Nonparametric Bayesian factor analysis of multiple time series

We propose a nonparametric Bayesian factor analysis framework for characterization of multiple time-series. The proposed model automatically infers the number of factors and the noise/residual variance, and it is also able to cluster time series which behave similarly over prescribed time windows. We use a Pitman-Yor process to impose such clustering. We also provide a general MCMC inference scheme and demonstrate the proposed framework on the analysis of multi-year stock prices of companies in the S & P 500.