Autocovariance Based Weighting Strategy for Time Series Prediction with Weighted LS-SVM

Abstract

Classic kernel methods (SVM, LS-SVM) use some arbitrarily chosen loss functions. These functions equally penalize errors on all training samples. In problem of time series prediction better results can be achieved when the relative importance of the samples is expressed in the loss function. In this paper an autocovariance based weighting strategy for chaotic time series prediction is presented. Proposed method can be considered a way to improve the performance of kernel algorithms by incorporating some additional knowledge and information on the analyzed learning problem.