole identifies some deviations of policy from the systematic prescriptions suggested by the rule that could, however, reflect a systematic response of the FOMC to its own projections. Our objective in this paper is to investigate this proposition. To this end we compare estimated policy rules that are based on recent economic outcomes with policy rules based on the economic projections of the FOMC. We investigate whether the federal funds rate target set by the FOMC when these projections are made responds systematically to these projections as opposed to recent economic data. Our results, which are based on real-time data and projections over the past 20 years, indicate that interest rates respond predominantly to FOMC projections and thus that a forecast-based rule better characterizes FOMC decisionmaking during this period. Furthermore, we check to what extent deviations from an outcome-based Taylor rule may be better explained by the information incorporated in FOMC forecasts. Our analysis suggests that by distinguishing between forecasts and outcomes one can explain a number of deviations of policy from the simple underlying rule, though it can also identify episodes where devi308 J U LY / A U G U S T 2008 ations remain. This includes episodes where one would expect systematic policy to deviate from a simple rule of thumb, such as the response to financial turbulence experienced in 1998. Overall, our analysis suggests that FOMC projections used in the context of a rule of thumb are quite informative for understanding historical monetary policy, whereas similar analysis based on economic outcomes can often be of much lower value. ON RULES OF THUMB FOR MONETARY POLICY Simple estimated rules can be useful devices for understanding historical monetary policy if central banks conduct policy sufficiently systematically to be captured by such rules. Poole (1971) suggested that it is reasonable for individual policymakers to behave in a systematic manner: Individual policy-makers inevitably use informal rules of thumb in making decisions. Like everyone else, policy-makers develop certain standard ways of reacting to standard situations. These standard reactions are not, of course, unchanging over time, but are adjusted and developed according to experience and new theoretical ideas. (p. 151) Though it did not attract much attention at the time, the particular rule of thumb proposed by Poole in 1971 is of interest in that it incorporated both a reaction of the interest rate to real economic activity (specifically the deviation of the unemployment rate from the Federal Reserve’s estimate of the full employment rate at the time), as well as a nominal variable in a way that would ensure price stability over the long run. The latter was not based on the response of the interest rate to inflation, as is commonly specified today. Rather, Poole’s rule specified that the money supply should always be contained within bounds as a robust means of controlling inflation and suggested adjusting the interest rate to respond to deviations of unemployment from full employment only when doing so would respect these bounds. In essence, Poole’s rule of thumb uses money growth to ensure the maintenance of price F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Orphanides and Wieland stability and, subject to that, provides countercyclical policy prescriptions. He provided the following summary description: The proposed rule assumes that full employment exists when the unemployment rate is in the 4.0 to 4.4 per cent range. The rule also assumes that at full employment, a growth rate of the money stock of 3 to 5 per cent per annum is consistent with price stability. Therefore, when unemployment is in the full employment range, the rule calls for monetary growth at the 3 to 5 per cent rate. The rule calls for higher monetary growth when unemployment is higher, and lower monetary growth when unemployment is lower. Furthermore, when unemployment is relatively high the rule calls for a policy of pushing the Treasury bill rate down provided monetary growth is maintained in the specified range; similarly, when unemployment is relatively low the rule calls for a policy of pushing the Treasury bill rate up provided monetary growth is in the specified range. Finally, the rule provides for adjusting the rate of growth of money according to movements in the Treasury bill rate in the recent past. (p. 183) process that consists largely of reactions to current developments. Only gradually will policy-makers place greater reliance on formal forecasting models. (pp. 152-53) In 2007, Poole used a version of the classic Taylor (1993) rule to describe Federal Reserve behavior over the past 20 years.1 As is well known, this rule posits that the systematic component of monetary policy may be described as a notional target for the federal funds rate, fˆ: (1) The FOMC, and certainly John Taylor himself, view the Taylor rule as a general guideline. Departures from the rule make good sense when information beyond that incorporated in the rule is available. For example, policy is forward looking, which means that from time to time the economic outlook changes sufficiently that it makes sense for the FOMC to set a funds rate either above or below the level called for in the Taylor rule, which relies on observed recent data rather than on economic forecasts of future data. Other circumstances—an obvious example is September 11, 2001—call for a policy response. These responses can be and generally are understood by the market. Thus, such responses can be every bit as systematic as the responses specified in the Taylor rule. (p. 6) It is not proposed that this rule of thumb or guideline be followed if there is good reason for departure. But departures should be justified by evidence and not be based on vague intuitive feelings of what is needed since the rule was carefully designed from the theoretical and empirical analysis...and from a careful review of post-accord monetary policy. (p. 183) Given the accuracy of forecasts at the current state of knowledge, it seems likely that for some time to come forecasts will be used primarily to supplement a policy-decisionmaking F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W ) where π and y reflect contemporaneous readings of inflation and a measure of the output gap, respectively. Following Taylor, Poole assumed a constant inflation target, π *, and a constant equilibrium real interest rate, r*. Poole’s rendition of the Taylor rule is reproduced in Figure 1. As in his work 36 years earlier, Poole (2007) explained potential sources of deviation from the rule and also the potential use of forecasts: Poole also explicitly recognized a scope for deviations from his suggested rule of thumb, even if policymakers had decided to adopt it in principle. What was more important in Poole’s view was transparency in explaining the rationale for such deviations: As to whether rules could usefully rely on economic projections, Poole (1971) argued that an important factor would be the accuracy of the forecasts: ( fˆ = r * + π + 0.5 π − π * + 0.5y , This last remark suggests that a better rule of thumb for understanding the behavior of the Federal Reserve over the past 20 years could be a version of the Taylor rule that is explicitly based 1 Taylor (1993) showed that the rule could describe Federal Reserve behavior from 1987 to 1992 quite well. Interest rate rules had also acquired a normative dimension at that time because of their success in a large-scale model comparison project reported in Bryant, Hooper, and Mann (1993) (see also Henderson and McKibbin, 1993). J U LY / A U G U S T 2008 309 Orphanides and Wieland Figure 1 Poole’s (2007) Version of the Taylor Rule Percent 12 BOG Output Gap: CPI, 1987:09–2000:10 Federal Funds Rate 10 CBO Output Gap: CPI, 2000:11–2006:06 8 6 4 2 0 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 FOMC Meeting Dates NOTE: The solid blue line shows the Taylor rule constructed using the BOG real-time output-gap estimate. The blue dashed line extends the rule using the output-gap estimate of the CBO for those years for which the BOG estimate is not yet public information. on the FOMC’s own projections. This is the subject of the investigation that follows. We begin by describing how to construct constant-horizon forecasts that can be used in estimating a policy rule from publicly available projections. The semiannual monetary policy reports to Congress (the Humphrey-Hawkins reports) have presented information on the range and central tendency of annual forecasts of FOMC members since 1979.2 Following Poole’s (2007) analysis, we create a dataset of FOMC projections and corresponding real-time data on observed outcomes that focuses our attention on the past 20 years.3 Regarding projections, we take the midpoints of the central tendencies reported in each of the reports, starting with the February 1988 report and ending with the July 2007 report, and use these as proxies for the modal forecasts of FOMC expectations. Our objective using these data is to examine whether deviations from an outcome-based Taylor rule may be explained by the additional information contained in policymakers’ forecasts. These include inflation, the rate of unemployment, and output growth. Because we could not make approximate inferences of the FOMC forecasts of the output gap from these variables, although we do have the FOMC’s unemployment projections, we focus on a version of the Taylor rule that substitutes the unemployment rate for the output gap. Consequently, in our dataset we focus on data and forecasts regarding inflation and unemployment. 2 3 FOMC ECONOMIC PROJECTIONS AND REAL-TIME OUTCOMES A month after this paper was first presented, on November 14, 2007, the Federal Reserve announced that going forward the FOMC would compile and release these economic projections four times a year instead of just two times a year, which was the practice until then. 310 J U LY / A U G U S T 2008 In earlier work, Lindsey, Orphanides, and Wieland (1997), we examined the implications of FOMC projections for understanding policy in the sample prior to 1988 and presented some comparisons with the 1988-96 period. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Orphanides and Wieland Figure 2 The Timing of Forecasts in Humphrey-Hawkins Reports: Unemployment Rates February Report HH ut+3|t Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 HH ut+1|t HH ut+3|t HH ut+5|t July Report Some of the particular measures have been redefined over the years. For inflation, the implicit deflator of the gross national product was used through July 1988, thereafter replaced by the consumer price index (CPI). In February 2000, the CPI was replaced by the personal consumption expenditures (PCE) deflator measure of inflation, and from July 2004 onward the FOMC decided to focus on the core PCE deflator that excludes food and energy prices because of their volatility. These changes are of particular interest because the alternative measures do not always provide similar summary readings of inflationary pressures. They may differ both in their level and in their variability over time, especially in small samples, which poses some interpretation challenges. Tables 1 and 2 provide two recent examples useful for understanding what information on projections is released with the monetary policy reports. Forecasts for 2007 were first reported in July 2006 (not shown). In February 2007, revised forecasts for 2007 and first forecasts for 2008 were reported (Table 1). The final updated forecasts for 2007 were then published in July together with updated forecasts for 2008 (Table 2). Although we have only two observations per year, it is convenient to describe our dataset in F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W terms of a quarterly frequency because the FOMC projections report either quarterly data or growth rates over four quarters. Denoting time (measured in quarters) with t, we associate the February Humphrey-Hawkins report with the first quarter of the year and the July Humphrey-Hawkins report with the third quarter. We construct a dataset containing two sets of forecasts for each year, covering four-quarter intervals that always end three quarters in the future. For any variable x, let xt +i|t denote the estimated outcome (for i ⱕ 0) or forecast (for i > 0) of the value of the variable x at t +i as of time t.4 Then, letting u denote the unemployment rate, ut+3|t represents the three-quarter-ahead forecast of the unemployment rate formed during quarter t, and ut–1|t the estimate as of quarter t of what the outcome for the unemployment rate was in the previous quarter. As shown on the time chart in Figure 2, using the unemployment rate as an example, the forecasts reported to Congress in the February 4 Importantly, because of the lags with which information about the past becomes available, we need to keep track not only of revisions of forecasts but also of revisions regarding outcomes when trying to understand the environment in which FOMC decisions were taken. We later describe the data we use for outcomes. J U LY / A U G U S T 2008 311 Orphanides and Wieland Table 1 FOMC Forecasts for 2007 and 2008 from the February 2007 Humphrey-Hawkins Report 2007 Indicator 2008 Memo 2006 actual Range Central tendency Range Central tendency Nominal GDP 5.9 4 3/4–5 1/2 5–5 1/2 4 3/4–5 1/2 4 3/4–5 1/4 Real GDP 3.4 2 1/2–3 1/4 2 1/2–3 2 1/2–3 1/4 2 3/4–3 2.3 2–2 1/4 2–2 1/4 1 1/2–2 1/4 1 3/4–2 4.5 4 1/2–4 3/4 4 1/2–4 3/4 4 1/2–5 4 1/2–4 3/4 Change, fourth quarter to fourth quarter* PCE price index excluding food and energy Average level, fourth quarter Civilian unemployment rate NOTE: *Change from average for fourth quarter of previous year to average for fourth quarter of year indicated. SOURCE: “Economic Projections of Federal Reserve Governors and Reserve Bank Presidents” from the February 2007 HumphreyHawkins report. Table 2 FOMC Forecasts for 2007 and 2008 from the July 2007 Humphrey-Hawkins Report 2007 Indicator 2008 Range Central tendency Range Central tendency 4 1/2–5 1/2 4 1/2–5 4 1/2–5 1/2 4 3/4–5 Real GDP 2–2 3/4 2 1/4–2 1/2 2 1/2–3 2 1/2–2 3/4 PCE price index excluding food and energy 2–2 1/4 2–2 1/4 1 3/4–2 1 3/4–2 4 1/2–4 3/4 4 1/2–4 3/4 4 1/2–5 About 4 3/4 Change, fourth quarter to fourth quarter* Nominal GDP Average level, fourth quarter Civilian unemployment rate NOTE: *Change from average for fourth quarter of previous year to average for fourth quarter of year indicated. SOURCE: “Economic Projections of Federal Reserve Governors and Reserve Bank Presidents” from the July 2007 Humphrey-Hawkins report. 312 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Humphrey-Hawkins report have exactly the desired timing. That is, when t is the first quarter, the three-quarter-ahead forecast of unemployment, ut+3|t, corresponds to the February HumphreyHawkins forecast of the unemployment rate in the fourth quarter of the same year. That is, when t represents the first quarter of a year, we have (2) ut +3|t ; utHH +3|t , where we employ the superscript HH to denote the Humphrey-Hawkins forecasts. Note that in Figure 2 under the heading ‘‘February Report” the solid arrow points to the quarter on the time line for which the unemployment rate is predicted (t +3) and the dotted line points to the quarter in which the forecast is made (t). Similarly, for inflation, when t represents the first quarter of a year, the three-quarter-ahead forecast corresponds to the rate of growth of prices from the fourth quarter of the previous year to the fourth quarter of the current year, exactly matching the horizon of the February HumphreyHawkins forecast. Letting π represent the rate of inflation over four quarters, when t is the first quarter of a year, we have (3) π t +3|t ; π tHH +3|t . For the July Humphrey-Hawkins reports, some additional work is required to obtain threequarter-ahead projections; that is, we combine available information to estimate the forecast of the unemployment rate for the second quarter of next year and the corresponding forecast of the four-quarter growth rate of prices that ends in the same quarter. The timing of the two July Humphrey-Hawkins forecasts and the constructed three-quarter-ahead unemployment forecast is also shown with respect to the time line in Figure 2. In this case, the dashed arrow refers to the threequarter-ahead observation for which an unemployment forecast is needed. To approximate the unemployment forecast for the second quarter of the following year, we simply take from the July Humphrey-Hawkins report the forecasted unemployment rates for the current year’s fourth quarter and next year’s fourth quarter and average them. That is, when t represents the third quarter of the year, we set F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W (4) ut +3|t = ( Orphanides and Wieland ) 1 HH ut +1|t + utHH +5|t . 2 Other than the rare occurrence of when a shock is known to have only transitory effects, for a four-quarter interval that starts two quarters later, it is doubtful that FOMC members would have strong views about the likelihood of different changes in the unemployment rate over the two halves of that period. Implicitly, we assume that the changes forecasted in July for the unemployment rate in each half of next year are about the same. The desired second-quarter-to-second-quarter forecasts of the growth rate of prices is obtained by constructing two forecasted half-year annualized growth rates and then averaging them. In other words, when t represents the third quarter of the year, we set (5) π t + 3|t = ) ( 1 S π t +1|t + π tS+3|t , 2 S where S stands for semiannual, so that π t+1|t is the inflation forecast for the second half of the S current year and π t+3|t is the forecast for the first half of the following year. The inflation forecasted for the second half S of the current year, π t+1|t , can be inferred from the forecast reported for all of the year from a base of HH last year’s fourth quarter, π t+1|t , and the estimated inflation over the first half of the current year from S a base of last year’s fourth quarter, π t–1|t . That is, expressing all terms as annualized growth rates, when t represents the third quarter of the year, (6) S π tS+1|t = 2π tHH +1|t − π t −1|t . S For π t+3|t , inflation over the first half of the next year, we simply set it equal to the July Humphrey-Hawkins forecast for all of next year. That is, we set (7) π tS+3|t = π tHH + 5|t . The July Humphrey-Hawkins report does not provide an estimate of inflation for the first half S of the current year, that is, for π t–1|t . Thus, instead we make use of alternative real-time data sources, which are discussed below. J U LY / A U G U S T 2008 313 Orphanides and Wieland To allow for a direct comparison of rules based on the forecasts described above with rules based on outcomes of these variables, we construct parallel variables reflecting the latest historical information available to the FOMC at the time of their meetings preceding the two HumphreyHawkins reports each year. Thus, for the unemployment rate, we create the variable ut–1|t, which for the February observation reflects the average level in the fourth quarter of the prior year and for the July observation reflects the average level in the second quarter of the current year. Similarly, for inflation, we create the variable πt–1|t, which reflects the four-quarter growth rate of prices ending in the fourth quarter of the prior year for the February observation and ending in the second quarter of the current year for the July observation. An important aspect of our analysis is to ensure that our definition of outcomes reflects only information available to the FOMC in real time. To that end, we rely only on data that would have been available to the FOMC by early February or early July. This implies that the data we use correspond either to preliminary estimates, firstreported quarterly data, or estimates based on partial data for the quarter. To match the timing of this information as closely as possible, for the years 1988 through 2001 inclusive, we use BOG staff estimates of outcomes ending in the prior quarter, which are contained in the Greenbook that is distributed to the FOMC prior to the early-February and earlyJuly FOMC meetings. Even so, because Greenbook data remain confidential for five years, we cannot rely on that source for the last few years of our sample. Instead, for 2002-07 we use real-time vintage data from the Federal Reserve Bank of St. Louis ALFRED database.5 For these dates we use the data vintage from ALFRED that was available one week after the respective February and July Humphrey-Hawkins meetings. We choose this timing because FOMC members have the 5 As a robustness check, we have investigated how much the ALFRED-based information differs from Greenbook information in the years until 2001, when both are available. Although the data source does influence the data values somewhat, the differences were small. 314 J U LY / A U G U S T 2008 opportunity to revise their projections during a window of a few days following the meetings. ESTIMATED POLICY RULES: FOMC PROJECTIONS VERSUS RECENT OUTCOMES Specification The interest rate rules we estimate all share the following underlying structure with Taylor’s (1993) rule. They posit that the systematic component of monetary policy can be described as a notional target for the federal funds rate, fˆ, which increases with inflation, π, and real activity. As already mentioned with regard to projections of real activity, we do not have information about the FOMC’s assessment of the output gap. Thus, we cannot directly estimate an exact counterpart of the rule proposed by Taylor. Instead, an indirect comparison is feasible using the unemployment rate, u, as a measure of the level of economic activity.6 Following Taylor, we restrict attention to a linear specification of the rule and posit that7 (8) fˆ = a0 + aπ π + auu. Note that we do not have direct information on the policymakers’ views regarding the equilibrium interest rate, r*, the inflation target, π *, or the natural rate of unemployment, u*. If these concepts are roughly constant over the sample period, then they would be subsumed in the estimated intercept, a0 = r * − (aπ − 1)π * − auu* . In estimating our specification, we need to take an explicit stand regarding the explanatory 6 The difference between the unemployment rate and a constant natural rate (NAIRU) can then be translated into an estimate of the output gap by means of Okun’s law. 7 The linearity assumption is purely for simplicity in the spirit of the Taylor rule. Nonlinear reaction functions, such as those characterizing “opportunistic disinflation” examined by Orphanides and Wilcox (2002) and Aksoy et al. (2006) and those incorporating asymmetric easing near the zero-bound for nominal interest rates as derived by Orphanides and Wieland (2000), would likely be more-realistic but more-complicated depictions of policy. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Orphanides and Wieland Figure 3 The Timing of the Explanatory Variables in Humphrey-Hawkins Reports: Outcomes and Forecasts of Unemployment February Report ut+3|t ut–1|t Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 ut–1|t ut+3|t July Report variable as well as the timing of the information about inflation and real activity that the FOMC takes into account in their policy decision. Regarding the FOMC’s policy instrument, that is, the interest rate on the left-hand side of the rule, we use the FOMC’s intended level of the federal funds rate as of the close of financial markets on the day after the February and July FOMC meetings. Regarding the information on the current or projected state of the economy, we set (9) fˆ = a0 + aπ π τ t + auuτ t , where τ captures the particular timing. The explanatory variables, πτ|t and uτ|t, are meant to encompass the information variables to which the FOMC may be reacting. In this specification, τ = t –1 if the rule of thumb is outcome based, whereas τ = t+3 if it is forecast based, that is, based on the three-quarter-ahead projections. Figure 3 again employs a time line to put the timing of the explanatory variables into perspective, using the unemployment outcomes and forecasts as an example. Again, the arrows point to the quarters to which the forecast or outcome applies, and the dotted lines indicate the dates F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W on which the forecast or the estimate of the outcome are made. In our estimation, we also allow for the possibility that the FOMC has a preference for policy inertia and perhaps only partially adjusts the intended federal funds rate, f, toward its notional target, fˆ. We introduce such inertial behavior by allowing the FOMC decision prior to a HumphreyHawkins report to be influenced by the level of the intended federal funds rate decided at the FOMC meeting before the previous HumphreyHawkins report. With our timing convention, this can be written as (10) ft = (1 − ρ ) fˆt + ρ ft − 2 , where ρ provides a measure of the degree of partial adjustment. Thus, the restriction, ρ = 0, would reflect an immediate adjustment of the intended federal funds rate to its notional target. Regression Estimates: 1988-2007 The results from our regression analysis using our sample of Humphrey-Hawkins report data from 1988 to 2007 are summarized in Table 3. The estimates shown are obtained by non-linear J U LY / A U G U S T 2008 315 Orphanides and Wieland Table 3 Policy Reaction to Inflation and Unemployment Rates: Outcomes versus FOMC Forecasts, 1988-2007:Q2 Regressions based on Outcomes Forecasts (1) (2) (3) (4) a0 8.29 1.08 10.50 3.07 6.97 0.69 8.25 0.85 aπ 1.54 0.16 1.29 0.43 2.34 0.12 2.48 0.14 au –1.40 0.21 –1.70 0.55 –1.53 0.14 –1.84 0.17 ρ 0 0 – R2 0.69 0.14 0.39 0.06 0.74 0.84 0.91 0.96 SEE 1.10 0.85 0.64 0.44 SW 1.00 1.03 1.74 1.94 ft = ρ ft −2 + (1 − ρ ) (a0 + aπ π τ|t + auuτ|t ), NOTE: The regressions shown are least-squares estimates of Here, f denotes the intended federal funds rate, π the inflation rate over four quarters, and u the unemployment rate. The horizon τ either refers to three-quarter-ahead forecasts, τ = t +3, or outcomes observed in the preceding quarter, τ = t –1. ft = ρ ft − 2 + (1 − ρ ) (a0 + aπ π τ|t + auuτ|t ). least-squares regressions applied to the equation (11) Columns 1 and 2 of Table 3 show the results for the outcome-based regressions with τ = t –1; columns 3 and 4 show the results for the forecastbased regressions with τ = t+3. Standard errors are shown under the parameter estimates. In columns 1 and 3 the restriction, ρ = 0, is imposed, whereas in columns 2 and 4 the unrestricted partial-adjustment specification is shown. In all regressions shown in the table, we find that the estimated rules of thumb suggest a systematic response to inflation and unemployment. The response to inflation is positive and noticeably greater than 1, suggesting that all of these rules satisfy the Taylor principle. And the response to unemployment is negative and also quite large, suggesting a strong countercyclical stabilization response. These findings are quite robust and hold regardless of whether we employ FOMC projections or recent economic outcomes and 316 J U LY / A U G U S T 2008 regardless of whether we allow for some degree of interest rate smoothing or not. However, not all specifications describe policy decisions equally successfully. A comparison of the regressions based on recent outcomes, columns 1 and 2, with those based on FOMC projections, 3 and 4, reveals that the forecast-based rules describe policy decisions quite a bit better than the corresponding outcome-based rules. We also estimate a richer but more complicated specification that nests the regressions with forecasts and outcomes as limiting cases.8 Estimates of this specification with an estimated weight on forecasts near unity (not shown) confirm the above result. Furthermore, our results suggest a substantial degree of inertia in setting policy. We conclude that a rule of thumb that is based 8 In this case, the measure of inflation conditions in the regression is defined as π τ|t ≡ (1 − φ ) π t −1|t + φπ t +3 t . Similarly, the measure on unemployment conditions depends on the weight φ. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Orphanides and Wieland on the FOMC’s own projections of inflation and unemployment and allows for inertial behavior can serve as a very good guide for understanding the systematic nature of FOMC decisions over the past 20 years. The improved fit of the forecast-based rule relative to the outcome-based rule also suggests that at least some of the apparent deviations of actual interest rates from an outcome-based Taylor rule, such as described in Poole (2007), may be easily explained once FOMC forecasts are examined. To explore this question further, Figure 4 plots the fitted values of the forecast-based and outcome-based rules estimated in Table 3. The upper panel of the figure contains the rules without interest rate smoothing, which correspond to columns 1 and 3 in Table 3. The black line denoted ‘‘Fed Funds’’ indicates the actual federal funds rate target decided at each of the February and July FOMC meetings from 1988 to 2007. The solid blue line indicates the outcome-based rule and the blue dashed line the forecast-based rule. The figure confirms visually that the forecastbased rule explains the path of the federal funds rate target better than the outcome-based rule. Of course, the fit is further improved once we allow for interest rate smoothing, in other words, partial adjustment of the funds rate depending on last period’s realization. This can be seen in the lower panel in the figure, where the paths implied by the fitted outcome- and forecast-based rules, respectively, are smoother because they take into account the estimated degree of partial adjustment. Based on the figure, we can identify five periods where the outcome- and forecast-based rules diverge from each other in an interesting manner and that can improve our understanding of the role of projections for FOMC policy decisions. Two of these episodes, around 1988 and 1994, correspond to periods of rising policy rates. In both of these periods, the FOMC was raising rates preemptively because of concerns regarding the outlook for inflation. Correspondingly, the forecast-based rules track policy decisions better, while the outcome-based rules only manage to describe policy with a noticeable lag. Two other episodes, in 1990-91 and in 2001, correspond to periods of falling policy rates. In F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W both of these periods, the FOMC was easing policy out of concern of a faltering economy, clearly influenced by its projections of relatively weak economic activity. Again, the forecast-based rules track policy decisions better, while the outcomebased rules exhibit a noticeable lag. The last episode is 2002-03, when the forecastbased rule correctly tracked the further policy easing at the early stages of the recovery from the recession, while the outcome-based rule suggested that policy should have been considerably tighter. Of interest are also two additional episodes when the forecast-based rule did not track the actual policy setting as well but where the resulting deviations can be explained by other factors that are not part of the rule. The first of these is the 1998 policy easing. On this occasion, the FOMC was responding to the underlying financial turbulence that intensified that fall, a factor not well reflected in the rule of thumb, even considering its forward-looking nature. The second and arguably more controversial episode is the “miss” reflected in the forecastbased rule during 2004. This is more controversial because of recent criticisms that policy was much easier during this episode than would have been suggested by simple Taylor rules. This is evident, for example, in Poole’s rendition of the classic Taylor rule, reproduced in Figure 1. It has been argued that this policy stance may have contributed to the subsequent housing boom and associated price adjustments and liquidity difficulties experienced in financial markets (Taylor, 2007). Indeed, as is well-known, around 2003-04, the FOMC was particularly concerned with the risks of deflation and perceived an important asymmetry in the costs associated with a possible policy misjudgment. In particular, the costs of policy proving too tight were perceived as considerably exceeding the costs of policy proving too easy.9 Under these circumstances, it should be expected that even a rule of thumb that might track policy nearly perfectly under normal circum9 The suggested rationale was the uncertainty arising with operating policy near the zero bound. See Orphanides and Wieland (2000) for a model demonstrating the optimality of unusually accommodative policy in light of the asymmetric risks associated with the zero bound on nominal interest rates. J U LY / A U G U S T 2008 317 Orphanides and Wieland Figure 4 Outcome-Based versus Forecast-Based Rules, 1988-2007 No Interest Rate Smoothing 10.0 7.5 5.0 2.5 Fed Funds Outcomes Forecasts 19 88 19 89 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 0.0 With Interest Rate Smoothing 10.0 7.5 5.0 2.5 Fed Funds Outcomes Forecasts 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 89 19 19 88 0.0 NOTE: “Fed Funds” refers to the federal funds rate target. “Outcomes” refers to fitted values of the outcome-based rule without and with interest rates smoothing, that is, columns 1 and 2 in Table 3, respectively. “Forecasts” refers to the fitted values of the forecastbased rule without and with interest rate smoothing, that is, columns 3 and 4 in Table 3, respectively. 318 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Orphanides and Wieland stances would not accurately characterize policy and that policy would be easier than suggested by the rule. Even so, we find that the forecast-based rule, which is based on FOMC projections, tracks the federal funds rate target quite well through the first half of 2004 and that the only noticeable deviation is that it would have already called for much more aggressive tightening starting in the second half of 2004 than actually took place. Time-Variation in Natural Rates One might have suspected that the FOMC projections-based rule of thumb, presented in Table 3, could have proved too simple to capture the contours of FOMC decisions during the past 20 years. In that light, the explanatory power of the rule shown in Figure 4 may be considered surprisingly good. One reason to suspect that a rule based on the notional target, (12) fˆt = a0 + aπ π t|t + 3 + auut|t + 3 , might be too simple is the constant intercept. As already mentioned, this would not be of concern if FOMC beliefs regarding its inflation objective and natural rates of interest and unemployment were roughly constant over the estimation sample. If any of the above exhibited time variation, however, a better description of FOMC behavior would be in terms of the following similar, but not identical, rule: ( ) ( ) (13) fˆt = rt* + π t* + aπ π t|t + 3 − π t* + au ut|t +3 − ut* , which suggests a time-varying intercept, a0,t = rt* − (aπ − 1) π t* − auut* . Unfortunately, absent the necessary information required to proxy the FOMC’s real-time assessments of π *, u*, and r* in our sample, it is difficult to examine if a version of the rule allowing for such variation could explain the data even better than the rule of thumb based on equation (12). As a simple check in that direction, however, we reestimated the rule using a possible proxy of the FOMC’s likely perceptions of the natural rate of unemployment, u*. Absent the FOMC’s own F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W assessment, we relied on the real-time estimates published by the Congressional Budget Office (CBO) over the past 20 years. This is the same source of real-time estimates used by Poole (2007) as a proxy for Federal Reserve staff estimates. The results (not shown) were broadly similar to those presented in Table 3 and Figure 4. As with the baseline specification, the data suggest that the FOMC projection-based rule can describe policy decisions quite well. However, the overall fit of our preferred forecast-based regression does not improve with the inclusion of the real-time CBO estimate of the natural rate of unemployment. Rather, the fit deteriorates slightly. Two possible explanations for this are as follows. First, the CBO estimate may not capture the updating patterns of the FOMC’s own real-time estimates of the natural rate. Second, even in the presence of time variation in the natural rate of unemployment, countervailing time variation in the natural rate of interest might keep the intercept in the rule of thumb, a0,t , roughly constant. If so, correcting for the time variation in u* without a parallel correction for the time variation in r* should result in a deterioration in the fit of the rule. Interpreting Changes in the FOMC’s Preferred Inflation Concept Another reason one might be concerned that the rule of thumb based on equation (12), as estimated in Table 3, might be too simple relates to the FOMC’s choice of inflation concept. The decisions of the FOMC to change its inflations projections, for example, from CPI to PCE in 2000 and from PCE to core PCE in 2004, may be due to changes in preference as to the most appropriate concept for the measurement of inflation for policy purposes. To the extent that the typical dynamic behavior of each new measure differs from the one used previously, FOMC members would probably have made adjustments in their systematic response to movements in the inflation measure. To gain some insight into the possible implications of the FOMC turning from the overall CPI measure of inflation, to overall PCE, and then the core PCE measure excluding food and energy J U LY / A U G U S T 2008 319 Orphanides and Wieland prices, we compare the three series in Figure 5. The top panel shows the three series (percentage change in the price index relative to four quarters earlier) for the full 1988-2007 sample. The lower panel provides a detailed view of the most recent 10 years, 1997-2007. As the top panel shows, from 1990 to 1998 the three alternative inflation series steadily declined more or less in lockstep with each other, with the CPI series starting from a higher level than the other two measures. The core PCE seems to best capture the downward trend over this period. The comparison suggests that, ex post, a policy rule could have delivered fairly similar policy implications regardless of which of these inflation measures was used over this period.10 From 1999 onward, the three series exhibit some important differences. For instance, although all three inflation rates indicate rising inflation in 1999, the inflationary surge seemed much stronger in the overall CPI and PCE measures than in the core PCE. In fact, core PCE inflation stayed largely within the Federal Reserve’s so-called “comfort zone” of 1 to 2 percent all the way through 2007. CPI and PCE inflation, however, surged up two more times, in 2002 and in 2004, with CPI inflation reaching 4 percent in 2006. The overall PCE measure more or less follows the movements of the CPI, albeit staying somewhat lower than the CPI throughout. Clearly, the greater increases in PCE and CPI relative to core PCE must have been related to the movements of food and energy prices. These differences pose a challenge in that the different statistical properties of the alternative measures could in principle influence, perhaps in subtle ways, the specification of a rule of thumb. One potential result of the switch from CPI to PCE, for instance, could have been a change in the operational definition of price stability embedded in the rule, that is π *. Stated in PCE terms, π * could be 50 or so basis points lower than the corresponding object stated in CPI terms, reflecting recent estimates of the 50-basis-point average difference in the two series. On the other hand, 10 Note, however, that these series are compared from the July 2007 vintage perspective and not the real-time policymaker perspective. 320 J U LY / A U G U S T 2008 given the uncertainty associated with price measurement and the quantitative definition of price stability most appropriate for monetary policy, it is not entirely clear that such a change in the π * embedded in a rule of thumb should be incorporated in the analysis when the FOMC changes its preferred inflation measure. In light of these uncertainties and the differential movements of core PCE, PCE, and CPI inflation—especially from 2000 onward—we decided to perform two experiments to help examine how changes in the inflation concept potentially influence policy. One way to examine whether the policy rule changed when the FOMC switched inflation measures is to allow for changes in the intercept and/or slope coefficients at those points in time. We did so by introducing the appropriate additive and multiplicative dummy variables in our regression equations and reestimating over the full 1988-2007 sample. We consider possible shifts in 2000:Q1 (for the switch to PCE) as well as in 2004:Q3 (for the switch to core PCE). The results (not shown) did not indicate any significant shifts, suggesting the use of a new inflation measure may not have resulted in a corresponding change in the rule of thumb the FOMC used to make decisions or that, because of the limited sample, the change may have been too small to identify. Another way to examine possible differences since 1999 is to reestimate the regressions presented in Table 3 using only the subsample 1988-99 to see if excluding the period following the switch to PCE and later to core PCE would materially influence the results. The regression estimates, based on equation (11), are reported in Table 4 in identical fashion as those in Table 3. A comparison of Tables 3 and 4 shows that the coefficients of the outcome-based rule change quite a bit. This instability reinforces the prior evidence that the outcome-based rule is misspecified as a description of FOMC policy because it does not account properly for forecasts. The key result in Table 4 is that the estimates corresponding to the forecast-based rule for the subsample ending in 1999 do not materially differ from those corresponding to the full sample. This F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Orphanides and Wieland Figure 5 CPI, PCE, and Core PCE Inflation (vintage July 2007) 1998-2007 6.4 5.6 4.8 4.0 3.2 2.4 CPI PCE Core PCE 1.6 19 88 19 89 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 0.8 1997-2007 4.0 3.5 3.0 2.5 2.0 1.5 CPI PCE Core PCE 1.0 98 19 99 19 99 20 00 20 00 20 01 20 01 20 02 20 02 20 03 20 03 20 04 20 04 20 05 20 05 20 06 20 06 20 07 19 97 98 19 19 19 97 0.5 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 321 Orphanides and Wieland Table 4 Policy Reaction to Inflation and Unemployment Rates: FOMC Forecasts of CPI Inflation, 1988-99 Regressions based on Outcomes Forecasts (1) (2) (3) (4) a0 9.78 1.38 12.73 4.57 6.31 0.99 7.34 1.16 aπ 1.11 0.19 0.72 0.62 2.32 0.20 2.54 0.23 au –1.35 0.25 –1.68 0.71 –1.41 0.17 –1.72 0.22 ρ 0 0 – R2 0.69 0.20 0.41 0.08 0.68 0.78 0.87 0.94 SEE 1.03 0.84 0.64 0.43 SW 0.98 1.18 1.65 1.96 ft = ρ ft −2 + (1 − ρ ) (a0 + aπ π τ|t + auuτ|t ), NOTE: The regressions shown are least-squares estimates of where f denotes the intended federal funds rate, π the inflation rate over four quarters, and u the unemployment rate. The horizon τ either refers to three-quarter-ahead forecasts, τ = t +3, or outcomes observed in the preceding quarter, τ = t –1. suggests that the change in inflation concepts may not have resulted in a corresponding change in the rule of thumb describing FOMC decisions or that this corresponding change may have been rather small. Indeed, this is confirmed in the top panel of Figure 6, which shows the estimated forecast-based rule (dashed line) over the subsample ending in 1999 and a simulation that uses the parameter estimates from this rule together with the FOMC projections through 2007. This simulation confirms that interest rate setting in the 2000-06 period seemed in line with a systematic interest rate response to FOMC projections with the same coefficients, despite the change in inflation concepts. Note that the results for the policy rules do not include interest rate smoothing. This finding is somewhat puzzling, especially in light of the average difference expected in measured inflation in terms of CPI as opposed to PCE or core PCE (approximately 50 basis points). One might have expected that the switch to PCE would be accompanied by a countervailing adjustment in the parameters of the rule. Instead, use of the identical rule with the PCE instead of the CPI, 322 J U LY / A U G U S T 2008 assuming that PCE inflation forecasts are lower on average than corresponding CPI forecasts, would result in lower interest rate prescriptions on average. To get a sense of the magnitude of this effect, we simulated the rule with parameters estimated over the subsample ending in 1999, using the Blue Chip consensus forecasts of CPI inflation from 1988 to 2007. The results, indicated by the dashed line in the lower panel of Figure 6, show that from 1988 to the first half of 2002 the interest rate prescriptions based on the Blue Chip CPI forecasts are broadly in line with those based on the FOMC projections. From the second half of 2002 to 2006, the rule simulated with Blue Chip CPI forecasts implies a higher federal funds rate target. In other words, if the FOMC had continued to forecast CPI inflation and if its forecasts had been similar to those of the Blue Chip consensus from 2002 onward, the FOMC projections-based rule of thumb would have suggested systematically tighter policy than the policy setting suggested with the PCE and core PCE projections. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Orphanides and Wieland Figure 6 Rules Estimated for 1988-99 and Extrapolated to 2007 Simulation Using PCE and Core PCE Inflation 10.0 9.0 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 Fed Funds Outcomes Forecasts 19 88 19 89 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 0.0 Simulation Using CPI Outcomes and Blue Chip CPI Forecasts 10.0 9.0 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 Fed Funds CPI Outcomes Blue Chip CPI Forecasts 19 88 19 89 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 0.0 NOTE: “Fed Funds” refers to the federal funds rate target. “Outcomes” refers to fitted values of the outcome-based rule without interest rates smoothing, that is, column 1 in Table 3. “Forecasts” refers to the fitted values of the forecast-based rule without interest rate smoothing, that is, column 3 in Table 3. In the lower panel, these two rules are simulated with CPI inflation outcomes and Blue Chip CPI forecasts, respectively, from 1988-2007. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 323 Orphanides and Wieland CONCLUSION Many analysts often rely on rules of thumb, such as Taylor rules, to describe historical monetary policy decisions and to compare current policy to historical norms. William Poole’s (1971) study, written explicitly to offer advice to the FOMC, serves as an early example of such work. Analyses along these lines also permit evaluation of episodes where policy may have deviated from a simple policy rule and examination of the reasons behind such deviations. But there is disagreement as to whether the canonical rules of thumb for such work should draw on forecasts or recent outcomes of key variables such as inflation and unemployment. Poole (2007) points out that deviations of the actual funds rate from the prescriptions of a Taylor rule that relies on current readings of inflation and the output gap may be the result of systematic responses of the FOMC to information not contained in these variables. He notes, however, that much of this additional information may be captured in economic projections. We investigate this proposition in the context of FOMC policy decisions over the past 20 years, using publicly available FOMC projections from the Humphrey-Hawkins reports that are published twice a year. Our results indicate that FOMC decisions can be predominantly explained in terms of the FOMC’s own projections rather than recent economic outcomes. Thus, a forecast-based rule better characterizes FOMC decisionmaking. We also identify a difficulty associated with the FOMC switching the inflation concept it has used to communicate its inflation projections. Finally, we confirm that many of the apparent deviations of the federal funds rate from an outcome-based Taylor-style rule may be viewed as systematic responses to information contained in FOMC projections. Bryant, Ralph; Hooper, Peter and Mann, Catherine, eds. Evaluating Monetary Policy Regimes: New Research in Empirical Macroeconomics. Washington, DC: Brookings Institution, 1993. Henderson, Dale and McKibbin, Warwick. “A Comparison of Some Basic Monetary Policy Regimes for Open Economies: Implications of Different Degrees of Instrument Adjustment and Wage Persistence.” Carnegie-Rochester Conference Series on Public Policy, December 1993, 39, pp. 221-318. Lindsey, David; Orphanides, Athanasios and Wieland, Volker. “Monetary Policy Under Federal Reserve Chairmen Volcker and Greenspan: An Exercise in Description.” Unpublished manuscript, Board of Governors of the Federal Reserve System, 1997. Orphanides, Athanasios and Wieland, Volker. “Efficient Monetary Policy Design Near Price Stability.” Journal of the Japanese and International Economies, December 2000, 14, pp. 327-65. Orphanides, Athanasios and Wilcox, David. “The Opportunistic Approach to Disinflation.” International Finance, 2002, 5(1), pp. 47-71. Poole, William. “Rules-of-Thumb for Guiding Monetary Policy,” in Open Market Policies and Operating Procedures—Staff Studies. Washington, DC: Board of Governors of the Federal Reserve System, 1971. Poole, William. “Understanding the Fed.” Federal Reserve Bank of St. Louis Review, January/February 2007, 89(1), pp. 3-13. Taylor, John B. “Discretion versus Policy Rules in Practice.” Carnegie-Rochester Conference Series on Public Policy, December 1993, 39, pp. 195-214. Taylor, John B. “Housing and Monetary Policy.” NBER Working Paper No. 13682, National Bureau of Economic Research, December 2007. REFERENCES Aksoy, Yunus; Orphanides, Athanasios; Small, David; Wieland, Volker and Wilcox, David. “A Quantitative Exploration of the Opportunistic Approach to Disinflation.” Journal of Monetary Economics, November 2006, 53(8), pp. 1877-93. 324 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Charles I. Plosser I t is indeed a pleasure to have the opportunity to be here today at the Thirty-Second Annual Economic Policy Conference of the Federal Reserve Bank of St. Louis. It is a conference I have attended a dozen times or more in various roles. In each and every case, it has proven to be a timely and thoughtful interchange among academic economists and policymakers. I am particularly pleased to be participating this year, since this is the last of these conferences that will be held during Bill Poole’s tenure as president of the Federal Reserve Bank of St. Louis. I have had the privilege of knowing Bill for at least 25 years, perhaps more. I am sure that many of you in the room can also make similar claims. Over the years I have learned a great deal from Bill. His seminal contributions in the area of monetary theory and policy are widespread and span four decades. Whether it be his contributions on monetary policy under uncertainty, his early investigations of simple rules for setting the federal funds rate, or his analysis of rational expectations models of the term structure for monetary policy, his theoretical contributions provided fundamental insights and played an important role in developing what we now view as the core of modern monetary theory. He has continued his contributions to monetary policy as a member of the Federal Open Market Committee (FOMC), bringing the same sound, thoughtful, and consistent economic analysis to policy deliberations. One of the themes of Bill’s work was the importance of uncertainty for monetary policy. One dimension of uncertainty involves our uncertainty regarding the nature of the macroeconomic model that governs the economy. In Poole (1971), Bill investigated the performance of simple “rules of thumb” for setting the federal funds rate. He argued that these simple rules appeared to be “robust” across various model specifications. This line of research has become increasingly active and has some important things to say about the conduct of monetary policy. I find the analysis of simple rules intriguing for a couple of reasons. First and foremost, they are rules. Second, in a framework where policy is decided by committee, simple rules that are robust across models provide a valuable focal point for discussion among people with different world views. Bill’s primary concern dealt with uncertainty in a given model, but he also discussed how optimal policy would vary when model parameters changed. Thus, his concern reflected a desire to analyze optimal policy under model uncertainty. As I indicated, this area of research has seen a resurgence in recent years, with various methodologies ranging from robust control to Bayesian model averaging being employed to analyze optimal policy under uncertainty. Perhaps even more interesting is the research that considers the robustness of simple rules across models that are non-nested and thus potentially very different.1 1 McCallum (1988) was one of the first to investigate the robustness properties of simple rules. Charles I. Plosser is president of the Federal Reserve Bank of Philadelphia. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 325-29. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 325 Plosser Before I discuss the contributions of the Orphanides and Wieland (2008) paper, which I found very interesting, I would like to backtrack and talk about why I believe explorations of rules of thumb are important and what we know about their performance. As I stated at the outset, one of the greatest attractions of simple rules is that they are in fact rules. Since the pioneering work of Kydland and Prescott (1977), we have come to understand the theoretical foundations for the importance of commitment by policymakers. One way that commitment manifests itself is that, in model economies, the optimal monetary policy typically takes the form of a rule. In these models the researcher looks for policies that deliver efficient allocations, that is, the allocations that would be selected by a Ramsey planner. In this context, optimal policy need not be simple, but it does need to be a rule. However, the Fed does not pick allocations like the Ramsey planner—its picks an instrument and moves that instrument to influence economic outcomes. Thus, there are important issues regarding the implementation of policy that must be considered, and this is where I believe that simple policy rules have a role to play. The question then is, why might we choose to adopt simple rules? If everyone had the same model of the economy, there would be no reason to do so. So the underlying attraction of investigating simple rules is twofold. First, everyone may not agree on a common model. Thus, optimal policy for one policymaker may not be optimal for another. Second, even if there is an agreedupon model, the economy is likely to be more complicated than the model, so the optimal policy for that model may not be the optimal policy for the true underlying economy. So it seems natural to ask if there are simple rules that capture the essence of optimal rules and that give good results in a variety of theoretical environments. In other words, how different are the allocations under simple rules from those obtained under optimal policy? How costly is a simple rule? A number of researchers have analyzed the performance of simple rules. One interesting 326 J U LY / A U G U S T 2008 approach is that by Schmitt-Grohé and Uribe (2006). The underlying model is quite rich, incorporating price stickiness, investment adjustment costs, habit persistence, variable capacity utilization, and monopolistic competition. The model considers three types of shocks: policy shocks, total factor productivity shocks, and investmentspecific technology shocks. They find that a simple Taylor-like rule that responds aggressively to inflation, wage growth, and very little to deviations of output growth from target comes very close to achieving the optimal allocations. In fact, a rule that responds solely to price inflation yields good results in this model. The basic message here is that, in a model with large non-neutralities, but primarily forwardlooking agents, a simple Taylor-like rule comes fairly close to implementing Ramsey allocations. Perhaps surprisingly, the properties of the rule place significant weight on an aggressive response to deviations of inflation from target. Although the model of Schmitt-Grohé and Uribe does a fairly good job of matching the data, it may not be the only model to do so. Other policymakers or researchers may wish to stress different model features. For example, I may not be as keen on models with the degree of price stickiness or as comfortable with the adjustment costs and habit formation built into the model, but I do place a lot of stock in models with forwardlooking rational agents. So it is of some interest and importance to question whether these intriguing results are robust to perturbations in the model not considered by the authors. Basically, we want to know how robust these findings are to models that accommodate very different views of behavior, models that perhaps fit the data as well as the one Schmitt-Grohé and Uribe consider. The question of robustness has been addressed in a number of ways. The most common strategy is to look at the performance of simple rules in a host of different, sometimes non-nested models. In these exercises, the most interesting questions in my mind are these: How similar are the optimal rules from different models? And are there simple rules that work well across models? Levin, Wieland, and Williams (1999 and 2003) have explored the performance of various F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Plosser simple rules in a number of model contexts. They characterize the optimal simple rule for each model and find that there are broad similarities that describe the best simple rule. In some cases they find the best simple rule that minimizes the average loss across models. In their 2003 article, this rule responds to smoothed inflation forecasts at most one-year ahead, current inflation, lagged interest rates, and, unlike the Schmitt-Grohé and Uribe analysis, in a significant way to the output gap. In part, this distinction is driven by differences in the loss functions. One feature of the robust rule is that it exhibits inertia in that the coefficient on the lagged interest rate is close to 1. This is in contrast to the results found in SchmittGrohé and Uribe, where inertia was not important. Besides uncertainty resulting from stochastic factors or from not knowing the true model, policymakers and the public may be unaware of the true processes governing the stochastic elements of the model. Orphanides and Williams (2002 and 2007), for example, examine the usefulness of simple rules when the processes for the natural rate of interest and employment are unobserved and not known. Further, agents form forecasts of relevant macroeconomic variables using a learning methodology. The learning mechanism along with persistent errors in estimating natural rates yields highly nonlinear behavior and implies significant departures from the rational expectations equilibrium. An important feature of this model is that there is the possibility that expectations of inflation can become unanchored. The main lesson I take away from their analysis is that, even in this environment, simple rules can work quite well, but those rules should be based on rates of change in output or employment. This avoids the notoriously difficult problems associated with estimating natural rates of output or unemployment. These analyses indicate that in the presence of very different types of uncertainty—stochastic disturbances, uncertainty about the correct model, and uncertainty about the nature of true driving processes—simple rules for monetary policy are able to deliver good economic outcomes. To many observers it comes as somewhat of a surprise that simple rules should do as well as F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W they do. The reason they do is not obvious. It is hard to know whether some form of Occam’s razor is at work or something else. This is somewhat uncomfortable from a theoretical perspective, yet I find the analyses to date convincing and useful from the perspective and experience of a policymaker and economist. Given that simple rules appear to have desirable properties in terms of delivering good outcomes in a variety of theoretical settings, I think they have other desirable characteristics that suggest they are of value to policymakers. First, policy guided by simple rules is easy to monitor and to communicate. Transparency is an important attribute of good monetary policy. In a world where expectations of the future play a critical role in economic outcomes, the transparency and predictability/credibility of monetary policy can reduce expectational errors and contribute to a more stable economic environment. Using simple rules, which can be easily communicated to the public, as benchmarks or guidelines can enhance both transparency and credibility of policy. I believe that there is another benefit from using simple rules as a guide for policy. Specifically, I believe that simple rules serve as a useful focal point in policy deliberations. The underlying models employed by various FOMC members can be quite different and, in some cases, may not even share the same set of state variables. Thus, deliberating the implications of various policy options or the workings of the economy can be quite complicated. Indeed, as I previously discussed, the optimal rule, even from a wellarticulated model, can be quite complex and quite different across models. If the underlying model is not well articulated or completely specified, then we may not even know what the optimal or best rule might be. Thus, trying to reach a consensus on appropriate policy can be difficult. By focusing on simple rules, deliberations can focus on a few key variables and our assumptions or forecasts that shape them. It also leads to a more focused discussion of the shape and parameters of the loss functions that may be applicable. I believe this would greatly improve policy deliberations by directing attention to the key factors J U LY / A U G U S T 2008 327 Plosser that matter for the policy choices. Of course, these benefits arise only if the simple rules have some good robustness properties associated with them. As I have indicated, my reading of the literature to date makes me optimistic that this is indeed the case. Finally, let me turn my attention to the Orphanides and Wieland (2008) paper. How does their paper fit into this broader literature? First, the investigation is a positive analysis rather than a normative one. That is, they seek to study and uncover the characteristics of FOMC decisions over the past 20 years in the context of a simple rule. The rules investigated are simple and, importantly, are based on the real-time information that policymakers actually possess at the time decisions are made. The rule that seems to explain Fed behavior the best is a forward-looking rule that responds aggressively to deviations of inflation from target. In this regard, Fed behavior seems in accord with the guidance of robust rules. If anything, the response seems more aggressive than is sometimes indicated by the more theoretical investigations. The Fed also appears to respond to forecasts of unemployment and its deviations from some natural rate. The authors do not directly estimate a natural rate of unemployment; they allow it to be subsumed in the constant term. This strategy may or may not be a good one. Indeed, some of Orphanides’s own work suggests that looking at growth rate rules might be a better practice, and it would have been interesting to see how they would have stacked up in this comparison. Of course, this might not make too much difference over this period if there was not much movement in the natural rate. Another interesting feature of the results is that the degree of inertia is markedly less when the rules are assumed to be forward looking, that is, based on forecasts, than when they are based on outcomes. The robust rules prescribed in Levin, Wieland, and Williams (1999 and 2003) or Orphanides and Williams (2007) suggest that policy should be more inertial. What might be the reasons for this finding? One possibility is that the robustness results may be relying too 328 J U LY / A U G U S T 2008 heavily on learning and the world may be more rational and forward looking than the models presume. Probably one of the more interesting findings in the paper concerns the shifting focus of the Fed’s preferred inflation measure. The puzzle is that, as the Fed shifted its emphasis from the consumer price index to the core personal consumption expenditures, it apparently did not change the parameters of the estimated rule. This is a puzzle because the personal consumption expenditures measure generally is about 50 basis points below the consumer price index, on average. Thus, changing the inflation measure allowed the Fed to maintain a lower funds rate for a given rate of inflation. I will not speculate on why this is the case. But it does suggest that policy was not as committed to a policy rule as the regression estimates seem to suggest. Overall, I found the paper interesting and useful in furthering our understanding of simple rules. It points to some unresolved issues, particularly those that pertain to the ability to describe actual policy as rule based. That does not mean that simple rules are any less useful or valuable. Indeed, it may suggest that policy can be improved by being more transparent and committed to systematic behavior. REFERENCES Kydland, Finn, and Prescott, Edward. “Rules Rather Than Discretion: The Inconsistency of Optimal Plans.” Journal of Political Economy, June 1977, 85(3), pp. 473-91. Levin, Andrew; Wieland, Volker and Williams, John C. “Robustness of Simple Monetary Policy Rules under Model Uncertainty,” in John Taylor, ed., Monetary Policy Rules. Chicago: University of Chicago Press, 1999. Levin, Andrew; Wieland, Volker and Williams, John C. “The Performance of Forecast-Based Monetary Policy Rules under Model Uncertainty.” American Economic Review, June 2003, 93(3), pp. 622-41. McCallum, Bennett. “Robustness Properties of a Rule for Monetary Policy.” Carnegie Rochester Conference F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Plosser Series on Public Policy, Autumn 1988, 29, pp. 173-203. Orphanides, Athanasios and Williams, John C. “Robust Monetary Policy Rules with Unknown Natural Rates.” Brookings Papers on Economic Activity, 2002, 2, pp. 63-118. Orphanides, Athanasios and Williams, John C. “Robust Monetary Policy with Imperfect Knowledge.” Unpublished manuscript, 2007. Orphanides, Athanasios and Williams, John C. “Economic Projections and Rules of Thumb for Monetary Policy.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 307-24. Poole, William. “Rules of Thumb for Guiding Monetary Policy,” in Open Market Policies and Operating Procedure—Staff Studies. Washington, DC: Board of Governors of the Federal Reserve System, 1971, pp. 135-89. Schmitt-Grohé, Stephanie and Uribe, Martin. “Optimal Simple and Implementable Monetary and Fiscal Rules.” Unpublished manuscript, 2006. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 329 330 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Patrick Minford T he Taylor rule is widely seen as a good summary of what the Federal Reserve does. Though the rule cannot easily be fitted to actual data as subsequently revised, at least for a full postwar sample, it can be fitted to real-time data (i.e., data as seen at the time), as shown by earlier work by Orphanides (2003). But in practice the Fed’s Federal Open Market Committee (FOMC), if it is using a Taylor rule, will look at its own forecasts or projections. Orphanides and Wieland (2008) examine whether a Taylor rule can be fitted to the FOMC’s own projections since 1988. They find that it can with appropriate parameters that satisfy the Taylor principle—that is, that give a unique stable solution under rational expectations. Furthermore, they find that the rule works better with these projections and resolves various puzzles regarding the data on outcomes. This is without question an interesting finding; the paper is clear, cogent, and persuasive. Many will be totally persuaded by it; however, I do have a few doubts. Let me begin with some issues of specification and estimation and then proceed with two wider issues. SPECIFICATION AND ESTIMATION The Specification of the Taylor Projections Rule for Changes in Targets and Definitions As the authors note, there remains a puzzle: In spite of the change in the inflation definitions, particularly that from the consumer price index (CPI) to the personal consumption expenditures (PCE) deflator, their estimates find no shift in the Fed rule. Their experiments with a rule estimated for the CPI in the 1990s shows that the rule should have shifted up on the move to the PCE in the 2000s. The rule might have also shifted with the natural rate of employment; however, when they included this rate in the equation along with the inflation definition, the rule did not shift in line with either or both together. Had the equilibrium rate of interest been known, there may have been no puzzle. However, the authors argue that they had no estimate of this to include as a test; but surely index-linked government bond yields provide some idea of shifting real rate equilibria? This puzzle is particularly odd when viewed side by side with the explicit 0.5 percent shift in target inflation that occurred when the United Kingdom made an essentially similar change— from the retail price index to CPI. The U.S. CPI, too, systematically grows 0.5 percent or so faster than the PCE. The absence of a noticeable shift in the rule makes one wonder exactly what the FOMC projections are—a topic I return to below. The logic of the Taylor projections rule absolutely requires that the rule shifts when the inflation definition changes; this shift should have been imposed on the equation, together with some estimate of changing real interest rate equilibria based on index-linked bond-yield trends. Patrick Minford is professor of applied economics at Cardiff University and a research fellow of the Centre for Economic Policy Research. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 331-38. