The Information Matrix Test for Markov Switching Autoregressive Models
Martes 12/11, 12h
Seminario presentado por Enrique Sentana
"The Information Matrix Test for Markov Switching Autoregressive Models with Covariate-Dependent Transition Probabilities"
Abstract
The EM principle implies that the moments underlying the information matrix test for multivariate Markov switching autoregressive models with a transition matrix that depends on observed variables are the smoothed values of the moments tested if the latent Markov chain were observed. Thus, we identify influence functions related to the conditional heteroskedasticity, skewness and kurtosis of the regression residuals for each of the regimes, and influence functions related to the conditionally demeaned outer product of the generalised multinomial logit residuals for each column of the transition matrix times the levels, squares and cross-products of the observed variables affecting the probabilities. We conduct an extensive Monte Carlo exercise to assess the finite sample behaviour of our tests under both correct and incorrect specification and apply them to the empirical exercise in Pouzo, Psaradakis and Solá (2022).
*Jointly written with Dante Amengual and Gabriele Fiorentini.
Enrique Sentana
Ph.D. in Economics, London School of Economics. Professor of Economics at CEMFI. Fellow of the Econometric Society. Research Fellow of the CEPR Financial Economics Programme. His research interests are Factor models for financial returns and macroeconomic time series, Non-normal distributions and copulas for portfolio allocation, Estimation by simulation, Identification tests in structural models.
Abstract
The EM principle implies that the moments underlying the information matrix test for multivariate Markov switching autoregressive models with a transition matrix that depends on observed variables are the smoothed values of the moments tested if the latent Markov chain were observed. Thus, we identify influence functions related to the conditional heteroskedasticity, skewness and kurtosis of the regression residuals for each of the regimes, and influence functions related to the conditionally demeaned outer product of the generalised multinomial logit residuals for each column of the transition matrix times the levels, squares and cross-products of the observed variables affecting the probabilities. We conduct an extensive Monte Carlo exercise to assess the finite sample behaviour of our tests under both correct and incorrect specification and apply them to the empirical exercise in Pouzo, Psaradakis and Solá (2022).
*Jointly written with Dante Amengual and Gabriele Fiorentini.
Enrique Sentana
Ph.D. in Economics, London School of Economics. Professor of Economics at CEMFI. Fellow of the Econometric Society. Research Fellow of the CEPR Financial Economics Programme. His research interests are Factor models for financial returns and macroeconomic time series, Non-normal distributions and copulas for portfolio allocation, Estimation by simulation, Identification tests in structural models.