Impulse Response Functions

Transcription

1 Impulse Response Functions Wouter J. Den Haan University of Amsterdam April 28, 2011

2 General definition IRFs The IRF gives the j th -period response when the system is shocked by a one-standard-deviation shock. Suppose y t = ρy t 1 + ε t and ε t has a variance equal to σ 2 Consider a sequence of shocks { ε t } t=1. Let the generated series for y t be given by {ȳ t } t=1. Consider an alternative series of shocks such that { εt + σ if t = τ ε t = ε t o.w. The IRF is then defined as IRF(j) = ỹ τ 1+j ȳ τ 1+j

3 IRFs for linear processes Linear processes: The IRF is independent of the particular draws for ε t Thus we can simply start at the steady state (that is when ε t has been zero for a very long time) IRF(j) = σρ j 1 Often you cannot get an analytical formula for the impulse response function, but simple iteration on the law of motion (driving process) gives you the exact same answer Note that the IRF is not stochastic

4 IRFs in theoretical models When you have solved for the policy functions then it is trivial to get the IRFs by simply giving the system a one standard deviation shock and iterating on the policy functions. Shocks in the model are structural shocks, such as productivity shock preference shock monetary policy shock

5 IRFs in empirical models What we are going to do? Describe an empirical model that has turned out to be very useful (for example for forecasting) Reduced form VAR Make clear we do not have to worry about variables being I(1) Describe a way to back out structural shocks (this is the hard part) Structural VAR

6 Reduced Form Vector AutoRegressive models (VARs) Let y t be an n 1 vector of n variables (typically in logs) where A j is an n n matrix. y t = J A j y t j + u t j=1 This system can be estimated by OLS (equation by equation) even if y t contains I(1) variables constants and trend terms are left out to simplify the notation

7 Spurious regression Let z t and x t be I(1) variables that have nothing to do with each other Consider the regression equation z t = ax t + u t The least-squares estimator is given by â T = T t=1 x tz t T t=1 x2 t Problem: lim T ât = 0

8 Source of spurious regressions The problem is not that z t and x t are I(1) The problem is that there is not a single value for a such that u t is stationary If z t and x t are cointegrated then there is a value of a such that z t ax t is stationary Then least-squares estimates of a are consistent but you have to change formula for standard errors

9 How to avoid spurious regressions? Answer: Add enough lags. Consider the following regression equation z t = ax t + bz t 1 + u t Now there are values of the regression coeffi cients so that u t is stationary, namely a = 0 and b = 1 So as long as you have enough lags in the VAR you are fine (but be careful with inferences)

10 Structural VARs Consider the reduced-form VAR y t = J A j y t j + u t j=1 For example suppose that y t contains the interest rate set by the central bank real GDP residential investment What affects the error term in the interest rate equation? the error term in the output equation? the error term in the housing equation?

11 Structural shocks Suppose that the economy is being hit by "structural shocks", that is shocks that are not responses to economic events Suppose that there are 10 structural shocks. Thus u t = Be t where B is a 3 10 matrix. Without loss of generality we can assume that E[e t e t] = I

12 Structural shocks Can we identify B from the data? E[u t u t] = BE[e t e t]b = BB We can get an estimate for E[u t u t ] using ˆΣ = T û t û t/(t J) t=j+1 But B has 30 and ˆΣ only 9 elements.

13 Identification of B Can we identify B if there are only three structural shocks? B has 9 distinct elements ˆΣ is symmetric Not all equations are independent. Σ 1,2 = Σ 2,1. For example Σ 1,2 = b 11 b 21 + b 12 b 22 + b 13 b 23 but also Σ 2,1 = b 21 b 11 + b 22 b 12 + b 23 b 13 In other words, different B matrices lead to the same Σ matrix

14 Identification of B We need additional identification assumptions u i t e 1 u y t t = B e 2 u r t t e mp t And suppose we impose B = Then I can solve for the remaining elements of B from In Matlab use B=chol(S) ˆB ˆB = ˆΣ

15 Identification of B Suppose instead we use And that we impose uy t u i t u r t = D D = e 1 t e 2 t e mp t This corresponds with imposing B = This does not affect the IRF of e mp t. All that matters for the IRF is whether a variable is ordered before or after r t

