Bayesian Inference for DSGE Models. Lawrence J. Christiano

Transcription

1 Bayesian Inference for DSGE Models Lawrence J. Christiano

2 Outline State space-observer form. convenient for model estimation and many other things. Bayesian inference Bayes rule. Monte Carlo integation. MCMC algorithm. Laplace approximation

3 State Space/Observer Form Compact summary of the model, and of the mapping between the model and data used in the analysis. Typically, data are available in log form. So, the following is useful: If x is steady state of x t : ˆx t x t x, x =) x t x = 1 + ˆx t xt =) log = log (1 + ˆx t ) ˆx t x Suppose we have a model solution in hand: 1 z t = Az t1 + Bs t s t = Ps t1 + e t, Ee t et = D. 1 Notation taken from solution lecture notes, Korea_212/lecture_on_solving_rev.pdf

4 State Space/Observer Form Suppose we have a model in which the date t endogenous variables are capital, K t+1, and labor, N t : z t = ˆK t+1 ˆN t, s t = ˆ# t, e t = e t. Data may include variables in z t and/or other variables. for example, suppose available data are N t and GDP, y t and production function in model is: so that y t = # t K a t N 1a t, ŷ t = ˆ# t + a ˆK t +(1 a) ˆN t = ( 1 a ) z t + ( a ) z t1 + s t From the properties of ŷ t and ˆN t : Yt data log yt log y = log N = t log N + ŷt ˆN t

5 State Space/Observer Form Model prediction for data: Yt data log y = log N + h log y = log N + 1 a 1 ŷt ˆN t i h a z t + i h i 1 z t1 + s t x t = z t z t1 ˆ# t = a + Hx! t, a = log y log N h 1 a a 1, H = 1 i The Observer Equation may include measurement error, w t : Y data t = a + Hx t + w t, Ew t w t = R. Semantics: x t is the state of the system (do not confuse with the economic state (K t, # t )!).

6 State Space/Observer Form Law of motion of the state, x t (state-space equation): x t = Fx t1 + u t, Eu t u t = Q z t+1 z t s t+1! = u t = B I! e t, Q = " A # BP I P " BDB BD DB D # z t z t1 s t, F =! + B I! " A BP I P e t+1, #.

7 Uses of State Space/Observer Form Estimation of q and forecasting x t and Yt data Can take into account situations in which data represent a mixture of quarterly, monthly, daily observations. Data Rich estimation. Could include several data measures (e.g., employment based on surveys of establishments and surveys of households) on a single model concept. Useful for solving the following forecasting problems: Filtering (mainly of technical interest in computing likelihood function): Smoothing: h P x t Y data t1, Ydata t2 h P x t Y data T,..., Ydata,..., Ydata 1 1 i, t = 1, 2,..., T. i, t = 1, 2,..., T. Example: real rate of interest and output gap can be recovered from x t using simple New Keynesian model. Useful for deriving a model s implications vector autoregressions

8 Mixed Monthly/Quarterly Observations Di erent data arrive at di erent frequencies: daily, monthly, quarterly, etc. This feature can be easily handled in state space-observer system. Example: suppose inflation and hours are monthly, t =, 1/3, 2/3, 1, 4/3, 5/3, 2,... suppose gdp is quarterly, t =, 1, 2, 3,... GDP 1 t monthly inflation t monthly inflation Yt data t1/3 = monthly inflation B hours t C A hours t1/3 hours t2/3, t =, 1, 2,.... that is, we can think of our data set as actually being quarterly, with quarterly observations on the first month s inflation, quarterly observations on the second month s inflation, etc.

9 Mixed Monthly/Quarterly Observations Problem: find state-space observer system in which observed data are: Y data t = Solution: easy! GDP t monthly inflation t monthly inflation t1/3 monthly inflation t2/3 hours t hours t1/3 hours t2/3 1, t =, 1, 2,.... C A

10 Mixed Monthly/Quarterly Observations Model timing: t =, 1/3, 2/3,... z t = Az t1/3 + Bs t, s t = Ps t1/3 + e t, Ee t e t = D. Monthly state-space observer system, t =, 1/3, 2/3,... x t = Fx t1/3 + u t, Eu t u t = Q, u t iid t =, 1/3, 2/3,... Y t = Hx t, Y t = y t p t h t!. Note: first order vector autoregressive representation for quarterly state z } { x t = F 3 x t1 + u t + Fu t1/3 + F 2 u t2/3, u t + Fu t1/3 + F 2 u t2/3 ~iid for t =, 1, 2,...!!

