Chaotic time series prediction using filtering window based least squares support vector regression

Abstract

When the traditional strategy of sliding window (SW) deals with the flowing data, the data far from current position are mechanically and briefly moved out of the window, and the nearest ones are moved into the window. To solve the shortcomings of this forgetting mechanism, the strategy of filtering window (FW) is proposed, in which adopted is the mechanism for selecting the superior and eliminating the inferior, thus resulting in the data making more contributions to the will-built model to be kept in the window. Merging the filtering window with least squares support vector regression (LSSVR) yields the filtering window based LSSVR (FW-LSSVR for short). As opposed to traditional sliding window based LSSVR (SW-LSSVR for short), FW-LSSVR cuts down the computational complexity, and needs smaller window size to obtain the almost same prediction accuracy, thus suggesting the less computational burden and better real time. The experimental results on classical chaotic time series demonstrate the effectiveness and feasibility of the proposed FW-LSSVR.