Robust nonparametric multiple changepoint detection for multivariate variability

Abstract

Two robust, nonparametric multiple changepoint detection algorithms are introduced: DWBS and MKWP. These algorithms can detect multiple changes in the variability of a sequence of independent multivariate observations, even when the number of changepoints is unknown. The algorithms DWBS and MKWP require minimal distributional assumptions and are robust to outlying observations and heavy tails. The DWBS algorithm uses a local search method based on depth-based ranks and wild binary segmentation. The MKWP algorithm estimates changepoints globally via maximizing a penalized version of the classical Kruskal–Wallis ANOVA test statistic. It is demonstrated that this objective function can be maximized via the well-known PELT algorithm. Under mild, nonparametric assumptions, both of these algorithms are shown to be consistent for the correct number of changepoints and the correct location(s) of the changepoint(s). A data driven thresholding method for multivariate data is introduced, based on the Schwartz information criteria. The robustness and accuracy of the new methods is demonstrated with a simulation study, where the algorithms are compared to several existing algorithms. These new methods can estimate the number of changepoints and their locations accurately when the data are heavy tailed or skewed and the sample size is large. Lastly, the proposed algorithms are applied to a four-dimensional sequence of European daily stock returns.