The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper Series Should Optimal Discretionary Monetary Policy Look at Money? WP 02-04 Michael Dotsey Federal Reserve Bank of Philadelphia Andreas Hornstein Federal Reserve Bank of Richmond This paper can be downloaded without charge from: http://www.richmondfed.org/publications/ Should Optimal Discretionary Monetary Policy Look at Money?¤ Michael Dotseyyand Andreas Hornsteinz Federal Reserve Bank of Richmond Working Paper No. 02-04 November 2002 Abstract This paper examines whether monetary indicators are useful in implementing optimal discretionary monetary policy when the policy maker has incomplete information about the environment. We …nd that money does not contain useful information for the policy maker, if we calibrate the model to the U.S. economy. If money demand were to be appreciably less variable, observations on money could be useful in response to productivity shocks but would be harmful in response to money demand shocks. We provide an incomplete information example where equilibrium welfare declines when the money demand volatility decreases. JEL Nos: C61, E52, E58 Keywords: monetary policy, sticky prices, optimal time-consistent policy, asymmetric incomplete information ¤ We would like to thank Per Krusell, Alex Wolman, Robert King and seminar participants at the Federal Reserve Bank of Richmond for helpful comments. Errors are our own. The views expressed in this paper are those of the authors and do not necessarily represent those of the Federal Reserve Banks of Philadelphia and Richmond, or the Federal Reserve System. y Federal Reserve Bank of Philadelphia (michael.dotsey@phil.frb.org) z Federal Reserve Bank of Richmond (andreas.hornstein@rich.frb.org) 1. Introduction Over the years we have participated in many discussions concerning appropriate monetary policy actions. While most central banks use an interest rate instrument in the pursuit of monetary policy, participants in these discussions frequently suggest that a key consideration for the setting of this instrument should be the behavior of money. One of the most prominent advocates of this approach, Friedman (1969), suggests that if money demand is well behaved, then monetary policy should respond to deviations of money growth from a preset target. Although there have been periods when the behavior of money has in‡uenced the setting of short term interest rates, Friedman’s prescription has not been followed in general. This is re‡ected in empirically estimated policy rules, such as Taylor (1993), which suggest that monetary authorities adjust their interest rate instrument in response to the behavior of in‡ation and some measure of real economic activity, but not in response to the behavior of money. Is the current neglect of money in the pursuit of monetary policy justi…ed? We reconsider this issue within the context of optimal monetary policy in an explicitly speci…ed general equilibrium environment. In particular we study optimal time-consistent monetary policy in an economy where prices are sticky. If the policymaker has complete information about the state of the economy, optimal policy does respond to the state of the economy, but that state does not include the nominal money stock. Although monetary policy does not respond to the behavior of money, it may appear to do so to an outside observer, if that observer does not have complete information on the state of the economy, and the behavior of money re‡ects the behavior of the state. If the policymaker has incomplete information about the state of the economy, then optimal policy may respond to the behavior of money, if that behavior contains useful information for the policymaker about the underlying state of the economy.1 For this to hold, money demand needs to be more stable than we observe for the U.S. economy. With incomplete information, however, it is not necessarily true that more information improves welfare. In particular, we provide an example where with more stable money demand the policymaker responds more aggressively to movements in money and thereby reduces economic welfare. The plan of our paper is as follows. In section 2 we give a brief overview of monetary policy in the postwar United States and describe the extent to which the Federal Reserve System has used money to guide policy. We also investigate whether adding money to a Taylor-type rule would help explain Federal Reserve behavior. Only for the early to mid 1 Recent work on optimal policy under incomplete information include Aoki (2000), Orphanides (1998), Smets(1998), Svensson and Woodford (2000, 2001), and Tetlow (2000). 2 1970s and the early 1980s do we …nd some evidence that the Federal Reserve raised interest rates in response to high money growth. In section 3 we describe an economy with sticky prices due to staggered price setting in the spirit of Taylor (1980). In the model, real balances enter the representative agent’s utility function, and we can derive a money demand function where velocity shocks represent preference shocks. Besides velocity shocks, we also consider productivity shocks. We parameterize the model based on the U.S. economy, and calculate a linear approximation to the optimal time-consistent monetary policy under full information. We characterize optimal monetary policy through impulse response functions and the behavior of estimated Taylor rules. We …nd that the optimal monetary policy responds to both shocks, but that the e¤ects of productivity shocks dominate. In response to a positive productivity shock, optimal policy lowers interest rates. Taylor rules estimated in the model capture some but not all features of empirically estimated Taylor rules: while higher in‡ation is associated with higher interest rates, above average output is associated with lower interest rates. The last feature re‡ects the response to productivity shocks which are the major source of ‡uctuations in the model economy. Estimated policy rules also seem to indicate a desire for interest rate smoothing, although the underlying true policy rule has no role for this behavior. Finally, for some of the estimated policy rules, interest rates are negatively associated with money growth. In section 4 we analyze our model when the private sector continues to have full information, but the monetary authority has incomplete information. In particular, we assume that the policymaker does not observe the state of the economy, but receives a noisy contemporaneous signal of the money stock and lagged noisy information about output. Our basic …nding is that for a parameterization of the signal noise in money and the volatility of velocity, which is consistent with that observed in the United States, observing money in addition to output does not change the dynamics of the economy substantially. However, for appreciably less money demand volatility, information on money does improve the reponse to productivity shocks. This improved reponse to productivity shocks comes at a cost, in the sense that money demand shocks are now misconstrued as negative productivity shocks, and the policymaker’s response to the signal can cause a recession. Overall, the economy with a more stable money demand is actually worse o¤ in terms of unconditional expected utility. We conclude with a brief summary and some thoughts for future work. In the appendix we characterize time-consistent optimal policy for linear-quadratic control problems with incomplete and asymmetric information. 3 2. Evidence of the use of money in monetary policy In this section we provide some evidence on the use of money in monetary policy. Our discussion concentrates on post-World War II Federal Reserve policy, but it is clear from work such as Bernanke and Mishkin (1992), Clarida and Gertler (1997), and Rich (1997) that other central banks occasionally do pay attention to the behavior of some monetary aggregate. We investigate the use of money in two ways. First, we ask if the federal funds rate was adjusted in response to deviations of actual money growth from an explicitly stated money growth target. Second, we ask if the inclusion of past money growth enters signi…cantly into Taylorrule type estimates of monetary policy. 2.1. Descriptive evidence There is evidence that U.S. monetary policy responded to the behavior of M1 during the …rst half of the seventies and the …rst half of the eighties. For the …rst period, Hetzel (1981) describes how beginning in September 1972 an M1 target was speci…ed in terms of a twoquarter growth rate. Target is somewhat of a misnomer because the Fed did not conduct monetary policy with the sole intention of hitting some …xed growth rate. However, the behavior of M1 did in‡uence the setting of the funds rate at FOMC meetings and served as a device for indicating how the open market desk should vary the funds rate between meetings. For example, during the joint intervals October 1972 to August 1974, April 1975 to October 1975, and February 1977 to September 1979, the projected growth of M1 at the prevailing funds rate was above the midpoint of its tolerance range at 44 meetings and below the midpoint only 4 times. As a result, the FOMC raised the funds rate at 37 of these meetings and lowered the rate only seven times. Conversely, over the intervals September 1974 to March 1975 and November 1975 to January 1977, projected M1 growth was below target 14 times and above it on only three occasions. Over this period, the FOMC lowered the funds rate 13 times and raised it only four times. In particular, from the spring of 1973 to the fall of 1974, one could argue that the behavior of M1 constrained monetary policy in the sense that the Fed successfully hit its money growth targets over that period. Although M1 targets were de-emphasized in 1982, there is still evidence that M1’s behavior in‡uenced policy over the period 1983-85 (see Dotsey [1996])2 . During this period the ordering of variables in the FOMC’s directive to the open market desk continually changed. The February, March, and May 1983 directives emphasized the behavior of money. As the year progressed, business activity and in‡ation became increasingly important, but in 1984 2 There also exists a debate whether M1’s behavior a¤ected policy during the 1980-81 period. For di¤ering views see Broaddus and Goodfriend (1984) and Bernanke and Mishkin (1992). 4 the growth rate of money was emphasized again. As documented in Goodfriend (1993), the Fed was faced with an in‡ation scare and as a result policy tightened. The Fed’s focus on strong growth in both M1 and M2 no doubt provided useful political cover for tighter policy. But with the containment of in‡ation, emphasis on money waned and by 1985 there is no mention of it in the directive. Unfortunately, by mid-1985 the ordering of variables in the directive no longer changes and thus provides no information concerning the importance of money’s behavior for policy. Further, M1 targets were formally abandoned in February of 1987. 2.2. Statistical evidence To further investigate whether the behavior of money played any role in policy, we look at a Taylor-type reaction function augmented by the growth rate of M1, Taylor (1993) . Speci…cally we run the following regression using quarterly data, Rt = a0 + a1 gapt + a2 (log Pt ¡ log Pt¡4 ) + a3 Rt¡1 + a4 (log M 1t ¡ log M1t¡4 ) + et ; where R is the federal funds rate, gap is the output gap de…ned as the deviation of the log output from a quadratic trend where the estimated trend only uses information available during the relevant period. Thus, the gap is reestimated at each date. Pt is the consumer price index, and M1 is the money stock. The regression uses 40-quarter rolling windows and the estimated values of the coe¢cients together with their two-standard-deviation-con…dence bands are depicted in Figure 1.3 We use rolling windows because of the well-known time varying behavior of monetary policy. The coe¢cient on M1 growth is positive and takes on its largest value from the early to mid 1970’s and in the mid 1980’s, which is consistent with the descriptive evidence presented in the previous section. However, the coe¢cient is only statistically signi…cant at the 5 percent level for the earlier sub-period, but is signi…cant at the 10 percent level in the latter sub-period. Furthermore, starting in the late 1980s the con…dence bands not only include zero, but the absolute value of the coe¢cient on money growth declines. As noted, this was when the Federal Reserve formally dropped M1 targets. 