The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
o r k m g r a p e r s e r ie s Solving Nonlinear Rational Expectations Models by Parameterized Expectations: Convergence to Stationary Solutions Albert Marcet and David A. Marshall Working Papers Series Macroeconomic Issues Research Department Federal Reserve Bank of Chicago December 1994 (WP-94-20) FEDERAL RESERVE BANK OF CHICAGO Solving Nonlinear Rational Expectations M o d e l s b y Parameterized Expectations: C o n v e r g e n c e to Stationary Solutions Albert Marcet Universitat Pompeu Fabra, Barcelona David A. Marshall Federal Reserve Bank of Chicago and Northwestern University March 1994 W e thank Larry Christiano, Wouter den Haan, Joan Ketterer, To m Sargent, Ken Singleton, and three anonymous referees for helpful suggestions. The first part of this paper was presented at Stanford University in January 1989; thanks are due to its attendants, including Ken Judd, Rody Manuelli, John Bryant and Richard Rogerson. Marcet’s research was partially supported by the NSF and the DGICYT (Spanish Ministry of Education). All errors are our own. Abstract This paper develops the Parameterized Expectations Approach (PEA) for solving nonlinear dynamic stochastic models with rational expectations. The method can be applied to a variety of models, including models with strong nonlinearities, sub-optimal equilibria, and many continuous state variables. In this approach, the conditional expectations in the equilibrium conditions are approximated by flexible functional forms of finite elements. The approach is highly efficient computationally because it incorporates endogenous oversampling and Monte-Carlo integration, and it does not impose a discrete grid on the state variables or the stochastic shocks. W e prove that PEA can approximate the correct solution with arbitrary accuracy on the ergodic set by increasing the size of the Monte-Carlo simulations and the dimensionality of the approximating family of functions. 1 Introduction During the last decade, the use of dynamic stochastic models has extended to all fields of economics. These models are difficult to analyze because they typically have no analytic solution. However, recent increases in the power of computer hardware now allow these models to be studied by numerical simulation techniques. Simulations can be used to study the model from both a theoretical and empirical perspective. Theoretically, one can see if the model reproduces some stylized facts and how it responds to a change in the environment or in policy. Empirically, one can perform calibration exer cises or do more formal testing using estimation by simulation or maximum likelihood procedures. The numerical algorithms most widely used for solving dynamic mod els with rational expectations in economics are value-function iteration and linear-quadratic approximation. They are based on dynamic programming, so they are not well-suited to models in which the equilibrium does not cor respond to the solution of a planner’s problem. Furthermore, value-function iteration is affected by the ’curse of dimensionality’, so is impractical for models with several continuous endogenous state variables. Linear-quadratic approximation can handle large models, but does not provide an arbitrarily good approximation in nonlinear models; this is especially a problem, for example, when the model has inequality constraints. This paper introduces the parameterized expectations approach (PEA) for calculating numerical solutions to stochastic nonlinear models with ratio nal expectations. In this approach, the equilibrium conditions of the model are written as a system of stochastic difference equations, and the conditional expectations in these equations are parameterized with flexible functional forms of finite dimensions, such as polynomials or splines. Simulations are then generated using these functions in place of the conditional expectations in the equilibrium conditions. A PEA solution corresponds to a parameter ization of the conditional expectations that is consistent with the series it generates. In this paper, we formally characterize the PEA, and we derive a set of conditions under which arbitrarily good accuracy can be obtained as the solution is refined (for example, as the degree of the polynomial goes to infinity); it contains some of our earlier work in Marcet [1988], Marshall [1988] and Marcet and Marshall [1992]. 3 The PEA algorithm has been successful in delivering solutions to a num ber of large and complicated models. Applications in the literature include monetary asset pricing models, such as Marshall [1988, 1992], and den Haan [1990, 1991]; models of exchange rates, such as Bansal [1990], Bansal, Gal lant, Hussey and Tauchen [1994] and McCurdy and Ricketts [1992]; asset pricing models with heterogeneous agents, such as Marcet and Singleton [1989], Ketterer and Marcet [1989], and den Haan [1993]; models of tax policy, such as Rojas [1991], and Otker [1992]; and nonstandard stochastic growth models, such as Marcet and Marimon’s [1992] model with participa tion constraints and Christiano and Fisher’s [1994] model with investment constraints. A related method in which the laws of motion of endogenous variables are parameterized has been used to solve asset-pricing models in papers by Heaton [1993] and Bekaert [1993]. This non-exhaustive list of PEA applications includes models with many continuous state variables, multiple heterogeneous agents, strong non-linearities, inequality constraints, incentive constraints, participation constraints, non-stationarities, discrete choice spaces, and suboptimal equilibria. In short, the PEA has been roadtested: its practical applicability to a wide range of economic models has been demonstrated . 1 We show that the PEA delivers an arbitrarily close approximation to the true equilibrium as the solution is refined. The set of assumptions we impose is very general: essentially, our proof applies to most models with continu ous laws of motion (including non-differentiable cases). To our knowledge, no such proof is yet available for other methods that compute approximate equilibria by solving systems of Euler equations. Our solution procedure is easily adapted to cases where one has to solve for the transition path to the stationary distribution starting from an arbitrary initial condition, but in the current paper we concentrate our discussion on solving for the stationary and ergodic distribution. In section 2, we formally introduce the parameterized expectations ap proach and discuss how it can be applied to three well-known examples. In section 3, we relate PEA to two strands of economic literature. We first dis cuss PEA’s relationship to models of least-squares learning; we then compare 'This is to be contrasted with proposals to use numerical methods taken from other sci ences, without demonstrating their applicability to the problems of concern to economists. For example, Judd (1993) lists a number of numerical techniques used in engineering and physics that have not been tried in full scale economic applications. 4 our method to other solution algorithms used in economic applications. In section 4 , we present the convergence result. Section 5 provides a brief dis cussion of certain practical considerations in applying the method. Section 6 contains some conclusions and suggestions for extensions. 2 T h e Parameterized Expectations A p p r o a c h ( P E A ) 2.1 A General F r a m e w o r k We assume that the economy is described by a vector of variables z t € Z C R n , and a vector of exogenously given shocks u t 6 U C R s . The equilibrium process {z(,u t} is known to satisfy a stochastic difference equation <7(£<[<£(z<+i )], z t , z t - i , u t ) = 0 (1 ) for all t, where g : R m x R n x R n x R 3 —» R q and <j> : R n —> R m . Once the parameters of the model have been fixed, g and <f> are functions known to the economist. The vector z t includes all endogenous variables, as well as those exogenous variables that appear inside the expectation. The process u t is assumed to be Markov of order one. As usual, E t is the conditional expectation given information up to period t . The system ( 1 ) will typically include Euler equations, resource constraints, laws of motion of exogenous processes, market clearing conditions, incentive constraints, and so forth. Inequality constraints of the form h { E t [<f>(zt + i ) ] , z t , Zt_i, ut) > 0 are incorporated into ( 1 ) by including equations (A(£t[<£(zt+i)],zt,zt_i,u())" = 0 as part of the system g , where (a:)- is the function ‘negative part of x \ These inequality constraints may include Kuhn-Tucker conditions, second order op timality conditions, or participation constraints in incentive problems. Often, the sufficient conditions for an equilibrium include, in addition to equation ( 1 ), a set of transversality conditions. For most models, a process 5 { z t } that satisfies ( 1 ) also satisfies the transversality conditions if and only if the process is stationary. Stationarity can be imposed as an additional side condition of the model. We assume that the past information that is relevant for predicting <j>(zt+1) can be summarized in a finite-dimensional vector of state variables x t € X C R } satisfying x t = f ( z t- u u t) , (2 ) where / is a known function. 2 This implies Et[4>(z t + 1 )] = E [< f > {z t+ i)\x t\. (3) Furthermore, we assume that the model is recursive, so that the conditional expectation is given by a time-invariant function £ such that (4) £ ( x t ) = E[<f>(zt + i ) \ x t}. By definition, the function £ = £ satisfies arS {h ™1k»} E I* ^ 2 ‘+i) II2 ~ all (5) since the conditional expectation is the best predictor in the mean square sense. Throughout the paper we refer to a s o l u t i o n as a stochastic process {zt(ut} satisfying equation (1), (2), and (3), given the exogenous Markov process for {u<} . Since we will are dealing with systems g that are invertible in their second argument z t , finding a solution is equivalent to finding a function £ such that, if { z t , u t } satisfies g ( £ { x t ), z ti z t - u i i t ) = 0, (6) then £ satisfies (4) for all t . Alternatively, finding a solution is equivalent to finding a law of motion H such that the z t generated by zt - H ( z t- U u t ) (7) 2We are assuming here that the researcher knows how to select a set of sufficient state variables i<. Sometimes it is difficult to determine the sufficient state vector z t beforehand; this is often the case, for example, in models with private information. In principle, it is possible to apply PEA and find a sufficient set of state variables numerically; we would then have to incorporate into the algorithm a search over functions of past variables that summarize past information. We do not pursue this avenue any further in this paper. 6 satisfies (1), (2) and (3). We now show how three well-known models fit into the above framework. These examples are chosen for their simplicity; applications of PEA to more complicated models can be found in the papers cited in the Introduction. • E xam ple 1.1 (Lucas [1978] asset pricing m odel) A representative consumer chooses a stochastic consumption process {ct} of a perishable good in order to maximize E q Y ^ o ° t )• The single asset in the economy is traded in a competitive market at price p t , and pays an exogenous dividend flow {d(}. The consumer also receives exogenous labor income {u;t} . The Euler equation for the maximization problem of the consumer is u ' { c t ) p t = 8 E t [ u ' ( c t+ 1 ) (pi+1 + dt+i)]. (8 ) To map this model into the above framework we let z t = (p<, c t, d t), and U( = (d t w t); the system of equations corresponding to ( 1 ) is given by (8 ) and the market clearing condition c t = d t + w t . The function 4>{z) is given by u ' ( c ) ( p + d ) . A time-invariant solution for the asset price can be found for which x t = (d t , w t) is a sufficient set of state variables. • E xam ple 1.2 (Sim ple S tochastic G row th M odel) Consider the simple growth model where an agent maximizes E o subject to it ^ k t = k t- i p — 8 lu (c t) /g\ + it , given, ' ' where c< denotes consumption, k t is the capital stock, it is investment, and 6 t is an exogenous stochastic productivity shock, Markov of order one. The first order condition for optimality is u'(ct) = 8 E t[ ^(ct+i) ( k t ~ 1a O t + 1 + p )] (10) To map this model into the above framework, set z t = (ct, k t , k t- i , i t , 0t), and x t = ( k t- i , 0 t ) . The function g is given by the resource con straints (9) and the Euler equation (10). The function <f>{z) is given by ut = Of, 7 ti'(c) (fco - 1 a0 + f i ) . Standard results from dynamic programming guar antee that the solution is characterized by a time-invariant conditional expectation function. % • E xam ple 1.3 (Sim ple G row th M odel w ith Lower B ounds on In v estm en t) This example shows how inequality constraints can easily be handled by PEA. Suppose we add a non-negativity constraint to the simple growth model of example 1 .2 : t<>0 (11) With this restriction, the first order condition ( 1 0 ) is replaced by the Kuhn-Tucker conditions u'(ct) - A, - 6 E t[ u'(ct+1) (fc“ ‘ q Oi+i + /x) - fiX t+ i] = 0 (12) tjAf = 0 (13) At > 0,t't > 0, (14) where At denotes the Lagrange multiplier associated with constraint ( 1 1 ). Mapping this model into framework ( 1 ) is accomplished by setting zt = u t = 0 t . The system g is given by (9), the Kuhn-Tucker conditions (12) and (13), and the negative parts of the inequalities in (14). Note that, in this case, <f>(z) = u (c) (A:“- 1 a0 + fi) —fiX . Again, the model can be shown to be recursive with standard dynamic programming techniques.3 2.2 Definition a n d Calculation of a n A p p r o x i m a t e P E A Solution System ( 1 ) and explicit formulas for g and <j> are easy to find in many mod els. Unfortunately, finding the solution analytically in any generality is not possible. The difficulty is that S cannot be determined unless the process 3The application of PEA with inequality constraints was developed by Marcet and Sin gleton [1991]. For a more detailed discussion of how to impose the Kuhn-Tucker conditions the reader is referred to that paper. 