Time Series Forecasting Using a Markov Switching Vector Autoregressive Model with Stochastic Search Variable Selection Method

Abstract

AbstractThis paper investigates forecasting performance using a Markov switching vector autoregressive (MSVAR) model with stochastic search variable selection (SSVS) method. An MSVAR model has been widely used for empirical macroeconomics. However, an MSVAR model usually fits in-sample better but forecasts poorly relative to a linear VAR model, and typically has a large number of parameters, leading to over-parameterization problem. The use of SSVS prior is expected to mitigate this over-parameterization problem by setting insignificant parameters to be zero in an automatic fashion. In recursive forecasting exercises of empirical study and Monte Carlo simulation, I find that implementing SSVS to unrestricted VAR or MSVAR model typically improve forecasting performance. However, the results show that the MSVAR model with the SSVS prior does not always provide superior forecasting performance relative to the linear VAR model with SSVS prior, and that the improvement is not significantly large enough to alleviate over-parametrization problem of an MSVAR model.