Time series prediction method based on LS-SVR with modified Gaussian RBF

Abstract

LS-SVR is widely used in time series prediction. For LS-SVR, the selection of appropriate kernel function is a key issue, which has a great impact with the prediction accuracy. Compared with some other feasible kernel functions, Gaussian RBF is always selected as kernel function due to its good features. As a distance functions-based kernel function, Gaussian RBF also has some drawbacks. In this paper, we modified the standard Gaussian RBF to satisfy the two requirements of distance functions-based kernel functions which are fast damping at the place adjacent to the test point and keeping a moderate damping at infinity. The simulation results indicate preliminarily that the modified Gaussian RBF has better performance and can improve the prediction accuracy with LS-SVR. © 2012 Springer-Verlag.