An Improved Method of Sequential Probability Ratio Test for Change Point Detection in Time Series

In this study, we deal with real-time analysis and the forecast of time series data. In this case, it is important to detect the gap between prediction and observed data (structural change) as immediately, correctly, and simply as possible. And after the change point is found, it is necessary to rebuild the next prediction model as soon as possible. In this study, we modified the algorithm of sequential probability ratio test (SPRT) for structural change detection in the conventional method. Then, we compare the conventional method and proposed method by using two types of generative models in simple linear regression. As one of the results, the detection accuracy improved about 30 % and the detection points became faster by about 3 points. Furthermore, the variability and slope change of the time series data in the generation models enabled the estimation of structural change points.