Title: Clustering Time Series and the Surprising Robustness of HMMs

Abstract: Suppose that you are given a time series where consecutive samples are believed to come from a probabilistic source, and that the source changes from time to time. Your objective is to learn the distribution of each source and to cluster the samples according to the source that generated them. A standard approach to this problem is to model the data as a hidden Markov model (HMM). However, due to the Markov property and stationarity of HMMs, simple examples can be given where this approach yields poor results for the clustering. We propose a more general, non-stationary model of the data, where the only restriction is that the sources can not change too often. Even though the model governing the sources may not be Markovian, we show that that a maximum likelihood HMM estimator can still be used. Specifically, we show that a maximum-likelihood HMM estimator produces the correct second moment of the data, and the results can be extended to higher moments. In contrast to the existing consistency and misspecification results involving maximum likelihood for HMMs, our approach yields bounds for finite sample sizes.