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 331 Minford Estimation I am concerned with the authors’ estimation. They use ordinary least squares, a single-equation estimator, which is open to bias because of the correlation of the error term (the FOMC’s monetary judgment, or interest rate “shock”) with both endogenous variables. These are defined as lagged variables; but truly they are the FOMC’s current view of the forecast environment at the time interest rates are set and use contemporaneous data and reports on both output and inflation. Given signal extraction and the semiannual frequency of the data, it is clear that current data will influence projections and so the current interest rate judgment; at the same time, the interest rate shock will affect output and inflation in the semiannual time frame. Furthermore, the error is autocorrelated, except in the projections version when a lagged interest rate is included for “adjustment” reasons. However, in principle, even if the FOMC revises its judgment each semiannual period, each new judgment is unlikely to be independent of the last one; given that it represents views on such things as asset price movements, exchange rate behavior, and special factors like the 9/11 attacks. The FOMC’s judgment should show some persistence, and indeed that is what most dynamic stochastic general equilibrium (DSGE) modelers assume about a monetary shock. Given these issues, I regard the estimation methods of this paper as rather casual. For a start, we need more information on the error process; does “adjustment” really eliminate autoregressiveness in the error? Second, we need some effort to estimate the equation in a bias-free manner; fullinformation methods are ruled out by the absence of the rest of the model, but on this front it would be helpful to see some instrumental variable or two-stage least-squares results. Third, however, there are difficulties with any single-equation estimator, as pointed out by Cochrane (2007a). To illustrate his point, consider a standard New Keynesian model with a strict inflation-targeting rule, Rt = ψπt + it (it, the shock, will in general be autocorrelated and also correlated with πt). If we solve the model by imposing a stable solution, inflation is an autoregression, 332 J U LY / A U G U S T 2008 say, πt = ρπt –1 + ut (where the error is also autocorrelated, say, with a root κ ), and it follows that the Fisher identity gives interest rates as Rt = rt + Et πt +1, which thus equals ρπt + [κ ut + rt ], where the term in square brackets is an autocorrelated error, correlated with πt. How can this regression be distinguished from the inflation-targeting regression? A systems estimator imposing all over-identifying restrictions on the model is the only way. Modeling the FOMC Projections Rule To use this FOMC projections rule in a model requires some transfer function relating the Fed’s projections to the actual state of the economy. Thus, if the version here is to be taken seriously as a representation of policy, we need to know its properties in a full model, but of course those properties depend on how the FOMC projections are related to the actual economy. It matters a lot whether they are, for example, biased and/or subject to learning or rational expectations. A key reason for knowing these details is that they would make it possible to estimate the rule appropriately by full-information methods, as already argued. SOME WIDER ISSUES There are some wider issues I see as interestingly raised by this paper. The first is what exactly the Taylor rule is and how it fits into economic thought on policy rules. The second is whether this paper and associated work clinches the debate on which monetary rule was actually being pursued by the FOMC; I will argue that this turns on a difficult issue of identification. What Exactly Is the Taylor Rule? Origins and Application John Taylor wrote his paper (Taylor, 1993) proposing the rule in the early 1990s. It seems to have been heavily influenced by a 1989/90 Brookings conference event, which discussed the performance of different monetary rules (money supply, exchange rate targeting, or pegging, mainly) within large models of the world F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Minford economy, one of which was John Taylor’s “Taylor World Model” (another was my “Liverpool World Model”). As a new departure, Dale Henderson and Warwick McKibbin asked the modelers— around a dozen teams—to evaluate a new suggestion that money be bypassed by setting interest rates directly in response to macro data. Various formulations were tried. The modeling teams drew a blank initially in solving their models under these rules; it seems that they were tripping over indeterminacy and had not discovered the Taylor principle, but it may also have been that the algorithms being used at that time (mostly variants of the Fair and Taylor, 1993, method) simply had difficulty homing in on the solution. These proposed rules, we may well now have forgotten, were a quite unfamiliar way of thinking about monetary policy. It is true that rules for setting interest rates had had a long history (as pointed out by Stanley Black at this Federal Reserve Bank of St. Louis conference); indeed such rules were dominant in the postwar Keynesian era up to the 1970s. However, there was a strong reaction against such ideas in the late 1970s and 1980s as the rational expectations revolution took effect; interest rate rules were felt to give a poor nominal anchor (and would give rise to indeterminacy unless tied to a nominal target) and instead the setting of the money supply was emphasized. This accounts for the fact that the primary rules investigated in the Brookings conference were either money supply rules or exchange rate rules. When the teams had succeeded in solving their models for these new rules, they were found to perform surprisingly well and the results were written up by Henderson and McKibbin (1993a,b) at great length (1993a is in Bryant, Hooper, and Mann, 1993, Chap. 2; 1993b was a version of this chapter given at the same Carnegie-Rochester conference where Taylor presented his own paper, Taylor, 1993). It seems that the success of these rules in a wide variety of models indicated a surprising robustness, and it was this robustness that Taylor later emphasized as a major attraction of his own rule. He elaborated on this in further tests on other models. After the Brookings conference, in any case, John Taylor formulated his rule, F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W which could reasonably be termed the HendersonMcKibbin-Taylor rule. Nevertheless, there seems to be a difference between these authors’ views. Whereas Henderson and McKibbin were solely discussing what would be a good rule and never, as far as I am aware, argued that it was actually pursued, John Taylor went further and argued not only that it worked well but also that monetary policy could be thought of as being done this way. A paraphrase of his distinctive message could be “Look, here is a rough approximation of what a good central bank actually does and has done in the United States in recent years.” Thus, the attraction of the Taylor rule was that it was descriptive as well as normative; this was the new ingredient that Taylor added.1 Orphanides has in his earlier (2003) work argued that it can indeed describe FOMC behavior for the whole postwar period if real-time data are used. Yet, as I shall argue below, it is this implicit claim that the rule is descriptive that is problematic. We can pursue this history further with a review of how New Keynesian authors use the Taylor rule to account for inflation in the postwar period. Here, I follow the points made by Cochrane (2007b). He notes that these authors (e.g., Clarida, Galí, and Gertler, 2000) have argued that up to around 1980 the Taylor rule being pursued by the Fed violated the Taylor principle and thus produced or permitted high inflation; after 1980, the Fed raised the coefficient on inflation above unity and inflation was brought down. Yet, if the Fed before 1980 had such a Taylor rule, then inflation would have been indeterminate. So, in what sense does this account for any inflation path at all?2 (This is resolved by Orphanides, who says that, throughout, the Fed had a good rule but just 1 Yet, there is ambivalence even here. For example, McCallum stated in answer to my question at the Carnegie-Rochester 2002 conference on the Taylor rule that the rule was essentially normative, not descriptive. Ireland (2003), at the same conference, however, took the view that it was both a normative rule (enabling monetary economists to coalesce around inflation targeting after years of wrangling about other rules) and positive, in that central banks actually thought of policy in terms of Taylor rules. 2 The Taylor principle and this stable-sunspot corrollary can be illustrated for a simple model in which real interest rates are an exogenous AR(1) process (more complex models can produce J U LY / A U G U S T 2008 333 Minford had bad estimates of the output gap in the 1970s; to account for inflation, then, a full model including private sector information and learning is needed, which then makes this a branch of the learning literature and not a rational expectations model like the New Keynesian one.) For the post1980 period, Cochrane (2007b) argues that the way the rule works to discipline inflation is in any case incredible: In effect, the Fed threatens to raise inflation and interest rates without limit should inflation deviate from the stable path. Because people believe this threat, inflation goes to this unique path. Yet, what stops them from choosing one of these deviant paths, so that the Fed has to go along with them? Deviant paths in models with money supply targets can be suppressed by Fed action on the money supply; here it is not clear what the Fed will do to rule out deviant paths. Thus, there is a doctrinal puzzle in the Taylor rule approach. The Taylor rule emerged from a money-supply-rule world because models were found to behave rather well when the rule was imposed together with some unspecified device to rule out unstable paths. However, it was forgotten that in previous models that device had involved action on the money supply. I think what this shows is that the Taylor rule is an essentially incomplete statement about monetary policy. One has to assume that the authorities have some additional tool in their locker to rule out unstable paths. Cochrane (2007b) argues this can be a non-Ricardian fiscal policy. It could also be a money supply policy of the central bank. Does This Work Compel Us To Believe the Fed Really Was Pursuing a Taylor Rule? Identification Across Possible Models. The problem with the claim that the Fed projections slightly different Taylor conditions): rt = ρrt –1 + εt . Now add a Taylor rule for inflation only, Rt = απt, and the Fisher identity, Rt = rt + Etπt +1. The general solution of the model is πt = krt + ξt , where k = 共1/α – ρ兲 and the sunspot ξt = αξt –1 + ηt with ηt chosen randomly (the solution can be verified by substituting it into rt = –Etπt +1 + απt ). If α ≥ 1, then the sunspot is ruled out by the condition that the solution must be stable. But if α < 1, then inflation is a stable process with a sunspot and hence indeterminate in that each period the path can jump anywhere. 334 J U LY / A U G U S T 2008 rule is descriptive is in a general sense one of identification across possible models. DSGE models give rise to the same correlations between interest rates and inflation, even if the Fed is doing something quite different, such as targeting the money supply. For example, Minford, Perugini, and Srinivasan (2002 and 2003) show this in a DSGE model with Fischer wage contracts. Gillman, Le, and Minford (2007) use a real business cycle growth model with cash and credit in advance to derive a steady-state, or cointegrating, relation between interest rates and inflation and the growth rate when money supply growth is fixed—a “speed limit” version of the rule. The route they use to obtain an apparent Taylor rule is the Fisher equation, which links nominal interest rates with expected future inflation and real interest rates; they then use the relation elsewhere in the model, equating growth with the real interest rates to obtain a “Taylor rule.” This identification would still be a problem under the projections rule because of how the transfer function relates the actual data to the projections; that is, any relationship between interest rates and FOMC projections could be translated by this function into one between interest rates and actual data. I will return below to what this transfer function might look like. For now, let us just compare the normal Taylor rule using actual data with the other rule’s implied Taylor-type equation. To illustrate the point in detail, consider a popular DSGE model but with a money supply rule instead of a Taylor rule: (IS) y t = γ E t −1 y t +1 − φ rt + v t ( ) (Phillips) π t = ζ y t − y ∗ + ν E t −1π t +1 + (1 − ν ) πt −1 + ut (Money supply target) ∆mt = m + µt (Money demand) mt − pt = ψ 1E t −1 y t +1 − ψ 2 Rt + εt (Fisher identity) Rt = rt + E t −1πt +1 This model implies a Taylor-type relationship that looks like F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W ( ) ( ) Rt = r ∗ + π ∗ + γ χ−1 π t − π ∗ + ψ 1 χ −1 y t − y ∗ + w t , where χ = ψ2γ – ψ1φ and the error term, wt , is both correlated with inflation and output and autocorrelated; it contains the current money supply/demand and aggregate demand shocks and also various lagged values (the change in lagged expected future inflation, interest rates, the output gap, the money demand shock, and the aggregate demand shock). This particular Taylor-type relation was created with a combination of equations—the solution of the money demand and supply curves for interest rates, the Fisher identity, and the IS curve for expected future output.3 But other Taylor-type relationships could be created with combinations of other equations, including the solution equations, generated by the model. They will all exhibit autocorrelation and contemporaneous correlation with output and inflation, clearly of different sorts depending on the combinations used. Identification is of course a quite separate matter from estimation; the usual assumption is that we have infinite amounts of data to carry out completely accurate estimation. In fact, OLS estimation would be inappropriate, as we have seen, because it forces the error term to be orthogonal 3 From the money demand and money supply equation, ψ 2 ∆Rt = π t − m + ψ 1∆Et −1y t +1 + ∆εt − µt : Substitute Et –1yt +1 from the IS curve and then inside that for real interest rates from the Fisher identity, giving ψ 2 ∆Rt = πt − m + ψ 1 (ψ ( ){φ ( ∆R − ∆E 1 γ t ) ∆ (R − R ) t −1π t +1 ) + ∆y t − ∆vt } + ∆εt − µt ; then, rearrange this as − ψ 1φ γ = (π t − m ) − 2 t ψ 1φ γ ∗ ∆Et −1π t +1 + ψ1 γ ( ) ∆ yt − y ∗ − ψ1 γ ∆vt + ∆εt − µt , where the constants R* and y* have been subtracted from Rt and yt, respectively, exploiting the fact that when differenced they disappear. Finally, obtain ( ) ( Rt = r ∗ + π ∗ + γ χ−1 π t − π ∗ + ψ 1 χ −1 y t − y ∗ ( ) ( ) Rt −1 − R ∗ − ψ 1φχ −1∆Et −1π t +1 − ψ 1 χ −1 y t −1 − y ∗ + −ψ 1 χ −1∆v t + γχ −1 ∆εt − γχ −1µt ), where we have used the steady-state property that R* = r* + π* and m = π*. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Minford to the regressors, yet because this cannot be the case, it induces bias. Instead, estimation is done by a full-information estimator, which allows for the model’s simultaneity, including of the error term in this equation. With infinite data, we retrieve the parameters exactly and also the error terms. The error term in the Taylor rule proper is, as we have seen, the “monetary shock” created by FOMC special judgments on current events. This is, therefore, like the errors in the Taylor-type relationships, correlated with current events, including the output gap and inflation, both because these influence FOMC judgments (even if they do not observe the correct values, they know enough to extract signals from current reports, snapshot statistics, etc.) and because these shocks may affect current output and inflation. Distinguishing between the two equations is likely to be difficult in general. The error terms of both the Taylor rule and Taylor-type relations are autocorrelated and correlated with output and inflation. The coefficients on output and inflation in both are positive and that on inflation in the Taylor rule will be higher than the one in the Taylor-type relation if ψ2γ – ψ1φ is less than γ . The constant in both is the steady-state value of inflation plus the real rate of interest. Identification by “Narrative Evidence” and by Projections? Could we nevertheless be confident that there is a Taylor rule because of what we definitely know about policymakers’ behavior (what we might call narrative evidence)? In his replies to my comments, Athanasios Orphanides stated that FOMC minutes during this sample period (from 1988) supported the interpretation that the projections determined interest rate setting. However, the problem is that we cannot see directly in this way what FOMC policymakers were doing. They vote and there are minutes, but we do not know what they are really trying to do. We are familiar from psychology that people may describe their actions in one way when in truth they are being compelled to act (in a “deterministic” way) by other forces; also there may be reasons of prudence or politics that lead people to disguise the motives for their actions. Even when there is a legal objective, as in the United Kingdom, policymakers pursue all sorts J U LY / A U G U S T 2008 335 Minford of private agendas. Thus, in the United Kingdom recently we have had different members of the Bank of England Monetary Policy Committee being particularly concerned with measures like house prices, other asset prices, the state of the labor market, and latterly “moral hazard.” All these have jostled in the voting for a place in interest rate setting. Furthermore, there have been many phases in U.S. policy, as in U.K. policy. Under Bretton Woods, the dollar’s fixed rate against the Deutsche mark put some brakes on U.S. policy. After the end of Bretton Woods, leading to the Louvre and Plaza accords there were still flurries of concern with exchange rates; intermittently right up to present times there has been policy concern with the current account deficit and the need for exchange-rate movement. In 1979-81 there was a big debate about money supply targets and an episode of reserve targeting. Congress mandated that the Fed give an account of its efforts to hit various money supply targets in the 1970s and 1980s. Electoral pressures seem to have played a part at times. Further, we know that for much of the earlier postwar period some policymakers believed that inflation could be contained by wage/price controls and interest rates could be used to bring down unemployment. Even in recent times, influential policymakers have been opposed to an inflation target—including some policymakers inside the Fed itself—on the grounds that there needs to be “flexibility” to deal with unemployment. Finally, I note that the Fed, more or less now alone among central banks within the Organisation for Economic Co-operation and Development, does not have a formal inflation target set by law. This certainly makes it harder, even in this recent sample period from 1988, to use narrative evidence to identify the FOMC’s rule. Can We Be Confident Because We See Such a Close Correlation Between Projections and Interest Rates? It may be argued that such a high correlation (an R 2 of over 90 percent) proves beyond doubt that Fed governors were using their projections to produce their view on interest rates. This too is problematic; indeed such a high 336 J U LY / A U G U S T 2008 R 2 arouses suspicion.4 We do not know how these projections are produced, only that each governor sends them to the meeting having produced them with the help of his or her staff. They are then cropped and averaged to give the published values for the Humphrey-Hawkins legislation’s requirements. A reasonable suspicion would be that the fit is so close because the governors want to present a plausible public case for their views on interest rates; hence, governors that wish to raise rates will generate forecasts of higher inflation and/or higher output gap (overheating). Their reasons for raising rates may be quite different from these. Thus, their projections are molded by their views, not as assumed here, the other way round, views by projections. On this skeptical view of such a close fit, we have no evidence of what was driving the governors’ views. It could be that they are closet monetarists. It could be that they worry about asset prices or their latest regional data—any number of things. In the end, it still comes out looking like they follow a Taylor projections rule. This way of thinking about FOMC decisions could account for the lack of shift in the inflation forecasts after the change from CPI to PCE: If the governors are just rationalizing their interest rate decisions by producing projections, they will choose numbers not in the spirit of a good forecast but more in order to signal clearly the need they perceive to raise or lower rates. The actual number would be of little significance; the direction would be solely what mattered. Consider now what the transfer function might look like. It translates the governors’ average inflation and unemployment projections into the state variables producing them. Hence, these variables would be a mixture that could include domestic asset prices, the exchange rate, the money supply, unemployment and its dispersion—any variables that governors believe would trigger their desired interest rate change. 4 I owe this point to Clemens Kool. In his conference comment, Steven Cecchetti (2008) also questioned the meaning of these forecasts. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Minford CONCLUSIONS This interesting paper shows that, if one thinks the Taylor rule definitely describes the FOMC’s behavior over the past two decades, then a rather convincing relationship can be found, though there are concerns about estimation, how the transfer function relates projections to the actual data, and the puzzling lack of shift in the projections in response to well-known shifts in the environment. Yet the Taylor rule, as its intellectual history suggests, is an incomplete description of monetary policy, at least within a New Keynesian model; it cannot account for determinate inflation before 1980, and after 1980 it lacks a clear mechanism for ruling out unstable paths. If one is not a priori convinced it describes the FOMC’s behavior in the past two decades, then there is a nontrivial issue of identification: Taylor-type relationships can emerge from a DSGE model where no Taylor rule is guiding monetary policy. To test the Taylor rule descriptive hypothesis convincingly, one really needs to compare results for a full model with alternative formulations of monetary policy. That way we can see whether the data rejects one or other policy formulation when embedded in a full-model structure. REFERENCES Bryant, Ralph; Hooper, Peter and Mann, Catherine, eds. Evaluating Monetary Policy Regimes: New Research in Empirical Macroeconomics. Washington, DC: Brookings Institution Press, 1993. Clarida, Richard; Galí, Jordi and Gertler, Mark. “Monetary Policy Rules and Macroeconomic Stability: Evidence and Some Theory.” Quarterly Journal of Economics, February 2000, 115(1), pp. 147-180. Cochrane, John H. “Identification with Taylor Rules: A Critical Review.” NBER Working Paper No. 13410, National Bureau of Economic Research, September 2007a. Cochrane, John H. “Inflation Determination with Taylor Rules: A Critical Review.” NBER Working F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Paper No. 13409, National Bureau of Economic Research, September 2007b. Fair, Ray C. and Taylor, John B. “Solution and Maximum Likelihood Estimation of Nonlinear Rational Expectations Models.” Econometrica, July 1983, 51(4), pp. 1169-86. Gillman, Max; Le, Vo Phuong Mai and Minford, Patrick. “An Endogenous Taylor Condition in an Endogenous Growth Monetary Policy Model.” Economics Working Paper E2007/29, Cardiff University, 2007. Henderson, Dale W. and McKibbin, Warwick J. “An Assessment of Some Basic Monetary Policy Regime Pairs: Analytical and Simulation Results from Simple Multi-Region Macroeconomic Models,” in Ralph Bryant, Peter Hooper, and Catherine Mann, eds., Evaluating Monetary Policy Regimes: New Research in Empirical Macroeconomics. Chap. 2. Washington, DC: Brookings Institution Press, 1993a, pp. 45-218. Henderson, Dale W. and McKibbin, Warwick J. “A Comparison of Some Basic Monetary Policy Regimes for Open Economies: Implications of Different Degrees of Instrument Adjustment and Wage Persistence.” Carnegie-Rochester Conference Series on Public Policy, December 1993b, 39, pp. 221-317. Ireland, Peter N. “Robust Monetary Policy with Competing Reference Models: Comment.” Journal of Monetary Economics, July 2003, 50(5), pp. 977-82. Minford, Patrick; Perugini, Francesco and Srinivasan, Naveen. “Are Interest Rate Regressions Evidence for a Taylor Rule?” Economics Letters, 2002, 76(1), pp. 145-50. Minford, Patrick; Perugini, Francesco and Srinivasan, Naveen. “How Different Are Money Supply Rules from Taylor Rules?” Indian Economic Review, JulyDecember 2003, 38(2), pp. 157-166; published version of “The Observational Equivalence of the Taylor Rule and Taylor-Type Rules.” CEPR Working Paper 2959, Centre for Economic Policy Research, 2001. J U LY / A U G U S T 2008 337 Minford Orphanides, Athanasios. “Historical Monetary Policy Analysis and the Taylor Rule.” Journal of Monetary Economics, July 2003, 50(5), pp. 983-1022. Orphanides, Athanasios and Wieland, Volker. “Economic Projections and Rules of Thumb for Monetary Policy.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 307-24. Taylor, John B. “Discretion versus Policy Rules in Practice.” Carnegie-Rochester Conference Series on Public Policy, December 1993, 39, pp. 195-214. 338 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W House Prices and the Stance of Monetary Policy Marek Jarociński and Frank R. Smets This paper estimates a Bayesian vector autoregression for the U.S. economy that includes a housing sector and addresses the following questions: Can developments in the housing sector be explained on the basis of developments in real and nominal gross domestic product and interest rates? What are the effects of housing demand shocks on the economy? How does monetary policy affect the housing market? What are the implications of house price developments for the stance of monetary policy? Regarding the latter question, we implement a Céspedes et al. (2006) version of a monetary conditions index. (JEL E3 E4) Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 339-65. T he current financial turmoil, triggered by increasing defaults in the subprime mortgage market in the United States, has reignited the debate about the effect of the housing market on the economy at large and about how monetary policy should respond to booming house prices.1 Reviewing the role of housing investment in post-WWII business cycles in the United States, Leamer (2007, p. 53) concludes that “problems in housing investment have contributed 26% of the weakness in the economy in the year before the eight recessions” and suggests that, in the most recent boom and bust period, highly stimulative monetary policy by the Fed first contributed to a booming housing market and subsequently led to an abrupt contraction as the yield curve inverted. Similarly, using counterfactual simulations, Taylor (2007) 1 See the papers presented at the August 30–September 1, 2007, Federal Reserve Bank of Kansas City economic symposium Housing, Housing Finance, and Monetary Policy in Jackson Hole, Wyoming; http://www.kc.frb.org/home/subwebnav.cfm?level=3&theID=105 48&SubWeb=5. A literature survey is presented in Mishkin (2007). shows that the period of exceptionally low shortterm interest rates in 2003 and 2004 (compared with a Taylor rule) may have substantially contributed to the boom in housing starts and may have led to an upward spiral of higher house prices, falling delinquency and foreclosure rates, more favorable credit ratings and financing conditions, and higher demand for housing. As the short-term interest rates returned to normal levels, housing demand fell rapidly, bringing down both construction and house price inflation. In contrast, Mishkin (2007) illustrates the limited ability of standard models to explain the most recent housing developments and emphasizes the uncertainty associated with housing-related monetary transmission channels. He also warns against leaning against rapidly increasing house prices over and above their effects on the outlook for economic activity and inflation and suggests instead a preemptive easing of policy when a house price bubble bursts, to avoid a large loss in economic activity. Even more recently, Kohn (2007, p. 3) says Marek Jarociński is an economist and Frank R. Smets is Deputy Director General of Research at the European Central Bank. The authors thank their discussants, Bob King and Steve Cecchetti, for their insightful comments. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, the regional Federal Reserve Banks, or the European Central Bank. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 339 Jarociński and Smets I suspect that, when studies are done with cooler reflection, the causes of the swing in house prices will be seen as less a consequence of monetary policy and more a result of the emotions of excessive optimism followed by fear experienced every so often in the marketplace through the ages…Low policy interest rates early in this decade helped feed the initial rise in house prices. However, the worst excesses in the market probably occurred when short-term interest rates were already well on their way to more normal levels, but longer-term rates were held down by a variety of forces. In this paper, we review the role of the housing market and monetary policy in U.S. business cycles since the second half of the 1980s using an identified Bayesian vector autoregressive (BVAR) model. We focus on the past two decades for a number of reasons. First, following the “Great Inflation” of the 1970s, inflation measured by the gross domestic product (GDP) deflator has been relatively stable between 0 and 4 percent since the mid-1980s. As discussed by Clarida, Galí, and Gertler (1999) and many others, this is likely partly the result of a more systematic monetary policy approach geared at maintaining price stability. Second, there is significant evidence that the volatility of real GDP growth has fallen since 1984 (e.g., McConnell and Pérez-Quirós, 2000). An important component of this fall in volatility has been a fall in the volatility of housing investment. Moreover, Mojon (2007) has shown that a major contribution to the “Great Moderation” has been a fall in the correlation between interest rate– sensitive consumer investment, such as housing investment, and the other components of GDP. This suggests that the role of housing investment in the business cycle may have changed since the deregulation of the mortgage market in the early 1980s. Indeed, Dynan, Elmendorf, and Sichel (2005) find that the interest rate sensitivity of housing investment has fallen over this period. We use BVAR to perform three exercises. First, we analyze the housing boom and bust in the new millennium using conditional forecasts by asking this question: Conditional on the esti340 J U LY / A U G U S T 2008 mated model, can we forecast the housing boom and bust based on observed real GDP, prices, and short- and long-term interest rate developments? This is a first attempt at understanding the sources of the swing in residential construction and house prices in the new millennium. In the benchmark VAR, our finding is that housing market developments can only partially be explained by nominal and real GDP developments. In particular, the strong rise in house prices in 2000 and the peak of house prices in 2006 cannot be explained. Adding the federal funds rate to the information set helps forecast the housing boom. Interestingly, most of the variations in the term spread can also be explained on the basis of the short-term interest rate, but there is some evidence of a long-term interest rate conundrum in 2005 and 2006. As a result, observing the long-term interest rate also provides some additional information to explain the boom in house prices. Second, using a mixture of zero and sign restrictions, we identify the effects of housing demand, monetary policy, and term spread shocks on the economy. We find that the effects of housing demand and monetary policy shocks are broadly in line with the existing empirical literature. We also analyze whether these shocks help explain the housing boom and its effect on the wider economy. We find that both housing market and monetary policy shocks explain a significant fraction of the construction and house price boom, but their effects on overall GDP growth and inflation are relatively contained. Finally, in the light of the above findings and following a methodology proposed by Céspedes et al. (2006), we explore the use of a monetary conditions index (MCI), which includes the federal funds rate, the long-term interest rate spread, and real house prices, to measure the stance of monetary policy. The idea of measuring monetary conditions by taking an appropriate weight of financial asset prices was pioneered by the Bank of Canada and the Reserve Bank of New Zealand in the 1990s. As both countries are small open economies, these central banks worried about how changes in the value of the exchange rate may F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets affect the monetary policy stance.2 The idea was to construct a weighted index of the short-term interest rate and the exchange rate, where the weights reflected the relative effect of those monetary conditions on an intermediate or final target variable, such as the output gap, output growth, or inflation. A number of authors have extended the idea of an MCI to other asset prices, arguing that those asset prices may be equally or more important than the exchange rate. A prominent example is Goodhart and Hofmann (2007), who argue that real house prices should receive a significant weight because of their large effect on the economy and inflation in particular. In contrast to this literature, the crucial feature of the MCI methodology proposed by Céspedes et al. (2006) is that it takes into account that interest rates and house prices are endogenous variables that systematically respond to the state of the economy. As a result, their MCI can more naturally be interpreted as a measure of the monetary policy stance. Using the identified BVAR, we apply the methodology to question whether the rise in house prices and the fall in long-term interest rates led to an implicit easing of monetary policy in the United States. In the next section, we present two estimated BVAR specifications. We then use both BVARs to calculate conditional forecasts of the housing market boom and bust in the new millennium. In the third section, we identify housing demand, monetary policy, and term spread shocks and investigate their effect on the U.S. economy. Finally, in the fourth section we develop MCIs and show using a simple analytical example how the methodology works and why it is important to take into account the endogeneity of short- and long-term interest rates and house prices with respect to the state of the economy. We then use the estimated BVARs to address whether long-term interest rates and house prices play a significant role in measuring the stance of monetary policy. A final section contains some conclusions and discusses some of the shortcomings and areas for future research. 2 See, for example, Freedman (1994 and 1995a,b) and Duguay (1994). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W A BVAR WITH HOUSING FOR THE U.S. ECONOMY In this section, we present the results from estimating a nine-variable BVAR of order five for the U.S. economy. In addition to standard variables, such as real GDP, the GDP deflator, commodity prices, the federal funds rate, and M2, we include real consumption, real housing investment, real house prices, and the long-term interest rate spread. To measure house price inflation, we use the nationwide Case-Shiller house price index, which limits our sample to 1987:Q1-2007:Q2. The two estimated BVAR specifications are as follows: One is a traditional VAR in levels (LVAR) that uses a standard Minnesota prior. The other is a differences VAR (DVAR) that is specified in growth rates and uses priors about the steady state (see Villani, 2008). More specifically, in the LVAR, the vector of endogenous variables is given by (1) y t ct pt HI t / Yt hpt − pt cpt it st mt , where all variables are in logs, with the exception of the federal funds rate (it), the long-term interest rate spread (st ), and the housing investment share of GDP (HIt /Yt); yt is real GDP; ct is real consumption; pt is the GDP deflator; hpt is house prices; cpt is commodity prices; and mt is the money stock.3 In the DVAR, the vector of endogenous variables is instead given by (2) ∆yt ∆ct ∆pt HI t / Yt ∆hpt − ∆pt ∆cpt it st ∆mt , where ∆ is the difference operator and the BVAR is parameterized in terms of deviations from steady state. The main difference between the two specifications is related to the assumptions one makes about the steady state of the endogenous variables. The advantage of the DVAR with a prior on the joint steady state is that it guarantees that the growth rates are reasonable and mutually consistent in the long run, in spite of the short sample 3 See the data appendix for the sources of the time series. J U LY / A U G U S T 2008 341 Jarociński and Smets Table 1 Prior and Posterior Means and Standard Deviations of the Steady States in the DVAR Real GDP Real GDP consumption deflator growth growth inflation Variable Housing investment/ GDP House price growth Commodity price growth Federal funds rate Term spread Money growth Prior mean 2.50 2.50 2.00 4.50 0.00 2.00 4.50 1.00 4.50 Standard deviation 0.50 0.71 0.20 1.00 2.00 2.00 0.62 1.00 1.00 Posterior mean 2.96 3.23 2.21 4.51 1.52 2.00 5.05 1.42 4.35 Standard deviation 0.22 0.22 0.15 0.07 1.08 1.54 0.34 0.24 0.51 used in the estimation. The cost is that it discards important sample information contained in the LVAR variables. As we discuss below, this may be the main reason behind the larger error bands around the DVAR impulse responses and conditional projections. Although the forecasts of the LVAR match the data better at shorter horizons, the longer-run unconditional forecasts it produces make less sense from an economic point of view. Because these considerations may matter for assessing the monetary policy stance, we report the findings using both specifications. In both cases the estimation is Bayesian. In the case of the DVAR, it involves specifying a prior on the steady state of the VAR and a Minnesota prior on dynamic coefficients, as introduced in Villani (2008). The Minnesota prior uses standard settings, which are the same as the settings used for the LVAR. In the DVAR, the informative prior on the steady state serves two roles: First, it regularizes the inference on the steady states of variables. Without it, the posterior distribution of the steady states is ill-specified because of the singularity at the unit root. Second, and this is our innovation with respect to the approach of Villani (2008), through it we use economic theory to specify prior correlations between steady states. The steady-state nominal interest rate is, by the Fisher equation, required to be the sum of the steady-state inflation rate and the equilibrium real interest rate. The steady-state real interest rate is, in turn, required to be equal to the steadystate output growth rate plus a small error reflecting time preference and a risk premium. The steady-state output and consumption growth 342 J U LY / A U G U S T 2008 rates are also correlated a priori, as we think of them as having a common steady state. The prior and posterior means and standard deviations of the steady states in the DVAR are given in Table 1. Figure 1 plots the data we use, as well as their estimated steady-state values from the DVAR. The steady-state growth rate of real GDP is estimated to be close to 3 percent over the sample period. Average GDP deflator inflation is somewhat above 2 percent. The steady-state housing investment– to-GDP ratio is about 4.5 percent. During the new millennium construction boom, the ratio rose by 1 percentage point, peaking at 5.5 percent in 2005 before dropping below its long-term average in the second quarter of 2007. Developments in real house prices mirror the developments in the construction sector. The estimated steady-state real growth rate of house prices is 1.5 percent over the sample period. However, changes in real house prices were negative during the early-1990s recession. The growth rate of house prices rose above average in the late 1990s and accelerated significantly above its estimated steady state, reaching a maximum annual growth rate of more than 10 percent in 2005 before falling abruptly to negative growth rates in 2006 and 2007. Turning to interest rate developments, the estimated steadystate nominal interest rate is around 5 percent. The estimated steady-state term spread, that is, the difference between the 10-year bond yield rate and the federal funds rate, is 1.4 percent. In the analysis below, we will focus mostly on the boom and bust period in the housing market starting in 2000. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets Figure 1 Data Used and Their Estimated Steady-State Values from the DVAR Output 8 6 4 2 0 Data Mean 16 Percentile 84 Percentile –2 –4 1990 5.6 5.4 5.2 5.0 4.8 4.6 4.4 4.2 4.0 3.8 3.6 1995 2000 Consumption 7 6 5 4 3 2 1 0 –1 –2 –3 2005 1990 Housing Investment 1995 2000 2005 House Prices 15 5 0 –5 –10 –15 1995 2000 2005 Interest Rate 10 9 8 7 6 5 4 3 2 1 0 1990 1995 2000 1990 1995 1995 2000 2005 Commodity Prices 1990 1995 2000 2005 Money 12 10 8 6 4 2 0 –2 –4 1990 1995 Using both BVAR specifications, we then ask the following question: Can we explain developments in the housing market based on observed developments in real and nominal GDP and the short- and long-term interest rates? To answer this question we make use of the conditional forecasting methodology developed by Doan, Litterman, and Sims (1984) and Waggoner and Zha (1999). Figures 2A and 2B report the results for the DVAR and the LVAR, respectively, focusing on the post-2000 period. Each figure shows the actual developments of the housing investment–to-GDP ratio (first column) and the annual real growth F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W 2005 Spread 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0 –0.5 –1.0 2005 2000 1990 50 40 30 20 10 0 –10 –20 –30 –40 10 1990 Prices 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 2000 2005 1990 1995 2000 2005 rate of house prices (second column). Dotted black lines denote unconditional forecasts, and blue lines denote conditional forecasts, conditioning on observed real and nominal GDP (first row), observed real and nominal GDP and the federal funds rate (second row), and observed real and nominal GDP, the federal funds rate, and the term spread (third row). Note that this is an in-sample analysis in that the respective VARs are estimated over the full sample period. The idea behind increasing the information set is to see to what extent short- and long-term interest rates provide information about developments in the housing J U LY / A U G U S T 2008 343 Jarociński and Smets Figure 2A Housing Investment–to-GDP Ratio and Annual House Price Growth Rate, 1995-2007: Actual Data and Unconditional and Conditional Forecasts from the DVAR 5.6 Housing Investment Conditional on y,p Conditional Forecast Mean 16 Percentile 84 Percentile Unconditional Forecast Mean Actual 5.4 5.2 5.0 4.8 4.6 10 5 0 4.4 4.2 –5 4.0 5.6 House Prices Conditional on y,p 96-01 98-01 00-01 02-01 04-01 06-01 96-01 98-01 00-01 02-01 04-01 06-01 Housing Investment Conditional on y,p,i House Prices Conditional on y,p,i 5.4 10 5.2 5.0 5 4.8 4.6 0 4.4 4.2 –5 4.0 96-01 98-01 00-01 02-01 04-01 06-01 96-01 98-01 00-01 02-01 04-01 06-01 Housing Investment Conditional on y,p,i,s House Prices Conditional on y,p,i,s 5.6 5.4 10 5.2 5.0 5 4.8 4.6 0 4.4 4.2 –5 4.0 96-01 98-01 00-01 02-01 04-01 06-01 market, in addition to the information already contained in real and nominal GDP. A number of interesting observations can be made. First, as discussed above, the unconditional forecasts of housing investment and real house price growth are quite different in both VARs. The DVAR projects the housing investment–to-GDP ratio to fluctuate mildly around its steady state, while the growth rate of house prices is projected to return quite quickly to its steady state of 1.5 percent from the relatively high level of growth of more than 5 percent at the end of 1999. The LVAR instead captures some of the persistent in344 J U LY / A U G U S T 2008 96-01 98-01 00-01 02-01 04-01 06-01 sample fluctuations and projects a further rise in housing investment and the growth rate of house prices before it returns close to the sample mean in 2007. Second, based on the DVAR in Figure 2A, neither GDP developments nor short- or long-term interest rates can explain why real house prices continued to grow at rates above 5 percent following the slowdown of the economy in 2000 and 2001. Real and nominal GDP developments can explain an important fraction of the housing boom in 2002 and 2003, but they cannot account for the 10 percent acceleration of house prices in 2004 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets Figure 2B Housing Investment–to-GDP Ratio and Annual House Price Growth Rate, 1995-2007: Actual Data and Unconditional and Conditional Forecasts from the LVAR 5.6 Housing Investment Conditional on y,p 5.4 10 5.2 5.0 5 4.8 4.6 0 4.4 4.2 –5 4.0 5.6 House Prices Conditional on y,p Conditional Forecast Mean 16 Percentile 84 Percentile Unconditional Forecast Mean Actual 96-01 98-01 00-01 02-01 04-01 06-01 96-01 98-01 00-01 02-01 04-01 06-01 Housing Investment Conditional on y,p,i House Prices Conditional on y,p,i 5.4 10 5.2 5.0 5 4.8 4.6 0 4.4 4.2 –5 4.0 96-01 98-01 00-01 02-01 04-01 06-01 96-01 98-01 00-01 02-01 04-01 06-01 Housing Investment Conditional on y,p,i,s House Prices Conditional on y,p,i,s 5.6 5.4 10 5.2 5.0 5 4.8 4.6 0 4.4 4.2 –5 4.0 96-01 98-01 00-01 02-01 04-01 06-01 and 2005. The low level of short- and long-term interest rates in 2004 and 2005 helps explain the boom in those years. In particular, toward the end of 2004 and in 2005, the unusually low level of long-term interest rates helps account for the acceleration in house prices. According to this model, there is some evidence of a conundrum: In this period, long-term interest rates are lower than would be expected on the basis of observed shortterm interest rates. The ability to better forecast the boom period comes, however, at the expense of a larger unexplained undershooting of house prices and housing investment toward the end of F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W 96-01 98-01 00-01 02-01 04-01 06-01 the sample. Overall, these results suggest that the unusually low level of short- and long-term interest rates may have contributed to the boom in U.S. housing markets in the new millennium. Third, the LVAR results in Figure 2B are, however, less clear. The part of the housing boom that cannot be explained by developments in real and nominal GDP is smaller. Moreover, adding short- and long-term interest rates to the data set does not change the picture very significantly. These findings suggest that the results of this analysis partly depend on the assumed steady-state behavior of the housing market and interest rates. J U LY / A U G U S T 2008 345 Jarociński and Smets IDENTIFYING HOUSING DEMAND, MONETARY POLICY, AND TERM SPREAD SHOCKS To add a bit more structure to the analysis, in this section we identify housing demand, monetary policy, and term spread shocks and analyze their effect on the economy. We use a mixture of a recursive identification scheme and sign restrictions. As usual, monetary policy shocks are identified by zero restrictions. They are assumed to affect economic activity and prices with a onequarter lag, but they may have an immediate effect on the term spread and the money stock. The housing demand shock is a shock that affects housing investment and house prices contemporaneously and in the same direction. Moreover, its immediate effect on output is roughly equal to the increase in housing investment (i.e., this shock has no contemporaneous effect on the other components of output taken together). We use sign restrictions to impose this identification scheme.4 For simplicity, we also assume that the housing demand shock affects the GDP deflator only with a lag. The shock that affects housing investment and house prices in opposite directions can be interpreted as a housing supply shock. However, it turns out that this shock explains only a small fraction of developments in the housing market, so we will not explicitly discuss this shock. Figure 3 shows for the DVAR (shaded areas) and the LVAR (dotted lines) the 68 percent posterior probability regions of the estimated impulses. A number of observations are worth making. Overall, both VAR specifications give similar estimated impulse response functions. One difference worth noting is that, relative to the LVAR specification, the DVAR incorporates larger and more persistent effects on house prices and the GDP deflator. In what follows, we focus on the more precisely estimated LVAR specification. According to Figure 3, a one-standard-deviation housing demand shock leads to a persistent rise in real house prices of about 0.75 percent and an increase in the housing investment share of about 0.05 per4 For a discussion of VAR identification with sign restrictions, see, for example, Uhlig (2005). 346 J U LY / A U G U S T 2008 centage points. The effect on the overall economy is for real GDP to rise by about 0.10 percent after four quarters, whereas the effect on the GDP deflator takes longer (about three years) to peak at 0.08 percent above baseline. Note that, in the DVAR specification, the peak effect on goods prices is quite a bit larger. The monetary policy response as captured by the federal funds rate is initially limited, but eventually the federal funds rate increases by about 20 basis points after two years. The initial effect on the term spread is positive, reflecting that long-term interest rates rise in anticipation of inflation and a rise in shortterm rates. To assess how reasonable these quantitative effects are, it is useful to compare them with other empirical results. One relevant literature is the empirical literature on the size of wealth/collateral effects of housing on consumption. As discussed in Muellbauer (2007) and Mishkin (2007), the empirical results are somewhat diverse, but some of the more robust findings suggest that the wealth effects from housing are approximately twice as large as those from stock prices. For example, Carroll, Otsuka, and Slacalek (2006) estimate that the long-run marginal propensity to consume out of a dollar increase in housing is 9 cents, compared with 4 cents for non-housing wealth. Similarly, using cross-country time series, Slacalek (2006) finds that it is 7 cents out of a dollar. Overall, the long-run marginal propensities to consume out of housing wealth range from 5 to 17 percent, but a reasonable median estimate is probably around 7 to 8 percent compared with a 5 percent elasticity for stock market wealth. How does this compare with the elasticities embedded in our estimated impulse response to a housing price shock? A 1 percent persistent increase in real house prices leads to a 0.075 percent increase in real consumption after four quarters. Taking into account that the housing wealth–to-consumption ratio is around 3 in the United States, this suggests a marginal propensity to consume about one-third of the long-run median estimate reported above. This lower effect on consumption may partly be explained by the fact that the increase in house prices is temporary. The mean elasticities embedded in the DVAR are somewhat lower. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets Figure 3 Impulse Responses to a Housing Demand Shock, DVAR and LVAR Output Consumption 0.30 0.30 0.20 0.20 0.10 0.10 0.00 0.00 –0.10 –0.10 –0.20 –0.20 –0.30 0.30 0.20 0.10 0.00 –0.10 5 10 15 20 –0.30 0 5 Housing Investment 0.10 10 15 20 House Prices 2.00 0.05 1.00 0.00 0.00 –1.00 –0.05 –2.00 –0.10 0 5 10 15 20 0 5 Interest Rate 0.50 10 15 Spread 0.30 0.20 0.20 0.10 0.10 0.00 0.00 –0.10 –0.10 –0.20 –0.20 0 5 10 15 20 10 15 20 Commodity Prices 3.00 2.00 1.00 0.00 –1.00 –2.00 –3.00 –4.00 –5.00 –6.00 0 5 10 15 20 Money 0.50 0.00 –0.50 –1.00 0 5 We can also compare our estimated impulse responses with simulations in Mishkin (2007) that use the Federal Reserve Bank U.S. (FRB/US) model. Mishkin (2007, Figure 5) reports that a 20 percent decline in real house prices under the estimated Taylor rule leads to a 1.5 percent deviation of real GDP from baseline in a version of the FRB/US with magnified channels, and to only a bit more than 0.5 percent in the benchmark version (which excludes an effect on real housing investment). Translating our results to a 20 percent real house price shock suggests a multiplier of 2.5 percent. This multiplier is quite a bit higher than that suggested by the FRB/US simulations, F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W 5 1.00 –0.30 –0.30 0 20 0.30 0.40 DVAR 68 Percent LVAR Mean LVAR 68 Percent –0.20 –0.30 0 Prices 0.40 10 15 20 0 5 10 15 20 but in our case this may be partly the result of the strong immediate response of housing investment. Finally, we can also compare the estimated impulse responses of Figure 3 with the impulse responses to a positive housing preference shock in the estimated structural DSGE model of the U.S. economy in Iacoviello and Neri (2007). They find that a 1 percent persistent increase in real house prices is associated with a 0.07 percent increase in consumption and a 3.6 percent increase in real housing investment. Whereas our estimated elasticity of real consumption is very similar, the elasticity of real housing investment J U LY / A U G U S T 2008 347 Jarociński and Smets Figure 4 Impulse Responses to a Monetary Policy Shock, DVAR and LVAR Output 0.30 Consumption DVAR 68 Percent LVAR Mean LVAR 68 Percent 0.20 0.10 0.30 –0.10 –0.10 –0.20 –0.20 –0.30 –0.30 5 10 15 0.20 0.10 0.00 0 0.30 0.20 0.00 20 0.10 0.00 –0.10 –0.20 –0.30 0 5 Housing Investment 0.10 10 15 House Prices 1.00 0.00 0.00 –1.00 –0.05 –2.00 –0.10 0 5 10 15 20 0 5 Interest Rate 0.50 15 Spread 0.30 0.20 0.20 0.10 0.10 0.00 0.00 –0.10 –0.10 –0.20 –0.20 0 5 10 15 20 2008 10 15 20 5 10 15 20 Money 0.50 0.00 –0.50 –1.00 0 5 is quite a bit lower at approximately 1.5 percent. It falls at the lower bound of the findings of Topel and Rosen (1988), who estimate that, for every 1 percent increase in house prices lasting for two years, new construction increases on impact between 1.5 and 3.15 percent, depending on the specifications. Turning to a monetary policy shock, the LVAR results in Figure 4 show that a persistent 25-basispoint tightening of the federal funds rate has the usual delayed negative effects on real GDP and the GDP deflator. The size of the real GDP response is quite small, with a maximum mean negative effect of about 0.1 percent deviation from baseJ U LY / A U G U S T 0 1.00 –0.30 –0.30 5 Commodity Prices 3.00 2.00 1.00 0.00 –1.00 –2.00 –3.00 –4.00 –5.00 –6.00 20 0.30 0.40 348 10 0 20 2.00 0.05 Prices 0.40 10 15 20 0 5 10 15 20 line after three years. This effect is even smaller and less significant in the DVAR specification. For the LVAR specification, the effect on housing investment is larger and quicker, with a maximum negative effect of 0.03 percentage points of GDP (which would correspond to approximately 0.75 percent change) after about two years. Real house prices also immediately start falling and bottom out at 0.5 percent below baseline after two and a half years. The housing market effects are somewhat stronger in the DVAR specification. The higher sensitivity of housing investment to a monetary policy shock is consistent with the findings in the literature. For example, F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets Figure 5 Impulse Responses to a Term Spread Shock, DVAR and LVAR Output Consumption 0.30 0.30 0.20 0.20 0.10 0.10 0.00 0.00 –0.10 –0.10 –0.20 –0.20 –0.30 –0.30 0 5 10 15 20 DVAR 68 Percent LVAR Mean LVAR 68 Percent –0.30 0 5 10 15 20 House Prices –1.00 –2.00 –0.10 0 5 10 15 20 0 5 Interest Rate 0.50 10 15 20 Spread 0.30 0.20 0.20 0.10 0.10 0.00 0.00 –0.10 –0.10 –0.20 –0.20 0 5 10 15 20 15 20 Commodity Prices 0 5 10 15 20 Money 0.00 –0.50 –1.00 0 5 using identified VARs, Erceg and Levin (2002) find that housing investment is about 10 times as responsive as consumption to a monetary policy shock. Our results are also comparable with those reported in Mishkin (2007) using the FRB/US model. In those simulations, a 100-basis-point increase in the federal funds rate leads to a fall in real GDP of about 0.3 to 0.4 percent, although the lags (6 to 8 quarters) are somewhat smaller than those in our estimated BVARs. Further, the effect on real housing investment is faster (within a year) and larger, but the estimated magnitude of these effects (between 1 and 1.25 percent) is quite a bit larger in our case (around 2.5 percent). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W 10 0.50 –0.30 –0.30 5 1.00 0.30 0.40 0 3.00 2.00 1.00 0.00 –1.00 –2.00 –3.00 –4.00 –5.00 –6.00 0.00 –0.05 0.10 –0.20 1.00 0.00 0.20 0.00 2.00 0.05 0.30 –0.10 Housing Investment 0.10 Prices 0.40 10 15 20 0 5 10 15 20 Dynan, Elmendorf, and Sichel (2005) argue that the interest rate sensitivity of real housing investment has fallen since the second half of the 1980s (partly the result of deregulation of the mortgage market in the early 1980s). Our results suggest elasticities that are more in line with Erceg and Levin (2002) than with the FRB/US simulations. Our results can also be compared with the impulse responses to an adverse interest rate shock in Iacoviello and Neri (2007). They find that a 50-basis-point temporary increase in the federal funds rate leads to a fall in real house prices of about 0.75 percent from baseline, compared with a delayed 1 percent fall in real house J U LY / A U G U S T 2008 349 Jarociński and Smets Table 2A Shares of Housing Demand, Monetary Policy, and Term Spread Shocks in Variance Decompositions, DVAR Horizon Variable Shock Output Consumption Prices Housing investment House prices Commodity prices Interest rate Spread Money 0 3 11 23 Housing 0.016 0.034 0.052 0.062 Monetary policy 0.000 0.004 0.021 0.039 Term premium 0.000 0.003 0.015 0.028 Housing 0.005 0.018 0.033 0.055 Monetary policy 0.000 0.003 0.015 0.029 Term premium 0.000 0.005 0.034 0.063 Housing 0.002 0.013 0.120 0.166 Monetary policy 0.000 0.003 0.014 0.037 Term premium 0.000 0.006 0.034 0.046 Housing 0.521 0.579 0.382 0.291 Monetary policy 0.000 0.015 0.175 0.136 Term premium 0.000 0.005 0.023 0.062 Housing 0.535 0.554 0.410 0.242 Monetary policy 0.000 0.010 0.068 0.083 Term premium 0.000 0.002 0.021 0.060 Housing 0.027 0.028 0.041 0.085 Monetary Policy 0.000 0.012 0.167 0.222 Term premium 0.000 0.004 0.018 0.055 Housing 0.037 0.061 0.165 0.178 Monetary policy 0.752 0.496 0.192 0.166 Term premium 0.000 0.023 0.076 0.088 Housing 0.090 0.050 0.177 0.186 Monetary policy 0.223 0.303 0.214 0.206 Term premium 0.336 0.245 0.146 0.134 Housing 0.060 0.044 0.062 0.099 Monetary policy 0.204 0.141 0.044 0.045 Term premium 0.013 0.042 0.129 0.135 NOTE: The reported shares are averages over the posterior distribution and relate to the (log) level variables. 350 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets Table 2B Shares of Housing Demand, Monetary Policy, and Term Spread Shocks in Variance Decompositions, LVAR Horizon Variable Output Consumption Prices Housing investment House prices Commodity prices Interest rate Spread Money Shock 0 3 11 23 Housing 0.019 0.049 0.073 0.106 Monetary policy 0.000 0.005 0.036 0.052 Term premium 0.000 0.005 0.026 0.026 Housing 0.005 0.021 0.051 0.093 Monetary policy 0.000 0.008 0.040 0.051 Term premium 0.000 0.005 0.021 0.024 Housing 0.002 0.017 0.127 0.153 Monetary policy 0.000 0.005 0.038 0.114 Term premium 0.000 0.005 0.012 0.016 Housing 0.582 0.554 0.357 0.351 Monetary policy 0.000 0.027 0.124 0.125 Term premium 0.000 0.015 0.021 0.019 Housing 0.586 0.610 0.360 0.229 Monetary policy 0.000 0.011 0.087 0.066 Term premium 0.000 0.003 0.010 0.014 Housing 0.030 0.044 0.154 0.149 Monetary policy 0.000 0.008 0.072 0.100 Term premium 0.000 0.005 0.012 0.015 Housing 0.032 0.055 0.217 0.211 Monetary policy 0.709 0.453 0.206 0.177 Term premium 0.000 0.007 0.018 0.018 Housing 0.072 0.048 0.129 0.150 Monetary policy 0.230 0.281 0.163 0.152 Term premium 0.355 0.215 0.114 0.085 Housing 0.040 0.036 0.053 0.066 Monetary policy 0.257 0.237 0.089 0.060 Term premium 0.015 0.020 0.021 0.025 NOTE: The reported shares are averages over the posterior distribution and relate to the (log) level variables. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 351 Jarociński and Smets Figure 6A, Panel 1 Counterfactuals, Shutting Down Each of the Identified Shocks, DVAR Housing Investment Feeding All Shocks Except Housing Demand 5.6 Counterfactual Unconditional Forecast Actual 5.4 5.2 5.0 House Prices Feeding All Shocks Except Housing Demand 10 5 4.8 0 4.6 4.4 –5 4.2 5.6 96-01 98-01 00-01 02-01 04-01 06-01 96-01 98-01 00-01 02-01 04-01 06-01 Housing Investment Feeding All Shocks Except Monetary Policy House Prices Feeding All Shocks Except Monetary Policy 5.4 10 5.2 5.0 5 4.8 0 4.6 4.4 –5 4.2 5.6 96-01 98-01 00-01 02-01 04-01 06-01 96-01 98-01 00-01 02-01 04-01 06-01 Housing Investment Feeding All Shocks Except Term Spread House Prices Feeding All Shocks Except Term Spread 5.4 10 5.2 5.0 5 4.8 0 4.6 4.4 4.2 –5 96-01 98-01 00-01 02-01 04-01 06-01 prices in our case (the delay is partly the result of our recursive identification assumption). According to the estimates of Iacoviello and Neri (2007), real investment responds six times more strongly than real consumption and two times more strongly than real fixed investment. Overall, this is consistent with our results. However, the effects in Iacoviello and Neri (2007) are immediate, whereas they are delayed in our case. (See also Del Negro and Otrok, 2007.) In conclusion, the overall quantitative estimates of the effects of a monetary policy shock 352 J U LY / A U G U S T 2008 96-01 98-01 00-01 02-01 04-01 06-01 are in line with those found in the empirical literature. Similarly to our results, Goodhart and Hofmann (2007) find that a one-standard-deviation shock to the real short-term interest rate has about the same quantitative effect on the output gap as a one-standard-deviation shock to the real house price gap. Finally, in the light of the discussion of the effects of developments in long-term interest rates on the house price boom and bust in the United States and many other countries, it is also interesting to look at the effects of a term spread shock F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets Figure 6A, Panel 2 Counterfactuals, Shutting Down Each of the Identified Shocks, DVAR 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 –0.5 GDP Feeding All Shocks Except Housing Demand Prices Feeding All Shocks Except Housing Demand 3.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 7 6 3.0 5 2.5 4 2.0 3 2 1.5 1.0 GDP Feeding All Shocks Except Monetary Policy 0 97-01 00-01 03-01 06-01 3.5 Counterfactual Unconditional Forecast Actual 1 97-01 00-01 03-01 06-01 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 Interest Rate Feeding All Shocks Except Housing Demand Prices Feeding All Shocks Except Monetary Policy 97-01 00-01 03-01 06-01 Interest Rate Feeding All Shocks Except Monetary Policy 7 6 3.0 5 2.5 4 2.0 3 2 1.5 1 0 1.0 97-01 00-01 03-01 06-01 97-01 00-01 03-01 06-01 GDP Feeding All Shocks Except Term Spread Prices Feeding All Shocks Except Term Spread 3.5 97-01 00-01 03-01 06-01 Interest Rate Feeding All Shocks Except Term Spread 7 6 3.0 5 2.5 4 2.0 3 2 1.5 97-01 00-01 03-01 06-01 1.0 1 0 97-01 00-01 03-01 06-01 on the housing market. Figure 5 shows that a 20basis-point increase in long-term interest rates over the federal funds rate has a quite significant effect on housing investment, which drops by more than 0.014 percentage points of GDP (which corresponds to a 0.3 percent change) after about a year. Also, real GDP falls with a bit more of a delay, by about 0.075 percent after six quarters. Both the GDP deflator and real house prices fall, but only gradually. Overall, the size of the impulse responses is, however, small. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W 97-01 00-01 03-01 06-01 Tables 2A and 2B report the contribution of the three shocks to the forecast-error variance at different horizons in both specifications. Overall, the housing demand, monetary policy, and term spread shocks account for only a small fraction of the total variance in real GDP and in the GDP deflator. Monetary policy and housing demand shocks do, however, account for a significant fraction of the variance in the housing market. This can be verified by looking at the contribution of the three shocks to the historical boom and bust episode since 2000, as depicted in J U LY / A U G U S T 2008 353 Jarociński and Smets Figure 6B, Panel 1 Counterfactuals, Shutting Down Each of the Identified Shocks, LVAR 5.6 Housing Investment Feeding All Shocks Except Housing Demand 5.4 Counterfactual Unconditional Forecast Actual 5.2 5.0 House Prices Feeding All Shocks Except Housing Demand 10 5 4.8 0 4.6 4.4 –5 4.2 5.6 96-01 98-01 00-01 02-01 04-01 06-01 96-01 98-01 00-01 02-01 04-01 06-01 Housing Investment Feeding All Shocks Except Monetary Policy House Prices Feeding All Shocks Except Monetary Policy 5.4 10 5.2 5.0 5 4.8 0 4.6 4.4 –5 4.2 5.6 96-01 98-01 00-01 02-01 04-01 06-01 96-01 98-01 00-01 02-01 04-01 06-01 Housing Investment Feeding All Shocks Except Term Spread House Prices Feeding All Shocks Except Term Spread 5.4 10 5.2 5.0 5 4.8 0 4.6 4.4 4.2 –5 96-01 98-01 00-01 02-01 04-01 06-01 Figure 6A for the DVAR and 6B for the LVAR. Panel 1 of each figure shows the developments of the real housing investment–to-GDP ratio (first column) and the annual change in real house prices (second column). Panel 2 of each figure shows output (first column), prices (second column), and interest rates (third column). Each graph includes the actual data (black lines), unconditional forecasts as of 2000 (black dotted lines), and the counterfactual evolution (blue dashed lines) when each of the following three identified shocks are put to zero: a housing demand shock 354 J U LY / A U G U S T 2008 96-01 98-01 00-01 02-01 04-01 06-01 (first row), monetary policy shock (second row), and term spread shock (third row). For the DVAR (Figure 6A), the term spread shock does not have a visible effect on the housing market or the economy as a whole. The housing demand shock has a large positive effect on the housing market in 2001 and 2002 and again in 2004 and 2005. A negative demand shock also explains a large fraction of the fall in construction and house price growth from 2006 onward. These shocks have only negligible effects on overall GDP growth, but do seem to have pushed up inflation F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets Figure 6B, Panel 2 Counterfactuals, Shutting Down Each of the Identified Shocks, LVAR 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 GDP Feeding All Shocks Except Housing Demand Prices Feeding All Shocks Except Housing Demand 3.5 Interest Rate Feeding All Shocks Except Housing Demand 7 6 3.0 5 2.5 4 2.0 3 2 1.5 1.0 0 97-01 00-01 03-01 06-01 GDP Feeding All Shocks Except Monetary Policy 97-01 00-01 03-01 06-01 3.5 Counterfactual Unconditional Forecast Actual 1 Prices Feeding All Shocks Except Monetary Policy 97-01 00-01 03-01 06-01 Interest Rate Feeding All Shocks Except Monetary Policy 7 6 3.0 5 2.5 4 2.0 3 2 1.5 1 0 1.0 97-01 00-01 03-01 06-01 97-01 00-01 03-01 06-01 GDP Feeding All Shocks Except Term Spread Prices Feeding All Shocks Except Term Spread 3.5 97-01 00-01 03-01 06-01 Interest Rate Feeding All Shocks Except Term Spread 7 6 3.0 5 2.5 4 2.0 3 2 1.5 97-01 00-01 03-01 06-01 1.0 1 0 97-01 00-01 03-01 06-01 by 10 to 20 basis points over most of the post-2000 period. Loose monetary policy also seems to have contributed to the housing boom in 2004 and 2005. Without the relatively easy policy of late 2003 and early 2004, the boom in house price growth would have stayed well below the 10 percent growth rate in 2005. Easy monetary policy also has a noticeable, though small effect, on GDP growth and inflation. The LVAR results depicted in Figure 6B give similar indications, although they generally attribute an even larger role to the housing demand shocks. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W 97-01 00-01 03-01 06-01 HOUSE PRICES AND THE MONETARY POLICY STANCE IN THE UNITED STATES The idea of measuring monetary conditions by taking an appropriate weight of interest rates and asset prices was pioneered by the Bank of Canada and the Reserve Bank of New Zealand in the 1990s. Because both countries are small open economies, these central banks worried about how changes in the value of the exchange rate J U LY / A U G U S T 2008 355 Jarociński and Smets may affect the monetary policy stance.5 The idea was to construct a weighted index of the shortterm interest rate and the exchange rate, where the weights reflected the relative effect of the exchange rate on an intermediate or final target variable, such as the output gap, output growth, or inflation. A number of authors have extended the idea of the MCI to other asset prices, arguing that those asset prices may be equally or more important than the exchange rate. One prominent example is Goodhart and Hofmann (2007), who argue that real house prices should receive a significant weight in an MCI because of their significant effect on the economy. For the United States, they argue that the relative weight of the short-term interest rate versus house prices should be of the order of 0.6 to 1.8. In the small literature that developed following the introduction of the MCI concept, a number of shortcomings have been highlighted.6 One difficulty is that the lag structure of the effects of changes in the interest rate and real house prices on the economy may be different. As noted above, according to our estimates, the effect of an interest rate shock on economic activity appears to take somewhat longer than the effect of a house price shock. In response, Batini and Turnbull (2002;BT) proposed a dynamic MCI that takes into account the different lag structures by weighting all current and past interest rates and asset prices with their estimated impulse responses. Another shortcoming of the standard MCI is that it is very difficult to interpret the MCI as an indicator of the monetary policy stance, because it does not take into account that changes in monetary conditions will typically be endogenous to the state of the economy. The implicit assumption of the standard MCI is that the monetary conditions are driven by exogenous shocks. This is clearly at odds with the identified VAR literature that suggests that most of the movements in monetary conditions are in response to the state of the economy. For example, changes in the federal funds rate will 5 See, for example, Freedman (1994 and 1995a,b) and Duguay (1994). 6 See, for example, Gerlach and Smets (2000). 356 J U LY / A U G U S T 2008 be typically in response to changing economic conditions and a changing outlook for price stability. An alternative way of expressing this drawback is that the implicit benchmark against which the MCI is measured does not depend on the likely source of the shocks in the economy. As a result, the benchmark in the standard MCI does not depend on the state of the economy, although clearly for given objectives the optimal MCI will vary with the shocks to the economy. A third shortcoming is that often the construction of an MCI does not take into account that the estimated weight of its various components is subject to uncertainty and estimation error. This uncertainty needs to be taken into account when interpreting the significance of apparent changes in monetary conditions. The methodology developed by Céspedes et al. (CLMM; 2006) addresses each of these shortcomings. In this section, we apply a version of the MCI proposed by CLMM to derive a measure of the monetary policy stance that takes into account movements in the short- and long-term interest rates and in real house prices. Using this index, we try to answer this question: Did the rise in house prices and the fall in long-term interest rates since 2000 lead to an implicit easing of monetary policy in the United States? We use the BVARs estimated in the previous section to implement the methodology. In the next subsection, we define the MCI and use a simple analytical example to illustrate its logic. Next, we apply it to the U.S. economy using the estimated BVARs. An MCI in a VAR: Methodology and Intuition For the sake of example, let the economy be described by a stationary VAR of order one: (3) X t A11 P = A t 21 A12 X t −1 B1 + ε, A22 Pt −1 B2 t where Xt is the vector of nonpolicy variables, such as output and inflation, and Pt is the vector of monetary policy and financial variables, which in our case are the short-term interest rate, the long-term interest rate spread, and the real house F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets price index. As in BT, a standard dynamic MCI with respect to a target variable j can then be defined as ( ) j s −1 * MCI BT ,t = S j ∑ A11 A12 Pt −s − Pt −s , H (4) s =1 where Sj is a selection vector that selects the target variable j from the list of non-policy variables. Typically, the target variable in the construction of an MCI is either output growth or the output gap. This is based on the notion that financial and monetary conditions affect inflation primarily through their effect on spending and output. However, inflation can be used as a target variable also. In this paper, we will present results for both output growth and inflation as target variables. The parameter H is the time period over which lags of the monetary conditions are considered. Pt*–s is typically given by the steady state of the monetary conditions. In our case, this would be the equilibrium nominal interest rate, the steady-state term spread, and steady-state real house price growth rate. Alternatively, it could also be given by the monetary conditions that would have been expected as of period t –H, if there had been no shocks from period t –H to t. Equation (2) illustrates that the standard MCI is a weighted average of the deviations of current and past policy variables from their steady-state values, where the weights are determined by the partial elasticity of output with respect to a change in the policy variable. As discussed above, a problem with this notion of the MCI is that the policy variables are treated as exogenous and independent from the underlying economic conditions, or, alternatively, they are assumed to be driven by exogenous shocks. As a result, it is very problematic to interpret this index as a measure of the monetary policy stance. For example, it may easily be the case that the policy rate rises above its steadystate value because of positive demand shocks. In this case, monetary policy may either be too loose, neutral, or too tight, depending on whether the higher interest rate is able to offset the effect of the demand shocks only partially, fully, or more than fully. Instead, the standard MCI will F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W always indicate that monetary conditions have tightened. In contrast to the standard MCI, the alternative MCI proposed by CLMM does take into account the endogeneity of the policy instruments. In this case the MCI is defined as (5) ( j * s −1 MCI CLMM ,t = S j ∑ A11 A12 Pt −s − Pt −s ( H s =1 ) ) * +S j ∑ A1s1−1B1 E εt −s P − E εt −s P . s =1 H The first part is the same as in the standard case (equation (4)), but the second part adds the effect of shocks that are most consistent with the observed path of monetary conditions. More specifically, the shocks are drawn from their distribution, subject to the restriction that they generate the observed path of monetary conditions. Doan et al. (1984) and Waggoner and Zha (1999) show that the mean of this constrained distribution is given by (6) ε stacked = R ′ ( RR ′ ) −1 ( P − E [ P ])stacked, P where ε stacked is a vector of stacked shocks over period H, R is a stacked matrix of impulse response coefficients of the monetary conditions with respect to the shocks, and P – E[P] is the vector of correspondingly stacked forecast errors associated with the observed or assumed monetary conditions over the same period H. To understand the intuition for why the MCI by CLMM is a potentially much better indicator of the stance of monetary policy, it is useful to go through a simple static analytical example. Assume the economy is given by the following set of equations: (7) y t = α 1st + α 2ht + εty (8) st = β1εty + β2εth + β3εts (9) ht = δ st + εth , where yt is the target variable, say, output growth, st , is the short-term policy rate, ht is real house prices, and there are three shocks: an output shock, J U LY / A U G U S T 2008 357 Jarociński and Smets a policy shock, and a housing shock. Equation (7) reflects the dependence of output on the monetary conditions and an output shock. For convenience, we have in this case assumed that there are no lags in the transmission process. Equation (8) is a monetary policy reaction function, and equation (9) shows how house prices depend on the short rate and a shock. In this case, the standard MCI (as in BT) is given by (10) MCI BT ,t = α 1st + α 2 ht and is independent of the monetary policy reaction function. If α1 is negative and α2 is positive, a rise in house prices will lead to an easing of monetary conditions unless the short-term interest rate rises to exactly offset the effect of house prices on the target variable. In contrast, the MCI of CLMM is given by (11) MCI CLMM ,t = α 1st + α 2 ht + E εty st, ht , where we have assumed that all variables are measured as deviations from the steady state. As in equation (6), the mean output shock needs to be consistent with the observed short-term interest rate and real house prices. Next, we derive that the expression of the last term in equation (11) is a function of the interest rate and house prices. From equations (6) and (7), it is clear that the relation between the interest rate conditions and the shocks is given by (12) β2 st β1 h = δβ 1 + δβ 2 t 1 εty β3 h ε = R εt . δβ3 t s εt As discussed above, given a joint standard normal distribution of the shocks, the mean of the shocks conditional on the observed interest rates is given by (13) st E εt st ,ht = R ′( RR ′ ) −1 , ht where R is given in equation (12). To simplify even further, assume that β3 = 0, that is, there is no policy shock. In this case, there 358 J U LY / A U G U S T 2008 is a one-to-one relationship between the shocks and the observed interest rate and house prices, given by (14) εty (1 + δβ2 ) β 1 h= δ − εt − β2 β1 st . 