Capturing correlation changes by applying kernel change point detection on the running correlations

Change point detection methods signal the occurrence of abrupt changes in a time series. Non-parametric approaches, such as the Gaussian kernel based change point (KCP) detection (Arlot et al., 2012), are especially attractive because they impose less assumptions on the data. Yet, a drawback of these methods is that most of them are sensitive to changes in the mean, the variance, etc., making them less sensitive to specific kinds of changes. We show that KCP can be adapted to detect a particular type of change only. We focus here on correlation change, which has been put forward by different theories but proved hard to trace in multivariate time series. We propose KCP-corr, which boils down to applying KCP on the running correlations. To confirm that KCP-corr is more sensitive than merely applying KCP on the raw data (KCP-raw), a simulation study was conducted in which the number of (noise) variables and the size of the correlation change were varied. KCP-corr emerged as the better method especially in the more difficult but realistic settings where the correlation change is minimal and/or noise variables are present. KCP-corr also outperforms Cusum, a non-parametric method that specifically targets the detection of correlation changes.