Time series segmentation with shifting means hidden markov models
- 1Aristotle University of Thessaloniki, School of Engineering, GR 541 24 Thessaloniki, Greece
- 2Numerical Weather Prediction Research, Meteorological Research Division, Environment Canada, Dorval, Quebec H9P 1J3, Canada
Abstract. We present a new family of hidden Markov models and apply these to the segmentation of hydrological and environmental time series. The proposed hidden Markov models have a discrete state space and their structure is inspired from the shifting means models introduced by Chernoff and Zacks and by Salas and Boes. An estimation method inspired from the EM algorithm is proposed, and we show that it can accurately identify multiple change-points in a time series. We also show that the solution obtained using this algorithm can serve as a starting point for a Monte-Carlo Markov chain Bayesian estimation method, thus reducing the computing time needed for the Markov chain to converge to a stationary distribution.