8 for z t is known, but that process can not be backed out from (6 ) unless the conditional expectation £ is known. The PEA algorithm addresses this dif ficulty by replacing £ with an approximating function ip which is chosen to resemble £ in a manner to be made precise. Formally, let P denote a class of functions that is dense in : R l —►f?m|. It is assumed that each element of P is characterized by a parameter vector /? with a finite number of nonzero elements. For example, P can be taken as the set of polynomials, splines, neural networks or finite elements. Then, (3 would be, respectively, the coefficients of the polynomial, the parameters charac terizing each partition for the spline, the elements of the neural net, or the parameters characterizing the finite elements. The element of P characterized by parameter vector /? is denoted either by •) or by the simpler notation rpp. The s e t o f a d m i s s i b l e p a r a m e t e r s with at most v non-zero elements is denoted A * That is, A c{/? (E R ° ° : i th element of /? is zero if i > j/|. In PEA, the function £ is approximated by choosing an element of P that satisfies a property close to (6 ) and (5). More precisely, for a given positive integer v , /? € A> and {u<}~0, define the process (zt(/3 ), u t } ~ 0 as the solution for all t of the system 9 (W , *<(/?)), z t ( 0 ) , x t{P ) z t - 1 (/?), ttt) = 0 (15) /(z»-i (/?),«<) = Now, let £ „ (£ ) = argmjn E || <j>(zt + l (/3 )) - x t ( 0 ) ) ||2 (16) We wish to choose € A as close as possible to G v (f3v ) , so the approximate PEA solution is characterized by a fixed point, denoted /?„> as follows: 0, = <A(&). (17) For the sake of simplicity we will assume in the remainder of this section that G u is a well-defined mapping, and that a fixed point of G „ exists; these issues will be addressed formally in section 4. With the appropriate substitutions, one can see that xppv plays the role of £ in (6 ), so our approximate solution is the process {zt(^),u< }^0. 9 From this point on, we address the problem of approximating the solution at the stationary and ergodic set. This is an interesting case, since station a r y and ergodicity are the basis of time series econometrics or calibration exercises. PEA has been applied to solutions outside of the stationary distri bution by replacing the long simulation in Step 2 below with many repeated short-run simulations.4 Therefore, the claim in Judd [1992] (page 447) that PEA can not solve for equilibria outside the ergodic set and it can not be used for policy comparisons is incorrect. We now give an operational algorithm for computing the approximate PEA solution for a fixed v: • Step 1 : Write the system g in ( 1 ) so that it is invertible with respect to its second argument. Find a set of state variables x that satisfies (3). If a unique solution exists, g and x should be selected so that (1) and (3) are satisfied only by the unique solution. Replace the true conditional expectation by the finitely-parameterized function •) for (3 € D v to obtain (15). • Step 2: Fix Z q . For a given yd € D„ and for T large enough, draw a sample of size T of the exogenous stochastic shock and recursively calculate{z*(/?), u<}£_0 using (15). • Step 3: Find the sample version defined by: G>,t (£)= argmin of G v. More precisely, G v< t is Ef=o II ^(*<+i(£)) - £ ^(£»*t(0)) l|2 (18) This minimization is easy to perform by computing a non-linear least squares regression with the sample {.?*(/?), ut} ^ 0, taking as the dependent variable and V>(*,xt(/?)) as the explanatory function. • Step 4: Find the fixed point P v,T = G „ , t { ^ v ,t ) (19) We will discuss the numerical calculation of 0„tT in the section 5. 4See Marshall [1988] and [1993], Marcet and Marimon [1992], Rojas [1993] and Christiano and Fisher [1994] for applications. Also, see Marcet and Marshall [1992] for a proof of convergence when short-run simulations are used. 10 We end this subsection by discussing how this algorithm would be applied to the three examples presented in section 2 .1 . • E xam ple 1.1 (Lucas [1978] asset pricing m odel) The only non-trivial endogenous variable is the stock price p t . In this and the remaining examples, <f>takes only positive values, so it is ap propriate to choose P as a set of functions that can take only positive values. Step 2 is easily accomplished by solving for the price in each period from u'(c«) P t(P ) = ip (P ,d u W t) which is the (15) version of the Euler equation. The fixed point can be found iteratively5. Notice that our choice of state variables guarantees that we are approximating the unique stationary (non-bubble) solution. Bubble solutions could be allowed by adding today’s price to the list of state variables. • E xam ple 1.2 (Sim ple S tochastic G row th M odel) Steps 1 and 2 follow similar considerations to the previous example. Since the solution is a time-invariant function of ( k t- i , 0 t) and it satisfies the Euler equation, we can be certain that the solution approximates the unique solution of the model. Notice that the (suboptimal and meaningless) solutions to the Euler equation that violate the transversality condition will never be approximated; in those solutions Co is fixed, and k t is a time-invariant function of (&t_ ! ,0£) and c£, so that the transversality condition is imposed by our choice of state variables and the fact that initial consumption is not fixed. • E xam ple 1.3 (Sim ple G row th M odel w ith Lower B ounds on In v estm en t) We show how to use the Kuhn-Tucker conditions to find (ct(/?), *<(/?), k t ( 0 ) } while imposing inequalities (14). Let us write the parameterized version of (12) and (14): u/(ct(/?)) - A,(/?) = 6 *(/?; Jfc*_,0®), 0t) (20) 5For this model, the mapping G y as well as the fixed point p y can be found analytically. See Marcet [1988] 11 At(/3) > 0 Notice that, for a fixed value of increasing in t<(/5). and *<(/?) > 0 (21) ; kt-i(/3 ),6 t), the left side of (21) is We can then proceed as follows: for each t: (a) compute (ct(/?), *<(/?)) from (20) under the conjecture that At(fi) = 0. If the corresponding investment turns out to be negative, then (b) set i t ( P ) = 0, find c t ((3) from the feasibility constraint, and then compute At(/?) from (20). It is clear that if step (a) delivers a negative investment, consumption will be lower when we go to step (b), so uf(ct(/3)) will be higher (relative to (a)). This insures that (20) delivers a positive A, so (21) is satisfied. This strategy requires that the Kuhn-Tucker condition be written in a way that the function g is envertible. For example, the Kuhn-Tucker condition could be expressed as, u'(ct)ct- Atct = 6 E t [c< u'(ct+i) (k?~l a6t+i + /*) - ^ctAt+1] (22) However, if steps (a) and (b) were applied to (22), there may be states in which the function u'(c)c is decreasing in i. In that case, step (b) would deliver a negative value for the multiplier A. Therefore, (22) would be an inappropriate choice for g if PEA is used as a solution algorithm. In terms of the step-by-step description of the algorithm, (22) violates step 1 , since it implies a representation for g that may not be invertible with respect to its second element. 3 Relation of P E A 3.1 A p p r o x i m a t e P E A Learning to t h e Literature Solution a n d Least Squares In section 4, below, we show that the PEA solution can approximate the rational expectations equilibrium arbitrarily well by letting u and T go to 12 infinity. In this subsection we give an alternative interpretation to the PEA solution for fixed v as the equilibrium of an economy with boundedly rational agents. Suppose it were infeasible for the agents to compute the true conditional equilibrium function £. For example, agents may not know the correct func tional form for 6, or they may face computation constraints. Instead, agents are constrained to use some function rfip to forecast of <f>(zt+i). In principle, they can choose any parameter P € D„, but they are restricted to a fixed degree v. The process {zt(P)} can be interpreted as the vector of endogenous variables that would be generated by such an economy. If these agents were to choose an arbitrary parameter vector P € D the economy would not be in equilibrium: the best forecaster of <f>(zt+1) within the set D u would in fact be given by i p ( G u( P ) , x t ( P ) ) , rather than by i p ( / 3 , x t ( ( 3 )), the function used by the agents ex ante. Presumably, agents would eventually recognize their systematic errors and would update the parameters in their forecast function. The equilibrium of this boundedly rational economy corresponds to /?„, the fixed point of G v defined in equation (17): if the boundedly rational agents use (3V they will eventually realize that this is their best alternative, given that they are restricted to staying in Dv. For this model of boundedly rational agents to be meaningful economi cally, it should be locally stable: if agents start at t = 0 using some beliefs Po near the fixed point /?„, and update ft as new information received at each period (so their best forecast at time t is given by il)(P t,xt-1 )), their forecast coefficients should converge to /?„. One way to model how agents update their forecasts is to assume a least squares learning model (LSL), in which p t is generated by a non-linear least squares recursive algorithm (see Appendix 2 ). Adapting some results in Ljung [1975], under certain regularity conditions, it is possible to show the following results: • For P G £)„, if P 7^ P „ , P r o b ( P t —►/?) = 0. That is, LSL will almost surely not converge to a parameter vector that is not a fixed point of • Prob(Pt —►p v) = 1 locally, if and only if the differential equation 0 = G,{P) - P 13 (23) is locally stable at /?„. That is, local stability of equation (23) is equiv alent to local stability of the equilibrium of boundedly rational agents under least squares learning. The second result will be used in section 5.1 to suggest some algorithms for computing the approximate solution. To prove the convergence result in Proposition 1 , we must impose suf ficient conditions to insure that if a law of motion H * is close to the true solution H in (7), then {z*} (the process generated by H*) is close to {z(} (This property is implied by assumptions 4* and 4, in section 4.) Not surpris ingly, this property is one of the regularity conditions that must be imposed for least squares learning to converge. Intuitively, if two H 's are close but generate z ’s very far apart, agents will be chasing a moving target; their observations on z will not tell them how to update their forecast function. A condition like this is likely to be needed for any learning scheme to converge, as long as the learning scheme is only based on observations of the realized process. 3.2 R e l a t i o n to O t h e r A l g o r i t h m s . PEA has developed from other solution approaches used previously in eco nomics. It is closely related to the backsolving procedure of Sims [1985], Novales [1991] and Ingram [1990]. In this approach, conditional expectations are fixed by assuming a process for {<£(zt+i)}. The remaining variables, in cluding exogenous processes, are then solved from a system like ( 1 ) using this assumption. Backsolving can be understood as PEA without Steps 3 and 4, with particular assumptions on the whole process for <j>(zt+1 ). An other predecessor to PEA can be found in Townsend [1983] in a linear model with private information; that paper uses v = 1 and replaces Steps 2 and 3 with an explicit calculation of Gi based on spectral densities, in which the linearity of the model is exploited. The idea of calculating the mapping G i with long-run simulations was previously discussed in Marcet and Sargent [1989]. Other authors have used approximations of Euler equations with flexible functional forms more recently. Examples can be found in Coleman (1988), Judd (1992), Baxter (1991) and McGratten (1993); some of these methods are based on numerical methods used in other sciences such as engineering 14 or physics. We now relate these methods to PEA in terms of the framework and the notation laid out in section 2. Define £(x; P ) as the analogue to £ for the P process. More precisely, £(x; P ) = E[ <j>{zt+l{p))\ x t(P ) = x ], (24) so, if F (• | u) denotes the distribution of ut+i conditional on u t = u, (24) implies .£ (/( * ,« );/? ) = I iF (u '\u ). Ju Assuming that follows: x = z, these solution procedures can be summarized as • Choose a class of functions P and a degree of approximation v . Param eterize the law of motion by an element ^t € P ,‘ this is equivalent to parameterizing expectations as H P , z) = (*■(/?» z > u) (25) where g f 1 represents the inverse of g with respect to its first argument, evaluated at g = 0 . • Fix a grid of p points (x 1, x 2, ..., xp) in the state space Rf and a set of weights to1,to2,...,top. • For a given /?, calculate £ (x J; P ) numerically for each j . As we can see from (24), this involves calculating p integrals over R 3 (recall that s is the number of exogenous shocks in ut). • Find /?