1 ht As a result, the MCI of CLMM is given by (15) MCI CLMM ,t = (α 1 ) ( ) + (1 + δβ2 ) β 1 st + α 2 − β2 β 1 ht . Comparing expressions (15) and (10), it is obvious that the MCIs of BT and CLMM have different weights on the short-term interest rate and house prices. The weights in the MCI of CLMM depend not only on the partial elasticities of output with respect to the short-term interest rate and house prices, but also on the coefficients in the policy reaction function and the elasticity of house prices with respect to the short-term interest rate. To see why the MCI of CLMM is a better indicator of the monetary policy stance, it is useful to investigate how the weights in (15) will depend on systematic policy behavior. From equations (7) and (9), one can easily show that, if the central bank targets output growth, the optimal interest rate reaction function is given by (16) st = − α2 1 εty − εth . α 1 + δα 2 α 1 + δα 2 If the interest rate elasticity of output is negative (α1 < 0) and elasticity with respect to house prices is positive (α2 < 0), then a central bank trying to stabilize output will lean against positive output and house price shocks, where the size of the reaction coefficient will depend on the strength and the channels of the transmission mechanism. Substituting the coefficients β1 and β2 in (15) with the coefficients in expression (16), it can be verified that the MCI of CLMM will be equal to zero. In other words, a policy that stabilizes output will be seen as a neutral policy according to this index. In contrast, it is obvious that such a F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets change in the policy reaction function will not affect the standard MCI. Instead, assume that the central bank reacts optimally to the output shock, as in equation (13), but does not respond to the shock to house prices (β2 = 0). In this case, it can be shown that the MCI of CLMM is given by (17) MCI CLMM ,t = α 2 ( ht − δ st ) = α 2εth . This result is very intuitive: When the central bank does not respond to house price shocks and a rise in house prices has a stimulative effect on output, the MCI of CLMM will indicate easy monetary conditions whenever there is a positive shock to house prices. This simple example makes it clear that, in order to have a meaningful indicator of the monetary policy stance, it is important to realize that the monetary conditions endogenously reflect all shocks that hit the economy. An Application to House Prices and the Policy Stance in the United States Obviously, the static example is too simple to bring to the data. In reality, monetary conditions will have lagged effects on output and inflation and the lag patterns may differ across the various components, as shown earlier. In this section, we use the two specifications of the BVAR—the LVAR and the DVAR—to calculate MCIs for the U.S. economy. Consistent with the MCI literature, we use real GDP growth and inflation as the target variables. Moreover, to take into account the lags in the transmission process of monetary policy that we documented in the third section, we assume that real GDP growth is expected annual GDP growth one year ahead, whereas inflation is expected annual inflation two years ahead. Figures 7A and 7B show the results of this exercise. To illustrate the effect of taking endogeneity of the indicators of stance into account, we also compare the MCI of CLMM (which incorporates the full set of shocks) with the MCI of BT. In the latter case, we assume that the observed interest rates and house prices are driven by only the three exogenous shocks identified in the third section. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Figure 7A shows for the DVAR and 7B for the LVAR the estimated 68 percent probability regions for the MCI of CLMM (blue dotted lines) and the MCI of BT (gray shaded areas) based on one-year-ahead annual output growth (left column) and two-year-ahead annual inflation (right column) using the following indicators of monetary conditions: the federal funds rate (first row); the federal funds rate and the term spread (second row); and the federal funds rate, the term spread, and real house prices (third row). The MCIs shown are basically the difference between the conditional forecast of the target variable based on the actual path of the chosen indicators of stance and the unconditional forecast of the target variable. A few observations are worth making on the basis of Figure 7A. First, overall, the MCI with expected output growth as a target variable and the MCI with inflation as the target variable give similar indications about stance. Financial conditions were relatively tight in 2000-01, then gradually became relatively loose in 2002-05 before turning tight again during 2006. Second, the uncertainty surrounding the MCIs is very high. Based on standard significance levels, the monetary conditions were not significantly different from neutral during the whole period. Third, taking house prices into account (third row of Figure 7A) does seem to matter for measuring the monetary policy stance. More specifically, buoyant growth in house prices in 2004 and 2005 suggests that monetary policy was relatively loose in this period, whereas it turned tight in 2007. During the housing boom, easy monetary conditions implied two-year-ahead annual inflation that was more than 0.5 percentage points above its steady state. Most recently, tight conditions imply expected inflation almost 0.5 percentage points below the target. These results differ marginally when the LVAR specification is used (compare Figure 7B with 7A). In Figure 7B, a comparison of the 68 percent posterior probability regions for the MCI of CLMM (blue dotted lines) with those for the MCI of BT (shaded areas) reveals that, although the broad messages of the estimated MCIs are similar, conditioning on the three identified exogenous shocks J U LY / A U G U S T 2008 359 Jarociński and Smets Figure 7A MCIs of CLMM and BT, DVAR GDP Conditional on i Prices Conditional on i 3 3 2 2 1 1 0 0 –1 –1 –2 –2 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 GDP Conditional on i,s Prices Conditional on i,s 3 3 2 2 1 1 0 0 –1 –1 –2 –2 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 Prices Conditional on i,s,hp–p GDP Conditional on i,s,hp–p 3 3 2 2 1 1 0 0 –1 –1 –2 –2 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 only (the MCI of BT) gives less-precise estimates. This is partly because these exogenous shocks contribute only to a limited degree to the forecast variance of output and inflation. As a result, the effects are also less precisely estimated. The point estimates are similar, which suggests that the developments in 2002-05 were strongly influenced by the policy and housing demand shocks and not much by the responses to other shocks. As explained earlier, the MCIs are a weighted average of current and past levels of the short360 BT 68 Percent CLMM Mean CLMM 68 Percent J U LY / A U G U S T 2008 term interest rate, the term spread (or the longterm interest rate), and real house price growth. To show the relative importance of the three components, Table 3 gives the sum of the weights on current and past (up-to-8-quarter) lagged values of each. As in Figure 7A, using annual GDP and inflation as target variable, the MCIs of CLMM and BT are, respectively, calculated based on the short-term interest rate (the first panel); the shortand long-term interest rates (the second panel); and the short- and long-term interest rates, and real house price growth (the third panel). A few F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets Figure 7B MCIs of CLMM and BT, LVAR GDP Conditional on i Prices Conditional on i 3 3 2 2 1 1 0 0 –1 –1 –2 –2 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 GDP Conditional on i,s BT 68 Percent CLMM Mean CLMM 68 Percent Prices Conditional on i,s 3 3 2 2 1 1 0 0 –1 –1 –2 –2 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 GDP Conditional on i,s,hp–p Prices Conditional on i,s,hp–p 3 3 2 2 1 1 0 0 –1 –1 –2 –2 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 –3 00-01 01-01 02-01 03-01 04-01 05-01 06-01 07-01 Table 3 8-Quarter Sum of MCI Weights, DVAR MCIi MCIi,s MCIi,s,hp–p Short rate (i) Short rate Long rate (s) Short rate Long rate CLMM-GDP –0.162 –0.201 0.090 –0.198 0.102 0.000 BT-GDP –0.198 –0.190 0.074 –0.194 0.102 0.003 CLMM-Inflation –0.046 –0.142 0.162 –0.148 0.250 0.056 BT-Inflation –0.154 –0.182 0.168 –0.087 0.180 0.083 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W House prices (hp–p) J U LY / A U G U S T 2008 361 Jarociński and Smets observations are noteworthy. First, taking only the short-term interest rate as an indicator of the policy stance, it is clear that on average an observed increase in the interest rate above its steady-state value indicates a restrictive policy stance with respect to both GDP growth and inflation. This is, in particular, the case when the short-term interest rate is assumed to be driven by the three identified exogenous shocks (as in the MCI of BT). However, if the full endogenous nature of the nominal interest rate is taken into account (as in the MCI of CLMM), this is less the case and more so for inflation than for growth. The reason is that, because of the central bank’s reaction function, the short-term interest rate is likely to increase in response to shocks that drive up future GDP growth and inflation. In this case, a rise in interest rates may even suggest an easing of the policy stance if interest rates do not rise enough to offset the pickup in growth and inflation. To the extent that changes in the nominal interest rate reflect higher inflation and inflation expectations, this argument is particularly strong when expected inflation is the target. In the second panel of Table 3, adding the long-term interest rate slightly changes the picture. Keeping the long-term interest rate constant, observing a 1-percentage-point increase in the short-term interest rate for 8 quarters signals a fall in GDP growth of about 20 basis points over the next year and a fall in inflation of somewhat less over the next two years. In contrast, keeping the short-term rate constant, a rise in the longterm interest rate by 1 percentage point signals lax monetary policy, as it predicts a rise in both GDP growth (up to 9 basis points) and inflation (up to 16 basis points) above steady state. Finally, the far-right panel of Table 3 shows the weights when real house prices are included in the MCIs also. Their addition has little effect on the weights on interest rates. The upper rows show that the weight on real house price growth is close to zero when the target variable is GDP growth. This is indeed similar to the results in Figure 7A, which show that the actual MCIs do not change very much. However, when annual inflation over two years is the target variable, there is a significant weight on house prices: A 362 J U LY / A U G U S T 2008 5-percentage-point rise in the growth rate of real house prices signals a 30- to 40-basis-point rise in annual inflation According to the weights, such a rise in house prices would call for a substantially higher short-term rate (of about 2 percentage points) in order to have neutral monetary conditions. CONCLUSIONS In this paper, we examine the role of housing investment and house prices in U.S. business cycles since the second half of the 1980s using an identified Bayesian VAR. We find that housing demand shocks have significant effects on housing investment and house prices, but overall these shocks have had only a limited effect on the performance of the U.S. economy in terms of aggregate growth and inflation in line with the empirical literature. There is also evidence that monetary policy has significant effects on housing investment and house prices and that easy monetary policy designed to stave off perceived risks of deflation in 2002-04 has contributed to the boom in the housing market in 2004 and 2005. However, again, the effect on the overall economy was limited. A counterfactual simulation suggests that without those policy shocks inflation would have been about 25 basis points lower at the end of 2006. In order to examine the effect of house prices on monetary conditions, we implement a methodology proposed by Céspedes et al. (2006). This methodology consists of calculating the forecast of a target variable (expected GDP growth or expected inflation) conditional on the observed path of monetary conditions, including the shortterm interest rates, the term spread, and house prices. We show that, in spite of the endogeneity of house prices to both the state of the economy and the level of interest rates, taking house prices into account may sharpen the inference about the stance of monetary policy. Given the uncertainty about the sources of business cycle fluctuations and the effect of the various shocks (including housing demand shocks) on the economy, uncertainty regarding the stance of monetary policy F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets remains high. Nevertheless, taking the development of house prices into account, there is some indication that monetary conditions may have been too loose in 2004 and were relatively tight in the summer of 2007. Various caveats regarding the methodology we use in this paper are worth mentioning. First, all the analysis presented in this paper is insample and ex post. Although this is helpful in trying to understand past developments, this does not prove the methodology is sufficient for realtime analysis. For this we need to extend the analysis to a real-time context. Second, the statistical model we use to interpret the U.S. housing market and business cycle is basically a linear one. It has been argued that costly asset price booms and busts are fundamentally of an asymmetric nature. Our linear methodology is not able to handle such nonlinearities. Third, the robustness of the analysis to different identification schemes for the structural shocks needs to be further examined. We hope to shed light on some of these issues in further analysis. Boom Across U.S. States.” Journal of Monetary Economics, October 2007, 54(7), pp. 1962-85. Doan, Thomas; Litterman, Robert B. and Sims, Christopher. “Forecasting and Conditional Projection Using Realistic Prior Distributions.” Econometric Reviews, 1984, 3(1), pp. 1-100. Duguay, Pierre. “Empirical Evidence on the Strength of the Monetary Transmission Mechanism in Canada: An Aggregate Approach.” Journal of Monetary Economics, February 1994, 33(1), pp. 39-61. Dynan, Karen E.; Elmendorf, Douglas and Sichel, Daniel E. “Can Financial Innovation Help to Explain the Reduced Volatility of Economic Activity?” Journal of Monetary Economics, January 2006, 53(1), pp. 123-50. Erceg, Christopher J. and Levin, Andrew T. “Optimal Monetary Policy with Durable and Non-durable Goods.” International Finance Discussion Paper No. 748, Board of Governors of the Federal Reserve System, 2002. Batini, Nicolleta and Turnbull, Kenny. “A Dynamic Monetary Conditions Index for the UK.” Journal of Policy Modeling, June 2002, 24(3), pp. 257-81. Freedman, Charles. “The Use of Indicators and the Monetary Conditions Index in Canada,” in T.J.T. Baliño and C. Cottarelli, eds., Frameworks for Monetary Stability: Policy Issues and Country Experiences. Washington, DC: International Monetary Fund, 1994, pp. 458-76. Carroll, Christopher D.; Otsuka, Misuzu and Slacalek, Jirka. “How Large Is the Housing Wealth Effect? A New Approach.” NBER Working Paper 12746, National Bureau of Economic Research, 2006. Freedman, Charles. “The Canadian Experience with Targets for Reducing and Controlling Inflation,” in L. Leiderman and L. Svensson, eds., Inflation Targets. London: Centre for Educational Policy Research, 1995a. Céspedes, Brisne; Lima, Elcyon; Maka, Alexis and Mendonça, Mario J.C. “Conditional Forecasts and the Measurement of Monetary Policy Stance in Brazil.” Unpublished manuscript, 2006. Freedman, Charles. “The Role of Monetary Conditions and the Monetary Conditions Index in the Conduct of Policy.” Bank of Canada Review, Autumn 1995b, pp. 53-59. Clarida, Richard; Galí, Jordi and Gertler, Mark. “The Science of Monetary Policy: A New Keynesian Perspective.” Journal of Economic Literature, December 1999, 37(4), pp. 1661-707. Gerlach, Stefan and Smets, Frank. “MCIs and Monetary Policy.” European Economic Review, October 2000, 44(9), pp. 1677-1700. REFERENCES Del Negro, Marco and Otrok, Christopher. “99 Luftballons: Monetary Policy and the House Price F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Goodhart, Charles and Hofmann, Boris. “Financial Conditions Indices,” in Charles Goodhart, ed., House Prices and the Macroeconomy: Implications J U LY / A U G U S T 2008 363 Jarociński and Smets for Banking and Price Stability. Chap. 3: Oxford: Oxford University Press, 2007. Iacoviello, Matteo and Neri, Stefano. “The Role of Housing Collateral in an Estimated Two-Sector Model of the U.S. Economy.” Working Papers in Economics 659, Boston College Department of Economics, 2007. Kohn, Donald L. “Success and Failure of Monetary Policy since the 1950s.” Presented at the Deutsche Bundesbank conference Monetary Policy over Fifty Years, a conference to mark the fiftieth anniversary of the Deutsche Bundesbank, Frankfurt, Germany, September 21, 2007. Leamer, Edward E. “Housing and the Business Cycle.” Presented at the Federal Reserve Bank of Kansas City symposium Housing, Housing Finance, and Monetary Policy, Jackson Hole, WY, August 30September 1, 2007. McConnell, Margaret M. and Pérez-Quirós, Gabriel. “Output Fluctuations in the United States: What Has Changed Since the Early 1980s?” American Economic Review, December 2000, 90(5), pp. 1464-76. Mishkin, Frederic S. “Housing and the Monetary Transmission Mechanism.” Presented at the Federal Reserve Bank of Kansas City symposium Housing, Housing Finance, and Monetary Policy, Jackson Hole, WY, August 30-September 1, 2007. Muellbauer, John. “Housing and Consumer Behaviour.” Presented at the Federal Reserve Bank of Kansas City symposium Housing, Housing Finance, and Monetary Policy, Jackson Hole, WY, August 30-September 1, 2007. Slacalek, Jirka. “What Drives Personal Consumption? The Role of Housing and Financial Wealth.” Unpublished manuscript, November 2006. Taylor, John B. “Housing and Monetary Policy.” Panel discussion at the Federal Reserve Bank of Kansas City symposium Housing, Housing Finance, and Monetary Policy, Jackson Hole, WY, August 30September 1, 2007. Topel, Robert. H. and Rosen, Sherwin. “Housing Investment in the United States.” Journal of Political Economy, August 1988, 96(4), pp. 718-40. Uhlig, Harald. “What Are the Effects of Monetary Policy on Output? Results from an Agnostic Identification Procedure.” Journal of Monetary Economics, March 2005, 52(2), pp. 381-419. Villani, Mattias. “Steady-State Priors for Vector Autogressions.” Journal of Applied Econometrics, 2008 (forthcoming). Waggoner, Daniel F. and Zha, Tao. (1999). “Conditional Forecasts in Dynamic Multivariate Models.” Review of Economics and Statistics, November 1999, 81(4), pp. 639-51. Mojon, Benoit. “Monetary Policy, Output Composition, and the Great Moderation.” Working Paper WP 2007-07, Federal Reserve Bank of Chicago, 2007. 364 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Jarociński and Smets APPENDIX: DATA AND SOURCES Real GDP: real GDP, 3 decimal (GDPC96), seasonally adjusted annual rate, quarterly, billions of chained 2000 dollars. SOURCE: U.S. Department of Commerce: Bureau of Economic Analysis (BEA) data from Federal Reserve Economic Data (FRED; http://research.stlouisfed.org/fred2/). Real consumption: real personal consumption expenditures (PCECC96), seasonally adjusted annual rate, quarterly, billions of chained 2000 dollars. SOURCE: BEA data from FRED. GDP deflator: GDP: implicit price deflator (GDPDEF), seasonally adjusted, quarterly, index 2000 = 100. SOURCE: BEA data from FRED. Federal funds rate: effective federal funds rate (FEDFUNDS), monthly, percent, averages of daily figures. SOURCE: Board of Governors of the Federal Reserve System data from FRED (averaged over 3 months of the quarter). Long-term interest rate: 10-year Treasury constant maturity rate (GS10), monthly, percent, averages of business days. SOURCE: Board of Governors of the Federal Reserve System data from FRED (averaged over 3 months of the quarter). S&P/Case-Shiller U.S. National Home Price Index: quarterly, based on repeated sales. SOURCE: http://www.standardandpoors.com, available since 1987. M2: M2 money stock (M2NS), not seasonally adjusted, monthly, billions of dollars. SOURCE: Board of Governors of the Federal Reserve System data from FRED (averaged over 3 months of the quarter). Real private residential fixed investment: 3 Decimal, (PRFIC96), seasonally adjusted annual rate, quarterly, billions of chained 2000 dollars. SOURCE: BEA data from FRED. Commodity price index: Dow Jones spot average, quarterly. SOURCE: Global Financial Data; www.globalfinancialdata.com. In the VAR, we use the interest rate spread, computed as the difference between the long interest rate and the federal funds rate, house prices deflated relative to the GDP deflator, and the ratio of real private residential fixed investment to real GDP. All the variables, except for the short-term interest rate, spread, and housing investment, enter either in log levels or log differences (annualized), depending on the VAR specification indicated. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 365 366 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Robert G. King W hen monetary historians look back at this decade, they will undoubtedly highlight the major increases in house prices over the early part of the decade and the sharp declines of recent years as posing a major challenge for monetary policy and central banking. “House Prices and the Stance of Monetary Policy” by Marek Jarociński and Frank Smets (JS) is a valuable early contribution to the understanding of this episode. It is extremely clear in spelling out and accomplishing its two major objectives: a retrospective econometric analysis of the role of housing markets in recent developments and a consideration of the potential role of house prices in a monetary conditions index. For my purposes in this discussion, there are three important pieces of evidence provided by JS. Using conditional forecasting methods, the second section of their paper shows that there may be an important component of house price variation that cannot be accounted for by shifts in output and interest rates or there may not: The qualification is necessary because the results of difference and level VAR specifications differ importantly. Their third section uses an identified VAR to suggest that loose monetary policy may have contributed to the continuing increase in house prices in 2004 and 2005. Their fourth section investigates the effect of identified “housing demand shocks,” with results that I will discuss further below. MONETARY POLICY From the standpoint of monetary policy, there are three key questions. First, was the behavior of house prices and quantities normal or unusual over the recent period? Second, did easy money cause a major portion of the rise in house prices and thus make house price declines a necessary outcome when monetary policy tightened? Third, could a regular response to housing—perhaps via the type of monetary conditions index discussed by JS—be desirable in smoothing out overall economic activity and housing markets themselves? House Prices It is important to stress that the second section of the JS study, about the extent to which movements in house prices are unusual, can be read in quite different ways. JS show that movements in interest rates and output largely explain variation in house prices if one uses a level (Bayesian) VAR. In this case, there are two implications for monetary policy. First, it seems unnecessary to think about potentially including house prices in the state vector to which monetary policy should respond, since house prices appear to be well explained by interest rates and output. Second, there is no sense in which there is a puzzle in recent years: House prices just moved with macroeconomic conditions in a fairly standard manner. From the standpoint of modern macroeconomic analysis Robert G. King is a professor of economics at Boston University. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 367-370. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 367 King and modern central bank practice based on simple rules, this is an attractive reading of the data. However, JS also show that it is possible to argue that developments in interest rates and output leave a great deal to be explained if one uses a first difference VAR. In this sense, there may be an unusual event in recent years, with house prices departing from output and interest rate fundamentals just during this period. Or house prices may be not too closely related to these fundamentals most of the time, so that may be a case for thinking about a separate monetary policy response to housing. That is, we need to know whether other periods of house price increases and decreases look similar to or different from those of recent years. Monetary Policy as a Cause of House Prices There has been a great deal of public discussion about “easy money” in the house price boom, and there is some evidence for this view in JS. By shutting down the identified monetary policy shock in panels 1 of their Figures 6A (differences VAR) and 6B (level VAR), they find that house prices would have been lower without monetary policy shocks during 2004 and early 2005. I have three observations on this finding. First, one would like to know the statistical confidence with which we can make this statement (my own sense based on work with VARs is that this might be low). Second, taking the result at face value, it is important to stress that monetary policy accounts for only a temporary interval of higher house price increases and little of the ultimate decline in house prices. Third, the JS accounting method does not automatically mean that a shock yields a contribution during this period, as may be seen by comparing this to the contribution that JS suggest for a term structure spread: There is nothing contributed to house prices by the yield curve. So, the method is potentially informative in this and other contexts. I think that we do not know the role that monetary policy played in these events, but there is good reason to be skeptical of the manner in which “easy money” is used in many public discussions. 368 J U LY / A U G U S T 2008 In the public eye (that of my neighbors and my real estate agent in a Boston real estate market that was a hot one starting in about 2000), there were two distinct parts to the house price boom. The first was based on income and wealth: As my real estate agent said in 2000, people were buying houses in the face of rapidly increasing house prices with “real money” from successful economic ventures. The second was later: People were buying houses or refinancing houses, taking advantage of the increasingly favorable terms offered by lenders. Using my agent’s terminology at the later time, this was “easy money.” But lender terms were sufficiently generous that it is hard to draw a connection to the Fed: The public definition of easy money is a statement about lending terms, not necessarily about monetary policy. Monetary Policy Response to Housing An unfortunate aspect of the JS paper is that the dynamic response to an identified housing demand shock—that object to which a monetary policy authority would potentially want to respond—just doesn’t look plausible to me. The key features of this shock, as described at the start of their third section, are that it raises housing prices; it raises private consumption and national product; and it has a positive effect on house investment with a timing that is curious. From the standpoint of designing a monetary policy response to the housing sector, this puzzling pattern of responses makes it problematic to address my third question (above), which is the critical one from the standpoint of monetary policy. THINKING ABOUT DYNAMIC RESPONSES IN HOUSING The analysis of the housing demand shock requires that we begin to think more carefully about the nature of housing dynamics. While macroeconomists use the “time to build” model of Kydland and Prescott (1980) much less now than some time ago, housing is surely a setting in which this model is the benchmark. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W King To sketch how such a model works and the potential conflict that I see with the impulse responses for the identified demand shock of JS, let’s think about a setting in which there is an unexpected increase in housing demand at a fixed stock of housing. We would see an increase in house prices, which in turn would stimulate housing starts and an interval of higher housing expenditure. If the housing starts were undertaken “on spec” by construction companies, then one would expect increased starts only if the future house prices were expected to be high enough to justify construction costs. JS cite the empirical estimates of Topel and Rosen (1988) and the simulations of a recent quantitative macro model developed by Iacoviello and Neri (2007) as guidance in terms of the effects of house prices on residential investment. The estimates of Topel and Rosen (1988), in particular, suggest an elasticity in the range of 1.5 to 3.15 for the response of investment two years later to a permanent change in house prices. And JS argue that their model captures this level of overall response, thus supporting the identification of the housing demand shock. However, in terms of deciding whether this measure of a housing demand shock is plausible, I think that we need more detailed dynamic information. Suppose that it takes three quarters of a year to complete a housing construction project and that the distribution of expenditure is uniform over the construction project. Then, housing investment (i) is an equally weighted moving average of starts (s), 1 it = α st + st −1 + st −2 , 3 where α is a parameter describing the size of investment projects. More generally, the time-tobuild model may suggest that the time-path of investment depends on the distribution of investment costs over the life of the construction process and the interaction of optimal “housing starts” with the anticipated path of house prices. Suppose further that starts increase permanently at date t = τ. Then, investment builds up to a new higher level, with one-third of the increase taking place in each period. Now, the factors F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W generating starts are not permanent, but if there is a sustained increase in house prices, then this calculation should capture the early part of the impulse response. From the standpoint of this type of model, then, the dynamics in Figure 3 seem curious. That is, a housing demand shock raises prices at a point in time by about 1 percent, by the same amount by a year later, by perhaps 1/2 percent after two years, and by nothing after three years. The investment dynamics are a response of about 0.05 for the first two quarters, then perhaps half that by year’s end, and zero by six quarters. A conventional view of the construction process is that at least a year is a reasonable horizon overall, with the first quarter devoted to planning and permits and the last three quarters involving the bulk of the expenditure. There is no question that construction is faster now than it was a couple of decades ago. But before accepting the identification of the housing demand shock, one would like to see that dynamics are consistent with estimates of the distribution of quarterly construction costs. Housing Permits, Starts, and Investment Housing permits have long been used as a leading indicator (included in the Conference Board’s series of leading economic indicators), as have housing starts. Both of these series have been historically treated as noisy ones, but also containing useful information about future economic activity. Figure 1 of this commentary shows why, starting in 1987 as do JS. The reader’s eye is drawn naturally to the most recent part of the period, where housing permits and starts (monthly data) move prior to investment (quarterly data). If there is a persistent decline in housing starts, caused by a negative housing demand shock, then there will be a persistent decline in investment in any time-to-build model, but it will take time for the full effect to build up. From this standpoint, the near-term forecasts for housing investment are not too rosy. The identification of housing demand shocks would benefit from using indicators of permits and starts. Such empirical work, expanding on J U LY / A U G U S T 2008 369 King Figure 1 Housing Starts and Permits and Residential Fixed Investment Billions of Chained 2000 $ Thousands of Units 2,500 710 2,000 580 1,500 450 1,000 320 HOUST (left axis) PERMIT (left axis) PRFIC1 (right axis) 500 1987 1992 1997 2002 2007 190 2012 NOTE: HOUST is housing starts: total: new privately owned housing units started; PERMIT is new private housing units authorized by building permit; and PRFIC1 is real private residential fixed investment, 1 decimal. Shaded areas indicate U.S. recessions as determined by the National Bureau of Economic Research. SOURCE: Federal Reserve Bank of St. Louis: research.stlouisfed.org. the study of JS, could lead to dynamic responses for investment flows in response to identified housing demand shocks that are more in line with the structural characteristics of housing market investment. In turn, this would provide a more secure basis for analysis of the monetary policy response to housing. CONCLUSION The events of the last few years will certainly stimulate much additional research on the nature of housing and mortgage markets, as well as their implications for monetary policy. The analysis of Jarociński and Smets highlights a series of important questions about these linkages, as well as providing some interesting early empirical evidence. 370 J U LY / A U G U S T 2008 REFERENCES Iacoviello, Matteo and Neri, Stefano. “The Role of Housing Collateral in an Estimated Two-Sector Model of the U.S. Economy.” Working Papers in Economics No. 659, Boston College Department of Economics, 2007. Jarociński, Marek and Smets, Frank. “House Prices and the Stance of Monetary Policy.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 339-65. Kydland, Finn and Prescott, Edward C. “Time To Build and Aggregate Fluctuations.” Econometrica, November 1982, 50(6), pp. 1345-70. Topel, Robert. H. and Rosen, Sherwin. “Housing Investment in the United States.” Journal of Political Economy, August 1988, 96(4), pp. 718-40. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Stephen G. Cecchetti I s housing the business cycle, as Leamer says? In their fascinating paper, Jarociński and Smets’s (2008) careful analysis suggests that the answer is no. Very briefly, they show that the recent U.S. housing boom is explained by a combination of increases in housing demand and loose monetary policy. However, once they adequately account for the myriad of dynamic interactions, they find that housing demand shocks have a very limited impact on the overall volatility of real growth and inflation. And, finally, Jarociński and Smets use their estimates to suggest that, since 2000, monetary conditions have been close to neutral. The paper is divided neatly into two parts. The first presents the results of a careful model estimation exercise—a Bayesian vector autoregression (BVAR) that includes real gross domestic product (GDP), the GDP deflator, real consumption, real residential investment, real house prices, real commodity prices, the money stock, the federal funds rate, and the long-term interest rate spread. The second part of the paper uses estimates from the first to estimate a monetary conditions index (MCI). Following this organization, I divide my comments into two parts. First, I discuss the role of housing in the business cycle; and second, I will make a number of comments about the use of MCIs. PART I: UNDERSTANDING THE ROLE OF HOUSING In the first part of the paper, Jarociński and Smets present a careful analysis of the dynamic properties of housing, monetary policy, and growth. They focus on the impact of shocks to housing demand, monetary policy, and the term spread, concluding that they account for a small fraction of real GDP and the real GDP deflator but a large fraction of the variation in house prices and residential construction. (I am referring to the variance decomposition results in their Table 2A. For reasons that will become clear later, I prefer the differences version of their VAR.) Importantly, the Jarociński and Smets estimates show that a combination of a positive housing demand shock and low interest rates accounts for the bulk of the rise in house prices and the increase in residential construction activity. (See the historical decompositions in their Figure 6A.) I have three separate points to make about this conclusion. First, the results are neatly consistent with my strongly held view that over the period that Jarociński and Smets study, 19872006, monetary policymakers stopped being the destabilizing force that they probably were in the 1970s and may even have been successfully neutralizing a variety of demand shocks.1 That 1 See Cecchetti, Flores-Lagunes, and Krause (2006). Stephen G. Cecchetti is a professor of global finance at Brandeis International Business School and a research associate at the National Bureau of Economic Research. The author thanks Marek Jarociński and Frank Smets for both correcting errors in the initial version of these comments and for providing their data. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 371-76. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 371 Cecchetti Figure 1 Rolling Regression of a 10-Year Bond on Federal Funds, 48-Month Window Coefficient 2.5 2.0 1.5 1.0 +/–2 Standard Error 0.5 0 –0.5 –1.0 –1.5 –2.0 1991 1993 1995 1997 1999 2001 2003 2005 2007 NOTE: The figure plots coefficients of a regression of the fixed-maturity 10-year U.S. Treasury bond rate on the federal funds rate using a 48-month rolling window. Estimates are plotted on the last date of the sample. SOURCE: Board of Governors of the Federal Reserve System. is, the authors use their VAR to allocate the volatility of growth and inflation to its various sources and find no role for monetary policy disturbances. I take this as evidence of the success of central bank stabilization policy. Second, there is the always-vexing question of whether the sample record used in estimating the model is representative of the experience during the more recent period for which we would like to use the model. A number of concerns arise here. First, there is the problem of trying to separate changes in the federal funds rate from changes in the term spread. To see the possible problem, I have run a very simple regression of the 10-year bond rate on the federal funds rate, using a 48month moving window, and plotted the results in Figure 1. I simply note that the late-1990s look very different from the period either before or after and suspect that the identification that allows Jarociński and Smets to estimate the 372 J U LY / A U G U S T 2008 impact of the spread is coming from this part of the sample. Continuing with the issue of the sample period, there is the question of how we should interpret house price data since 2000. Figure 2 plots the ratio of the value of the U.S. housing stock (from the Federal Reserve Flow of Funds data) to the housing rental service flow (imputed for the computation of the National Income and Product Accounts). The results are striking. The post-2000 data look dramatically different from what came before. Finally, like others before them, Jarociński and Smets find significant housing wealth effects. Their estimate is that a persistent 1-percentagepoint increase in house prices leads to a 0.1 percent increase in real GDP after four quarters—an elasticity of 0.1. Interestingly, because of the richness of their model, Jarociński and Smets are able to estimate that this effect is split roughly F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Cecchetti Figure 2 The Ratio of the Value of the U.S. Housing Stock to the Rental Service Flow 20 2006 = 18.4 18 16 14 1978-1999 Average = 14.3 12 10 1952 1955 1958 1961 1964 1967 1970 1973 1976 1979 1982 1985 1988 1991 1994 1997 2000 2003 2006 SOURCE: Value of residential real estate: Federal Reserve Flow of Funds data, line 4 of Table B100 plus line 4 of Table B103. Rental service flow: the National Income and Product Accounts estimate of the total housing services in personal consumption expenditure, Table 2.3.5, line 14. equally between investment in residential construction and consumption. So, the elasticity of consumption with respect to housing wealth is only about 0.05, which is at the low end of the range found by previous researchers (and cited in the paper). To digress only slightly, I should note that it is not obvious that changes in the value of housing should affect nonhousing consumption at all. We all have to live somewhere. When home prices rise, it does not signal any increase in the quantity of economy-wide output. Although someone with a bigger house could sell it and move into a smaller one, there must be someone else on the other side of the trade. That is, for each person trading down and taking wealth out of their house, someone is trading up and putting wealth in. And renters planning to purchase should save more. All of this should cancel out so that in the aggregate there is no change! F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Put another way, people own their homes to hedge the risk arising from potential changes in the price of purchasing housing services. They want to ensure that they can continue to live in the same size home. A rise in property prices means people are consuming more housing, not that they are wealthier. And yet, everyone finds that when the housing market booms, people raise their consumption. Is this increase justified? Well, it depends. If the consumption and house price increases are both a consequence of higher estimated long-run growth, then the answer is yes. That is, if everyone now expects higher future incomes, then they will demand more housing along with more of everything else, and there is no bubble. So, if the house price boom is accompanied by an increase in the rate of growth of potential output, then it is not a bubble. An equity price boom would have to accompany this as well. And, importantly, this J U LY / A U G U S T 2008 373 Cecchetti would likely imply an increase in the long-run real interest rate, too. So, if housing, equity, and bonds all boom at the same time, we probably need not be concerned. Regardless of my fairly minor concerns, I am convinced by Jarociński and Smets’s conclusion: Stabilizing real growth requires at least some focus on residential construction and housing demand. Housing may not be the business cycle, but it does play a measurable role. But, as Jarociński and Smets show, this depends primarily on long-term interest rates and housing demand, both of which seem to have a life of their own. Monetary policymakers are left wondering what tools they have at their disposal to do anything about this. PART II: MCIs The second part of the Jarociński and Smets paper presents a very clear discussion of MCIs. They conclude that, since 2000, Federal Reserve policy has been roughly neutral. Before working through this paper, I had not understood what MCIs are. Now I do, so I will make some attempt to share this new-found insight. As Jarociński and Smets describe, in the past, several (but not many) central banks used MCIs as guides to policy formulation. More recently, business economists have been churning these out, combining a variety of financial indicators into something that is supposed to measure conditions in financial markets (the Goldman-Sachs Financial Conditions Index, DeutscheBank Financial Conditions Index, Morgan Stanley Financial Conditions Index, etc.). The idea behind what I will call the “traditional MCI” is that it should provide a measure of the relative ease or tightness of monetary conditions. For policymakers, this MCI is supposed to answer the following question: Given the current state of the economy, how should policymakers set their operational instrument? The traditional MCI employed by the Bank of Canada, for example, was of the following type: (1) 374 MCE = α ( r − r * ) + β (e − e * ) , J U LY / A U G U S T 2008 where r is the interest rate instrument, e is the exchange rate, and the “*” signifies an equilibrium level. In practice, the problem is that (1) implies the same reaction to any deviation of the exchange rate from its equilibrium, regardless of the source. This creates problems, because supply shocks should (one assumes) require different responses from demands shocks. It matters why the exchange rate has moved. As Jarociński and Smets describe in clear detail, this led researchers to suggest the computation of a “conditional MCI”—that is, conditional on some sort of information. A conditional MCI is the forecast k periods ahead for the output gap (actual output, y, less potential output, y*) or the inflation gap (the deviation of inflation, π, from its target, π*): (2) E ( y t + k − y *t + k ) I t and (3) E π t + k I t . Importantly, these expectations are conditional on the policymaker’s implied monetary policy reaction function. But, the information set used to compute the expectations need not have everything in it. Looking at (2) and (3) leads me to ask the following question: If policymakers are doing their job, why would the conditional MCI ever deviate from zero? Because the conditional MCI should be zero, what might we get from computing it? As it turns out, quite a bit. To see, we can start with a generic formulation of the policymaker’s problem. Assume that monetary policy sets the interest rate, r, to minimize the quadratic loss function, (4) L = E απ t2 + y t2 , subject to the constraints imposed by the dynamic structure of the economy: (5) yt εt π = A ( L ) r , t t F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Cecchetti where A(L) is a polynomial in the lag operator, L, and ε is a vector of disturbances. This problem yields a policy “rule” of the form (6) rt* = φ (L ) εt . Substituting (6) into (5) yields a reduced form: (7) yt % π = A ( L ) εt . t The conditional MCI is related to the properties of (7). Jarociński and Smets note that when α = 0 and A(L) in (5) has no lags, then E ( yt + k − y *t + k ) I t = 0 for all k. They interpret this as neutral policy. Although this is fine as far as it goes, the conditional MCI is actually capable of addressing two additional questions: (i) Does the central bank need to change its reaction function to meet its stated goal? Is the reaction function (6) appropriate to minimize the loss function, (4)? (ii) What is the tradeoff or relative weight, a, in the central bank’s loss function, (4)? Looking at question (i), we see that this is not a question of whether policy is loose, tight, or neutral. The issue is whether it is properly responding to the shocks that are hitting the economy. Are policymakers moving their instrument to neutralize demand shocks completely? Are they changing the short-term interest rate to offset supply shocks appropriately? It is not about action, it is about reaction. To understand (ii), take a look at the following static version of (5) written as an aggregate demand–aggregate supply model: (8) y = – λ r + εd (aggregate demand) (9) π = ω y + ε s (aggregate supply). The parameters λ and ω represent the slopes of the aggregate demand and aggregate supply curves, respectively. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W This setup implies a simple policy rule of the form (10) r* = aεd + bε s . Using this, we can now compute the implied conditional MCI for output and inflation (conditional on the optimal policy response, that is): (11) and (12) ( ) (1 +αωαω ) ε E y r* = − 2 s 1 ε . ( ) (1 + αω ) E π r* = 2 s ( ) = −αω. E (π r ) Now, take the ratio of (11) to (12) to obtain E y r* (13) * So, once we know the slope of the aggregate supply curve, the ratio of these two conditional MCIs tells us the relative importance of inflation variability in the policymaker’s objective function—their inflation volatility aversion, if you will. To figure out a reasonable value for ω, take a look at the impulse responses in their Figure 4. The first row tells us that an interest rate shock (which is basically an aggregate demand shock) has roughly the same impact on inflation and output. This leads to the conclusion that ω ≈ 1. Next, take a look at the first row of their Figure 7A— the MCI conditional on monetary policy, but not on other financial conditions. (Because my very simple construction really models the unconditional, steady-state behavior, I have chosen to use the differences VAR estimates.) The implied time series for α is plotted in Figure 3. These point estimates move around quite a bit. But the primary problem is that they are negative. That is, inflation and output seem to be moving in the same direction at the horizons over which Jarociński and Smets report their conditional MCI computations. There are several possible reasons for this. The first is that Jarociński and Smets’s Figure 7A reports the conditional MCI over different horizons for output and inflation. For the former it is J U LY / A U G U S T 2008 375 Cecchetti Figure 3 Implied Inflation Volatility Aversion 50 40 30 20 10 0 –10 –20 –30 Jan 00 Jul 00 Jan 01 Jul 01 Jan 02 Jul 02 Jan 03 Jul 03 Jan 04 Jul 04 Jan 05 Jul 05 Jan 06 Jul 06 Jan 07 SOURCE: Author’s calculations using data from Jarociński and Smets (2008, Figure 7A). one year, whereas for the latter it is two. So, although there might be a contemporaneous volatility tradeoff, it isn’t showing up here. A second possibility is that monetary policymakers were not in fact acting appropriately to neutralize the housing demand shock. This interpretation is consistent with Jarociński and Smets’s results that the boom which began in fall 2001 was the consequence of a combination of an increase in housing demand and expansionary monetary policy. My conclusion is that this means Federal Reserve policy was not on the output-inflation volatility frontier. In conclusion, I found this a very rewarding paper to read. Although I may not subscribe to Jarociński and Smets’s interpretation of the conditional expectation of output or inflation as an indicator of monetary conditions, I do agree with their conclusion that housing is at the core of the business cycle, so it should have a prominent role in the formulation of monetary policy. 376 J U LY / A U G U S T 2008 REFERENCES Cecchetti, Stephen G.; Flores-Lagunes, Alfonso and Krause, Stefan. “Has Monetary Policy Become More Efficient? A Cross-Country Analysis.” Economic Journal, April 2006, 116(4), pp. 408-33. Jarociński, Marek and Smets, Frank R. “House Prices and the Stance of Monetary Policy.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 339-65. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Assessing Monetary Policy Effects Using Daily Federal Funds Futures Contracts James D. Hamilton This paper develops a generalization of the formulas proposed by Kuttner (2001) and others for purposes of measuring the effects of a change in the federal funds target on Treasury yields of different maturities. The generalization avoids the need to condition on the date of the target change and allows for deviations of the effective fed funds rate from the target as well as gradual learning by market participants about the target. The paper shows that parameters estimated solely on the basis of the behavior of the fed funds and fed funds futures can account for the broad calendar regularities in the relation between fed funds futures and Treasury yields of different maturities. Although the methods are new, the conclusion is quite similar to that reported by earlier researchers— changes in the fed funds target seem to be associated with quite large changes in Treasury yields, even for maturities of up to 10 years. (JEL: E52, E43) Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 377-93. E conomists continue to debate how much of an effect monetary policy has on the economy. But one of the more robust empirical results is the observation that changes in the target that the Federal Reserve sets for the overnight federal funds rate have been associated historically with large changes in other interest rates, even for the longest maturities. This paper contributes to the extensive literature that tries to measure the magnitude of this effect. One of the first efforts along these lines was by Cook and Hahn (1989), who looked at how yields on Treasury securities of different maturities changed on the days when the Federal Reserve changed its target for the fed funds rate. Let is,d denote the interest rate (in basis points) on a Treasury bill or Treasury bond of constant maturity s months as quoted on some business day, d, and let ξd denote the target for the fed funds rate as determined by the Federal Reserve for that day. Using just those days between September 1974 and September 1979 on which there was a change in the target, Cook and Hahn estimated the following regression by ordinary least squares (OLS): (1) is ,d − is ,d −1 = α s + λs (ξd − ξd −1 ) + usd . Their estimates of λs for securities of several different maturities are reported in the first column of Table 1. These estimates suggest that, when the Fed raises the overnight rate by 100 basis points, short-term Treasury yields go up by over 50 basis points and there is a statistically significant effect even on 10-year yields. Subsequent researchers found that the magnitudes of the estimated coefficients for λs were significantly smaller when later data sets were used. For example, column 2 of Table 1 reports Kuttner’s (2001) results when the Cook-Hahn regression (1) was reestimated using data from June 1989 to February 2000; see also Nilsen (1998). However, Kuttner (2001) also identified some conceptual problems with regression (1). For one James D. Hamilton is a professor of economics at the University of California, San Diego. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 377 Hamilton Table 1 Alternative Estimates of the Response of Interest Rates to Changes in the Federal Funds Target Study Cook-Hahn Specification Sample Kuttner Kuttner Poole-Rasche (1) (1) (2)-(3) (4)-(3) 1974-79 1989-2000 1989-2000 1988-2000 0.73** s = 3 months 0.55** 0.27** 0.79** s = 6 months 0.54** 0.22** 0.72** s = 1 year 0.50** 0.20** 0.72** s = 5 years 0.21** 0.10* 0.48** s = 10 years 0.13** 0.04* 0.32** — 0.78** — 0.48** NOTE: *indicates statistically significant with p-value < 0.05; **denotes p-value < 0.01. thing, the market may have anticipated much of the change in the target ξd that occurred on day d many days earlier, in which case those expectations would have already been incorporated into is,d –1. In the limiting case when the change was perfectly anticipated, one would not expect any change in is,d to be observed on the day of the target change. To isolate the unanticipated component of the target change, Kuttner used fd , the interest rate implied by the spot-month fed funds contract on day d. These contracts are settled on the basis of what the average effective fed funds rate turns out to be for the entire month containing day d. Because much of the month may already be over by day d, a target change on day d will have only a fractional effect on the monthly average. Kuttner proposed the following formula to identify the unanticipated component of the target change on day d: (2) Nd ξ%du = ( fd − fd −1 ), N d − td + 1 where Nd is the number of calendar days associated with the month in which day d occurs and td is the calendar day of the month associated with day d. Kuttner then replaced (1) with the regression ( ) (3) is,d − is,d −1 = α s + γ s ξd − ξd −1 − ξ%du + λs ξ%du + usd , with additional modifications if d were the first day or one of the last three days of a month. 378 J U LY / A U G U S T 2008 Kuttner found that the values for γs were essentially zero, meaning that if target changes were anticipated in advance, then they had no effect on other interest rates. Kuttner’s estimates of λs , the effects of unanticipated target changes, are reported in column 3 of Table 1 and turn out to be a bit larger than the original Cook-Hahn estimates. Poole and Rasche (2000) proposed to sidestep the issues associated with a mid-month target change by using not the spot-month contract on day d but instead the one-month-ahead contract, that is, the interest rate implied by a contract purchased on day d for settlement based on the average fed funds rate prevailing in the following month, denoted fd1. They then replaced the expression in (2) with (4) ξ%du = fd1 − fd1−1 . Their estimates for λs using this formulation turned out to be similar to Kuttner’s and are reported in column 4 of Table 1. However, mid-month target changes remain an issue for the Poole-Rasche estimates because there is always the possibility of a second (or even a third) change in the target some time after day d and before the end of the following month; indeed, this turned out to be the case for about half of the target changes observed between 1988 and 2006. Gürkaynak, Sack, and Swanson (2007) developed an analog to Kuttner’s formula (2) based on the date of the next target change that followed F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Hamilton after the one implemented on day d; see also Gürkaynak (2005). Another potential drawback to either (2) or (4) was raised by Poole, Rasche, and Thornton (2002). These authors noted that, particularly prior to 1994, market participants may not have been perfectly aware of the target change even at the end of day d, in which case these formulas would include a measurement error that would bias the coefficients downward. Poole, Rasche, and Thornton developed corrections for the estimates to allow for this measurement error. A related issue is that the series for ξd, the actual target change, is itself subject to measurement error, as indeed Kuttner (2001) and Poole, Rasche, and Thornton (2002) used slightly different series. Learning about the target change presumably also began well before day d. For both reasons, one would think that data both before and after day d should typically be used. In this paper I develop a generalization of the Kuttner (2001) and Poole, Rasche, and Thornton (2002) adjustments for purposes of estimating the parameter λs . The basic idea is to suppose that there exists some day within the month at which the target may have been changed, but to choose deliberately not to condition on this day for purposes of forming an econometric estimate. The paper also generalizes the earlier approaches by explicitly modeling the difference between the effective fed funds rate and the actual target. The next section begins with an examination of the relation between the target rate chosen by the Fed and the actual effective fed funds rate. The third section develops a simple statistical description of how these deviations, along with the process of learning by the market about what the fed funds target is going to be for this month, would determine the volatility of the spot-month futures rate. The fourth section shows how the parameters estimated from the behavior of the effective fed funds rate and the spot-month futures rate can be used to predict calendar regularities in the estimated values for a generalization of the coefficient λs . The final section finds such calendar regularities largely borne out in the observed relation between Treasury rates and daily changes in the spot-month futures rate and develops new F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W estimates of this parameter. Although the method and data set are rather different from the earlier researchers, my estimates in fact turn out to be quite similar to those originally found by Kuttner (2001) and Poole and Rasche (2000). THE EFFECTIVE AND TARGET FEDERAL FUNDS RATES In this paper, time is indexed in two different ways, using calendar days t for developing theoretical formulas and business days d to apply these ideas to actual data. The theoretical formulas will be developed for a typical month consisting of N calendar days indexed by t = 1,2,...,N, whereas the data set will consist of those days d = 1,2,...,D for which there are data on both Treasury interest rates and fed funds futures rates; d = 1 corresponds to October 3, 1988, and d = D = 4,552 corresponds to December 29, 2006. The empirical sample for all estimates reported in this paper also excludes the volatile data from September 13 to September 30, 2001. The effective fed funds rate for calendar day t, denoted rt , is a volume-weighted average of all overnight interbank loans of Federal Reserve deposits for that day. All numbers in this paper will be reported in basis points, so that, for example, a 5.25 percent interest rate would correspond to a value of rt = 525. Since October 1988, the Chicago Board of Trade has offered futures contracts whose settlement is based on the average value for the effective fed funds rate over all the calendar days of the month (with Friday rates, for example, also imputed to Saturday and Sunday). For a month that contains N calendar days, settlement of these futures contracts would be based on the value of S = N −1 ∑ rt . N (5) t =1 The terms of a given fed funds futures contract can be translated1 into an interest rate, ft, such that, if S (which is not known at day t but will 1 Specifically, if Pt is the price of the contract agreed to by the buyer and seller on day t, then ft = 100 × 共100 – Pt 兲. J U LY / A U G U S T 2008 379 Hamilton Figure 1 Effective Fed Funds Rate, Target Fed Funds Rate, and Fed Funds Futures Rate, December 1990 could use the change in the spot-month contract price on day n to infer how much of the change in the target interest rate caught the market by surprise according to the formula (7) Fed Funds Rate 900 800 700 Effective Target Futures 600 500 1 8 15 22 29 Calendar Date become known by the end of the month) turns out to be bigger than ft, the buyer of the contract has to compensate the seller by a certain amount for every basis point by which S exceeds ft . If the marginal market participant were risk neutral, it would be the case that (6) ft = E t (S ), where Et共.兲 denotes an expectation formed on the basis of information available to the market as of day t. This paper will consider only spot-month contracts, that is, contracts for which by day t we already know some of the values for r (namely, rτ for τ ⱕ t) that will end up determining S. My forthcoming paper (Hamilton, forthcoming) demonstrates that, for futures contracts at short horizons (the spot-month, 1-month-ahead, and 2-monthahead contracts), expression (6) appears to be an excellent approximation to the data, though Piazzesi and Swanson (forthcoming) note potential problems with assuming that it holds for longer-horizon contracts. Suppose that the Fed changes the target for the effective fed funds rate on calendar day n of this month. Kuttner (2001) suggested that we 380 J U LY / A U G U S T 2008 N (f − f ) . N − n + 1 n n −1 I will provide a formal derivation of (7) as a special case of a more general statistical inference problem explored below, but would first like to comment on one potential drawback of (7), which is that it implies a huge reweighting of observations that come near the end of the month (n near N ). Kuttner (2001, p. 529) recognized that this is a potential concern here because (7) abstracts from the deviation between the Federal Reserve’s target for the effective fed funds rate and the actual effective rate, and as a result magnifies the measurement error for observations near the end of the month. Kuttner himself avoided using (7) for the last three days of the month. Other researchers like Gürkaynak (2005) avoid applying it to data from the last week. Figure 1 plots the relevant variables for December 1990, which was a particularly wild month as banks adjusted to lower reserve requirements (Anderson and Rasche, 1996). Although the Fed had lowered the target to 725 basis points on December 7, the effective fed funds rate was trading well above this the week after Christmas, and speculators seemed to be allowing for a possibility of a big end-of-year spike up, such as the 584-basis-point increase in the effective fed funds rate that was seen in the last two days of 1985 or the 975-basis-point spike between December 28 and December 30, 1986. In the event, however, the effective funds rate plunged 200 basis points on December 31, 1990. Because the December 1990 futures contract was based on the effective rate rather than the target, speculators were watching these events closely. The futures rate was trending well above the new target of 725 basis points in the latter part of December, partly because the month’s average would include the first week’s 750-basis-point target values, partly because the effective rate had been averaging above the new target subsequently, and partly in anticipation of an end-of-year spike F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Hamilton Figure 2 Squared Residuals by Day of Month Average Squared Residual 3,000 2,500 2,000 1,500 1,000 500 0 –500 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 Day of Month NOTE: The figure plots the average squared residuals from a regression of the deviation of the fed funds rate from the target on its own lagged value, by day of the month (in basis points). 95 percent confidence intervals are indicated by the upper and lower box lines, and predicted values from regression (9) are indicated by the dashed line. up. When it became clear on December 31 that the last day of the year generated a big move down rather than up, the December futures contract fell by 23 basis points on a single day. Formula (7) would call for us to multiply this number by 31, to deduce that the interest rate surprise on this day was some 713 basis points, plausible perhaps if the market was anticipating a spike up to 1,250 rather than the plunge down to 550 that actually transpired. Although this is an extreme example, it drives home the lesson that one really wants to downweight the end-of-month observations rather than magnify them in the manner suggested by the expression in (7). The next section proposes a more formal statement of this problem and its solution. A necessary first step is to document some of the properties of the deviation between the target that the Fed F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W has in place for business day d (denoted ξd) and the actual fed funds rate. The effective fed funds rate, rd, was taken from the FRED database of the Federal Reserve Bank of St. Louis (which in turn is based on Board of Governors release H.15), and the target ξd prior to 1994 is from the FRED series that comes from Thornton (2005) and since 1994 is from Federal Open Market Committee (FOMC) transcripts. I first estimated the following regression (similar to the models in Taylor, 2001, and Sarno, Thornton, and Valente, 2005) by OLS (standard errors are in parentheses): (8) rd − ξd = 2.45 + 0.300 ( rd −1 − ξd −1 ) + eˆ d . (0.29 ) ( 0.014) This regression establishes that there is modest serial correlation in deviations from the target. Of particular interest in the next section will be J U LY / A U G U S T 2008 381 Hamilton the calendar variation in the variability of êd. Let ωjd = 1 if day d occurs on the j th calendar day of the month and zero otherwise. A regression of êd2 on {ωjd }31 j =1 then gives the average squared residual as a function of the calendar day of the month: êd2 = ∑ βˆ jω jd + νˆ d . The effective fed funds rate for each day is the sum of the target for that day plus the deviation from the target, denoted ut: rt = ξt + ut . It follows from (5) and (6) that 31 j =1 The estimated values βˆj are plotted as a function of the calendar day j in Figure 2 along with the 95 percent confidence intervals for each coefficient. A big outlier on January 23, 1991, (when the funds rate spiked up nearly 300 basis points on a settlement Wednesday) is enough to skew the results for day 23. Apart from this, the most noticeable feature is an increased volatility of the deviation of the funds rate from the target toward the end of a month. One can represent this tendency parametrically through the following restricted regression: (9) 31−t eˆ d2 = 283 + 1,746 × 0.5( d ) + νˆ d , ( 40) (232) where td is the calendar day of the month associated with business day d. The predicted values from (9) are also plotted in Figure 2. In the next section, a simple theoretical formulation based on (8) and (9) will be used to characterize the modest predictability of deviations from the target and their tendency to become more pronounced at the end of the month. ACCOUNTING FOR THE VOLATILITY OF SPOT-MONTH FUTURES PRICES Suppose that market participants know that, if the Fed is going to change the target within a given month consisting of N calendar days, it would do so on calendar day n, so that its target is a step function: ξt = ξ0 ξt = ξn 382 forÄ Å t = 1,2,...,n − 1 Ä forÅ Ä t = n,n + 1,..., N . J U LY / A U G U S T 2008 (10) N ft = Et N −1 ∑ (ξτ + uτ ) τ =1 N − n + 1 n − 1 ξ + = E t (ξn ) N 0 N + N −1 ∑ uτ + N −1 t τ =1 ∑ N τ =t +1 E t (uτ ). On the day before the target change, I presume that market participants had some expectation of what the target was going to be, denoted En –1共ξn 兲. The actual target would deviate from this by some magnitude hn: ξn = E n −1 (ξn ) + hn . If the equilibrium fed funds price is determined by risk-neutral rational speculators, the forecast error hn would be a martingale difference sequence that represents the content of the news about ξn that arrived on the day of the target change itself. Similarly, ξn = E n −2 (ξn ) + hn + hn −1, where hn –1 is the news that arrived on day n –1 of the Fed’s intentions on day n, and ξn = hn + hn −1 + hn −2 + L + h1 + E 0 (ξn ) . Under rational expectations, {ht} should be a sequence of zero-mean, serially uncorrelated variables, whose unconditional variance is denoted σh2. Notice that h1 represents the information that the market receives on day 1 about the value for the target that the Fed will adopt on day n, h2 represents the new information received on day 2, and so on, with E 0 (ξn ) + h1 + h2 + L + ht (11) E t (ξn ) = Å ξn forÄ t ≤ n forÄ t > n . F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Hamilton Given (8) and (9), I assume that deviations follow an AR(1) process with an innovation variance that increases at the end of the month: ut = φ ut −1 + εt E ( ) εt2 = γ 0 + γ 1δ ( N −t ) , where the empirical results suggest values of φ = 0.30, γ0 = 283, γ1 = 1,746, and δ = 0.5. Then ∑ N (12) τ =t +1 Et (uτ ) = φut + φ ut + L + φ 2 N −t ut = ( φ 1 − φ N −t 1−φ )u . t Substituting (11) and (12) into (10) gives n − 1 ft = ξ N 0 N − n + 1 + E 0 (ξn ) + h1 + h2 + L + ht N + N −1 ∑ uτ + N −1 t (13) τ =1 ( φ 1−φ N −t 1−φ )u t for t ≤ n t N − n + 1 n − 1 ft = ξ0 + ξn + N −1 ∑ uτ N N τ =1 +N −1 ( φ 1 − φ N −t 1−φ )u for t > n. t From (13) we can then calculate the change in the spot-month futures rate for t ⱕ n to be N − n + 1 ft − ft −1 = ht N +N −1 (1 − φ N −t +1 1−φ ) u − N φ (1 − φ −1 t N −t +1 ( 1−φ ( ) ) )u t −1 1 − φ N −t +1 N − n + 1 −1 (14) = (φut −1 + εt ) ht + N N 1−φ − N −1 ( φ 1 − φ N −t +1 1−φ )u t −1 1 − φ N −t +1 N − n + 1 −1 = + h N εt for t ≤ n, t N 1−φ F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W whereas for t > n, changes in futures prices are driven solely by the deviation of the effective fed funds rate from the target: ft − ft −1 = N −1 (1 − φ N −t +1 1−φ )ε forÅ t > n. t It follows that the variance of daily changes in the spot-month futures rate would be given by 2 E ( ft − ft −1 ) targetÄ changeÄ onÄ dayÄ n 2 2 2 N −t +1 2 / ( N − n + 1) / N σ h + σ ε ,t 1 − φ 2 (15) 2 = N (1 − φ ) forÄ Å t ≤ n N t 2 − σ 1 − φ +1 2 / N 2 (1 − φ )2 for t > n ε ,t ( ) ( ) σ ε2,t = γ 0 + γ 1δ ( N −t ) . Prior to 1992, the day of a target change would often (but not always) occur the day after an FOMC meeting. Since 1994, it usually has occurred on the day of an FOMC meeting, but there are exceptions: Three times in 2001 (January 3, April 18, and September 17) the Fed changed the target without a meeting, and in August and September of 2007 there was active speculation that the Fed was considering or possibly had even already implemented an intermeeting rate cut. Rather than treat day n as if always known to the econometrician, I have followed a different philosophy, which is to ask, How would the data look if they were generated by (15) but the econometrician does not condition on knowledge of the particular value of n? Suppose that the day of the target change (which the formula assumed was known to market participants as of the start of the month) could have occurred with equal probability on any one of the calendar days n = 1,2,...,N. If we let η denote the unknown day of the target change, then the unconditional data would exhibit a calendar regularity in the variance that is described by J U LY / A U G U S T 2008 383 Hamilton Figure 3 Squared Spot-Month Change by Day of Month Average Squared Change 20 15 10 5 0 –5 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 Day of Month NOTE: The figure plots the average squared change in the spot-month futures rate, by day of the month (in basis points). 