11 Mixed Monthly/Quarterly Observations Consider 1the following 2 system: 3 x x t 1 3 A = 4 F3 x t1 F 2 x t 4 3 x t 2 3 F x t 5 3 Define x t so that x t x t 1 3 x t A, F = 2 4 F3 F 2 F 3 1 A + 5, ũ t = " I F F 2 I F I " I F F 2 I F I x t = F x t1 + ũ t, ũ t iid in quarterly data, t =, 1, 2,... u t u t 1 3 u t 2 3 u t u t 1 3 u t A. 1 A, Eũ t ũ t = Q = " I F F 2 I F I #" D D D #" I F F 2 I F I #

12 Mixed Monthly/Quarterly Observations Conclude: state space-observer system for mixed monthly/quarterly data, for t =, 1, 2,... x t = F x t1 + ũ t, ũ t iid, Eũ t ũ t = Q, Y data t = H x t + w t, w t iid, Ew t w t = R. Here, H selects elements of x t needed to construct Y data t can easily handle distinction between whether quarterly data represent monthly averages (as in flow variables), or point-in-time observations on one month in the quarter (as in stock variables). Can use Kalman filter to forecast ( nowcast ) current quarter data based on first month s (day s, week s) observations.

13 Connection Between DSGE s and VAR s Fernandez-Villaverde, Rubio-Ramirez, Sargent, Watson Result Vector Autoregression where u t is iid. Y t = B 1 Y t1 + B 2 Y t u t, Matching impulse response functions strategy for building DSGE models fits VARs and assumes u t are a rotation of economic shocks (for details, see later notes). Can use the state space, observer representation to assess this assumption from the perspective of a DSGE.

14 Connection Between DSGE s and VAR s System (ignoring constant terms and measurement error): ( State equation ) x t = Fx t1 + De t, D = B!, I ( Observer equation ) Y t = Hx t. Substituting: Y t = HFx t1 + HDe t Suppose HD is square and invertible. Then e t = (HD) 1 Y t (HD) 1 HFx t1 () Substitute latter into the state equation: x t = Fx t1 + D (HD) 1 Y t D (HD) 1 HFx t1 = h i I D (HD) 1 H Fx t1 + D (HD) 1 Y t.

15 Connection Between DSGE s and VAR s We have: x t = Mx t1 + D (HD) 1 Y t, M = If eigenvalues of M are less than unity, h i I D (HD) 1 H F. x t = D (HD) 1 Y t + MD (HD) 1 Y t1 + M 2 D (HD) 1 Y t Substituting into () e t = (HD) 1 Y t (HD) 1 HF h i D (HD) 1 Y t1 + MD (HD) 1 Y t2 + M 2 D (HD) 1 Y t or, where Y t = B 1 Y t1 + B 2 Y t u t, u t = HDe t, B j = HFM j1 D (HD) 1, j = 1, 2,... The latter is the VAR representation.

16 Connection Between DSGE s and VAR s The VAR repersentation is: Y t = B 1 Y t1 + B 2 Y t u t, where u t = HDe t, B j = HFM j1 D (HD) 1, j = 1, 2,... Notes: e t is invertible because it lies in space of current and past Y t s. VAR is infinite-ordered. assumed system is square (same number of elements in e t and Y t ). Sims-Zha (Macroeconomic Dynamics) show how to recover e t from current and past Y t when the dimension of e t is greater than the dimension of Y t.

17 Quick Review of Probability Theory Two random variables, x 2 (x 1, x 2 ) and y 2 (y 1, y 2 ). Joint distribution: p (x, y) x 1 x 2 y 1 p 11 p 12 y 2 p 21 p 22 = x 1 x 2 y y where Restriction: p ij = probability x = x i, y = y j. Z x,y p (x, y) dxdy = 1.