16 First-order VAR y t = A 1 y t 1 + Be t Now we can calculate IRFs, variances, and autocovariances analytically Mainly because you can easily calculate the MA representation y t = Be t + A 1 Be t 1 + A 2 1 Be t 2 +

17 State-space notation Every VAR can be presented as a first-order VAR. For example let [ ] [ ] [ ] [ ] y1,t y1,t 1 y1,t 2 e1,t = A y 1 + A 2,t y 2 + B 2,t 1 y 2,t 2 e 2,t y 1,t y 2,t y 1,t 1 y 2,t 1 = [ A1 A 2 I ] y 1,t 1 y 2,t 1 y 1,t 2 y 2,t 2 [ + B ] e 1, e 2, 0 0

18 VAR used by Gali z t = z t = ε t = J A j z t j + Bε t j=1 with [ ] ln(yt /h t ) ln(h t ) [ ε t,technology ε t,non-technology ]

19 Identifying assumption (Blanchard-Quah) Non-technology shock does not have a long-run impact on productivity Long-run impact is zero if Response of the level goes to zero Responses of the differences sum to zero

20 Get MA representation Note that D 0 = B z t = A(L)z t + Bε t = (I A(L)) 1 Bε t = D(L)ε t = D 0 ε t + D 1 ε t 1 +

21 Sum of responses D j = D(1) = (I A(1)) 1 B j=0 Blanchard-Quah assumption: [ D j = j=0 0 ]

22 If you ever feel bad about getting too much criticism...

23 If you ever feel bad about getting too much criticism... be glad you are not a structural VAR

24 Structural VARs & critiques From MA to AR Lippi & Reichlin From prediction errors to structural shocks Fernández-Villaverde, Rubio-Ramirez, Sargent Problems in finite samples Chari, Kehoe, McGratten

25 From MA to AR Consider the two following different MA(1) processes y t = ε t + 1 ] 2 ε t 1, E t [ε t ] = 0, E t [ε 2 t = σ 2 ] x t = e t + 2e t 1, E t [e t ] = 0, E t [e 2 t = σ 2 /4 Different IRFs Same variance and covariance E [ y t y t j ] = E [ xt x t j ]

26 From MA to AR AR representation: Solve for a j s from y t = (1 + θl) ε t 1 (1 + θl) y t = ε t 1 (1 + θl) = a j L j j=0 1 = a 0 + (a 1 + a 0 θ) L + (a 2 + a 1 θ) L 2 +

27 From MA to AR Solution: a 0 = 1 a 1 = a 0 θ a 2 = a 1 θ = a 0 θ 2 You need θ < 1

28 Prediction errors and structural shocks Solution to economic model x t+1 = Ax t + Bε t+1 y t+1 = Cx t + Dε t+1 x t : state variables y t : observables (used in VAR) ε t : structural shocks

29 Prediction errors and structural shocks From the VAR you get e t+1 = y t+1 E t [y t+1 ] = Cx t + Dε t+1 E t [Cx t+1 ] = C (x t E t [x t ]) + Dε t+1 Problem: Not guaranteed that x t = E t [x t ]

30 Prediction errors and structural shocks Suppose: y t = x t that is, all state variables are observed Then x t = E t [x t ]

31 Prediction errors and structural shocks Suppose: y t = x t F-V,R-R,S show that x t =E t [x t ] if the eigenvalues of A BD 1 C must be strictly less than 1 in modulus

32 Finite sample problems Summary of discussion above Life is excellent if you observe all state variables But, we don t observe capital (well) even harder to observe news about future changes If ABCD condition is satisfied, you are still ok in theory Problem: you may need -order VAR for observables recall that k t has complex dynamics

33 Finite sample problems 1 Bias of estimated VAR apparently bigger for VAR estimated in first differences 2 Good VAR may need many lags

34 Alleviating finite sample problems Do with model exactly what you do with data: NOT: compare data results with model IRF YES: generate N samples of length T calculate IRFs as in data compare average across N samples with data analogue This is how Kydland & Prescott calculated business cycle stats