The extended switching regression model: Allowing for multiple latent state variables

In this paper we extend the widely followed approach of switching regression models, i.e. models in which the parameters are determined by a latent discrete state variable. We construct a model with several latent state variables, where the model parameters are partitioned into disjoint groups, each one of which is independently determined by a corresponding state variable. Such a model is called an extended switching regression (ESR) model. We develop an EM algorithm to estimate the model parameters, and discuss the consistency and asymptotic normality of the maximum likelihood estimates. Finally, we use the ESR model to combine volatility forecasts of foreign exchange rates. The resulting forecast combination using the ESR model tends to dominate those generated by traditional procedures.