Relearning Process for SPRT in Structural Change Detection of Time-Series Data

Abstract

This paper proposes a process for relearning a prediction model after de- tecting a structural change point. Three problems are encountered in the detection of structural change points in time-series data: (1) how to generate a prediction model, (2) how to detect a structural change point rapidly, and (3) how to relearn the prediction model after detection. This paper targets the third problem and pro- poses five relearning methods and a process that embeds the relearning process in the sequential probability ratio test. The two experiment using 20 generated data set and TOPIX which consists of 1104 time-series data points between 1991 and 2012 show that it is useful to relearn the model by using past and future data from detected structural change point.