Learning Segmentation from Time Series Clustering

DOI: 10.5176/2251-1911_CMCGS15.41

Authors: Maha Bakoben, Tony Bellotti, and Niall Adams

Abstract:  Clustering models that rely on independence assumptions are inapplicable to time series data. We propose an appropriate multivariate time series clustering model which considers serial correlation, trend and seasonal variations. In this paper, we present the application of time series clustering in constructing segmentation models which are aimed at developing a separate prediction model for each segment. The proposed model is tested on a real data set of credit card account behaviours and successfully identifies distinct risk groups. A high prediction performance of account defaults is obtained by segmentation, when a different prediction model is built for each cluster.

Keywords: multivariate time series clustering; segmentation model; vector autoregression model