3. Optimal monetary policy with full information We have seen that monetary policy occasionally responds to the behavior of money, but not always. We now ask whether monetary policy should respond to the behavior of money and, if 3 The main message of the results are not appreciably di¤erent if we replace the output gap with growth in the output gap or output growth, or if we extend our rolling windows to 60 quarters. 5 yes, under what circumstances. To answer this question we study optimal monetary policy in an explicit dynamic general equilibrium framework. This approach has the advantage that the objective of the policymaker is well-de…ned, namely the welfare of the agents in the economy. We choose to look at time-consistent or discretionary monetary policy because we feel it may be a better representation of actual policy than full commitment, although both are extreme cases. At least in the United States, policy is decided as a sequence of individual policy actions rather than adherence to a well-de…ned rule. Indeed much academic policy advice suggests that a rule-like behavior be adopted. It is, however, true that current policy actions are somewhat constrained by long-run in‡ation objectives, but concerns about future in‡ation also enter into the discretionary policymaker’s decision through his value function. He, however, takes policy decisions with respect to future in‡ation as outside his control. Therefore, policy issues that concern reputation and credibility are outside the scope of our modeling strategy. Given the experience with in‡ation scares and the variability of monetary policy over the last 30 years, neither the time-consistent nor the full commitment approach seems adequate. We feel there is something to be learned from both approaches. 3.1. The model We wish to develop a model where the behavior of money is potentially important in the design of optimal policy. We do this by incorporating two important channels through which money may in‡uence policy. First, changes in the demand for money arise from changes in preference parameters. In this setting, a monetary authority concerned with maximizing welfare may wish to react to changes in the demand for money. Second, in a world of incomplete information, money may convey useful information about state variables in the monetary authority’s reaction function. In this section we abstract from the second channel and study the case of full information. Our basic model includes an in…nitely lived representative household with preferences over consumption, leisure, and real balances. The consumption good is produced with a large number of di¤erentiated intermediate goods. Each intermediate good is produced by a monopolistically competitive …rm with labor as the single input. Each intermediate goods …rm sets a nominal price for its product, and this price is …xed for a …nite number of periods. In particular, an equal number of …rms can change their price each period. This type of staggered time-dependent pricing behavior, referred to as a Taylor contract, is a common methodology for introducing price stickiness into an otherwise neoclassical model. 6 3.1.1. The household The representative household’s utility is a function of consumption ct , real money balances mt , and the fraction of time spent working nt , "1 ( )# 1¡Á X (1 ¡ n ) ¡ 1 t 1¡½ ½ 1=½ U = E0 ¯ t log[µc½t + (1 ¡ µ) zmt mt ] +  ; (3.1) 1 ¡ Á t=0 where Â; Á ¸ 0, ½ · 1, 0 < ¯ < 1, and zmt is a preference shock. The notation Et is used to denote the expectations conditional on the information available to the household at time t. The household’s period budget constraint is Pt ct + Bt+1 + Mt+1 · Wt nt + Rt¡1 Bt + Mt + ¦t ; (3.2) where Pt (Wt ) is the money price of consumption (labor), Bt+1 (Mt+1 ) are the end-of-period holdings of nominal bonds (money), and Rt¡1 is the gross nominal interest rate on bonds.4 The agent owns all …rms in the economy, and ¦t is pro…t income from …rms. In the following we will use the term real to denote nominal variables de‡ated by the price of consumption goods, and we use lower case letters to denote real variables. In particular, real balances are de‡ated end-of-period nominal balances mt = Mt+1 =Pt . The …rst order conditions of the representative household’s problem can be written as ¸t wt =  (1 ¡ nt )¡Á ; · ¸ Pt ¸t = ¯Rt Et ¸t+1 ; Pt+1 µ ¶1=(1¡½) 1 ¡ µ Rt mt = zmt ct ; µ Rt ¡ 1 ¸t = ¤ (Rt ; zmt ) =ct ; ) ( µ ¶1=(1¡½) µ ¶½=(1¡½) 1¡µ Rt ¤ (Rt ; zmt ) ´ 1= 1 + zmt . µ Rt ¡ 1 (3.3) (3.4) (3.5) (3.6) Equation (3.3) states that the marginal utility of leisure equals the real wage weighted by the marginal utility of consumption. Everything else unchanged, the consumer will work more with higher wages. Equation (3.4) describes the optimal savings behavior of individuals. If the return to saving rises, then households will consume less today, save more, and consume more in the future. Equation (3.5) is a money demand relationship, where real money demand depends on consumption and the opportunity cost of holding money. Equation (3.6) de…nes the Lagrange multiplier on the resource constraint, that is the marginal value of consumption. 4 In an equilibrium, bonds are in zero net supply. 7 3.1.2. Firms The consumption good is the …nal output of a constant returns to scale technology, which uses a continuum of di¤erentiated intermediate goods as inputs, indexed j 2 [0; 1]. Total i"=("¡1) hR 1 ("¡1)=" , where dj output as a function of intermediate goods y (j) used is c = 0 y(j) " > 1. Producers of the …nal good behave competitively in their markets, and given prices P (j) for the intermediate goods, the nominal unit cost and price of the …nal good is ·Z 1 ¸1=(1¡") 1¡" P = P (j) dj : (3.7) 0 For a given level of production, the cost-minimizing demand for an intermediate good j is y(j) = [P (j)=P ]¡" c: (3.8) Each intermediate good is produced by a single …rm, and j indexes both the …rm and the good. Output of …rm j is a function of labor n (j) only, yt (j) = zyt nt (j), (3.9) where zy is an aggregate technology shock. Each …rm behaves competitively in the labor market, and takes wages as given. Real marginal cost in terms of …nal goods is then à t = wt =zyt . Alternatively, the average mark-up in the economy is 1=à t . Since each intermediate good is unique, intermediate goods producers have some monopoly power, and they face downward sloping demand curves (3.8). A …rm is allowed to adjust its nominal price once every J periods, and it chooses a price that will maximize the expected value of the discounted stream of pro…ts over that period. Since all intermediate goods producers are identical except for the time when they can adjust their price, we consider only symmetric equilibria where producers di¤er only according to ¤ how much time has elapsed since they last changed their price. Let p¿ ;t = Pt¡¿ =Pt denote ¤ the time t relative price of a …rm which has set its price ¿ periods ago, Pt¡¿ . These relative prices change with the price level p¿+1;t+1 = p¿ ;t for ¿ = 0; : : : ; J ¡ 2. Pt+1 =Pt (3.10) Given the sequence of relative prices the expected present value of intermediate goods producers can be de…ned recursively as5 v0;t = max f¼t (p0;t ) + Et [¢t;t+1 v1;t+1 (p1;t+1 )]g p0;t v¿ ;t (p¿ ;t ) = ¼t (p¿ ;t ) + Et [¢t;t+1 v¿ +1;t+1 (p¿ +1;t+1 )] for ¿ = 1; : : : ; J ¡ 2 5 vJ¡1 (pJ¡1;t ) = ¼t (pJ¡1;t ) + Et [¢t;t+1 v0;t+1 ] for ¿ = J ¡ 1, We have suppressed the dependence of the …rm’s value on variables other than it’s own relative prices. 8 ¿ where real pro…ts are ¼ t (p¿ ;t ) = (p¿ ;t ¡ à t ) p¡" t;¿ ct , and ¢t;t+¿ = ¯ ¸t+¿ =¸t is the discount factor according to which the representative household evaluates future consumption relative to current consumption. Assuming that the value functions are di¤erentiable, we get a recursive de…nition of the marginal values of intermediate goods producers £ ¤ 0 0 = ¼ 0t (p0;t ) + Et ¢t;t+1 v1;t+1 (p1;t+1 ) (Pt =Pt+1 ) (3.11) £ ¤ v¿0 ;t (p¿ ;t ) = ¼ 0t (p¿;t ) + Et ¢t;t+1 v¿0 +1;t+1 (p¿ +1;t+1 ) (Pt =Pt+1 ) for ¿ = 1; : : : ; J ¡ 2 0 vJ¡1;t (pJ¡1;t ) = ¼ 0t (pJ¡1;t ) : Repeated substitution for the marginal value of a …rm with preset prices yields the following representation of the pro…t-maximizing relative price choice ¤ PJ¡1 £ " " Pt¤ ¿ =0 Et ¢t;t+¿ à t+¿ (Pt+¿ =Pt ) yt+¿ = : PJ¡1 "¡1 y Pt "¡1 t+¿ ] ¿ =0 Et [¢t;t+¿ (Pt+¿ =Pt ) (3.12) If there is zero in‡ation and marginal cost is constant, the optimal relative price is a constant markup over marginal cost, ¹ = "= (" ¡ 1). In general, however, a …rm’s pricing decision depends on future marginal costs, the future aggregate price level, future aggregate demand, and future discount rates. For example, if a …rm expects marginal costs to rise in the future, or if it expects higher rates of in‡ation, it will choose a relatively higher current price for its product. We obtain an “aggregate” production function from the demand function (3.8) for intermediate goods, the production function of intermediate goods (3.9), and labor market clearing " J¡1 1 X ¡" p yt = zyt at nt with at = J ¿ =0 ¿ ;t #¡1 (3.13) Production e¢ciency requires equal use of intermediate inputs, given that the …nal goods production function is symmetric and concave with respect to intermediate inputs. Yet, in an economy with sticky prices and in‡ation, di¤erent intermediate goods producers charge di¤erent prices and therefore sell di¤erent quantities. Production in the economy with sticky prices is then in general ine¢cient, and the allocational e¢ciency coe¢cient at · 1 re‡ects the distortion introduced by unequal relative prices. Finally, these relative prices have to satisfy the following adding-up constraint, derived from the aggregate price index (3.7), J¡1 1 X 1¡" p : 1= J ¿ =0 ¿ ;t (3.14) To complete the model, we assume that aggregate productivity and the money demand 9 preference shifter follow stationary AR(1) processes ln zyt = ° y ln zy;t¡1 + uyt (3.15) ln zmt = ° m ln zm;t¡1 + umt with j° i j < 1, E [uit ] = 0, E [u2it ] = ¾ 2i , for i = y; m. 3.2. Optimal time-consistent monetary policy The policymaker follows a policy which maximizes the expected present value of the representative agent’s lifetime utility subject to the restriction that the allocation can be supported as a competitive equilibrium. The agent’s current period utility u (x; y; R) is a function of the state of the economy x, other non-predetermined variables y, and the policy instrument, which we take to be the nominal interest rate R. We assume that the policymaker cannot commit to future policy actions, and for this reason we study Markov-perfect equilibria. In a Markov-perfect equilibrium we can view policy as being determined by a sequence of independent policymakers, and today’s policymaker assumes that future policymakers will select the policy instrument as a given function of the state, Rs = Fs (xs ) for s > t. Also, given the decision rules of future policymakers, next period’s non-predetermined variables and the lifetime utility of the representative agent from period t + 1 on will be given by the functions Gt+1 (xt+1 ) and Vt+1 (xt+1 ). We represent the competitive equilibrium restrictions through a system of equations that represent the law of motion for state variables Cx , and restrictions on non-predetermined variables derived from market-clearing and optimizing behavior Cy . Given these restrictions, the policymaker chooses the nominal interest rate and non-predetermined variables optimally6 Vt (xt ) = maxRt ;yt ;xt+1 u(xt ; yt ; Rt ) + ¯Et [Vt+1 (xt+1 )] s.t. xt+1 = Cx (xt ; yt ; Rt ; ux;t+1 ) Et [Cy1 (xt+1 ; yt+1 ; Rt+1 )] = Cy0 (xt ; yt ; Rt ) yt+1 = Gt+1 (xt+1 ) and Rt+1 = Ft+1 (xt+1 ): (3.16) The utility maximizing choice implies policy functions for today’s instrument and nonpredetermined variables, Rt = Ft (xt ) and yt = Gt (xt ), and a value function Vt (xt ) which re‡ects maximal lifetime utility of the representative agent from today on. A stationary Markov-perfect equilibrium for this problem is characterized by the triple Z = (F; G; V ) such that (3.16) maps Z into itself. 