A dynamic matrix-variate model for clustering time series with multiple sources of variation

In this paper we present a dynamic matrix-variate model for clustering online several multivariate time series. The specific matrix state-space structure of the model allows for the inclusion of multiple sources of variation, such as those due to dynamics typical of time series, or the presence of time-dependent regressors. The use of matrix-variate normal distributions for the error terms and the introduction of an unknown selection matrix among the model matrices for clustering ensure flexibility, interpretability and usability in many contexts, while keeping the number of parameters small. We outline the estimation method and analyze, as an example, the training activities of one athlete using biometric data collected with a sportwatch.