We introduce a non-parametric approach for the segmentation in regime-switching time-series models. The approach is based on spectral clustering of target-regressor tuples and derives a switching regression tree, where regime switches are modeled by oblique splits. Our segmentation method is very parsimonious in the number of splits evaluated during the construction process of the tree--for a candidate node, the method only proposes one oblique split on regressors and a few targeted splits on time. The regime-switching model can therefore be learned efficiently from data. We use the class of ART time series models to serve as illustration, but because of the non-parametric nature of our segmentation approach, it readily generalizes to a wide range of time-series models that go beyond the Gaussian error assumption in ART models. Experimental results on S\&P 1500 financial trading data demonstrates dramatically improved predictive accuracy for the exemplifying ART models.

PDF file

In  ICPRAM-2012: 1st International Conference on Pattern Recognition Applications and Methods

Publisher  SciTePress Digital Library

Details