6 When the policy maker has full information, the distinction between policy instruments and nonpredetermined variables has no substantive content. 10 We study a linear-quadratic approximation of the model, and a full description of our methodology can be found in the technical appendix of Dotsey and Hornstein (2001). Brie‡y, conditional on an initial guess of the steady state nominal interest rate, we can solve for the steady state of the competitive equilibrium. We then construct a linear-quadratic approximation of the objective function and a linear approximation of the competitive equilibrium constraints around this steady state. We solve for the Markov-perfect equilibrium of the linear-quadratic approximation following Svensson and Woodford (2000) and obtain an optimal policy Rt = F xt and yt = Gxt . We use the optimal policy together with the linearized law of motion for the state variables to determine the steady state of the approximation, which includes the steady state of the nominal interest rate. We adjust the initial guess of the steady state nominal interest rate until the two rates are the same. 3.2.1. The representation of our speci…c problem In a standard rational expectations equilibrium, when the policy rule speci…es the choice of policy instrument Rt as some given function of state and ‡ow variables, we treat the lagged relative prices p¿ ;t¡1 for ¿ = 0; : : : ; J ¡ 2, which have been set by …rms in the past J ¡ 1 periods as state variables. From the point of view of the planning problem, however, nominal levels are of no concern. The equations that characterize the competitive equilibrium in any given period involve real variables, relative prices, the nominal interest rate, and the in‡ation rate, but not the current-period price level. Given that the price level is arbitrary, past nominal prices impose restrictions on the current allocation only through their relative prices. To clarify this point de…ne the normalized lagged prices as q¿ ;t = p¿ ¡1;t¡1 =pJ¡2;t¡1 , for ¿ = 1; : : : ; J ¡ 2. Using the transition equation for relative prices (3.10), we can rewrite the constraint on relative prices (3.14) as " à J¡1 !µ ¶1¡" # X Pt¡1 1 1¡" 1= p0;t + q¿ ;t 1¡" pJ¡2;t¡1 J Pt ¿ =1 with qJ¡1;t ´ 1. Since the policymaker is free to choose the current in‡ation rate, the level of lagged relative prices, that is pJ¡2;t¡1 does not represent a restriction on the policymaker’s choices; it is not pay-o¤ relevant. Only normalized lagged prices constrain the policymaker’s choices and should therefore be included as state variables in a Markov-perfect equilibrium. Finally, the normalized lagged prices evolve according to q¿ ;t+1 = p¿ ¡1;t for ¿ = 1; : : : ; J ¡ 2. pJ¡2;t (3.17) We use the equations that de…ne the competitive equilibrium, (3.3), (3.4), (3.5), (3.6), (3.11), (3.13), (3.14), (3.15), and (3.17) to de…ne the dynamic constraints of the planning 11 problem. The state variables are the normalized lagged prices and the exogenous shocks, xt = [q1t ; : : : ; qJ¡2;t ; zyt ; zmt ]. A convenient choice of the ‡ow variables includes consumption, the relative price of the current price adjusting …rm, and the marginal value of …rms that 0 0 have changed their prices in previous periods, yt = [ct ; p0t ; v1t ; : : : ; vJ¡2;t ]. Solving for the behavior of these variables allows us to recover the behavior of all the other variables in the model. In order to perform our numerical analysis we have to parameterize the model economy. Table 1 lists the parameter values and implied steady state values. We choose the time preference parameter ¯ such that the annual real rate of interest is 4 percent. We select the leisure parameters  and Á such that agents work approximately 25 percent of total hours and the implied labor supply elasticity is slightly greater than one, which is consistent with estimates in Mulligan (1998). We choose ½ so that the interest elasticity of money demand is ¡0:1. This elasticity is within the bounds of most empirical estimates for M1. Likewise we choose µ to make the velocity of money equal to 1:12, which is the average value of M1 velocity over the period 1959-99. Finally, consistent with the work of Basu and Fernald (1997), ² is calibrated to yield a markup of 10 percent. The autocorrelation coe¢cients and variances for the technology shock are roughly consistent with the values used in the literature on quantitative dynamic general equilibrium models. The process for the money demand shock is derived from an M1 demand function estimated for the United States from 1970 to 1999. Thus our parameterization is broadly consistent with both the literature and U.S. data. 3.3. The optimal time-consistent policy function We now characterize our approximation of the Markov-perfect equilibrium. We have outlined our solution procedure in section 3.2. We …rst describe the steady state nominal interest rate, and then interpret the behavior of the nominal interest rate o¤ the steady state. The two parts are interrelated since in Markov-perfect equilibria the steady state depends on o¤-steady state policy decisions. Conditional on the parametrization of our economy we …nd that the steady state nominal interest rate is 3:52 percent on a quarterly basis. Given the 1 percent real interest rate, this implies an annual in‡ation rate of about 10 percent. This number is large relative to current in‡ation rates in most OECD countries, although many countries have experienced in‡ation rates of that magnitude over the last 30 years. The steady state in‡ation rate is also high relative to what would occur under a policy of full commitment. The high optimal in‡ation in this model occurs for the following reasons. In our economy, 12 the policymaker faces several distortions. There is the standard feature that real balances are too low unless the nominal net-interest rate equals zero. There are also two other distortions that tend to become more important in an economy where the policymaker cannot commit to future policy choices. First, monopolisticaly competitive …rms set their price as a markup over marginal cost, and production is ine¢ciently low. This distortion creates a desire to in‡ate, which lowers the average markup because not all …rms can adjust their price. Second, with in‡ation and staggered price setting, …rms di¤er according to the prices they charge, and the quantities they produce and sell. Because the di¤erent …rms’ intermediate goods enter …nal goods production symmetrically, this dispersion in the production allocation is ine¢cient. The optimal response to this distortion is to lower the in‡ation rate, which reduces the price and output dispersion across intermediate goods-producing …rms.7 For a policymaker who cannot commit to future actions, the relative importance of the markup and relative price distortion changes directly with the number of periods for which prices are …xed. Loosely speaking, if prices are preset for a longer duration, then the impact of contemporaneous in‡ation on the current average markup increases, and the impact on the current relative price distortion decreases. This happens because there are relatively more …rms with preset prices, and the fraction of …rms whose relative price can be a¤ected declines. Since the policymaker cannot commit to future actions, he tends to focus on the contemporaneous impact of his actions and discounts the impact on future average markups and relative price distortions. In experiments, not reported here, we have found that the steady state in‡ation rate is reduced by one-half if J = 2, and that the steady state in‡ation rate is close to the Friedman rule if there is full commitment. Thus our solution leads to two observations. One is that monetary authorities have done better than the discretionary equilibrium, perhaps because they have access to some form of commitment technology. The other is that time-dependent pricing may not be an accurate depiction of the way …rms behave. We conjecture that a state-dependent pricing mechanism would result in greater price ‡exibility and less incentive for the monetary authority to in‡ate. Optimal policy lowers the interest rate in response to higher productivity and increases the interest rate in response to higher money demand. The response of the interest rate to percentage deviations of the state variables from their steady state values is8 ^ t = 0:22^ R q1t + 0:019^ q2t ¡ 0:13^ zyt + 0:0058^ zmt : We have found Figure 2 useful for interpreting how policy responds to deviations of 7 8 For a detailed discussion see Khan, King, and Wolman (2000). Hats denote percentage deviations from steady state values. 13 state variables from their steady state. Figure 2 displays the inital relative price p0t and marginal cost à t , conditional on the contemporaneous policy and state variables Rt , q1t ,. . . , 0 qJ¡2;t , zyt , zmt , and next period’s variables ¸t+1 , p1;t+1 , v1;t+1 . The initial relative price and marginal cost re‡ect the two distortions the policymaker tries to a¤ect: a higher marginal cost is equivalent to a lower average markup 1=à t , and a higher initial relative price implies a lower allocational e¤ciency at .9 The curve labeled LL is based on the household’s optimality conditions (3.3), (3.4), and (3.6), and market clearing conditions (3.13), (3.14). The curve labeled PP is based on the optimal price-setting equation (3.11). For our discussion we assume that the utility function is logarithmic in leisure, Á = 1, and the two curves are given by ¤ (Rt ; zmt ) 1 ; Rt ¯¸t+1 p1;t+1 = Â= f¤ (Rt ; zmt ) (zyt =ct ¡ 1=at )g ; ct = p0t Ãt £ ¤ 1¡" 0 p0t (" ¡ 1) ¡ "à t p¡1 = ¯¸t+1 v1;t+1 p1;t+1 =¤ (Rt ; zmt ) : 0t (LL) (PP) The intersection of the two curves represents a “temporary” equilibrium in which both the …rm’s and the agent’s …rst order conditions are satis…ed. It is a “temporary” equilibrium, since it is conditional on given values of future variables that may depend on current choices. Both curves are upward sloping, but numerical analysis shows that at the steady state the LL curve is steeper than the PP curve. Regarding the LL curve, a higher p0 implies a higher in‡ation rate Pt+1 =Pt = p0t =p1;t+1 , that is a lower real rate, and therefore higher consumption today. Labor demand increases because production increases, and due to the higher p0 production is less e¢cient (a declines). Overall the real wage and therefore marginal cost has to increase along LL. The left-hand side of equation (PP) is proportional to the negative of marginal pro…t of a price-adjusting …rm, and because the steady state marginal pro…t of such a …rm is negative and period pro…ts are concave in p0 , the direct e¤ect of an increase in p0 is positive. For given values of future variables, marginal cost has to increase to maintain the equilibrium. The policymaker faces a trade-o¤ between the markup and the relative price distortion. A higher nominal interest rate increases the real interest rate, consumption declines, and the LL curve shifts down. A higher nominal interest rate also shifts down the PP curve, since ¤ 9 From equations (3.13) and (3.14), the elasticity of the relative price distortion with respect to the relative price is @a¡1 p0 p1¡² p0 a¡1 ¡ 1 0 =J = ² > 0: 1¡² ¡1 @p0 a 1 ¡ p0 =J p0 a¡1 Notice that the magnitude of the derivative declines with the number of periods J over which prices cannot be adjusted. 