GARCH-LSSVM Coupled Predication Model and Its Application on Stock Index Forecasting

It is usually a challenge for traditional time series prediction models to combine the linear and non-linear factors effectively, which might cause a problem that the trend forecast is not accurate. In this paper we propose a new prediction model based on GARCH and LSSVM. The GARCH model is used to deal with the heteroscedasticity of the residual series of the closing price of the stock index data. At the same time, a number of technical indicators are constructed to train the LSSVM model and the corresponding predicted values and residuals are obtained. Further, the predicted value and the residual obtained respectively from GARCH and LSSVM are used as the training set to modify the LSSVM model, with the predicted value of the logarithmic closing price of the stock index obtained. This coupled model not only contains the linear trend of historical information, but also integrates the non-linear features such as market volatility information which is closely related to the target data. Numerical experimental results show that the prediction accuracy of the model is 98.22% on the testing data set. This model performs also better than the GARCH 97.84% and the LSSVM 98.04% respectively in accuracy deviation.