Condition for Implementing Stacked Generalization in Time Series Prediction

K.P. Pang (Australia)


CUSUM, stacked generalization, structural change


An analysis of Pang and Ting (2003) reveals that the traditional approach of Centered CUSUM of square that use all available data in one statistical test fails to detect the structural change when the size of data after the break is apparently larger than the size of data before the break. That research paper introduces another approach by reducing the data size after the break for improving the structural detection ability. This paper further extends this new conceptualization for structural change detection and integrates this new concept with stacked generalization (Wolpert 1992). Structural change is one of important concerns in time series. Time series data can be classified into two categories: "without structural change" and "with structural change". Better structural detection ability leads to have better classification. The results of this paper show that stacked generalization can improve the time series prediction when using the data which is classified as "without structural change" by our new conceptualization and classification. In the experiments, we compare the performance of stacked generalization using the data classified by traditional approach with the new approach to structural change detection. In the new approach, we do not take all available data in one statistical test. Instead, we take the multiple statistical tests using various data sizes. The result of this research enables us to find out the favorable condition for applying stacked generalization into time series prediction.

Important Links:

Go Back