14 is decreasing in the nominal interest rate. For our parametrization, the e¤ect of a nominal interest rate change on the PP curve is small relative to the e¤ect on the LL curve, and we will ignore the e¤ect on the PP curve. Given the positive slope of the PP curve, marginal cost and the initial relative price increase. We interpret the response of optimal policy to deviations of state variables from their steady state as an attempt of the policymaker to return the economy to a desired combination of markup and relative price distortion. Consider …rst an increase in the normalized lagged price qj for j = 1; 2. Near the steady state the relative price distortion is increasing in qj , that is allocational e¢ciency a declines. The decline in a reduces the denominator of the term on the right hand side of (LL) and marginal cost must rise for any given value of p0 : Therefore, the LL curve shifts up. This shift reduces the marginal cost à and the initial relative price p0 , that is it increases the average markup and and increases allocational e¢ciency. In order to o¤set the impact on the distortions, the policymaker increases the nominal interest rate and reverses the shift of the LL curve. Notice that as the number of periods increases for which a …rm’s price is preset, the policymaker’s impact on the distribution of relative prices declines, because much of the distribution is inherited, cf footnote 99. In terms of Figure 2 this means that the impact of the initial relative price p0 on the relative price distortion becomes smaller, and the (LL) curve becomes less steep. Therefore, a nominal rate increase has a bigger e¤ect on the average markup compared to the relative price distortion, and the monetary authority on average chooses a higher nominal interest rate and in‡ation rate. Now consider the interest rate response to a productivity increase. Higher productivity shifts the LL curve down, which in equilibrium lowers the mark-up distortion and increases the relative price distortion. In order to return the economy towards the original combination of distortions, the interest rate is lowered and the LL curve shifts back towards its original position. In response to a money demand increase, the PP curve shifts down very slightly implying a higher markup and lower p0 : In order to move back in the direction of the steady state trade-o¤ between the two distortions, the LL curve must also shift down. This shift is accomplished by an increase in the nominal interest rate. In each case the interest rate response is to return the economy toward its steady state equilibrium. 3.4. The behavior of the economy with full information The preceding discussion gives some intuition on how a policymaker will respond to various shocks. The discussion, however, holds …xed the values of future variables, and it does not consider the dynamic interaction of private sector decisions and policy decisions. We now 15 describe the qualitative and quantiative dynamic features using impulse response functions. We also discuss whether our model generates reduced form policy rules which are similar to Taylor-rules estimated for the U.S. economy. 3.4.1. Impulse response functions We display the impulse reponse functions for a productivity shock (panel A) and a money demand shock (panel B) in Figure 3. Following a productivity shock the policymaker lowers the nominal interest rate, and the economy’s behavior is remarkably close to that of an economy without frictions. In an economy with ‡exible prices the income and substitution e¤ects of higher real wages cancel, and employment does not move. The output response then just mirrors the time path for productivity. In an economy with sticky prices, optimal monetary policy with commitment almost perfectly replicates the outcome for the economy without frictions, Khan, King, and Wollman (2000). As we can see here, optimal policy without commitment appears to be a bit more opportunistic in the sense that it generates a small but noticable increase of employment, and therfore a bigger output response. In lowering the nominal interest rate, the policymaker is able to reduce the markup, as is evident from the small increase in marginal costs, and simultaneously achieves an increase in e¢ciency, as indicated by the decline of the initial relative price. The increase in output, as well as the decline in the nominal interest rate, increases real balances. One period after the shock, adjusting …rms also lower their price as do the remaining …rms when their turn for price adjustment occurs. The e¤ects of a money demand shock are quantitatively small, relative to the e¤ects of a productivity shock. The only exception are real balances.10 For output movements, a comparison of panels A and B in Figure 3 shows that productivity shocks are much more important than are money demand shocks. Thus without observations on the underlying shocks, one is more likely to attribute a given increase in output to a productivity shock. This feature will be important when we study optimal policy under incomplete information in the next section. 3.4.2. Reduced form feedback rules In this section we wish to investigate whether monetary policy derived from optimizing behavior gives rise to estimated behavior that resembles a Taylor rule. Surprisingly, estimation 10 Our result are therefore little di¤erent from what occurs in a model where money enters separably. This is consistent with McCallum (2000), who shows that for reasonably calibrated money demand behavior, separability is not a bad approximation. 16 of policy rules with model-generated data yields statistical relationships that are somewhat similar to policy rules estimated for the U.S. economy, in the sense that monetary policy apparently increases interest rates in response to in‡ation, and that policy apparently smooths the behavior of interest rates. We present two examples, …rst the policy rule (T) based on Taylor (1993) and second the rule (CGG) based on Clarida, Gali, and Gertler (2000): ¶ µ ¶ µ Mt Pt (T ) ¡ 0:47 log yt + 0:16 log Rt¡1 ¡ 0:001 log log Rt = 0:13 + 0:20 log (0:03) (0:04) (0:001) (0:01) (0:12) Pt¡4 Mt¡1 ¶ µ ¶ µ Mt Pt+1 (CGG) ¡ 0:39 Et¡1 log yt + 0:19 log Rt¡1 ¡ 0:014 log : log Rt = 0:09 + 1:74 Et¡1 log (0:04) (0:10) (0:004) (0:01) (0:60) Pt Mt¡1 The regression coe¢cients represent averages of 200 regressions run on 100 quarters worth of data, with standard deviations in parentheses. Our model does not contain a trend, we therefore do not correct output by a potential output measure to obtain an output gap measure as is usual in the literature on policy equations11 . For speci…cation (CGG) we estimate a bivariate VAR with output and prices for each sample, and use the VAR to generate one-quarter-ahead forecasts of in‡ation and output, with standard deviations in parentheses. The qualitative properties of the two speci…cations are very similar. In both speci…cations output appears with the “wrong” sign, the coe¢cient on in‡ation has the “correct” sign, and the coe¢cient on output tends to be more signi…cant than the coe¢cient on the in‡ation rate. The positive coe¢cient on in‡ation re‡ects the fact that under optmal policy the interest rate and in‡ation behave similarly to technology and money demand disturbances. The negative coe¢cient on output re‡ects the fact that optimal policy lowers the nominal interest rate relatively aggressively when there is a positive shock to technology. Hence, output increases are inversely related to the interest rate. Both speci…cations appear to indicate a desire for interest rate smoothing, which re‡ects the high degree of autocorrelation in the nominal interest rate. This autocorrelation is a result of the autocorrelation in the shock processes and not of any fundamental desire of the monetary authority to smooth rates12 . For the (CGG) speci…cation, the regressions seems to indicate that the behavior of nominal money is important for monetary policy: nominal money growth enters with a signi…cant negative coe¢cient. 11 For each speci…cation quarterly output, the nominal interest rate, and growth rates are annualized. For speci…cation (T) in‡ation is the average in‡ation over the last four quarters. 12 Model-generated data indicate that output, in‡ation, and the nominal interest rate are all highly autocorrelated. Their …rst-order autocorrelation coe¢cients are 0.87, 0.95, and 0.90, respectively. The high autocorrelation is due to the highly autocorrelated shocks. 17 4. Optimal monetary policy with incomplete information We have seen that optimal monetary policy is unresponsive to the behavior of money when the monetary authority has full information about the state of the economy. It is unlikely, however, that the monetary authority ever sees technology or money demand shocks, nor does it observe the true price distribution. In an environment where the policymaker does not have complete information on the state of the economy, optimal policy may respond to money because the behavior of money contains information about the unobserved state. We model the notion of money as a signal for an information-constrained policymaker who never observes the true state of the economy as follows. In each period the policymaker receives two signals: a signal s0t on money at the beginning of the period, contemporaneously with the decision to be made, and a signal s1t on output at the end of the period, after a decision has been made. Given our linear approximation of the economy, the policymaker’s information about the state xt is summarized by his conditional expectations. The policymaker starts out with an expectation about the current state x¹tjt¡1 conditional on information from past periods. Given the signal s0t the policymaker updates the expectation to x¹tjt . After the equilibrium has been determined the policymaker receives the signal s1t and he updates his conditional expectation to x¹tjt+ . Using the law of motion for state variables he also forms expectations about next period’s state to x¹t+1jt . We continue to assume that private agents have full information. We do so largely to make the problem more tractable.13 We show that certainty equivalence holds in the sense that the policymaker’s decision depends only on his expected value of the state variables, Rt = Ft x¹tjt .14 We can therefore use the results on optimal full-information policy rules from the previous section, and simply replace actual values of state variables with the policymaker’s conditional expectations. The equilibrium values of all nonpredetermined variables will depend on the actual state and the policymaker’s conditional expectation of the state yt = G1;t xt + G2;t x¹tjt . We solve the inference problem for the policymaker using a methodology based on the Kalman-…lter. The inference problem is fairly complicated and the equilibrium outcome depends on the policymaker’s policy rule. Thus, the separation of inference from optimization which Svensson and Woodford (2000) have shown to hold under common incomplete information is no longer 13 As pointed out in Aoki (2000), private agents may face measurement problems less serious than those that face the central bank. Firms may have a better idea of the state of technology and consumers may not need to know the price of all goods when optimizing. Thus modeling private agents as having information superior to that possessed by the monetary authority seems a reasonable strategy. Even so, our information assumptions are severe. 