95 percent confidence intervals are indicated by upper and lower box lines, and predicted values from regression (19) are indicated by the dashed line. 2 2 E (ft − ft −1 ) = N −1 ∑ E ( ft − ft −1 ) η = n n =1 N (16) (1 − φ = ) N −t +1 2 N 2 (1 − φ ) (1 − φ = ) 2 σ ε2,t + N −1 ∑ N −t +1 2 N 2 (1 − φ ) 2 N n =t γ 0 + γ 1δ ( N −t ) + = κ 1 (t ) + κ 2 (t ) σ 2h , where κ1 (1 − φ (t ) = 384 ) N −t +1 2 N (1 − φ ) N2 h ∑ τ2 2 σh N3 N −t +1 τ =1 γ 0 + γ 1δ ( N −t ) ( N − t + 1) ( N − t + 2) (2N − 2t + 3) . 2 κ 2 (t ) = ( N − n + 1)2 σ 2 J U LY / A U G U S T 2 6N 3 2008 Expression (16) describes the variance of changes in the spot-month rate as the sum of two terms. The first term (κ1共t兲) represents solely the contribution of deviations of the effective funds rate from the target. For days near the beginning of the month (N – t large), this is essentially equal to γ0/共1 – φ兲2 (the unconditional variance of ut) divided by N 2 (because each ut contributes with weight 1/N to the monthly average). This declines gradually during the month (because there are fewer days remaining for which the serial correlation in ut contributes to the variance) but then rises quickly at the end of the month because of the large value of γ1, reflecting the increased volatility of the deviations from the target at month end. The second term (κ2共t兲) represents the contribution of target changes to the volatility of the spot-month rate. This contribution declines monotonically as F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Hamilton the day of the month t increases. This is because, as the month progresses, it becomes increasingly likely that the target change for the month has already occurred and there is no more uncertainty about the value of ξn for that month. Added together, expression (16) implies that the variance of changes in the spot-month futures rate should decline over most of the month but then increase at the very end. Expression (16) was derived under the assumption that at the beginning of every month market participants are certain that there will be a target change on day n of the month. If instead there is a fraction ρ of months for which people anticipate a change on some day n and a fraction 1 – ρ for which they are certain there will be no change, the result would be that the last term in (16) would be multiplied by ρ: (17) E ( ft − ft −1 ) = κ 1 (t ) + γ 2κ 2 (t ), 2 where γ 2 = ρσ h2 . This model can be tested using daily data on fed funds futures contracts.2 Figure 3 plots regression coefficients along with 95 percent confidence intervals from a regression of the squared change in the spot-month futures rate on calendar day j : (fd − fd −1 )2 = ∑ βˆ j ω jd + νˆ d , 31 (18) j =1 where ω jd = 1 if business day d occurs on calendar day j and is zero otherwise. In other words, βˆj is the average squared change for observations falling on the jth day of a month. These indeed exhibit a tendency to fall over most of the month but then rise at the end. 2 Data for October 3, 1988, through June 30, 2006, were purchased from the Chicago Board of Trade; data for July 3 through January 29, 2007, were downloaded from the now-defunct web site spotmarketplace.com. For d corresponding to the first day of the month (say the first day of February for illustration), fd – fd –1 was calculated as the change in the February contract between February 1 and the last business day in January. For all other days of the month, it was simply the change in the spot-month contract between day d and the previous business day. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Let td denote the calendar day associated with business day d (in other words, if ω jd = 1, then td = j). I then tested whether the specific function derived in (17) could account for this pattern by estimating via OLS the following relation: (19) .9κ 2 (td ) + νˆ d . (fd − fd −1 )2 = κ 1 (td ) + 27 (3.3) Note that all the parameters appearing in the functions κ1共t兲 and κ2共t兲 are known as described above on the basis of the observed behavior of deviations of the effective fed funds rate from its target, so that only a single parameter—the coefficient on κ2共td 兲 in equation (19)—was estimated directly from the behavior of the futures data. This parameter, γ2, has the interpretation of being the variance of daily news the market receives in a typical month about the upcoming Fed target (recall equation (17)): ˆ ˆ h2 = 27.9 . γˆ 2 = ρσ (3.3) Note also that (19) imposes 30 separate restrictions on the 31 parameters of the unrestricted regression (18). The F(30, 4,521) = 0.59 test statistic leads to ready acceptance of the null hypothesis that this relation is indeed described by the function given in (16) with a p-value of 0.96 (again, treating κj 共t兲 as known functions). The model thus successfully accounts for the tendency of the volatility of the spot-month futures rate to decline over most of the month but then increase the last few days. The actual volatility seems to increase more at the end of the month than the model predicts, though it is possible to attribute this entirely to sampling error. INFERRING MARKET EXPECTATIONS OF TARGET CHANGES FROM THE SPOTMONTH FUTURES RATE We are now in a position to answer the primary question of this paper, which is, What does an observed movement in the spot-month futures rate signal about market expectations about the target rate that is going to be set for this month? J U LY / A U G U S T 2008 385 Hamilton E Yt ( ft − ft −1 ) = N −1 ∑ E Yt ( ft − ft −1 ) η = n N Figure 4 n =1 Plot of κ 4(t) as a Function of t N −t +1 N − n + 1 2 τ = N −1 ∑ ρ E ht = N −1 ∑ γ2 (21) N N τ =1 n =t N κ 4std 2.5 = 2.0 1.5 1.0 2N 2 2= γ 2κ 3 (t ) . Substituting (21) and (17) into (20) establishes 0.5 0 6 1 11 21 16 26 31 Day of Month Let Λt denote the information set available to market participants as of date t, and let Ωt = {ft , ft –1,…} be the information set that is going to be used by the econometrician to form an inference, where it is assumed that Ωt is a subset of Λt , the previous target ξ0 is an element of both Ωt and Λt, and the day n target change is an element of Λt but not of Ωt. Our task is to use the observed data Ωt to form an inference about how the market changed its assessment of ξn based on information it received at t, that is, to form an assessment about ( ) ( Yt = E ξn Λt − E ξn Λt −1 ht forÄ Å t ≤ n = 0 forÄ Å t > n. ) We can calculate the linear projection of Yt on Ωt as follows (e.g., Hamilton, 1994, equation [4.5.27]): (20) ( N − t + 1) ( N − t + 2) γ ( ) Ê Yt Ωt = E Yt ( ft − ft −1 ) 2 E ( ft − ft −1 ) ( ft − f t − 1 ) . Recalling (14), the numerator of (20) can be found from (22) ( ) Eˆ Yt Ωt = κ 3 (t )γ 2 (f − f ) κ 1 (t ) + κ 2 (t ) γ 2 t t −1 = κ 4 (t ) (ft − ft −1 ) . The parameters determining κ4共t兲 have all been estimated above from the properties of the deviations of the fed funds rate from the target and squared changes in the spot-month futures rate. Figure 4 plots the function κ4共t兲 for these parameter values. To understand the intuition for this function, consider first the case in which the fed funds rate is always identically equal to the target, so that σ ε2,t and κ1共t兲 are both zero. From (14), the expected squared change in the spotmonth rate conditional on knowing that the target change will occur on day n would be given by E (ft − ft −1 ) η = n,σ ε2,t = 0 2 σ 2 ( N − n + 1) / N forÅ Ä t ≤ n = h , forÄ Å t > n 0 2 (23) whereas the covariance of the spot-month futures rate change with the expected target rate change would for this case be E (ft − ft −1 )Yt η = n,σ ε2,t = 0 σ h2 ( N − n + 1) / N forÅ Ä t ≤ n . = forÄ Å t > n 0 Thus, if we knew both the day of the target change and that there were no targeting errors, the inference would be Ê ht Ωt ,η = n,σ ε2,t = 0 = βn (t ) (ft − ft −1 ), 386 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Hamilton where N / ( N − n + 1) forÅ Ä t ≤ n , βn (t ) = forÄ Å t > n 0 which reproduces Kuttner’s (2001) formula (7) for the special case considered by Kuttner, namely, t = n. If we don’t know the day of the target change, but still impose no targeting error, we’d use the unconditional moments: 2 E ( ft − ft −1 ) σ ε2,t = 0 = N −1 ∑ σ h2 ( N − n + 1) / N N 2 n =t E ( ft − ft −1 )Yt σ ε2,t = 0 = N −1 ∑ σ h2 ( N − n + 1) / N N n =t Eˆ ht Ωt ,σ ε2,t = 0 = β (t ) (ft − ft −1 ) (24) β (t ) = N −1 ∑ n =t ( N − n + 1) / N N N −1 ∑ n =t ( N − n + 1) / N N 2 . For N large and t = 1, the numerator of (24) would be approximately (1/2) and the denominator about (1/3), so that the coefficient β 共1兲 would be close to 1.5. This is bigger than Kuttner’s expression (7), which equals unity at n = 1, because a one-unit increase in h1 will increase the expected target on day n > 1 by one unit but increase the futures rate on day t = 1 by only [共N – n + 1兲/N ] < 1. Kuttner’s formula assumes that, if we use the day t = 1 change in the futures, the target change occurs on day n = 1, whereas our formula assumes that in all probability the actual change is going to be implemented on some day n > 1. Going from t to t + 1, we drop N –1[共N – t + 1兲/ N ] from the numerator and drop the smaller magnitude N –1[共N – t + 1兲/ N ]2 from the denominator, so that the ratio (24) monotonically increases in t until it finally reaches the same value as (7) on the last day of the month: β (N ) = N . F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W In the presence of targeting errors, expression (22) adds the term κ1共t兲/γ 2 to the denominator of (24), so, as noted by Poole, Rasche, and Thornton (2002), the optimal inference in the presence of targeting errors always puts a smaller weight on ft – ft –1 than does (24). This explains why the function κ4共t兲 in Figure 4 begins at a value below 1.5 for t = 1. The function κ4共t兲 then begins to increase monotonically in t for the same reason as in (24). However, as t increases, both the numerator and denominator in (24) become smaller, whereas κ1共t兲/γ 2 is approximately constant (at least for small t). This latter effect eventually overwhelms the tendency of (22) to increase in t, and it begins to fall after the 20th day of the month. This decline accelerates toward the very end of the month as κ1共t兲 starts to spike up from the endof-month targeting errors. RESPONSE OF INTEREST RATES TO CHANGES IN FEDERAL FUNDS FUTURES We’re now ready to return to the original question of how interest rates for Treasuries of various maturities seem to respond to the spot-month fed funds futures rate. Deviations of the funds rate from the target should have a quite negligible effect on maturities greater than three months, because the autocorrelation implied by (8) dies out within a matter of days. We should therefore find that, if we regress the change in Treasury yields on the change in the spot-month futures rate, the value of the regression coefficient should exhibit exactly the same pattern over the month as the function in Figure 4—the impact should rise gradually through the first half of the month and fall off quickly toward the end of the month. As a first step in evaluating this conjecture, divide the calendar days of a month into j = 1,2, ...,8 octiles and let ψjd = 1 if business day d is associated with a calendar date in the jth octile of the month. For example, ψ1d = 1 if day d falls on one of the first four days of the month, whereas ψ8d = 1 if it falls on the 29th, 30th, or 31st. Let is,d denote the yield in basis points on day d for a Treasury bill or bond of constant maturity s J U LY / A U G U S T 2008 387 Hamilton Figure 5 The Effect of Federal Funds Rate Changes on 1-Year Treasury Yields 3.0 2.5 2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 Day of Month NOTE: The figure plots the coefficients and 95 percent confidence intervals for the OLS regression of daily change in the 1-year Treasury yield on daily change in the spot-month futures rate, with different coefficients for each octile based on the calendar day of the month (denoted by the rectangles and vertical lines) and the predicted values for the coefficients for each day of the month as implied by (26) (denoted by the dashed line). months; for example, i12,d would be the 1-year rate. (Daily Treasury yields were taken from the St. Louis FRED database.) Consider OLS estimation of is ,d − is ,d −1 = ∑ α jsψ jd ( fd − fd −1 ) + usd . 8 (25) j =1 In Figure 5, the OLS estimates, α̂ j共t兲,s, along with their 95 percent confidence intervals, are plotted as a function of calendar day t = 1,2,...,31 for s = 12, which corresponds to a 1-year Treasury security. These indeed display very much the predicted pattern—an increase in the fed funds futures rate around the middle of the month has a slightly bigger effect on the 1-year Treasury rate than it would have at the beginning of the month, 388 J U LY / A U G U S T 2008 and a much bigger effect than it would have toward the end of the month. The same pattern holds for shorter yields (Figure 6) and longer yields (Figure 7). According to the theory, we can capture the exact effect predicted for each calendar day by regressing the change in interest rates on the product between the change in fed funds futures and the function in (22): (26) is ,d − is ,d −1 = λsκ 4 (td )( fd − fd −1 ) + usd , where td is the calendar day of the month associated with business day d, λs is the effect of a onebasis-point increase in the target rate on a Treasury security of maturity s, and usd results from factors influencing yields that are uncorrelated with changes in the expected target rate. Note that all F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Hamilton Figure 6 The Effect of Federal Funds Rate Changes on 3-Month and 6-Month Treasury Yields Effect on 3-Month Treasury Yield 3.0 2.5 2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 24 26 28 30 Day of Month Effect on 6-Month Treasury Yield 3.0 2.5 2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 2 4 6 8 10 12 14 16 18 20 22 Day of Month NOTE: The figure plots the coefficients and 95 percent confidence intervals for the OLS regression of daily change in the 3-month and 6-month Treasury yields on daily change in the spot-month futures rate, with different coefficients for each octile based on the calendar day of the month (denoted by the rectangles and vertical lines) and the predicted values for the coefficients for each day of the month as implied by (26) (denoted by the dashed line). the parameters governing κ4共t兲 have been inferred from the behavior of the fed funds rate and futures alone. Estimates of λs for different maturities, s, are reported in the first column of Table 2, and values of λ̂ sκ4共t兲 for different maturities, s, are plotted as a function of t in Figures 5 to 7. The adequacy of (26) was investigated in a number of different ways. One obvious question is how important the function κ4共td 兲 is for the regression. This can be explored by comparing (26) with a specification in which changes in futures prices have the same effect on interest rates regardless of when within the month they occur: (27) is ,d − is ,d −1 = cs ( fd − fd −1 ) + usd . The specifications (26) and (27) are non-nested, but it is simple enough to generalize to a model that includes them both as special cases: F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W (28) is,d − is,d −1 = c s (fd − fd −1 ) + λsκ 4 (td ) (fd − fd −1 ) + usd . If model (26) is correct, then we should be able to accept the null hypothesis that cs = 0, whereas if (27) is correct, we should accept the null hypothesis that λs = 0. If neither specification is correct, then we should reject both null hypotheses. The second and third columns of Table 2 report the OLS coefficient estimates and standard errors for (28). For maturities greater than two years, we accept the null hypothesis that cs = 0 and strongly reject the hypothesis that λs = 0. For maturities less than two years, both hypotheses are rejected, suggesting that there is more to the response of short-term interest rates to fed funds futures than is captured by (26) alone. Even in these cases, however, the term involving κ4共t兲 makes by far the more important contribution statistically. I conclude that the model successfully captures a J U LY / A U G U S T 2008 389 Hamilton Figure 7 The Effect of Federal Funds Rate Changes on 2-Year, 3-Year, and 10-Year Treasury Yields Effect on 2-year Treasury Yield 3.0 2.5 2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 22 24 26 28 30 22 24 26 28 30 Day of Month Effect on 3-year Treasury Yield 3.0 2.5 2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 2 4 6 8 10 12 14 16 18 20 Day of Month Effect on 10-year Treasury Yield 3.0 2.5 2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 2 4 6 8 10 12 14 16 18 20 Day of Month NOTE: The figure plots the coefficients and 95 percent confidence intervals for the OLS regression of daily change in 2-year, 3-year, and 10-year Treasury yields on daily change in the spot-month futures rate, with different coefficients for each octile based on the calendar day of the month (denoted by the rectangles and vertical lines) and the predicted values for the coefficients for each day of the month as implied by (26) (denoted by the dashed line). 390 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Hamilton Table 2 The Effect of Federal Funds Futures on Interest Rates Restricted effect Maturity λs With constant effect added λs cs With separate octile effects added (p-value for indicated H0) α 1s = … = α 8s = 0 λs = 0 3 Months 0.658** (0.022) 0.256** (0.089) 0.499** (0.060) (0.00)** (0.98) 6 Months 0.706** (0.021) 0.286** (0.084) 0.529** (0.056) (0.00)** (0.45) 1 Year 0.748** (0.023) 0.226** (0.095) 0.608** (0.063) (0.00)** (0.60) 2 Years 0.685** (0.029) 0.159 (0.112) 0.586** (0.079) (0.00)** (0.74) 3 Years 0.641** (0.030) 0.143 (0.122) 0.552** (0.081) (0.01)** (0.62) 10 Years 0.426** (0.028) 0.082 (0.115) 0.375** (0.077) (0.05)* (0.45) NOTE: This table shows the regression coefficients relating change in the interest rate on securities with maturity s to change in the fed funds futures rate. *indicates statistically significant with p-value <0.05; **denotes p-value <0.01. OLS standard errors are in parentheses. clear tendency in the data for the impact to vary across the month, although it seems to leave something out in the description of the response of short-term interest rates. In the same spirit, we can nest (26) and (25): is ,d − is ,d −1 = (29) ∑α js ψ jd ( fd − fd −1 ) + λsκ 4 (td )( fd − fd −1 ) + usd . 8 j =1 The results, shown in the last two columns of Table 2, are not as encouraging. In every case, we strongly reject the hypothesis that α 1s = … = α 8s = 0, meaning that for each maturity, s, there are statistically significant deviations from the broad monthly pattern that is predicted by (26), and in every case readily accept the hypothesis that λs = 0, meaning that the specific variation within octiles that is predicted by (26) is not particularly found in the data. These last results are perhaps not too surprising given the many approximations embodied in (26), which assumed among other things that all months have N = 31 calendar days and ignored F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W both weekend effects and that some business days convey much more important economic news than others (on this last point, see Poole and Rasche, 2000, and Gürkaynak, Sack, and Swanson, 2005). We can in fact carry that last point a step further and estimate a separate coefficient λjs for every calendar day j = 1,…,31: is ,d − is ,d −1 = ∑ λ jsω jd ( fd − fd −1 ) + usd , 31 (30) j =1 where ωjd = 1 if day d falls on the jth day of the month. Figure 8 plots the OLS estimates of λjs as a function of the calendar day j along with 95 percent confidence intervals and the predicted values for the function λjs implied by (26) for 1-year Treasuries. Again the broad pattern seems to fit well, though again there are large deviations on some days that are well beyond what could be attributed to sampling error, and formal hypothesis tests comparing (30) with (26) (which the former formally nests as a special case) lead to overwhelming rejection, with a p-value less than 10–10 for each s. In addition to the details noted J U LY / A U G U S T 2008 391 Hamilton Figure 8 Effects of Federal Funds Rate Changes on 1-Year Treasury Yields 3.0 2.5 2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 Day of Month NOTE: The figure plots the coefficients and 95 percent confidence intervals for the OLS regression of daily change in the 1-year Treasury yields on daily change in the spot-month futures rate, with different coefficients for each calendar day of the month (denoted by the rectangles and vertical lines) and the predicted values for the coefficients for each day of the month as implied by (26) (denoted by the dashed lines). above, individual outliers are highly influential for the daily regression (30), and one would want to carefully model these non-Gaussian innovations, usd , and GARCH effects before trying to build a more detailed model that could reproduce more of the unrestricted pattern. This and related tasks, such as trying to use information about the actual date of the target change when it is unambiguously known, using one-month or two-month futures contracts in place of the spot rate, and exploring the consequences of a secular change in σ h2 (e.g., Lang, Sack, and Whitesell (2003) and Swanson, 2006), we leave as topics for future research. Although there is much more to be done before having a completely satisfactory understanding of these relations, I believe that the 392 J U LY / A U G U S T 2008 approach developed here gives us a plausible interpretation of the broad regularities found in the data and a sound basis for generalizing the Kuttner (2001) and Poole, Rasche, and Thornton (2002) approaches. Although the methods involve some new uses of the data, the conclusion I draw is quite consistent with earlier researchers— changes in the fed funds target seem to be associated with quite large changes in Treasury yields, even for maturities of up to 10 years. REFERENCES Anderson, Richard G. and Rasche, Robert H. “A Revised Measure of the St. Louis Adjusted Monetary Base.” Federal Reserve Bank of St. Louis Review, March/April 1996, 78(2), pp. 3-14. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Hamilton Cook, Timothy and Hahn, Thomas. “The Effect of Changes in the Federal Funds Rate Target on Market Interest Rates in the 1970s.” Journal of Monetary Economics, November 1989, 24(9), pp. 331-51. Gürkaynak, Refet S. “Using Federal Funds Futures Contracts for Monetary Policy Analysis.” Working paper, Board of Governors of the Federal Reserve System, 2005. Gürkaynak, Refet S.; Sack, Brian P. and Swanson, Eric T. “Do Actions Speak Louder Than Words? The Response of Asset Prices to Monetary Policy Actions and Statements.” International Journal of Central Banking, June 2005, 1(1), pp. 55-93. Gürkaynak, Refet S.; Sack, Brian P. and Swanson, Eric T. “Market-Based Measures of Monetary Policy Expectations.” Journal of Business and Economic Statistics, April 2007, 25(2), pp. 201-12. Hamilton, James D. Time Series Analysis. Princeton, NJ: Princeton University Press, 1994. Hamilton, James D. “Daily Changes in Fed Funds Futures Prices.” Journal of Money, Credit, and Banking (forthcoming). Kuttner, Kenneth N. “Monetary Policy Surprises and Interest Rates: Evidence from the Fed Funds Futures Market.” Journal of Monetary Economics, June 2001, 47(3), pp. 523-44. Lang, Joe; Sack, Brian P. and Whitesell, William. “Anticipations of Monetary Policy in Financial Markets.” Journal of Money, Credit, and Banking, December 2003, 35(6, part 1), pp. 889-909. Piazzesi, Monika and Swanson, Eric T. “Futures Prices as Risk-Adjusted Forecasts of Monetary Policy.” Journal of Monetary Economics (forthcoming). Poole, William and Rasche, Robert H. “Perfecting the Market’s Knowledge of Monetary Policy.” Journal of Financial Services Research, December 2000, 18(2/3), pp. 255-98. Poole, William; Rasche, Robert H. and Thornton, Daniel L. “Market Anticipations of Monetary Policy Actions.” Federal Reserve Bank of St. Louis Review, July/August 2002, 84(4), pp. 65-94. Sarno, Lucio; Thornton, Daniel L. and Valente, Giorgio. “Federal Funds Rate Prediction.” Journal of Money, Credit, and Banking, June 2005, 37(3), pp. 449-71. Swanson, Eric T. “Have Increases in Federal Reserve Transparency Improved Private Sector Interest Rate Forecasts?” Journal of Money, Credit, and Banking, April 2006, 38(3), pp. 791-819. Taylor, John B. “Expectations, Open Market Operations, and Changes in the Federal Funds Rate.” Federal Reserve Bank of St. Louis Review, July/August 2001, 83(4), pp. 33-47. Thornton, Daniel L. “A New Federal Funds Rate Target Series: September 27, 1982–December 31, 1993.” Working Paper 2005-032, Federal Reserve Bank of St. Louis, 2005; http://research.stlouisfed.org/wp/2005/2005-032.pdf. Nilsen, Jeffrey H. “Borrowed Reserves, Fed Funds Rate Targets, and the Term Structure,” in Ignazio Angeloni and Riccardo Rovelli, eds., Monetary Policy and Interest Rates. London: Macmillan Press, 1998. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 393 394 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Alec Chrystal I am very pleased to be asked to participate in this conference that honors the career of Bill Poole. As a student of monetary economics I, like all my generation, was substantially influenced by Poole (1970). I was pleased to meet him many times at the annual St. Louis economic policy conference, both before and after he became president of the Federal Reserve Bank of St. Louis. We were also very grateful that he came to London in 2000 to give the annual Henry Thornton Lecture at the Cass Business School of City University. I now turn to my comments on Professor Hamilton’s (2008) paper. The paper uses data from daily movements in federal funds futures to test for links between futures prices, the policy rate itself, and the behavior of market interest rates. I first comment on the empirical work presented and then suggest additional avenues of research to further enlighten the topic. I then ask this: Who might be interested in these results and what might they learn from them? Some of the difficulty in using the federal funds futures price as an indicator of market expectations arises from the fact that the contract settles on an average daily price over a month. I do not wish to get into the institutional detail here or into the econometric problems this causes, not least because I am dominated in institutional knowledge by Ken Kuttner, the other discussant, and in econometric expertise by Professor Hamilton himself. However, as a naive outsider, I cannot help but ask whether there is some “cleaner” money market interest rate that contains the same information but avoids the complexities of moving-average valuation. Could we, for example, do roughly the same exercise with shortterm Treasury bill discount rates, short maturity Treasury bond yields, or indeed interbank loan rates? If we could, then it would surely be simpler to use these rates and parsimony would lean in their favor. Assuming now that the federal funds futures prices are the best proxy for market expectations, what do the results tell us, and what else might we like to know? The results reported in this paper confirm two earlier findings: First, market rates anticipate actual policy rate changes and, second, other market rates (yields to maturity) move with the federal funds rate, including those of up to 10-year maturity. I will discuss each of these in turn. It is not a major surprise to find that markets anticipate policymakers’ decisions. That this is highly likely has been central to economics since the rational expectations revolution of the 1970s. However, it would be interesting to know if markets have become better at doing this over time and whether this ability has been affected by improved transparency about the target rate, the stated biases in the policy stance, and what is being targeted. Similar questions apply to the unexpected component of policy changes: Has the impact of policy changed over time, and are the results for the full sample dependent on specific periods or specific sets of events? Alec Chrystal is a professor of money and banking and associate dean and head of the faculty of finance at Sir John Cass Business School, City University, London, England. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 395-97. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 395 Chrystal A slightly different point arises in the context of testing the effects of federal funds futures on other market rates. It makes sense that short rates move most closely with the federal funds rate. However, the theoretical link with long rates is rather more ambiguous. In what way should long rates react to changes (and expected changes) in short-term policy rates? This could go either way, and indeed there could be no link at all. Suppose the Fed is tightening rates in order to bring down inflation in the future. This will raise implied forward rates up to some term, but it may lower interest rate expectations further out, that is, tip the yield curve. In such cases, forward rates further out will change but the change could quite logically be in the opposite direction. Tighter policy now could lower inflation expectations further out and, hence, create expectations of lower interest rates in the future. Hence, it might be interesting to test the reaction of long forward rates, in addition to yields to maturity, as the former will strip out the impact at the short end of the yield curve. What actually happens in each case will be dependent on the complexity of the environment, and the reaction might be asymmetric— rises may have a different impact than falls. Why and to whom is all this likely to be interesting? Potentially there are three groups who may be able to learn something from the relationships that emerge from this and similar studies: first, the monetary authorities themselves; second, market participants who trade in these and related markets; and third, those in the economics profession who want to understand how monetary policy works, that is, those with an interest in the transmission mechanism. The monetary authorities may be interested in all this for two possible reasons. First, by monitoring the federal fund futures they can see what the markets expect policy to be and can factor that into their decisions. Second, they could understand what impact an unexpected rate change has on the markets. (I will return later to the issue of whether these results mean that only unexpected rate changes matter.) I do not know for sure, but my guess is that Federal Open Market Committee members have reliable ways of backing out market expectations and of estimating 396 J U LY / A U G U S T 2008 the impact of their policy rate changes without having to rely on this evidence from the federal funds futures market. Hence, I suspect that the contribution of these results to policymakers’ decisionmaking is quite small. Market participants have little to learn from these results because the federal funds futures prices reflect their behavior in the first place, so they are not going to learn about their own expectations from a price that their behavior has created. There may be something that these players could learn from federal funds futures prices, but only if the data were much more finely sampled. Tick-by-tick data for this and other closely linked money markets might help to identify exactly where changes in sentiment first appear. Market traders probably know this already, but it is also possible that the news for some episodes appears to some segments of the market first. However, it is more likely that market participants get new information more or less simultaneously and the timing of market movements is purely a product of how we measure the “market price.” That is, all prices respond as quickly as is technically possible to the same information. So what can we as economists learn from all this about the transmission mechanism of monetary policy? I suggest that this evidence does nothing but confirm what we already knew: Markets anticipate what policymakers are going to do, and markets move most when the policy change is most unexpected. However, I should emphasize that this evidence neither supports nor confounds the old notion of the Lucas aggregate supply curve, by which only unexpected policy changes have real effects. To see this, I hypothesize that monetary policy works through a number of channels to influence aggregate demand in the economy. Only one of these channels is the direct effects on other market interest rates. Other channels include asset prices (and thus wealth effects), expectations and confidence, and international financial markets (and thus the exchange rate). The fact that market rates anticipate policy rate changes does not mean that the changes have no effect; it just means that the effects happen sooner. Market rate changes will still affect saving and investment decisions and F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Chrystal thus also aggregate demand. They will also affect asset valuations and thus create wealth effects. Unexpected policy rate changes may well have a bigger measurable impact on market rates of all maturities, but this does not prove either that only unexpected rate changes have real effects or that unexpected rate changes have bigger real effects. It remains possible that unexpected policy changes have a bigger impact on aggregate demand, but the evidence adduced here does not address this issue. In short, this paper contains some outstanding innovative econometric work that throws much light on the links between federal funds futures prices, the policy rate, and other market rates. However, the results have no apparent implications that should cause us to revise our view of how monetary policy works. REFERENCES Poole, William. “Optimal Choice of Monetary Policy Instrument in a Simple Stochastic Macro Model.” Quarterly Journal of Economics, 1970, 84(2), pp. 197-216. Hamilton, James D. “Assessing Monetary Policy Effects Using Daily Federal Funds Futures Contract.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 377-93. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 397 398 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Kenneth N. Kuttner R ecent efforts to understand the transmission of monetary policy have spawned a growing literature examining the response of financial markets to monetary policy.1 Most of these studies assess the likely impact of unanticipated changes in the target federal funds rate, typically in a sample of well-defined policy “events” consisting of Federal Open Market Committee (FOMC) meeting days, plus the days of unscheduled funds rate changes. The problem is that economists do not always know the days on which the policy actions took place, especially in the early 1990s. Before the FOMC began announcing its policy actions in February 1994, there was often some confusion in the financial markets as to whether there had been a change in the funds rate target. This ambiguity has been largely dispelled by the FOMC’s announcements, although, as Hamilton (2008) notes, there has been occasional speculation that the Fed has surreptitiously changed the target rate.2 Hamilton’s (2008) paper is primarily an effort to address the issue of unknown event dates. It departs from the usual assumption that the days of policy actions (or possible actions) are known 1 The first paper in this literature was Cook and Hahn (1989). Subsequent work includes Poole and Rasche (2000), Kuttner (2001), Poole, Rasche, and Thornton (2002), Gürkaynak, Sack, and Swanson (2005), and Bernanke and Kuttner (2005). 2 So far, none of this speculation has proved to be correct. and uses instead a signal-extraction approach to determine the market’s reaction without conditioning on this information. His elegant approach allows the market’s reaction to be estimated using the entire sample, not just event days. Moreover, the approach allows for the measurement of financial markets’ response to evolving expectations of future Fed actions, a feature that allows him to extract information even when the Fed does not surprise the markets. The analysis focuses on the response of term interest rates, as in Kuttner (2001), although there is no reason the same approach could not also be applied to stock prices or exchange rates. The paper’s key empirical results largely confirm those reported elsewhere, which is good news for those of us who have used the much simpler event-study approach. The response of term interest rates to changes in the funds rate is uniformly less than one for one, and the effect on longer-term interest rates is generally less than it is for short-term rates. It is interesting to note, however, that this latter tendency is less pronounced than it is in Kuttner’s (2001) results. My discussion will focus on two issues. The first point is somewhat technical, as it concerns the details of how the “noise” in the federal funds rate is modeled. The second is a more conceptual discussion of how the interpretation of the shocks identified by Hamilton’s procedure might differ from those in conventional event-study analyses. Kenneth N. Kuttner is a professor of economics at Williams College and a research associate at the National Bureau of Economic Research. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 399-403. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 399 Kuttner Figure 1 The Target and Effective Funds Rates, 1995 Percent 7.5 7.0 6.5 6.0 5.5 5.0 4.5 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 1995 NOTE: Vertical lines denote settlement Wednesdays. MODELING FUNDS RATE NOISE Unlike the more common event-study analysis, Hamilton’s signal-extraction method requires statistically modeling the noise process present in the daily effective federal funds rate.3 Intuitively, the reason for this is that calculating the likely signal present in any given funds rate change requires some estimate as to the amount of noise likely to be present on any given day: The noisier the effective funds rate, the less likely it is that the observed change in the rate (and, by extension, the expected rate implied by the current-month futures contract) represents a policy change. Observing that the magnitude of these deviations tends to increase over the course of a month, Hamilton models the targeting error as an auto3 This noise results from the fact that the New York Fed’s control over the funds rate is not absolute: Their Trading Desk injects just enough reserves to hit the target funds rate, given its assessment of the factors affecting reserve demand and supply. However, because of unanticipated changes in demand or supply, the actual (“effective”) funds rate may differ from the target. 400 J U LY / A U G U S T 2008 regressive process whose innovation variance is a linear function of the day of the month (equations (8) and (9)). To get a sense of the magnitude of these targeting errors, his estimated parameters imply a 45-basis-point standard deviation on the 31st day of the month, 34 basis points on the 30th day, and 17 basis points on the 1st day. Although this is not an unreasonable first pass, some refinements are possible. First, because there is no reason to think that the end-of-month volatility in 31-day months is greater than it is for 30-day months, it would be desirable to relax the assumption of 31-day months and replace equation (9) with eˆ d2 = a + b × 0.5( N i −td ) + νˆ d , where Ni is the number of days in month i. A second important refinement would be to account for the “settlement Wednesday” effect. Especially in the early part of the sample, the Wednesdays associated with the final day of the reserve maintenance period were often associated F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Kuttner Figure 2 The Target and Effective Funds Rates, 2002 Percent 2.00 1.75 1.50 1.25 1.00 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jul Aug Sep Oct Nov Dec 2002 Percent 2.00 1.75 1.50 1.25 1.00 Jan Feb Mar Apr May Jun 2002 NOTE: The vertical lines in the top panel denote settlement Wednesdays; in the bottom panel they mark the last day of the month. with extremely large funds rate spikes, as shown in Figure 1. (The vertical lines denote settlement Wednesdays.) To account for this pattern, a reasonable specification for the targeting error might be something like rd − ξd = 0.3( rd −1 − ξd −1 ) + 14Wd + 19M d + eˆ d eˆ d2 = 179 + 196Wd +1 + 1,422Wd − 27M d +1 +1,481M d + νˆ d , where Wd is a dummy equal to 1 on settlement Wednesdays and Md is a dummy equal to 1 on the last day of the month. The other notation is the same as Hamilton’s.4 Three features of this alternative specification are particularly interesting. One is that there are 4 The parameter estimates are estimated using ordinary least squares from May 17, 1989, through October 12, 2007, excluding September 2001 and December 1999. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W significant level effects associated with settlement Wednesdays and with the last day of the month: Errors on these days tend to be positive. The second is that the standard deviation of the targeting error is 27 basis points higher on settlement Wednesdays. Third, unlike in Hamilton’s specification, there is no evidence of a month-end effect, except on the very last day of the month. Other information about changes in the federal funds market can also be brought to bear to further refine the specification. One such change is the shift to lagged reserve accounting as of July 30, 1998. Partly as a result of this change, the monthend and settlement-Wednesday volatility of the funds rate, as well as the overall variance, has fallen sharply in recent years. Post 1998, the standard deviation of the last-day-of-month targeting error is only 22 basis points (compared with Hamilton’s last-day estimate of 45 basis points), and there is no longer any evidence of a settlement-Wednesday spike. That the Federal Reserve J U LY / A U G U S T 2008 401 Kuttner Bank of New York’s Trading Desk has improved control over the funds rate is readily apparent in Figure 2. (The vertical lines in the top panel denote settlement Wednesdays; in the bottom panel they mark the last day of the month.) Finally, in refining the estimates of the targeting error process, one would want to make allowances for special circumstances affecting the federal funds market. Hamilton already makes one such allowance, omitting September 2001 from the sample used for estimating the model. December 1999 should be dropped for similar reasons: With the Y2K changeover approaching, the Fed flooded the market with reserves in an effort to assuage liquidity concerns. Consequently, the funds rate traded as much as 150 basis points below its target as the end of the month approached. Including atypical episodes, such as this one, could overestimate the amount of noise normally present in the effective funds rate. It is important to emphasize that none of these observations undercuts in any way the soundness of Hamilton’s basic approach. In particular, I can think of no reason to suspect that any misspecification in equations (8) or (9) would necessarily bias the parameter estimates reported in Hamilton’s Table 2. Instead, it is more akin to the problem of choosing inappropriate weights in a weighted-least-squares procedure: In that case, while the parameter estimates may not be biased, the procedure is not making optimal use of the information contained in the data. ON INTERPRETING THE “HAMILTON SHOCKS” The second part of my remarks concerns the interpretation of the funds rate shocks underlying Hamilton’s procedure. By way of background, it may be useful to distinguish between two different regimes. In the first regime, changes in the funds rate target are equally likely on any day—but changes in the target are not announced by the FOMC. This regime plausibly corresponds to the pre-1994 world, in which policy actions were generally not disclosed and a significant fraction of rate changes took place between meetings. In 402 J U LY / A U G U S T 2008 this regime, day-to-day changes in the futuresimplied rate on any particular day would plausibly represent the market’s inference as to whether the Fed had changed its target on that day. In the second regime, which is more relevant post 1994, the days of the rate changes are largely known; and even when policy actions are taken between FOMC meetings, the changes are announced, and not in response to any specific news that might have arrived on that day. In this case, the day-to-day change in the futures rate on days other than “event” days (i.e., days of rate changes or FOMC meetings) would reflect changes in the market’s expectation of the target funds rate on some future date. Now consider the sources of news that could affect policy expectations. One source is new macroeconomic information: higher-than-expected employment, for example, or lower-than-expected inflation. The other source would be changes in the Fed’s perceived preferences regarding inflation vis à vis output—the presumed source of monetary policy “shocks,” as the term is commonly used in the literature. These distinctions bear on how we should interpret the information contained in alternative measures of monetary policy shocks or surprises. In the second regime, policy surprises (i.e., changes in the futures rate) occurring on event days are more likely to be driven by the second category of news: changes in the Fed’s perceived preferences.5 Changes occurring on days other than event days would, for the most part, be associated with the arrival of economic news. In the first regime, however, day-to-day changes in the futures rate could result from either source: changes in policy preferences or macro news. Thus, conditioning on known event days allows the econometrician to distinguish between the endogenous response of policy expectations to new economic information and otherwise inexplicable policy shocks. This distinction can be critically important in assessing the financial market response to monetary policy. As shown in 5 It is also possible that the change would be interpreted as the Fed reacting to private information, although the evidence for this view is weak; see Faust, Swanson, and Wright (2004). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Kuttner Bernanke and Kuttner (2005), the stock market’s reaction to unanticipated funds rate changes is effectively zero when those changes occurred on the same day as an employment report, a pattern that was common in the early 1990s. Any analysis that failed to make this distinction could provide a misleading answer to the primary question of interest to policymakers: how the market will react to an unexpected change in the fund rate target. Within Hamilton’s framework, it would be easy to make this distinction. In fact, it suggests an interesting test of the null hypothesis that the response to (calendar-adjusted) changes in the futures rate is the same on event days as it is on non-event days. The alternative, of course, would be that the reaction differs in a systematic way. Given the richness of the dataset, it would be possible to go further and distinguish between event days and the days of specific economic news releases (e.g., inflation, employment, gross domestic product). This assumes that the relevant days are known, of course—but after 1994, this is not such a bad assumption. Econometrically, the only modification to Hamilton’s procedure would be to allow the relevant dummy variables to interact with the slope coefficient in equation (26). CONCLUSION None of these points takes away from the bottom line: The paper is a classic Hamilton timeseries tour de force. It addresses an important question using elegant econometrics, and it incorporates a detailed knowledge of the market for federal funds. Using more of that knowledge to refine the targeting-error specification would enhance an already fine paper, as would further efforts to understand what the shocks in the model really represent. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W REFERENCES Bernanke, Ben S. and Kuttner, Kenneth N. “What Explains the Stock Market’s Reaction to Federal Reserve Policy?” Journal of Finance, June 2005, 60(3), pp. 1221-57. Cook, Timothy and Hahn, Thomas. “The Effect of Changes in the Federal Funds Rate Target on Market Interest Rates in the 1970s.” Journal of Monetary Economics, November 1989, 24(3), pp. 331-51. Faust, Jon; Swanson, Eric T. and Wright, Jonathan H. “Do Federal Reserve Policy Surprises Reveal Private Information About the Economy?” Contributions to Macroeconomics, 2004, 4(1), pp. 1-29. Gürkaynak, Rafet S.; Sack, Brian P. and Swanson, Eric T. “Do Actions Speak Louder Than Words? The Response of Asset Prices to Monetary Policy Actions and Statements.” International Journal of Central Banking, June 2005, 1(1), pp. 55-93. Hamilton, James D. “Assessing Monetary Policy Effects Using Daily Federal Funds Futures Contracts.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 377-93. Kuttner, Kenneth N. “Monetary Policy Surprises and Interest Rates: Evidence from the Fed Funds Futures Market.” Journal of Monetary Economics, June 2001, 47(3), pp. 523-44. Poole, William and Rasche, Robert H. “Perfecting the Market’s Knowledge of Monetary Policy.” Journal of Financial Services Research, December 2000, 18(2/3), pp. 255-98. Poole, William; Rasche, Robert H. and Thornton, Daniel L. “Market Anticipations of Monetary Policy Actions.” Federal Reserve Bank of St. Louis Review, July/August 2002, 84(4), pp. 65-93. J U LY / A U G U S T 2008 403 404 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Panel Discussion The Importance of Being Predictable John B. Taylor I t is a pleasure to participate in this conference and join in the recognition of Bill Poole. My remarks build on two of Bill Poole’s important contributions to monetary theory: his 1970 Quarterly Journal of Economics (QJE) paper on monetary policy under uncertainty and his more recent series of lucid short papers on predictability, transparency, and policy rules, many of which were adapted from speeches and published in the Review of the Federal Reserve Bank of St. Louis. At the same time I want to express my appreciation for Bill’s extraordinary service in public policy: starting in the 1960s as a member of the staff of the Federal Reserve Board, where he wrote his 1970 QJE paper and many others; then later as a member of the President’s Council of Economic Advisers during the difficult disinflation of the early 1980s, where his role in explaining and supporting the Fed’s price stability efforts was essential; and most recently as president of the Federal Reserve Bank of St. Louis, where his emphasis on good communication and good policy has contributed, and will continue to contribute, to improvements in the conduct of monetary policy. Regarding these contributions I give two of my favorite examples of Bill Poole’s many pithy phrases which I hope will ring in monetary policymakers’ ears for many years to come: “We ignore the behavior of the monetary aggregates at our peril” (Poole, 1999); and “Clearly, more talk does not necessarily mean more transparency” (Poole, 2005a). THE BEGINNINGS OF RESEARCH ON POLICY RULES IN STOCHASTIC MODELS CIRCA 1970 Let me begin by reviewing Bill Poole’s deservedly famous 1970 QJE article. In my view, that paper conveyed two novel messages, one about dealing with uncertainty and the other about reducing uncertainty. An Approach to Monetary Policy That Could Deal with Existing Uncertainty The first message was presented in the form of a simple graphical ISLM analysis, and soon after textbook writers incorporated this analysis in their macroeconomics and money and banking textbooks. At the time Poole wrote his paper, the typical IS and LM curves were drawn without a notion that they could move around stochastically. Bill Poole showed how adding exogenous disturbances to the curves provided a simple framework for monetary policy decisionmaking under uncertainty. While the framework was simple, the message was extremely useful: When shocks to money demand are very large, central banks should target the interest rate because those shocks would otherwise cause harmful swings in interest rates. When John B. Taylor is a professor of economics at Stanford University. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 405-10. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 405 Panel Discussion shocks to investment demand or consumption demand are very large, central banks should target the money supply because the interest rate will move to mitigate these demand shocks. Hence, the Poole analysis showed explicitly how policymakers could deal with exogenous uncertainty in a formal mathematical way. An Approach to Monetary Policy That Could Reduce Uncertainty The second message was more complex and profound, and also more relevant for my purpose here. Poole investigated what he called a “combination policy” involving both the interest rate and the money supply, and he examined its properties in an economy-wide dynamic stochastic model. The model, with the combination policy inserted, could be written as a vector autoregression. Poole showed how to compute the steadystate stochastic distribution implied by the model. He also showed how to find the optimal policy to minimize the variance of real gross domestic product (GDP) around the mean of this stochastic steady-state distribution. The method involved finding the homogeneous and particular parts of the solution and then writing the endogenous variable as an infinite weighted sum of lagged shocks—what is now commonly called an impulse response function. The combination policy had key features of active monetary policy rules in use today. The policy involved the money supply (M), the interest rate (r), and lagged values of real GDP (Y ). Poole wrote it algebraically as M = c1′ + c2′ r + lagged values of Y , where the coefficients c1′ and c2′ were determined to minimize the variance of real GDP in the steadystate stochastic distribution. He showed that the optimal policy yielded a smaller loss than the fixed interest rate policy, the fixed money supply policy, or a combination policy that ignored the reactions to lagged real GDP. Note that, although the rule was active, there was no discretion here. Once those parameters were chosen, they would stay for all time. People criticized Poole for this rule approach and argued 406 J U LY / A U G U S T 2008 instead in favor of discretion. They said that policymakers could see or forecast the shocks to the LM curve and the IS curve and adjust the policy instruments as they saw fit without having to stick to any one policy rule. For example, I have a vivid memory of discussing the Poole paper with Franco Modigliani after I presented a paper at MIT later in the decade. He insisted that there was no reason to constrain policymakers the way Poole did. There was still an enormous resistance to policy rules, even the active sort, at this time. However, although discretionary actions might improve performance in a given situation, the possibility of discretion, and especially its misuse, could add to the uncertainty already in the markets. The advantage of Poole’s active policy rules was that they were more predictable and could therefore reduce uncertainty. The second lesson from Poole’s 1970 paper was thus that policymaking based on rules would improve economic performance by reducing uncertainty compared with policymaking based on pure discretion. This same basic stochastic dynamic modeling approach was applied again and again in the 1970s and 1980s, eventually to more complex empirically estimated models with rational expectations and sticky prices. Optimal rules were computed in these newer models. Over time the resistance to active policy rules began to weaken. Most surprising was that actual monetary policy decisions became more predictable and could even be described closely by policy rules. Most rewarding was that the more predictable rule-like behavior yielded improved policy performance. And most interesting is that we can now look back at this period of greater predictability and learn from it. RULES OF THUMB IN THE PRIVATE SECTOR An unanticipated advantage—at least from the vantage point of 1970—of the more predictable behavior by central banks has been the response of the private sector. Recognizing that the central bank’s interest rate settings are following more regular rule-like responses to such variables as F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Panel Discussion inflation and real GDP, the private sector has taken these responses into account in projecting future variables and in developing their own rules of thumb for making decisions. An important example is the formation of expectations of future interest rates, which affect bond traders’ and investors’ decisions and thereby influence longterm interest rates, as has been emphasized by Poole in his more recent writings. I quote from a paper he gave earlier this year (Poole, 2007, p. 6): What our analysis missed a generation ago was that the typical model with only one interest rate could not possibly allow for stabilizing market responses in long rates when the central bank set the short rate. Of course, macro econometric models did have both short and long rates, but the structure of the models did not permit analysis of the sort I am discussing because the typical term structure equation made the long rate a distributed lag on the short rate. The model’s short rate, in turn, was determined by monetary policymakers setting it directly or by the money market under a policy determining money growth. Once we allow expectations to uncouple the current long rate from the current short rate, the situation changes dramatically. The market can respond to incoming information in a stabilizing way without the central bank having to respond. Long bond rates can change, and change substantially, while the federal funds rate target remains constant. In this example, the private sector has adapted to a particular policy rule in which the short-term interest rate rises by a predictable amount when inflation rises. Thus, if expectations of inflation rise, the private sector will predict that the central bank will raise short-term interest rates in the future; traders will then bid down bond prices, raising long-term interest rates, and thereby mitigating the inflationary impulse before the central bank action is needed. There are other examples where private sector behavior has adapted to rule-like behavior of the central bank. Consider foreign exchange markets. Empirical studies show that when there is a surprise increase in inflation, the immediate reaction in foreign exchange markets is an appreciation of the currency. Yet conventional price theory F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W would predict the opposite, a negative correlation between exchange rates and inflation, because higher prices make goods at home relatively expensive, requiring a depreciation of the currency to keep purchasing power from moving too far away from parity. But the regular central bank interest rate response to inflation explains the empirical correlation. How? An increase in inflation implies that the central bank will raise the interest rate, which makes the currency more attractive, bidding up the exchange rate. There are many other examples where individuals and institutions in the private sector adapt to policy-induced correlations. In effect, they are creating their own rule-like behavior, their own rules of thumb, and we are probably unaware of most of them. Indeed, the individuals who act on them may not even know that they derive from the rule-like behavior of policymakers. Of course, it is not only the private sector in the United States. Markets all over the world follow closely what the Fed is likely to do. And it is not only the private sector. Central banks take account of the predictable behavior of the other central banks and in particular the behavior of the Federal Reserve, which matters greatly for their own decisions. For example, the recent June 2007 Monetary Policy Report of the Norges Bank states that “It cannot be ruled out that a wider interest rate differential will lead to an appreciation of the krone. This may suggest a gradualist approach in interest rate setting.” In other words, actions by the Federal Reserve that affect the interest rate differential will in turn influence interest rates set by other central banks. This effect can also occur automatically— another rule of thumb—if model simulations used to set interest rates at central banks assume, as they usually do, that other central banks follow such policy rules. An implication of this development is that if central banks depart from their regular responses, then they run the risk of disrupting private sector rules of thumb. Even if they explain the reason for the irregular behavior as clearly as possible, emphasizing that it is temporary, some individuals or institutions may continue operating with the old rules of thumb unaware that these rules have J U LY / A U G U S T 2008 407 Panel Discussion anything to do with the monetary policy–induced correlations. For example, during the period from 2002 to 2005, the interest rate in the United States fell well below levels that would have been predicted from the behavior of the Federal Reserve during most of the period during the Great Moderation. Using modern time-series methods, Frank Smets and Marek Jaronciński (2008) showed in their paper for this conference that there was such a deviation, and they linked the deviation to the boom and bust in housing prices and construction. In Taylor (2007), I argued that the resulting acceleration of housing starts and housing prices, as well as the low interest rates, may have upset rules of thumb that mortgage originators were using to assess the payment probabilities based on various characteristics of the borrower. Their programs are usually calibrated in a cross section at a point in time. If housing prices start rising rapidly, the cross section will show increased payment probabilities, but the programs will miss this time-series element. When housing prices reverse, the models will break down. It would have been very difficult to predict a breakdown in the rules of thumb such as the mortgage underwriting programs, but if it had not been that rule of thumb, it might have been another. Another related example was the negligible response of long-term interest rates when the Federal Reserve raised short-term interest rates in 2004 and 2005. This might be explained by this same deviation. Investors may have felt that the Fed had departed from the kind of rule that formed the basis of the longer-term interest rate responses of the kind discussed in the above quote by Poole. Two examples from international monetary policy issues are also worth noting. Following the Russian debt default and financial crisis of 1998 there was a global contagion that affected emerging markets with little connection to Russia. The contagion even reached the United States, led to the Long Term Capital Management crisis, and caused enough of a freeze-up in U.S. markets that the Federal Reserve reduced the interest rate by 75 basis points. In contrast, following a very similar default and financial crisis in Argentina in 2001, there was virtually no contagion. The 408 J U LY / A U G U S T 2008 main difference between these two episodes in my view is predictability. In the case of Russia, the International Monetary Fund suddenly removed financial support, only one month after renewing it. This surprise disrupted the world’s financial markets. In contrast, in the case of Argentina, the International Monetary Fund gradually reduced support and was as clear as it possibly could be in its intentions. Hence, there was little surprise. The default and currency crises were discounted by the time they happened. Another international example is the currency intervention policy of the United States and the other key currency countries. There has been no intervention by the United States or Europe in these markets since September 2000. And since March 2004, Japan has not intervened. Moreover, most policymakers in these countries have suggested a strong aversion to intervention in the currency markets. In effect, compared with a policy of frequent intervention, as in the 1980s and 1990s, the currency policy has become much more predictable. The assumption of zero intervention in most circumstances is a good one. What has been the result? The behavior of the major currencies has been less volatile and even the volatility of volatility has come down. It is difficult to prove causality in any of these examples, and certainly more research is needed. Our experience with different degrees of predictability is increasing and strongly suggests advantages of policy predictability and risks of unpredictability. Toward Greater Predictability There have been great strides in improving monetary policy predictability at the Federal Reserve and other central banks in recent years, as Bill Poole has documented and explained (Poole, 2003 and 2005a,b; Poole and Rasche, 2003). Can we make monetary policy even more predictable? One suggestion is to publish the Fed’s balance sheet on a daily basis, or at least the Fed balances that commercial banks hold at the Fed. This would make it easier to interpret episodes where the central bank decides to provide additional liquidity in the overnight money market, as on August 9 and 10 of this year. The available data on repos F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Panel Discussion do not provide the information that analysts need to interpret these actions and to distinguish them from monetary policy actions aimed at overall macroeconomic goals of price stability and output stability. Another suggestion would be to publish some of the key assumptions used in formulating policy, including potential GDP and/or the GDP gap, or at least publish these with a shorter lag. This would make it easier for the private sector to assess the deviations from policy rules. In this regard, it is interesting that Bill Poole’s (2006) recent analysis of the Fed’s policy rule could not go beyond 2001, because the data on the GDP gap were not released beyond that date. What about the Federal Reserve formally announcing numerical inflation targets as other central banks have done? I have suggested moving slowly in this direction because a sudden change could be misunderstood, and because policy has worked well for two decades with a more informal inflation target. A further lengthening of the inflation forecast horizon for the Monetary Policy Report would be an example of a more gradual change and would be a good step in my view. I have been concerned that placing more emphasis on a numerical inflation target could take emphasis away from predictability in setting the instruments. From the perspective of a policy rule approach, publishing one part of the rule— the inflation target—and not publishing other parts—the reaction coefficients—would create an asymmetry in a direction away from the regular reactions of the instruments that I have stressed in these remarks. Perhaps there is a way to prevent creating such an asymmetry. For example, the possibility of a joint announcement might be considered, perhaps both a target range for the inflation rate, from 1.5 to 2.5 percent, and a target range for the reaction coefficient of the interest rate to the inflation rate, from 1.5 to 2.5 percent, but there are many other possibilities. CONCLUSION In these remarks I have tried to convince you of the importance of being predictable in monetary F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W policy, building on Bill Poole’s paper written nearly four decades ago and on more recent experience with different degrees of predictability in practice. One of the key points, which needs much more research, is how the private sector and other public sector institutions develop rules of thumb that are based, perhaps unknowingly, on the systematic rule-like behavior of the monetary authorities. These private sector rules of thumb can improve the operation of the economy, but they can be broken in unanticipated and disruptive ways if policy becomes less predictable even for a short time and even if policymakers make their very best efforts to explain why. REFERENCES Jaronciński, Marek and Smets, Frank R. “House Prices and the Stance of Policy.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 339-65. Poole, William. “Optimal Choice of Monetary Policy Instruments in a Simple Stochastic Macro Model.” Quarterly Journal of Economics, May 1970, 84(2), pp. 197-216. Poole, William. “Monetary Policy Rules?” Federal Reserve Bank of St. Louis Review, March/April 1999, 81(2), pp. 3-12. Poole, William. “Fed Transparency: How Not Whether?” Federal Reserve Bank of St. Louis Review, November/December 2003, 85(6), pp. 1-8. Poole, William. “FOMC Transparency.” Federal Reserve Bank of St. Louis Review, January/February 2005a, 87(1), pp. 1-9. Poole, William. “How Predictable Is Fed Policy?” Federal Reserve Bank of St. Louis Review, November/December 2005b, 87(6), pp. 659-68. Poole, William. “The Fed’s Monetary Policy Rule.” Federal Reserve Bank of St. Louis Review, January/ February 2006, 88(1), pp. 1-12. Poole, William. “Milton and Money Stock Control.” Presented at the Milton Friedman luncheon, co- J U LY / A U G U S T 2008 409 Panel Discussion sponsored by the University of Missouri–Columbia Department of Economics, the Economic and Policy Analysis Research Center, and the Show-Me Institute; Columbia, MO, July 31, 2007. Taylor, John B. “Housing and Monetary Policy.” Panel discussion at the Federal Reserve Bank of Kansas City Symposium Housing, Housing Finance, and Monetary Policy, Jackson Hole, WY, September 1, 2007. Poole, William and Rasche, Robert H. “The Impact of Changes in FOMC Disclosure Practices on the Transparency of Monetary Policy: Are Markets and the FOMC Better ‘Synched’?” Federal Reserve Bank of St. Louis Review, January/February 2003, 85(1), pp. 1-10. Monetary Policy Under Uncertainty Ben S. Bernanke B ill Poole’s career in the Federal Reserve System spans two decades separated by a quarter of a century. From 1964 to 1974, Bill was an economist on the staff of the Board’s Division of Research and Statistics. He then left to join the economics faculty at Brown University, where he stayed for nearly 25 years. Bill rejoined the Fed in 1998 as president of the Federal Reserve Bank of St. Louis, so he is now approaching the completion of his second decade in the System. As it happens, each of Bill’s two decades in the System was a time of considerable research and analysis on the issue of how economic uncertainty affects the making of monetary policy, a topic on which Bill has written and spoken many times. I would like to compare the state of knowledge on this topic during Bill’s first decade in the System with what we have learned during his most recent decade of service. The exercise is interesting in its own right and has the added benefit of giving me the opportunity to highlight Bill’s seminal contributions in this line of research. DEVELOPMENTS DURING THE FIRST PERIOD: 1964-74 In 1964, when Bill began his first stint in the Federal Reserve System, policymakers and researchers were becoming increasingly confident in the ability of monetary and fiscal policy to smooth the business cycle. From the traditional Keynesian perspective, which was the dominant viewpoint of the time, monetary policy faced a long-term tradeoff between inflation and unemployment that it could exploit to keep unemployment low over an indefinitely long period at an acceptable cost in terms of inflation. Moreover, improvements in econometric modeling and the importation of optimal-control methods from engineering were seen as having the potential to tame the business cycle. Of course, the prevailing optimism had its dissenters, notably Milton Friedman. Friedman believed that the inherent complexity of the economy, the long and variable lags with which monetary policy operates, and the political and bureaucratic influences on central bank decisionmaking precluded policy from fine-tuning the level of economic activity. Friedman advocated the use of simple prescriptions for monetary policy— such as the k percent money growth rule—which Ben S. Bernanke is Chairman of the Board of Governors of the Federal Reserve System. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 410-15. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. 