18 Quick Review of Probability Theory Joint distribution: p (x, y) x 1 x 2 y 1 p 11 p 12 y 2 p 21 p 22 = x 1 x 2 y y Marginal distribution of x : p (x) Probabilities of various values of x without reference to the value of y: or, p (x) = n p11 + p 21 =.4 x = x 1 p 12 + p 22 =.6 x = x 2. Z p (x) = p (x, y) dy y

19 Quick Review of Probability Theory Joint distribution: p (x, y) x 1 x 2 y 1 p 11 p 12 y 2 p 21 p 22 = x 1 x 2 y y Conditional distribution of x given y : p (x y) Probability of x given that the value of y is known or, ( p (x1 y 1 ) p (x y 1 ) = p (x 2 y 1 ) p (x y) = p 11 p 11 +p 12 = p 11 p(y 1 ) =.5.45 =.11 p 12 p 11 +p 12 = p 12 p(y 1 ) =.4.45 =.89 p (x, y) p (y).

20 Quick Review of Probability Theory Joint distribution: p (x, y) Mode x 1 x 2 y p (y 1 ) =.45 y p (y 2 ) =.55 p (x 1 ) =.4 p (x 2 ) =.6 Mode of joint distribution (in the example): argmax x,y p (x, y) = (x 2, y 1 ) Mode of the marginal distribution: argmax x p (x) = x 2, argmax y p (y) = y 2 Note: mode of the marginal and of joint distribution conceptually di erent.

21 Maximum Likelihood Estimation State space-observer system: x t+1 = Fx t + u t+1, Eu t u t = Q, Y data t = a + Hx t + w t, Ew t w t = R Reduced form parameters, (F, Q, a, H, R), functions of q. Choose q to maximize likelihood, p Y data q : p Y data q = p Y1 data,..., YT data q = p Y1 data q p Y2 data Y1 data, q computed using Kalman Filter z } { p Yt data Yt1 data Ydata 1, q p,, Ydata 1, q YT data Ydata T1 Kalman filter straightforward (see, e.g., Hamilton s textbook).

22 Bayesian Inference Bayesian inference is about describing the mapping from prior beliefs about q, summarized in p (q), to new posterior beliefs in the light of observing the data, Y data. General property of probabilities: p Y data, q = which implies Bayes rule: p Y data q p (q) p q Y data p Y data, p q Y data = p Y data q p (q) p Y data, mapping from prior to posterior induced by Y data.

23 Bayesian Inference Report features of the posterior distribution, p q Y data. The value of q that maximizes p q Y data, mode of posterior distribution. Compare marginal prior, p (q i ), with marginal posterior of individual elements of q, g q i Y data : g q i Y data Z = p q Y data dq j6=i (multiple integration!!) q j6=i Probability intervals about the mode of q ( Bayesian confidence intervals ), need g q i Y data. Marginal likelihood for assessing model fit : p Y data Z = p Y data q p (q) dq (multiple integration) q

24 Monte Carlo Integration: Simple Example Much of Bayesian inference is about multiple integration. Numerical methods for multiple integration: Quadrature integration (example: approximating the integral as the sum of the areas of triangles beneath the integrand). Monte Carlo Integration: uses random number generator. Example of Monte Carlo Integration: suppose you want to evaluate Z b a f (x) dx, - a < b. select a density function, g (x) for x 2 [a, b] and note: Z b a f (x) dx = Z b a f (x) f (x) g (x) dx = E g (x) g (x), where E is the expectation operator, given g (x).

25 Monte Carlo Integration: Simple Example Previous result: can express an integral as an expectation relative to a (arbitrary, subject to obvious regularity conditions) density function. Use the law of large numbers (LLN) to approximate the expectation. step 1: draw x i independently from density, g, for i = 1,..., M. step 2: evaluate f (x i ) /g (x i ) and compute: Exercise. µ M 1 M M f (x i ) Â g (x i=1 i )! M! E f (x) g (x). Consider an integral where you have an analytic solution available, e.g., R 1 x2 dx. Evaluate the accuracy of the Monte Carlo method using various distributions on [, 1] like uniform or Beta.

