We are testing Markov switching models to forecast risk regimes, similar to the paper by Kritzman, Page and Turkington. We find that in some cases the Baum-Welch algorithm converges very slowly or not at all. Apparently this issue is known for hidden Markov models in speech recognition and other fields, where several alternatives to Baum-Welch have been proposed: for example, Bayesian estimation by M. Johnson or estimators based on an entropic prior by M. Brand.
Have these alternatives been applied to financial times series? Is there a reason to favor one particular approach, apart from slow convergence?