Segmentation of Nonstationary Time Series with Geometric Clustering

We introduce a non-parametric method for segmentation in regimeswitching 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. Such models can be learned efficiently from data, where clustering is used to propose one single split candidate at each split level. 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.