Clustering Multiple Time Series with Structural Breaks

Abstract

Time series clustering pattern could change over time. In this article we develop a new Bayesian approach to handle clustering analysis of multiple time series with structural breaks. The number of breaks is treated as a random variable, with group membership and group‐specific parameters allowed to change on these breaks. Group‐specific parameters in each regime can be integrated analytically, so we only have a small number of parameters to be handled by posterior simulation. We further discuss prediction, identification, clustering, and detection of the number of groups. Using Monte Carlo simulation, we document the performance of the proposed approach in statistical efficiency, forecasting, and detection of the structural breaks. An application on quarterly industrial production growth rates of 21 countries links regimes to historical business cycles. Prediction performance and economic gains are illustrated based on the proposed method.