14 A detailed explanation of our procedure is contained in the appendix. Similar derivations can be found in Svensson and Woodford (2001). 18 present. In most developed countries data on real GDP, nominal GDP, and the money stock are readily available for a policymaker. Furthermore, reliable information on money is available earlier than is information on GDP. We represent this information structure in our model as follows.15 At the beginning of the period, before the policy instrument is set, the policymaker receives a signal on real balances ln(Mt =Pt )o + ´ R Rt = yt + zmt + À mt , where À mt is the measurement error. In reality policymakers observe separate signals on the nominal money stock and the price level. To keep the problem tractable, we have combined both signals into the real balance signal, which conveys more information than would the nominal money stock alone. At the end of period, after the policy instrument has already been determined, the policymaker receives a signal on real output, yto = yt + À yt , where À yt is the measurement error on real output. The standard deviation of the measurement error of the nominal disturbance is 0:003, and the one for the output signal is 0:005. These values are consistent with the variance of revisions for real GDP and nominal M1.16 Using these two signals, the monetary authority forms expectations of the economy’s current state and conducts policy accordingly. 4.1. The behavior of the economy under incomplete information We …rst study how the information structure a¤ects the response of monetary policy to productivity shocks. We …nd that when the policymaker has only lagged information on output, a productivity shock results in an initial decline of output with subsequent humpshaped convergence towards the full-information path. If the policymakers receives additional contemporaneous information on money, the policymaker can respond to the productivity shock and the initial negative e¤ect on output is weakened. The more stable money demand is, the better can the policymaker respond to a productivity shock. We then show that with a stable money demand the policymaker is more likely to confuse money demand disturbances for productivity shocks. Since the policymaker responds strongly to perceived productivity shocks, this introduces undesirable volatility into the economy. For our example, having more information can reduce welfare. Finally, we estimate Taylor rules for our model economy and …nd that money growth enters signi…cantly with a negative coe¢cient in these equations. For 15 We assume that these measures are never updated, and that the authority never observes the actual state variables. For an alternative information setting in which the monetary authority observes the true state with a one period lag, see Aoki (2000) and Swensson and Woodford (2001). 16 Because we parameterize the “nominal” noise using only the standard deviation of the money stock noise, we underestimate the magnitude of the “nominal” noise, implicitely giving money more signal value than it actually has. 19 the analysis of the model economy under incomplete information, we maintain the calibration of the previous section. 4.1.1. The e¤ects of a productivity shock Figure 4 displays the economy’s response to a productivity shock: the actual values and the policymaker’s expectations of state variables in panel A, and the actual values of nonpredetermined variables in panel B. Because the policymaker has no contemporaneous information, there is no immediate response to the shock. From the standpoint of the full information case this inaction means that the nominal rate is roughly 50 basis points too high. With a one-period delay, the policymaker becomes aware of the productivity increase and understands that his inaction has resulted in a decline in q1 : In response to this new information, the policymaker now drastically lowers the nominal interest rate, but brings it back up to the full information case within three periods. We can see how incomplete information magni…es the policymaker’s response to a shock: relative to full information the nominal interest rate movements are about four times as big. In turn, the policymaker’s delayed and magni…ed response also introduces additional output and employment volatility. Firms who adjust their price in the period that the shock occurs know that it will take time for output to rise. Hence the path of wages and marginal cost will be lower than under full information and as a result …rms slash prices. This price reduction reduces the anticipated in‡ation rate, and since the nominal interest rate is unchanged, it increases the real rate of interest. Although, future consumption is expected to increase, after the policymaker recti…es the mistake, the real rate increase is su¢cient to induce a reduction in current-period consumption and output. Could the excess volatility that occurs under incomplete information relative to what occurs under full information be reduced if the policymaker were to obtain contemporaneous information on money in addition to lagged information on output? It turns out that for our parameterization of money demand uncertainty, the nominal money stock does not contain much useful information about output. The impulse response functions with contemporaneous information are essentially the same as the ones in Figure 4. Basically, money demand is too volatile for observations on money to be of much value. Since observed movements in money can be largely attributed to the money demand shock, the inference of the technology shock and of relative prices is unchanged, and observing money does not improve the policymaker’s knowledge of the state of the economy. The property that money does not contain any useful information about output and hence technology disturbances is an empirical matter. If money demand is more stable, for 20 example if we reduce the variance of the money demand shock by a factor of ten, the outcome is quite di¤erent (see Figure 5). In this situation contemporaneous money is sending a much sharper signal of contemporaneous output, and following a productivity shock the policymaker attributes much of the movement in money to that shock. The policymaker, however, also attributes some of the increased output to relatively low normalized lagged prices and a greater degree of inherited e¢ciency. The combination of beliefs leads to a decline in the nominal interest rate of 50 basis points, almost the identical decline as under full information. Yet output does not respond in nearly the same way as it does under full information. The key to the di¤erence is subtle and involves the fact that the policymaker only gradually learns the true nature of the shock, but that the private sector already anticipates further nominal rate reductions induced by this learning process. As the policymaker learns more about the productivity shock, he implements further substantial nominal rate reductions that increase output further and reduce prices. The anticipation of big price reductions in the future induces current producers to reduce their prices substantially, and the expected in‡ation rate declines and the real rate increases. The real rate increase is then consistent with the hump-shaped response of output. 4.1.2. The e¤ects of a money demand shock Money demand shocks do not play an important role under full information. We would then expect that in an incomplete information environment a policymaker might ignore money demand shocks alltogether. This is actually the case when the policymaker has only lagged information on output.17 We have also seen that the policymaker’s response to a productivity shock is improved when he has contemporaneous information on money and money demand volatility is low. However, more information is not necessarily welfare improving in our environment. When the policymaker believes that the contemporaneous signal is a more reliable indicator of productivity shocks, the consequences are more severe when he mistakes a money demand shock for a productivity shock. Figure 6 shows the e¤ects of a money demand shock when the central bank observes money and money demand is stable; that is the variance of the money demand shock is one-tenth of our baseline speci…cation. With this speci…cation, the policymaker believes not only that a money demand shock has occurred, but also believes that the economy has been subject to a negative technology disturbance. This latter result occurs, because the inference problem and the policy response are no longer separate. Note that the contemporaneous 17 We do not report the results here, but in this case the policy maker never updates his estimate of the money demand shock, and he responds only to perceived productivity and normalized lagged price shocks. 21 signal contains the sum of the money demand shock and the output response (ct + zmt + À mt < 0). In our case the planner sees a negative signal, which he interprets as a negative productivity shock and a positive money demand shock. The policymaker ignores the money demand shock, but responds to the productivity shock and increases the nominal rate; the higher interest rate, in turn, reduces consumption, which leads to a negative signal. Thus the response and the inference are internally consistent. Unlike the case when the shock is actually a productivity shock, more information is not preferable and less variability in money demand actually inhibits accurate inference of the true disturbance. Simulations of the model economies show that the economy with contemporaneous observation on money and low money demand volatility is much more volatile than the economy with information on lagged output only. This indicates that more information is not necessarily welfare improving in our environment.18 We can verify this for our linear-quadratic approximations of the utility functions. In particular, we calculate the consumption equivalent welfare loss relative to full information (1) with lagged output only as 0:0072 percent of steady state consumption; (2) with contemporaneous money and high money demand volatility as 0:0058 percent; and (3) with contemporaneous money and low money demand volatility as 0:0713 percent. 4.1.3. Reduced form feedback rules Policy rules estimated from data generated by a model where the policymaker has incomplete information show a policymaker who is more responsive to money stock changes. When we estimate the two equations (T) and (CGG) from section 3.4.2 for the baseline parameterization of the incomplete information model with a contemporaneous money signal and a lagged output signal, we obtain µ ¶ µ ¶ Pt Mt (T ) ¡ 0:76 log yt + 0:33 log Rt¡1 ¡ 0:11 log log Rt = 0:14 + 0:10 log (0:17) (0:10) (0:02) (0:02) (0:086) Pt¡4 Mt¡1 µ ¶ µ ¶ Pt+1 Mt (CGG) log Rt = 0:12 + 1:75 Et¡1 log ¡ 0:65 Et¡1 log yt ¡ 0:15 log Rt¡1 ¡ 0:06 log : (0:02) (0:20) (0:16) (0:09) (0:02) Pt Mt¡1 The Taylor speci…cation (T) now displays a weaker response to annual in‡ation, a stronger negative response to the output gap, more interest rate smoothing, and a signi…cant negative response to money growth. The Clarida-Gali-Gertler speci…cation displays an unchanged positive response to predicted in‡ation, a stronger negative response to the predicted output 18 This is not an entirely new observation. Pearlman (1992) has pointed out that more information need not be welfare improving in optimal control problems with partial information, even when the inference and the control problem are separable. 