410 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Panel Discussion he felt would work reasonably well on average while avoiding the pitfalls of attempting to finetune the economy in the face of pervasive uncertainty (Friedman, 1968). Other economists were more optimistic than Friedman about the potential benefits of activist policies. Nevertheless, they recognized that the fundamental economic uncertainties faced by policymakers are a first-order problem and that improving the conduct of policy would require facing that problem head on. During this decade, those researchers as well as sympathetic policymakers focused especially on three areas of economic uncertainty: the current state of the economy, the structure of the economy (including the transmission mechanism of monetary policy), and the way in which private agents form expectations about future economic developments and policy actions. Uncertainty about the current state of the economy is a chronic problem for policymakers. At best, official data represent incomplete snapshots of various aspects of the economy, and even then they may be released with a substantial lag and be revised later. Apart from issues of measurement, policymakers face enormous challenges in determining the sources of variation in the data. For example, a given change in output could be the result of a change in aggregate demand, in aggregate supply, or in some combination of the two. As most of my listeners know, Bill Poole tackled these issues in a landmark 1970 paper, which examined how uncertainty about the state of the economy affects the choice of the operating instrument for monetary policy (Poole, 1970). In the simplest version of his model, Bill assumed that the central bank could choose to specify its monetary policy actions in terms of a particular level of a monetary aggregate or a particular value of a short-term nominal interest rate. If the central bank has only partial information about disturbances to money demand and to aggregate demand, Bill showed that the optimal choice of policy instrument depends on the relative variances of the two types of shocks. In particular, using the interest rate as the policy instrument is the better choice when aggregate demand is relaF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W tively stable but money demand is unstable, with money growth being the preferable policy instrument in the opposite case. Bill was also a pioneer in formulating simple feedback rules that established a middle ground between the mechanical approach advocated by Friedman and the highly complex prescriptions of optimal-control methods. For example, Bill wrote a Federal Reserve staff paper titled “Rulesof-Thumb for Guiding Monetary Policy” (Poole, 1971). Because his econometric analysis of the available data indicated that money demand was more stable than aggregate demand, Bill formulated a simple rule that adjusted the money growth rate in response to the observed unemployment rate. Bill was also practical in noting the pitfalls of mechanical adherence to any particular policy rule; in this study, for example, he emphasized that the proposed rule was not intended “to be followed to the last decimal place or as one that is good for all time [but]…as a guide—or as a benchmark—against which current policy may be judged” (p. 152). Uncertainty about the structure of the economy also received attention during that decade. For example, in his elegant 1967 paper, Bill Brainard showed that uncertainty about the effect of policy on the economy may imply that policy should respond more cautiously to shocks than would be the case if this uncertainty did not exist. Brainard’s analysis has often been cited as providing a theoretical basis for the gradual adjustment of policy rates of most central banks. Alan Blinder has written that the Brainard result was “never far from my mind when I occupied the Vice Chairman’s office at the Federal Reserve. In my view…a little stodginess at the central bank is entirely appropriate” (Blinder, 1998, p. 12). A key source of uncertainty became evident in the late 1960s and 1970s as a result of highly contentious debates about the formation of expectations by households and firms. Friedman (1968) and Ned Phelps (1969) were the first to highlight the central importance of expectations formation, arguing that the private sector’s expectations adjust in response to monetary policy and therefore preclude any long-run tradeoff between unemployment and inflation. However, Friedman J U LY / A U G U S T 2008 411 Panel Discussion and Phelps retained the view that monetary policy could exert substantial effects on the real economy over the short to medium run. In contrast, Robert Lucas and others reached more dramatic conclusions, arguing that only unpredictable movements in monetary policy can affect the real economy and concluding that policy has no capacity to smooth the business cycle (Lucas, 1972; Sargent and Wallace, 1975). Although these studies highlighted the centrality of inflation expectations for the analysis of monetary policy, the profession did not succeed in reaching any consensus about how those expectations evolve, especially in an environment of ongoing structural change. DEVELOPMENTS DURING THE SECOND PERIOD: 1998-2007 Research during the past 10 years has been very fruitful in expanding the profession’s understanding of the implications of uncertainty for the design and conduct of monetary policy. On the issue of uncertainty about the state of the economy, Bill’s work continues to provide fundamental insights regarding the choice of policy instrument. Money-demand relationships were relatively stable through the 1950s and 1960s, but, in the wake of dramatic innovations in banking and financial markets, short-term money-demand relationships became less predictable, at least in the United States. As a result, consistent with the policy implication of Bill’s 1970 model, the Federal Reserve (like most other central banks) today uses the overnight interbank rate as the principal operating target of monetary policy. Bill’s research also raised the possibility of specifying the operating target in other ways, for example, as an index of monetary or financial conditions; and it provided a framework for evaluating the usefulness of intermediate targets— such as core inflation or the growth of broad money—that are only indirectly controlled by policy. More generally, the task of assessing the current state of the economy remains a formidable challenge. Indeed, our appreciation of that chal412 J U LY / A U G U S T 2008 lenge has been enhanced by recent research using real-time data sets.1 For example, Athanasios Orphanides has shown that making such real-time assessments of the sustainable levels of economic activity and employment is considerably more difficult than estimating those levels retrospectively. His 2002 study of U.S. monetary policy in the 1970s shows how mismeasurement of the sustainable level of economic activity can lead to serious policy mistakes. On a more positive note, economists have made substantial progress over the past decade in developing new econometric methods for summarizing the information about the current state of the economy contained in a wide array of economic and financial market indicators (Svensson and Woodford, 2003). Dynamic-factor models, for example, provide a systematic approach to extracting information from real-time data at very high frequencies. These approaches have the potential to usefully supplement more informal observation and human judgment (Stock and Watson, 2002; Bernanke and Boivin, 2003; and Giannone, Reichlin, and Small, 2005). The past decade has also witnessed significant progress in analyzing the policy implications of uncertainty regarding the structure of the economy. New work addresses not only uncertainty about the values of specific parameters in a given model of the economy but also uncertainty about which of several competing models provides the best description of reality. Some research has attacked those problems using Bayesian optimalcontrol methods (Brock, Durlauf, and West, 2003). The approach requires the specification of an explicit objective function as well as of the investigator’s prior probabilities over the set of plausible models and parameter values. The Bayesian approach provides a useful benchmark for policy in an environment of well-defined sources of uncertainty about the structure of the economy, and the resulting policy prescriptions give relatively greater weight to outcomes that have a higher probability of being realized. In contrast, other researchers, such as Lars Hansen and Thomas Sargent (2007), have developed robust1 A recent example is Faust and Wright (2007). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Panel Discussion control methods—adapted from the engineering literature—that are aimed at minimizing the consequences of worst-case scenarios, including those with only a low probability of being realized. An important practical implication of all this recent literature is that Brainard’s attenuation principle may not always hold. For example, when the degree of structural inertia in the inflation process is uncertain, the optimal Bayesian policy tends to involve a more pronounced response to shocks than would be the case in the absence of uncertainty (Söderstrom, 2002). The concern about worst-case scenarios emphasized by the robustcontrol approach may likewise lead to amplification rather than attenuation in the response of the optimal policy to shocks (Giannoni, 2002; Onatski and Stock, 2002; and Tetlow and von zur Muehlen, 2001). Indeed, intuition suggests that stronger action by the central bank may be warranted to prevent particularly costly outcomes. Although Bayesian and robust-control methods provide insights into the nature of optimal policy, the corresponding policy recommendations can be complex and sensitive to the set of economic models being considered. A promising alternative approach—reminiscent of the work that Bill Poole did in the 1960s—focuses on simple policy rules, such as the one proposed by John Taylor, and compares the performance of alternative rules across a range of possible models and sets of parameter values (Levin, Wieland, and Williams, 1999 and 2003). That approach is motivated by the notion that the perfect should not be the enemy of the good; rather than trying to find policies that are optimal in the context of specific models, the central bank may be better served by adopting simple and predictable policies that produce reasonably good results in a variety of circumstances. Given the centrality of inflation expectations for the design of monetary policy, a key development over the past decade has been the burgeoning literature on the formation of these expectations in the absence of full knowledge of the underlying structure of the economy.2 For example, considerations of how the public learns about the econ2 See Bernanke (2007) and the references therein. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W omy and the objectives of the central bank can affect the form of the optimal monetary policy (Gaspar, Smets, and Vestin, 2006; and Orphanides and Williams, 2007). Furthermore, when the public is unsure about the central bank’s objectives, even greater benefits may accompany achieving a stable inflation rate, as doing so may help anchor the public’s inflation expectations. These studies also show why central bank communications is a key component of monetary policy; in a world of uncertainty, informing the public about the central bank’s objectives, plans, and outlook can affect behavior and macroeconomic outcomes (Bernanke, 2004; and Orphanides and Williams, 2005). CONCLUSION Uncertainty—about the state of the economy, the economy’s structure, and the inferences that the public will draw from policy actions or economic developments—is a pervasive feature of monetary policymaking. The contributions of Bill Poole have helped refine our understanding of how to conduct policy in an uncertain environment. Notably, we now appreciate that policy decisions under uncertainty must take into account a range of possible scenarios about the state or structure of the economy, and those policy decisions may look quite different from those that would be optimal under certainty. For example, policy actions may be attenuated or augmented relative to the “no-uncertainty benchmark,” depending on one’s judgments about the possible outcomes and the costs associated with those outcomes. The fact that the public is uncertain about and must learn about the economy and policy provides a reason for the central bank to strive for predictability and transparency, avoid overreacting to current economic information, and recognize the challenges of making real-time assessments of the sustainable level of real economic activity and employment. Most fundamentally, our discussions of the pervasive uncertainty that we face as policymakers is a powerful reminder of the need for humility about our ability to forecast and manage the future course of the economy. J U LY / A U G U S T 2008 413 Panel Discussion REFERENCES Bernanke, Ben S. “Fedspeak.” Presented at the Meetings of the American Economic Association, San Diego, January 3, 2004; www.federalreserve.gov/ boarddocs/speeches/2004/200401032/default.htm. Bernanke, Ben S. “Inflation Expectations and Inflation Forecasting.” Presented at the Monetary Economics Workshop of the National Bureau of Economic Research Summer Institute, Cambridge, MA, July 10, 2007; www.federalreserve.gov/newsevents/speech/ bernanke20070710a.htm. Bernanke, Ben S. and Boivin, Jean. “Monetary Policy in a Data-Rich Environment.” Journal of Monetary Economics, April 2003, 0(3), pp. 525-46. Blinder, Alan S. Central Banking in Theory and Practice. Cambridge, MA: MIT Press, 1998. Brainard, William C. “Uncertainty and the Effectiveness of Policy.” American Economic Review, May 1967, 57(2), pp. 411-25. Brock, William A.; Durlauf, Steven N. and West, Kenneth D. “Policy Analysis in Uncertain Economic Environments.” Brookings Papers on Economic Activity, 2003, 1, pp. 235-322. Faust, Jon and Wright, Jonathan H. “Comparing Greenbook and Reduced Form Forecasts Using a Large Realtime Dataset.” Presented at the Federal Reserve Bank of Philadelphia conference RealTime Data Analysis and Methods in Economics, April 19-20, 2007; www.phil.frb.org/econ/conf/ rtconference2007/papers/Paper-Wright.pdf. Friedman, Milton. “The Role of Monetary Policy.” American Economic Review, March 1968, 58(1), pp. 1-17. Gaspar, Vitor; Smets, Frank and Vestin, David. “Adaptive Learning, Persistence, and Optimal Monetary Policy.” Journal of the European Economic Association, April-May 2006, 4(2/3), pp. 376-85. Giannone, Domenico; Reichlin, Lucrezia and Small, David. “Nowcasting GDP and Inflation: The RealTime Informational Content of Macroeconomic Data Releases.” Finance and Economics Discussion Series 414 J U LY / A U G U S T 2008 2005-42, Board of Governors of the Federal Reserve System, October 2005; www.federalreserve.gov/ pubs/feds/2005. Giannoni, Marc P. “Does Model Uncertainty Justify Caution? Robust Optimal Monetary Policy in a Forward-Looking Model.” Macroeconomic Dynamics, February 2002, 6(1), pp. 111-44. Hansen, Lars Peter and Sargent, Thomas J. Robustness. Princeton, NJ: Princeton University Press, 2007. Levin, Andrew; Wieland, Volker and Williams, John. “Robustness of Simple Monetary Policy Rules Under Model Uncertainty,” in John B. Taylor, ed., Monetary Policy Rules. Chicago: University of Chicago Press, 1999, pp. 263-99. Levin, Andrew; Wieland, Volker and Williams, John. “The Performance of Forecast-Based Monetary Policy Rules under Model Uncertainty.” American Economic Review, June 2003, 93(3), pp. 622-45. Lucas, Robert E. Jr. “Expectations and the Neutrality of Money.” Journal of Economic Theory, April 1972, 4(2), pp. 103-24. Onatski, Alexei and Stock, James H. “Robust Monetary Policy under Model Uncertainty in a Small Model of the U.S. Economy.” Macroeconomic Dynamics, February 2002, 6(1), pp. 85-110. Orphanides, Athanasios. “Monetary-Policy Rules and the Great Inflation.” American Economic Review, May 2002, 92(2), pp. 115-20. Orphanides, Athanasios and Williams, John C. “Inflation Scares and Forecast-Based Monetary Policy.” Review of Economic Dynamics, April 2005, 8(2), pp. 498-527. Orphanides, Athanasios and Williams, John C. “Robust Monetary Policy with Imperfect Knowledge.” Journal of Monetary Economics, July 2007, 54(5), pp. 1406-35. Phelps, Edmund S. “The New Microeconomics in Inflation and Employment Theory.” American Economic Review, May 1969, 59(2), pp. 147-60. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Panel Discussion Poole, William. “Optimal Choice of Monetary Policy Instruments in a Simple Stochastic Macro Model.” Quarterly Journal of Economics, May 1970, 84(2), pp. 197-216. Stock, James H. and Watson, Mark W. “Forecasting Using Principal Components from a Large Number of Predictors.” Journal of the American Statistical Association, December 2002, 97(460), pp. 1167-79. Poole, William. “Rules-of-Thumb for Guiding Monetary Policy,” in Open Market Policies and Operating Procedures: Staff Studies. Washington, DC: Board of Governors of the Federal Reserve System, 1971, pp. 135-89. Svensson, Lars E.O. and Woodford, Michael. “Indicator Variables for Optimal Policy.” Journal of Monetary Economics, April 2003, 50(2), pp. 691-720. Sargent, Thomas J. and Wallace, Neil. “‘Rational Expectations,’ the Optimal Monetary Instrument, and the Optimal Money Supply Rule.” Journal of Political Economy, April 1975, 83(2), pp. 241-54. Tetlow, Robert J. and von zur Muehlen, Peter. “Robust Monetary Policy with Misspecified Models: Does Model Uncertainty Always Call for Attenuated Policy?” Journal of Economic Dynamics and Control, June/July 2001, 25(6/7), pp. 911-49. Söderstrom, Ulf. “Monetary Policy with Uncertain Parameters.” Scandinavian Journal of Economics, February 2002, 104(1), pp. 125-45. The Importance of Being Predictable William Poole T his has been an absolutely wonderful occasion for me. I deeply appreciate all those who have come: friends that I’ve known from way, way back, newer friends recently formed. And I am very gratified that Ben Bernanke and John Taylor joined on the panel. I especially want to thank, above all, Bob Rasche and the Research Division here, both for organizing and executing this event— but even more than that for the support that I’ve gotten and the intellectual excitement over my almost 10 years here. We’ve really worked together in a very collegial way. It’s going to be hard to imagine being productive without hav- ing a staff like that behind me. They have been coauthors, really—staff is really the wrong way to put it—coauthors on the speeches, some of which have been published in the Federal Reserve Bank of St. Louis Review. Well, nostalgia takes you only so far. And, so I want to talk about business, if you will, going back to some of the earlier literature. How we got to where we are today does help to inform us about some very important current issues. I was fascinated—totally unexpected—that my obscure, 1971 paper would become a centerpiece of some of the discussion. It’s interesting to reflect on that because the times were so different. When I was working on that paper, the policy of the Federal Reserve was sort of unspecified. It was calculated meeting by meeting. And what struck me was that there are (at least the way I looked at it with my Chicago background) some powerful business William Poole was the president of the Federal Reserve Bank of St. Louis at the time this conference was held. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 415-19. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 415 Panel Discussion cycle regularities—just the very crude sort of thing that Friedman and Schwartz demonstrated, the patterns of money growth and interest rates over the business cycle. And so my idea was that we’ve got to find a way to avoid making exactly the same mistake over and over again. When you go back and look at the business cycles in the 1920s and 50s and 60s, it just looked like the same mistake over and over again. So there had to be some way of formalizing something as a—call it a rule of thumb—a baseline, and say we should depart from some baseline behavior only if we had some pretty good reason for doing so. Otherwise, we’re just going to be making the same mistake over and over again. And, in fact, we did make the same mistake, in a business cycle sense, several times more after that. So that was the origin of that paper; it was nothing more complicated than that. There was one other piece of it. In this era there was tremendous disagreement between those who viewed policy in the context of setting a monetary aggregate (the Chicago background that I had) and those who looked at policy entirely through an interest rate filter. And, of course, the origin of my 1970 paper was an attempt to make sense of those different views. You could not make sense of it in a deterministic model. It had to have something to do (if there was anything valid in this debate) with the uncertainty in the model, the nature of the disturbances. And that was the origin of my 1970 paper. And the origin of this sort of combination money growth/interest rate rule that was discussed earlier here at this conference was really an effort to try to bridge the gap between these two very different schools of thought and how they approached monetary policy. And obviously John Taylor did a much, much better job with that later on. Now, in the discussion of the Svensson and Williams (2008) paper at this conference, which I had not seen before, there was something that sort of rubbed me the wrong way and I couldn’t put my finger on it right away. I raised the issue about the model’s assumption about central bank behavior, the assumption of the state of knowledge in the private sector. And the answer was the model assumes complete knowledge of the 416 J U LY / A U G U S T 2008 central bank. As I reflect on that, that’s equivalent to saying that the central bank has permanent credibility—no one will ever doubt what the central bank is going to do. Put another way, that everyone knows exactly what the central bank is going to do. And I just don’t believe that’s a valid assumption. I think credibility has been very costly to create among central banks around the world, and I think it’s a terrible mistake to take it for granted. Credibility is potentially very fragile; indeed, one of the central things that we need to pay attention to is how to maintain credibility. And the way in which you maintain credibility is a very important part of a rules-based monetary policy. An important part of maintaining credibility is to say what you are going to do and then do it. The central bank does what it said it would do unless it has a very good explanation for why it departs from what it said it was going to do. Of course, we’ve had important institutional developments here with central bank independence, and, really, I think we’ve strengthened independence in the Federal Reserve although the law hasn’t changed very much—strengthened independence in a practical sense. And that’s been very important. But we should always keep in mind that the central bank is a political institution established by law or by treaty—by laws that can be changed. But even more than that: John Taylor’s served in the government, I’ve served at the Council of Economic Advisers (CEA); any of you who’ve been there—Murray Weidenbaum, who recruited me to the CEA—anybody who’s worked in the government knows that there are all sorts of things that are done around the edges of the law, behind the scenes, that are not exactly in line with what the law might call for. And there’s a natural view, which I think is correct, to be suspicious because central banks in the past have not always been immune from behavior that is secret or around the edges of the law. So, to maintain the confidence that people need to have in the central bank, you need to do things with a great deal of careful planning and you have to maintain a very high level of integrity; you have to have people there who can be trusted not to be a part of the political process. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Panel Discussion That problem is not going to go away. We live in a vigorously democratic society where people try to use government for various purposes that are not always in the national interest. And central banks are inevitably going to remain part of the political system, as they should. But we need to maintain the highest possible integrity in order to maintain confidence. And that’s what I think bothers me about the Svensson and Williams paper: It misses a critical component. To me, one of the ways in which you could lose confidence is if you started to run experiments. I can’t imagine having to write a press statement or to give a speech to explain why it was that we conducted an experiment that had predictable (and I mean predictable) consequences of a recession or mild recession to the purpose of learning more about some parameter. It’s hard to imagine anything that we might do that would be more damaging to long-run credibility. I’ll describe one of the things that happened to me when I came to St. Louis, talking a little bit about my journey here, having been an academic for most of my professional career. When the St. Louis Fed’s board of directors recruited me, John McDonnell was the chairman of the board; he said, “Bill you have to understand, you are not going to be able to do any research in this job.” Well, it hasn’t turned out that way. And, in fact, I think that with my research productivity jointly with the economists here (they’ve done all the hard work), I’ve accomplished more in these 10 years than I had in the previous 20. The research started as a consequence, really, of dealing with issues that I needed to understand as a part of doing my job. But I didn’t know where I could turn in the journal literature with which I was familiar to get any help with any of these issues. One of the very first issues was this question: If I am going to give a speech, what am I going to write, what am I going to talk about? And then how to deal with press contacts and the Q&A with press coverage and so forth. So I started to think abstractly about the whole process of central bank communication, and I went back to what I would regard as the two first principles that come out of the rational expectations literature. One is that the private sector needs F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W to know what the central bank is doing. And you can’t have a good equilibrium if the private sector doesn’t know what the central bank is doing. A part of that requires that the central bank itself knows what it is doing. That’s the place to start. So, anyone who has been in a classroom, and most of you here probably have, know that you hone your own ideas and develop a great deal of clarity when you are forced to actually stand up and talk about them. And part of the effort to understand on a more systematic basis what sort of policy adjustments we should make comes from pinning down the fundamental nature of the policy rule. And John, particularly, of course, has led the way on that. So that’s part of the process, and the private sector learns about what the central bank is doing in good part just from observing what it is doing and trying to put some system into that, which you in principle can extract from what’s done without any words on our part. But there are certainly lots of cases where this signal extraction might go a lot more smoothly and a lot more quickly if the central bankers would actually talk intelligently about what they are doing. Think about the context of a simple learning model, for example. Suppose that the central bank has been operating on some value of a parameter and then decides for whatever reason that it wants to have a different value of that parameter. Well, it might take someone with Jim Hamilton’s skills to generate an enormous number of observations to actually discover from central bank behavior that the parameter has changed. And in the meantime, that means that the private sector is operating under a different understanding of that parameter than the central bank is operating under. That produces expectational problems in terms of the equilibrium and efficiency. If we could explain what we are doing, why this parameter has changed, we ought to be able to move that equilibrium to the correct point much more quickly. So that’s part of the task of central bank talk: Explain what we’re doing and why were doing it, to help promote a good equilibrium between the central bank and the private sector—an equilibrium in which the meshing of knowledge is really critical to the efficiency of the outcome. J U LY / A U G U S T 2008 417 Panel Discussion But it seems to me that there is a second principle that’s extremely important from the rational expectations literature: The central bank ought not to be purveyors of random disturbance. We ought not to add random noise to the system either in terms of the actions we take or in terms of what we say. So you’re standing up in front of an audience and you’ve got these two things you’re struggling with: first, trying to convey genuine information, and second, trying not to say something that causes a market disturbance that is decidedly not helpful. Some of the press people might love it, but it’s not what I ought to be doing. Now, when I came, I had no professional guidance in any of the economics literature about how to do this. I knew what the basic principles were. But what do you actually do when you are standing up in front of an audience? I had no guidance whatsoever. Probably I didn’t read enough memoirs; I don’t know. But I don’t think that people generally talk about this kind of thing in their memoirs, either. So, I started to think a lot about the communications process, and I know that one approach that some people take is that, they’re so worried about the second problem, they give up on the first and so they really don’t say much of anything. That didn’t seem to me to be satisfactory, because I thought the first principle of trying to produce a better understanding in the marketplace of what the central bank is doing was really an important responsibility of my office. Another issue is that behavior of the markets is obviously driven by active and by-and-large pretty well-informed market participants, not primarily by Main Street. And a lot of what we do out here in the Reserve Banks is to wander around the Districts or, more broadly speaking, to audiences of all sorts, with different backgrounds and degrees of expertise. One of the communications challenges is to be able to give a speech that says something to well-informed people and at the same time doesn’t pass completely over the heads of people who are not so well informed. Of course, that puts a lot of constraints on what you say, but also how you say it. But why would we care about Main Street? Well, one of the very important reasons is that 418 J U LY / A U G U S T 2008 this is the business we’re in. The monetary policy business that we’re in is designed to improve the welfare, and maintain a high degree of welfare, for all the citizens. For Main Street as well as Wall Street. We need to talk to bankers and traders and portfolio managers, but we also need to talk to Main Street because these are our constituents. At any moment in time, all the time, the interests of various people in the markets are in conflict: Some people are long, some people are short, some people have short-term investments, some long-term investments, some equity, some bonds, and so forth. There are a lot of different interests, and it is extremely important that we serve the “general interest.” I think that what that means is that we have broad macroeconomic objectives that we can summarize quite well in talking about the dual mandate: maintaining the stable purchasing power of the currency and reducing fluctuations in GDP and employment from equilibrium paths. If we are successful with these explanations, we will have done 99.9 percent of what we can do and what we ought to do. Another important reason for talking to Main Street can be illustrated by a story from when I was at the CEA. That was a difficult period, in the early 1980s, and there was a lot of commentary on the part of Congress and to some extent the administration about Federal Reserve policy. Knowing a lot about the Fed, I was trying to explain to people that pushing the Fed was counterproductive from the point of view of the interests of the politicians themselves. I remember that a senator, who often made comments about the Fed, wanted lower interest rates. (By the way, you may have seen the comment that Alan Greenspan made, in one of his interviews, that not once while he was in office did he ever get a phone call or a letter from a politician recommending higher interest rates. Not once. There is an asymmetry here.) So, a lot of the politicians are not all that well informed about monetary policy, and I remember going up to Capitol Hill and talking to a very prominent senator and saying, “You have to understand that the Fed values its independence and it is extremely important that the Fed not appear to be responding to the entreaties of politicians. And, therefore, if you F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Panel Discussion want interest rates lower, you will not get the result that you want by blasting the Fed, because the Fed can’t respond to that blast. It’s not in your interest.” He said, “I understand that, but it plays well on Main Street.” It was very simple. Very simple. So, one of the reasons to talk to Main Street is to help people understand why that position is wrong. I hope that people will develop a tin ear to language like that that comes from all sorts of different directions. Another issue that I’ve been quite concerned about during my time in St. Louis has been trade issues. I have made a good number of speeches where I’ve talked about trade and capital flows, the importance of world markets, and trying to resist—I don’t even like to use the word “protectionism,” because it is good to try to protect people—the kind of economic isolationism that many of these policies encourage. So, we need to talk to Main Street as well as the monetary experts, and that makes the communications issues both challenging and very, very interesting. Very interesting. I do want to say one other thing. John referred to a lecture that I gave earlier this year. He did not mention that this was an event on Milton Friedman’s birthday. It was out at the University of Missouri, and a point that I remember vividly from those days in Chicago, and a point that I think has tremendous importance today, is that Milton always argued—and Brunner and Meltzer, and others, but Milton sort of led this analysis— that one of the great advantages of a monetary aggregates rule is that it allows maximum scope for the market to respond to disturbances and F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W move interest rates in a way that will be stabilizing. Built-in stability is very important. Of course, there is a huge literature about the built-in stabilizers in the fiscal policy area. The current policy stance has the advantages of high credibility, wellanchored inflation expectations, and the possibility of understanding in a more formal way with several decoupled interest rates in the model of allowing the market to do a great deal of the stabilization work. That has enormous advantages in producing efficient results. It allows the Federal Reserve at many critical times to sit back and watch until the situation is clearer in terms of the arriving evidence. And you can go through lots of recent cases where there have been very substantial fluctuations in long-term interest rates that do an enormous amount of the stabilization work for us. That provides great clarity in the stance of policy and at the same time is a framework that produces a tremendous amount of built-in stabilization. REFERENCES Poole, William. “Optimal Choice of Monetary Policy Instruments in a Simple Stochastic Macro Model.” Quarterly Journal of Economics, May 1970, 84(2), pp. 197-216. Poole, William. “Rules-of-Thumb for Guiding Monetary Policy,” in Open Market Policies and Operating Procedures: Staff Studies. Washington, DC: Board of Governors of the Federal Reserve System, 1971, pp. 135-89. J U LY / A U G U S T 2008 419 420 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Announcements and the Role of Policy Guidance Carl E. Walsh By providing guidance about future economic developments, central banks can affect private sector expectations and decisions. This can improve welfare by reducing private sector forecast errors, but it can also magnify the impact of noise in central bank forecasts. I employ a model of heterogeneous information to compare outcomes under opaque and transparent monetary policies. While better central bank information is always welfare improving, more central bank information may not be. (JEL E52, E58) Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 421-42. S tandard models used for monetary policy analysis typically assume that households and firms and the central bank share a common information set and economic model, yet actual policy decisions are taken in an environment in which heterogeneous information is the norm and many alternative models coexist. The resulting heterogeneity in views can play an important role in affecting both policy choices and the monetary transmission process. Transparency in the conduct of policy can help to reduce heterogeneous information. Inflation-targeting central banks, for example, make significant attempts to reduce uncertainty about policy objectives, such as through the release of detailed inflation and output projections, to ensure the public shares central bank information about future economy developments. By being transparent about its objectives and its outlook for the economy, central banks help provide the public with guidance about the future. But providing guidance carries risks. As Poole (2005, p. 6) has expressed it, “[F]or me the issue is whether under normal and routine circumstances forward guidance will convey information or whether it will create additional uncertainty.” Because any forecast released by the central bank is subject to error, being more transparent may simply lead the private sector to react to what was, in retrospect, noise in the forecast. The possibility that the private sector may overreact to central bank announcements does capture a concern expressed by some policymakers. For example, in discussing the release of Federal Open Market Committee (FOMC) minutes, Janet Yellen expressed the view that “Financial markets could misinterpret and overreact to the minutes” (Yellen, 2005, p. 1). In this paper, I explore the role of economic transparency—specifically, transparency about the central bank’s assessment of future economic conditions—in altering the effectiveness of monetary policy. I do so in a framework in which central bank projections may convey useful information but may also introduce inefficient fluctuations into the economy. A focus on economic transparency seems appropriate for understanding the issues facing many central banks. The recent concerns about Carl E. Walsh is a professor of economics at the University of California, Santa Cruz, and a visiting scholar at the Federal Reserve Bank of San Francisco. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 421 Walsh the implications of the subprime mortgage market reflect, in part, private sector uncertainty about the Fed’s view of the economic outlook and the way the outlook for inflation and real economic activity may be affected by financial market conditions. Throughout 2007, for example, many financial market participants appeared to hold more pessimistic views than the Federal Reserve about future economic developments;1 and in recent months, market participants have often expected significant interest rate cuts, while some members of the FOMC have emphasized concerns about the outlook for inflation, suggesting they saw less need for rate reductions. News reports speculating on possible interest rate cuts by the Fed or the European Central Bank focused very little on uncertainty about central bank preferences but a great deal on the uncertainty about the outlook for the economy. These reports reveal heterogeneity among private forecasters and uncertainty about the Fed’s (or the European Central Bank’s) outlook for the economy. And public statements by central bankers were designed to communicate their views on future economic developments. Jean-Claude Trichet’s statement that the markets “have gone progressively back to normal” (Atkins, Mackenzie, and Davies, 2007, p. 1) and Ben Bernanke’s (2007) comment that housing remains a “significant drag” on the economy, both exemplify how central bankers signal their assessment of economic conditions, and this assessment is one factor that influences the (heterogeneous) outlooks among members of the private sector. The uncertainty in financial markets in recent months illustrates clearly the significant differences that can arise between the central bank and private market participants. This is a classic example of heterogenous information about the economy. Much of the debate has been focused on the question of future interest rate cuts, but the underlying issues appear to be related to differing views among private forecasters and between private 1 “Even as Wall Street analysts ratchet up their worries about a recession, Fed officials are far from convinced that a true downtown is likely” (Andrews, 2007). A more vivid example of disagreement was provided by CNBC commentator Jim Cramer, whose blast that the Fed is clueless about “how bad it is out there” was reportedly seen by more than a million viewers on YouTube. 422 J U LY / A U G U S T 2008 forecasters and the Fed over the likely impact of financial market disturbances on the real economy and the likelihood of a future recession. The next section discusses the two goals of transparency Bill Poole (2005) has stressed— accountability and policy effectiveness. The third section develops a model of asymmetric and heterogeneous economic information that can be used to model the implications of transparency. Two policy regimes are considered. In the first, the public observes the policy instrument of the central bank but the central bank provides no further information to the public. In the second, the central bank provides information on its outlook for future economic developments. The welfare implications of these regimes are discussed in the fourth section. Within each regime, better quality central bank information is always welfare improving (the pro-transparency aspect of Morris and Shin, 2002, emphasized by Svensson, 2006). However, across regimes, more central bank information has ambiguous effects. THE GOALS OF TRANSPARENCY Transparency requires asymmetric information, but the nature of this asymmetry can take many forms. In fact, Geraats (2002) has classified five types of transparency—political, procedural, economic, policy, and operational. Briefly, these correspond to central bank transparency about objectives, the internal decisionmaking process, forecasts and models, policy actions, and instrument setting and control errors. Each of these dimensions of transparency is important and has been studied extensively (see Geraats, 2002, for a survey). In recent years, central banks have become more transparent along all these dimensions, and levels of transparency that would have been viewed as exceptional 20 years ago are today accepted as best practice among modern central banks.2 The trend toward independent central 2 See Eijffinger and Geraats (2006) and Dincer and Eichengreen (2007) for indices of central bank transparency. Cukierman (2006) discusses some of the factors that might place limits on how transparent central banks should (or can) be. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh banks with explicit mandates assigned to them and the widespread adoption of inflation targeting has contributed greatly to political transparency. The Bank of England is among the most procedurally transparent central banks, publishing minutes and individual votes of its Monetary Policy Committee discussions. Central banks, such as the Federal Reserve, that were formerly reluctant to communicate policy actions directly now do so clearly, timely, and directly. The most transparent central banks, such as the Reserve Bank of New Zealand and the Bank of Norway, publish their projections for the policy interest rate. The use of a short-term interest rate as the policy instrument has greatly enhanced operational transparency. But although most central banks today are transparent about their policy stance and operational procedures—something hard to avoid when the policy instrument is a short-term market interest rate—there is much greater variation in the extent to which central banks are transparent about their decisionmaking process, their internal forecasts, and their policy objectives. But what is the point of being transparent? As noted earlier, Poole (2006) has articulated two goals of transparency: to meet the Fed’s “responsibility to be politically accountable” and “to make monetary policy more effective.” The next two subsections discuss each of these goals. Transparency and Accountability The role transparency plays in supporting accountability can differ depending on whether the ultimate objectives of monetary policy are observable or unobservable. Consider first the case in which the objectives of monetary policy are, ex post, clearly measurable and observable. For concreteness, assume inflation is the only objective of the central bank and there is agreement on the appropriate measure of inflation that the central bank should control. In this environment, it is in principle straightforward to ensure accountability. Observing the ex post rate of inflation would seem to provide a simple means for judging the performance of the central bank. However, even under the conditions specified (a single measurable objective), the ex post realizaF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W tion of inflation is not a sufficient performance measure. The reason is that inflation is not directly controllable—even under an optimal policy (where the central bank is doing exactly what it should be doing), the realized inflation rate can differ from the desired value. This difference may be small, but as long as there is any random variation that is beyond the ability of the central bank to eliminate, public accountability based solely on inflation outcomes will punish some good central bankers and reward some lucky ones. Transparency can help promote accountability by allowing the public to base its evaluation of the central bank not just on observed inflation but on the information that was available to the central bank when it had to make its policy decision. Having access to internal bank forecasts, for example, allows outsiders to evaluate the decisions made by the central bank. This can mitigate some of the problems associated with evaluations based solely on realized inflation. Having access to the information on which decisions were based helps remove the influence of random uncontrollable events that affect inflation and therefore supports a better system of accountability.3 In general, however, policy objectives are not directly observable, and they may even be inherently unmeasurable. Certainly, recent theoretical models, which have emphasized the use of the welfare of the representative agent as the appropriate objective of policy, have defined optimal policy in terms of unmeasurable objectives. It is not clear that we could reach agreement on the correct way to measure welfare, as that depends on the specific model we believe characterizes the economy, even if we could agree on how to define welfare. It certainly is not observable. Transparency can be especially critical when objectives are unobserved. Assessing, or holding accountable, an economic agent when objectives are unobservable is not a situation unique to monetary policy and central banks. Education is perhaps the most prominent field in which public 3 As Tim Harford (2007, Part 2, p. 3) pointed out in a recent “Dear Economist” column in the Financial Times, it might seem sensible for a company to judge its ice cream sales force on total sales, but having information about the weather allows for a better assessment of the contribution of the sales team to actual sales. J U LY / A U G U S T 2008 423 Walsh policy must deal with this situation; the objectives are high quality education and teaching but there exists wide disagreement over how to define and measure these qualities. Because social welfare does depend on inflation and inflation can be observed, one might use inflation as a type of performance measure, holding the central bank accountable for achieving a low and stable inflation rate. Inflation targeting can be thought of as defining a performance measure for the central bank. The critical issue in choosing any performance measure, however, is how powerful one wants to make the incentives. If accountability is based strictly on realized inflation and the consequences of missing the target are large, then the central bank will naturally focus on achieving the target, even if this means sacrificing other, more difficult to measure, aspects of social welfare. The concern that inflationtargeting produces too much of a focus on inflation control is at the heart of most criticisms of inflation targeting in the United States. But this is where transparency becomes particularly important. Greater transparency can lessen the need to rely on a single easily measured performance indicator. When there is greater transparency, and the public is able to assess the same information the central bank has used to set policy, it is no longer necessary to base central bank accountability on inflation outcomes only (Walsh, 1999). Transparency and the Effectiveness of Monetary Policy Poole’s second goal of transparency, promoting policy effectiveness, requires that private sector decisions be influenced, and influenced systematically, by the information central banks provide. With the development of New Keynesian models and their emphasis on the importance of forwardlooking behavior, managing expectations to improve policy effectiveness has taken on a new importance. Woodford (2005) has gone so far as to state that “not only do expectations about policy matter, but, at least under current conditions, very little else matters.”4 The intuition for Woodford’s statement is straightforward. Policymakers control directly only a short-term interest rate. Yet rational agents are forward looking and so base their spending and pricing decisions on their assessment of future interest rates, not just current rates. The recognition that expectations matter is not confined to academics; a recent article in the Financial Times (Guha, 2007) states that “What really matters, both for the markets and the economy, is not the current policy rate but the expected path of future rates.” Transparency and its relationship to policy effectiveness played a key role in the large literature that focused on the average inflation rate bias that could arise under optimal discretionary policy. By and large, this literature emphasized political and operational transparency, and it employed models in which policy surprises were the source of the real effects of monetary policy. Geraats (2002) provides an excellent survey of the literature. In these models, the central bank preferences were generally treated as stochastic and unknown. The policy instrument was also taken to be observed with error or subject to a control error. For example, the central bank might control nonborrowed reserves, but this allowed only imperfect control of the money supply.5 Observing money growth would not provide enough information for the public to disentangle the effects of control errors from shifts in central bank preferences. Thus, there was opaqueness about political objectives and operational implementation. Transparency was typically modeled as a reduction in the noise in the signal on the policy instrument. The optimal degree of transparency ensured the public would learn quickly when the central bank preferences shifted, but still left open the possibility that the bank could create a surprise if one was needed to aid stability. Cukierman and Meltzer (1986) showed that the central bank may prefer to adopt a less efficient operating procedure than 5 4 Italics in the original. 424 J U LY / A U G U S T 2008 See, for example, Cukierman and Meltzer (1986) and Faust and Svensson (2002). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh is technically feasible (i.e., not reduce the control error variance to its minimum possible level).6 As emphasized in recent discussions of transparency, however, New Keynesian models imply that it is predictable monetary policies, not surprises, that are most effective in achieving policy goals. In such an environment, transparency, rather than reducing the efficacy of policy can actually increase it. Central bank announcements about future policy actions, or about future economic developments, can affect private sector expectations of future interest rates, inflation, and economic activity. With spending and pricing decisions dependent on these expectations, using announcements to influence expectations gives the central bank an additional policy instrument. As such, it serves to make policy more effective. The argument that transparency can increase the effectiveness of monetary policy is certainly more consistent with the modern practice of central banks, which has been uniformly to move in the direction of greater transparency. But providing information to the public may have potential costs. These costs are associated with the conditional nature of any forecast. Some economists have worried that the public will not understand the distinction between a conditional and an unconditional forecast.7 Particularly because reputation is important, deviating from a previously announced policy path may be interpreted as a deviation from a commitment equilibrium rather than as an appropriate response based on new information. If a central bank fails to raise interest rates after signaling that it planned to, the private sector may believe the bank has become less concerned about inflation, causing inflation expectations to rise. Financial market participants may underestimate the conditionality of the announced rate path and so view deviations as introducing unwarranted uncertainty into financial markets. These factors may make the central 6 7 See also Faust and Svensson (2002), who show that, when the choice of transparency is made under commitment, patient central banks with small inflation biases will prefer minimum transparency. They argue that this result might account for the (then) relatively low degree of transparency that characterized the U.S. Federal Reserve System. Goodhart (2006). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W bank reluctant to adjust rates, producing a lock-in effect that would reduce flexibility and limit policy effectiveness. Even when the public understands the conditional nature of the guidance provided by the central bank, announcements may introduce new sources of volatility. The influential paper by Morris and Shin (2002) has highlighted one channel through which central bank announcements may have a detrimental effect. Unlike standard models that assume all private agents share the same information, Morris and Shin focus on the more realistic case in which private agents have individual, heterogeneous sources of information and must attempt to forecast what others are expecting.8 Morris and Shin have argued that there can be a cost to providing more-accurate public information; agents may overreact to public information, making the economy more sensitive to any forecast errors in the public information. Subsequent research (e.g., Hellwig, 2004, and Svensson, 2006) has suggested that the MorrisShin result is not a general one and that better, more accurate, central bank information is welfare improving. However, just as the earlier literature on transparency employed models at odds with current policy frameworks (only surprises mattered, the money supply was the instrument), the Morris-Shin analysis is conducted within a framework that fails to capture important aspects of actual monetary policy. For example, the issue facing most central banks is not whether to provide more-accurate forecasts. Instead, the issue is whether or not to provide more information by, for example, announcing forecasts. And even in the absence of explicit announcements or guidance, central banks already provide information through the setting of the policy instrument. The impact of a change in the policy instrument will depend, in part, on the information that it conveys about the central bank’s view of the economy. The work by Morris and Shin has been extended by Amato and Shin (2003), who cast the Morris-Shin analysis in a more standard macro model. In their model, the central bank has per8 Woodford (2003) has investigated the role of higher-order expectations in inducing persistent adjustments to monetary shocks in the Lucas-Phelps islands model. See also Hellwig (2002). J U LY / A U G U S T 2008 425 Walsh fect information about the underlying shocks. This ignores the uncertainty policymakers themselves face in assessing the state of the economy. Nor do Amato and Shin allow the private sector to use observations on the policy instrument to draw inferences about central bank information. They also assume one-period price setting and represent monetary policy by a price level–targeting rule. In Hellwig (2004), prices are flexible and policy is given by an exogenous stochastic supply of money; private and public information consists of signals on the nominal quantity of money. The potential costs and benefits of releasing central bank forecasts have also been analyzed by Geraats (2005). However, Geraats assumes agents do not observe the bank’s policy instrument prior to forming expectations and employs a traditional Lucas supply function. Her focus is on reputational equilibria in a two-period model with a stochastic inflation target. Thus, the model and the issues addressed differ from the focus on the role of information in a Morris-Shin-like environment. Rudebusch and Williams (2006) and Gosselin, Lotz, and Wyplosz (forthcoming) focus specifically on the provision of future interest rate projections. Rudebusch and Williams explore the role of interest rate projections in a model of political transparency—the asymmetry of information pertains to policy preferences and the central bank inflation target. Transparency is modeled as reducing noise in central bank projections. In contrast to the model I develop in the next section, Rudebusch and William incorporate learning and find that the public’s ability to learn and welfare increase when interest rate projections are provided. Gosselin, Lotz, and Wyplosz (forthcoming) adopt a quite different approach and focus on what they characterize as creative opacity. In their model, the private sector learns from the information released by the central bank, but the central bank also learns about private sector information by observing long-term interest rates. By providing its projection for the short-term interest rate, the central bank is able to recover private sector information from the long-term rate. This aligns expectations but may require the central bank to distort its current interest rate setting to achieve 426 J U LY / A U G U S T 2008 the desired long-term rate. If central bank information is poor, it may be better to remain opaque. Although the role of central bank learning is a critical one, I ignore it in the model in the next section in order to focus on the way inflation and output are affected by central bank announcements. Thus, several questions remain unresolved concerning the role of transparency in an environment in which agents have heterogeneous information and central bank actions and announcements are commonly available. Specifically, how does the information conveyed by the central bank instrument affect the central bank’s incentives and alter the effectiveness of policy?9 What is the effect of more information as opposed to better information? And are concerns about the added uncertainty of greater transparency warranted? These questions are addressed in the model in the next section. WELFARE EFFECTS OF OPAQUENESS AND TRANSPARENCY To investigate the role of economic transparency, I employ a simple model motivated by New Keynesian models based on Calvo-type pricing adjustment by monopolistic firms and by Morris and Shin’s (2002) demonstration of the role heterogenous information can play.10 Like Gosselin, Lotz, and Wyplosz (forthcoming), I assume the central bank’s preferences are known. Unlike their model, however, I incorporate the common-knowledge effect central to the Morris and Shin model. However, I focus on how the private sector learns from information provided by the central bank and ignore the reverse inference, where the central bank learns from private sector information, which is key in the Gosselin, Lotz, and Wyplosz model. The basic model is similar to the one employed in Walsh (2007a,b). In these earlier papers, how9 In Walsh (2007b), I show that this incentive effect under discretion can make it socially optimal to appoint a Rogoff-conservative central banker, that is, a central banker who places less weight on outputgap stabilization than society does. 10 As noted earlier, in the basic Morris-Shin model, Svensson (2006) shows that for almost all parameter values, better central bank information is welfare improving. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh ever, only demand and cost shocks were present, so it was necessary to make just a single projection (of inflation or the output gap) to fully reveal the central bank information (because the public also observed the policy instrument). The primary focus was also on partial transparency in the sense of Cornand and Heinmann (2004). The chief contributions of the present paper are to enrich the information structure, to account fully for the welfare costs of relative price dispersion created by heterogeneous information, and to assess transparency in terms of both quantity (the role of providing more information) and quality (the effect of better information). Firms receive private signals on the fundamental shocks affecting the economy. Each period, a fraction of firms adjust their prices. In doing so, they are concerned with their relative price and so must attempt to forecast what other priceadjusting firms are doing. But this requires the individual firm to predict what other firms are predicting about the shocks hitting the economy. Hence, higher-order expectations will matter, as in Morris and Shin (2002). The central bank, like individual firms, is assumed to possess potentially noisy information on the economic outlook. I consider two policy regimes. In the first, the opaque regime, denoted by superscript o, the central bank makes no announcements. However, even in this regime, the central bank reveals something about its outlook for the economy when it sets its policy instrument. In the absence of other information, the private sector forms expectations by combining the observation on the instrument with their own private information. A rise in the policy interest rate, for example, will be interpreted partially as a central bank attempt to offset a projected positive demand shock and partially as an attempt to contract real output to offset a positive cost shock. When deciding on its policy, the central bank needs to take into account how the public will interpret its actions because the instrument conveys information. The second regime, denoted by superscript f, corresponds to full transparency. In this regime, the central bank releases its projections on future economic developments. Because it is on this F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W information that the central bank bases its policy decision, the actual setting of the instrument conveys no additional information. The benefits of this regime are that private sector forecasts are improved and, because there is more common information across firms, relative price dispersion is reduced. The potential cost is that private expectations react to what may turn out ex post to be central bank forecast errors. While I assume the central bank operates in a discretionary manner in setting its policy instrument, I also assume it can commit to a policy regime (opaque or transparent). The Basic Model The underlying model of price adjustment is based on Calvo, combined with the timing assumptions of Christiano, Eichenbaum, and Evans (2005) and the addition of firm-specific information. The Christiano, Eichenbaum, and Evans timing implies that firms who adjust their price for period t do so based on t–1 information. Expressed alternatively, firms in period t make decisions about their prices for period t +1. Because information differs across firms, price-setting firms will not all set the same price as in the standard common-information framework that is employed in most models. In addition, because firms care about their relative price, they must forecast the aggregate t +1 price level when they set their individual price for that period. This also differs from standard specifications in which firms are assumed to know the aggregate equilibrium price level when they set their price level. Three types of shocks are considered: (i) costs shocks that are assumed to represent inefficient volatility in real marginal costs; (ii) aggregate demand shocks; and (iii) shocks to the gap between the economy’s flexible-price equilibrium level of output and its efficient level of output. The last one will be referred to as a welfare-gap shock. The model differs from standard New Keynesian models in that the same information is not commonly available to all firms and firms must set prices before observing the current realizations of shocks. The basic timing is as follows: J U LY / A U G U S T 2008 427 Walsh (i) At the end of period t, the central bank forms projections about t+1 economic conditions and sets its policy instrument, θt . (ii) Firms observe πt , xt, and θt as well as individual specific signals about t +1 shocks. Firms may also observe announcements made by the central bank. (iii)Those firms that can adjust their price set prices for t +1. (iv) Period t +1 shocks occur and πt +1 and xt +1 are realized. A randomly chosen fraction 1 – ω of firms optimally set their price for period t +1. If β is the discount factor (see Walsh, 2007b), one can show that (1) π*j ,t +1 = (1 − ω ) E tj π*t +1 + (1 − ωβ )κ E tj xt +1 ωβ j + (1 − ωβ ) Etj ets+1 + E π , 1 − ω t t +2 where π *j,t +1 is the log price firm j sets for period t+1 relative to the period t average log price level (i.e., p*j,t +1 – pt ); E tjπ–t*+1 is firm j ’s expectation about the average π * i,t +1 being set by other adjusting j firms; E t xt +1 is firm j ’s expectation about the output gap in t+1; est+1 is the aggregate, common cost shock; and E tj π t +2 is firm j ’s expectation about future inflation. For simplicity, I assume (1) is linearized around a zero-inflation steady state. To keep the model simple, I represent the demand side of the model in a very stylized, reduced-form manner. Monetary policy is represented by the central bank’s choice of θt and by any announcements the central bank might make. I assume θt is observed at the start of the period so that any firm that sets its price in period t can condition its choice on the central bank’s policy action. The output gap is then equal to (2) xt +1 = θt + etv+1 , where evt+1 is a demand shock. Although I will call θt the central bank instrument, it essentially represents the central bank’s intended output gap. Information. As noted, there are three fundamental disturbances in the model: ets represents cost factors that, for a given output gap and expectations of future inflation, generate ineffi428 J U LY / A U G U S T 2008 cient inflation fluctuations; evt the aggregate demand disturbance; and eut a shock to the gap between the flexible-price output gap and the efficient output gap. I assume each is serially and mutually uncorrelated. Firms must set their prices and the central bank must set its policy instrument before learning the actual realizations of the aggregate shocks. i Firm j ’s idiosyncratic information, e j,t+1 for i = s, v, u, is related to the aggregate shock according to e ij ,t +1 = eti+1 + φ ij ,t +1, i = s,v ,u. i The φ j,t+1 terms are identically and independently distributed across firms and time. These signals are private in that they are unobserved by other i agents. For convenience, each φj,t+1 will be referred s to as a noise term, even though φ j,t+1 is actually the idiosyncratic component of the firm’s cost shock. All stochastic variables are assumed to be normally distributed. Define the signal-to-noise ratio, γ ji = σ 2i/共σ 2i + σ 2j,i 兲, where σ 2i is the variance 2 of ei and σ j,i is the variance of φ ji. Let Ωj,t +1 denote the vector of private signals received by firm j, and let Ωt+1 = 兰Ωj,t+1 be the information aggregated across firms. The central bank combines its information, models, and judgment to obtain forecasts of future economic disturbances. It will be convenient to represent this information, in parallel with the treatment of firm information, as signals on the three aggregate disturbances: i i i ecb ,t = et +1 + φcb,t , i = s,v ,u. i The noise terms φ cb are assumed to be independently distributed and to be independent of φ ji i 2 for all i, j, and t. Define γ cb = σ 2i/共σ 2i + σ i,cb 兲, where 2 i σ i,cb is the variance of φ cb . Let Ωcb,t +1 denote the innovation to the central bank information set. Let Z′t = [ets evt eut ]. Then Etcb Zt+1 = Γcb Ωcb,t+1, where E cb denotes expectations conditional on central bank information and Γcb s γ cb = 0 0 0 v γ cb 0 0 0 . u γ cb The central bank’s objective is to minimize, under discretion, a standard quadratic loss funcF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh tion that depends on inflation variability and output-gap variability. Specifically, loss is given by (3) ( ) cb i 2 u Lcb t = Et ∑ β π t + i + λ x xt + i − et + i , ∞ i =0 where eut is equal to stochastic variation in the gap between the flexible-price output gap (x) and the welfare-maximizing output gap. With staggered price adjustment, New Keynesian models imply that the welfare costs of inflation variability arise from the dispersion of relative prices it generates (Rotemberg and Woodford, 1997, Woodford, 2003a). Relative price dispersion can arise from inflation (because of staggered price adjustment) and because of heterogeneous information across firms. It can be shown (see the appendix) that the variance of relative prices across firms depends on π t2 and on the noise in the signals received by individual firms. Thus, social loss is given by ( ) (4) Lst = Et ∑ β i π t2+ i + λI zt2+i + λx xt + i − etu+ i , ∞ i =0 zt2 is where relative price dispersion arising from heterogenous information across individual firms, with the appropriate weight on this source of loss relative to π t2 given by λI = (1 − ω )2 . ω The loss associated with heterogeneous information can be reduced if the central bank provides more information. However, this loss is not affected by the period-by-period policy choice the central bank makes in setting its instrument (conditional on the policy regime that defines the type of announcements the central bank makes). Thus, under discretion, the central bank takes as given the term zt2 in (4), which is due to heterogeneous information, and minimizes (3). We can now evaluate equilibrium under each policy regime. (equivalently, its signals) directly to the public.11 In the absence of central bank announcements, firm j ’s new information is given by its private signals and the policy instrument. The new information available to firm j consists of Ωj,t+1 and θt . Assume beliefs about monetary policy are θt = δ o E tcbψ t +1 = δ o Γcb Ωcb ,t +1 , where δ 0 is 1 × 3. These beliefs are consistent with a rational expectations equilibrium under discretionary monetary policy. Define Θo = [Θo1 Θo2 ] such that Θo1 is 3 × 3 and Θo2 is 3 × 1, where the ij th element of Θo1 gives the effect of the firm’s j th signal on its forecast of the i th shock. Similarly, the i th element of Θo2 is the effect of θt on the firm’s forecast of the i th shock. Firm j ’s expectation of Zt+1 is Etj Zt +1 = Θo1 Ω j ,t +1 + Θo2θt . Because the firm’s signals on the different shocks are uncorrelated, Θo1 would, in the absence of the observation of θt, consist of a diagonal matrix with signal-to-noise ratios along the diagonal. The offdiagonal elements of Θo1 can be nonzero when the firm combines its own information with θt to forecast the shocks. For example, suppose θt > 0. This might indicate a response by the central bank to a negative demand shock, a negative cost shock, or a positive welfare-gap shock. If the firm’s signal on the demand shock is positive, then given θt, this makes it less likely the central bank is reacting to a negative demand shock. The firm will therefore alter its forecast of cost and target shocks. As shown in the appendix, the equilibrium strategy for firm j will take the form (5) π*j ,t +1 = b1o Ω j ,t +1 + b2oθt , where bo1 is 1 × 3. Under both regimes, the expression for the coefficients on Ωj,t+1 in the firm’s equilibrium strategy takes the same form.12 Equilibrium Under the Opaque Regime 11 In regime o, firms observe their own private signals and the central bank instrument. In regime f, the central bank provides its forecasts Alternatively, the central bank could announce its inflation and output-gap forecasts; combined with the observed instrument setting, these announcements would fully reveal the central bank’s signals. 12 Of course, their values differ under the two regimes to the extent that the information available to firms differs. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 429 Walsh The appendix shows also that the impact of the instrument on an individual firm’s pricing decision is (6) 1 − ωβ b2o = κ ω 1 + (1 − ω )b1o + (1 − ωβ ) (ι1 + κι2 ) Θo2 , ω where ιi is a 3 × 1 vector of zeros with a 1 in the i th place. Equation (6) illustrates the channels through which a policy action affects the pricing decisions of firms. The first term, 共1 – ωβ 兲κ /ω is the standard effect operating through the output gap. Because inflation is 共1 – ω 兲 times the pricing decision of the individual firm in a standard New Keynesian model, the effect on aggregate inflation operating through this terms would be 共1 – ω 兲 共1 – ωβ 兲κ/ω, which is the normal coefficient on the output gap in a New Keynesian model based on Calvo pricing. The remaining terms on the right side in (6) represent the informational effects of policy actions. For example, observing θt affects the firm’s expectations about cost, given by the term 共1 – ωβ 兲ιiΘo2 , and demand shocks, given by the term 共1 – ωβ 兲κιiΘo2 . Observing θt also affects individual pricing decisions through the firm’s expectations of what other firms are expecting, the 共1 – ω 兲bo1 term. Equilibrium inflation is given by (7) and ( πt +1 = (1 − ω ) π*t +1 = (1 − ω ) b1o Ωt +1 + b2oθt ) ∂πt +1 = (1 − ω )b2o . ∂θt The information channel can significantly affect the extent to which the central bank instrument impacts inflation. I calibrate the model by setting ω = 0.65 (as a compromise between micro evidence suggesting ω on the order of 0.5 and time-series estimates typically on the order of 0.8), β = 0.99, and κ = 1.8. These values imply 共1 – ω兲 共1 – ωβ 兲κ /ω = 0.3455. The standard deviations of all shocks are set equal to 1. Figure 1 shows how 共1 – ω兲b2o varies with the quality of private sector information, as measured by the signal-to-noise 430 J U LY / A U G U S T 2008 ratio, γ ji. When firms have perfect information on the shocks 共γ ji = 1兲, the policy instrument, θ, conveys no information and its effect on inflation equals 0.3455, which is shown by the horizontal line in Figure 1. However, when θ conveys information (i.e., when γ ji < 1), its impact on inflation is significantly reduced. Movements in θ are partially attributed to the central bank’s response to the various shocks. A rise in θ, for example, lowers firms’ forecasts of demand shocks. Because the net effect on the expected output gap is θt + E jievt+1, the effect on price-setting behavior and inflation is less than the change in θ. A rise in θ also leads firms to reduce their forecast of cost shocks, partially offsetting the positive impact of a rise in θ on inflation. For a given quality of private sector information, the information channel becomes more important as central bank information improves and private firms place more weight on the information conveyed by policy actions. The informational effects are larger, therefore, when the central bank has better quality information (in Figure 1, compare the solid line for i i γ cb = 0.5 with the dashed line for γ cb = 0.9). Operating in a discretionary regime, the central bank sets policy optimally in each period based on its current forecasts about the future state of the economy. The first-order condition for minimizing the expected value of the central bank’s loss function (3) subject to (2) and (7) is given in the appendix. This first-order condition can be solved for the optimal policy responses, and their values are also given in the appendix. The solution to the model is obtained numerically by beginning with initial values for the policy coefficients, using these to obtain Θo, bo1, and bo2, and then obtaining new values for the policy coefficients. This process continues until convergence. Once the equilibrium values of bo1 and bo2 and the policy coefficients are obtained, aggregate inflation is given by ( ) b1o + b2oδ o Γcb ψ t +1 , π t +1 = (1 − ω ) +b2o δ o Γ cbφcb ,t +1 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh Figure 1 Elasticity of Inflation with Respect to the Policy Instrument in the Opaque Regime as a Function of the Quality of Private Information o (1 – ω)*b 2 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Signal-to-Noise Ratio for Private Information i = 0.5; dotted line, γ i = 0.9. NOTE: Solid line, γ cb cb whereas the welfare gap is given by xt +1 − etu+1 = θt + etv+1 − etu+1 = δ o Γ cb Ωcb,t +1 + (ι2 − ι3 ) Zt +1 ( ) = δ o Γcb + ι2 − ι3 Zt +1 + δ o Γcbφcb ,t +1 . Equilibrium Under a Transparent Regime I interpret full transparency as a regime in which the central bank shares its information on the economy. Within the context of the model, this would mean that the central bank publishes its signals on the various disturbances so that Ωcb,t+1 becomes known to all firms. Equivalently, the central bank could publish its forecasts for inflation and the output gap. In a transparent regime, the instrument is no longer a source of information to the private sector. This alters the impact of θt on inflation and affects the central bank’s incentives for setting policy. When the cenF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W tral bank provides its information to the public, the central bank information set is a subset of the public information set. In this context, Svensson and Woodford (2003) have shown that certainty equivalence holds and the policy decision of the central bank depends only on the expected values of the shocks. In particular, this implies that the optimal policy will be independent of the quality of either central bank information or private sector information. Let Θ f = [Θ 1f Θ 2f ] be the appropriate 3 × 6 coefficient matrix such that Etj Zt +1 = Θ1f Ω j ,t +1 + Θ1f Ωcb ,t +1 . The appendix shows that the equilibrium strategy for price setting firms is π*j ,t +1 = b1f Ω j ,t +1 + b2f θt + b3f Ωcb,t +1, where bf1 takes the same form as bo1 (except when Θ 1f replaces Θo1 in the expression for bf1). Although the formula bf1 is the same as for bo1, their values J U LY / A U G U S T 2008 431 Walsh Table 1 Loss Under Alternative Regimes (σ 2s = σ 2v = σ 2u = 1) i γ cb γ ji 0.2 0.4 0.6 0.8 1.0 8.83 10.20 11.70 13.52 16.79 = 0.5 Opaque regime Transparent regime π Equivalent i γ cb 9.52 10.33 11.49 13.35 16.79 3.32 1.39 1.80 1.66 0 4.56 5.55 6.50 7.21 7.64 6.11 6.15 6.22 6.40 7.64 4.97 3.08 2.11 3.60 0 = 0.9 Opaque regime Transparent regime π Equivalent NOTE: Bold indicates the regime with the least loss. will differ as Θ f ⬆ Θ o. The effects of the central bank instrument and information are given by b2f = ( 1 − ωβ ) κ ω and 1 b3f = (1 − ω )b1f + (1 − ωβ )(ι1 + κι2 ) Θ2f . ω Inflation will equal 共1 – ω 兲π–t*+1, so ∂πt +1 (1 − ω )(1 − ωβ )κ = (1 − ω )b2f = ∂θt ω and is independent of any informational effects. The exact expressions for the optimal policy response to each type of signal are given in the appendix. THE VALUE OF RELEASING INFORMATION We can now compare the effects of providing information by comparing outcomes under the opaque regime and the transparent regime. To assess outcomes under the two regimes, the model is solved using the same calibrated parameters as employed earlier (i.e., ω = 0.65, β = 0.99, 432 J U LY / A U G U S T 2008 κ = 1.8). I initially set the variances of all shocks equal to 1. For the loss function, I set λx = 1/16, reflecting the use of quarterly inflation rates. Table 1 shows the loss under each regime for different combinations of the signal-to-noise ratios for both the private sector and the central bank. The first thing to note is the loss is increasing in the quality of private sector information (moving across rows from left to right) and decreasing in the quality of central bank information (comparing the top panel to the bottom panel). Better private information makes expectations more sensitive to signals and so increases the volatility of expectations. Greater volatility of expectations produces more inflation volatility. This is welfare decreasing. Better central bank information is welfare improving because it allows the central bank to engage in more effective stabilization policies that reduce the volatility of inflation and the output welfare gap. Although Morris and Shin (2002) suggest that improved commonly available information could reduce welfare, the results in Table 1 are consistent with Hellwig (2004) and Svensson (2006), who argue that better quality central bank information generally improves welfare. When γ ji = 1, firms observe the true shocks perfectly. In this case, the release of information or projections by the central bank is irrelevant and the loss is the same under both regimes, as shown in the last column of Table 1. When private F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh Table 2 Components of Loss (σ 2s = σ 2v = σ 2u = 1) 0.2 i γ cb = 0.9 γ ji 0.4 0.6 0.8 1.0 Opaque regime L 4.56 5.55 6.50 7.21 7.64 2 u λxσ x–e 1.82 2.15 2.85 3.72 3.22 1.68 1.71 1.84 2.20 4.42 1.07 1.70 1.81 1.30 0 L σ π2 6.11 6.15 6.22 6.40 7.64 2 u λxσ x–e 1.66 1.66 1.67 1.73 3.22 λIσ z2 4.42 4.42 4.42 4.42 4.42 0.02 0.06 0.13 0.26 0 σ π2 λIσ z2 Transparent regime information is imperfect, loss differs under the two regimes (the regime with the least loss is indicated in bold). The rows labeled “π equivalent” express the reduction in loss under the optimal regime in terms of the reduction in average inflation (expressed at annual rates) that would yield a similar reduction in loss. For example, if γ ji = 0.8 i and γ cb = 0.5, the improvement of moving from an opaque regime to a transparent one is equivalent to a reduction in inflation of 1.66 percentage points. The general results are similar in both the top panel, when central bank information is relatively poor (the signal and the noise have equal i variances so that γ cb = 0.5), and the bottom panel, when central bank information is relatively good i (γ cb = 0.9). What matters is the quality of private information. If this is low, then the expectations of firms (and what individual firms expect that other firms are expecting) are sensitive to any commonly available information released by the central bank. The results in Table 1 are robust to different values for the variances of the underlying shocks.13 The finding that transparency can lower welfare when private information is poor is suggestive of the Morris and Shin (2002) argument that noisy 13 For each σ i2, the value was changed between 2 and 0.01, whereas the other variances were held fixed at 1. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W public information can decrease welfare. To investigate whether this is the effect that accounts for the relative performance of the two regimes, one can calculate the sources of loss under each regime. From (4), loss arises from inflation variability, welfare-gap variability, and relative price dispersion caused by heterogeneous information. Table 2 shows each of these components for the i = 0.9, which corresponds to the lower case γ cb i = 0.5). panel of Table 1 (results are similar for γ cb Table 2 reveals three differences between the equilibria for the opaque and transparent regimes that are independent of the quality of private information. First, inflation is less volatile when policy is transparent. Second, the contribution of welfare-gap volatility to the overall loss is much larger when policy is transparent. And third, the welfare cost of relative price dispersion is much smaller when policy is transparent. When γ ji is very low, opacity is the preferred regime because the welfare gap is much more stable. As will be discussed further below, the informational effects of policy actions are larger when the quality of private information is poor and thus these effects distort the incentive of the central bank such that policy reacts too little to cost shocks. This makes inflation more volatile but leaves the welfare gap more stable. Both inflation and output-gap volatilJ U LY / A U G U S T 2008 433 Walsh Table 3 Optimal Policy Coefficients (σ 2s = σ 2v = σ 2u = 1) δ i γ cb = 0.5 γ ji = 0.4 s δ v u δ s δ v δ u Opaque regime –0.0947 –0.8884 0.7179 –0.2510 –0.9764 0.5246 Transparent regime –0.3647 –1.0000 0.3436 –0.3647 –1.0000 0.3436 i γ cb = 0.9 Opaque regime –0.0816 –0.8944 0.7475 –0.1865 –0.9713 0.6356 Transparent regime –0.3647 –1.0000 0.3436 –0.3647 –1.0000 0.3436 ity in the opaque regime increase as γ ji rises, so that the transparent regime becomes preferred when private sector information is good.14 Table 3 shows the optimal policy responses to three central bank signals for γ ji equal to 0.4 i and 0.8 and for γ cb equal to 0.5 and 0.9. Response coefficients in the transparent regime are independent of the quality of both private sector and the central bank information. This result follows from the demonstration by Svensson and Woodford (2003) that the central bank’s decision problem satisfies the conditions for certainty equivalence if the private sector has more information than the central bank. This is the case in the transparent regime because the private sector knows both the central bank signals and their own private signals. The way informational effects in the opaque regime distort stabilization policy is clear from the muted response (in absolute value) to signals on the cost shock and amplified response to signals on the welfare-gap shock. The tradeoff between inflation and welfare-gap volatility is clearly present—policy under the transparent regime responds more to stabilize inflation and, as a result, the welfare gap is more volatile, as was shown in Table 2. 14 δ γ ji = 0.8 Also apparent in Table 2 is that, in the transparent regime, the volatility of the welfare gap is independent of the quality of private sector information. This reflects the certainty equivalence property that characterizes the policy choice of the central bank in the transparent regime. The central bank’s setting of its instrument is independent of γ ji and, as a result, so is the behavior of the output and welfare gaps. 434 J U LY / A U G U S T 2008 In addition, transparency allows the central bank to more efficiently neutralize the effects of expected demand shocks. This can be seen by comparing the policy reaction coefficients under the two regimes. Under the transparent regime, expected demand shocks are completely offset (i.e., δ v = –1) regardless of the quality of private sector or central bank information. Under the opaque regime, δ v = –1 only when the public sector has perfect information on the shocks. Otherwise, δ v is less than 1 in absolute value and demand shocks are not fully offset. Under the opaque regime, when the policy instrument is moved, the public will confuse movements designed to offset forecasted demand shocks with movements designed to offset either cost or welfare-gap shocks. As a consequence, movements aimed at offsetting demand shocks can affect inflation expectations and cause actual inflation to fluctuate as the public attributes part of the instrument change to the other shocks. This makes it optimal to not offset demand shocks completely. Once the public can infer the central bank estimate of demand shocks, as it can under transparency, there is no longer any reason not to fully react to insulate the output gap and inflation from projected demand shocks, so δ 2v = –1. In New Keynesian models, the welfare costs of inflation are the result of the relative price dispersion that arises with staggered price adjustment. Heterogeneous information among firms will also create relative price dispersion. Because information provided by the central bank is comF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh Figure 2 Relative Price Dispersion Due to Heterogeneous Information σ z2 0.15 0.10 Opaque Regime, γcb Low Opaque Regime, γcb High Transparent Regime, γcb Low Transparent Regime, γcb High 0.05 0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Quality of Private Information mon to all firms, it can help reduce relative price dispersion. Figure 2 shows the measure of relative prices dispersion that results from heterogeneous information among firms. The solid line with asterisks corresponds to the case of poor-quality i = 0.5) under the central bank information (γ cb opaque regime, and the unconnected asterisks correspond to the opaque regime with high-quality central bank information (γ cb = 0.9). The diamonds indicate the outcomes under the transparent regime with poor-quality central bank information (the solid line) and high-quality central bank information (the unconnected diamonds). When γ ji = 1, all firms share the same information, so dispersion due to heterogeneous information goes to zero under either policy regime. When firms have very poor-quality information (i.e., for low initial values of γ ji ) the heterogeneity of the information is high, but because the information is of poor quality, firms do not respond strongly to it. As information quality improves, firms react more strongly to their own private information and this increases price dispersion. Hence, relative price dispersion is initially increasing in γ ji . Now consider the role of quality central bank information under the opaque regime. Relative price dispersion is lower when central bank inforF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W mation is good than when it is poor, though the loss from relative price dispersion actually constitutes a larger fraction of total social loss when central bank information is good. This is the result of the better stabilization the central bank can achieve when it has high-quality information on the economy. Not surprisingly, relative price dispersion is always lower under the transparent regime. For the same reason, high-quality central bank information reduces relative price dispersion under the transparent regime. CONCLUSIONS Under an opaque policy regime, where the private sector and the central bank do not share the same information, policy actions become a source of information to the public. And these policy actions have both direct effects on the output gap and indirect informational effects. Under an opaque regime, however, certainty equivalence does not hold and information channels affect the central bank’s incentives. Optimal policy will depend on the quality of both central bank information and public information. In an opaque regime, the central bank stabilizes inflation less J U LY / A U G U S T 2008 435 Walsh and the welfare gap more than it would in a transparent regime. Under a completely transparent regime, the public sector has access to the central bank assessment of the economy. In this case, policy actions no longer provide any additional information. Optimal policy is independent of the quality of central bank information. Consistent with the work of Svensson (2006) and Hellwig (2004), better central bank information was found to improve welfare. With better information, the central bank can implement more effective stabilization policies. The effect of providing more information by making announcements about projected inflation and the output gap is more ambiguous. Transparency always acts to lower relative price dispersion across firms by expanding the set of commonly available information, but central bank announcements can make expectations more volatile, particularly if firms have relatively poor information. Transparency dominates opacity when the private sector has relatively good information because in this case firms do not overreact to the information contained in central bank announcements. However, if private sector information is poor, central bank announcements can reduce welfare. So although better central bank information is desirable, more central bank information may not be. REFERENCES Amato, Jeffrey D. and Shin, Hyun Song. “Public and Private Information in Monetary Policy Models.” Working Paper No. 138, Bureau of Labor Statistics, September 2003. Andrews, Edmund L. “Bad News Puts Political Glare Onto Economy.” New York Times, September 8, 2007; http://www.nytimes.com/2007/09/08/business/ 08policy.html. Atkins, Ralph; Mackenzie, Michael and Davies, Paul J. “ECB Chief Fails to Reassure Markets.” Financial Times, August 15, 2007, p. 1. Bernanke, Ben S. “The Recent Financial Turmoil and its Economic and Policy Consequences.” Presented 436 J U LY / A U G U S T 2008 at the Economic Club of New York, New York, NY, October 15, 2007. Blinder, Alan S. “Monetary Policy Today: Sixteen Questions and About Twelve Answers.” Presented at the Banco de España conference Central Banks in the 21st Century, June 2006. Christiano, Lawrence J.; Eichenbaum, Martin and Evans, Charles. “Nominal Rigidities and the Dynamic Effects of a Shock to Monetary Policy.” Journal of Political Economy, 2005, 113(1), 1-45. Cornand, Camille and Heinemann, Frank. “Optimal Degree of Public Information Dissemination.” Working Paper No. 1353, CESifo, December 2004. Cukierman, Alex. “The Limits of Transparency.” Presented at the session “Monetary Policy Transparency and Effectiveness” at the American Economic Association, January 2006. Cukierman, Alex and Meltzer, Allan H. “A Theory of Ambiguity, Credibility, and Inflation under Discretion and Asymmetric Information.” Econometrica, September 1986, 54(5): pp. 1099-128. Dincer, Nergiz N. and Eichengreen, Barry. “Central Bank Transparency: Where, Why, and With What Effects?” NBER Working Paper No. 13003, National Bureau of Economic Research, March 2007. Eijffinger, Sylvester C.W. and Geraats, Petra M. “How Transparent Are Central Banks?” European Journal of Political Economy, March 2006, 22(1), pp. 1-21. Faust, Jon and Svensson, Lars E.O. “The Equilibrium Degree of Transparency and Control in Monetary Policy.” Journal of Money, Credit, and Banking, May 2002, 34(2), pp. 520-39. Geraats, Petra M. “Central Bank Transparency.” Economic Journal, November 2002, 112(483), pp. 532-65. Geraats, Petra M. “Transparency and Reputation: The Publication of Central Bank Forecasts” Topics in Macroeconomics, 2005, 5(1), pp 1-26. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh Goodhart, Charles A.E. “Letter to the Editor.” Financial Times, June 29, 2006. Gosselin, Pierre; Lotz, Aileen and Wyplosz, Charles. “The Expected Interest Rate Path: Alignment of Expectations vs. Creative Opacity Should Central Banks Reveal Expected Future Interest Rates?” International Journal of Central Banking (forthcoming). Guha, Krishna. “Debate Unfolds on Likely Impact of Cut in US Interest Rates.” Financial Times, September 6, 2007, p. 2. Harford, Tim. “Dear Economist.” Financial Times, September 1, 2007, Part 2, p. 3. Hellwig, Christian. “Public Announcements, Adjustment Delays and the Business Cycle.” UCLA, November 2002. Hellwig, Christian. “Heterogeneous Information and the Benefits of Transparency.” UCLA, December 2004. Morris, Stephen and Shin, Hyun Song. “Social Value of Public Information.” American Economic Review, December 2002, 92(5), pp. 1521-34. Poole, William. “Communicating the Fed’s Policy Stance.” Presented at the HM Treasury/GES conference Is There a New Consensus in Macroeconomics? London, November 30, 2005; www.stlouisfed.org/ news/speeches/2005/11_30_05.htm. Poole, William. “Fed Communications.” Presented to the St. Louis Forum, February 24, 2006; www.stlouisfed.org/news/speeches/2006/ 02_24_06.htm. Rotemberg, Julio J. and Woodford, Michael. “An Optimizing-Based Econometric Model for the Evaluation of Monetary Policy,” in Ben S. Bernanke and Julio J. Rotemberg, eds., NBER Macroeconomic Annual 1997. Cambridge, MA: MIT Press, 1997, pp. 297-346. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Rudebusch, Glenn D. and Williams, John C. “Revealing the Secrets of the Temple: The Value of Publishing Central Bank Interest Rate Projections.” October 2006; also in J.Y. Campbell, ed., Asset Prices and Monetary Policy. Chicago: University of Chicago Press (forthcoming). Svensson, Lars E.O. “Social Value of Public Information: Morris and Shin (2002) Is Actually Pro Transparency, Not Con.” American Economic Review, March 2006, 96(1), pp. 448-51. Svensson, Lars E. O. and Woodford, Michael. “Optimal Policy With Partial Information in a Forward-Looking Model: Certainty-Equivalence Redux.” NBER Working Paper No. w9430. National Bureau of Economic Research, January 2003. Walsh, Carl E. “Announcements, Inflation Targeting and Central Bank Incentives.” Economica, May 1999, 66(262), pp. 255-69. Walsh, Carl E. “Transparency, Flexibility, and Inflation Targeting,” in Frederic Mishkin and Klaus Schmidt-Hebbel, eds., Monetary Policy Under Inflation Targeting. Santiago, Chile: Banco Central de Chile, 2007a. Walsh, Carl E. “Optimal Economic Transparency.” International Journal of Central Banking.” March 2007b, 3(1), pp. 5-30. Woodford, Michael. Money, Interest, and Prices, Princeton: Princeton University Press, 2003. Woodford, Michael. “Central Bank Communications and Policy Effectiveness.” Presented at the Federal Reserve Bank of Kansas City symposium The Greenspan Era: Lessons for the Future, Jackson Hole, WY, August 2005. Yellen, Janet L. “Policymaking on the FOMC: Transparency and Continuity.” Federal Reserve Bank of San Francisco Economic Letter, No. 2005-22, September 2, 2005. J U LY / A U G U S T 2008 437 Walsh APPENDIX Welfare Weight on Information Dispersion The welfare loss in New Keynesian models arises from inefficient price dispersion across firms. – Let pj,t denote firm j ’s price and let Pt be the aggregate price level. Then ( ∆t ≡ varj logp j ,t − Pt −1 ( = E ( logp = Et logp j ,t − Pt −1 t ( ) j ,t − Pt −1 ( ) ) − ( E logp − P ) ) − (P − P ) . 2 2 t −1 t ) 2 t −1 j ,t t 2 ( Using the assumptions of the Calvo model, the first term on the right can be written as E t logp j ,t −1 − Pt −1 Now 2 + (1 − ω ) E t logp*j ,t − logPt −1 2 ) = ω ∆t −1 + (1 − ω ) Et logp*j ,t − Pt −1 . 2 logp*j ,t − P t −1 = logp*j ,t − logp*t + logp*t − Pt −1, where the first term on the right is zero in the standard New Keynesian model with common information across firms. Hence, ( Et logp*j ,t − Pt −1 ) ( ) + (logp = E t logp*j ,t − logp*t 2 2 * t − Pt −1 ) 2 because the idiosyncratic noise is independent of the fundamental shocks. From the definition of inflation, ( so E t logp*j ,t − Pt −1 Combining these results, ( ) π t = (1 − ω ) logp*t − Pt −1 , ) 2 ( = E t logp*j ,t − logp*t ( ∆t = ω ∆t −1 + (1 − ω ) E t logp*j ,t − logp*t ( = ω ∆t −1 + (1 − ω ) E t logp*j ,t − logp*t It follows that where ) 2 ) 2 ) 2 1 2 π . + 1 − ω t 2 1 2 + π − πt2 1 − ω t ω 2 + π . 1 − ω t ( ) ∞ ∞ 2 ω 1 i 2 * * E t ∑ β i ∆t + i = E t ∑ β π t + i + λ I logp j ,t + i − logpt + i , 1 − ω 1 − ωβ i = 0 i =0 λ I = (1 − ω ) / ω . 2 438 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh The Opaque Regime Let σ s2 0 Σ = 0 σ v2 0 0 Σj and Σcb σ s2 + σ 2j ,s = 0 0 2 σ s2 + σ cb ,s 0 = 0 0 0 , σ u2 0 σ v2 + σ 2j ,v 0 0 2 σ v2 + σ cb ,v 0 , 0 σ u2 + σ 2j ,u 0 0 . 2 2 σ u + σ cb,u 0 In the absence of central bank announcements, firm j ’s new information is given by ets+1 + φ sj ,t +1 etv+1 + φ vj ,t +1 Ω j ,t +1 = , etu+1 + φ uj ,t +1 θt θt where θt = δ o Γcb Ωcb ,t +1 . Define Θ = Σ o Σj ΣΓcb ′ δ o′ ΣΓcb δ ′ o δ Γ cb Σ δ o Γ cb Σcb Γc′ b δ o ′ o′ −1 = Θo1 Θo2 , where Θo1 is 3 × 3 and Θo2 is 3 × 1. Thus, firm j ’s expectation of Zt +1 is Etj Zt +1 = Θo1 Ω j ,t +1 + Θo2θt = Θo1 Ω j ,t +1 + Θo2δ o Γ cb Ωcb ,t +1 . The aggregate information (i.e., aggregated across all firms) is ets+1 Ωt +1 etv+1 Zt +1 θ = u = θ . t et +1 t θ t F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 439 Walsh Defining ιi as a 1 × 3 vector with a 1 in the i th place and zeros elsewhere, we can write (1), a firm’s price adjustment, as ωβ j π*j ,t +1 = (1 − ω ) E tj π*t +1 + (1 − ωβ )κθt + (1 − ωβ )(ι1 + κι2 ) Etj Zt +1 + E π 1 − ω t t +2 = (1 − ω ) Etj π t*+1 + (1 − ωβ )κθt ( ) ωβ j + (1 − ωβ )(ι1 + κι2 ) Θo1 Ω j ,t +1 + Θo2θt + E π . 1 − ω t t +2 An equilibrium strategy for firm j will take the form π*j ,t +1 = b1o Ω j ,t +1 + b2oθt , where bo1 is 1 × 3. In forming expectations about the pricing behavior of other firms adjusting in the current period, firm j ’s expectation of π–t*+1 is given by E tj π*t +1 = b1o Etj+1Ωt +1 + b2oθt = b1o Etj+1Zt +1 + b2oθt = b1o Θ1o Ω j ,t +1 + Θo2θt + b2oθt ( ) = b1o Θ1o Ω j ,t +1 + b1o Θo2 + b2o θt . Because π t +1 = (1 − ω )π*t +1 , E tj π t +2 = (1 − ω ) Etj π t*+2 it follows that ( ) = (1 − ω ) Etj b1o Θ1o Ω j ,t +2 + b1o Θ2o + b2o θt +1 = 0. Substituting these into the equation for π * j,t +1 and collecting terms, π*j ,t +1 = (1 − ω )b1o Θ1o Ω j ,t +1 + (1 − ωβ ) (ι1 + κι2 ) Θo1 Ω j ,t +1 ( ) + (1 − ωβ )κθt + (1 − ω ) b1o Θo2 + b2o θt + (1 − ωβ )(ι1 + κι2 ) Θo2θt . Equating coefficients with the proposed solution yields b1o = (1 − ωβ ) (ι1 + κι2 ) Θ1o I 3 − (1 − w ) Θ1o . −1 The expression for bo2 is reported in the text. The objective function under discretion involves minimizing ( ) 2 E tcb π t2+1 + λx xt +1 − etu+1 440 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Walsh (1 − ω )bθ Etcbπ t +1 + λx (θt + Etcbetv+1 − Etcbetu+1 ) = 0. subject to (2) and (7). The first-order condition for the central bank decision problem under discretion is Etcb π t +1 = (1 − ω )b1o Etcb Ω j ,t +1 + (1 − ω )b2oθt From (7), = (1 − ω )b1o Γcb Ωcb ,t +1 + (1 − ω )b2oθt because E tcb Ω j ,t +1 = E tcb Zt +1 = Γ cb Ωcb ,t +1 . ( ) Hence, the first-order condition becomes λ + (1 − ω )2 b o 2 θ = (1 − ω )bo E cb eu 2 2 t t +1 x t − (1 − ω ) b2o b1o Γ cb Ωcb ,t +1 − λx Etcbetv+1 . 2 This in turn implies that ( ) θ λ + (1 − ω )2 b o 2 x Hence, where δ o = [δ s δ v δ u ] and 2 t = − (1 − w ) b2ob1o Γcb Ωcb ,t +1 2 + 0 − λx (1 − ω )b2o Γ cb Ωcb,t +1 . θt = δ o Γcb Ωcb ,t +1 , (8) 2 o 1 − ω ) b2ob11 ( δ =− + 1− 2 o λx ( ω ) b2 (9) 2 o λx + (1 − ω ) b2ob12 δ =− + − 2 o 2 λx (1 ω ) b2 s v 2 ( ) ( ) 2 o λ − (1 − ω ) b2ob13 . δu = x + 1− 2 o 2 λx ( ω ) b2 ( ) The Transparent Regime In regime f, the central bank announces its signals so that firms observe Ωcb,t+1 directly. Firms’ expectations now depend on Ωcb,t+1 and not directly on θt. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 441 Walsh Guess an equilibrium strategy of the form π ∗j ,t +1 = b1f Ω j ,t +1 + b2f Ωcb ,t +1 + b3f θt . Then, following the same procedures as used to solve the model without announcements, one finds that b1f = (1 − ωβ )(ι1 + κι2 ) Θ1f I 3 − (1 − ω ) Θ1f 1 b2f = (1 − ω )b f + (1 − ωβ )(ι1 + κι2 ) Θ2f . ω −1 b3f = (1 − ωβ ) κ . ω (1 − ω )b3f Etcbπ t +1 + λx (θt + Etcbetv+1 − Etcbetu+1 ) = 0. Optimal policy in this regime satisfies the first-order condition ( Note that ) E tcb π t +1 = (1 − ω ) b1f Γ cb + b2f Ωcb ,t +1 + b3f θt because E tcb Ω j ,t +1 = Γcb Ωcb ,t +1 . Solving the first-order condition yields θt = d f Γcb Ωcb ,t +1 , with d f = [d s d v d u ] and ( ) bf / γ s + bf 21 11 cb 2 d s = − (1 − ω ) b3f 2 f 2 λx + h (1 − w ) ( ) λ + (1 − ω ) b (b / γ ) + b = − λ + (b ) (1 − ω ) λ − (1 − ω ) b (b / γ ) + b = . λ + b ω 1 − ( ) ( ) 2 d v f 3 x 2 d x x 442 J U LY / A U G U S T 2008 v cb f 2 3 x u f 22 f 3 f 2 3 f 23 f 12 2 u cb f 13 2 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Commentary Marvin Goodfriend B ack in 1984, I was invited by Bill Poole, then a member of the President’s Council of Economic Advisors, to work as a senior staff economist for money and banking at the Council. When I arrived at the Old Executive Office Building that fall, I brought with me an early draft of a paper on central bank secrecy that I had just finished. I gave a copy to Bill, who I knew had a longstanding interest in central bank communications. I remember his reaction: Bill put the paper in an envelope, signed it, wrote on it “for my eyes only,” and had his secretary put it in a safe. How appropriate, I thought! Later, Bill asked for a briefing and I described among other things the substance of the FOMC defense of monetary policy secrecy in a recently concluded Freedom of Information Act lawsuit. My interest in the topic was initiated by a headline in the American Banker that read “Secrecy Primary Tool of Monetary Policy.” How could that be? The assertion seemed at odds with everything Bill taught us in graduate school at Brown—that, according to rational expectations theory, more information should be better than less. Bill emphasized that private agents have an incentive to use to their advantage whatever information they have, whatever its source. I wrote the paper to explore under what circumstances, if any, central bank secrecy could be justified. I would never have predicted that “information policy” would have generated so much interest among central bankers or as much research as it does today. The main difference is that today we speak of central bank “transparency” or “policy guidance” rather than central bank “secrecy.” I remember thinking that it was unlikely that central bank secrecy would ever be debated openly, and if it ever were, then I thought the case for transparency would quickly win the day. I wasn’t quite correct on either outcome. In any case, it is useful to recall how far we’ve come. For the most part, central banks have moved away from secrecy toward transparency, partly by being more explicit about longer-run inflation objectives and partly by communicating shortterm policy concerns and intentions more explicitly. Few would now claim that secrecy is a tool of monetary policy. Quite the contrary, communication is today widely recognized to play a central role in monetary policy. That said, we have come to the point where even those who favor transparency in principle worry that excessive forward guidance on monetary policy might be counterproductive. This is the thrust of the concern expressed by Bill that motivates Carl Walsh’s (2008) paper. The balance of my remarks addresses the limits of forward guidance by drawing a distinction between two dimensions of information policy: (i) transparency with regard to a long-run inflation objective and (ii) discretionary announcements used by central banks to substitute for transparency about a long-run inflation objective. Transparency and communication help to implement interest rate policy in two ways— Marvin Goodfriend is a professor of economics at the Tepper School of Business, Carnegie Mellon University. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 443-45. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 443 Goodfriend first, because a central bank uses a nominal interest rate policy instrument to manage real interest rates, and second, because a central bank uses a short-term interest rate to manage longerterm rates. The primary role of transparency and communication must be to convey clearly the central bank’s long-run inflation objective in order to anchor inflation expectations firmly, so that a central bank can manage short- and longer-term real interest rates reliably with its nominal interest rate policy instrument. The secondary role of communication is to help exercise leverage over longer-term interest rates with a short-term interest rate policy instrument. This a central bank can do because financial markets price longer-term interest rates (up to a possibly time-varying term premium) as an average of expected future short rates. To manage expectations of future short rates, a central bank must take markets into its confidence by communicating its intentions based on its forecast of economic conditions, its structural view of the economy, and its medium-term objectives for employment, financial stability, and inflation. Naturally, central banks are reluctant to reveal much of their current concerns or intentions because judgments about such things are necessarily imperfect, tentative, and subject to frequent revision. And some central banks are reluctant to announce explicitly their longer-run inflation objectives too. On the other hand, central banks recognize that interest rate policy benefits from transparency and communication, and central banks are inclined to be evermore revealing of their thinking in an effort to better manage longerterm interest rates. The reluctance of central bankers to take markets systematically into their confidence creates a reliance on announcements to convey their concerns about the economy and their intentions for short-term interest rates. Discretionary announcements employed to guide markets in lieu of systematic transparency about underlying objectives and concerns would appear to provide a degree of flexibility in communication policy. The point I wish to make, however, is that it is an illusion to think that discretionary announcements can 444 J U LY / A U G U S T 2008 substitute reliably for systematic strategic transparency. Inevitably, the public will find it hard to interpret announcements made without strategic guidance and, therefore, a central bank will find it hard to predict the public’s reaction to such announcements. My point is nothing more than to apply rational expectations reasoning, made famous by Robert Lucas, to announcements. It is difficult for a central bank to predict how either a policy action or a discretionary announcement will be interpreted by markets when undertaken with insufficient strategic guidance, that is, when either is undertaken independently of a policy rule. The Federal Reserve’s experience in May and June 2003 is a case in point. The Fed famously accompanied a cut in its federal funds rate target at the May 2003 FOMC meeting with a surprise announcement that significant further disinflation would be “unwelcome.” The statement was intended to alert the market to the fact that the Fed would act to deter deflation. The Fed was taken by surprise by what it considered an overreaction in the media and markets to its concern for deflation. The Fed rectified matters by dropping the federal funds rate by only 25 basis points at the June FOMC meeting instead of the expected 50 basis points. The market reaction to the surprise May 2003 FOMC announcement was excessive relative to what the Fed expected, but it could have been just as easily insufficient relative to what the Fed intended. Either way, such misunderstandings are potentially costly for the implementation of interest rate policy because they whipsaw markets, create confusion, and weaken a central bank’s ability to manage interest rates. Failing to convey a monetary policy message accurately in the first place can produce an extended period of policyinduced volatility as the mutual understanding between markets and the central bank on interest rate policy is gradually and painfully restored. Arguably, the confusion in 2003 could have been avoided if an explicit numerical lower bound on the Fed’s tolerance range for core personal consumption expenditures inflation had been in place. Markets would have been prepared for interest rate actions the Fed would take as F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Goodfriend inflation neared the 1 percent lower bound on its tolerance range. And longer-term interest rates would have drifted down as inflation drifted lower in early 2003 in anticipation of the Fed’s reaction. In that context, announcements could have reinforced reliably the Fed’s concern about further disinflation and the credibility of its commitment to prevent inflation from falling below 1 percent. In conclusion, and returning to Bill Poole’s concern about excessive forward guidance, we can say this: Forward guidance on interest rate policy is likely to be most effective when it reinforces a well-articulated monetary policy strategy anchored by an explicit numerical long-run inflation target. Otherwise, forward guidance should be undertaken with care and only with good reason given that discretionary announcements are difficult if not impossible to calibrate consistently to achieve their intended effect. REFERENCE Walsh, Carl E. “Announcements and the Role of Policy Guidance.” Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 421-42. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 445 446 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Rules-of-Thumb for Guiding Monetary Policy William Poole This article was originally published in the Board of Governors of the Federal Reserve System Open Market Policies and Operating Procedures—Staff Studies, July 1971. It is reprinted here as an addendum to these conference proceedings. Federal Reserve Bank of St. Louis Review, July/August 2008, 90(4), pp. 447-97. INTRODUCTION T his study has been motivated by the recognition that the key to understanding policy problems is the analysis of uncertainty. Indeed, in the absence of uncertainty it might be said that there can be no policy problems, only administrative problems. It is surprising, therefore, that there has been so little systematic attention paid to uncertainty in the policy literature in spite of the fact that policymakers have repeatedly emphasized the importance of the unknown. In the past, the formal models used in the analysis of monetary policy problems have almost invariably assumed complete knowledge of the economic relationships in the model. Uncertainty is introduced into the analysis, if at all, only through informal consideration of how much difference it makes if the true relationships differ from those assumed by the policymakers. In this study, on the other hand, uncertainty plays a key role in the formal model. Since this study is so long, a few comments at the outset may assist the reader in finding his way through it. The remainder of this introduc- tory section outlines the structure of the study so that the reader can see how the various parts fit together. The reader interested only in a summary of the analysis and empirical findings should read this introductory section and then turn directly to the summary in Section V. This summary concentrates on the theoretical analysis while only briefly stating the most important empirical findings. It omits completely the technical details of both the theoretical and empirical work. The reader interested in the technical details should, of course, turn to the appropriate parts of Sections I through IV. Insofar as possible these sections have been written so that the reader can understand any one section without having to wade through all of the other sections. Section I contains the theoretical argument comparing interest rates and the money stock as policy-control variables under conditions of uncertainty. The analysis is verbal and graphical, using the simple Hicksian IS-LM model with random terms added. This model is general enough to include both Keynesian and monetarist outlooks, depending on the specific assumptions as to the shapes of the functions. Since the theoretical analysis emphasizes the importance of the relative William Poole is a former president of the Federal Reserve Bank of St. Louis. At the time this article was written, he was a senior economist in the special studies section of the division of research and statistics at the Board of Governors of the Federal Reserve System. Joan Walton, Lillian Humphrey, and Debra Bellows provided research assistance. © 2008, The Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 447 Poole stability of the expenditures and money demand functions, an examination of the evidence on relative stability appears in Section II. Given the conclusion of Section II on the superiority of a policy operating through adjustments in the money stock, the next question is how the money stock should be adjusted to achieve the best results. While policymakers generally look askance at suggestions for policy rules, the only way that economists can give long-run advice is in terms of rules. That is to say, the economist is not being helpful at all if he in effect says, “Look at the rate of inflation, at the rate of unemployment, at the forecasts of the government budget deficit, and at other relevant factors, and then act appropriately.” Advice requires the specification of exactly how policy should be adjusted, and for this advice to be more than an ad hoc recommendation for the current situation, it must involve specification of how the money stock or some other control variable should be adjusted under hypothetical future conditions of inflation, unemployment, and so forth. The purpose of Section III is to develop such a rule-of-thumb, or policy guideline, based on the theoretical and empirical analyses of Sections I and ll. A number of technical problems of monetary control are examined in Section IV. After a short introduction to the issues, the first part of this section discusses the relative merits of a number of monetary aggregates including various reserve measures, the narrowly and broadly defined money stocks, and bank credit. The second part examines whether policy should specify desired rates of change of an aggregate in terms of weekly, monthly, or quarterly averages, or in some other manner. The third part examines in a very incomplete fashion a few of the problems of adjusting open market operations so as to reach the desired level of an aggregate. Finally, Section V consists of a summary of Sections I through IV. To avoid undue repetition, woven into this summary section are a number of general observations not examined in the other sections. 448 J U LY / A U G U S T 2008 I. THE THEORY OF MONETARY POLICY UNDER UNCERTAINTY Basic Concepts The theory of optimal policy under uncertainty has provided many insights into actual policy problems (Theil, 1964; Brainard, 1967; Holt, 1962; Poole, 1970). While much of this theory is not accessible to the nonmathematical economist, it is possible to explain the basic ideas without resort to mathematics. The obvious starting point is the observation that with our incomplete understanding of the economy and our inability to predict accurately the occurrence of disturbing factors such as strikes, wars, and foreign exchange crises, we cannot expect to hit policy goals exactly. Some periods of inflation or unemployment are unavoidable. The inevitable lack of precision in reaching policy goals is sometimes recognized by saying that the goals are “reasonably” stable prices and “reasonably” full employment. While the observation above is trite, its implications are not. Two points are especially important. First, policy should aim at minimizing the average size of errors. Second, policy can be judged only by the average size of errors over a period of time and not by individual episodes. Because this second point is particularly subject to misunderstanding, it needs further amplification. Since policymakers operate in a world that is inherently uncertain, they must be judged by criteria appropriate to such a world. Consider the analogy of betting on the draw of a ball from an urn with nine black balls and one red ball. Anyone offered a $2 payoff for a $1 bet would surely bet on a black ball being drawn. If the draw produced the red ball, no one would accuse the bettor of a stupid bet. Similarly, the policymaker must play the economic odds. The policymaker should not be accused of failure if an inflation occurs as the result of an improbable and unforeseeable event. Now consider the reverse situation from that considered in the previous paragraph. Suppose the bettor with the same odds as above bets on the red ball and wins. Some would claim that the bet was brilliant, but assuming that the draw was not F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole rigged in any way, the bet, even though a winning one, must be judged foolish. It is foolish because, on the average, such a betting strategy will lead to substantially worse results than the opposite strategy. Betting on red will prove brilliant only one time out of 10, on the average. Similarly, a particular policy action may be a bad bet even though it works in a particular episode. There is a well-known tendency for gamblers to try systems that according to the laws of probability cannot be successful over any length of time. Frequently, a gambler will adopt a foolish system as the result of an initial chance success such as betting on red in the above example. The same danger exists in economic policy. In fact, the danger is more acute because there appears to be a greater chance to “beat the system” by applying economic knowledge and intuition. There can be no doubt that it will become increasingly possible to improve on simple, naive policies through sophisticated analysis and forecasting and so in a sense “beat the system.” But even with improved knowledge some uncertainty will always exist, and therefore so will the tendency to attempt to perform better than the state of knowledge really permits. Whatever the state of knowledge, there must be a clear understanding of how to cope with uncertainty, even though the degree of uncertainty may have been drastically reduced through the use of modern methods of analysis. The principal purpose of this section is to improve understanding of the importance of uncertainty for policy by examining a simple model in which the policy problem is treated as one of minimizing errors on the average. Particular emphasis is placed on whether controlling policy by adjusting the interest rate or by adjusting the money stock will lead to smaller errors on the average. The basic argument is designed to show that the answer to which policy variable—the interest rate or the money stock—minimizes average errors depends on the relative stability of the expenditures and money demand functions and not on the values of parameters that determine whether monetary policy is in some sense more or less “powerful” than fiscal policy. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Figure 1 r LM1 r0 LM2 IS Y Yf SOURCE: Originally published version, p. 139. Monetary Policy Under Uncertainty in a Keynesian Model 1 The basic issues concerning the importance of uncertainty for monetary policy may be examined within the Hicksian IS-LM version of the Keynesian system. This elementary model has two sectors, an expenditure sector and a monetary sector, and it assumes that the price level is fixed in the short run.2 Consumption, investment, and government expenditures functions are combined to produce the IS function in Figure 1, while the demand and supply of money functions are combined to produce the LM function. If monetary policy fixes the stock of money, then the resulting LM function is LM1, while if policy fixes the interest rate at r0 the resulting LM function is LM2. It is assumed that incomes above “full employment income” are undesirable due to inflationary pressures while incomes below full employment income are undesirable due to unemployment. If the positions of all the functions could be predicted with no errors, then to reach full employment income, Yf , it would make no differ1 For the most part this section represents a verbal and graphical version of the mathematical argument in Poole (1970). 2 Simple presentations of this model may be found in Reynolds (1969, pp. 275-82) and Samuelson (1967, pp. 327-32). J U LY / A U G U S T 2008 449 Poole Figure 2 Figure 3 SOURCE: Originally published version, p. 141. SOURCE: Originally published version, p. 140. ence whether policy fixed the money stock or the interest rate. All that is necessary in either case is to set the money stock or the interest rate so that the resulting LM function will cut the IS function at the full employment level of income. Significance of Disturbances. The positions of the functions are, unfortunately, never precisely known. Consider first uncertainty over the position of the IS function—which, of course, results from instability in the underlying consumption and investment functions—while retaining the unrealistic assumption that the position of the LM function is known. What is known about the IS function is that it will lie between the extremes of IS1 and IS2 in Figure 2. If the money stock is set at some fixed level, then it is known that the LM function will be LM1, and accordingly income will be somewhere between the extremes of Y1 and Y2. On the other hand, suppose policymakers follow an interest rate policy and set the interest rate at r0. In this case income will be somewhere between Y1′ and Y2′, a wider range than Y1 to Y2, and so the money stock policy is superior to the interest rate policy.3 3 In Figure 2 and the following diagrams, the outcomes from a money stock policy will be represented by unprimed Y’s, while 450 J U LY / A U G U S T 2008 The money stock policy is superior because an unpredictable disturbance in the IS function will affect the interest rate, which in turn will produce spending changes that partly onset the initial disturbance. The opposite polar case is illustrated in Figure 3. Here it is assumed that the position of the IS function is known with certainty, while unpredictable shifts in the demand for money cause unpredictable shifts in the LM function if a money stock policy is followed. With a money stock policy, income may end up anywhere between Y1 and Y2. But an interest rate policy can fix the LM function at LM3 so that it cuts the IS function at the full employment level of income, Yf . With an interest rate policy, unpredictable shifts in the demand for money are not permitted to affect the interest rate; instead, in the process of fixing the interest rate the policymakers adjust the stock of money in response to the unpredictable shifts in the demand for money. In practice, of course, it is necessary to cope with uncertainty in both the expenditure and monetary sectors. This situation is depicted in the outcomes from an interest rate policy will be represented by primed Y’s. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole Figure 4 Figure 5 SOURCE: Originally published version, p. 141. SOURCE: Originally published version, p. 141. Figure 4, where the unpredictable disturbances are larger in the expenditure sector, and in Figure 5 where the unpredictable disturbances are larger in the monetary sector. The situation is even more complicated than shown in Figures 4 and 5 by virtue of the fact that the disturbances in the two sectors may not be independent. To illustrate this case, consider Figure 5 in which the interest rate policy is superior to the money stock policy if the disturbances are independent. Suppose that the disturbances were connected in such a way that disturbances on the LM1 side of the average LM function were always accompanied by disturbances on the IS2 side of the average IS function. This would mean that income would never go as low as Y1, but rather only as low as the intersection of LM1 and IS2, an income not as low as Y1′ under the interest rate policy. Similarly, the highest income would be given by the intersection of LM2 and IS1, an income not so high as Y2′.4 4 The diagram could obviously have been drawn so that an interest rate policy would be superior to a money stock policy even though there was an inverse relationship between the shifts in the IS and LM functions. However, inverse shifts always reduce the margin F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Importance of Interest Elasticities and Other Parameters. So far the argument has concentrated entirely on the importance of the relative sizes of expenditure and monetary disturbances. But is it also important to consider the slopes of the functions as determined by the interest elasticities of investment and of the demand for money, and by other parameters? Consider the pair of IS functions, IS1 and IS2, as opposed to the pair, IS3 and IS4, in Figure 6. Each pair represents the maximum and minimum positions of the IS function as a result of disturbances, but the pairs have different slopes. Each pair assumes the same maximum and minimum disturbances, as shown by the fact that the horizontal distance between IS1 and IS2 is the same as between IS3 and IS4. For convenience, but without loss of generality, the functions have been drawn so that under an interest rate policy represented by LM2 both pairs of IS functions produce the same range of incomes. To keep the diagram from becoming of superiority of an interest rate policy, possibly to the point of making a money stock policy superior. Conversely, positively related shifts favor an interest rate policy. J U LY / A U G U S T 2008 451 Poole Figure 6 Figure 7 SOURCE: Originally published version, p. 141. SOURCE: Originally published version, p. 141. Figure 8 Figure 9 SOURCE: Originally published version, p. 143. SOURCE: Originally published version, p. 143. 452 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole too messy, only one LM function, LM1, under a money stock policy has been drawn. Now consider disturbances that would shift LM1 back and forth. From Figure 6 it is easy to see that if shifts in LM1 would lead to income fluctuations greater than from Y1′ to Y2′—which fluctuations would occur under an interest rate policy—then an interest policy would be preferred regardless of whether we have the pair IS1 and IS2, or the pair IS3 and IS4. The importance of the slope of the LM function is investigated in Figure 7 for the two LM pairs, LM1 and LM2, and LM3 and LM4. The functions have been drawn so that each pair represents different slopes but an identical range of disturbances. It is clear that if shifts in IS1 are small enough, then an interest rate policy will be preferred regardless of which pair of LM functions prevails. Conversely, if a money stock policy is preferred under one pair of LM functions because of the shifts in the IS function, then a money stock policy will also be preferred under the other pair of LM functions. The upshot of this analysis is that the crucial issue for deciding upon whether an interest rate or a money stock policy should be followed is the relative size of the disturbances in the expenditure and monetary sectors. Contrary to much recent discussion, the issue is not whether the interest elasticity of the demand for money is relatively low or whether fiscal policy is more or less “powerful” than monetary policy. To avoid possible confusion, it should be emphasized that the above conclusion is in terms of the choice between a money stock policy and an interest rate policy. However, if a money stock policy is superior, then the steeper the LM function is, the lower the range of income fluctuation, as can be seen from Figure 7. It is also clear from Figure 6 that under an interest rate policy an error in setting the interest rate will lead to a larger error in hitting the income target if the IS function is relatively flat than if it is relatively steep. But these facts do not affect the choice between interest rate and money stock policies. The “Combination” Monetary Policy. Up to this point the analysis has concentrated on the choice of either the interest rate or the money F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W stock as the policy variable. But it is also possible to consider a “combination” policy that works through the money stock and the interest rate simultaneously. An understanding of the combination policy may be obtained by further consideration of the cases depicted in Figures 2 and 7. In Figure 8 the disturbances, as in Figure 2, are entirely in the expenditure sector. As was seen in Figure 2, the result obtained by fixing the money stock so that LM1 prevailed was superior to that obtained by fixing the interest rate so that LM2 prevailed. But now suppose that instead of fixing the money stock, the money stock were reduced every time the interest rate went up and increased every time the interest rate went down. This procedure would, of course, increase the amplitude of interest rate fluctuations.5 But if the proper relationship between the money stock and the interest rate could be discovered, then the LM function could be made to look like LM0 in Figure 8. The result would be that income would be pegged at Yf . Disturbances in the IS function would produce changes in the interest rate, which in turn would produce spending changes sufficient to completely offset the effect on income of the initial disturbance. The most complicated case of all to explain graphically is that in which it is desirable to increase the money stock as the interest rate rises and decrease it as the interest rate falls. In Figure 9 the leftmost position of the LM function as a result of disturbances is LM1 when the money stock is fixed and is LM2 when the combination policy of introducing a positive money-interest relationship is followed. The rightmost positions of the LM functions under these conditions are not shown in the diagram. When the interest rate is pegged, the LM function is LM3. If either LM1 or LM2 prevails, the intersection with IS1 produces the lowest income, which is below the Y1′ level 5 The increased fluctuations in interest rates must be carefully interpreted. In this model the IS function is assumed to fluctuate around a fixed-average position. However, in more complicated models involving changes in the average position of the IS function, perhaps through the operation of the investment accelerator, interest rate fluctuations may not be increased by the policy being discussed in the text. By increasing the stability of income over a period of time, the policy would increase the stability of the IS function in Figure 8 and thereby reduce interest rate fluctuations. J U LY / A U G U S T 2008 453 Poole obtained with LM3. But in the case of LM2, income at Y1 is only a little lower than at Y1′, whereas when IS2 prevails, LM2 is better than LM3 by the difference between Y2 and Y2′. Since the gap between Y2 and Y2′ is larger than that between Y1 and Y1′, it is on the average better to adopt LM2 than LM3 even though the extremes under LM2 are a bit larger than under LM3. Extensions of Model. At this point a natural question is that of the extent to which the above analysis would hold in more complex models. Until more complicated models are constructed and analyzed mathematically, there is no way of being certain. But it is possible to make educated guesses on the effects of adding more goals and more policy instruments, and of relaxing the rigid price assumption. Additional goals may be added to the model if they are specified in terms of “closer is better” rather than in terms of a fixed target that must be met. For example, it would not be mathematically difficult to add an interest rate goal to the model analyzed above, if deviations from a target interest rate were permitted but were treated as being increasingly harmful. On the other hand, it is clear that if there were a fixed-interest target, then the only possible policy would be to peg the interest rate, and income stabilization would not be possible with monetary policy alone. The addition of fiscal policy instruments affects the results in two major ways. First, the existence of income taxes and of government expenditures inversely related to income (for example, unemployment benefits) provides automatic stabilization. In terms of the model, automatic stabilizers make the IS function steeper than it otherwise would be, thus reducing the impact of monetary disturbances, and reduce the variance of expenditures disturbances in the reduced-form equation for income. This effect would be shown in Figure 6 by drawing IS1 so that it cuts LM2 to the right of Y1′ and drawing IS2 so that it cuts LM2 to the left of Y2′. The second major impact of adding fiscal policy instruments occurs if both income and the interest rate are goals. Horizontal shifts in the IS function that are induced by fiscal policy adjustments, when accompanied by a coordinated mone454 J U LY / A U G U S T 2008 tary policy, make it possible to come closer to a desired interest rate without any sacrifice in income stability. An obvious illustration is provided by the case in which the optimal monetary policy from the point of view of stabilizing income is to set the interest rate as in Figure 5. Fiscal policy can then shift the pair of IS functions, IS1 and IS2, to the right or left so that the expected value of income is at the full employment level. If the interest rate is not a goal variable, then fiscal policy actions that shift the IS function without changing its slope do not improve income stabilization over what can be accomplished with monetary policy alone, provided the lags in the effects of monetary policy are no longer than those in the effects of fiscal policy. An exception would be a situation in which reaching full employment with monetary policy alone would require an unattainable interest rate, such as a negative one. These comments on fiscal policy have been presented in order to clarify the relationship between fiscal and monetary policy. While monetary policymakers may urge fiscal action, for the most part monetary policy must take the fiscal setting as given and adapt monetary policy to this setting. It must then be recognized that an interest rate goal can be pursued only at the cost of sacrificing somewhat the income goal.6 All of the analysis so far has taken place within a model in which the price level is fixed in the short run. This assumption may be relaxed by recognizing that increases in money income above the full employment level involve a mixture of real income gains and price inflation. Similarly, reductions in money income below the full employment level involve real income reductions and price deflation (or a slower rate of price inflation). The model used above can be reinterpreted entirely in terms of money income so that departures from what was called above the “full employment” level of income involve a mixture of real income and price changes. Stabi6 An interest rate goal must be sharply distinguished from the use of the interest rate as a monetary policy instrument. By a goal variable is meant a variable that enters the policy utility function. Income and interest rate goals might be simultaneously pursued by setting the money stock as the policy instrument or by setting the interest rate as the policy instrument. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole lizing money income, then, involves a mixture of the two goals of stabilizing real output and of stabilizing the price level. However, interpreted in this way the structure of the model is deficient because it fails to distinguish between real and nominal interest rates. Price level increases generate inflationary expectations, which in turn generate an outward shift in the IS function. The model may be patched up to some extent by assuming that price changes make up a constant fraction of the deviation of income from its full employment level and assuming further that the expected rate of inflation is a constant multiplied by the actual rate of inflation. Expenditures are then made to depend on the real rate of interest, the difference between the nominal rate of interest and the expected rate of inflation. The result is to make the IS function, when drawn against the nominal interest rate, flatter and to increase the variance of disturbances to the IS function. These elects are more pronounced: (a) the larger is the interest sensitivity of expenditures; (b) the larger is the fraction of price changes in money income changes; and (c) the larger is the effect of price changes on price expectations. The conclusion is that since price flexibility in effect increases the variance of disturbances in the IS function, a money stock policy tends to be favored over an interest rate policy. II. EVIDENCE ON THE RELATIVE MAGNITUDES OF REAL AND MONETARY DISTURBANCES Nature of Available Evidence Little evidence is available that directly tests the relative stability of the expenditure and money demand functions. It is necessary, therefore, to proceed somewhat indirectly. First, simulation of the FR-MIT model7 is used to show the probable size of the effect on gross national product (GNP), the GNP deflator, and the unemployment rate of an assumed expenditure disturbance. This evi- dence provides some indication of the extent to which the impact of an expenditure disturbance depends on the choice between the money stock and the Treasury bill rate as monetary policy control variables. This evidence bears only on the question of what happens if an expenditure disturbance occurs, not on the relative stability of the expenditure and money demand functions. However, this approach is useful when combined with intuitive feelings about relative stability. The second type of evidence, derived from reduced-form studies, is more directly related to the question of relative stability; nevertheless, it is not entirely satisfactory because the studies examined were not designed to answer the question at hand. To supplement these studies by other investigators, there follows a simple test of the stability of the demand for money function. Impact of an Expenditure Disturbance Simulation of the FR-MIT model provides some insight as to how the size of the impact of an expenditure disturbance depends on the choice of the monetary policy instrument. The simulation technique is necessary because the FR-MIT model is nonlinear, making it impossible to obtain an explicit expression for the reduced form.8 However, comparison of two sets of simulations provides some interesting results. Except as indicated below, the simulations all used the actual historical values of the model’s exogenous variables and all simulations started with 1962-I, a starting date selected arbitrarily. The first set of five simulations assumes an exogenous money stock that grows by 1 percent per quarter, starting with the actual money stock in 1961-IV as the base. To investigate the impact of a disturbance in an exogenous expenditures variable, the exogenous variable “federal expenditures on defense goods” was set in one simulation at its actual level minus $10 billion; in another at actual minus $5 billion; and in three further simulations at actual, actual plus $5 billion, and actual plus $10 billion. This procedure produces 8 7 For a general description of the model, see de Leeuw and Gramlich (1968). