VAR Model Based Clustering Method for Multivariate Time Series Data

Abstract

In this study, we develop a clustering method for multivariate time series data. In practical situations, such problems can arise in finance, economics, control theory, and health science. First, we propose to use a simulation based approximation to the test statistic and develop a method to test if two multivariate time series are coming from same VAR process. Then, the testing method is extended to a group of multivariate time series objects. Finally, a new clustering algorithm is developed using the testing method. The proposed algorithm does not use a predetermined number of clusters and finds the best possible clustering from the data. Empirical studies are provided in this paper, and they establish the accuracy of the algorithm. Finally, as a practical example, the algorithm is implemented to identify activities of different persons from a real-life data obtained from single chest-mounted accelerometers worn by different individuals.