26 Monte Carlo Integration: Simple Example Standard classical sampling theory applies. Independence of f (x i ) /g (x i ) over i implies:! M 1 f (x var M Â i ) = v M g (x i ) M, v M var f (xi ) g (x i ) i=1 ' 1 M M Â i=1 f (xi ) g (x i ) µ M 2. Central Limit Theorem Estimate of R b f a (x) dx is a realization from a Nomal distribution with mean estimated by µ M and variance, v M /M. With 95% probability, r Z r vm b µ M 1.96 M vm f (x) dx µ M M Pick g to minimize variance in f (x i ) /g (x i ) and M to minimize (subject to computing cost) v M /M. a

27 Markov Chain, Monte Carlo (MCMC) Algorithms Among the top 1 algorithms "with the greatest influence on the development and practice of science and engineering in the 2th century". Reference: January/February 2 issue of Computing in Science & Engineering, a joint publication of the American Institute of Physics and the IEEE Computer Society. Developed in 1946 by John von Neumann, Stan Ulam, and Nick Metropolis (see

28 MCMC Algorithm: Overview compute a sequence, q (1), q (2),..., q (M), of values of the N 1 vector of model parameters in such a way that h i lim Frequency q (i) close to q = p q Y data. M! Use q (1), q (2),..., q (M) to obtain an approximation for Eq, Var (q) under posterior distribution, p q Y data g q i Y data = R p q Y data dqdq q i6=j p Y data = R q p Y data q p (q) dq posterior distribution of any function of q, f (q) (e.g., impulse responses functions, second moments). MCMC also useful for computing posterior mode, arg max q p q Y data.

29 MCMC Algorithm: setting up Let G (q) denote the log of the posterior distribution (excluding an additive constant): G (q) = log p Y data q + log p (q) ; Compute posterior mode: q = arg max G (q). q Compute the positive definite matrix, V : V 2 G (q) q q 1 q=q Later, we will see that V is a rough estimate of the variance-covariance matrix of q under the posterior distribution.

30 MCMC Algorithm: Metropolis-Hastings q (1) = q to compute q (r), for r > 1 step 1: select candidate q (r), x, draw {z} x from q (r1) + N1 step 2: compute scalar, l : jump distribution z } 1{ k {z}, VA, k is a scalar N1 p Y data x p (x) l = p Y data q (r1) p q (r1) step 3: compute q (r) : q (r) = q (r1) if u > l x if u < l, u is a realization from uniform [, 1]

31 Practical issues What is a sensible value for k? set k so that you accept (i.e., q (r) = x) in step 3 of MCMC algorithm are roughly 23 percent of time What value of M should you set? want convergence, in the sense that if M is increased further, the econometric results do not change substantially in practice, M = 1, (a small value) up to M = 1,,. large M is time-consuming. could use Laplace approximation (after checking its accuracy) in initial phases of research project. more on Laplace below. Burn-in: in practice, some initial q (i) s are discarded to minimize the impact of initial conditions on the results. Multiple chains: may promote e ciency. increase independence among q (i) s. can do MCMC utilizing parallel computing (Dynare can do this).

32 MCMC Algorithm: Why Does it Work? Proposition that MCMC works may be surprising. Whether or not it works does not depend on the details, i.e., precisely how you choose the jump distribution (of course, you had better use k > and V positive definite). Proof: see, e.g., Robert, C. P. (21), The Bayesian Choice, Second Edition, New York: Springer-Verlag. The details may matter by improving the e ciency of the MCMC algorithm, i.e., by influencing what value of M you need. Some Intuition the sequence, q (1), q (2),..., q (M), is relatively heavily populated by q s that have high probability and relatively lightly populated by low probability q s. Additional intuition can be obtained by positing a simple scalar distribution and using MATLAB to verify that MCMC approximates it well (see, e.g., question 2 in assignment 9).

33 MCMC Algorithm: using the Results To approximate marginal posterior distribution, g q i Y data, of q i, compute and display the histogram of q (1) i, q (2) i,..., q (M) i, i = 1,..., M. Other objects of interest: mean and variance of posterior distribution q : Eq ' q 1 M M Â q (j), Var (q) ' 1 M h ih M Â q (j) q q (j) qi. j=1 j=1

34 MCMC Algorithm: using the Results More complicated objects of interest: impulse response functions, model second moments, forecasts, Kalman smoothed estimates of real rate, natural rate, etc. All these things can be represented as non-linear functions of the model parameters, i.e., f (q). can approximate the distribution of f (q) using Var (f (q)) ' 1 M f q (1),..., f q (M)! Ef (q) ' f 1 M M h  f q (i) f i=1 M  f q (i), i=1 ih f q (i) f i

35 MCMC: Remaining Issues In addition to the first and second moments already discused, would also like to have the marginal likelihood of the data. Marginal likelihood is a Bayesian measure of model fit.

