A Constructive Competitive Regression Method for Analysis and Modeling of Non-Stationary Time Series

A simple constructive learning algorithm is proposed for regime discovery in non-stationary time series. The approach is based on competition among regression models aided by averaging of their squared residuals over neighboring data points and incremental introduction of additional models when needed. A case study on California’s deregulated power market discovered that 4 regimes existed in market clearing price for April 1998 – June 1999 period, which could be helpful information for better understanding of the market. Discovered local regression models outperformed a global modeling approach indicating that the proposed method can be used in practice for forecasting applications.