Mixtures of autoregressions with an improper component for panel data

N.T. Longford and P. D'Urso


In the analysis of multiple time-series or panel data, we come across problems similar to the analysis of other multivariate data -- a particular simple model fits well to a subset of the cases, but not to all of them. We apply mixture models to such a setting, assuming that the cases in the study represent a small number of distinct subpopulations (or clusters) to which different models apply. In particular, we consider mixtures of autoregressive models of possibly different orders. Further, we consider a subset of the cases to which no model applies, and associate them with an improper (black-hole) distribution. An example with monthly income from outlets of a retail company is presented.

Journal of Classification 29; to appear.
July 2012