„ that solves P* = arg {min} Ejf= i II £(*J;P) ~ H P , *J) II* • (26) The difference ||£(xJ;/?) —i p ( P , H ) \ \ is often called the ’residual’ of the Euler equation at grid-point j , and a generic name for these methods is ’Minimum Weighted Residuals’. The methods differ in the way the approxi mating class P , the grid-points, and the weights are chosen, how the system 15 is written, the way iterations are performed to find the above minimum, and in the method for computing the integrals involved in £ ( x ; f l ) . It is clear that the steps involved in PEA are similar to the steps in MWR methods. In particular, (26) is analogous to (18). There are, however, several fundamental differences: g 1. In MWR, the grid-points and weights are exogenous in the sense that they are independent from the approximating parameter /?. Notice that, in PEA, the sum in (18) is evaluated at the generated series, so that { x t ( P ) } J = \ plays the role of the grid, and the weights are given by the empirical probability of each value of x in the simulated series; hence, the grid and weights are chosen by the algorithm, as a function of fS. This is known as e n d o g e n o u s o v e r s a m p l i n g . 2. The conditional expectation involved in S is never calculated explicitly in PEA. Instead, the approximation in (18) tries to find ^ close to <f>(zt+ 1 ), so that the calculation of the r e s i d u a l (a fundamental step in MWR methods), is entirely bypassed. 3. all integrals are calculated by Monte-Carlo instead of quadrature. The endogenous oversampling feature implies that PEA only pays atten tion to those points that actually happen in the solution. This avoids the problem that researchers face when they must choose grid-points without knowing which points are likely in equilibrium. With endogenous oversam pling, only the economically relevant region of the state space is explored, so no computer time is spent on states that never happen in equilibrium. Fur thermore, the resulting approximation fits more closely at those states that happen more frequently. This is one reason why PEA has been successful in models with a large number of state variables. On the other hands, methods based on exogenous state-space grids are impractical for such models, since the number of points in the grid increases exponentially with the number of state variables. In numerical analysis, the grid (x1, x2, ..., x p) is sometimes adapted in the course of the algorithm as a result of the iterations on /?. This is a form of endogenous oversampling. Unfortunately, these procedures have not yet been used by economists applying MWR. Indeed, some economists have 16 claimed that endogenous oversampling is a bad feature of an algorithm 6 . In fact, endogenous oversampling is an important feature of many algorithms in numerical analysis, and it is likely to be an essential element in any algorithm for solving models with many state variables. By eliminating the integral computations involved in the calculation of the residual, PEA further reduces computation time, since standard quadrature integration is often very costly even in two-dimensional problems. (Notice that, strictly speaking, this makes PEA a non-MWR method.) The use of Monte-Carlo integration means that the integrals can be calculated even in models with a large number of stochastic shocks. PEA parameterizes the conditional expectations directly, while economists using MWR often parameterize the laws of motion it. Formally, the two al ternatives are equivalent: we showed in section 2 how to find a law of motion consistent with V’/3j equation (25) shows how to find V’/s consistent with a proposed law of motion it. However, parameterizing conditional expectations often has some practical advantages: i ) sometimes we know that x contains fewer variables than z , so that /? contains fewer parameters; i i ) when the degree of the polynomial is increased, one can see if some higher-order ele ments need to be introduced beforehand by testing the predictive power of those elements7; H i) if the shocks have a continuous distribution, the condi tional expectation is obtained by integrating over H , so that £ is likely to be smoother than H and easier to approximate with low degree polynomials; i v ) inequality constraints are often easier to impose when the conditional expectation function is parameterized8. The method in Heaton [1993]9 fits into the framework of Section 2 except that the approximate solution satisfies y £ W *+.(fe r» - - H A . J =o Simple algebra shows that this is exactly equivalent to PEA with the pa rameterization (25) if the objective function defining G „ in (9) is modified 6See, for example, Judd [1992]. 7See den Haan and Marcet (1994) for a full description of this idea. 8Christiano and Fisher (1994) make this point in their comparison of solution techniques in a model close to our Example 1.3. 9Also used in Bekaert [1993]. 17 to E \ \ [<f>(zt+ i ( 0 ) ) - r J > (£ ,X t( 0 ))] h ( z t ( 0 ) ) ||2, where h = |d0 1 1 /d*'| . In other words, Heaton’s method is a special case of our algorithm where Step 3 is modified to be a w e i g h t e d non-linear least squares minimization, and where h- is used as a weighting function. Heaton’s method performs endoge nous oversampling, it uses Monte-Carlo integration, and it does not calculate any residual. A small modification of our convergence proof accounts for this case, but is not included in this paper. 10 4 C o n v e r g e n c e to the A p p r o x i m a t e Solution as v — > oo Convergence will be proved in the strong sense that the approximate law of motion H p v , defined in (28) below, converges uniformly to the true law of motion H on the support of the stationary distribution. Proposition 1 establishes the convergence result when v and T are chosen sufficiently large. Proposition 2 provides a partial converse to proposition 1 . A formal proof of convergence is essential for any proposed approximation method. It is not sufficient to show that typ can, in principle, approximate any fixed function arbitrarily well. In PEA, the approximating function xjjp is required to approximate a conditional expectation function £(•; 0 ) which itself depends on 0 through the simulated process {z*(/3), x t { 0 ) } . In other words, £(•;/?) is a moving target. Furthermore, one can easily construct examples of plausible approxima tion schemes which fail to converge to the function being approximated. For example, Judd (1994) points to a standard non-convergence example where, if a polynomial of degree v was fitted exactly at v equally-spaced grid-points, the approximation becomes arbitrarily bad as v becomes larger. He suggests that PEA may suffer from a similar problem of nonconvergence; Proposition 1 formally disposes of Judd’s criticism . 11 10Details on the algebra and the convergence proof for the weighted case are in Marcet and Marshall [1992]. u In any case, this example does not apply to PEA for two reasons: i) in the example, the function is evaluated at v points, while PEA evaluates the function at T points, where T is much larger than v \ ii) the v points in the example cited by Judd are chosen independently of /?, while in PEA this task is endogenous. If anything, this non-convergence example may be indicative of convergence problems for algorithms, including some suggested in Judd 18 To demonstrate convergence of the approximate solution, several con ceptual issues must be addressed: i ) the class P has to be dense in the appropriate space, i i ) Compactness of the function space P is needed in order to:guarantee that a sequence of approximators contains a convergent subsequence, H i) In general, G vj need not have a fixed point, so the defini tion of an approximate solution must be generalized to insure that a solution exists for all v and T, and that a fixed point exists approximately for large v and T . i v ) It must be shown that G „ tT is a good approximation to G „ , for large T . v ) Simulations are generated endogenously in Step 2 , so we must insure that {zt(/?)} is well behaved. In particular, we must show that the series does ro t explode, and that the effect of the initial conditions chosen for z dies out sufficiently rapidly. Point v ) and the choice of P are related problems since one must insure that P is dense in the set of laws of motion that generate well-behaved series. The first set of assumptions are regularity conditions on the functions defining the equilibrium (1). Let <7,_1 be the implicit function that defines the i th argument of g ; for example, (27) g (a ,g 2 l ( a ,z ,u ) ,z ,u ) = 0 The process { z t ((3), x t ( ( 3 ) } defined by equation (15) is a first-order Markov process satisfying z o ( /3 ) = zq and z t ((3) = H ( z t„ x ( p ) , u t\ p ) = For notational convenience we will let that H p is analogous to H } 2 A ssum ption 1 *<(/?)), Zt-i(/?),«t) Hp denote the function H (28) (•, •; /?). Note ( a ) F u n c t i o n s <j>,g, a n d f a r e u n i f o r m l y 13 L i p s c h i t z - c o n t i n u o u s i n a l l a r g u m e n t s , a n d a r e d i f f e r e n t i a b l e a . e .* 123 (1992), where n = v and where the grid is exogenous. To our knowledge, no convergence proof for such methods is yet available. 12We use M H” to denote both the Markov operator defined in (7) and that defined in (28). It will be clear which usage is intended since the operator defined in (28) has /? as an argument, while the true solution does not. 13A function with multiple arguments is said to be uniformly Lipschitz-continuous if the Lipschitz coefficient with respect to the i th argument does not depend on the value of the j th argument. 19 (b) For all (a ,z ,u ) € <f>{Z) x Z x U, g2 l satisfying (27) exists and is uniformly Lipschitz-continuous in its second argument. (c) Tl\e true conditional expectation £ is Lipschitz-continuous a.e. where ’a.e.’ is with respect to the Lebesgue measure. Notice that parts (a) and (b) can be checked directly, while part (c) can not be verified directly from functions <f>, g , and / . It should be kept in mind, however, that in most models parts (a) and (b) imply part (c) . 14 Assumption 1 (b) insures that H can be derived from knowledge of g and £; similarly, it insures that Hp is well defined and that it can be derived from knowledge of g and xpp. The assumption that g fx is Lipschitz-continuous insures that a small change in zt- 1 does not necessitate an arbitrarily large change in zt to maintain equilibrium condition (1). If g is differentiable, this assumption requires the partial derivative of g with respect to its second argument to be uniformly bounded away from zero. Assumption 1 implies that H is Lipschitz-continuous. The next two assumptions insure that the conditional expectation terms in ( 1 ) can be well approximated by some function xp. Approximating classes of functions, such as polynomials, exponentiated polynomials, or splines, ap proximate a given function arbitrarily well only over a compact set. There fore, we impose A ssum ption 2 Z x U is a compact set. Assumption 2 implies that both exogenous and endogenous state variables have compact support. Compact support for endogenous processes can be implied by economic models in a number of ways. In models with capital accumulation, depreciation often implies a bounded capital stock if U is bounded. Asset pricing models are only well defined if a lower bound on asset-holdings is imposed on agents; this is usually assumed through shortsale constraints. Finally, if the model is stationary, bounded support can be achieved by directly imposing exogenous constraints on the model’s variables at levels which will be attained with very small probability. For example, in the simple growth model with zero depreciation the capital stock can be HSee, for example, Santos’s [1991] proof that £ is differentiable in dynamic programming models. 20 arbitrarily large. However, a very high level of the capital stock would be achieved in equilibrium with extremely low probability, so the technology of the model can be modified by imposing a very large upper bound on the level of the capital stock. Presumably, the difference between the behavior of the original model and this modified model will be negligible. The next set of assumptions describes the class of approximating func tions, { i p : D„ x R p —►R m}, where D „ is a compact subset of the space of sequences with at most the first u elements non-zero. 15 We will construct the sequence {A/}£Li such that D „ C A /+ i,V i/, and we define D = (J„>i A / • A ssum ption 3 (a) V x <E X , f o r each v , th e r e s tr ic tio n o f i p ( - , x ) to D „ s a tis f ie s a L i p s c h i tz c o n d itio n u n i f o r m l y in x . ( b) |i p ( f i , -)| < M, V/? € A w here supz€z <p(z) < M < oo. ( c ) ip((3 , •) i s c o n t i n u o u s , d i f f e r e n t i a b l e a l m o s t e v e r y w h e r e , a n d su ch th a t | j < K y V/? G D , x GX : A x R p —►R m s u c h t h a t | {/?