A Nonparametric Bayesian Approach to Detect the Number of Regimes in Markov Switching Models

The literature on Markov switching models is increasing and producing interesting results both at theoretical and applied levels. Most often the number of regimes, i.e., of data generating processes, is considered known; this strong hypothesis is adopted to somewhat bypass the nuisance parameter problem which affects hypothesis testing for the number of regimes. In this paper we take the view that some results derived from a nonparametric Bayesian approach provide a convenient way to deal with the issue of detecting the number of components in the mixture density, based on the assumption that the parameter distributions are generated by a Dirichlet process. The advantage is that we need no testing (in a classical sense) for the number of regimes, and the approach is not affected by a change point at the beginning or at the end of the sample. A Monte Carlo experiment provides some insights into the performance of the procedure. The potentiality of the approach is illustrated in reference with some well known results on exchange rate modelling.