Model selection criteria for model-based clustering of categorical time series data: A Monte Carlo study

Abstract

An open issue in the statistical literature is the selection of the number of components for model-based clustering of time series data with a finite number of states (categories) that are observed several times. We set a finite mixture of Markov chains for which the performance of selection methods that use different information criteria is compared across a large experimental design. The results show that the performance of the information criteria vary across the design. Overall, AIC3 outperforms more widespread information criteria such as AIC and BIC for these finite mixture models.