Reconnecting the Markov Switching Model with Economic Fundamentals

This paper seeks to investigate and remedy the apparent inability of Markov regime switching models to predict future states in the medium to long term. We show that projected time varying transition probability series in the model may be biased towards predicting regime switches with high probability in the short run, and as a consequence it is hard or impossible to obtain longer run inference. We propose a penalized maximum likelihood estimator where non-smoothness in the transition series has negative influence on the likelihood function, which is shown to remedy the short run bias. In an empirical investigation of U.S. real GDP, the penalized model works better in terms of forecasting future recessions as defined by the NBER business cycle dating.