„}£ii, € D „ , su ch th at (d ) F o r a n y co n tin u o u s fu n c tio n q K , th e re e x is ts a se q u e n c e 3 K < oo w h ere the d e r iv a tiv e e x ists. <7 | < K' < dip(/3^,x) q(x) = 0. dx Assumption 3 (b) is nonrestrictive, since ip is only used to approximate the conditional expectation of <p. Notice that in assumption 3(d) we assume that the d e r i v a t i v e s of the approximating function sequence approximate any con tinuous function. In Lemma 1 of Appendix 1 we show that assumption 3 (d) implies that any absolutely continuous function can be uniformly approxi mated by some sequence {ip p In practice, assumption 3 (d) is not very restrictive. For most commonly-used approximation functions, such as poly nomials or splines, the derivatives are themselves a class of approximating function. (Step functions are one exception.) lsThe restriction that D v be compact is without loss of generality. In the case that the parameters may have to be arbitrarily large for obtaining an approximation (as it may be the case, for example, with polynomials), the {A} sequence is constructed so that there is a bound on all elements of {A} but this bound goes to infinity as v grows. 21 There is a popular misconception that the derivatives of polynomial ap proximators must diverge from the derivatives of the target function when the approximators become uniformly close to the target. This is true only for certain ways of constructing polynomial approximators; for example, if the approximation is required to fit a v th order polynomial to the target function exactly on v points. It is not true, however, if the approximating sequence is chosen to minimize other criteria, such as the l? distance, that take into ac count the fit of the function at many points. For example, polynomials would fit Assumption 3, since the derivative of a polynomial is itself a polynomial, and our Lemma 1 shows that an approximating sequence can be chosen with bounded derivatives. In this paper, we restrict our attention to stationary and ergodic pro cesses. D efin itio n {zt°(0 )} and {zt°°} are stationary and ergodic processes sat isfying equations (28) and (7), respectively, for t = 0 , ± 1 , ± 2 , ... That is, { z t} and (z<(/?)} denote processes generated respectively by (25) and (12) starting from a fixed initial condition zo, while {zf°}t^_oo and {z(° °(/?)}£-oo denote stationary and ergodic processes. The processes {zt{P)} and {zt) are, in general, non-stationary, since the initial condition is fixed, rather than a draw from the stationary distribution. There is a final set of issues to be addressed. In common with virtually all solution methods discussed in section 3, PEA delivers an approximation, Hp, to the equilibrium law of motion H . This approximate law of motion generates a time series { zt((3)}J= l, starting from some arbitrary initial con dition, which is used to obtain inferences about the stochastic properties of the true stationary equilibrium process zf°. For this procedure to be valid, three conditions must hold: i) {zf°(/?)} and { z f5} must exist; it) the effect of the initial condition must decline as t grows; iii) if Hp is close to H , then zf°(/?) must be close to zf°. These three conditions are closely related; most processes either satisfy all three conditions or violate all of them. None of these conditions hold, for example, if { z t} were an explosive process or a random walk. Insuring that these conditions hold is not only a problem for PEA. They must hold for any proposed approximation procedure if the approximator is 22 to converge to the true stationary equilibrium process. For example, if H were explosive, an approximation method that used a fixed grid (x i, ...jX,,) could never converge to the true equilibrium, since the endogenous process would eventually explode out of the pre-specified discrete grid. These considerations require additional regularity conditions both on the true equilibrium and on the space of admissible PEA approximators. We propose two distinct sets of conditions, either of which is sufficient to prove proposition 1. In the first, more general, approach, we simply assume the •needed conditions directly. Formally, for an arbitrary law of motion H a (where a is an index) let { z°} denote the process generated by H a starting at fixed initial condition z0, and let {zf00} denote the stationary process associated with H a. D efin itio n S is a closed set (in the sup norm) of laws of motion such that, for all H a 6 S (a) a stationary process with support in the set Z U ° exists (b) if initial condition (z0,u i) € Z U °, sup0 |zf — zf00! —* 0 almost surely as t —►oo, uniformly in the initial condition zq. (c) for any sequence of functions such that H k —* H a in the sup norm as k —» oo, we have zf°k —> zf00 almost surely as k —» oo, uniformly in t. Note that condition (b) allows for a process with several disjoint ergodic sets. All that is required is that, once the process is in the ergodic set, the effect of the initial conditions disappears. A ssu m p tio n 4* The law o f motion H is in the interior o f S . This assumption can not be verified directly from knowledge of system (1). There is a large literature on how to verify parts (a) and (b) of the definition of S analytically.16 Part (c) in this definition is essentially a robustness condition; if it did not hold, one can make the case that the model at hand is not a particularly interesting one, since the time-series properties of the 16For example, Marimon (1989) shows existence of a stationary and ergodic solution of the growth model under very general assumptions. Santos [1991] verifies part b) for dynamic programming models. 23 model’s endogenous variables would be highly sensitive to small deviations from fully rational behavior. Alternatively, one can verify assumption 4* by using proposition 2 below: if assumption 4* is the only assumption to fail, this will;be detected with PEA by the absence of an asymptotic fixed point. Under assumptions 1, 2, 3, and 4* , the approximate PEA solution con verges to the true stationary equilibrium, in the sense of Proposition 1 below, if the set D„ of admissible parameters is restricted to the subset B f , defined by B f = {fi € D„:HP € S ) . (29) A potential problem with using assumption 4* is that it is difficult to check formally if j3 € B f . In principle, (5 €B ^ can be checked infoimally, as follows: part (a) can be tested by observing if the solution settles around a stationary distribution; part (b) can be tested by re-doing the calculations with different z0; finally, if part (c) were not satisfied, the series will not settle down even with small changes in and the user will notice that the iterations to find the fixed point will not converge. Our second set of regularity conditions avoids this problem: the needed restrictions on the /?„’s can be checked formally. The cost is that this second set of regularity conditions is somewhat less general than assumption 4*. The alternative conditions uses the following version of Duffie and Singleton’s (1993) asymptotic unit circle (AUC) condition: A ssu m p tio n 4 (a) H has a stationary and ergodic solution with support in the set Z U . (b) There exists a sequence o f positive random variables {/>(««)} satisfying Elog[p(ut)] = a < 0 a. s. (30) such that H (',u ) has Lipschitz coefficient p(u) in Z U . Assumption 4 is a nonlinear analogue to the unit circle condition in linear time series models. While this is more restrictive than 4*, it is still satisfied in most models of interest. First, the Lipschitz constants on H need to be imposed only on the ergodic set of the true solution. Second, as long as the expectation in (30)is less than zero, the condition permits Lipschitz constants greater than unity for a subset of U with positive (possibly substantial) 24 probability measure. Finally, the assumption is silent about the particular transformation of the endogenous variables used to construct the z t series. For example, the simple growth model of Example 1.2 for p = 0 is known to satisfy this condition if we write the law of motion in terms of log(kt), even though this may not be true if H is written in terms of k t directly. Under assumption 4, we must impose an analogous AUC condition on the set of admissible approximators, which will be denoted B ^ u c : D efin ition : For all i/y 1&*UC is a closed subset of £>„ with the prop erty that, V/? € B * u c , there exist psitive constants Sp and positive random variables {p/?(ut)} satisfying Elog[p/3(u()] < a < 0 a.s. such that, for all || (31—/? || < (31) H (•, u t; /?') has Lipschitz coefficient pp(ut). The expectation in (31) is with respect to the stationary distribution of u t. For any given /?, condition (31) can be verified numerically: set pp(ut) equal to the maximum derivative of //(•, u t; /?'), and integrate numerically over the u 's.17 Under either assumption 4* or assumption 4, we must restrict the ap proximate solution f3vj to a subset B u € D„, where B„ equals B f or H * uc depending on which assumption is used. Under this restriction it is not easy to guarantee that G ^ t has a fixed point, since G u<t now maps B„ into a larger set D„. One way to guarantee the existence of a fixed point is to re strict the minimization that defines G v> t in (18) to the set B„. The function G u>t would then map B,. into itself, and existence of a fixed point would follow from Brouwer’s theorem. It turns out that this strategy does not de liver a proof of Proposition 1: we must use the fact that G>tr minimizes the 17Another consequence of using assumption 4 and B^yc is that the proof becomes much more involved than if B^and assumption 4* were used. The reason is that the proof needs to show that H can be arbitrarily well approximated by a sequence of H j s with P € B * u c . In other words, we need that PC\{rpp : P G U ^ xB^yc} is dense in the subset of { h : R l —* flm} containing S . Now, since U£Lj B AUC has an empty interior, this does not follow immediately from the fact that P is dense in { h : R 1 - * R m } . On the other hand, denseness of Pfl [ipp : /? GU^LjBf} in the subset of [ h : R 1 —► containing 8 follows immediately from assumption 4* and the definition of this set. 25 mean square error among all admissible /? € D„. (See, for example, Lemma 8, equation 62). Instead, we use the following definition of an approximate solution, which generalizes equation (19): % D efin ition : An a p p ro x im a te so lu tio n o f order v, sa m p le size T , is a parameter vector /?„,t satisfying Pv,T = argmin ££f=i I0(/?,*<(/?)) ~ 'P{G„,T(/3),xt(/3))\2 (32) Under assumptions 1 to 3 an approximate solution always exists by con tinuity and boundedness of the objective function. If G„tr had a fixed point in the set B„, the two definitions would obviously coincide. In practice, a fixed point usually can be found. Even if no fixed point exists for finite v and T, lemma 9 in the appendix implies that as v, T —►00 4 IVKA'.t , zt(/?i/,r)) - V ^ . H / W ) , “ ♦O* 1 «=i This makes precise the sense in which and C?„,r(/3) can be made arbitrarily close. In that sense, the approximate solution (32) represents a fixed point asymptotically. The minimization problem in the above definition is over a restricted set. Finding constrained minima numerically in non-linear setups is often difficult. In practical applications of PEA, one can avoid imposing this restriction directly by solving an unconstrained minimization along a homotopy path for which the constraint does not bind. More details are given at the end of next section. Notice that, either under assumption 4 or under 4*, we may have non uniqueness. This can happen because system (1) may be satisfied for several laws of motion H , or because a given law of motion has several ergodic sets ZU . Our discussion of examples 1.1,1.2 and 1.3 shows how to pick out unique solutions in some cases. Now we present the fundamental proposition in this section.18 l8It is understood that the double limits in the paper are defined as: v lim C„,r = v lim (Jim G , t /) — oo \T - > 0 0 ,T — oq 26 P ro p o sitio n 1 (C onvergence of A pproxim ate Solutions): A ssum e th a t th e re is a u n iq u e s o lu tio n w ith a u n iq u e s t a t io n a r y a n d e r g o d ic d i s t r i b u tio n ; d e n o te th e s u p p o r t a s Z U . either a s s u m p t i o n B„ = B ^yc, t h e n a d d itio n , and lim ( sup U n d e r a s s u m p t i o n s 1, 2 , a n d 3 , i f , i n 4 * h o ld s a n d | H (z, u B„ = B f or \(3 ^ t ) — H ( z , u ) \(z,u )e zff lim I sup a s s u m p tio n 4 h o ld s | 1 = 0 , a.s. ) | d’ if li'T , z) —S { x ) | I = 0, a.s. \xe/(zU ) ) (All proofs are in Appendix 1 .) Notice that convergence obtains in the strong sense of uniform convergence. Also notice that convergence is guaranteed only in the ergodic set ZU . In case that the solution is non-unique or has several ergodic sets, a trivial modification of the proof of Proposition 1 would show that the approximate solution becomes arbitrarily close to the set of ergodic solutions The following corollary asserts that all properties of the model that are of interest in time series applications are appropriately approximated by PEA. C orollary 1 U n d e r t h e c o n d i t i o n s o f P r o p o s i t i o n 1, i f ( z0 ,u i) € ZU , we have (a) (S im u la te d S o lu tio n P a th s con verge) Z i {(3v>t ) (b) —►Zt a .s., u n if o r m ly in t, ( E q u ilib r iu m C o n d i t i o n s a re s a tis f ie d in th e l i m i t ) g ( E [<j>(zt+ 1 ( p v tT ) ) \ x t ( 0 „tT)\, zt(&,r (c) ), ( S i m u l a t e d S a m p l e M o m e n t s c o n v e r g e ) : I f d : Z —* «t) -» 0 R ? is a n y L ip s c h itz fu n c tio n , th en -» £(d(zt°°)) 1 t=i 27 a .s. a .s. a s u , T —* oo. We now provide a partial converse to Proposition 1 . It shows how the conclusion of Proposition 1 depends on assumption 4* or 4, and provides a way to check these assumptions in the limit. Specifically, if B„ = B f , part ( a) says that if H does not satisfy assumption 4*, this can be detected because min ||Vv3 —V’G„(/?)