Time Series Prediction And Application Based On Multi-Kernel Support Vector Machine Method

The estimation of model parameters is an important point in the study of nonlinear time series models.Generally,both the maximum likelihood method and the conditional least square method have the best estimation effect when the data are normally distributioned.Support vector regression(SVR)is a powerful regression estimation method,and does not need to determine the data distribution in advance.Therefore,this paper estimates the parameters of the nonlinear time series model by using SVR to improve the predictive performance of the model in further study.The study on nonlinear time series models is generally from two perspectives.one is the nonlinearity of the variance series,the representative model is the generalized autoregressive conditional heteroscedasticity(GARCH)model.Based on previous research and the development of multi-kernel learning in recent years.In this article,a multi-kernel support vector regression(MRMRKA-SVR)base on kernel alignment(KA)and minimum redundancy maximum correlation(MRMR)algorithm is used to estimate the parameters of the GARCH model.First,the MRMRKA algorithm is used to select a set of high-quality basic kernels,then use multi-kernel learning to establish a multi-kernel function,then this multi-kernel function is used as the kernel of the SVR to establish a MRMRKA-SVR model,and finally useing this model for GARCH parameter estimation.Empirical analysis is carried out on the stock price data of China Unicom and ICBC,and compared with the GARCH model which estimation of the mixed Gaussian kernel SVR.The result shows that the prediction accuracy of the MRMRKA-SVR-GARCH model varies with the change of the basic kernel number m.It fluctuates slightly.But the prediction errors of MRMRKA-SVR-GARCH model decrease by 1.42%--16.2% compared with the SVR-GARCH model with mixed Gaussian kernel.Therefore,using MAMAKA-SVR to estimate the parameters of the GARCH model can further improve the prediction performance of the model,and the MRMRKA-SVR-GARCH model can provide a certain degree of reference in analyzing the trend and forecast of volatility.The other is the nonlinearity of the mean series,and the most used model is the self-exciting threshold autoregressive(SETAR)model.This article first proposes to use SVR instead of maximum likelihood and conditional least squares to estimate the parameters of the SETAR model.Empirical analysis on the stock price data of China Unicom and ICBC shows that,on China Unicom's data,the prediction error of the Gaussian kernel SVR-SETAR model is 8% lower than that of the ML-SETAR model and the CSS-SETAR model,and on ICBC data,the prediction error of the Gaussian kernel SVR-SETAR model is 26.32% and 20%lower than that of the ML-SETAR model and CSS-SETAR model,respectively,which proves that SVR can be used to estimate the parameters of the SESAT model,the SVR-SETAR model has higher predictive ability.