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W In a reduced-form equation, an endogenous (that is, simultaneously determined) variable is expressed as depending only on exogenous and predetermined variables (variables taken as given for the current period). J U LY / A U G U S T 2008 455 Poole Figure 10 Reduced-Form Parameter Estimates for Federal Defense Expenditures from FR-MIT Model SOURCE: Originally published version, p. 146. four hypothetical observations on “disturbances” in defense expenditures, of –10, –5, +5, and +10, and the simulation provides four corresponding observations for the change in income (and other endogenous variables). By using income as an example, the change in an endogenous variable in response to a disturbance in defense expenditures is the difference between income simulated by the model when defense expenditures were set at actual historical values and when set at actual plus 10, plus 5, and so forth. The income obtained in the simulations, even when defense expenditures are set at actual levels, is not the same as the actual historical level of income both because the assumed monetary policy differs from 456 J U LY / A U G U S T 2008 the policy actually followed and because of errors in the model itself. By calculating the ratio of the change in an endogenous variable to the disturbance in defense expenditures for the four observations, four estimates of the linear approximation to the reduced-form parameter, or multiplier, of defense expenditures are obtained, and these four estimates have been averaged to produce a single estimate. Since the effects of a disturbance accumulate over time, the reduced-form parameter estimate has been calculated for the 12 quarters from 1962-I through 1964-IV. Exactly the same procedure has been used for the simulations with a fixed rate for 3-month Treasury bills. Finally, the ratio of the parameter estimates for the reduced forms under the money stock and interest rate policies has been calculated with the parameter estimates from the simulations with the exogenous money stock in the numerator of the ratio. The reduced-form parameter estimates under the two monetary policies, and the ratios of these estimates, have been plotted in Figure 10 for 12 quarters for the reduced forms for nominal GNP, for the unemployment rate, and for the GNP deflator. The results are striking. A substantial difference appears in the parameters of reduced forms for the fourth quarter following the initial disturbance, and the differences in the parameters become steady thereafter. By the 12th quarter the reduced-form parameters for the money stock policy are only about 40 percent of those for the interest rate policy. The interpretation of these results is that employment, output, and the price level are far more sensitive to disturbances in defense expenditures under an interest rate policy than under a money stock policy. This conclusion presumably generalizes to expenditures variables other than defense expenditures, but the results would differ in detail because each expenditures variable enters the FR-MIT model in a somewhat different way. It might be argued that these results suggest that there is no significant difference between interest rate and money stock policies because the reduced-form parameters are essentially identical up to about four quarters. Surely, so this F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole argument goes, mistakes could be discovered and onset within four quarters. There are two difficulties with this argument. The first is that the FR-MIT model may overstate the length of the lags and therefore understate the differences in reduced-form parameters for the two policies for the quarters immediately following a disturbance. But the second and more important reason is that it may not be easy to reverse the effects of the disturbance after the disturbance has been discovered. With an interest rate policy, a very large change in the rate might be required to offset the effects appearing after the fourth quarter, and such a change might not be feasible, or at least not desirable in terms of its effects on security markets and on income in the more distant future. The numerical results reported above depend, of course, on the FR-MIT model, and this model is deficient in a number of respects. But any model in which, other things being equal, investment and other interest-sensitive expenditures decline when interest rates rise will show results in the same direction. These results may be extended to analyze the significance of errors in forecasting exogenous variables. Consider an explicit expression for the reduced form for income. Let the exogenous variables such as government expenditures, perhaps certain categories of investment, strikes, weather, population growth, and so forth, be X1, X2, …, Xn, and let the coefficients of these variables be α1, α2, …, αn when the interest rate is the policy instrument, and λ1, λ2, …, λn when the money stock is the instrument. Then the reduced form for income when the interest rate is the instrument is (1) Y = α 0 + α 1X 1 + α 2 X 2 + …+ α n X n + α r r + u where αr is the coefficient of the interest rate and u is the random disturbance. On the other hand, when the money stock is the instrument, the reduced form is (2) Y = λ0 + λ1X 1 + λ2X 2 +…+ λn X n + λM M + ν As discussed in Section II, the disturbance νt may have either a larger or a smaller variance than the disturbance ut . One factor tending to make νt smaller than ut is that a money stock policy F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W reduces the impact of expenditures disturbances, but another factor, the introduction into the reduced form of money demand disturbances, tends to make νt larger. The net result of these two factors cannot be determined a priori. But in formulating policy it is not possible to reason directly from equations 1 and 2 because many of the Xi cannot be predicted in advance with perfect accuracy. For scientific purposes ex post it may be possible to say that a change in income was caused by a change in some Xi; for policy purposes ex ante this scientific knowledge is useless unless the change in Xi can be predicted. It is necessary to think of each Xi as being composed of a predictable part, X̂i , and an unpredictable part, Ei. X i = Xˆ i + E i For policy purposes the error term in the reduced form includes both the disturbances to the equation and the errors in forecasting exogenous variables. The two types of errors ought to be treated exactly alike in formulating policy. Equations 1 and 2 can then be rewritten as follows: (3) Y = α 0 + α i Xˆ 1 + α 2Xˆ 2 +…+ α n Xˆ n + α r r + α 1E 1 + α 2 E 2 + …+ α n E n + u (4) Y = λ0 + λi Xˆ 1 + λ 2Xˆ 2 +…+ λn Xˆ N + λM M + λ1E 1 + λ 2 E 2 + …+ λn E n + ν For policy purposes the error term in the reducedform equation 3 is the sum of the terms from α1E1t through ut and in the reduced-form equation 4 the sum of the term λ1E1t through νt . A systematic study of the importance of the Ei terms cannot be made because no formal record of errors in forecasting exogenous variables exists insofar as the author knows. However, some insight into the problem may be obtained by listing the variables that must be forecast. Which variables have to be forecast depends, of course, on the model being used. The larger econometric models generally have relatively few exogenous variables that raise forecasting problems because so many variables are explained endogeneously by the model itself. The FR-MIT model has 63 J U LY / A U G U S T 2008 457 Poole exogenous variables; some of these are relatively easy to forecast, but others are subject to considerable forecasting error. The latter include such variables as exports, number of mandays idle due to strikes, armed forces, and federal expenditures. Furthermore, this model involves lagged endogenous variables in many equations; hence an inaccurate forecast of GNP next quarter will increase the error in forecasting GNP two quarters into the future, which in turn will lead to errors in forecasting GNP three quarters into the future, and so forth. Errors in forecasting exogenous variables, therefore, produce cumulative errors in forecasting GNP in future quarters. In simpler models the forecasting problem is more severe. Consider, for example, the opposite extreme from the large econometric model, the single-equation model. Convenient representatives of such models are those spawned in the controversy over the Friedman-Meiselman paper (1963) on the stability of the money/income relationship. The various definitions of exogenous, or “autonomous,” spending utilized by the various authors in this controversy are as follows : a) Friedman-Meiselman definition: Autonomous expenditures consist of the “net private domestic investment plus the government deficit on income and product account plus the net foreign balance” (Friedman and Meiselman, 1963, p. 184). b) Ando-Modigliani definition: Autonomous expenditures consist of two variables which enter the reduced form with different coefficients. One variable is “property tax portion of indirect business taxes” plus “net interest paid by government” plus “government transfer payment” minus “unemployment insurance benefits” plus “subsidies less current surplus of government enterprises” minus “statistical discrepancy” minus “excess of wage accruals over disbursement.” The second variable is “net investment in plant and equipment, and in residential houses” plus “exports” (Ando and Modigliani, 1965a, pp. 695-96, 702). c) DePrano-Mayer definition: The basic definition is “investment in producers’ durable 458 J U LY / A U G U S T 2008 equipment, nonresidential construction, residential construction, federal government expenditures on income and product account, and exports. One variant of this hypothesis subtracts capital consumption estimates, and the other does not” (DePrano and Mayer, 1965a, p. 739). DePrano and Mayer also tested 18 other definitions of autonomous expenditures (DePrano and Mayer, 1965a, pp. 739-40). d) Hester definition: Autonomous expenditures consist of the “sum of government expenditure, net private domestic investment, and the trade balance” (Hester, 1964a, p. 366). Hester also experimented with three other definitions involving alternative treatments of imports, capital consumption allowances, and inventory investment (Hester, 1964a, pp. 366-67). To a considerable extent the diversity in these definitions is misleading because except for the Friedman-Meiselman definition all the definitions are in fact rather similar. But whichever definition is used, it is impossible to escape the feeling that inaccurate forecasting of exogenous variables is likely to be a major source of uncertainty. And while this discussion has taken place within the context of formal models, exactly the same problem plagues judgmental forecasting. Every forecasting method can be viewed as starting from forecasts of “input,” or exogenous, variables and then proceeding to judge the implications of these inputs for GNP and other dependent, or endogenous, variables. Regardless of what type of model is used, it appears that for the foreseeable future it will be necessary to forecast exogenous variables that simply cannot be forecast accurately by using present methods. As a result, it seems very likely that the error term including forecast errors has a far smaller variance in equation 4 than in equation 3. Indeed, it might be argued that as a source of uncertainty the Ei terms are far more important than the u or ν terms, and therefore that the smaller size of the λi parameters as compared to the α i parameters is of great importance. If the parameter estimates from the FR-MIT model are accepted, the standard deviation of the total ranF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole dom term relevant for policy (that is, including errors in forecasting exogenous variables) would be over twice as large under an interest rate policy as under a money stock policy. If this argument is correct, shifting from the current policy of emphasizing interest rates to one of controlling the money stock might cut average errors in half, where errors are measured in terms of the deviations of employment, output, and price level from target levels for these variables. Evidence from Reduced-Form Equations Additional insight into the relative sizes of disturbances under interest rate and money stock policies may be obtained by examining the controversy generated by the Friedman-Meiselman paper on the stability of the money/income relationship (Friedman and Meiselman, 1963). In this paper equations almost the same as equations 1 and 2 above were estimated. The equation corresponding to equation 1 differs in that the exogenous variables were assumed to consist only of a single autonomous spending variable, as defined above. The equation corresponding to equation 2 has the same disability for our purposes, but it also did not include an interest rate as a variable. Before examining the implications of the Friedman-Meiselman findings for this study, it should be noted that their approach was sharply criticized in papers by Donald D. Hester (1964a), Albert Ando and Franco Modigliani (1965a), and Michael DePrano and Thomas Mayer (1965a). These critics particularly attacked the FriedmanMeiselman definition of autonomous expenditures, and proposed and tested the alternative definitions listed above. However, they also attacked the single-equation approach and recommended the use of large models instead. The tests of alternative equations must be regarded as inconclusive in terms of which variable—the money stock or autonomous spending— is more closely related to the level of income.9 9 For reasons that need not be explained here, most of this controversy was conducted in terms of equations with consumption rather than GNP as the dependent variable. In the FriedmanMeiselman study, however, results are reported for equations with GNP (Friedman and Meiselman, 1963, p. 227). Such results are also reported in Andersen and Jordan (1968, p. 17). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Both approaches achieve values for R 2 of 0.98 or 0.99 so that the unexplained variance is very small in both cases. It seems very unlikely that the addition of an interest rate variable to the equations by using autonomous expenditures as the explanatory variable, which addition would make the equations correspond to equation 1 above, would make any substantial difference. From this evidence it appears that ex post explanations of the level of income are about as accurate by using autonomous expenditures alone as are those by using money stock alone. But given the inaccuracies in forecasting autonomous expenditures, it must be concluded that ex ante explanations by using the money stock are substantially more accurate than those with forecasts of autonomous expenditures. From this evidence, the total random term in equation 4 appears to have a substantially smaller variance than the total random term in equation 3. For the reasons mentioned by the FriedmanMeiselman critics, evidence from single-equation studies cannot be considered definitive. But neither can the evidence be ignored, especially in light of the difficulties encountered in the construction and the use of large econometric models such as the FR-MIT model. Evidence on Stability of Demand for Money Function One of the shortcomings of the single-equation studies discussed above is that their authors paid too little attention to the stability of regression coefficients over time. Consider the following statement by Friedman and Meiselman: The income velocity of circulation of money is consistently and decidedly stabler than the investment multiplier except only during the early years of the Great Depression after 1929. There is throughout, including those years, a close and consistent relation between the stock of money and consumption or income, and between year-to-year changes in the stock of money and in consumption or income (Friedman and Meiselman, 1963, p. 186). This conclusion is based on correlation coefficients between money and income (or consumpJ U LY / A U G U S T 2008 459 Poole Figure 11 Velocity and Interest Rate Regressions (regressions fitted to quarterly data, 1947-60) SOURCE: Originally published version, p. 150. tion), but what is relevant for policy is the regression coefficient, which determines how much income will change for a given change in the money stock. In the Friedman-Meiselman study, a table (Friedman and Meiselman, 1963, p. 227) reports the regression coefficient for income on money as being 1.469 for annual data 1897-1958. However, the same table reports regression coefficients for 12 subperiods, some of which are overlapping, ranging from 1.092 to 2.399. With a few exceptions, most economists agree that velocity changes can be explained in part by interest rate changes.10 Thus, variability in the regression coefficients when income is regressed on money is not evidence of the instability of the demand for money function. To obtain 10 For a convenient review of evidence on this subject, see Laidler (1969). 460 J U LY / A U G U S T 2008 some evidence on the stability of this function, the following simple procedure was used. Quarterly data were collected on the money stock, GNP, and Aaa corporate bond yields for 1947 through 1968. A demand for money function was fitted by regressing the log of the interest rate on the log of velocity, and vice versa. The regressions were run for the four periods, 1947 through 1960, 1947 through 1962, 1947 through 1964, and 1947 through 1966. The results inside each estimation period were then compared with the results outside the estimation period. The results of this process for the 1947-60 estimation period are shown in Figure 11. The observations for 1947 through 1960 are represented by dots, and the observations for 1961 through 1968 by X’s. The two least-squares regressions—log interest rate on log velocity and vice versa—fitted for the 1947-60 period have been drawn. From Figure 11 it appears that the relationship since 1960 has been quite similar to the one prior to 1960. Table 1 presents the results of applying a standard statistical test to the regression and postregression periods to determine whether the demand for money function was stable. To understand this table, refer first to section A of the table, and to the 1947-60 estimation period. Section A reports results from regressing the log of velocity on the log of the Aaa corporate bond rate, and the first row refers to the regression for 1947 through 1960. The square of the regression’s standard error of estimate is 0.00517 with 54 degrees of freedom. There were 32 quarters in the postregression period 1961 through 1968, and for this period the mean-square error of velocity from the velocity predicted by the regression is 0.00836. The ratio of the mean-square errors from regression outside to those inside the estimation period is given in the column labeled “F.” Since the ratio of two mean squares has the F distribution under the hypothesis that both mean squares were produced by the same process, an F test may be used to test whether the demand for money function has been stable. If the function has been stable, then errors from regression outside the period of estimation F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole Table 1 Tests of the Stability of the Demand for Money Function by Using Quarterly Data Regression Estimation period (SEE)2 Progression d.f. MSE d.f. F Significance level .10 A. Log velocity regressed on log Aaa corporate bond yield 1947-60 .00517 54 .00836 32 1.62 1947-62 .00484 62 .00746 24 1.54 .10 1947-64 .00509 70 .00587 16 1.15 >.25 1947-66 .00502 78 .00986 8 1.96 .10 B. Log Aaa corporate bond yield regressed on log velocity 1947-60 .00684 54 .00589 32 1.16* >.25 1947-62 .00614 62 .00723 24 1.18 >.25 1947-64 .00570 70 .01162 16 2.04 .025 1947-66 .00537 78 .02192 8 4.08 .005 NOTE: *MSE < (SEE)2. should be, on the average, the same size as the errors inside the period of estimation. For the 1947-60 regression being discussed, F = 1.62 and is significant at the 10 percent level but not at the 5 percent level. Looking at Table 1 as a whole it can be seen that, for three of the regressions, the errors outside the period of estimation are not statistically significantly larger than those inside the period of estimation. Indeed, for the bond rate regression for the 1947-60 period, the errors outside the period of estimation were actually smaller, on the average, than those inside the period of estimation. Overall, however, these results taken at face value cast some doubt on the stability of the demand for money function. However, there is reason to believe that there are problems in applying the F test in this situation. The reason is that the residuals from regression exhibit a very high positive serial correlation as indicated by Durbin-Watson test statistics of around 0.15 for all of the regressions. What this means is that the effective number of degrees of freedom is actually less than indicated in the table, and with fewer degrees of freedom the F ratios computed have less statistical significance than F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W the significance levels reported in the table. The only way around this problem is to run a more complex regression that removes the serial correlation of the residuals, but there is no general agreement among economists as to exactly what variables belong in such a regression. The virtue of the simple regressions of velocity on an interest rate and vice versa is that this form has been used successfully by many investigators starting in 1954 (Latané, 1954). The appropriate conclusion to be drawn from this evidence would seem to be that the relationship between velocity and the Aaa corporate bond rate is too close and too stable to be ignored, but not close enough and stable enough to eliminate all doubts. However, the question is not whether an ironclad case for a money stock policy exists but rather whether the evidence taken as a whole argues for the adoption of such a policy. While there is certainly room for differing interpretations of Figure 11 and Table 1, and of the other evidence examined above, on the whole all of these results seem to point in the same direction. It appears that the money stock rather than interest rates should be used as the monetary policy control variable. J U LY / A U G U S T 2008 461 Poole III. A MONETARY RULE FOR GUIDING POLICY Rationale for a Rule-of-Thumb The purpose of this section is to develop a rule-of-thumb to guide policy. Such a rule—not meant to be followed slavishly—would incorporate advice in as systematic a way as possible. The rule proposed here is based upon the theory and evidence in Sections II and III and upon a close examination of post-accord experience. Individual policymakers inevitably use informal rules-of-thumb in making decisions. Like everyone else, policymakers develop certain standard ways of reacting to standard situations. These standard reactions are not, of course, unchanging over time, but are adjusted and developed according to experience and new theoretical ideas. If there were no standard reactions to standard situations, behavior would have to be regarded as completely random and unpredictable. The word “capricious” is often, and not unfairly, used to describe such unpredictable behavior. There are several difficulties with relying on unspecified rules-of-thumb. For one thing, the rules may simply be wrong. But an even more important factor, because formally specified rules may also be wrong, is that the use of unspecified rules allows little opportunity for cumulative improvements over time. A policymaker may have an extremely good operating rule in his head and excellent intuition as to the application of the rule but unless this rule can be written down there is little chance that it can be passed on to subsequent generations of policymakers. An explicit operating rule provides a way of incorporating the lessons of the past into current policy. For example, it is generally felt that monetary policy was too expansive following the imposition of the tax surcharge in 1968. Unless the lesson of this experience is incorporated into an operating rule, it may not be remembered in 1975 or 1980. How many people now remember the overly tight policy in late 1959 and early 1960 that was a result of miscalculating the effects of the long steel strike in 1959? Since the FOMC membership changes over time, many of the cur462 J U LY / A U G U S T 2008 rent members will not have learned firsthand the lesson from a policy mistake or a policy success 10 years ago. If the FOMC member is not an economist, he may not even be aware of the 10-yearold lesson. It is for these reasons that an attempt is made in this section to develop a practical policy rule that incorporates the lessons from past experience. The rule is not offered as one to be followed to the last decimal place or as one that is good for all time. Rather, it is offered as a guide—or as a benchmark—against which current policy may be judged. A rule may take the form of a formal model that specifies what actions should be taken to achieve the goals decided upon by the policymakers. Such a model would provide forecasts of goal variables, such as GNP, conditional on the policy actions taken. The structure of the model and the estimates of its parameters would, of course, be derived from past data and in that sense the model would incorporate the lessons of the past. But in spite of advances in modelbuilding and forecasting, it is clear that forecasts are still quite inaccurate on the average. In a study of the accuracy of forecasts by several hundred forecasters between 1953 and 1963, Zarnowitz concluded that the mean absolute forecast error was about 40 percent of the average year-to-year change in GNP (Zarnowitz, 1967, p. 4). He also reported, “there is no evidence that forecasters’ performance improved steadily over the period covered by the data” (Zarnowitz, 1967, p. 5). Not only are forecasts several quarters ahead inaccurate but also there is considerable uncertainty at, and after, the occurrence of businesscycle turning points as to whether a turning point has actually occurred. In a study of FOMC recognition of turning points for the period 1947-60, Hinshaw concluded that (Fels and Hinshaw, 1968, p. 122): The beginning data of the Committee’s recognition pattern varied from one to nine months before the cyclical turn…On the other hand, the ending of the recognition pattern varied from one to seven months after the turn…With F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole the exception of the 1948 peak, the Committee was certain of a turning point within six months after the NBER date of the turn. At the date of the turn, the estimated probability was generally below 50; it reached the vicinity of 50 about two months after the turn. This recognition record, which is as good as that in 10 widely circulated publications whose forecasts were also studied in (Friedman and Meiselman, 1963) casts further doubt on the value of placing great reliance on the forecasts.11 Given the accuracy of forecasts at the current state of knowledge,12 it seems likely that for some time to come forecasts will be used primarily to supplement a policy-decisionmaking process that consists largely of reactions to current developments. Only gradually will policymakers place greater reliance on formal forecasting models.13 While a considerable amount of work is being done on such models, essentially no attention is being paid to careful specification of how policy should react to current developments. While sophisticated models will no doubt in time be developed into highly useful policy tools, it appears that in the meantime relatively simple approaches may yield substantial improvements in policy. Given that knowledge accumulates rather slowly, it can be expected that carefully specified but simple methods will be successful before large-scale models will lie. Careful specification of policy responses to current developments is but a small step beyond intuitive policy responses to current developments. This step surely represents a logical evolution of the policyformation process. 11 For further analysis of forecasting accuracy, see Mincer (1969). 12 The accuracy of forecasts may now be better than in the periods examined in the studies cited above. But without a number of years of data there would be no way of knowing whether forecasts have improved, and so forecasts must in any case be assumed to be subject to a wide margin of error at the present time. 13 It may be objected that great reliance is already placed on forecasts, at least on judgmental forecasts. However, these forecasts typically involve a large element of extrapolation of current developments. It seems fair to say that in most cases in which conditions forecast a number of quarters ahead differ markedly from current conditions, policy has followed the dictates of current conditions rather than of the forecasts. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Post-Accord Monetary Policy That an operating guideline is needed can be seen from the experience since the Treasury– Federal Reserve accord. In order that this experience may be understood better, subperiods were defined in terms of “stable,” “easing,” or “firming” policy as determined from the minutes of the Federal Open Market Committee. The minutes used are those published in the Annual Reports of the Board of Governors of the Federal Reserve System for 1950 to 1968. The definitions of “stable,” “easing,” and “firming” periods are necessarily subjective as are the determinations of dates when policy changed.14 The dating of policy changes was based primarily on the FOMC minutes, although the dates of changes in the discount rate and in reserve requirements were used to supplement the minutes. “Stable” periods are those in which the policy directive was unchanged except for relatively minor wording changes. In some cases the directive was essentially unchanged although the minutes reflected the belief that policy might have to be changed in the near future. While the Manager of the System Open Market Account might change policy somewhat as a result of such discussions, the unchanged directive was taken at face value in defining policy turning points. More difficult problems of interpretation were raised by such directives as “unchanged policy, but err on the side of ease,” or “resolve doubts on the side of ease.” Such statements were used to help in defining several periods during which policy was progressively eased (or tightened). For example, in one meeting the directive might call for easier policy, the next meeting might call for unchanged policy but with doubts to be resolved on the side of ease, and a third meeting might call for further ease. These three meetings would then be taken together as defining an 14 The author was greatly assisted in these judgments by Joan Walton of the Special Studies Section of the Board’s Division of Research and Statistics. Miss Walton, who is not an economist, carefully read the minutes of the entire period and in a large table recorded the principal items that seemed important at each FOMC meeting. Having a noneconomist read the minutes tempered the inevitable tendency for an economist to read either too much or too little into the minutes. However, the final interpretation of the minutes rested with the author. J U LY / A U G U S T 2008 463 Poole “easing” period. However, unless accompanied by other FOMC meetings clearly calling for a policy change, statements such as those calling for an “unchanged policy with doubts resolved on the side of ease” were interpreted as not calling for a policy change. Some important monthly economic time series for the post-accord period are plotted in Figure 12. The heavy vertical lines represent periods of “stable,” “easing,” and “firming” policy as indicated by “S,” “E,” and “F” at the bottom of the figure. Except for the unemployment rate, the average of each series for each policy period has been plotted as a horizontal line. The two features of the post-accord experience are especially noteworthy. First, decisions to change policy have been taken about as close to the time when, in retrospect, policy changes were needed as could be expected in the light of existing knowledge.15 There have been mistakes in timing, but the overall record is impressive. The second major feature of this period is that policy actions, as opposed to policy decisions, have been in the correct direction if policy actions are defined by either free reserves or interest rates, but not if policy actions are defined in terms of either the money stock or bank credit. To examine the timing question in more detail, a useful comparison is that between business cycle turning points (as defined by the National Bureau of Economic Research) and decisions to change policy. The post-accord period begins at a time when the U.S. economy was beset by inflation stemming from the war in Korea. the dates of the principal changes in policy and of the business cycle peaks and troughs are listed in Table 2. The policy dates are those that define the beginning of the “stable,” “easing,” and “firming” periods indicated in Figure 12. The decision to ease policy was made prior to the business cycle peaks of July 1953 and May 1960. The decision in 1957 was made in the fourth month following the cycle peak in July, but as can be seen from Figure 12, the unemployment rate had not risen very much through October. Given 15 For additional views on the timing of Federal Reserve decisions, see Brunner and Meltzer (1964) and Fels and Hinshaw (1968). 464 J U LY / A U G U S T 2008 the amount of uncertainty always present in interpreting business conditions, this lag must be considered to be well within the margin of error to be expected for stabilization policy. However, the easing policy decision in 1968 was clearly a mistake in retrospect but not in prospect given the expectations held by the majority of economists that the tax increase would significantly temper the economic boom. Firming policy decisions were also generally well timed. Following the 1953-54 recession, decisions to firm policy in small steps were taken from December 1954 to September 1955, as unemployment declined to about 4 percent of the labor force. During the recovery period after the 1957-58 recession, firming decisions were taken from July 1958 to May 1959. There was also a series of firming decisions taken from the end of 1961 to 1966. Especially noteworthy are those taken from December 1965 to August 1966, in response to the beginning of inflation associated with the escalation of military activity in Vietnam. The easing policy decisions taken in late 1966 and early 1967 were fully appropriate in light of the economic slack that developed in 1967. Even from the point of view of those who doubt the importance of fiscal policy, this record of the timing of policy decisions in the post-accord period is remarkably good. The timing record does not suggest that much attention was paid to forecasts, but this lack of attention was perhaps not unfortunate given the accuracy of forecasts during the period. From this point of view, the only real mistake was the easing decision taken in 1968. Of course, those who believe that a steady rate of growth of the money stock is better than any discretionary policy likely to be achieved in practice may read this record as supporting their thesis. But the post-accord record of the timing of policy decisions is certainly encouraging to those who believe that the lags in the effects of policy are short enough, and the effects predictable enough, to make discretionary monetary policy a powerful stabilization tool if only decisions can be made promptly. While the System’s performance in the timing of policy decisions has been commendable, the same cannot be said for the actions taken in F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole Table 2 Dates of Principal Monetary Policy Decisions and of Business Cycle Peaks and Troughs Business cycle Turning point FOMC policy decisions Date Policy Starting date Accord 1951 March 1-2 Firming 1952 September 25 Stable Peak 1953 July Trough 1954 August Peak 1957 July Trough 1958 April Peak 1960 May Trough 1961 February Easing December 8 1953 Stable December 15 Firming 1954 December 11 Stable 1955 October 4 Easing 1957 November 12 Stable 1958 April 15 Firming July 29 Stable 1959 June 16 Easing 1960 March 1 Stable Firming August 16 1961 Stable Firming October 24 November 14 1962 June 19 Stable July 10 Firming December 18 Stable 1963 Firming January 8 May 7 Stable August 20 Firming 1964 August 18 Stable 1965 March 2 Firming Stable December 14 1966 Easing Stable Stable September 13 November 1 1967 Firming May 2 November 27 1968 April 30 Easing July 16 Stable August 13 Firming December 17 Stable F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W June 11 1969 April 29 J U LY / A U G U S T 2008 465 Poole Figure 12 Post-Accord Monetary Policy SOURCE: Originally published version, p. 154. 466 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole Figure 12, cont’d Post-Accord Monetary Policy SOURCE: Originally published version, p. 155. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 467 Poole response to the decisions. In the earlier discussion the purposely vague terms “easing,” “firming,” and “stable” were used to describe policy decisions. These terms were meant to convey the notions that policymakers wanted, respectively, to accelerate, decelerate, or maintain the pace of economic advance. The question that must now be examined is whether policy actions did in fact tend to accelerate, decelerate, or maintain the level of economic activity. Policy actions were in accord with policy decisions if these actions are measured by either the 3-month Treasury bill rate or free reserves. The bill rate rose in “firming” periods, fell in “easing” periods, and tended to remain unchanged in “stable” periods. However, there was some tendency for the bill rate to rise in “stable” periods following “firming” periods, and to fall in “stable” periods following “easing” periods, a pattern not inconsistent with the interpretation of policy being offered in this study. Similar comments apply to free reserves. But the picture is quite different if policy actions are measured by the rate of growth of the money stock. Careful study of Figure 12 will make this point clear. The growth rate declined in response to the “firming” policy decision in late 1952, and again in the “stable” period in early 1953. This behavior was, of course, consistent with the “firming” decision. But the rate of growth declined further following the “easing” decision in June 1953 and remained low until the middle of 1954. The unemployment rate rose rapidly from its low of 2.6 percent at the cycle peak in July 1953 to 6.0 percent in August 1954, the cycle trough; the money stock was at the same level in April 1954, 9 months following the cycle peak and 10 months following the decision to adopt an “easing” policy, as it had been at the peak. The same pattern that had appeared during the 1953-54 recession appeared again at the time of the 1957-58 recession. The rate of growth of the money stock declined in 1957 prior to the cycle peak. (The Treasury bill rate also rose substantially.) But after the decision to adopt an “easing” policy in November 1957, the growth rate of the money stock declined further. From October 1957 to January 1958, the money stock 468 J U LY / A U G U S T 2008 fell at a 2.9 percent annual rate; from the cycle peak in July to October it had fallen at a 1.5 percent annual rate. The rate of growth of the money stock increased substantially in February 1958, and it remained at the higher level during the “stable” policy period April to July. There followed a period of “firming” policy decisions from the end of July 1958 to May 1959; however, the average growth rate of the money stock during this period was virtually identical to the average in the preceding “stable” period. But in the “stable” period from June 1959 to February 1960, the rate of growth of money, at –2.2 percent, was much lower than in the preceding “firming” period. This rate of growth of money can hardly be considered appropriate in the light of the fact that except for one month the unemployment rate was continuously above 5 percent. However, the picture was confused by a long steel strike. The decision to ease policy was taken on March 1, 1960, but the rate of growth of the money stock remained negative until July. The rate of growth of money fell following the “firming” policy decisions of October 1961 and June 1962. In spite of another firming decision in December 1962 the rate of growth then increased, and it continued to rise during the “firming” period in 1963, maintaining the same rate in the following “stable” period. In August 1964, another “firming” decision was taken, and the growth rate trended down during the “firming” period from August 1964 to February 1965. During the “stable” period from March to November 1965, the Vietnam war heated up. In the second half of 1965 the growth rate of money was 6.1 percent compared with 3.0 percent during the first half. The “firming” policy decision came in December, but the rate of growth of money averaged over 6 percent for the months December through April 1966. At this point monetary growth ceased. In January 1967 the money stock was actually less than in May 1966—there having been no increase in the growth rate in the months immediately following the “easing” decision of November 1, 1966. The growth rate of money then accelerated during the “stable” period from May through F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole October 1967; for the period as a whole growth averaged 8.7 percent. In the following “firming” period November 1967 through April 1968, the rate of growth of the money stock was lower but it was still relatively high at 5.1 percent. The growth rate then rose to 9.6 percent in the “stable” period May through July 1968 and thereafter fell to a little less than 6 percent in the July-November 1968 period following the “easing” decision of July 16, 1968. There ensued a “firming” period from December 1968 through April 1969. Although original figures indicated that monetary growth was relatively little during this period, a revision in the money stock series showed that the rate averaged 5.5 percent for the period as a whole. The rate following April was lower, especially in the June-December 1969 period, which saw no net growth in the money stock. A broadly similar view of the timing of policy actions is obtained from a careful examination of the rate of growth of total bank credit. However, as shown in Figure 12, this series is quite erratic and much more difficult to interpret than the series on the rate of growth of the money stock. The proper way to interpret these results would seem to be as follows. When interest rates fell in a recession, policy was easier than it would have been if interest rates had not been permitted to fall. But if the money stock was also falling, or growing at a below-average rate, policy was tighter than it would have been had money been growing at its long-run average rate. Similar statements apply to rising interest rates and aboveaverage monetary growth in a boom. A Monetary Rule Given the arguments of Sections I and II on the advantages of controlling the money stock as opposed to interest rates, a logical first step in developing a policy guideline is to examine cases clearly calling for ease or restraint. Consider first a recession. To insure that monetary policy is expansionary, the rule might be that interest rates should fall and the money stock should rise at an above-average rate. This policy avoids two possible errors. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W The first is illustrated in Figure 13. If the IS function shifts down from IS1 to IS2 while the LM function shifts from LM1 to LM2, the interest rate will fall from r1 to r2. The shift from LM1 to LM2 could be caused by a shift in the demand for money with the stock of money unchanged. But this shift could also be caused by a decline in the stock of money, perhaps because of an attempt by policymakers to keep the interest rate from falling too rapidly. However, in terms of income it is clearly better to permit the interest rate to fall to r3 by maintaining the stock of money fixed, and better yet to shift the LM function to the right of LM1 by increasing the stock of money. The point is the simple one that monetary policy should not rely simply on a declining interest rate in recession but should also insure that the money stock is growing at an adequate rate. The LM function may still shift to LM2 in spite of monetary growth because of an increased demand for money; without the monetary growth, however, this shift in the demand for money would push the LM function to the left of LM2 and income would be even lower. The second type of error avoided by the proposed policy rule is illustrated in Figure 14. Again, it is assumed that the situation is one of recession. With a fixed money stock, an increase in the demand for money will shift the LM function from LM1 to LM2, tending to reduce income. However, if the interest rate is prevented from rising above r1, the increased demand for money is met by an increased supply of money. Maintaining monetary growth and a declining interest rate in recession insures that the contribution of monetary policy is expansive. Increases in the demand for money, unless accompanied by a falling IS function, are fully offset by preventing increases in the interest rate. The greater the fall in the IS function the smaller the offset to an increased demand for money. However, in no case should a fall in the IS function be permitted to cause a fall in the money stock. The policy proposed does not, of course, guarantee an expansion of income. No such guarantee is possible because downward shifts in the IS function may exceed any specified shift in the LM function. But more important than theoretical J U LY / A U G U S T 2008 469 Poole Figure 13 Figure 14 SOURCE: Originally published version, p. 159. SOURCE: Originally published version, p. 159. possibilities are empirical probabilities. For all practical purposes the problem is not how to insure expansion in a recession but how to trade off the risks of too much expansion against too little. The discussion of Figures 13 and 14 was entirely in terms of encouraging income expansion, or limiting further declines, in the face of depressing disturbances. But disturbances may be expansionary in a recession, and such disturbances may combine with expansionary policy to create overly rapid recovery from the recession. Consider again Figure 13, but suppose the initial position is as shown by IS2 and LM2. If the interest rate is not permitted to rise, a shift to IS1 will lead to a large increase in income to the level given by the intersection of IS1 with a horizontal LM function drawn at r2. This situation can be avoided only if the interest rate is permitted to rise. The natural question is how the interest rate can be permitted to rise within a recession policy of pushing the interest rate down and maintaining above-average monetary growth. The answer is that the recession policy should be followed 470 J U LY / A U G U S T 2008 only if the interest rate can be kept from rising with a monetary growth rate below some upper bound. Exactly the same analysis running in reverse applies to a policy for checking an inflationary boom. In a boom interest rates should rise and monetary growth should be below average. However, there must be a lower limit on monetary growth to avoid an unduly contractionary policy. Having presented the basic ideas behind the formulation of a monetary rule, it is now necessary to become more specific about the rule. After specifying the rule in detail, it will be possible to discuss the considerations behind the specific numbers chosen. The proposed monetary policy rule-of-thumb is given in Table 3. The rule assumes that full employment exists when unemployment is in the 4.0 to 4.4 percent range and that monetary growth in the 3 to 5 percent range is consistent with price stability. At full employment the Treasury bill rate may rise or fall, either because of market pressures or because of small adjustments in monetary F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole Table 3 Proposed Monetary Policy Rule-of-Thumb (Percent) Rule for month* Unemployment rate previous month Direction of Treasury bill rate (3-month) Growth of money stock (annual rate) 0-3.4 Rising 1-3† 3.5-3.9 Rising 2-4† 4.0-4.4 Rising or falling 3-5 4.5-4.9 Falling 4-6‡ 5.0-5.4 Falling 5-7‡ 5.5-5.9 Falling 6-8‡ 6.0-100.0 Falling 6-8 NOTE: *The 3-month bill rate is to be adjusted in the indicated direction provided that monetary growth is in the indicated range. If the bill rate change cannot be achieved within the monetary growth rate guideline, then the bill rate guideline should be abandoned. †If the bill rate the previous month was below the bill rate 3 months prior to that, then the upper and lower limits on monetary growth are both increased by 1 percent. ‡If the bill rate the previous month was above the bill rate 3 months prior to that, then the upper and lower limits on monetary growth are both reduced by 1 percent. policy; however, monetary growth should remain in the 3 to 5 percent range. When unemployment drops below 4 percent, the rule calls for a restrictive monetary policy. The bill rate should rise and monetary growth should be reduced. If the bill rate and monetary growth guidelines are not compatible, then the monetary guideline should be binding. For example, suppose that unemployment is in the 3.5 to 3.9 percent range. If monetary growth below 2 percent would be required to obtain a rising bill rate, then monetary growth should be 2 percent and the bill rate be permitted to fall. If this situation persists so that the bill rate falls for several months in spite of the low monetary growth, then the limits on monetary growth should be increased as indicated in footnote 2 to Table 3. The reason for this prescription is that the bill rate on the average turns down 1 month before the peak of the business cycle (Holt, 1962, p. 111). Unemployment, on the other hand, may increase relatively little in the early months following a cycle peak. Tying monetary growth to the bill rate in the way indicated in footnote 2 of Table 3 produces a more timely adjustment of policy than relying on the unemployment rate alone. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W The proposed rule calls for a falling bill rate and a relatively higher rate of monetary growth as unemployment rises above the 4.0 to 4.4 percent range. The rule for high unemployment situations calls for adjusting the monetary growth rate downward when the bill rate is consistently rising as indicated by footnote 3 to Table 3. The reasoning behind this adjustment is exactly parallel to the reasoning above for low unemployment situations. The proposed monetary rule has the virtues of simplicity and dependence on relatively wellestablished economic doctrine. Because of its simplicity, the basic ideas behind the rule can be explained to the noneconomist. The simplicity of the rule also will make possible relatively easy evaluations of the rule’s performance in the future if the rule is followed. With more complicated rules it would be much more difficult to know how to improve the rule in the future because it would be difficult to judge what part of the rule was unsatisfactory. Since, as has been repeatedly emphasized above, the rule is not proposed as being good for all time, it is best to start with a simple rule and then gradually to introduce more variables into the rule as experience accumulates. J U LY / A U G U S T 2008 471 Poole In designing the rule, the attempt was made to base the rule on fairly well-established economic knowledge. There is, of course, a great deal of debate as to just what is and what is not well established. What can be done, and must be done, is to explain as carefully as possible the assumptions upon which the rule is based, with full recognition that other economists may not accept these assumptions. First, the evidence for the importance of money is impressive. It seems fair to say that very few economists believe today that changes in the stock of money have nothing to do with business fluctuations. Rather, the argument is over the extent to which monetary factors are important. Some no doubt will feel that the 2-percentagepoint ranges on monetary growth specified by the rule are excessively narrow; however, it should be noted that a 4 percent growth rate is double a 2 percent growth rate. Also important is the fact that the rule is meant to serve as a guideline rather than be absolutely binding. Since policy should deviate from the rule if there is good and sufficient reason—such as wartime panic buying—a further element of flexibility exists within the framework of the rule. The rule is specified in terms of changes in the bill rate and the monetary growth rate, with the monetary growth rate being tied to the unemployment rate and to changes in the bill rate in the recent past. This formulation has been designed to avoid what seem to be the most obvious errors of the past. Over the years the monetary growth rate has been lowest at business cycle peaks and in the early stages of business contractions, and highest at cycle troughs and in the middle stages of business expansions. The highest rate of monetary growth since the Treasury–Federal Reserve accord has been during the inflation associated with escalation of military operations in Vietnam. For purposes of smoothing the business cycle, so far as this author knows, there is no theory propounded by any economist that would call for high monetary growth during inflationary booms and low monetary growth during recessions. Such behavior of the money stock could only be optimal within a theory in which money had little or no effect on business fluctuations and in which other 472 J U LY / A U G U S T 2008 goals such as interest rate stability were important. Being based on the unemployment rate and bill rate changes in the recent past, the proposed monetary rule does not rely on forecasting. Nor does the rule depend on the current and projected stance of fiscal policy. Both of these factors ought to be included in applying the rule by adjusting the rate of growth of the money stock within the rule limits, or even by going outside the limits. But given the accuracy of economic forecasts under present methods, and given the current uncertainty over the size of the impact of fiscal policy (not to mention the hazards in forecasting federal receipts and expenditures), it does not appear that these variables can be systematically incorporated into a rule at the current state of knowledge. Tests of the Proposed Rule Three types of evidence on the value of the rule are examined below. The first approach involves a simple comparison of the rule with the historical record to show that the rule would generally have been more expansionary (contractionary) than actual policy when actual policy—in the light of subsequent economic developments—might be judged to have been too contractionary (expansionary). The second approach examines the cyclical behavior of the estimated residuals from a simple demand for money function to show that it is unlikely that the proposed rule would interact with the disturbances to produce an excessively inflationary or deflationary impact. Both these approaches are deficient because they rely heavily on the historical record, a record that would have been quite different had the rule been followed in the past. To avoid this difficulty, a third approach uses simulation of the FR-MIT model, but the results do not appear very useful because of shortcomings in this model. An Impressionistic Examination of the Rule. Broadly speaking, the results of comparing the rule with the historical record since the Treasury– Federal Reserve accord in March 1951 are these. The rule would have provided a substantially tighter monetary policy than the actual during the inflationary period from the accord until about September 1952. At that point, actual F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole policy as measured both by the rate of growth of the money stock and by the 3-month bill rate became considerably tighter. In the last quarter of 1952, actual policy was in accord with the rule, but thereafter it tightened even further. In the 9 months following the cyclical peak in July 1953, the money stock had a zero rate of growth while the unemployment rate rose from 2.6 percent to 5.9 percent. Under the rule the rate of growth of the money stock would never have gone below 1 percent and would have steadily increased as unemployment rose. Actual policy became more expansive in the second quarter of 1954, and the cycle trough was reached in August. However, the rule would have been considerably more expansive, and it would have remained more expansive than the actual all through the 1955-56 boom. Inasmuch as the unemployment rate remained near 4.0 percent from May 1955 through August 1957, the rule would have been too inflationary during this period. However, it can be argued that monetary policy was overly restrictive before the cycle peak in July 1957, since in the year prior to the peak the money stock grew only by 0.7 percent. Less subject to dispute is the fact that policy was far too restrictive after the peak; in the 6 months following the peak the money stock fell at an annual rate of 2.2 percent, and at the same time the unemployment rate rose from 4.2 percent to 5.8 percent. The rule would have been considerably more expansive all during the high unemployment period of 1958-59, and it would have prevented the declines in the money stock in late 1959 and early 1960. At the peak in May 1960 the unemployment rate was 5. 1 percent, and the money stock had fallen by 2.1 percent in the previous 12 months. Unlike the periods following peaks in 1954 and 1957, policy became more expansive immediately after the May 1960 peak, although not so expansive as called for by the proposed rule. From the trough in February 1961 through June 1964, the unemployment rate never declined below 5 percent. Under the rule, policy would have been more expansive than the actual policy F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W followed throughout this period, especially as compared with the March-September 1962 period, during which the money stock fell slightly. Unemployment fell rapidly in 1965 with the Vietnam build-up; the rule would have been more expansive than actual through July 1965 and then less expansive than actual through April 1966. Indeed, in the 9-month period prior to April 1966, with the unemployment rate falling from 4.4 percent to 3.8 percent, monetary growth accelerated to a 6.6 percent annual rate; the proposed rule would have first called for monetary growth in the 3 to 5 percent range, and then in the 2 to 4 percent range starting in February 1966, following the drop in the unemployment rate below 4.0 percent in January. Finally, the negative growth rates of money in the 1966 credit crunch would have been avoided under the rule, as would the high rates of growth in 1967 and 1968. This impressionistic look at the proposed rule may be supplemented by a simple scoring system for judging when the rule would have been in error. For each month during the sample period it was determined whether the rule would have been more or less expansive than the actual policy, or about the same as the actual policy. The unemployment rate 12 months from the month in question was used to indicate whether or not the policy was correct, with a desired range of unemployment of 4.0 to 4.4 percent. The rule was deemed to have made an error if: (1) the actual policy was in accord with the rule, but unemployment 12 months later was not in the desired range; (2) the rule called for a more expansive policy than the actual, and unemployment 12 months later was below the desired range; and (3) the rule called for a less expansive policy than the actual, and unemployment 12 months later was above the desired range. Since the latest data used in this analysis were for July 1969, comparison of the rule with actual policy ends July 1968. Starting the sample with 1952, the first full year after the accord, provides a total of 199 months. Based on the criterion described above, the rule would have been in error in 63 months. If the criterion is changed by substituting the unemployment rate 9 months J U LY / A U G U S T 2008 473 Poole ahead instead of 12 months ahead, the rule has 62 errors; using the unemployment rate 6 months ahead yields 59 errors. Some of these errors are of negligible import. For example, in March 1953 the rule calls for a money growth rate of 2 to 4 percent, but the actual was 1.9 percent. Thus, the rule would have been more expansive than the actual this particular month, a mistake since unemployment was too low and inflation too high during this period. However, the rule would have been less expansive than actual in every one of the preceding 6 months and in all but one of the 6 months following this “mistake.” Except for scattered errors such as the one just discussed, most of the rule errors occurred in two separate periods. The first is the 2-year period following the cycle trough in August 1954, during which time the rule would have been too expansive. The second is the last half of 1964 and the first half of 1965, when the rule would have been too expansive in light of the subsequent sharp decline in unemployment. Unless one has completed a careful examination of the data, there is a tendency to underestimate how rapidly the economy can change. For example, from the cycle peak in July 1953 to the cycle trough 13 months later, the unemployment rate rose by 3.4 percentage points; and from the peak in July 1957 to the trough 9 months later in April 1958, it rose by 3.2 percentage points. Changes in the other direction have tended to be somewhat less rapid, but significant nonetheless. In the year following the trough in August 1954, the unemployment rate declined 2.0 percentage points, and it declined 2.2 percentage points in the year following the trough in April 1958. In January 1965 unemployment was 4.8 percent and the problem was still one of how to reach full employment. A year later the rate was 3.9 percent and the problem was inflation. Thus, it appears that for the most part the rule would have been superior to policy actually followed. Of course, the rule is not infallible and would have erred on a number of occasions. But in spite of these errors—and it should be recognized that some errors are inevitable no matter what rule or which discretionary policymakers are in charge—the proposed rule has the great 474 J U LY / A U G U S T 2008 virtue of turning policy around promptly as imbalances develop. Relationship of the Rule to Monetary Disturbances. Since the rule was developed on the basis of the theoretical and empirical analysis of Sections I and II, which emphasized the relative stability of the demand for money, it is appropriate to conduct a systematic examination of the disturbances in the demand for money. It will be recalled that the rule was formulated in such a way as to insure expansionary policy action in a recession and contractionary policy action in a boom. However, it was recognized that disturbances in the expenditure sector and/or in the monetary sector might reinforce policy actions leading to an excessively expansionary or contractionary effect on income. If there were a significant chance of these excessive effects occurring, then the rule proposed would be overly “aggressive” and a rule involving a smaller range of monetary growth rates would be in order. To provide some evidence of the effect of disturbances in the money demand function, the residuals from the simple velocity function tested in Section II were examined carefully. The technique involved regressing velocity on the Aaa corporate bond rate, and vice versa, for the 1947-68 period and then comparing the residuals with turning points in the business cycle. The reader may make these comparisons visually from Figure 15. At the bottom of this figure cycle peaks and troughs are identified by “P” and “T,” respectively. The residuals from the estimated equations suggest that the demand for money has contractionary disturbances near business cycle peaks and expansionary disturbances near cycle troughs. The residuals have the same turning points for the regression of velocity on the interest rate as for the regression of the interest rate on velocity. The residual peaks occur at or before the cycle peaks, while the residual troughs occur at or after the cycle troughs. To assess the significance of these endings, consider the following simple view as to the dynamics of monetary elects. In the short run, income is a predetermined variable in the demand for money function. An increase in the money F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole Figure 15 Residuals from Velocity Regression Compared with Business-Cycle Turning Points SOURCE: Originally published version, p. 164. stock makes the interest rate lower than it would be otherwise, and this eventually leads to expansion in investment and income. A downward disturbance in the demand for money function has the same effect. Given this view of monetary dynamics, Figure 15 suggests the following conclusions. Shifts in the demand for money tend to be contractive in their effect on income in the late stages of a business cycle expansion, implying that a restrictive monetary policy must not be pushed too hard. Then, shortly before the cycle peak, the shifts apparently tend to become expansive. This effect is fortunate since it is only after the cycle peak that rising unemployment would trigger a F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W policy change under the proposed rule. However, there appears to be little danger that the rule would be overly expansionary because after the cycle trough, while policy is still expansionary, contractive shifts in the demand for money occur. Simulations of the FR-MIT Model. The final technique used to test the proposed monetary rule was to simulate the FR-MIT model under the rule. As explained below, the results are of questionable value but are presented anyway for the sake of completeness and in order not to suppress results unfavorable to the proposed rule. To simplify the computer programming, the rule used in the simulations is not exactly the same as the one proposed in Table 3 above. The J U LY / A U G U S T 2008 475 Poole Figure 16 Simulations of Unemployment in FR-MIT Model SOURCE: Originally published version, p. 165. 476 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole proposed rule, it will be recalled, involved a bill rate guideline and a money stock guideline. If, for example, the bill rate cannot be pushed up without pushing monetary growth below the lower limit in the money guideline, the proposed rule calls for setting monetary growth at its lower limit. The simulation rule, on the other hand, ignores the bill rate guideline and simply sets the monetary growth rate at the midpoint of the range specified by the proposed rule. Another difference, and no doubt a more important one, between the proposed rule and the simulation rule is that the simulation rule had to be specified in terms of quarterly data since the FR-MIT model uses quarterly data. In the simulation rule, the growth rate of the money stock depends on the level of unemployment determined by the model in the previous quarter. The growth rate of the money stock was modified by past changes in the bill rate, as in footnotes 2 and 3 to Table 3, except that the relevant bill rate change was in terms of the previous quarter before that. The simulation rule, then, reacts somewhat more slowly to unemployment trends than does the proposed rule. In order to investigate the importance of the starting point, simulations were run with starting dates in the first quarters of 1956, 1958, 1960, 1962, and 1964. The simulated unemployment rate for the five simulations is shown in the five panels of Figure 16 by the curves marked “S.” The actual unemployment rate is shown by the curves marked “A” and control simulations, to be explained below, by the unconnected points. It is clear from Figure 16 that the simulation rule for money growth produces an unstable unemployment rate. However, because of deficiencies in the model this result is probably not very meaningful. That the model is defective can be seen by comparing unemployment in the control simulations with the actual unemployment. In the control simulations all of the model’s exogenous variables, including the money stock, were set at their actual levels.16 Even with the exoge16 The FR-MIT model was estimated with the money stock as an endogenous variable. There are separate equations for currency and demand deposits, both of which are endogenous, while unborrowed reserves are exogenous. In the simulations the money stock F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W nous variables set at their actual levels, the simulated level of unemployment at times differs from the actual level. Because of the role of the stochastic disturbances in the model, especially as they feed through lagged endogenous variables, it cannot be expected that control simulations will exactly duplicate the actual results. But the fact that the control simulations differ from the actual by considerable margins over long periods of time strongly suggests that the money rule simulations do not provide much useful information on the properties of the proposed rule. The simulations are valuable in one respect, however. An examination of Figure 16 strongly suggests that the money rule is interacting with the rest of the model to produce a cycle of 5 to 6 years. Such a cycle is particularly evident in the simulations starting in 1956 and 1958. That the monetary rule has very powerful effects in the model is shown by the simulations beginning in 1960 and 1962. In both simulations unemployment reaches a trough in 1964 and then rises in spite of the 1964-65 tax cuts and the stimulus of spending for military operations in Vietnam starting at the end of 1965. There is no doubt that the monetary rule is too aggressive within the context of the FR-MIT model. A simulation of a perfectly steady rate of growth of money is shown in Figure 17. The rate of growth in this simulation is 2.76 percent per year, the same as the actual rate of growth over the period 1955-IV through 1969-I. In Figure 17, the curve labeled S2 is the simulated unemployment rate with the steady rate of growth of money. The simulated unemployment rate under the monetary rule is shown by S1, which is the same as S in panel A of Figure 16. The unconnected points show the same control simulation as shown in panel A of Figure 16. was made exogenous by suppressing the equation that makes demand deposits depend on unborrowed reserves. To simulate the effects of a particular rate of growth of money, the currency equation was retained, but demand deposits were set at whatever level was required to obtain the desired rate of growth of demand deposits plus currency. In the control simulations demand deposits were set at their actual levels, but currency remained an endogenous variable and differed somewhat from actual since simulated GNP differed somewhat from actual GNP. J U LY / A U G U S T 2008 477 Poole Figure 17 Simulations of FR-MIT Model SOURCE: Originally published version, p. 166. It appears impossible to draw any firm conclusions from the simulations. However, the simulations clearly raise the possibility that the proposed monetary rule may produce economic instability. If anything, the proposed rule is too aggressive, and so policy should probably err on the side of producing growth rates in money closer to a steady 3 to 5 percent rather than farther from the extremes in the proposed rule. IV. SELECTION AND CONTROL OF A MONETARY AGGREGATE Basic Issues Up to this point, the analysis has been entirely in terms of optimal control of the money stock. The theoretical analysis has been general enough that no precise definition of the money stock has been required. The empirical work, however, has used the narrow definition of demand deposits adjusted plus currency, for the simple reason that this definition seems to be the most appropriate one. 478 J U LY / A U G U S T 2008 In principle there is no reason not to look simultaneously at all of the aggregates and, of course, at all other information as well. But in practice, at the present state of knowledge, there simply is no way of knowing how all of these various measures ought to be combined.17 Furthermore, the selection of a single aggregate for operating purposes would permit the FOMC to be far more precise in its policy deliberations and in its instructions to the Manager of the Open Market Account. Thus, the best procedure would seem to be to select one aggregate as the policy control variable, and insofar as the state of knowledge permits, to incorporate other information into policy by making appropriate adjustments in the rate of growth of the aggregate selected. In principle the aggregate singled out as the control variable should be subject to exact determination by the Federal Reserve. The reason is that errors in reaching an aggregate that cannot be precisely controlled may interact with disturbances in the relationships between the aggregate and goal variables such as GNP to produce a suboptimal policy. However, as argued later in this section, this consideration is likely to be quite unimportant in practice for any of the aggregates commonly considered. Therefore, the analysis of which aggregate should be singled out will be conducted under the assumption that all of the various aggregates can be precisely controlled by the Federal Reserve. Selection of a Monetary Aggregate At the outset it must be emphasized that the various aggregates frequently discussed are all highly correlated with one another in the postwar period. This is true for total bank credit, the narrow money stock, the broad money stock (narrow money stock plus time deposits), the bank credit proxy (total member bank deposits), the 17 This point is an especially important one since those favoring simple approaches are frequently castigated for ignoring relevant information, and for applying “simplistic solutions to inherently complex problems.” For this charge to be upheld, it must be shown explicitly and in detail how this other information is to be used, and evidence must be produced to support the proposed complex approach. As far as this author knows, there is essentially no evidence sorting out the separate effects of various components of monetary aggregates. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole monetary base (member bank reserves plus currency held by the public and nonmember banks), and several other figures that can be computed. While these various aggregates are highly correlated over substantial periods of time, they show significantly different trends for short periods. In selecting an aggregate, the most important considerations are the theoretical relevance of the aggregate and the extent to which the theoretical notions have been given empirical support. Both of these considerations point to the selection of the narrowly defined money stock. The most important theoretical dispute is between those who emphasize the importance of bank deposit liabilities—the “monetary” view— and those who emphasize the importance of banks’ earning assets—the “credit” view. This controversy, which dates back well into the 19th century, is difficult to resolve because historically banks have operated on a fractional reserve basis and so have had both earning assets and deposit liabilities. Since balance sheets must balance, bank credit and bank deposits are perfectly correlated except insofar as there are changes in nonearning assets—such as reserves—or nondeposit liabilities—such as borrowing from the Federal Reserve System. If these factors never changed, the perfect correlation between bank deposits and bank credit would make it impossible ever to obtain evidence to distinguish between the monetary and the credit views. Since the correlation, while not perfect, has historically been very high, it has been very difficult to obtain evidence. Hence, it is still necessary to place major reliance on theoretical reasoning. There would be little reason to examine the issue closely if we could be confident that the very high correlation between deposits and bank credit would continue into the indefinite future. But there are already substantial differences in the short-run movements of bank credit and bank deposits, and these differences are likely to become greater and of a longer-term character in the future. Banks are raising increasingly large amounts of funds through nondeposit sources such as sales of commercial paper and of capital certificates and through borrowing from the Euro-dollar market and the Federal Reserve System. (Borrowings F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W from the System would probably expand significantly if proposed changes in discount-window administration were implemented.) The easiest way to examine the theoretical issues is to consider some hypothetical experiments. Consider first the experiment in which the Federal Reserve raises reserve requirements by $10 billion at the initial level of deposits but simultaneously buys $10 billion in U.S. government securities in the open market. Deposits need not change, but banks must hold more reserves and fewer earning assets. Under the monetary view the effects would be nil (except for very minor effects examined below) because deposits would be unchanged, but under the credit view the effect would be a tendency for income to contract because bank credit would be lower. The monetary view is easily explained. Suppose first that the banks initially hold U.S. government securities in excess of $10 billion. When reserve requirements are raised, the banks simply sell $10 billion of these securities, and this is exactly the amount being purchased by the Federal Reserve. Thus, since deposits are unchanged and bank loans to the nonbank private sector— hereinafter called simply the “private sector”— are also unchanged, there should be no effects on that sector. Now suppose that the banks do not have $10 billion in government securities. In this case they must sell private securities, say corporate bonds, to the private sector. The private sector obtains the funds to buy these bonds from the sale of $10 billion of government securities to the Federal Reserve. The amount of credit in the private sector is again unchanged. The banks own fewer private securities, while the public owns more private securities and fewer government securities. Thus, the amount of credit extended to the private sector need not change at all even though bank credit falls. However, two minor effects are possible: First, the Federal Reserve purchase of government securities changes the composition of portfolios. Thus, even if banks have over $10 billion of government securities, they may be expected to adjust their portfolios by selling some government securities and some private securities. For ease of exposition, run-offs of loans may be J U LY / A U G U S T 2008 479 Poole included in the sale of private securities. The net result, then, is that the banks have more reserves, fewer government securities, and fewer private securities; the private sector has fewer government securities and fewer liabilities to the banks. The private sector may have—but it will not necessarily have—fewer claims within the sector. It is quite possible that private units may substitute claims on other private units for the government securities sold to the Federal Reserve. Looked at from the liability side, those units initially with liabilities outstanding to banks may have those liabilities shifted to other private sector units. This occurs, of course, when banks sell securities to the private sector or allow loans to run off that are then replaced by firms selling commercial paper to other firms, drawing on sources of trade credit, and/or borrowing from nonbank financial institutions. A net effect can occur only when the combined portfolios of banks and the private sector contain fewer government securities, though more reserves, than before; such a change may be looked upon as a reduction in liquidity and thereby lead to a greater demand for money and a reduced willingness to undertake additional expenditures on goods and services. The second effect of the hypothetical experiment being discussed is that bank earnings will be reduced by the increase in reserve requirements. Banks will eventually adjust by raising service charges on demand deposits and/or reducing interest paid on time deposits. For simplicity, assume that the change in reserve requirements applies only to demand deposits so that there is no reason for banks to change the interest paid on time deposits. With higher service charges on demand deposits, lower interest rates on securities are required if people are to hold the same stock of money as before. Since the hypothetical experiment assumed that deposits did not change, interest rates must fall by the same amount as the increase in service charges, an effect that will tend to expand investment and national income. The portfolio effect tends to contract income while the service charge effect tends to expand income. These effects individually seem likely to be small, and the net effect may well be nil. In this regard, it is interesting to note that the rela480 J U LY / A U G U S T 2008 tionship of velocity to the Aaa corporate bond rate is about the same for observations in the 1950’s as in the 1920’s (Latané, 1954, 1960) in spite of the enormous changes in financial structure and in government bonds outstanding. Consider another hypothetical experiment— one that is in fact not so hypothetical at the current time. Suppose that banks suddenly start issuing large amounts of commercial paper and investing the proceeds in business loans. It is possible that the loans simply go to corporations that have stopped issuing their own commercial paper. In this case the bank would be purely a middleman with no effect on the aggregate amount of commercial paper outstanding. The increase in bank credit would not represent an increase in total credit. But, of course, banks issuing commercial paper must perform some function. This function is clearly that of increasing the efficiency of the financial sector in transferring funds from the ultimate savers to the ultimate borrowers. The efficiencies arise in several ways. First, under fractional reserve banking, banks have naturally developed expertise in lending. It is efficient to make use of this expertise by permitting banks to have more lendable funds than they would have if restricted to demand deposits alone. The efficiency takes the form of fewer administrative resources being required to transfer funds from savers to borrowers. The second form of efficiency results from the fact that financial markets function best when there is a large amount of trading in a standardized instrument. For example, the shares of large corporations are much more easily marketed than those of small corporations. Many investors want, and require, readily marketable securities, and they can be persuaded to buy securities in small firms only if the yields are high. As a result funds may go to large corporations to finance relatively low-yielding investment projects while highyielding projects available to small firms cannot be financed. Commercial banks, and other financial intermediaries, improve the allocation of capital by issuing relatively standardized securities with good markets and lending the proceeds to small firms. