Title: Capturing correlation changes by applying kernel change point detection on the running correlations
Authors: Jedelyn Cabrieto - KU Leuven (Belgium) [presenting]
Francis Tuerlinckx - KU Leuven (Belgium)
Peter Kuppens - Leuven (Belgium)
Eva Ceulemans - University of Leuven (Belgium)
Abstract: Change point detection methods signal the occurrence of abrupt changes in a time series. Non-parametric approaches are especially attractive in this regard because they impose less assumptions on the data. Yet, a drawback of these methods is that they are expected to be sensitive to changes in the mean, the variance, the correlation, and even the higher moments. This implies that one is not certain what kind of change the methods pick up, whereas this is often important from a substantive point of view. We focus on signaling correlation change, because this is put forward by different theories but proved hard to trace in multivariate time series. We demonstrate how correlation change can be detected by applying the best non-parametric method, kernel change point detection, on the running correlations rather than on the raw data. We inspect the detection performance of this approach in a simulation study and provide an illustrative example on detecting synchronicity in reactivity data.