| is eventually bounded away from zero; part ( b) tells us that those equilibria that fail to satisfy assumption 4* will not be approximated by PEA. (An analogous interpretation holds for B„ = B ^1707.) For example, this implies that rational expectations bubble equilibria in Ex ample 1 .1 will not be approximated by PEA with the choice of state variables proposed in section 2 . P ro p o sitio n 2 (a) U n d e r a s s u m p t i o n s 1, 2 a n d 3, t/B*, = Bjf a s s u m p tio n 4 is n o t sa tisfie d , th en th e re ex ist th a t W > or i f B„ = H * u c 7 > 0 and N > 0 a n d a s s u m p tio n 4 * is n o t s a tis fie d and such N min E [ r J > ( P ,x ™ ( 0 ) ) - t{G „ {(3 ),x ? (0 ))]2 > 7 PfcOl/ (b) i f B„ = B J a n d th e r e e x ists an eq u ilib riu m H A s s u m p t i o n 4*> o r i f B„ = th a t d o e s n o t s a tis f y a n d th e re e x ists a n e q u ilib r iu m H t h a t d o e s n o t s a t i s f y A s s u m p t i o n 4 , t h e n H p vT d o e s n o t c o n v e r g e t o H . Proposition 2 can be used to verify assumption 4* or 4 by calculating i f ; 11 1 ii! t = 1 for v and T arbitrarily large. If B„ = B f , assumption 4* is satisfied if and only if this infinite sum can be made arbitrarily close to zero. (An analogous verification of assumption 4 holds for B„ = B ^170.) This is important be cause verifying assumption 4* or 4 analytically may be difficult if one simply inspects the equilibrium conditions ( 1 ). 28 5 S o m e 5.1 P r a c t i c a l Issues. S i m p l e A l g o r i t h m s for f i n d i n g (3u>t In order to calculate the fixed point of Step 4 (or the arg min in equation (32)) one can use standard hill-climbing algorithms for solving non-linear systems of equations. Nevertheless, this may not always be the best alternative. First, calculation of the gradient of can become very expensive in models with many coefficients. Second, is only well defined if /3 G B u . We will discuss the second issue in the subsection on homotopy, below. In order to avoid the first problem, the following algorithm based on modified successive approximations has been used successfully in many applications M r + 1) = (1 - A)fl,,T(r) + AG „ , t { M t )) (33) for some A > 0 . It is clear that no gradient has to be calculated, so that the iterations are extremely easy to program, and each iteration is done very quickly. This algorithm needs more iterations to converge than a gradient algorithm, but there is a trade-off between ease of performing each iteration and number of iterations needed to converge. In practice, the above algorithm is often at least as fast as gradient algorithms. Although one can construct examples where iterations on (33) are locally unstable, this has not been the case in most practical applications up to date. Furthermore, since the least-squares learning model is locally stable if and only if the differential equation (23) is stable, iterations on (33) with A small diverge only if the model is locally unstable under learning. If this were the case, the model would be uninteresting from an economic standpoint. Another fast algorithm that avoids calculating the gradient is to simulate directly the model under least squares learning. More precisely, Steps 2, 3 and 4 are substituted by • Step 2’: For T large enough, draw a sample of size T of the exoge nous stochastic shock and calculate { z t , U t } J =Q using the least squares learning model of Appendix 2 . Again, if the model is stable under learning /?t will converge to the fixed point. 29 5.2 S p e e d o f C o m p u t a t i o n , Initial C o n d i t i o n s a n d H o m o topy. The design of S t e p s 1 t o 4 in s e c t i o n 2 i s d e s c r i b e s t h e s i m p l e s t p o s s i b l e a p p l i c a t i o n of PEA. A number of elementary modifications can be used to speed up computations, for example: one iteration on the algorithm for running the non-linear regressions of Step 3 is sufficient to deliver the fixed point; g can be often rewritten in a way that solving for z t in Step 2 is very simple; the class of functions P can be chosen to match some properties of the con ditional expectation (for example, it can be set to take only positive values); and, as suggested by the definition in equation (32), it is better to place the convergence criterion used to decide that the algorithm has arrived at the fixed point on the values of 0 ^, instead of on the values of /?. Suppose a researcher has calculated the solution of order u and wants to calculate the solution of order j/ + 1. In this case, using (3v>t as initial condition for the iterations to find the v + 1 approximation is not a good alternative, since the elements of the higher degree will usually be correlated with those of a lower degree; instead, it is best to start the algorithm at Gv+i,T(/9i/,T)Furthermore, it is not always necessary to introduce all elements of degree v + 1. Only those higher-order elements that have some incremental predictive power for <f>(zt+1(/?)) need be included. It is likely that the introduction of textbook techniques from numerical analysis will be useful for finding minima, computing efficient Monte-Carlo integrals, introducing alternative flexible functional forms, and setting homotopy paths. One has to be careful not to introduce these techniques unless there is a good reason; oftentimes, the simplest approach will be sufficient for solving the model of interest. In many cases, a simple algorithm is also the fastest alternative. Many successful pplications of PEA make use of homotopy techniques; this is a simple way of obtaining starting values for the algorithm and of keeping the simulations in the stable set B„. Along a homotopy, the desired solution is obtained by moving slowly from a known solution to the solution of the model we are interested in. For example, den Haan and Marcet (1990) calculate the simple growth model of example 1 . 2 with partial depreciation by starting out at the solution for the case p = 0 (a case for which we know the analytic solution), and then solving a series of models increasing p gradually. At each step along the homotopy, the solution for the previous 30 step is used as initial condition. In this way, one insures that all iterations stay close to the true law of motion and, therefore, away from the boundary of B„. Therefore, there is no need to use c o n s t r a i n e d minimization routines in order:to solve the minimization problem in (32). PEA has turned out to be a fast and simple method in many applications. A comparison of speed of convergence, accuracy and convenience in a highly non-linear model has been made by Christiano and Fisher (1994); they use a simple growth model with irreversible investment (our example 1.3) and as sume that the shock 6 t is a discrete Markov chain. Their model is particularly unsuitable to PEA, since it only has one continuous state variable and the stochastic shock can take only two possible values: in this particular model, endogenous oversampling is not very useful, and Monte-Carlo integration is very inefficient, since integrals £ ( x i ] ( 3 ) are given by a simple formula that is exploited in the other algorithms but not in PEA. The computation times are higher for PEA but, even for their model, only by factors of five or seven. Furthermore, of the several algorithms they tried, PEA was the only one where the computation time increased only slightly when a non-negativity constraint was introduced. Finally, Christiano and Fisher find that PEA delivers a very accurate answer for low-degree polynomials and it is, by far, the easiest method to implement. The claim in Judd (1992) that MWR methods are ’hundreds of times faster’ than PEA is based on comparing the computation times of den Haan and Marcet (1990) with his own computations. These computation times are not comparable because different starting values were used: den Haan and Marcet (1990) (who were unaware that they were in a race) deliberately started the algorithm at a very incorrect initial condition to illustrate the use of homotopy. The solutions reported in Judd (1992) use the non-stochastic s t e a d y - s t a t e s o l u t i o n as starting value, which happens to be very close to the stochastic solution. Despite their model choice, the comparison done in Christiano and Fisher (1994) is more informative, since they start out all the algorithms at equivalent initial conditions. 6 Conclusion We have presented the PEA algorithm for solving dynamic stochastic non linear models with rational expectations. This approach is highly flexible, 31 and it has been used successfully in many applications with suboptimal equi libria, strong non-linearities and inequality constraints. It is particularly fast in those models with a large number of state variables and stochastic shocks, and is quite easy to implement. We prove that, for models with continuous laws of motion and under some mild regularity assumptions, the approximate solution converges to the true solution with arbitrary accuracy as the approximation is refined. Con vergence has been proved for the case of the stationary distribution. Similar results for other techniques solving Euler equations with finite dimensional approximations are not yet available. A very general condition requiring, essentially, ergodicity of the true pro cess, is enough for convergence. We have also proved convergence under the more restrictive assumption that Duffie and Singleton’s AUC condition has to hold in the support of the ergodic distribution in order to guarantee that the stability condition on the approximate process can be verified formally. Future research might deliver similar convergence results for PEA under dis continuous laws of motion, but certain technical details, such as a condition for compactness of the class of approximating functions and an equivalent stability condition, need to be addressed. Our restriction to the support of the stationary distribution is non-essential: a result that guarantees con vergence of the law of motion outside the ergodic set can be found in the predecessor to this paper when the algorithm uses short-run simulations. The convergence theorem in this paper represents a first step in studying the properties of PEA and other numerical methods for solving systems of Euler equation. Additional work needs to be done. For example, it would be useful to understand the optimal rate at which u and T should be allowed to grow. A more formal method of selecting the higher degree elements to include when increasing u would be valuable. Finally, the use of more sophis ticated homotopy techniques can improve the speed of convergence in large models, and it can provide a basis for a theorem proving that the iterations to find the approximate solution converge globally from the initial condi tions. We believe that this area of research will pay substantial dividends in expanding the range of economic models accessible to quantitative analysis. 32 References • Baxter, M. (1991); "Approximation of Suboptimal Dynamic Equilib rium: An Euler Equation Approach”, J o u r n a l o f M o n e t a r y E c o n o m i c s , October, pp 173-200. • Bansal, R. (1990); ”Can Nonseparabilities Explain Exchange Rate Move ments and Risk Premia?” Ph.D. thesis, Carnegie-Mellon University. • Bansal, R., R. Gallant, R. Hussey and G. Tauchen (1994); ”Nonparametric Estimation of Structural Models for High-Frequency Cur rency Market Data”, forthcoming :n J o u r n a l o f E c o n o m e t r i c s . • Bekaert, G. (1993); ’’The Time-Variation of Risk and Return in Foreign Exchange Markets: A General Equilibrium Approach,” Working Paper No. 1276, Stanford University, Graduate School of Business. • Christiano, L. and J. Fisher (1994); ’’Prototype Algorithms for Solv ing Dynamic Models with Occasionally Binding Constraints”, working paper, Northwestern University and University of Western Ontario. • Coleman, W.J. (1988) "Equilibrium in a Production Economy with an Income Tax” , working paper, Board of Governors of the Federal Reserve System. • den Haan, W. (1990); ’’The Optimal Inflation Path in a Sidrausky Model with Uncertainty,” J o u r n a l o f M o n e t a r y E c o n o m i c s vol 25, 389410. • den Haan, W. (1991); "The Term Structure of Interest Rates in Real and Monetary Production Economies” working paper, Carnegie Mellon University. • den Haan, W. (1993); "Aggregate Uncertainty, Heterogeneous Agents, and Short-term Interest Rates”, working paper, UC San Diego. • den Haan, W. and Marcet A. (1994); "Accuracy in Simulations”, v i e w o f E c o n o m i c S t u d i e s , January. 33 R e • den Haan and Marcet (1990); ’’Solving a Simple Growth model by Pa rameterizing Expectations” Journal of Business and Economic Statis tics, January, pp 31-34. • Diiffie, D., and K.J. Singleton (1993); "Simulated Moments Estimation of Markov Models of Asset Prices,” Econometrica, vol. 61, 929-952. • Hansen, L. P. (1982); "Large Sample Properties of Generalized Method of Moments Estimators”, Econometrica, vol. 50, 1029-1054. • Heaton, J. (1993); ”An Empirical Investigation of Asset Pricing with Temporally Dependent Preference Specifications,” Working Paper, Al fred P. Sloan School of Management, Massachusetts Institute of Tech nology. • Ingram, Beth (1990); ’’Equilibrium Modelling of Asset Prices: Ratio nality Versus Rules of Thumb”, Journal of Business and Economic Statistics. • Judd, K. (1992); ’’Projection Methods in Optimal Growth Models”, Journal of Economic Theory, December, pp 410-453. Appeared as a working paper (1989) ’’Minimum Weighted Residual Methods in Dy namic Economic Models”, Hoover Institution. • Judd, K. (1993); ’Comments on Marcet, Pakes and Rust’, discussion, forthcoming in Advances in Econometrics, ed. C.A. Sims, Cambridge University Press. • Judd, K. (1994); ’Numerical Analysis in Economics’, mimeo. • Ketterer and Marcet (1989); ’’Introducing Derivative Securities: A General Equilibrium Approach”. Preliminary Draft, Carnegie Mellon U. • Kolmogorov and Fomin (1970); Introductory Real Analysis, Dover Pub lications, Inc. • Ljung, L. (1975); ’’Theorems for the Asymptotic Analysis of of Recur sive Stochastic Algorithms”, Report 7522, Department of Automatic Control, Lund Institute of Technology, Sweden. 