CLS-SVM: a local modeling method for time series forecasting

High computational complexity and low generalization ability seriously limit the forecasting application of LS-SVM in large scale time series. Aiming at this problem, a local modeling method called Clustering LS-SVM (CLS-SVM) is proposed. CLS-SVM uses the K-means algorithm to cluster time series dataset and adopts the variance ratio criterion to find the optimal clustering number. Then in each cluster, local LS-SVM modeling is implemented using Cholseky decomposition method instead of Conjugate Gradient method to improve the efficiency in solving the linear equation problem. Simulation experiment and real application test show that CLS-SVM can improve the modeling efficiency by 5 to 28 times without obvious precision dropping, and also effectively increase generalization ability and decrease computational complexity.