Using Chernoff's bounding method for high-performance structural break detection and forecast error reduction

Abstract

In this paper, a new method for detecting multiple structural breaks, i.e. undesired changes of signal behavior, is presented and applied to artificial and real-world data. It will be shown how Chernoff Bounds can be used for high-performance change-point detection after preprocessing arbitrary time series to binary random variables using adequate transformation routines. The algorithm is evaluated on artificial time series and compared to state of the art methods. The developed algorithm is competitive to state of the art methods in terms of classification errors but is considerably faster especially when dealing with long time series. Theoretical results on artificial data from part one of this paper are applied to real-world time series from a pharmaceutical wholesaler and show striking improvement in terms of forecast error reduction, thereby greatly improving forecast quality. In order to test the effect of structural break detection on forecast quality, state of the art forecast algorithms are applied to time series with and without previous application of structural break detection methods. © 2013 Springer-Verlag.