22 gap, no signi…cant interest rate e¤ect, and also a statistically negative dependence on money growth. Incomplete information makes the dynamics of the model su¢cienctly complicated such that the simple policy rules no longer capture the dynamics of the interest rate so well.19 The increased importance of the model’s internal propagation relative to the exogenous productivity shock is also re‡ected in the reduced persistence of variables. The …rst-order serial correlation coe¢cients on output, in‡ation, and the nominal interest rate are now .63, .66, and .51. Thus incomplete information and its subsequent e¤ect on policy has a signi…cant e¤ect on how an econometrician would view this economy. This is true even though the private sector’s behavior is unchanged. 5. Conclusion In this paper we have attempted to evaluate how useful money is for the pursuit of monetary policy. We have done so for an environment where we have explicitly speci…ed why money has real e¤ects (sticky prices due to staggered price setting), and what are the policymaker’s objective (maximize the expected utility of the representative agent) and constraints (choose a time-consistent policy rule). We have found that even though money communicates information on aggregate output, it is of limited use for a policymaker. We should emphasize that money’s usefulness as a signal is an empirical matter, and that if the money demand was more stable than it appears to be, the value of money as a signal could dramatically increase. In particular money would be a useful signal in an environment driven by productivity shocks, but using it as a signal would have adverse consequences in the presence of money-demand disturbances. This …nding suggests that time variation in the behavior of money-demand disturbances and in the types of shocks primarily a¤ecting the economy could imply time variation in a policymaker’s responsiveness to money. Importantly, however, the dynamic behavior of an economy can be very di¤erent depending on what information is available to the policymaker. The dynamic behavior di¤ers because policy responds very di¤erently under full information from the way it responds under incomplete information. As stressed in Dotsey (1999), the form of the policy rule can be crucial in the type of model studied here. Thus, it may pay to model the economy’s information structure accurately if one is to explain economic behavior. Information and signal extraction problems were a centerpiece of the early literature on rational expectations, these same items may again prove to be important for the analysis of optimal monetary policy. 19 The average adjusted R2 for the two speci…cations are 0:73 (0:05) for (T) and 0:85 (0:04) for (CGG). Under complete information they were 1:03(0:01) and :91(:05); respectively. 23 Finally, the joint speci…cation of private pricing decisions and discretionary policy tends to produce somewhat high steady state rates of in‡ation. It would be interesting to relax both of these features of the model by allowing for state-dependent pricing decisions or some form of partial commitment. We hope to investigate these avenues in future work. 24 References [1] Aoki, K. 2000. Optimal commitment policy under noisy information. Kobe University working paper. [2] Basu, S., Fernald, J., 1997. Returns to scale in U.S. production: estimates and implications. Journal of Political Economy 105, 249-83. [3] Bernanke, B., Mishkin, F. 1992. Central bank behavior and the strategy of monetary policy: observations from six industrialized countries. In Blanchard, O., Fischer, S. (Ed.) NBER Macroeconomics Annual 1992, MIT Press, Cambridge, 183-227. [4] Brayton, F., Levin, A., Tryon, R., Williams, J. 1997. The evolution of macro models at the Federal Reserve Board. Carnegie-Rochester Conference Series on Public Policy 47, 43-82. [5] Broaddus, A., Goodfriend, M.,1984. Base drift and the longer run growth of M1: experience from a decade of monetary targeting. Federal Reserve Bank of Richmond Economic Review 70, 3-14. [6] Calvo, G., 1983. Staggered prices in a utility maximizing framework. Journal of Monetary Economics 12, 383-398. [7] Clarida, R., Gali J., Gertler M., 2000. Monetary policy rules and macroeconomic stability: evidence and some theory. Quarterly Journal of Economics 115, 147-180. [8] Clarida, R., Gertler, M., 1997. How the Bundesbank conducts monetary policy. In Romer, C., Romer, D. (Ed.) Reducing in‡ation: motivation and strategies. University of Chicago Press, Chicago, 363-403. [9] Dedola, L. 2000. Essays in monetary and exchange rate policies. University of Rochester Ph.D. Thesis. [10] Dotsey, M.1996. Changing policy orientation in the United States. In: Siebert H. (Ed.) Monetary policy in an integrated world economy. Institut für Weltwirtschaft an der Universität Kiel, 95-113. [11] Dotsey, M.. 1999. The importance of systematic monetary policy for economic activity. Federal Reserve Bank of Richmond Economic Quarterly, 85, 41-60. [12] Goodfriend, M. 1993. Interest Rate Policy and the In‡ation Scare Problem: 1979-1992; Federal Reserve Bank of Richmond Economic Quarterly, Winter 1993, v. 79, iss. 1, 1-24 25 [13] Hetzel, R. 1981. The Federal Reserve System and control of the money supply in the 1970’s. Journal of Money, Credit, and Banking 13, 31-43. [14] Khan, A., King, R., Wolman, A. 2000. Optimal monetary policy. Federal Reserve Bank of Richmond Working Paper No. 00-10. [15] McCallum, B., 2000. Monetary policy analysis in models without money. Manuscript. [16] Mulligan, C., 1998. Substitution over time: another look at life-cycle labor supply. In: Bernanke, B., Rotemberg, J. (Ed.) NBER Macroeconomics Annual. MIT Press, Cambridge, pp [17] Orphanides A., 1998. Monetary policy evaluation with noisy information. Finance and Economics Discussion Series Working Paper No. 1998-50, Federal Reserve Board, Washington D.C. [18] Pearlman, J.G. 1992. Reputational and nonreputational policies under partial information. Journal of Economic Dynamics and Control 16, 339-357. [19] Rich, G., 1997. Monetary targets as a policy rule: lessons from the Swiss experience. Journal of Monetary Economics 39, 113-141. [20] Smets, F., 1998. Output gap uncertainty: does it matter for the Taylor rule? Bank for International Settlements Working Paper No.60 [21] Svensson, L., Woodford, M., 2000. Indicator variables for optimal policy. NBER working paper 7953. [22] Svensson, L. and Woodford, M., 2001. Indicator variables for optimal policy under asymmetric information. manuscript. [23] Swanson, E., 2000. On signal extraction and non-certainty equivalence in optimal monetary policy rules. manuscript. [24] Taylor, J. B. 1980. Aggregate dynamics and staggered contracts. Journal of Political Economy 88, 1-23 [25] Taylor J. 1993. Discretion versus policy rules in practice. Carnegie Rochester Conference Series on Public Policy 39, 195-214. [26] Tetlow, R. 2000. Uncertain potential output and monetary policy in a forward-looking model. Manuscript. 26 Technical Appendix A1. Introduction This appendix describes the algorithms we use to …nd the steady state of an economy with discretionary optimal policy and incomplete information. We …rst describe a Markovperfect equilibrium for the optimal control problem when the policymaker cannot commit to future policy choices. We describe a simple algorithm to …nd a linear approximation to the equilibrium. For this case we assume that the policymaker has complete information. We then consider the case where the policymaker has incomplete information, in particular he has less information than the private sector. This analysis takes as a starting point the linear approximation of the environment. The description of the linear-quadratic optimal control and estimation under incomplete information essentially follows Svensson and Woodford (2000). The section on incomplete information extends SW to the case of asymmetric information between the private sector and the policymaker. A2. The model with full information In this section we describe the optimal control problem of a policymaker under full information. The policymaker chooses allocations which satisfy the constraints that support the outcome of a market equilibrium given the actions of the policymaker. The policymaker maximizes an intertemporal objective function. For our applications this objective function is the expected present value of the representative agent’s utility. We assume that the policymaker cannot commit to future policy choices, therefore we study a Markov-perfect equilibrium. The constraints of the policymaker, are the …rst order conditions and market clearing conditions of the competitive equilibrium. It is useful to divide these constraints into two blocks: one that contains the evolution of the predetermined state variables, x; and denoted Cx , and the other that involves the non-predetermined ‡ow variables, y; and denoted Cy : Formally, the constraints can be represented by xt+1 = Cx (xt ; yt ; Rt ; ux;t+1 ) Et Cy1 (xt+1 ; yt+1 ; Rt+1 ) = Cy0 (xt ; yt ; Rt ) (5.1) (5.2) where R denotes the policy instruments and u is an iid random variable with mean zero. There are nx state variables, ny ‡ow variables, and nR instruments. De…ne Zt ´ [x0t ; yt0 ; Rt0 ]0 . Equation (5.1) de…nes the law of motion for the state variables, and equation (5.2) re‡ects the fact that the chosen allocation has to satisfy the private sector’s optimality conditions 27 in a market economy. The function Cx is vector-valued of dimension nx, Cy0 and Cy1 are vector-valued of dimension ny , and Et denotes the private sector’s expectations conditional on the information set It . In this section we assume that the policymaker and the private sector have complete information at time t, It = fZ¿ : ¿ · tg. Note that (5.1) and (5.2) de…ne an incomplete dynamic system, since the policymaker’s decision rule with respect to the policy variables has not been speci…ed. If we were to specify the policy variables as a function of the state and/or ‡ow variables, such as in a Taylor-rule, we could solve (5.1) and (5.2) for the implied rational expectations equilibrium. The objective function of the policymaker is E0 1 X ¯ t U (Zt ) (5.3) t=0 where 0 < ¯ < 1. In our example U is the utility function of the representative agent, which generates the competitive equilibrium. We are looking for a time-consistent policy. For this purpose we solve for the set of Markov-perfect equilibria where the policymaker’s decision depends only on pay-o¤ relevant state variables, that is Rt = Ft (xt ) (5.4) and the equilibrium outcome is such that the ‡ow variables depend only on the current state yt = Gt (xt ) : (5.5) For the Markov-perfect equilibrium, the policymaker takes next period’s outcome functions Ft+1 and Gt+1 and the continuation value function Vt+1 (xt+1 ) as given, and the optimization problem is Vt (xt ) = max Et [U (xt ; yt ; Rt ) + ¯Vt+1 (xt+1 )] xt+1 ;yt ;Rt (5.6) s.t. xt+1 = Cx (xt ; yt ; Rt ; ux;t+1 ) (5.7) Et [Cy1 (xt+1 ; Gt+1 (xt+1 ) ; Ft+1 (xt+1 ))] = Cy0 (xt ; yt ; Rt ) : (5.8) A Markov-perfect equilibrium is characterized by the triple (F; G; V ) such that (5.6), (5.7), and (5.8) maps (Ft+1 ; Gt+1 ; Vt+1 ) = (F; G; V ) to (Ft ; Gt ; Vt ) = (F; G; V ). For the solution we proceed in several steps. We …rst construct a linear-quadratic approximation of the problem around a steady state indexed by the policy instrument R¤ . We then ¤ solve the LQ approximation, and derive the steady state of the approximation RLQ . This de…nes a mapping from R¤ to R¤LQ . We solve for a steady state choice of policy instruments ¤ such that R¤ = RLQ . 