36 MCMC Algorithm: the Marginal Likelihood Consider the following sample average: M 1 h q (j) M Â j=1 p Y data q (j) p q (j), where h (q) is an arbitrary density function over the N dimensional variable, q. By the law of large numbers, M 1 h q (j) M Â j=1 p Y data q (j) p q (j)! E M!! h (q) p Y data q p (q)

37 MCMC Algorithm: the Marginal Likelihood 1 M Z = M Â j=1 q h q (j) p Y data q (j) p h (q) p Y data q p (q) q (j)! M! E! h (q) p Y data q p (q)! p Y data q p (q) p Y data dq = 1 p Y data. When h (q) = p (q), harmonic mean estimator of the marginal likelihood. Ideally, want an h such that the variance of h q (j) p Y data q (j) p q (j) is small (recall the earlier discussion of Monte Carlo integration). More on this below.

38 Laplace Approximation to Posterior Distribution In practice, MCMC algorithm very time intensive. Laplace approximation is easy to compute and in many cases it provides a quick and dirty approximation that is quite good. Let q 2 R N denote the Ndimensional vector of parameters and, as before, G (q) log p Y data q p (q) p Y data q ~likelihood of data p (q) ~prior on parameters q ~maximum of G (q) (i.e., mode)

39 Laplace Approximation Second order Taylor series expansion of G (q) log p Y data q p (q) about q = q : G (q) G (q ) + G q (q )(q q ) 1 2 (q q ) G qq (q )(q q ), where G qq (q ) = 2 log p Y data q p (q) q q Interior optimality of q implies: Then: ' p p q=q G q (q ) =, G qq (q ) positive definite p (q) Y data q Y data q p (q ) exp 1 2 (q q ) G qq (q )(q q ).

40 Laplace Approximation to Posterior Distribution Property of Normal distribution: Z 1 G q (2p) N qq (q ) 1 2 exp (q q ) G qq (q )(q q ) dq = 1 Then, Z p Y data q p (q) dq ' Z p Y data q p (q ) exp 1 2 (q q ) G qq (q )(q q ) d = p Y data q p (q ). 1 G qq (q ) 1 2 (2p) N 2

41 Conclude: Laplace Approximation p Y data ' p Y data q p (q ). 1 G qq (q ) 1 2 (2p) N 2 Laplace approximation to posterior distribution: p Y data q p (q) p Y data ' 1 (2p) N 2 exp G qq (q ) (q q ) G qq (q )(q q ) So, posterior of q i (i.e., g q i Y data )isapproximately h q i ~N qi, G qq (q ) 1i. ii

42 Modified Harmonic Mean Estimator of Marginal Likelihood Harmonic mean estimator of the marginal likelihood, p Y data : M h q (j) 31 M Â j=1 p Y data q (j) p q (j) 5, with h (q) set to p (q). In this case, the marginal likelihood is the harmonic mean of the likelihood, evaluated at the values of q generated by the MCMC algorithm. Problem: the variance of the object being averaged is likely to be high, requiring high M for accuracy. When h (q) is instead equated to Laplace approximation of posterior distribution, then h (q) is approximately proportional to p Y data q (j) p being averaged in the last expression is low. q (j) so that the variance of the variable

43 The Marginal Likelihood and Model Comparison Suppose we have two models, Model 1 and Model 2. compute p Y data Model 1 and p Y data Model 2 Suppose p Y data Model 1 > p Y data Model 2. Then, posterior odds on Model 1 higher than Model 2. Model 1 fits better than Model 2 Can use this to compare across two di erent models, or to evaluate contribution to fit of various model features: habit persistence, adjustment costs, etc. For an application of this and the other methods in these notes, see Smets and Wouters, AER 27.