Bayesian Semiparametric Multivariate Change Point Analysis

We develop a general Bayesian semiparametric change-point model in which separate groups of parameters (for example, location and dispersion) can each follow a separate multiple change-point process, driven by time-dependent transition matrices among the latent regimes. The distribution of the observations within regimes dened by the various change-points is unknown and given by a Dirichlet process mixture prior. The prior-posterior analysis by Markov chain Monte Carlo techniques is developed on a multivariate forward-backward algorithm for sampling the various regime indicators.