28 Solving for an approximate steady state Step 1. Suppose the steady state values of the policy instruments are given by R¤ . Conditional on R¤ we can solve (5.1) and (5.2) for the steady state values of the state and ‡ow variables (x¤ ; y ¤ ). Now derive a linear approximation of the constraints (5.1) and (5.2) xt+1 = Cx;Z Zt + ux;t+1 Et [Cy1;Z Zt+1 ] = Cy0;Z Zt (5.9) (5.10) and a quadratic approximation of the period utility function20 2 32 3 Uxx Uxy Uxz xt 6 76 7 Zt0 UZt = [x0t ; yt0 ; Rt0 ] 4 Uyx Uyy Uyz 5 4 yt 5 : Uzx Uzy Uzz Rt Step 2. Given the linear-quadratic structure, we guess that next period’s non-predetermined variables are linear functions of next period’s state variable and that the continuation value from next period on is a quadratic function of next period’s state variable. Speci…cally, Rt+1 = Ft+1 xt+1 (5.11) yt+1 = Gt+1 xt+1 (5.12) and x0t+1 Vt+1 xt+1 is the continuation value: We now determine the equilibrium outcome for the current period and next period, conditional on the current state and policy choice. Assume that the matrix " # Ix 0xy C1;t = Cy1;x + Cy1;R Ft+1 Cy1;y is invertible, then we can rewrite the constraints as " # " # " #" # " # " # xt+1 C A A x B u x xx xy t x x;t+1 ¡1 = C1;t Zt = + Rt + : (5.13) Et yt+1 Cy0;Z Ayx;t Ayy;t yt By;t 0 We now proceed as in Svensson and Woodford (2000). Substitute the transition equations for the state variables (5.13) in our guess for next period’s ‡ow variables (5.12) and take conditional expectations Et [yt+1 ] = Gt+1 (Axxxt + Axy yt + Bx Rt ) : 20 In the following we will interpret the variables in terms of deviations from the steady state. Furthermore, the …rst derivative of the utility function is implicitly included by adding a constant term to the vector of 0 state variables [1; x0 ] . We also have normalized the covariance matrix of ux such that the derivative of Cx;ux = I. 29 Now substitute these expectation for next period’s ‡ow variables in the optimality conditions from the competitive equilibrium (5.13) and get Gt+1 (Axx xt + Axy yt + Bx Rt ) = Ayx;t xt + Ayy;t yt + By;t Rt : Assuming that Ayy;t ¡ Gt+1 Axy is invertible, we can solve for this period’s ‡ow variables ~t Rt with yt = A~t xt + B A~t ´ (Ayy;t ¡ Gt+1 Axy )¡1 (Gt+1 Axx ¡ Ayx;t ) (5.14) ~t ´ (Ayy;t ¡ Gt+1 Axy )¡1 (Gt+1 Bx ¡ By;t ) : B Substituting (5.14) for this period’s ‡ow variables in the transition equation (5.9) yields xt+1 = A¤t xt + Bt¤ Rt + ux;t+1 with (5.15) A¤t ´ Axx + Axy A~t Bt¤ ´ Bx + Axy;t Bt : After substituting for yt from (5.14) and for xt+1 from (5.15) in the period t optimization problem we get © £ ¤ª max Zt0 UZt + ¯Et x0t+1 Vt+1 xt+1 s.t. (5.14) and (5.15). h i 0 0 max xt ; Rt Qxx;t Qxz;t #" xt # Qzx;t Qzz;t Rt £ ¤ ¤ +¯E (At xt + Bt¤ Rt + ux;t+1 )0 Vt+1 (A¤t xt + Bt¤ Rt + ux;t+1 ) Rt with " Qxx;t Qxz;t Qzz;t i h = Ix ; A~0t " Uxx Uxy #" Ix A~t # Uyx Uyy (" # " #" #) h i U U U 0 xz xx xy = Ix ; A~0t + ~t Uyz Uyx Uyy B " #" # " # h i U h i U U 0 xx xy xz ~t ~t0 = Uzz + 0; B + 0; B ~ Uyx Uyy Bt Uyz " # 0 + [Uzx ; Uzy ] : ~t B The …rst order condition for the optimal choice of the policy instrument is 0 = (Qzz;t + ¯Bt¤0 Vt+1 Bt¤ ) Rt + (Qzx;t + ¯Bt¤0 Vt+1 A¤0 t ) xt 30 (5.16) (5.17) Assuming that Qzz;t + ¯Bt¤0 Vt+1 Bt¤ is invertible we can solve for the policy instrument and de…ne this period’s policy function Rt = Ft xt and ‡ow variable equation yt = Gt xt with ¡1 Ft = ¡ (Qzz;t + ¯Bt¤0 Vt+1 Bt¤ ) ~t Ft : Gt = A~t + B (Qzx;t + ¯Bt¤0 Vt+1 A¤0 t ) (5.18) (5.19) Substituting for the policy function in (5.16) yields the current period value as a quadratic function of the current period state de…ned by the matrix ( " # )" # ¤0 A I t Vt = [I; Ft0 ] Qt + ¯ Vt+1 [A¤t ; Bt¤ ] ¤0 Bt Ft (5.20) A stationary Markov-perfect equilibrium to the policymaker’s decision problem is a triple of matrices (F; G; V ) which is a …xed point of the mapping de…ned by equations (5.13) through (5.20). Step 3. Calculate the steady state of the Markov-perfect equilibrium. Substituting the policy rule F and the equilibrium function G into the transition equation for the state variables we get x¤LQ = Axx x¤LQ + Axy Gx¤LQ + Bx F x¤LQ : Recall that constant terms are implicit in this equation through the de…nition of the state variable which contains a constant term. We can solve this expression for the steady state ¤ value of the linear approximation x¤LQ and RLQ = F x¤LQ . Since we started with the assumption that the steady state of the Markov perfect equilibrium is R¤ we adjust R¤ until ¤ RLQ = R¤ . Implementation of the algorithm The implementation of the algorithm is straightforward, with two minor exceptions. The derivation of the Markov-perfect equilibrium suggests that we use a simple iteration scheme: given an assumption on the triple (Ft+1 ; Gt+1 ; Vt+1 ) use equations (5.13) through (5.20) to obtain values for the triple (Ft ; Gt ; Vt ), and iterate until convergence. There are two problems with simple iterations, which relate to the fact that we have not shown that such a process will indeed converge. First, we have found that frequently we obtain convergence on the linear terms of the functions F , G, and V , but we cannot obtain convergence once we include the constant terms in these functions. This problem is easily dealt with. Given the linear quadratic structure we know that the linear terms are independent of the constant terms, and in a …rst run we ignore the constant terms and interpret the model in terms of deviations from the steady state. Essentially this means ignoring the linear terms of the utility function. Usually we 31 obtain convergence on the linear terms for this simpli…ed model. After we have obtained the linear terms, the constant terms can be obtained as a solution to a linear system of equations. We need to know the constant terms since they will determine the steady state of the approximation economy. The second problem is that occasionally the simple iteration scheme does not converge for the linear terms. In this case we solve the equations (5.13) through (5.20) as one big system of non-linear equations. This procedure tends to be slower than simple iterations and good starting values are required. A3. The model with incomplete and asymmetric information We now consider a special case of incomplete information: we assume that the private sector of the economy has complete information, whereas the policymaker has incomplete information. In particular, we assume that the policymaker receives two types of signals about the economy. The …rst signal s0t is received at the beginning of the period, and its value is determined simultaneously with the values of the ‡ow variables and the policy instrument. The second signal s1t is received at the end of the period, after the ‡ow variables and the policy instrument have been determined. We call the information contained in the …rst and second signal contemporaneous respectively lagged. The signals are related to the true state of the economy according to 0 sit = Si [x0t ; yt0 ] + eit (5.21) where eit is iid with mean zero and covariance matrix §ei , for i = 0; 1. The policymaker’s information set at the beginning of period t is I¹t = f(Rt ; s0t ) and (R¿ ; s0¿ ; s1¿ ) for ¿ < tg and at the end of the period it is I¹t+ = I¹t [fs1t g. We denote the policymaker’s expectations about £ ¤ the current state of the economy conditional on the available information as x¹tjt = E xt jI¹t , £ ¤ respectively x¹t+1jt = E xt+1 jI¹t+ . The private sector has complete information, that is the beginning and end-of-period information sets are It = f(xt ; yt ; zt ; s0t ) and (x¿ ; y¿ ; z¿ ; s0¿ ; s1¿ ) for ¿ < tg and It+ = It [ fs1t g, respectively, and the conditional expectations are denoted xtjt and xt+1jt . The optimal control problem. We guess that next period’s policymaker’s decision is Rt+1 = Ft+1 x¹t+1jt+1 (5.22) and that next period’s ‡ow variables depend on the actual values of the state variables and 32 the policymaker’s perceptions of the state variables yt+1 = G1;t+1 xt+1 + G2;t+1 x¹t+1jt+1 : (5.23) Also assume that the continuation value of next period’s policymaker is given by £ ¤ E x0t+1 Vt+1 xt+1 jI¹t+1 : When the policymaker forecasts next period’s ‡ow variables, the distinction between actual and forecasted variables is irrelevant y¹t+1jt = Gt+1 x¹t+1jt and Gt+1 ´ G1;t+1 + G2;t+1 : Proceeding as before we can obtain the policymaker’s expectation about the current ‡ow variables ~t Rt y¹tjt = A~t x¹tjt + B and his forecast of next period’s state variables x¹t+1jt = A¤t x¹tjt + Bt¤ Rt ~t , A¤t , and Bt¤ are de…ned as in (5.15) and (5.14). where A~t , B The policymaker’s value function is now given by where lt ´ E £ ¤ 0 E Zt0 UZt + ¯x0t+1 Vt+1 xt+1 j I¹t = Z¹tjt UZ¹tjt + ¯ x¹0t+1jt Vt+1 x¹t+1jt + lt h¡ i ¡ ¢0 ¡ ¢ ¢0 ¡ ¢ Zt ¡ Z¹tjt U Zt ¡ Z¹tjt + ¯ xt+1 ¡ x¹t+1jt Vt+1 xt+1 ¡ x¹t+1jt j I¹t : (5.24) (5.25) Notice that if the policymaker’s inference and forecast errors are independent of current policy, then the term lt is independent of current policy. For this “certainty equivalence” case, optimization proceeds with respect to the policymaker’s forecasts of the state variables, and we can use the results from the full information case where we replace actual values with the policymaker’s best forecasts. Inference: the Kalman …lter In this section we solve the optimal signal extraction problem while assuming that the policy is as in the full information case. From the solution to the optimal control problem we have Rt = Ft x¹tjt and y¹tjt = Gt x¹tjt . In order to determine the realized values of the ‡ow variables we also need the components of Gt as de…ned in (5.23). Below we will construct G1t and G2t , but for now we take them as given. In the following we construct the Kalman-…lter 33 recursively. The equilibrium we study in the paper is a …xed point of the mapping implied by this procedure. We …rst use the policy rule (5.22) and the equilibrium ‡ow variables equation (5.23) to eliminate Rt and yt from the transition equation (5.13) and the signal equation (5.21): xt+1 = Jt x¹tjt + Ht xt + ux;t+1 with (5.26) Ht ´ Axx + Axy G1;t Jt ´ Axy G2;t + Bx Ft and sit = Mit x¹tjt + Lit xt + ei;t with (5.27) Lit ´ Six + Siy G1;t Mit ´ Siy G2;t These equations essentially de…ne the Kalman-…lter problem: equation (5.26) is the transition equation for the unobserved state, and equation (5.27) is the measurement equation which relates the signals to the state. The only non-standard feature is that the policymaker’s updated estimate of the state appears on the right-hand side of each equation. We now introduce the following notation for the derivation of the Kalman …lter. First, de…ne the forecast errors for the state and signals as x~t ´ xt ¡ x¹tjt¡1 and s~0t ´ s0t ¡ s¹0;tjt¡1 at the beginning of the period, and x~+ ´ xt ¡ x¹tjt t and s~1t ´ s1t ¡ s¹0;tjt at the end of the period. The forecast update is then x^t = x¹tjt ¡ x¹tjt¡1 . Second, in the standard Kalman-…lter the update of the state variable is a linear function of the signal forecast error. Here we follow Svensson and Woodford (2000) and guess that at the beginning of the period the policymaker’s update of the state is a linear function of a transformation of the signal, s~0t ¡M0t x¹tjt , x^t = K0t (L0t x~t + e0t ) (5.28) This guess will be veri…ed below. We construct the Kalman-…lter in two steps, for each signal separately. Beginning of the period. From equation (5.27) and the law of iterated expectations, the policymaker’s forecast of this period’s contemporaneous signal, based on the information available at the end of last period, is s¹0;tjt¡1 = (M0t + L0t ) x¹tjt¡1 34 We subtract this expression for the prior expectation of the contemporaneous signal from its actual value (5.27), and obtain the forecast error for the signal ¡ ¢ ¡ ¢ s~0t = L0t xt ¡ x¹tjt¡1 + M0t x¹tjt ¡ x¹tjt¡1 + e0;t = L0t x~t + M0t x^t + e0t : Substituting our guess (5.28) for x^t , the expression for the contemporaneous forecast error simpli…es to (5.29) s~0t = N0t x~t + º t with N0t ´ (I + M0t K0t ) L0t º t ´ (I + M0t K0t ) e0;t and §º;t ´ Et [º t º 0t ] = (I + M0t K0t ) §e0 (I + M0t K0t )0 : From equation (5.29) we get the update of the state variable x^t as a linear projection of x~t onto s~0t £ ¤ £ ¤ + + ¡1 ^ 0t s~0t s~0t = K (5.30) x^t = E x~t s~00t jI¹t¡1 E s~0t