34 • Lucas, R. E. (1978); "Asset Prices in an Exchange Economy”, Econometrica, vol. 46, 1426-1446. • Marcet, A. (1988); "Solving Non-linear Stochastic Models by Parame terizing Expectations”, working paper, Carnegie Mellon University. • Marcet, A. (1990) "Simulation Analysis of Stochastic Dynamic Mod els: Applications to Theory and Estimation”, Invited Symposium of the VI World Congress of the Econometric Society. Forthcoming in Advances in Econometrics, ed. by C.A. Sims. • Marcet, A. and Marimon, R. (1992); "Communication, Commitment and Growth”, Journal of Economic Theory, December, pp 219-250. • Marcet, A. and D.A. Marshall (1992); "Convergence of Approximate Model Solutions to Rational Expectations Equilibria Using the Method of Parameterized Expectations,” Working Paper No. 73, Department of Finance, Kellogg Graduate School of Management, Northwestern University. • Marcet, A. and T. J. Sargent (1989); "Convergence of Least Squares Learning in Environments with Hidden State Variables and Private Information”, Journal of Political Economy, vol 97, 1306-1322. • Marcet, A. and K.J. Singleton (1990) "Equilibrium Asset Prices and Savings of Heterogeneous Agents with Portfolio Constraints” Working Paper, Carnegie Mellon University. • Marimon, R. (1989); "Stochastic Turnpike Property and Stationary Equilibrium”, Journal of Economic Theory, pp. 287-306. • Marshall, D. (1988); "Inflation and Asset Returns in a Monetary Econ omy with Transaction Costs” . Ph.D. thesis, Carnegie Mellon Univer sity. • Marshall, D. (1992); "Inflation and Asset Returns in a Monetary Econ omy”. Journal of Finance, vol. 47, 1315-1342. • McCurdy T. and Ricketts, N. (1992); ” An International Economy with Country-Specific Money and Productivity Growth Processes”, working paper, Queen’s University. 35 • McGratten, E. (1993); ” Solving the Stochastic Growth Model by a Finite Element Method”, Staff Report, Federal Reserve Bank of Min neapolis. % • Novales, A. (1990):”A Stochastic Monetary Equilibrium Model of the Interest Rate” Econometrica, vol.58, 93-112. • Otker (1991); "The Welfare Implications of Fiscal Policy with Dis tortionary Taxation; a Numerical Analysis” , working paper, Carnegie Mellon University. • Rojas, G. (1991); "Optimal Taxation in a Model with Public Capital”, working paper, Universitat Pompeu Fabra. • Santos, M. (1991); "Smoothness of the Policy Function in Discrete Time Economic Models”, E c o n o m e t r i c a , vol. 59, 1365-1382. • Sims, C. A. (1985); "Solving Non-Linear Stochastic Equilibrium Models ‘Backwards’” , manuscript, University of Minnesota. • Townsend, R. (1983); "Forecasting the Forecasts of Others” , o f P o l i t i c a l E c o n o m y , vol 91, 546-565.• • Wheeden and Zygmund (1977); Inc. M e a s u r e a n d I nt egr al , 36 Journal Marcel Dekker APPENDIX 1 PROOFS AND LEMMAS For ease of exposition and notation, the propositions and lemmas are proved for the case where G u and are well-defined functions (that is, the least-squares minimizers in (16), (18), (32) and , (54) are unique). This avoids dealing explicitly with convergence of correspondences. All propositions are valid for the general case where these objects may be non-unique. First, suppose G vj { P ) were multi-valued. For any £ € G ^ r i P ) , fakes the same value, so the minimand in (32) is well-defined. If the minimizer in (32) were non-unique, there would be multiple equivalent PEA approximators for each value of v and T . In that case, one could still construct a doubly-indexed sequence {/?k,t } by selecting any one of the equivalent PEA approximators for each v and T . Any such sequence converges to the true equilibrium, in the sense of proposition 1. Finally, throughout this appendix, || • || denotes the L 2 norm, and B = U„>iB„. The first lemma proves that P is dense in the space of conditional expec tations. If € was differentiable, the lemma would follow almost trivially from the fundamental theorem of integral calculus. Most of the derivations in the Lemma are to handle nondifferentiabilities in £ . L em m a 1 The re exist s a sequence /?„ € D*, such that (*) d£(x) lim ^ fe l */->oo (jx Ox ’ a.e. in X . (b) Jjm sup |V>(/?„,:r)-£(x)| = 0 (c ) l i m s u p v-><x> sup x€X dx 37 —sup xex d£(x) dx Pr oof. To accommodate points of non-differentiability, our proof strategy in volves smoothing the function £ . We first define and characterize kernel approximator functions. For k = 1,2, • • •, let n k : R p —►R m be a continuous function for which f Nk Kk{t)dt = 1, and let A C X be a set of Lebesgue measure zero. Let q : (X — A) —►R m be any arbitrary bounded continuous function. The kernel approximator function q k is defined as follows: q k( x ) = J k <l(x + t ) n k( t ) d t In a technical appendix, available upon request, we prove that limsupsup | k—>oo x£X q k( x ) | = sup | x£X q(x) |. (34) Property (34) will be used in proving part (c) of the lemma. To prove the lemma, choose x € X . By theorem (7.29) in Wheeden and Zygmund (1977)), since £ is absolutely continuous, the derivative of £ exists a.e., it is uniformly bounded, and we have for all x € X . Let d £ k denote the kernel approximator to be defined as follows: and let £ k{ x ) = J X _ d £ k( x ' ) d x ' + M k, where the scalar sequence Clearly, { M k} is chosen so that d £ k( x ) = so, in particular, d£k £ k( x ) = £ ( x ) £k (35) for all k. d £ k( x ) ’ is differentiable everywhere and, lim k—*oo dx d£(x) d £ k(x) dx dx = 0 pointwise, except on a set of Lebesgue measure zero. 38 (36) We first prove that lim sup \ £ k( x ) —£ { x ) \ = 0. *€* 1 1 (37) Given our construction (35), it is enough to show that the integrals of d £ k converge uniformly to the integrals of |£ . Let K denote the Lipschitz coefficient on £, which exists by virtue of assumption 1(c). It follows that K is a bound on and (by virtue of (34) on 8gg . Using Egorov’s theorem (see Wheeden and Zygmund (1977) page 57), we know that convergence a.e. of measurable functions implies uniform convergence except in a set of arbitrarily small measure. Therefore, given any e > 0, we can find A C X such that | A |< c/(4 K ) (where, | A \ represents the Lebesgue measure of the set A ) and sup x£X—A d £ k{ x ) d£(x) dx dx 0 as k —►oo In particular, there exists a k such that, for all k > k , s u p * ^ ^ | d£Q ^ ~ c/(2 |X |). For such a k , Egorov’s theorem implies that, sup |^ (a:) —£(z)| = sup xex 1 1 rgX < sup x€X d£(x') Jx dx Jx UX dx' d £ k{x') d£(x') d £ k{x') d£(x') dx dx dx dx < sup ( f , , e/{2 | A' xex \Jlx,x]-A \)dx'+ | < dx I A | 2A') <| A 11/(2 | A |)+2A e/(4 J ■) K) This completes the proof of (37). Parts (a), (b), and (c) of the Lemma can now be demonstrated. Fix By Assumption 3 there exists a sequence f3k € D v such that lim sup xex di>(ff,,x) d £ k{ x) dx dx 39 0. = e, k. (38) Equation (38) implies Jfirn sup |v>(/?*,x) — £fc(x)| = 0. (39) and 9 rp (^ l,x ) lim sup *'-00*6X 8 £ k(x) — sup dx (40) dx xex We know from the properties of kernel approximators that £ k —> £ k —► oo and that, for each k ; from (39) we have that v —►oo. Furthermore, (34) implies that Pi/ as —* £ k uniformly as suPa:€A' sup^g^ |^ f^ | as k —* oo. Therefore, a sequence {/3 „} can be constructed by taking the appropriate elements from the doubly indexed sequence {/?*} so that parts (a) , (b), and (c) of the Lemma are satisfied. Q E D The next lemma proves that restricting the approximation to stay in the set of well-behaved laws of motion B approximation. does not preclude an arbitrarily good Formally, the lemma could be stated as saying that P is dense in B . L em m a 2 L e t G B„ f o r v s u f f i c i e n t l y l a r g e . b e a s i n l e m m a 1, t h e n Proof. In the case where assumption 4* is made and B„ = B^, the lemma follows immediately from part (c) of the definition of 5, since H is in the interior of S . The remainder of the proof is for the case where assumption 4 is made and B„ = 3 * u c . We use the following result: sup xex £ { x ) ~ £ (*) — sup x —x xex d £ (x) dx (41) A proof of equation (41) is in a technical appendix available from the authors. Assumptions 1(a) and 3(a) imply that H ( z , u \ ‘) is continuous in /?. There fore, to prove the lemma, it is sufficient to show that, for each v , the minimal Lipschitz coefficient of H ( - , u ] P u ) (which we will denote p p (tt)) satisfies limsuppjj (tt) < p ( u ) , 1/-+00 40 (42) where p(u) is the Lipschitz coefficient on H defined in Assumption 4. Un der Assumptions 1(a) and 1(b), the functions <7 , ( 7 2 and / are uniformly Lipschitz in all arguments. Therefore, the following coefficients are finite: g{oc,z',z, u ) - g ( a ,z ', z,u) a —a Of,Qft2#,Z K \{u) = max I<3 ( u ) = g{<*, z',z,u) - g (a,z',z,u) z —z a,7,z',z max K ( u ) = max CXyZyZ g i l (a,z,u) - g?x{cc, z,u) z —z K / ( u ) = max Z,Z where z, z, z € Z, and a € <f>(Z). /(*»“ ) - /(* ,« ) z —z Assumption 4 then implies that \l<\(u)I<j(u)K -I- A'3(u)] K ( u ) < p(u) where K = supx6^ |d |. Let (43) Kp be defined as follows: 50(/?,x) dx Kp = sup z€A' It follows that pp(u) = [K \(u )K j(u )K p + A'3(u)] I<(u) (44) Equations (43), (44), and Lemma 1(c) immediately imply (42). QED. The next Lemma collects a number of results that will be used later. Let X \ let H* : Z x U —> Z be defined £ * denote an arbitrary function defined on by $(£*(/(*»«))» H ‘ ( z , u), z, u) = 0, so H mis the law of motion consistent with £*; let coefficient of //*(•, u). p’ (u) be the Lipschitz If assumption 4* is made, we will assume that H* € S. If assumption 4 is made, we will assume that H* satisfies the A U C condition in (31), so that 41 J3[log[/>*(u<)]] < a" < 0 a.s. (45) Let zf3'* b e a sta tio n a ry an d ergodic process satisfying zf3'" = H*( For any 'given zeZ, define {zt(r,z; recursively as follows: Z-T(T,Z]0) = z Zt(T,z-,/3) = t f ( z t_ i ( r , J ; / ? ) , u t; 0 ), t> -r, so z< (r,z;/?) is o b tain ed w ith th e law of m otion Hp, sta rtin g a t z a t d ate —r . Sim ilarly, {zf(T,z)}?l_T is o b tain ed w ith th e law of m otion H * and initial condition z_T(r, z) = z. N ote th a t zt(r ,z ;/3 ) an d zt* ( r ,z ) depend only on {uj}j-__T+1, while th e statio n ary processes zf°’’ an d zf°(/?) depend on Finally, let {/?*} be a sequence such th a t lim sup \xj){Pl,x) — £ ' ( x ) \ = 0 xgX (N ote th a t, according to L em m a 1 , {/?„} plays th e role of th is sequence if £• = S). L em m a 3 (a) zf°-* exists, an d lim r _oo | (b) For all (3 G B , 0 a.s. z *{t ,z) — zf°-* | = 0 a .s. zf°(/?) exists, and lim T_oo | zt( r , z ; /?) — zf°(/?) | = (c) For any v and any t, the restriction of zf°((3) to B „ is continuous in /3 a.s. (d) lim^—oo su p l€ 0 u6t/ | H(z,u;0*) — H*(z,u) | = (e) lim^_»oo s u p je ^ y T>j | z ,(r, z; ft) - z* (r, z) | = 0. 0 a.s. Proof. U nder assum ption 4* w ith B „ = B f , p a rts (a) an d (b) follow im m ediately from p a rts (a) an d (b) of th e definition of S. U nder assu m p tio n 4 w ith B „ = B * u c , p a rts (a) an d (b) follow d irectly from Duffie an d S ingleton’s [1993] L em m a 3. P a r t (c) is proven as follows: Let e > 0 b e given. It is sufficient to show th a t 35 > 0 such th a t | /? — 0 | < 5 im plies th a t | z£°(/?) — z ” (/3 ) | < e 42 a.s. For any z € Z and any r, . I ZV 0 ) ~ ZF(P) I £ I *i°°(0) “ zt(T,z-,P) | + I zt{T,z\p) - zt(T,z;(3) | + | zt(T,z;/3) - z f ° 0 ) \ (46) According to part (b) of this lemma, r can be chosen big enough so the first and third terms on the right-hand side of (46) are less than e/3 almost surely. Furthermore, H ( z , w , •) : B —» Z is continuous, so z t ( r , z ; •) : B —* Z is continuous. It follows that there exists S > 0 such that | /? — /? | < S, and the second term on the right-hand side of (46) is less than e/3. Part (d) of the lemma is proved as follows. Assumption 1(b) insures that g 2 * is uniformly Lipschitz. Since and H ’ (z ,u ) = g ^ 1 ( £ ’ ( f ( z , u ) ) , z , u ) H { z , u; P I ) = g ^ f ( z ,«)), z , u) we have, for some K < oo, sup | H " ( z , u ) - H { z , u \ 0 ’ ) | 2,U < - I < sup | £ ‘ ( f ( z , u ) ) - x j > ( 0 ; , f ( z , u ) ) e | -+ 0 z,u as v —» oo. We now turn to part (e) of the Lemma. Under assumption 4* with B„ = B f , part (e) follows immediately from part (d) of the lemma along with part (c) of the definition of S . The remainder of this proof treats the case where assumption 4 is made and where B„ = B * u c . In this part of the proof, it is useful to notate explicitly the dependence of the z ’s on the draw from the sample space, so we use the notation ” zt(r, "z,u>; 0 ) " and ”z*(r, z,u>)” . According to part (d) of this lemma, for any <5 > 0 3 N ( S ) < o o such that supI U | H * ( z , u ) — H ( z , u ] (3‘ ) | < S, Vis > N ( 6 ) . Therefore, for date —r + 1, |z_ T+i(r,J,a > ;/? ;) - z 1 t+1( t , z ,w )| = | H ( z , u - t + \ ( u))] P I ) - H m( z , u - T + i ( u j ) ) | < 6, ' i u > N ( S ) . Proceeding recursively forward to date t , one obtains: <+T-l j I zt(r,z,u;p*)-z;(T,z,u) < S 43 h - 2 n />*(«<-*+!m ) 3=1 k=l W > N (S). Since delta can be made arbitrarily small, part (e) of the lemma follows if M(t,u>,r) = n i= i />*(u«-/fc+i(w)) is bounded for al most all w, with the bound uniform in r. (Notice that M (t,w ,r ) does not depend on z . ) Since log(/j*(u<(a;)) is stationary and ergodic, the sam ple mean of log[p*(u*(u>))] converges almost surely to its population mean. Therefore, equation (45) there exists a J ( t , u ) < oo such that, for almost all J E i= i log[?