Multivariate change-point analysis for climate time series

The aim of this work is to find individual and joint change-points in a large multivariate database of climate data. We model monthly values of precipitation, minimum and maximum temperature recorded in 360 stations covering all Italy for 60 years (12 × 60 months). The proposed three variate Gaussian change-point model exploits the Hierarchical Dirichlet process, allowing for a formalization that lets us estimate a different change-point model for each station. As stations possibly share some of the parameters of the trivariate normal emission distribution, this model framework provides an original definition of the change-points corresponding to changes in any subset of the 12 model parameters. In this paper, results for two stations in Southern Italy are shown as an example.