s~00t jI¹t¡1 ¡ ¢ ¡1 0 ^ 0t ´ Ptjt¡1 N0t0 N0t Ptjt¡1 N0t + §º;t and with K £ 0 + ¤ Ptjt¡1 ´ E x~t x~t jI¹t¡1 . Based on the posterior of the state variable we get an expression for the state variable forecast error ^ 0t s~t = x~t ¡ K ^ 0t (N0t x~t + º t ) x~+ ¹tjt = xt ¡ x¹tjt¡1 ¡ K (5.31) t = xt ¡ x which we can use to get an estimate of the variance of the update error Ptjt = E h¡ xt ¡ x¹tjt ¢¡ ´ ¢0 i ³ ^ 0t N0t Ptjt¡1 : xt ¡ x¹tjt jI¹t = I ¡ K (5.32) We now have two expressions for how to use the contemporaneous signal to update the estimate of the state variables: our initial guess from equation (5.28) and the implied Kalman …lter from equation (5.30). For our initial guess to be correct we need that ^ 0t (I + M0t K0t ) : K0t = K (5.33) ^ 0t from (5.30) and for N0t and §v;t from equation (5.29) Substituting for the de…nition of K and simplifying we get ¡ ¢¡1 K0t = Ptjt¡1 L00t L0t Ptjt¡1 L00t + §e0 : (5.34) Finally note that (5.33) together with the de…nition of N0t from (5.29) also implies that ^ 0t N0t : K0t L0t = K 35 (5.35) End of the period. We now turn to the second stage when the policymaker receives signal s1t . From equation (5.27) and the law of iterated expectations, the policymaker’s forecast of this period’s lagged signal based on the information available at the beginning of the period is s¹1;tjt = (M1t + L1t ) x¹tjt . We subtract this expression for the prior expectation of the lagged signal from its actual value (5.27), and obtain the forecast error for the signal ¡ ¢ s~1t = L1t xt ¡ x¹tjt + e1;t = L1t x~+ t + e1t : (5.36) From equation (5.36) we get the update of the state variable x^+ ~+ t as a linear projection of x t onto s~1t £ ¤ £ ¤¡1 ^ 1t s~1t with x^+ = E x~+ ~01t jI¹t E s~1t s~01t jI¹t s~1t = K t t s ¡ ¢ ^ 1t ´ Ptjt L01t L1t Ptjt L01t + §e1 ¡1 and K £ ¤ Ptjt ´ E x~t x~0t jI¹t . (5.37) The updated conditional expectation of this period’s and next period’s state variables then are © ª ^ ¹tjt ¹+ x¹+ tjt + K1t s1t ¡ (M1t + L1t ) x tjt = x x¹t+1jt = Jt x¹tjt + Ht x¹+ tjt : From this expression and the law of motion for state variables (5.26) we obtain the forecast error for the state variables at the beginning of next period as a function of this period’s beginning-of-period forecast error x~t+1 = xt+1 ¡ x¹t+1jt ³ ´ = Ht xt ¡ x¹+ tjt + ux;t+1 ³ ´ ^ 1t L1t x~t + ux;t+1 ¡ Ht K ^ 1t e1t ; = Ht I ¡ K (5.38) and we can use this expression to update the period’s beginning-of-period h¡ estimate of ¢next i ¡ ¢0 forecast error covariance matrix Pt+1jt = E xt+1 ¡ x¹t+1jt xt+1 ¡ x¹t+1jt jI¹t+ ³ ´ ³ ´0 0 ^ 1t L1t Ptjt I ¡ K ^ 1t L1t Ht0 + Ht K ^ 1t §e1 K ^ 1t Pt+1jt = Ht I ¡ K Ht0 + §u : It remains to show how the equilibrium function (5.23) can be de…ned. 36 (5.39) The equilibrium relation G Suppose the equilibrium rule is as de…ned in equation (5.23). Then the policymaker’s and the private sector’s forecast of next period’s ‡ow variables, and the implied di¤erence in their forecasts is y¹t+1jt = (G1;t+1 + G2;t+1 ) x¹t+1jt £ ¤ yt+1jt = G1;t+1 xt+1jt + G2;t+1 E x¹t+1jt+1 jIt ¡ ¢ ¡ £ ¤ ¢ yt+1jt ¡ y¹t+1jt = G1;t+1 xt+1jt ¡ x¹t+1jt + G2;t+1 E x¹t+1jt+1 jIt ¡ x¹t+1jt : (5.40) We can write the policymaker’s conditional expectation at the beginning of next period as this period’s beginning-of-period conditional expectations plus the sum of this period’s three updates ³ ´ ³ ´ ¡ ¢ + + x¹t+1jt+1 = x¹tjt + x¹tjt ¡ x¹tjt + x¹t+1jt ¡ x¹tjt + x¹t+1jt+1 ¡ x¹t+1jt : The public’s expectations of these three updates are in turn h i ^ 1;t L1;t x~+ E x¹+ ¡ x ¹ jI = K tjt t t tjt i h ^ 1;t L1;t x~+ = x¹t+1jt ¡ x¹tjt + (Ht ¡ I) K E x¹t+1jt ¡ x¹+ t tjt jIt ³ ´ £ ¤ ^ 0;t+1 N0;t+1 Ht I ¡ K ^ 1;t L1;t x~+ E x¹t+1jt+1 ¡ x¹t+1jt jIt = K t : Substituting these expressions yields h ³ ´i £ ¤ ^ ^ ^ E x¹t+1jt+1 jIt ¡ x¹t+1jt = ªt x~+ with ª = H K L + K N H I ¡ K L : t t 1;t 1;t 0;t+1 0;t+1 t 1;t 1;t t (5.41) From the transition equation for the state variables (5.13), we get that the di¤erence between the private sector’s forecast and the policymaker’s forecast of the state variables is ¡ ¢ ¡ ¢ xt+1jt ¡ x¹t+1jt = Axx xt ¡ x¹tjt + Axy yt ¡ y¹tjt : (5.42) Using (5.42) and (5.41) in (5.40) yields ¡ ¢ ¡ ¢ yt+1jt ¡ y¹t+1jt = (G1;t+1 + G2;t+1 ªt ) xt+1jt ¡ x¹t+1jt + G1;t+1 Axy yt ¡ y¹tjt On the other hand, the optimality constraints on allocations (5.13), imply that the di¤erence in forecasts is given by ¡ ¢ ¡ ¢ yt+1jt ¡ y¹t+1jt = Ayx;t xt ¡ x¹tjt + Ayy;t yt ¡ y¹tjt : Equating the last two expressions and solving for the current value of the ‡ow variables yields ¡ ¢ yt = y¹tjt + (Ayy;t ¡ G1;t+1 Axy )¡1 (G1;t+1 Axx + G2;t+1 ªt ¡ Ayx;t ) xt ¡ x¹tjt 37 or (5.43) yt = G1;t xt + G2;t x¹tjt where G1;t ´ (Ayy;t ¡ G1;t+1 Axy )¡1 (G1;t+1 Axx + G2;t+1 ªt ¡ Ayx;t ) G2;t ´ Gt ¡ G1;t . Note that the matrix G1;t is also implicitly de…ned as Ayy;t G1;t + Ayx;t = G1;t+1 Ht + G2;t+1 ªt+1 : (5.44) The interaction of optimal control and inference We have solved the optimal control problem under the assumption that certainty equivalence holds, that is the policymaker’s decision is a function of the policymaker’s expectation of the state of the economy, conditional on the available information at the time the decision is made. For this to be true it remains to be shown that the term lt in the policymaker’s value function (5.24) is indeed independent of the policy choice. From equation (5.25) it follows that lt is a function of the update error xt ¡ x¹tjt ; xt ¡ x¹tjt = x~t ¡ x^t ¡ ¢ yt ¡ y¹tjt = G1;t xt ¡ x¹tjt ¡ ¢ ¡ ¢ xt+1 ¡ x¹t+1jt = Axx xt ¡ x¹tjt + Axy yt ¡ y¹tjt + ut+1 ; and it is enough to show that the update error for the state variable xt ¡ x¹tjt is independent of policy. The update error has two components, the forecast error from the previous period x~t , and the update of the state variable estimate x^t . As of time t the forecast error x~t is independent of the current policy choice. This follows from equation (5.38). Notice, that the only way the policy rule a¤ects the signal extraction problem is through the determination of Gt in equation (5.19). Thus if we can show that the update of the state variable estimate does not depend on Gt , then the update error is independent of current policy. From equation (5.28) the time t update of the state variable depends on the matrices K0t and L0t , and the forecast error x~t . From equation (5.34), K0t depends on Ptjt¡1 and L0t , and L0t in turn depends on G1;t . From equation (5.39), Ptjt¡1 depends on past outcomes such as Ht¡1 , but not current choices, so it is independent of current policy. Thus we have to show that G1t is independent of current policy. From the de…nition of G1t in equations ^ 1;t L1t in (5.37), L1t in (5.27), (5.41) and (5.43), together with the de…nition of Ht in (5.26), K and Ptjt in (5.32) it follows that G1t is a function of G1;t+1 and Gt+1 , but not of Gt . Therefore G1t is independent of current policy Ft . 38 We have just shown that certainty equivalence holds, that is conditional on past and future decision rules, the inference problem of the current policymaker is not a¤ected by the policymaker’s own choices. This does not mean that the signal extraction problem is independent of the optimal control problem. Since the Markov-perfect equilibrium we study is a …xed point of the mapping de…ned by the inference problem, and since G is a function of the policy rule F , the inference problem cannot be separated from the control problem. 39 ¯ = 0:99, Table 1. Parameters and Steady State Parameters µ = 0:73, ½ = ¡17:52,  = 1:30, Á = 3:0, " = 10, ° y = 0:9, y = 0:265, c = 0:265, q0 = 1:042, ° m = 0:9, ¾ y = 0:01, Steady State n = 0:265, q1 = 1:016, y0 = 0:175, y1 = 0:225, R = 1:036, q2 = 0:991, ¾ m = 0:026 P^ = 1:025, q3 = 0:966; y2 = 0:289, y3 = 0:370 40 w = 0:897; J = 4, Figure 1. Estimated Taylor-Rule for the U.S. Economy 41 LL ψ PP p0 Figure 2. Temporary Equilibrium 42 Panel A. The Response to a Productivity Shock Output, Employment, and Real Balances Nominal Interest Rate, Inflation Rate, and Real Rate 1.5 1.5 0.0 0.0 Pt+1/Pt 1.0 ct 0.5 rt 1.0 mt nt 2 3 4 5 -1.0 -1.0 0.0 0.0 1 0 6 Relative Prices 1 2 3 4 5 p3t 0.05 0.10 0.020 0.05 0.015 0.00 0.010 -0.05 0.005 -0.10 0.000 0.020 0.00 p1t p0t 0 1 2 3 4 5 6 0.015 at p2t -0.10 6 Allocational Efficiency and Marginal Cost 0.10 -0.05 -0.5 0.5 zyt 0 Rt -0.5 0.010 ψt 0.005 0.000 0 1 2 3 4 5 6 Panel B. The Response to a Money Demand Shock Output, Employment, and Real Balances Nominal Interest Rate, Inflation Rate, and Real Rate 1.5 1.5 1.0 0.00 -0.5 1 2 3 4 5 -0.05 6 -0.05 0 Relative Prices 1 2 3 4 0.05 0.000 6 0.00 -0.005 -0.05 -0.010 0.000 ψt p0t 0.00 5 Allocational Efficiency and Marginal Cost 0.05 p1t 0.00 rt 0.0 -0.5 0 0.05 Rt 0.5 nt ct Pt+1/Pt 0.05 mt 0.0 0.10 1.0 zmt 0.5 0.10 at p2t -0.005 p3t -0.05 0 1 2 3 4 5 6 -0.010 0 1 2 3 Figure 3. The Full-Information Case 43 4 5 6 Panel A. Conditional Expectations of State Variables Lagged Relative Price 1.0 0.0 1.0 1.0 0.0 0.0 Lagged Relative Price 1.0 0.0 _ -1.0 -1.0 -1.0 -2.0 -2.0 -3.0 -3.0 q2,t|t-1 -1.0 _ q1,t|t-1 -2.0 q1,t -3.0 0 1 2 3 4 5 Productivity Shock 1.0 1.0 zy,t -3.0 0 6 -2.0 q2,t 1 2 3 4 5 6 Money Demand Shock 0.1 0.1 zm,t 0.0 0.5 0.0 0.5 _ zm,t|t-1 -0.1 _ -0.1 zy,t|t-1 0.0 0.0 0 1 2 3 4 5 -0.2 6 -0.2 1 2 3 4 5 6 7 Panel B. Realizations of Flow Variables Output, Employment, and Real Balances 2.0 1.0 mt 0.0 ct nt -1.0 -2.0 0 1 2 3 4 5 Nominal Interest Rate, Inflation Rate, and Real Rate 2.0 4.0 1.0 2.0 0.0 0.0 -1.0 -2.0 -2.0 -4.0 6 p3t 0.0 p1t -1.0 p0t -2.0 -3.0 0 1 2 3 0.0 Rt -2.0 Pt+1/Pt -4.0 1 2 3 4 5 6 Allocational Efficiency and Marginal Cost 2.0 p2t 2.0 rt 0 Relative Prices 1.0 4.0 4 5 2.0 1.0 1.0 0.0 0.0 -1.0 -1.0 -2.0 -2.0 -2.0 -3.0 -3.0 -3.0 -4.0 6 1.0 at 0.0 -1.0 ψt -4.0 0 1 2 3 4 5 6 Figure 4. The Response to a Productivity Shock with Information on Lagged Output 44 Panel A. Conditional Expectations of State Variables Lagged Relative Price Lagged Relative Price 0.5 0.0 -0.5 _ q1,t|t-1 -1.0 _ q1,t|t -1.5 q1,t -2.0 0 1 2 3 4 5 0.5 0.0 0.0 -0.5 -0.5 -1.0 -1.0 -1.5 -1.5 -2.0 -2.0 1.0 zy,t 0.5 0.0 -2.0 1 0.0 5 5 6 0.1 0.0 zm,t|t-1 -0.1 _ zm,t|t zy,t|t-1 4 4 _ -0.1 3 3 zm,t _ 2 2 Money Demand Shock 0.0 1 -1.0 -1.5 0.5 0 -0.5 q2,t 0.1 zy,t|t 0.0 q2,t|t-1 q2,t|t _ 0.5 _ _ 0 6 Productivity Shock 1.0 0.5 -0.2 6 -0.2 0 1 2 3 4 5 6 Panel B. Realizations of Flow Variables Nominal Interest Rate, Inflation Rate, and Real Rate Output, Employment, and Real Balances 1.5 1.5 mt 1.0 1.0 1.0 0.5 0.0 0.0 0.5 ct Rt -1.0 0.0 0.0 -0.5 -1.0 2 3 4 5 1.0 -3.0 0 1 2 3 4 5 6 Allocational Efficiency and Marginal Cost 2.0 1.0 1.0 at 1.0 p3t 0.0 p2t 0.0 -3.0 6 Relative Prices 2.0 -2.0 Pt+1 /Pt -1.0 1 -1.0 -2.0 -0.5 nt 0 1.0 rt 0.0 0.0 ψt p1t -1.0 -1.0 -1.0 -1.0 p0t -2.0 -2.0 0 1 2 3 4 5 -2.0 6 -2.0 0 1 2 3 4 5 6 Figure 5. The Response to a Productivity Shock with Information on Contemporaneous Money and Lagged Output 45 Panel A. Conditional Expectations of State Variables Lagged Relative Price Lagged Relative Price 0.5 0.5 0.5 _ q1,t|t 0.0 0.0 0.0 -0.5 -0.5 -0.5 -1.0 -1.0 _ q1,t|t-1 -1.0 -1.5 q1,t -2.0 0 1 2 3 4 5 -1.5 -2.0 -2.0 0.5 1.0 q2,t|t 0.0 -0.5 q2,t|t-1 -1.0 q2,t -1.5 -2.0 1 2 3 4 5 6 Money Demand Shock 1.0 zm,t _ _ zm,t|t zy,t|t-1 zy,t 0.0 _ 0 6 Productivity Shock 0.5 -1.5 0.5 _ 0.0 0.5 0.5 _ zm,t|t-1 _ zy,t|t -0.5 -0.5 0 1 2 3 4 5 0.0 6 0.0 0 1 2 3 4 5 6 Panel B. Realizations of Flow Variables Nominal Interest Rate, Inflation Rate, and Real Rate Output, Employment, and Real Balances 2.0 2.0 1.0 1.0 3.0 3.0 2.0 mt 0.0 0.0 yt 1.0 0.0 -2.0 -2.0 1 2 3 4 5 -3.0 0 Relative Prices 1.0 4 5 6 1.0 at 0.0 -1.0 ψt 0.0 p1t -1.0 -1.0 p0t -2.0 -2.0 2 3 -1.0 p2t 1 2 0.0 1.0 p3t 0 1 Allocational Efficiency and Marginal Cost 2.0 1.0 -2.0 Pt+1/Pt -3.0 6 2.0 0.0 -1.0 -1.0 -2.0 0 0.0 Rt -1.0 nt -1.0 2.0 rt 1.0 3 4 5 -2.0 -2.0 -3.0 -3.0 -4.0 6 -4.0 0 1 2 3 4 5 6 Figure 6. The Response to a Money Demand Shock with Information on Contemporaneous Money and Lagged Output 46