A Recursive Partitioning Approach for Detecting Nonstationarity of Levels and Trends in Time Series Data.

In this study a method for detecting level stationarity in time series data based on recursive partitioning (regression trees) is examined. Recursive partitioning iteratively splits the time series into two segments with the goal of finding the most optimal difference between segment means, given certain criteria related to the length of the segments and the magnitude of the level differences. The performance of the method to recover the true transition points and the true stationary levels in simulated time series data is evaluated in terms of sensitivity analyses for different parameter settings. The performance is compared to a change point detection algorithm based on Binary Segmentation. The results show the recursive partitioning method, with an additional post-analysis adjustment which removes likely false positives, performs equally well as the change point analysis in detecting the true transition time. The recursive partitioning method appears to perform better when the estimation of the true stationary level is concerned.