A ovel Parameters Selection Approach for Support Vector Machines to Predict Time Series

Aimed to solve the problem that there is no structural approach to select the best free parameters for Support Vector Machines when being used in time series prediction, a novel approach is proposed. In this method, best parameters for SVM is not got at the minimal MSE (Mean Squared Error) of validation set, but that the residue of training set is in White Noise form. This conclusion is deduced from the fact that the targets of training set have inherent correlations with each other. This approach is also effective to predict time series with nolinear and non-stationary characteristics. Furthermore, by using this method, confidence interval can be computed under any given confidence degree ^{1 - alpha} which is an important value for many applications. Two algorithms to compute confidence interval are given under different circumstances. Program is also given about how to make dynamic on-line prediction. Experiment was made to predict the annual sunspot number and perfect result was achieved.