Structural Vector Autoregressions with Markov Switching. Markku Lanne University of Helsinki. Helmut Lütkepohl European University Institute, Florence

Transcription

1 Structural Vector Autoregressions with Markov Switching Markku Lanne University of Helsinki Helmut Lütkepohl European University Institute, Florence Katarzyna Maciejowska European University Institute, Florence Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

2 Motivation Identifying shocks is a central problem in SVAR models. Previous proposals: Restrictions on instantaneous effects of shocks (Sims, 1980). Restrictions on long-run effects of shocks (Blanchard/Quah, 1989; King/Plosser/Stock/Watson, 1991). Sign restrictions (Canova/DeNicoló, 2002; Uhlig, 2005). Bayesian methods (Koop, 1992). Using statistical data properties Heteroskedasticity (Rigobon, 2003; Lanne/Lütkepohl, 2008) Nonnormal residual distribution (Lanne/Lütkepohl, 2009) This paper: Use Markov regime switching Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

3 Overview 1 Model setup 2 Estimation 3 Inflation, Unemployment, Interest Rate 4 Conclusions Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

4 Overview Model setup 1 Model setup 2 Estimation 3 Inflation, Unemployment, Interest Rate 4 Conclusions Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

5 Model setup Reduced form VAR model Variables of interest: y t = (y 1t,..., y Kt ) y t = Dd t + A 1 y t A p y t p + u t d t deterministic term. u t (0, Σ u ). Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

6 Model setup Structural form VAR model AB-model à la Amisano & Giannini (1997): Ay t = D (s) d t + A (s) 1 y t A (s) p y t p + Bε t A typically has ones on its main diagonal. A or B may be identity matrix. ε t (0, Σ ε ). Σ ε diagonal. Σ u = A 1 BΣ ε B A 1. Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

7 Markov Switching Model setup Markov process s t {1,..., M} (t = 0, ±1, ±2,... ) Transition probabilities p ij = Pr(s t = j s t 1 = i), i, j = 1,..., M. Conditional distribution of u t u t s t N (0, Σ st ). Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

8 Model setup Identification of Shocks Assumption: M = 2 A matrix decomposition result Σ 1, Σ 2 positive definite a (K K) matrix B and Λ = diag(λ 1,..., λ K ) such that Σ 1 = BB and Σ 2 = BΛB Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

9 Model setup Identification of Shocks Assumption: M = 2 A matrix decomposition result Uniqueness of B Σ 1, Σ 2 positive definite a (K K) matrix B and Λ = diag(λ 1,..., λ K ) such that Σ 1 = BB and Σ 2 = BΛB B is (locally) unique if the λ j s are distinct and ordered in a specific way (e.g., smallest to largest) Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

10 Model setup Identification of Shocks Assumption: M = 2 A matrix decomposition result Uniqueness of B Identification assumption Σ 1, Σ 2 positive definite a (K K) matrix B and Λ = diag(λ 1,..., λ K ) such that Σ 1 = BB and Σ 2 = BΛB B is (locally) unique if the λ j s are distinct and ordered in a specific way (e.g., smallest to largest) Instantaneous effects of shocks are the same across all states Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

11 Model setup More than two states I Assumption: M > 2 Matrix decomposition Test for state invariant B Σ 1 = BB, Σ i = BΛ i B, i = 2,..., M imposes restrictions on covariance matrices which can be tested. LR test has asymptotic χ 2 -distribution with 1 2 MK(K + 1) K 2 (M 1)K degrees of freedom. Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

12 Model setup More than two states II Uniqueness of B Alternative decomposition B is (locally) unique if for each pair of equal diagonal elements, for example, in Λ 2 = diag(λ 21,..., λ 2K ) there is a corresponding pair of distinct diagonal elements in one of the other Λ i = diag(λ i1,..., λ ik ), i = 3,..., M. Σ i = A 1 Λ i A 1, i = 1,..., M Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

13 Overview Estimation 1 Model setup 2 Estimation 3 Inflation, Unemployment, Interest Rate 4 Conclusions Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

14 ML estimation Estimation Use EM algorithm to optimize (pseudo) log likelihood. Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

15 Overview Inflation, Unemployment, Interest Rate 1 Model setup 2 Estimation 3 Inflation, Unemployment, Interest Rate 4 Conclusions Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

16 Inflation, Unemployment, Interest Rate Research question and data Question of interest Has US monetary policy changed or just the volatility of shocks? (Primiceri, 2005) Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

17 Inflation, Unemployment, Interest Rate Research question and data Question of interest Has US monetary policy changed or just the volatility of shocks? (Primiceri, 2005) Data and variables quarterly US data Sample period: 1953Q1 2001Q3 Variables: π t inflation rate based on chain weighted GDP price index u t unemployment rate r t yield on three-months treasury bills Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

18 Inflation, Unemployment, Interest Rate Primiceri s identifying restrictions A = π t u t r t Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

19 Inflation, Unemployment, Interest Rate Table: Estimates of Structural Parameters of MS Models for (π t, u t, r t ) with Lag Order p = 2 and Intercept Term (Sample Period: 1953Q1 2001Q3) unrestricted model restricted model parameters estimates std.dev. estimates std.dev. λ λ λ log L T Note: Standard errors are obtained from the inverse Hessian of the log likelihood function. Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

20 Inflation, Unemployment, Interest Rate Table: Wald Tests for Equality of λ i s for Unrestricted Model from Table 1 H 0 test value p-value λ 1 = λ λ 1 = λ λ 2 = λ Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

21 Inflation, Unemployment, Interest Rate Test of Primiceri s identifying restrictions Estimated B matrix ˆB = (0.024) (0.026) (0.029) (0.027) (0.025) (0.026) (0.049) (0.086) (0.031) Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

22 Inflation, Unemployment, Interest Rate Test of Primiceri s identifying restrictions Estimated B matrix ˆB = LR test of lower triangularity (0.001) (0.024) (0.026) (0.029) (0.027) (0.025) (0.026) (0.049) (0.086) (0.031) Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

23 Inflation, Unemployment, Interest Rate Inflation responses to monetary policy shock Unrestricted Restricted Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

24 Inflation, Unemployment, Interest Rate Unemployment responses to monetary policy shock Unrestricted Restricted Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

25 Inflation, Unemployment, Interest Rate Interpretation of states  = , ˆΛ 1 = and ˆΛ 2 = Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

26 Inflation, Unemployment, Interest Rate Probabilities of State 2 (Pr(s t = 2 Y T )) for the unrestricted model for (π t, u t, r t ) Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

27 Overview Conclusions 1 Model setup 2 Estimation 3 Inflation, Unemployment, Interest Rate 4 Conclusions Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

28 Conclusions Conclusions Identification assumptions: MS in the reduced form residuals. The impulse responses are invariant across regimes. Advantages: Identifying assumptions in standard SVAR framework become overidentifying and can be tested. MS structure can be investigated with statistical methods. Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25

29 Extensions Conclusions Remaining problems: Tests for number of states. Confidence intervals for impulse responses. Models with many variables and regimes are computationally difficult to handle because of difficult likelihood function. Theory for asymptotic inference for models with cointegrated variables is not complete. Helmut Lütkepohl (EUI Florence) Carlo Giannini Conference Rome, January / 25