Multi-Scaled Forecasting Model Based on Support Vector Machines

Abstract

The theories of phase space reconstruction and Support Vector Machines (SVM) are introduced firstly. A novel time series forecasting model based on wavelet and SVM is proposed. It first performances multi-scaled decomposition on complex time series using discrete wavelet transformation. Then the reconstructed approximate series and detail series are forecasted respectively using SVM. Finally, the outcomes are coalesce together. The forecasting model is constructed and applied to the stock index data. Experimental results indicate that the proposed forecasting model has superiority over simple SVM and Artificial Neural Network (ANN) for it has lower forecast error.