*(ut-Jfc+iH)] < a*,V ; > < M (t,u ,T ) and />•(«*-*+!(«)) < E / i 'r ’ n i_ , />•(«,-»+.(«))+E 5 , [«*•]'■ = E /IV * n i_ , />-(«,_*+!(«)) + e » 7 (l - e°‘) = K ( t , w ) < oo, for almost all u . This proves part (e) of the lemma. Q E D . The following lemma insures compactness of P in the supnorm. L e m m a 4 T h e f a m i l y o f f u n c t i o n s P = {V>(/?, •),/? € D } i s e q u i c o n t i n u o u s . Proof. Assumption 3 states that the derivative of elements of P is uniformly bounded, so I ~ V’(^,x) g0(/?,(g<) dx dx' < K X — X I QED. The next lemma is a uniform strong law of large numbers for the mean square prediction error. For / 3 , ( e D let h t ( ( 3 , £ ) = <f>(zt + \ { P ) ) — i l > ( £ , x t { ( } ) ) , let = ^(z1 ° ?,(^)) - V>(f,x~(/?)) and let 7 „,t be defined by: it? L em m a = 5 sup P£Bu£(zDt/ { ifE L .N ^ W -E lW .O F l} - lim r—oo' f t . T ~ 0 a.s. Proof 44 (47) The proof is sketched. 7t,T < SUP \ t 'ET=i h t ( P , 0 2 ~ h fiP iO 2 (48) ? + sup If T j = i h ? ( P , 0 2 ~ E hV { P , t Y h f ’(/?, £) satisfies the conditions of Hansen’s uniform strong law of large num bers for stationary and ergodic processes. (See Hansen [1982], Lemma 4.5.) It follows that the second term on the right-hand side of (48) converges to zero almost surely as T —►oo. Under assumption 4* with B„ = B^, the first term on the right-hand side of (48) converges to zero almost surely as an immediate implication of part (b) of the definition of S. On the other hand, if assumption 4 is made and B„ = 3 *uc, it can be shown that the first term on the right-hand side of (48) also converges to zero almost surely by using a slight modification of the proof of Lemma 4 in Duffie and Singleton (1993). QED. The following lemma insures that the non-linear regressions converge to the population least squares minimizer. Lemma 6 lim sup | G ut {P) —G„{P) | = 0, T-°°/3€Bv a.s. Proof. Let e > 0 be given. Let set Q(e, /?) £ D„ be the set of elements of D„ which are bounded away from G U(P). Formally: Q ( e , 0 ) = { ( z D l, s . t . \ ( - G M \ >0 Define 7r(e,/?) > 0 by *(£,/?) = - E [h r(P ,G M ,v )? } (50) and let 7r(c) be defined by 7r(e) = inf 7r(e, 3) /?gB„ (51) 7r(e) > 0 (52) Compactness of D v implies that 45 Equation (52) and Lemma 5 imply that 3 A 6 fl with prob (A) = 1 and 3 T ( u >, e) such that, for V u t A and V T > T(u>, e), o < 1;,t M : < ^ (53) where 7 *(T(u>) is defined in (47). Choose some w e A , T > trary /? 6 B„. iFrom the definition of T(u>, e), and arbi 0 < E[h?(0,G„,T{i),» ) , * ) ? - E { h ? { 0 , G M , u ) \ 2 1 t=1 T T + ? E IW , g ,,t W , » ) } 2 - ^ c .W ) ^ ) ) 2 1 t=1 1 t=l + 4 BM /J,G,,( 0 ),")]! - E \h r (fi,G M ,u )-p 1 t- 1 E[hT(0, + G,.r(/>, w), W)]:2- i; D M A <?,(/»), «)]J - i £ M l> , 1 (=1 G,.t W , E {h ? (0 ,G M «)]’ ,»)] 2 1 t=l < 2 7 *t (u>) < 7r(e) < tt( c, P ) where the third inequality follows from the definition of G ^ t ( P , u>), the fourth inequality follows from (53), and the fifth inequality follows from (51). This implies that G v ,t (/?,u ;) ft Q ( P , e), proving the lemma. QED. Lemma 7 states that the approximate solution for large enough T will be arbitrarily close to the "population” approximate solution of order u : /?„ = arg min || L em m a 7 Let g i v e n v, w e h av e rp(p, * V ( P ) ) ~ xT(P)) II (54) he the a p p r o x i m a t e s ol ut ion , as in e q ua t io n ( 3 2 ) . T he n, lim | T— +oo Pu,T — Pu 46 | = 0, a.s. Proof. We first show that the minimand in (54) converges to the minimand in (18) uniformly in (3 as T —*■ oo: 2 -E sup i t sup 1 t = 1PeBv i! | < w ,x,(0)) - t ( c , . T(fi),*tW) I2 (55) + sup PzBv 1 <=i We first show that the first term on the right-hand side of (55) converges to zero. According to assumption 3(a), tl>(-,x) is Lipschitz, uniformly in x £ X. Furthermore, the xj>: B„ x X —►R m is a continuous function of a compact set, so is bounded. These conditions imply that |0(/?, xt((3)) — ip(£,xt(P)) |2 is Lipschitz in £ € B„, uniformly in /? 6 B„, so 3A" < oo such that 11 *i>(fi,x,(p)) - ih g .,t (/3),x,(/3))i 2 - \ < K 0 , x , m - < H G . m , x , m ) i2 < 1< I G„r(/3) - G „ ( t S ) | (56) Lemma 6 implies that the right-hand side of (56) converges uniformly to zero as T —» oo, so, given e > 0,3 T ( e ) such that V T > T(e), sup /JeB* | G„,T { P ) ~ G M | < £ (57) Equations (56) and (57) imply that the first term on the right-hand side of (55) is less than e for all T > T(e), which proves that term converges to zero as T —» oo. We now complete the proof. By an argument analogous to that used in the proof of Lemma 5, | ^(/^ £«(/?)) — 0(Gv(/?),xt(/?)) |2 satisfies a uniform strong law of large numbers. This implies that the second term on the righthand side of (55) converges to zero as T —► oo. Thus, the left-hand side of 47 (55) converges to zero, so the minimand in (18) converges uniformly to the minimand of (54). Since the minimization is over the compact set B „, this implies the conclusion of the lemma. Q E D . L em m a 8 L e t lim 1/—+QO b e a s e q u e n c e s u c h t h a t /?„ G B „,V j/. T h e n 0 £ ( x r ( h ) J v ) - tf(c M & ),* r(& )) (58) Proof. According to Lemma 4, { r p ( ( 3 , - ) , / 3 G B } is an equicontinuous family of functions. Assumptions 1(a) and 1(b) then imply that {£(•,/?),/? € B } is also equicontinuous. Furthermore, since <j> is a bounded function, it follows that {£(•,/?), (5 € 5 } is uniformly bounded and Arzela’s theorem (Kolmogorov and Fomin (1970, p. 102)) then implies that there exists a subsequence Vk and a continuous function £ such that lim sup | £ (x ,/?„ ) — £(x) | = 0. (59) k->°° xex The function £ is the uniform limit of a sequence of uniformly Lipschitz functions, so £ satisfies a Lipschitz condition. By a proof analogous to Lemma 1(b), there exists a sequence with G D„k such that lim sup |£(x) — VK/^n*) I = 0. (60) iex (Notice that this sequence is different from \ P u k ] in the statement of the lemma) Equation (58) is now demonstrated along subsequence i/*. The definition of G implies that ^ A j-a rg n u n |£ ( * r ( A J » f l J “ | » (61) since the expectational error is orthogonal to the least squares predictor. The following chain of inequalities hold: 48 II £ ( x r ( l k) M - r P ( G Vk( i k) , x r 0 , k)) ii $ II £ ( * r ( « , - H ^ x r i ^ m < supieA- I £ ( x , j 3 y k ) - I < supx6X I £ ( x , j 3 „ k ) - S ( x ) | + supl€JC I S ( x ) - I p 0 „ k , x ) I - * 0 (62) where the first inequality follows from (61) and the third inequality follows from the triangle inequality. The first term on the right-hand side of (62) converges to zero as k —► oo by (59). The second term on the right-hand side of (62) converges to zero as k —» oo by (60). Since (62) holds for any arbitrary convergent subsequence, this result implies the conclusion of the lemma. Q E D . The next and last lemma shows that /?„ gets arbitrarily close to (?„(/?„) as v grows. In this sense, we say that an asymptotic fixed point exists. L em m a 9 Km | min || V>(/?, * W ) ) “ ^ ( G v(fi)t x?(0)) || | = 0 Proof. According to Lemma 1(b), a sequence {/?„}£!j exists such that lim sup | £ ( x ) - ^ ( ^ . x ) | = 0 ‘/-,0° xex (63) Recall that £ is the true conditional expectation. According to Lemma 2, /?„ € B„, so the minimand in the statement of the lemma is dominated by (64) and it is enough to show that (64) goes to zero. Now, < | | w , . * r ( « ) - f ( * r ( 0 | | + [5 (* r(A .)) - £ ( * r ( A .) ; /u | + | W ( « ; A . ) - * ( G .(/U * r(/U )| (65) 49 The first term in the right side goes to zero as The second term is bounded by | / « » ( * .» ') ) <U VI«) - / v —►oo by the choice of A,)) ^ M « ) where F(*|u) is the conditional distribution of given u t = u. The above expression converges to zero by Lemma 3 (d) and the Lipschitz-continuity of <p. Finally, the third term goes to zero by Lemma 8. QED. P ro o f of P ro p o sitio n 1 In this proof, convergence of a function of (z, x ) in the supnorm is taken over the ergodic set of the true process Z U C Z x U. Let xpp = xp((3, •) and Hp = //(•, •; /?). We first show that xppv converges to £ pointwise. Since P is a uniformly bounded, equicontinuous class of functions, it is a compact set in the sup norm. Therefore, there exists a subsequence (indexed by k ) and a limiting function xp" such that xppVk —* xp’ in the supnorm. It is enough if we show that, for any such convergent subsequence, xp* = £ a.e. in YU. Let H * be the law of motion consistent with xp’, and £ ’ the true condi tional expectation consistent with H ’ . We have Hp —*■ H* uniformly, by an argument analogous to that used in the proof of Lemma 3(d). Note that dF{u'\u). (66) All the functions inside the integral are continuous and bounded, so Lebesgue Dominated Convergence implies (/(*> as J v $ (H * ( # “(*« u)>u')] d F { u ' \ u ) = £ ’ { f { z , u)) (67) k —* oo. We will eventually show that H = H ’ \ to this end, we first have to prove that H ’ satisfies assumption 4* (if B„ = B f) or assumption 4 (if B„ = B * u c ). In the former case, H p Vk € S,Vfc, by the definition of B f. Since S is closed under the sup norm, and H p v H* uniformly, it follows that H’ €S. 50 We now turn to the latter case. It is sufficient to show that, if B „ = B ^ 1707, H * has a Lipschitz coefficient satisfying (45). Let = p*{u) sup H * (z,u ) z,z£Z - z H -{z,u ) — z Fix t > 0, and any u € U . By definition of p*(u), there exist z and z such that ,■(«) < m - H "(z, u) * 'u? ~ ^ H Pl/k( z , u ) | + \ H 0 „k ( z , u ) - H/}„k ( z , u ) \ + Iz "H ‘ { z , u ) - H p„k ( z , u )| Iz ’u)l(i + t ) - zI \z + —z \ H p „k ( z , u ) - H m( z , u ) \ (1+0 —z I **»(*■“)-«•(*- “)l' 1 | + rt i ( it 1 PPvk \U) (1+c) (68) where p p „ k (u ) is the function defined in (31) Now, fix z, z; take logs on both sides of (68). Since inequality (68) holds for all we can take the liminf* on both sides and use the uniform convergence of H p Uk —> H * to obtain log p ' ( u ) < (lirninf log ftj.,(u )j + lo g(I + e) (69) which holds for all e > 0 and all u . Now, given any e, taking expectations over u we get B [log />*(“ )] < E [lhn inf log + log(l + e) < lirninf [lSlog p/j^(u)] + log(l + e) < a + log(l + c) where the second inequality follows from Fatou’s lemma, and the last in equality follows from the fact that (3Uk € B „ . Since the above inequality holds for all e > 0, we conclude E log p * ( u ) 51 < a This completes the demonstration that H * satisfies (45) when B „ = B £ u c . We have shown that H * satisfies assumption 4* or 4 (depending on the definition of B„ being used). It follows that a stationary and ergodic process { z f 00} ejcists satisfying z ^ ° ° = so the L 2 norm of {-2*°°} is well defined. We can write, ||V >-(*n - £ '(* ;~ )|| < + I W A ..,* r to J ) - <HGV.(ft,)>*r(/U)ll + aj i i We can now apply Lemma 3, parts (a), (b), and (e), to conclude Ixr-zfW I-^ O a.s. (71) The limsup of the right side of (70) equals zero, as follows: The limsup of the first term is equal to zero by the definition of V>*; the second term goes to zero by (71), along with the continuity of and Lebesgue Dominated Convergence; the third term goes to zero by Lemma 9; the fourth term goes to zero by Lemma 8, the fifth term goes to zero by (71), along with the continuity of £ ( - , / 3 „ k ) , and Lebesgue Dominated Convergence; the sixth term goes to zero by (67). We have proved that the left side of (70) is equal to zero, which implies V>*(x*°°) = £ m( X f ° ° ) almost surely in the support of x*°°. We can summarize this derivation as follows: - r ( x ) = E (^ (*;~ ) I = x) (72) This is an important result in itself; it says that the limit of the pa rameterized expectations V’* >s equal to the true expectation of the process generated by ij>m. Finally, since 9 ( ' K 0 » k i x t ( P ' * ) ) * z t ( 0 * k ) i Z t - i ( P » k ) , U t ) = 0, VA: 52 by construction, and g is continuous, (72) and lemma 3(e) imply s ( £ ( ^ ; +i) I * ;) , = °, with probability one, which implies that H * is a rational expectations equi librium or, equivalently, that £ * = € and H = H * . To summarize, we have shown that any convergent subsequence of H p v goes to H in the supnorm, which implies that the same holds for the whole sequence. To complete the proof, we have to show the same type of convergence for the sample version of the approximate solution. It is enough to show Lim T —* oo i sup 6A' ^ (x , P v,t ) — V’(~, P v ) \ = 0. (73) Lemma 7 and continuity of xpp in /? imply pointwise. Since P is equicontinuous, by Lemma 4, and bounded it is compact in the sup norm, which implies uniform convergence as in (73). Finally, (72), (73), and Lemma 3(d) imply that H p T converges uniformly to //*. This completes the proof of Proposition 1. P ro o f of C orollary a) Follows from lemma 3(e) and the uniform convergence of H p v T —* H as in Proposition 1. b) Lemmas 8 and 9 imply || f ( x f ° ( ^ ir),i9„,r) || —> 0. Part b) of the corollary follows from part a) and continuity of g . c) Follows from proposition 1 and part (b) of the definition of S (if as sumption 4* is imposed and B„ = B^) or from the A U C condition, Proposition 1 and Lemma 4 in Duffie and Singleton (if assumption 4 is imposed and B „ = T 5 * u c ). P ro o f of P ro p o sitio n 2 We give the proof for the case where B„ = B * u c . The proof for the case where B^ = B f is analogous, and is therefore omitted. • We first prove that if min E \ m * W ) ) - * « ? ,( « , * r ( « ) ] 2 - “ C-Oi/ P 53 (’ 4) then assumption 4 is satisfied. Consider any convergent subsequence of {Vvj„r }; as in the proof of proposition 1, this subsequence has a limit xf>* for which the corresponding law of motion H * satisfies the AUC condition (30). As in the proof of proposition 1, (70) holds for this V’* and the corresponding conditional expectation S* (defined as in (67)) . Now, consider the right side of (70); with the exception of the third term, all terms go to zero as k —►oo if Assumptions 1, 2 and 3 are satisfied19. The third term goes to zero because of (74). This proves the existence of an equilibrium H mthat satisfies AUC and that Assumption 4 is satisfied. To prove part (x) assume, towards a contradiction, that liminf p £ b v E [ r p ( 0 , x f = { p ) ) - ^ ( ^ ( Z ? ) ,! ^ ^ ) ) ] 2 = 0. Then there would be a subsequence {/?