Bayesian Hidden Markov Models for Early Warning

We show how Bayesian hidden Markov models may be employed to build early warning systems of particular risky events. The adopted model formulation assumes that every binary response variable depends only on the latent state further to the lagged covariates and response. A Markov chain Monte Carlo algorithm is proposed for estimation and forecasting, where the latter is based on the optimisation of the F-score. An application referred to banking crisis of countries based on an unbalanced panel dataset is used as an illustration.