A Hybrid Approach to Regime Shift Detection

Abstract

In this article, we propose a method for the analysis of regime shifts in frequency data. This method identifies those points in the develop- ment of a process for which deviations are most extreme. Based on a statistical model, functions are estimated that describe the process. This description can represent either the entire series of scores or the series before and after a shift point. The shift point can be either given a priori or estimated from the data. The method is hybrid in that it first uses standard models for the estimation of parameters of the process that is examined and then, in a second step, elements of Configural Frequency Analysis. Uni- and multivariate versions of the method are proposed. In data examples, road traffic data from California and Germany are analyzed before and after particular shift points. Extensions of the proposed method are discussed.