„*} for which the min goes to zero and we can find a sub-subsequence {/?„*.} such that Hp„ —* H*. By the argument given above, this would imply existence of H* that satisfies Assumption 4, which is a contradiction. • We now prove part (b). Let Q = { laws of motion for z that satisfy AUC condition}. By an argument analogous to that used in proving that H * of Propo sition 1 satisfies AUC, we can prove that Q is a closed set. Since each HpvT € Q and H Q , part (b) follows immediately. QED 19Notice that assumption 4 has only been used in results leading to Lemma 9, which causes the third term in (70) to go to zero. 54 APPENDIX 2 NON-LINEAR LEAST SQUARES LEARNING The recursive least squares estimator to predict <f>(zt+1 ) can be written as A = A -i + <* R T l W*<) - iH A ,i.-i)l) R , = R t-1 + a , *♦<*;*-■> (75) Notice that if we fixed /?, replaced 5* by z*(/?), and for the particular choice a* = (1/0) A ls the non-linear least squares estimator to predict <f>(zt (l3)) with V>(*)£*-i(/?)), where the series z are not affected by /?<, except that the residuals are calculated recursively. In the system defined by (75), however, the V s depend on the estimate /3t . In words, (75) describes an economy where agents use recursive non linear least squares to form their expectations about the future, and where equilibrium is generated by such expectations; the first equation says that today’s beliefs are updated according with last period’s beliefs and the pre diction error made this period, the second equation gives the gradient for how the prediction error affects the beliefs /?*, and the last equation defines the equilibrium z t in terms of the expectations of the agents. This is a non linear version of economies studied Marcet and Sargent (1989), where the non-linearities appear both in the equation generating the observations and in the expectational rules. Using notation DMP) = |W , * t - i ( P ) ) / W \ 0=1; and R(0) = E{ D M 0 ) W P ) ' ) and adapting results from Ljung (1975], it is possible to show that the only learning can converge with positive probability are 0 ' s where least squares 0 <= D v that satisfy R 0 )-'E (D A 0 ) ( ? ) ) ] ) = 0. (76) This equation is obtained by taking the object multiplying a t in the recursive algorithm, fixing 0 — 0 , and taking the expectation with respect to the 55 stationary distribution of { z t (/3)}. Now, since the first order conditions for the maximization problem of (16) imply that ; E (D M P ) W t m - r p ( G M , x t. ,(/?))]) = 0. it is clear that (76) is satisfied, precisely, at the fixed point /?„. Further results in Ljung [1975] guarantee that j3t —* /?„ if the differential equation W (r),x ,_ ,(/? (r)» )) ? M . = R(0( t ) )-'E { D M 0 ( r ) ) (77) is stable. To analyze this differential equation, notice that the right side of (77) is equal to W > * .-iW )l) R (0T 'e ( D M 0 ) by the definition of G„. By the mean value theorem, in a neighborhood of /?„ this expression can be written as m u r 'E ( d m 0) D M 0 ) ' (o ,( 0 ) - 0)) for some /3 in this neighborhood. Therefore, /?t —►/?„ locally if and only if the differential equation = R(0( t ) ) - 'E ( D M 0 ( r ) ) D M 0 (t )Y) (G,(0( t ) ) - 0 ( t )) = G„(0( t ) ) - 0 ( t ) is locally stable at /?„. 56 Working Paper Series A series o f research studies on regional economic issues relating to the Seventh Federal Reserve District, and on financial and econom ic topics. REGIONAL ECONOMIC ISSUES Estimating Monthly Regional Value Added by Combining Regional Input With National Production Data WP-92-8 P h ilip R. Is r a ile v ic h a n d K e n n e th N . K u ttn e r Local Impact o f Foreign Trade Zone WP-92-9 D a v id D . W e iss Trends and Prospects for Rural Manufacturing WP-92-12 W illia m A . T e sta State and Local Government Spending--The Balance Between Investment and Consumption WP-92-14 R ic h a r d H . M a tto o n Forecasting with Regional Input-Output Tables WP-92-20 P .R . Is r a ile v ic h , R. M a h id h a r a , a n d G .J .D . H e w in g s A Primer on Global Auto Markets WP-93-1 P a u l D . B a lle w a n d R o b e r t H. S c h n o r b u s Industry Approaches to Environmental Policy in the Great Lakes Region WP-93-8 D a v id R. A lla r d ice, R ic h a r d H. M a tto o n a n d W illia m A . T e sta The Midwest Stock Price Index—Leading Indicator o f Regional Economic Activity WP-93-9 W illia m A . S tr a u s s Lean Manufacturing and the Decision to Vertically Integrate Some Empirical Evidence From the U.S. Automobile Industry WP-94-1 T h o m a s H . K lie r Domestic Consumption Patterns and the Midwest Economy WP-94-4 R o b e r t S c h n o r b u s a n d P a u l B a lle w 1 Workingpaperseriescontinued To Trade or Not to Trade: Who Participates in RECLAIM? WP-94-11 T h o m a s H . K lie r a n d R ic h a r d M a tto o n Restructuring & Worker Displacement in the Midwest WP-94-18 P a u l D . B a lle w a n d R o b e r t H . S c h n o r b u s ISSUES IN FINANCIAL REGULATION Incentive Conflict in Deposit-Institution Regulation: Evidence from Australia WP-92-5 E d w a r d J. K a n e a n d G e o r g e G. K a u fm a n Capital Adequacy and the Growth o f U.S. Banks WP-92-11 H e r b e r t B a e r a n d J o h n M c E lr a v e y Bank Contagion: Theory and Evidence WP-92-13 G e o r g e G. K a u fm a n Trading Activity, Progarm Trading and the Volatility o f Stock Returns WP-92-16 J a m e s T. M o s e r Preferred Sources o f Market Discipline: Depositors vs. Subordinated Debt Holders WP-92-21 D o u g la s D . E v a n o f f An Investigation o f Returns Conditional on Trading Performance WP-92-24 J a n ie s T. M o s e r a n d J a c k y C. S o The Effect o f Capital on Portfolio Risk at Life Insurance Companies WP-92-29 E lija h B r e w e r III, T h o m a s H . M o n d s c h e a n , a n d P h ilip E. S tr a h a n A Framework for Estimating the Value and Interest Rate Risk o f Retail Bank Deposits WP-92-30 D a v id E. H u tc h is o n , G e o r g e G. P e n n a c c h i Capital Shocks and Bank Growth-1973 to 1991 WP-92-31 H e r b e r t L. B a e r a n d J o h n N . M c E lr a v e y The Impact o f S&L Failures and Regulatory Changes on the CD Market 1987-1991 WP-92-33 E lija h B r e w e r a n d T h o m a s H . M o n d s c h e a n 2 Workingpaperseriescontinued Junk Bond Holdings, Premium Tax Offsets, and Risk Exposure at Life Insurance Companies WP-93-3 E lija h B r e w e r I I I a n d T h o m a s H . M o n d s c h e a n Stock Margins and the Conditional Probability o f Price Reversals WP-93-5 P a u l K o fm a n a n d J a m e s T. M o s e r Is There Lif(f)e After DTB? Competitive Aspects o f Cross Listed Futures Contracts on Synchronous Markets WP-93-11 P a u l K o fm a n , T o n y B o u w m a n a n d J a m e s T. M o s e r Opportunity Cost and Prudentiality: A RepresentativeAgent Model o f Futures Clearinghouse Behavior WP-93-18 H e r b e r t L. B a e r, V ir g in ia G. F r a n c e a n d J a m e s T. M o s e r The Ownership Structure o f Japanese Financial Institutions WP-93-19 H esn a G enay Origins o f the Modern Exchange Clearinghouse: A History o f Early Clearing and Settlement Methods at Futures Exchanges WP-94-3 J a m e s T. M o s e r The Effect o f Bank-Held Derivatives on Credit Accessibility WP-94-5 E lija h B r e w e r III, B e r n a d e tte A . M in to n a n d J a m e s T. M o s e r Small Business Investment Companies: Financial Characteristics and Investments WP-94-10 E lija h B r e w e r III a n d H e s n a G e n a y MACROECONOMIC ISSUES An Examination o f Change in Energy Dependence and Efficiency in the Six Largest Energy Using Countries--1970-1988 WP-92-2 J a c k L. H e r v e y Does the Federal Reserve Affect Asset Prices? WP-92-3 V efa T a rh a n Investment and Market Imperfections in the U.S. Manufacturing Sector WP-92-4 P a u la R. W o r th in g to n 3 Workingpaperseriescontinued Business Cycle Durations and Postwar Stabilization o f the U.S. Economy WP-92-6 M a r k W. W a ts o n A Procedure for Predicting Recessions with Leading Indicators: Econometric Issues and Recent Performance WP-92-7 J a m e s H . S to c k a n d M a r k W. W a tso n Production and Inventory Control at the General Motors Corporation During the 1920s and 1930s WP-92-10 A n i l K. K a s h y a p a n d D a v id W. W ilc o x Liquidity Effects, Monetary Policy and the Business Cycle WP-92-15 L a w r e n c e J. C h r is tia n o a n d M a r tin E ic h e n b a u m Monetary Policy and External Finance: Interpreting the Behavior o f Financial Flows and Interest Rate Spreads WP-92-17 K e n n e th N. K u ttn e r Testing Long Run Neutrality WP-92-18 R o b e r t G. K in g a n d M a r k W. W a tso n A Policymaker's Guide to Indicators o f Economic Activity WP-92-19 C h a r le s E v a n s , S te v e n S tr o n g in , a n d F r a n c e s c a E u g e n i Barriers to Trade and Union Wage Dynamics WP-92-22 E lle n R. R is s m a n Wage Growth and Sectoral Shifts: Phillips Curve Redux WP-92-23 E lle n R. R is s m a n Excess Volatility and The Smoothing o f Interest Rates: An Application Using Money Announcements WP-92-25 S te v e n S tr o n g in Market Structure, Technology and the Cyclicality o f Output WP-92-26 B r u c e P e te r s e n a n d S te v e n S tr o n g in The Identification o f Monetary Policy Disturbances: Explaining the Liquidity Puzzle WP-92-27 S te v e n S tr o n g in 4 Workingpaperseriescontinued Earnings Losses and Displaced Workers WP-92-28 L o u is S. J a c o b s o n , R o b e r t J. L a L o n d e , a n d D a n ie l G. S u lliv a n Some Empirical Evidence o f the Effects on Monetary Policy Shocks on Exchange Rates WP-92-32 M a r tin E ic h e n b a u m a n d C h a r le s E v a n s An Unobserved-Components Model of Constant-Inflation Potential Output WP-93-2 K e n n e th N . K u ttn e r Investment, Cash Flow, and Sunk Costs WP-93-4 P a u la R. W o r th in g to n Lessons from the Japanese Main Bank System for Financial System Reform in Poland WP-93-6 T a k e o H o sh i, A n il K a s h y a p , a n d G a r y L o v e m a n Credit Conditions and the Cyclical Behavior o f Inventories WP-93-7 A n il K. K a s h y a p , O w e n A . L a n io n t a n d J e r e m y C. S te in Labor Productivity During the Great Depression WP-93-10 M ic h a e l D. B o r d o a n d C h a r le s L. E v a n s Monetary' Policy Shocks and Productivity Measures in the G-7 Countries WP-93-12 C h a r le s L. E v a n s a n d F e r n a n d o S a n to s Consumer Confidence and Economic Fluctuations WP-93-13 J o h n G. M a ts u s a k a a n d A r g ia M . S b o r d o n e Vector Autoregressions and Cointegration WP-93-14 M a r k W. W a tso n Testing for Cointegration When Some o f the Cointegrating Vectors Are Known WP-93-15 M ic h a e l T. K. H o r v a th a n d M a r k W. W a tso n Technical Change, Diffusion, and Productivity WP-93-16 J e ffr e y R. C a m p b e ll 5 Workingpaperseriescontinued Economic Activity and the Short-Term Credit Markets: An Analysis o f Prices and Quantities WP-93-17 B e n ja m in M . F r ie d m a n a n d K e n n e th N . K u ttn e r Cyclical Productivity in a Model o f Labor Hoarding WP-93-20 A r g ia M . S b o r d o n e The Effects o f Monetary Policy Shocks: Evidence from the Flow o f Funds WP-94-2 L a w r e n c e J. C h ris tia n o , M a r tin E ic h e n b a u m a n d C h a r le s E v a n s Algorithms for Solving Dynamic M odels with Occasionally Binding Constraints WP-94-6 L a w r e n c e J . C h r is tia n o a n d J o n a s D .M . F is h e r Identification and the Effects o f Monetary Policy Shocks WP-94-7 L a w r e n c e J . C h r is tia n o , M a r tin E ic h e n b a u m a n d C h a r le s L. E v a n s Small Sample Bias in GMM Estimation o f Covariance Structures WP-94-8 J o s e p h G. A lto n ji a n d L e w is M . S e g a l Interpreting the Procyclical Productivity o f Manufacturing Sectors: External Effects o f Labor Hoarding? WP-94-9 A r g ia M . S b o r d o n e Evidence on Structural Instability in Macroeconomic Time Series Relations WP-94-13 J a m e s H. S to c k a n d M a r k W. W a tso n The Post-War U.S. Phillips Curve: A Revisionist Econometric History WP-94-14 R o b e r t G. K in g a n d M a r k W. W a tso n The Post-War U.S. Phillips Curve: A Comment WP-94-15 C h a r le s L. E v a n s Identification o f Inflation-Unemployment WP-94-16 B e n n e tt T. M c C a llu m The Post-War U.S. Phillips Curve: A Revisionist Econometric History Response to Evans and McCallum WP-94-17 R o b e r t G. K in g a n d M a r k W. W a tso n 6 Workingpaperseriescontinued Estimating Deterministic Trends in the Presence o f Serially Correlated Errors WP-94-19 E u g e n e C a n je ls a n d M a r k W. W a tso n Solving Nonlinear Rational Expectations Models by Parameterized Expectations: Convergence to Stationary Solutions WP-94-20 A lb e r t M a r c e t a n d D a v id A . M a r s h a ll 1