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole The question is whether there is any effect on economic activity from an increase in bank credit financed by commercial paper—assuming that the money stock is not affected. To begin with, it must be emphasized that an increase in the efficiency of investment does not necessarily affect the total of investment. The same resources may be absorbed either in building a factory that will produce a product that cannot be sold or in building a factory to produce a highly profitable product in great demand. Banks, and financial intermediaries in general, have the effect of reducing somewhat the cost of capital for small firms. Because intermediaries bid funds away from large corporations, the cost of capital for large corporations tends to be somewhat higher than it would be if there were no intermediaries. At this stage in the analysis the net effect on investment is impossible to predict since it depends on whether the reduction in investment by large corporations is larger or smaller than the increase in investment by small corporations. In examining the effects of intermediation, however, another factor must be considered. Suppose it is assumed that the interest rates relevant for the demand for money are rates on high-quality securities. It was argued above that intermediation tends unambiguously to raise the yields on highquality securities above what they otherwise would be. Since the assumption throughout has been that the stock of money is unchanged, the level of income must increase if the quantity of money demanded is to be unchanged with the higher interest rate of high-quality securities. The conclusion, therefore, is that the increase in bank credit is expansionary in the hypothetical experiment being discussed. This conclusion, however, does not warrant the further conclusion that bank credit is the appropriate monetary aggregate for policy purposes. The effect examined above occurs when any financial intermediary expands. Not only is there the problem that data for all intermediaries are simply not available on a current basis but also there are serious problems in even defining an intermediary. A particularly good example of this difficulty is afforded by trade credit. A large F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W nonfinancial corporation may advance trade credit to customers, many of whom may be small, and may also advance funds to suppliers through prepayments. The large corporation finances these forms of credit through the sale of securities, or through retained earnings diverted from its own investment opportunities and/or from dividends. In this case the large corporation is serving exactly the same function as the financial intermediaries are. But tracing these credit flows is obviously impossible at the present time. Another problem with bank credit as a guide to policy is that changes in bank credit depend both on changes in bank deposits and on changes in nondeposit sources of funds. As demonstrated by the hypothetical experiments examined above, the effect of a change in bank credit depends heavily on whether or not deposits change. One final hypothetical experiment will be considered. Suppose the U.S. Treasury sells additional government securities to the public to finance an increase in cash balances at commercial banks. Since banks have received no additional reserves, total deposits cannot change. Deposits owned by the public are transferred to the Treasury. Bank credit is unchanged, but the impact on the private sector is clearly contractionary. The private sector holds more government bonds and fewer deposits. Equilibrium can be restored only through some combination of a rise in interest rates and a decline in income. The conclusion is that it appears to be fundamentally wrong for policymakers to place primary reliance on bank credit. This is not to say that there is no information to be gained from analysis of bank and other credit flows. However, selection of bank credit as the monetary aggregate would be a mistake. Instead, information on credit flows may be used to adjust the desired rate of growth of the money stock, however it is defined, although it is not clear that the knowledge presently exists as to how to interpret credit flows. From this analysis it appears that neither bank credit nor any deposit total that includes Treasury deposits is an appropriate monetary aggregate for monetary policy purposes. Before considering the narrow and broad definitions of J U LY / A U G U S T 2008 481 Poole the money stock, let us examine the monetary base, total reserves, and unborrowed reserves. It is clear that different levels of the money stock may be supported by the same level of the monetary base. Given the monetary base, different levels of the money stock result from changes in reserve requirement ratios; from shifts of deposits between demand and time, which of course are subject to different reserve requirement ratios; from shifts of deposits among classes of banks with different reserve ratios; and from shifts between currency and deposits. These effects are widely understood, and they have led to the construction of monetary base figures adjusted for changes in reserve requirements. Similar adjustments are applied to total and nonborrowed reserves. If enough adjustments are made, the adjusted monetary base is simply some constant fraction of the money stock, while adjusted reserves are some constant fraction of deposits. It is obviously much less confusing to adopt some definition of the money stock as the appropriate aggregate rather than to use the adjusted monetary base or an adjusted reserve figure. There can be no doubt that FOMC instructions to the Manager in terms of nonborrowed reserves would be more precise and more easily followed than instructions in terms of the money stock. But the simplicity of reserve instructions would disappear if adjusted reserves were used, for then the Manager would have to predict such factors as shifts between demand and time deposits, the same factors that must be predicted in controlling the money stock. No one would argue that such factors—and others such as changes in bank borrowings and shifts in Treasury deposits—should be ignored. lf the FOMC met daily, instructions could go out in unadjusted form with the FOMC making the adjustments. But surely this technical matter should be handled not by the FOMC but by the Manager and his staff in order to permit the FOMC to concentrate on basic policy issues. The only aggregates left to consider are the narrowly and broadly defined money stocks. There is a weak theoretical case favoring the narrow definition because time deposits must be transferred into demand deposits or currency before they can be spent. The case is weak because 482 J U LY / A U G U S T 2008 the cost of this transfer is relatively low. If the cost were zero, then there would be no effective distinction between demand and time deposits. Indeed, since time deposits earn interest, all funds would presumably be transferred to time deposits. No strong empirical case exists favoring one definition over the other. The broad and narrow money stocks are so highly correlated over time that it is impossible to distinguish separate effects. It appears, however, that there is a practical case favoring the adoption of the narrow money stock. Time deposits include both passbook accounts, which can be readily transferred into demand deposits, and certificates of deposit, which cannot. Since CD’s appear to be economically much more like commercial paper than like passbook time accounts, they ought to be excluded from the broadly defined money stock. There is, of course, no reason why CD’s cannot be excluded from the definition of money. The problem is that banks may in the future invent new instruments that will be classified as time deposits for regulatory purposes but that are not really like passbook accounts. In retrospect it may be clear how the new instrument should be treated, but the situation may be confused for a time. The same sort of problem exists with demand deposits—consider the compensating balance requirements imposed by many banks— but it seems likely that the problem will remain more serious for time deposits. In summary, there is a strong case favoring the selection of some definition of the money stock as the monetary aggregate, and there appears to be a marginal case for preferring the narrowly defined money stock. Technical Problems of Controlling Money Stock In the preceding sections it has been argued that the monetary policy control instrument should be the money stock. The purpose of this section is to investigate some of the technical problems in controlling the money stock. The first topic examined is that of the form of instructions to the Manager of the System Open Market Account. Following this discussion is an examiF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole nation of the feedback method of control. Finally, there is an examination of the significance of data revisions. All of this discussion is in terms of the narrowly defined money stock, but much of it also applies to other aggregates. Specification of the Desired Money Stock. There are two major issues connected with the form of FOMC instructions to the Manager. The first is whether the desired money stock should be expressed in seasonally adjusted or unadjusted form, while the second is whether the desired money stock should be expressed in terms of a complete path week by week over time or of an average over some period of time. The first issue turns out to be closely related to the question of data revisions, and so its discussion will be deferred for the moment. It is to the second issue that we now turn. Since required reserves are specified in terms of a statement-week average, the statement week is the natural basic time unit for which to measure the money stock, and the measure takes the form of the average of daily money stock figures over the statement week. The fact that daily data may not be available on all components of the money stock does not affect the argument; however estimated, the weekly-average figure is the most appropriate starting point in the analysis. The weekly money stock is clearly not subject to precise control because of data lags and uncontrollable random fluctuations. Furthermore, no one believes that these weekly fluctuations have any significant impact. The natural conclusion to be drawn is that there is no point in specifying instructions in terms of weekly data but rather that some average level over a period of weeks should be used. Upon closer examination, however, this conclusion can be shown to be unjustified. The difficulty in expressing the instructions in terms of averages can be explained very simply by two examples. To keep the examples from becoming too complicated, it will be assumed that instructions take the form of simple rates of growth on a base money stock of $200 billion. The neglect of compounding makes no essential difference to the argument. For the first example, assume that the policy instruction is for a growth rate of 4 percent per F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W annum, which is $8 billion per year or about $154 million per week. If the money stock grew by $154 million per week for 8 weeks, then the figure for the eighth week would be above the base week figure by an amount representing a 4 percent annual growth rate. The average of weeks 5 through 8 would be above the average of weeks 1 through 4 by $616 million, an amount also representing a 4 percent annual growth rate. So far, there is no reason to favor the path specification over a specification in terms of 4-week averages. Now suppose that the increase in weeks 1 through 4 was on schedule, but that a large uncontrollable increase of $500 million occurred in the fifth week. Starting from a base-week figure of $200 billion, the average money stock for weeks 1 through 4 would be $200.385 billion, and if the instruction were in terms of 4-week averages it would specify an average money stock of $201.001 billion for weeks 5 through 8. Since by hypothesis the money stock grew by $154 million in each of the first 4 weeks, in the fourth week the level was $200.616 billion. The jump of $500 million in the fifth week would take the level to $201.116 billion, a figure already above the desired average of $201.001 billion for weeks 5 through 8. To reach this desired average given the jump in week 5, the money stock in weeks 6 through 8 would have to average less than $201.001 billion, and so the money stock would have to be forced below the level of the fifth week for weeks 6 through 8. Furthermore, as the reader may calculate, it would be necessary to have higher than normal weekly growth in weeks 9 through 12 if the average of these weeks were to be above the average of weeks 5 through 8 by $616 million. On the other hand, if the instruction were in terms of the desired weekly path, the instruction would read that the desired money stock in the eighth week was $201.232 billion, and therefore the Manager would not have to force the money stock down in weeks 6 through 8. Instead, he could aim for a growth of about $39 million in each of the weeks 6 through 8 to bring the level in week 8 to the desired figure of $201.232 billion. From this example it can be seen that specification in terms of averages of levels of the money J U LY / A U G U S T 2008 483 Poole stock forces the Manager to respond to random fluctuations in a whipsawing fashion. Since week-by-week fluctuations have essentially no significance, there is no point in wrenching the financial markets in order to undo a random fluctuation. If averaging is to be used, the average should be specified in terms of the desired average weekly change over, say, the next 4 weeks rather than in terms of the average level of the next 4 weeks. Specification in terms of the average weekly change is equivalent to a specification stating that the Manager should aim for a particular target level in the fourth week. The second example illustrating the hazards of specification in terms of the average level will show what happens when policy changes. As before, assume that the money stock in the base week is $200 billion and that the desired growth is at a 4 percent rate in weeks 1 through 4. In this example it is assumed that there are no errors in hitting the desired money stock. Thus, the money stock is assumed to grow by $154 million per week, reaching a level of $200.616 billion in the fourth week and an average level of $200.385 billion for weeks 1 through 4. Now suppose that in week 4 the FOMC decides on a policy change and specifies a 1 percent growth rate for the money stock for weeks 5 through 8. If the specification were in terms of the average level, then it would require an increase in the average level of $154 million, which would bring the average level to $200.539 billion for weeks 5 through 8. But the figure for week 4 is already $200.616 billion, and so the money stock in weeks 5 through 8 would have to average less than the figure already achieved in week 4. Thus, after a steady 4 percent growth week by week, an average-level policy specification would actually require a negative week-by-week growth before the new 1 percent growth rate could be achieved. On the other hand, a policy specification in terms of the weekly path would require a weekly growth of $38.5 million each week for weeks 5 through 8. To make the point clear, this example was constructed so that the policy shift from a 4 to a 1 percent growth rate would actually require a negative growth rate for a time on a week-by-week 484 J U LY / A U G U S T 2008 basis when the instructions are in terms of average levels. In general, when average levels are used, a policy shift to a lower growth rate will require in the short term a growth rate lower than the new policy rate set, and a policy shift to a higher growth rate will require a short-term growth rate above the new policy rate. Since policymakers will typically want to shift policy gradually, the levels specification is especially damaging because it in fact instructs the Manager to shift policy more rapidly than the policymakers had desired. It should be noted that the larger the number of weeks included in the average-level specification, the more severe this problem becomes. Because the money stock cannot be controlled exactly, there is a natural tendency to feel that instructions stated in terms of averages are more attainable. In actuality, of course, this effect is illusory; averaging produces a smaller number to measure the errors, but does not improve control. Nevertheless, if averages are to be used in the instructions, the above examples demonstrate that the averages should be calculated in terms of weekly (or perhaps monthly) changes but not in terms of averages of levels. Use of average changes does have one advantage, however. An instruction in this form permits the Manager to correct an error in week 1 over the next few weeks rather than instructing him to correct the error entirely in week 2. As explained above, an instruction in terms of the average weekly change over the next 4 weeks is equivalent to an instruction in terms of the desired level in week 4, leaving unspecified the desired levels in weeks 1 through 3. Control Through the Feedback Principle. It is useful to begin by comparing the problems of controlling the money stock with the problems of controlling interest rates. In controlling interest rates, the availability of continuous readings on rates makes it possible for the Manager to exercise very accurate control without understanding the causes of rate changes. Being in continuous contact with the market, the Manager can intervene with open market purchases or sales as soon as the federal funds rate, the Treasury bill rate, or any other rate starts to change in an undesirable fashion. This feedback control is not exact F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole since interest rate information arrives with some lag, and there are other lags such as the time required to decide upon and execute an open market transaction and the time it takes for the market to react to the transaction. More precise control over interest rates could be achieved if the Manager were willing to announce Federal Reserve buying and selling prices for, say, 3-month Treasury bills available to all comers. This is essentially the way in which government securities were pegged during World War II. In principle, there is no reason why such a peg could not be operated in peacetime, although it would certainly be desirable to change the peg frequently, perhaps as often as every day or even every hour. However, in terms of actual behavior of interest rates there is no significant difference between a frequently adjusted peg and continuous intervention by the Manager as described in the previous paragraph. The main point of this discussion of interest rate control is to emphasize that with frequent interest rate readings it is not necessary to know exactly what causes interest rate changes. In time the Manager develops a feel for the market that enables him to guess accurately which interest rate changes are temporary and which are likely to be “permanent” and so require offsetting open market operations. Furthermore, his feel for the market will enable him to know how large the operations should be. Finally, when he guesses wrongly on these matters, his continuous contact with the market enables him to correct mistakes rapidly. The same arguments apply to controlling the money stock. The difference between interest rate control and money stock control is a matter of degree rather than kind. Data on the money stock become available with a greater lag, and the data are more subject to revision. But since it is not necessary to control the money stock down to the last dollar, the question is whether it is technically possible to have control that is accurate enough for policy purposes. The answer to this question would certainly appear to be in the affirmative. The weekly-average figure for the money stock is released to the public 8 days following the end F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W of the week to which the average refers. Of course, data are available internally with a shorter lag. Since the policy rule in the previous section is based on controlling the monthly-average money stocks it would appear that the data are at the present time available with a short enough lag that feedback methods of control are feasible. To see how feedback control would work, suppose that the Manager were instructed to come as close as possible to a target money stock of M4* in week 4 of a 4-week operating horizon. The Manager knows that the weekly change in the money stock depends on open market purchases, P, which he controls, and many other factors as well, which for simplicity of exposition will be denoted by one factor, z. These factors cannot be predicted exactly, and so the Manager will think of z as consisting of a predictable part, ẑ, and an unpredictable part, u. These relationships may be expressed as (5) ∆M = α P + z = α P + zˆ + u where α is the coefficient giving the change in money per dollar of open market purchases. If there were no errors in measuring the money stock, the analysis could be completed on the basis of equation 5. But of course there are errors in measuring the money stock. To analyze the significance of measurement errors, let Mi be the money stock for week i as measured at the end of week i.18 Also, let Mif be the final “true” money stock figure for week i, and let ei = Mif – Mi . The Manager starts out the 4-week period with an estimated money stock of M0 for week zero. Of course, the figure for M0 is a preliminary one, but revisions in this figure as more data accumulate will affect the estimates for the money stock in later weeks and so affect the Manager’s actions in later weeks. It will be assumed that he wants to increase the money stock by equal amounts in each week to reach the desired figure of 18 If a money stock estimate is not directly available at the end of week i, one can be constructed by taking the estimate from actual deposit data for week i – 1 and adding to it a projection for the effects of open market operations and other factors for week i. This projection would, of course, come from equation 5. J U LY / A U G U S T 2008 485 Poole M4* in week 4. In week 1, therefore, he wants to produce a change in the money stock 1/4(M4* – M0). Substituting this figure into equation 5 we obtain ( ) 1 M 4∗ − M 0 = α P1 + zˆ 1 + u1 4 ( ) Thus, the Manager sets P1 according to P1 = (6) 1 1 M 4∗ − M 0 − zˆ 1 α 4 At the end of the first week the Manager has the estimate, M1, for the money stock for that week, and again it is assumed that he wants to spread the desired change M* – M1 equally over the next 3 weeks. Thus, the Manager sets P2 according to P2 = (7) ( ) 1 1 M 4∗ − M 1 − zˆ 2 α 3 Similarly, he sets P3 and P4 according to equations 8 and 9. ( ) (8) P3 = 1 1 M 4∗ − M 2 − zˆ 3 α 2 (9) P4 = 1 ∗ M − M 3 − zˆ 4 α 4 ( ) From equations 9 and 5 it can be seen that the actual money stock in week 4 is (10) M 4f = M 3f + M ∗ − M 3 − zˆ 4 + z = M 4∗ + e3 + u4 This expression for the fourth week of a planning period generalizes to the nth week of a planning period of any length merely by replacing the subscript 4 by the subscript n. We can, therefore, express the annual rate of growth, g, over an n week period by g= (11) = 52 M nf − M 0f n M 0f 52 M n∗ − M 0f 52 + n (e n − 1 + u n ) n M 0f From equation 11 it can be seen that the actual growth rate, g, equals the desired growth rate plus an error term that becomes smaller as n becomes larger. 486 J U LY / A U G U S T 2008 This analysis shows that a feedback control system that continuously adjusts open market operations as data on the money stock in the recent past become available can achieve a target rate of growth with a margin of error that is smaller the longer the period over which the rate of growth is calculated. It also provides a framework in which to examine the relative importance of operating errors, the ui , and data errors, the ei . To obtain an accurate estimate of the sizes of these errors is beyond the scope of this study. However, a very crude method may be used to obtain an estimate of the maximum size of the total error. Monthly money stock changes at annual rates were computed for the period January 1951 through September 1969 on the basis of seasonally adjusted data. This time period yields a total of 225 monthly changes. Then each monthly change was expressed in terms of its deviation from the average of the changes for the previous 3 months. For example, the September deviation was calculated by subtracting from the September monthly change the average of the changes for August, July, and June. The use of deviations allows in part for longer-run trends in the money stock, which trends are assumed to be readily controllable. Since the deviations were calculated over a period during which little or no attention was paid to controlling the money stock, they surely represent an upper limit to the degree of volatility in the money stock to be expected under a policy directed at control of the money stock. These monthly deviations have a standard deviation of 3.12 percent per annum. Applying equation 11, except for replacing 52 by 1 to reflect the fact that the rates of change were expressed at annual rates in the first place, it is found that the standard deviation over a 3-month period would be 1.04 percent per annum. If it is assumed that these deviations are normally distributed, the conclusion is that over 3-month periods the actual growth rate would be within plus or minus 1.04 percent of the desired growth rate about 68 percent of the time, and would be within plus or minus 2.08 percent about 95 percent of the time. Inasmuch as these limits would be cut in half over 6-month periods, the actual growth rate 95 percent of the time would be in the range of plus F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole or minus 1.04 percent of the desired growth rate.19 When it is recalled that these calculations are based on an estimate of variability over a period in which very little attention was paid to stabilizing money stock growth rates, it is clear that fears as to the ability of the Federal Reserve to control the money stock accurately are completely unfounded.20 This conclusion justifies the approach used at the beginning of this section on the selection of a monetary aggregate, at least for the narrowly defined money stock and most probably for other aggregates as well. That approach, it will be recalled, analyzed the selection issue on the assumption that every one of the aggregates considered could be precisely controlled for all practical purposes. There can be no doubt that errors in reaching targets for goal variables such as GNP, at the present state of knowledge, are due almost entirely to incomplete knowledge of the relationships between instrument variables (such as various aggregates and interest rates) and the goal variables, and hardly at all to errors in setting instrument variables at desired levels. Problems of Data Revisions and Changing Seasonality. Another topic that needs examination is the effect of data revisions. While weeklyaverage data are released with an 8-day lag, these figures are subject to revision. Not much weight can be given to early availability of data that are later revised substantially. To investigate this problem, two money stock series were compared, one “preliminary” and one “final.” Since the analysis below is based on published monthly data, it obviously provides little insight into the accuracy of weekly data. However, since policy instructions may be based on monthly data, the analysis is of some value in assessing data accuracy. Furthermore, the conclusions on the importance of revisions in seasonal factors can be expected to hold for the weekly data. A “preliminary” series of monthly growth rates of the money stock was constructed by calculating the growth rate for each month from data reported in the Federal Reserve Bulletin for the following month. For example, the Bulletin dated September reports money stock data for 13 months through August; it is the annual rate of change of August over July that is called the “preliminary” August rate-of-change observation. The “final” series is the annual rate of growth calculated from the monthly money stock series covering 1947 through September 1969, reported in the Federal Reserve Bulletin for October 1969, pp. 790-93. Data were gathered on both a seasonally adjusted basis and an unadjusted basis for January 1961 through August 1969. The correlation between the preliminary and final seasonally adjusted series is 0.767, while for the unadjusted series the correlation is 0.997. Another way to compare the preliminary and final series is to examine the differences in the two series.21 For the seasonally adjusted data, the differences have a mean of 0.122 and a standard deviation of 3.704, and the mean absolute difference is 2.891. On the other hand, for the seasonally unadjusted data the differences have a mean of 0.150 and a standard deviation of 1.366, and the mean absolute difference is 0.955.22 These results make it abundantly clear that the major reason why the preliminary and final 19 21 The analysis of the differences inadvertently runs from February 1961 through August 1969 while the correlation analysis runs from January 1961 through August 1969. 22 To take account of the fact that the “final” money stock series may be further revised for months near the October 1969 publication date of this series, the analysis of differences between the preliminary and final series was also run on the period February 1961 through December 1968. The mean difference, the standard deviation of the differences, and the mean absolute difference, are, respectively, for the seasonally adjusted data 0.026, 3.779, and 2.922, while the figures for the seasonally unadjusted data are 0.038, 1.280, and 0.890. In spite of the fact that the “final” series is not really final for 1969 data, the average differences are generally larger for the longer period due to the relatively large data revisions in the middle of 1969. 20 If the calculations are based on the variability of the monthly changes themselves rather than on the deviations of the monthly changes, the results are not greatly changed. The standard deviation of the monthly changes over the same period used before is 3.53 per cent per annum, which yields a 95 percent chance of the growth rate being in a range around the desired rate of plus or minus 2.36 (1.18) per cent per annum for 3-month (6-month) periods. Compare “First, however, it may be worthwhile to touch on the extensively debated subject whether the Federal Reserve, if it wanted to, could control the rate of money supply growth. In my view, this lies well within the power of the Federal Reserve to accomplish provided one does not require hair-splitting precision and is thinking in terms of a time span long enough to avoid the erratic, and largely meaningless, movements of money supply over short periods” (Holmes, 1969, p. 75). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 487 Poole figures on the money stock differ is revision of seasonal adjustment factors. While such revisions may produce substantial differences between preliminary and final monthly growth rates, the differences must be lower for the average of several months’ growth rates. The reason, of course, is that revision of seasonal factors must make the figures for some months higher and those for other months lower, leaving the annual average about unchanged. The significance of revisions in seasonal factors can be understood only after a discussion of the significance of seasonality for a money stock rule. If the monetary rule were framed in terms of the seasonally unadjusted money stock, the result would be to introduce substantially more seasonality into short-term interest rates than now exists. It can be argued not only that greater seasonality in interest rates would not be harmful but also that it would be positively beneficial. Greater seasonality in interest rates would presumably tend to push production from busy, highinterest seasons into slack, low-interest seasons. Although the argument for seasonality in interest rates could be pushed further, there is an important practical reason for not initially adopting a money rule stated in terms of the seasonally unadjusted money stock. The reason is that the rule ties the growth rate of the money stock to the seasonally adjusted unemployment rate and to the interest rate. The rule has been developed through an examination of past experience. If the seasonal were taken out of the money stock, a different seasonal would be put into interest rates, and possibly into the unemployment rate as well. Seasonal factors for these variables, especially for the unemployment rate, determined from past data would no longer be correct if the money stock seasonal were removed. Seasonally adjusting the unemployment index by the old factors could produce considerable uncertainty over the application of the monetary rule. Thus, application of the rule through the seasonally unadjusted money stock, if desirable at all, should only come about through gradual reduction rather than immediate elimination of seasonality. A further reason for a gradual approach would be to permit the financial markets to adjust more easily to changed seasonality. 488 J U LY / A U G U S T 2008 The point of this discussion is not to urge acceptance of a rule framed in terms of the unadjusted money stock, since this step would not be initially desirable in any case. Rather, the point is to emphasize that seasonality is in the money stock only in order to reduce the seasonality of other variables, primarily interest rates. The seasonality of the money stock, unlike variables such as agricultural production, is not inherent in the workings of the economy but rather exists because the Federal Reserve wants it to exist. The money stock can be made to assume any seasonal pattern the Federal Reserve wants it to assume. The monetary rule should be framed, at least initially, in terms of the seasonally adjusted money stock—using the latest estimated seasonal factors. In subsequent years changes in these seasonal factors should not result from mechanical application of seasonal adjustment techniques to the money stock data but rather should be the result of a deliberate policy choice. The policy choice would be based on the desire to change seasonality of other variables. For example, if it were thought desirable to take the seasonality out of short-term interest rates, the seasonal factors for the money stock would then be changed to take account of changes in tax dates and other factors. Under a money stock policy, whether or not guided by a monetary rule, revised seasonal factors cannot properly be applied to past data. If the changes are applied to past data with the result that some monthly growth rates of adjusted data become relatively high while others become relatively low, the conclusion to be drawn is not that policy was mistaken as a result of using faulty seasonal factors. Instead, the conclusion is merely that seasonal policy differed in the past from current policy or from the seasonal pattern assumed by the investigator who computed the seasonal factors. Seasonal policy can be shown to be “wrong” only by showing that undesirable seasonals exist in other variables. One final problem deserves discussion. While it appears from the analysis of seasonally unadjusted money stock data that revisions of the data F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole are relatively unimportant, at least from the evidence for 1961-69, how should the policy rule be adjusted when there are major data revisions— as in the middle of 1969? For example, suppose that revisions indicate that monetary growth has been much higher than had been expected, and higher than was desirable. On the one hand, policy could ignore the past high rate of growth and simply maintain the current rate of growth of the revised series in the desired range. On the other hand, the policy could be to return the money stock to the level implied by applying the desired growth rate to the money stock in some past base period. The first alternative involves ratifying an undesirable high past rate of growth, while the second may involve a wrenching change in the money stock to return it to the desired growth path. The proper policy would no doubt have to be decided on a case-by-case basis. However, a useful presumption might be to adopt the second alternative, but to set as the base the money stock 6 months in the past and to return to the desired growth path over a period of several months. Improving Control Over the Money Stock. The analysis above has shown that under present conditions the money stock can be controlled quite accurately. However, it should be emphasized that there are numerous possibilities for improving control. Although detailed treatment of this subject is beyond the scope of this study, a few very brief comments appear appropriate. There are three basic methods for improving control. The first method is that of improving the data. The more quickly the deposit data are available, the more quickly undesirable movements in the money stock can be recognized and corrected. And the more accurate the deposit data, the fewer the mistakes caused by acting on erroneous information. It is clear that expenditures of money on expanding the number and coverage of deposit surveys and on more rapid processing of the raw survey data can improve deposit data. The second method of improving control is through research, which increases our understanding of the forces making for changes in the money stock. For example, transfers between F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W demand and time deposits might be more accurately predicted through research into the causes of such transfers. The third method of improving control is through institutional changes. To reduce fluctuations in excess reserves and thereby achieve a more dependable relationship between total reserves and deposits, the federal funds market might be improved by making possible transfers between the East and West Coasts after east coast banks are closed. Also helpful would be a change from lagged to contemporaneous reserve requirements. More radical reforms such as equalization of reserve requirements for city, country, and nonmember banks and elimination of reserve requirements on time deposits should also be considered. V. SUMMARY Purposes of the Study The primary purpose of this study has been to argue that a major improvement in monetary policy would result through a systematic policy approach based on adjustments in the money stock. Equal emphasis has been placed on the “systematic” part and the “money stock” part of this approach. The analysis has proceeded first by showing why policy adjustments should be made through money stock adjustments, and second by showing how these policy adjustments might be systematically linked to the current business situation through a policy guideline or rule-of-thumb. A third, and subsidiary, part of this study is an analysis of the reasons for preferring the money stock over other monetary aggregates, and of some of the problems in reaching desired levels of the money stock. It has been emphasized throughout that this policy approach is one that is justified for the intermediate-term future on the basis of knowledge now available. The specific recommendations are not intended to be good for all time. Indeed, the approach has been designed to encourage evaluation of the results so that the information obtained thereby can be incorporated into policy decisions in the future. J U LY / A U G U S T 2008 489 Poole The Theory of Monetary Policy Under Uncertainty Since policymakers have repeatedly emphasized the importance of uncertainty, it is necessary to analyze policy problems within a model that explicitly takes uncertainty into account. In particular, only within such a model is it possible to examine the important current issue of whether policy adjustments should proceed through interest rate or money stock changes. A monetary policy operating through interest rate changes sets interest rates either through explicit pegging as was used in World War II or through open market operations directed toward the maintenance of rates in some desired range. Under such a policy the money stock is permitted to fluctuate to whatever extent is necessary to keep interest rates at the desired levels. On the other hand, a policy operating through money stock changes uses open market operations to set the money stock at its desired level while permitting interest rates to fluctuate freely. If there were perfect knowledge of the relationships between the money stock and interest rates, the issue of money stock versus interest rates would be nonexistent. With perfect knowledge, changes in interest rates would be perfectly predictable on the basis of policy-induced changes in the money stock, and vice versa. It would, therefore, be a matter of preference or prejudice, but not of substance, whether policy operated through interest rates or the money stock. To analyze the interest versus money issue, then, it is necessary to assume that there is a stochastic link between the two variables. And, of course, this is in fact the case. There are two fundamental reasons for the stochastic link. First, the demand for money depends not only on interest rates and the level of income but also on other factors, which are not well understood. As a result, the demand for money fluctuates in a random fashion even if income and interest are unchanged. If the stock of money is fixed by policy, these random demand fluctuations will force changes in interest and/or income in order to equate the amount demanded with the fixed supply. The second source of disturbances between money and interest stems from disturbances in 490 J U LY / A U G U S T 2008 the relationship between expenditures—especially investment-type expenditures—and interest rates. Given an interest rate fixed by policy, these disturbances produce changes in income through the multiplier process, and these income changes in turn change the quantity of money demanded. With interest fixed by policy, the stock of money must change when the demand for money changes. On the other hand, if the money stock were fixed by policy, since the expenditure disturbance changes the relationship between income and interest, some change in the levels of income and/or interest would be necessary for the quantity of money demanded to equal the fixed stock. Money stock and interest rate policies are clearly not equivalent in their effects, given that disturbances in money demand and in expenditures do occur. Since the effects of these policies are different, which policy to prefer depends on how the effects differ and on policy goals. At this level of abstraction, it is clearly appropriate to concentrate on the goals of full employment and price stability. Unfortunately, the formal model that has been worked out, which is examined carefully in Section I above, applies only to the goal of stabilizing income. If “income” is interpreted to mean “money income,” then the goals of employment and price level stability are included but are combined in a crude fashion. The basic differences in the effects of money stock and interest rate policies can be seen quite easily by examining extreme cases. Suppose first that there are no expenditure disturbances, so there is a perfectly predictable relationship between the interest rate and the level of income. In that case, a policy that sets the interest rate sets income, and policymakers can choose the level of the interest rate to obtain the level of income desired. When the interest rate is set by policy, disturbances in the demand for money change the stock of money but not the level of income. On the other hand, if policy sets the money stock, then the money demand disturbances would affect interest and income leading to less satisfactory stabilization of income than would occur under an interest rate policy. The other extreme case is that in which there are disturbances in expenditures but not in money F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole demand. If policy sets the interest rate, expenditure disturbances will produce fluctuations in income. But if the money stock is fixed, these income fluctuations will be smaller. This point can be seen by considering a specific example such as a reduction in investment demand. This disturbance reduces income. But given an unchanged money demand function, with the fall in income, interest rates must fall so that the amount of money demanded will equal the fixed stock of money. The decline in the interest rate will stimulate investment expenditures, thus offsetting in part the impact on income of the initial decline in the investment demand function. With expenditures disturbances, then, to stabilize income, it is clearly better to follow a money stock policy than an interest rate policy. The conclusion is that the money versus interest issue depends crucially on the relative importance of money demand and expenditures disturbances. It is especially important to note that nothing has been said about the size of the interest elasticity of the demand for money, or of the interest elasticity of investment demand. These coefficients, and others, determine the relative impacts of changes in money demand and in investment and government expenditures when the changes occur. The interest versus money issue does not depend on these matters, however, but only on the relative size and frequency of disturbances in the money demand and expenditures functions.23 The analysis above is modified in detail by considering possible interconnections between money demand and expenditures disturbances. It is also true that in general the optimal policy is not a pure interest or pure money stock policy, but a combination of the two. These matters, and a number of others, are discussed in Section I. Evidence on Relative Magnitudes of Real and Monetary Disturbances Resolution of the money versus interest issue depends on the relative size of real and monetary disturbances. Unfortunately, there is no com23 For a full understanding of this important point, the reader should refer to the analysis of Section I. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W pletely satisfactory body of evidence on this matter. Indeed, because of the conceptual difficulties of designing empirical studies to investigate the issue, the evidence is unlikely to be fully satisfactory for some time to come. Nevertheless, by examining a number of different types of evidence, a substantial case can be built favoring the use of the money stock as the policy control variable. Before discussing the evidence, it is necessary to define in more detail what is meant by “disturbance.” Consider first a money demand disturbance. The demand for money depends on the levels of income and of interest rates, and on other variables. The simplest form of such a function uses GNP as the income variable, and one interest rate—say the Aaa corporate bond rate—and all other factors affecting the demand for money are treated as disturbances. To the extent possible, of course, these other factors should be allowed for, but for policy purposes these factors must be either continuously observable or predictable in advance so that policy may be adjusted to offset any undesirable effects on income of these other factors. Factors not predictable in advance must be treated as random disturbances. Similarly, expenditures disturbances are defined as the deviations from a function linking income to the interest rate and other factors. These other factors would include items such as tax rates, government expenditures, strikes, and population changes. Again, for policy purposes these factors must be forecast, and so errors in the forecasts of these items must be included in the disturbance term. It is important to realize that the disturbances will be defined differently for scientific purposes ex post because the true values of government spending and so forth can be used in the functions once data on these items are available. In the discussion of the theoretical issues above it was noted that an expenditure disturbance would have a larger impact on income under an interest rate policy than under a money stock policy. Simulation of the FR-MIT model provides the estimate that the impact on income of an expenditures disturbance, say in government spending, is over twice as large under an interest rate policy as under a money stock policy. An J U LY / A U G U S T 2008 491 Poole error in forecasting government spending, then, would lead to twice as large an error in income under an interest rate policy. Since there is no systematic record of forecasting errors for variables such as government spending and strikes, there is no way of producing evidence on the size of such forecasting errors. However, after listing the variables that must be forecast, as is done in Section II, it is difficult to avoid feeling that errors in forecasting are likely to be quite significant. These real disturbances, including forecast errors in government expenditures, strikes, and so forth, must be compared with the disturbances in money demand. The reduced-form studies conducted by a number of investigators provide some evidence on this issue. These studies compare the relative predictive power of monetarist and Keynesian approaches in explaining fluctuations in income. From these studies the predictive power of both approaches appears about equal. However, the predictive power of the Keynesian approach relies on ex post observation of “autonomous” expenditures, and it is clear that these expenditures are subject to forecasting errors ex ante whereas the money stock can be controlled by policy. The evidence from the reduced-form studies suggests that when forecast errors of autonomous expenditures are included in the disturbance term, the disturbances are larger on the real side than on the monetary side. There are many difficulties with the reduced-form approach and so these results must be interpreted cautiously. Nevertheless, the results cannot be ignored. The final piece of evidence offered in Section II is a study by the author of the stability of the demand for money function over time. Using a very simple function relating the income velocity of money to the Aaa corporate bond rate, he found that a function fitted to quarterly data for 1947-60 also fits data for 1961-68 rather well. The reader interested in the precise meaning of “rather well” should turn to the technical discussion in Section II. Evidence on relative stability is difficult to obtain and subject to varying interpretations. No single piece of evidence is decisive, but all the various scraps point in the same direction. The 492 J U LY / A U G U S T 2008 evidence is not such that a reasonable man can say that he has no doubts whatsoever. But since policy decisions cannot be avoided, the reasonable decision based on the available evidence is to adopt the money stock as the monetary policy control variable. A Monetary Rule for Guiding Policy The conclusion from the theoretical and empirical analysis is that the money stock ought to be the policy control variable. For this conclusion to be very useful, it must be shown in detail how the money stock ought to be used. It is not enough simply to urge policymakers to make the “appropriate” adjustments in the money stock in the light of all “relevant” information. There is no general agreement on exactly what types of adjustments are appropriate. However, it would probably be possible to obtain agreement among most economists that ordinarily the money stock should not grow faster than its long-run average rate during a period of inflation and should not grow slower than its long-run average rate during recession. But many economists would want to qualify even this weak statement by saying that there may at times be special circumstances requiring departures from the implied guideline. Others would say that there is no hope at present of gauging correctly the impact of special circumstances (or even of “standard” circumstances) so that policy should maintain an absolutely steady rate of growth of the money stock. The basic issues are, first, whether policymakers can forecast disturbances well enough to adjust policy to offset them, and second, the extent to which money stock adjustments to offset short-run disturbances will cause undesirable longer-run changes in income and other variables. The theoretical possibilities are many, but the empirical knowledge does not exist to determine which theoretical cases are important in practice. It is for this reason that a systematic policy approach is needed so that policy can be easily evaluated and improved with experience. Policy could be linked in a systematic way to a large-scale model of the economy. Target values of GNP and other goal variables could be selected F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole by policymakers, and then the model solved for the values of the money stock and other control variables (for example, discount rate) needed to achieve policy goals. While this approach may be feasible in the future, it is not feasible now because a sufficiently accurate model does not exist. Instead, policy decisions are now made largely on the basis of intuitive reactions to current business developments. Given this situation, the obvious approach is to specify precisely how policy decisions ought to depend on current developments, and this is the approach taken in Section III. The specification there takes the form of a policy guideline, or rule-of-thumb. The proposed rule is purposely simple so that evaluation of its merits would be relatively easy. Routine evaluation of an operating guideline would over time produce a body of evidence that could be used to modify and complicate the rule. But it is necessary to begin with a simple rule because the knowledge that would be necessary to construct a sophisticated rule does not exist. The proposed rule assumes that full employment exists when the unemployment rate is in the 4.0 to 4.4 percent range. The rule also assumes that at full employment, a growth rate of the money stock of 3 to 5 percent per annum is consistent with price stability. Therefore, when unemployment is in the full employment range, the rule calls for monetary growth at the 3 to 5 percent rate. The rule calls for higher monetary growth when unemployment is higher, and lower monetary growth when unemployment is lower. Furthermore, when unemployment is relatively high the rule calls for a policy of pushing the Treasury bill rate down provided monetary growth is maintained in the specified range; similarly, when unemployment is relatively low the rule calls for a policy of pushing the bill rate up provided monetary growth is in the specified range. Finally, the rule provides for adjusting the rate of growth of money according to movements in the Treasury bill rate in the recent past. The exact rule proposed is in Table 3 and the detailed rationale for the various components of the rule is explained in the discussion accompanying that table. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W The rule is specified throughout in terms of 2 percent ranges for the rate of growth of the money stock on a month-by-month basis. By expressing the rule in terms of a range, leeway is provided for smoothing undesirable interest rate fluctuations and for minor policy adjustments in response to other information. Furthermore, it is not proposed that this rule-of-thumb or guideline be followed if there is good reason for a departure. But departures should be justified by evidence and not be based on vague intuitive feelings of what is needed since the rule was carefully designed from the theoretical and empirical analysis of Sections I and II, and from a careful review of post-accord policy. There is no way of really testing the proposed rule short of actually using it. However, it is useful to compare the rule with post-accord policy. A detailed comparison may be found in Section III. A summary comparison suggests, however, that for the period January 1952 through July 1968 the rule would have provided a less appropriate policy than the actual policy in only 63 of the 199 months in the period. The rule was judged to be less appropriate if it called for a higher— lower—rate of monetary growth than actually occurred and unemployment 12 months hence was below—above—the desired range of 4.0 to 4.4 percent. The rule was also judged less appropriate than the actual policy if actual policy was not within the rule but unemployment nevertheless was in the desired range 12 months hence. The rule actually has slightly fewer errors if the criterion is unemployment either 6 or 9 months following the months in question. The rule has the great virtue of turning policy around promptly as imbalances develop and of avoiding cases such as the 2.2 percent rate of decline in the money stock from July 1957 through January 1958, during which time the unemployment rate rose from 4.2 percent to 5.8 percent. Furthermore, it seems most unlikely that the rule would produce greater instability than the policy actually followed. Actual policy has, as measured by the money stock, been most expansionary during the early and middle stages of business cycle expansions and most contractionary during the last stages of business expansions and early J U LY / A U G U S T 2008 493 Poole stages of business contractions. Unless a very improbable lag structure exists, the rule would surely be more stabilizing than the actual historical pattern of monetary growth. Selection and Control of a Monetary Aggregate The analysis in this study is almost entirely in terms of the narrowly defined money stock. The reasons for using the narrowly defined money stock as opposed to other monetary aggregates may be stated fairly simply. Some economists favor the use of bank credit as the monetary aggregate because they view policy as operating through changes in the cost and availability of credit. The major difficulty with this view is that there is no unambiguous way of defining the amount of credit in the economy. And even if a satisfactory definition could be worked out, there is no current possibility of obtaining timely data on the total amount of credit or of controlling the total amount. The definitional problem arises largely from the activities of financial intermediaries. Suppose, for example, that an individual sells some corporate debentures and invests the proceeds in a fixed-income type of investment fund, which in turn uses the funds to buy the very same debentures sold by the individual. If both the debentures and the investment fund shares are counted as part of total credit, then in this example total credit has risen without any additional funds being made available to the corporation to finance new facilities and so forth. As another example, it is difficult to see that it would make any substantial difference to aggregate economic activity whether a corporation financed inventories through sales of commercial paper to the public or through borrowing from banks that raised funds through sales of CD’s to the public. Since there are numerous close substitutes for bank credit, the amount of bank credit is most unlikely to be an appropriate figure to emphasize. Furthermore, since bank credit is only a small part of total credit there is essentially no possibility of controlling total credit, however defined, through adjustments in bank credit. 494 J U LY / A U G U S T 2008 Ultimately the issue again becomes that of the stability of various functions. If the demand and supply functions for all of the various credit instruments, including those of financial intermediaries, were stable and were known, then it would be possible to focus on any aggregate that was convenient. For if all the functions were known, then there would be known relationships among various credit instruments, the money stock, and stocks and flows of goods. But the demand and supply functions for the various credit instruments are not known, and it is unlikely that they ever will be known with any degree of precision. There are two basic reasons for this state of affairs. The first, and less important, is that given the great degree of substitutability among credit instruments, substitutions are constantly taking place as a result of changes in regulations, including tax regulations. But second and more important, individual credit instruments are greatly influenced by changes in tastes and technology, factors that economists do not understand well. As an example of the effects of regulations, consider the substitution in recent years of debentures for preferred stock as a result of the tax laws permitting deduction of interest. As examples of the effects of changes in tastes and technology, consider the inventions of new instruments such as CD’s and the shares in dual-purpose investment funds. Furthermore, the relationships among credit instruments will change as attitudes toward risk change due to numerous factors including perhaps fading memories of the last recession or depression. Money viewed as the medium of exchange seems to be substantially less subject to changes in tastes and technology than do other financial assets. Of course, money is not immune to these problems, as shown by the uncertainty presently existing over the impact of credit cards. But a great deal of empirical work on money has been completed and the major findings have been substantiated by a number of different investigators. And the interpretation of the empirical findings is usually clear because the empirical work has been conducted within the framework of a welldeveloped theory of money. There is, on the other F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole hand, no satisfactory theory of bank credit to guide empirical work and to permit interpretation of the significance of empirical findings. For these reasons, and others, bank credit does not appear to be an appropriate monetary aggregate for policy to control. However, because bank credit and the money stock were so highly correlated in the past, it must be admitted that it probably would not have made much difference which one was used. From recent experience, however, it appears that changes in banks’ nondeposit sources of funds are likely to become more, rather than less, important, and so in the future the correlation between money and bank credit is likely to be lower than in the past. If this prediction is correct, then the issue is a significant one. As a monetary aggregate, to be used for policy adjustments, the money stock has clear advantages over the monetary base and various reserve measures. These aggregates are almost always examined in adjusted form, where the adjustments allow for such factors as changes in the currency/ deposit ratio, in reserve requirements, and in shifts between time and demand deposits. The adjustments are made because the effects of these various factors are understood and are thought to be worth offsetting. The adjustments have the effect of making the base an almost constant fraction of the money stock, or making total reserves an almost constant fraction of demand deposits. It obviously makes more sense to look directly at the money stock, especially since given the nature of the adjustments it is no easier to control the adjusted base or adjusted total reserves than to control the money stock. The final aggregate to be considered is the broadly defined money stock—the narrow stock plus time deposits. No strong case can be made against the broad money stock. From existing empirical work both definitions of money appear to work equally well. The theoretical distinction between demand deposits and passbook savings deposits depends on the costs of transferring between the two types of deposits, and these costs appear to be quite low. However, CD’s do appear to be theoretically different and probably should be excluded from the definition of money. The major reason for excluding all time deposits from F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W the definition is that in the future banks may invent new instruments that will be classified as time deposits for regulatory purposes but for which the matter of definition as money may not be at all clear. The issue of controllability is a technical one and need not be discussed carefully in this summary. However, two conclusions may be stated. First, instructions from the FOMC to the Manager of the Open Market Account should take the form of a specified average weekly change in the money stock over the period between FOMC meetings. Such an instruction must be distinguished from one in terms of the average level of the money stock over the period between FOMC meetings. The average-level specification has several technical difficulties and should be avoided. The second conclusion is that it is possible to control the rate of growth of the money stock over a 3-month period in a range of 1 percent on either side of a desired rate of growth. This conclusion is based on an analysis of monthly changes in the money stock over the 1951-68 period, a period during which little or no attention was paid to stabilizing monetary growth, and it takes the historical record at face value. Assuming that efforts to control the money stock would in fact succeed in part rather than make money growth less stable than in the past, the estimate of plus or minus 1 percent is an upper limit to the errors in controlling the growth rate of money over 6month periods. Concluding Remarks The orientation throughout this study has been the redirection of monetary policy on the basis of currently available theory and evidence. The recommendations are not utopian; in the author’s view they are supported by current knowledge and are operationally feasible. The approach has been in terms of what ought to be done in the near future, rather than in terms of what might be done eventually if enough information accumulates. No effort has been made to slide over gaps in our knowledge; rather, the emphasis has been on how policy should be formed given the huge gaps in our knowledge. Indeed, it is precisely these J U LY / A U G U S T 2008 495 Poole gaps in our knowledge that lead to the conclusion favoring policy adjustments through the money stock. It is the contention of this study that policy can be improved if there is explicit recognition of the importance of uncertainty. As much attention should be given to the consequences of errors in projections as to the projections themselves. Policy may be improved more by “don’t know” answers to questions than by projections believed by no one. This is the static view. If policy can be improved now through greater attention to uncertainty, in the long run it can be improved further only through a reduction in uncertainty. This longer view underlies the proposal for a policy rule-of-thumb. Policy successes and failures ought to be incorporated into a policy design in a form that will repeat the successes and prevent the recurrence of the failures. Policymaking will always require judgment, but the judgment will be applied to changing problems at a moving frontier of knowledge. A systematic formulation of policy will speed the accumulation of knowledge so that the policy problems of today will become the technical staff problems of tomorrow. REFERENCES Andersen, Leonall C. and Jordan, Jerry L. “Monetary and Fiscal Actions: A Test of Their Relative Importance in Economic Stabilization.” Federal Reserve Bank of St. Louis Review, November 1968, pp. 11-24. Ando, Albert and Modigliani, Franco. “The Relative Stability of Monetary Velocity and the Investment Multiplier.” American Economic Review, September 1965a, 55, pp. 693-728. Ando, Albert and Modigliani, Franco. “Rejoinder.” American Economic Review, September 1965b, 55, pp. 786-90. Brainard, William. “Uncertainty and the Effectiveness of Policy.” American Economic Review: Papers and Proceedings of the 79th Annual Meeting of the American Economic Association, May 1967, 57, pp. 411-25. 496 J U LY / A U G U S T 2008 Brunner, Karl and Meltzer, Allan H. “The Federal Reserve’s Attachment to the Free Reserve Concept.” Subcommittee on Domestic Finance, Banking, and Currency Committee, House of Representatives, 88th Congress, 2nd Session. Washington, DC: Government Printing Office, 1964. de Leeuw, Frank and Gramlich, Edward. “The Federal Reserve–MIT Econometric Model.” Federal Reserve Bulletin, January 1968, 54, pp. 11-40. DePrano, Michael and Mayer, Thomas. “Tests of the Relative Importance of Autonomous Expenditures and Money.” American Economic Review, September 1965a, 55, pp. 729-52. DePrano, Michael and Mayer, Thomas. “Rejoinder.” American Economic Review, September 1965b, 55, pp. 791-92. Fels, Rendigs and Hinshaw, C. Elton. Forecasting and Recognizing Business Cycle Turning Points. New York: National Bureau of Economic Research, 1968. Friedman, Milton and Meiselman, David. The Relative Stability of Monetary Velocity and the Investment Multiplier in the United States, 18971958. Commission on Money and Credit, Stabilization Policies. Englewood Cliffs, NJ: Prentice-Hall, Inc., 1963. Friedman, Milton and Meiselman, David. “Reply to Donald Hester.” Review of Economics and Statistics, November 1964, 46, pp. 369-76. Friedman, Milton and Meiselman, David. “Reply to Ando and Modigliani and to DePrano and Mayer.” American Economic Review, September 1965, 55, pp. 753-85. Hester, Donald D. “Keynes and the Quantity Theory: A Comment on the Friedman-Meiselman CMC Paper.” Review of Economics and Statistics, November 1964a, 46, pp. 364-68. Hester, Donald D. “Rejoinder.” Review of Economics and Statistics, November 1964b, 46, pp. 376-77. Holmes, Alan R. “Operational Constraints on the Stabilization of Money Supply Growth,” Controlling F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole Monetary Aggregates. Boston: Federal Reserve Bank of Boston, 1969. Holt, Charles C. “Linear Decision Rules for Economic Stabilization and Growth.” Quarterly Journal of Economics, February 1962, 76, pp. 20-45. Laidler, David E.W. The Demand for Money: Theories and Evidence. Scranton, PA: International Textbook Company, 1969. Latané, Henry A. “Cash Balances and the Interest Rate—A Pragmatic Approach.” Review of Economics and Statistics, November 1954, 36, pp. 456-60. Latané, Henry A. “Income Velocity and Interest Rates—A Pragmatic Approach.” Review of Economics and Statistics, November 1960, 42, pp. 445-49. Mincer, Jacob, ed. Economic Forecasting and Expectations. New York: National Bureau of Economic Research, 1969. Moore, Geoffrey H. and Shiskin, Julius. “Indicators of Business Expansions and Contractions.” Occasional Paper 103, National Bureau of Economic Research, 1967. Poole, William. “Optimal Choice of Monetary Policy Instruments in a Simple Stochastic Macro Model.” Quarterly Journal of Economics, May 1970, 84, pp. 197-216. Reynolds, Lloyd G. Economics. Third Edition. Homewood, IL: Richard D. Irwin, Inc., 1969. Samuelson, Paul A. Economics. Seventh Edition. New York: McGraw-Hill, 1967. Theil, Henri. Optimal Decision Rules for Government and Industry. Amsterdam: North-Holland Publishing Company, 1964. Zarnowitz, Victor. “An Appraisal of Short-Term Economic Forecasting.” Occasional Paper 104, National Bureau of Economic Research, 1967. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W J U LY / A U G U S T 2008 497 498 J U LY / A U G U S T 2008 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W