A study of non-periodic short-term random walk forecasting based on RBFNN, ARMA, or SVR-GM(1,1|τ) approach

This paper introduces several prediction models for short-term random walk forecasting like stock price indexes forecasting. The radial basis function neural net (RBFNN) is widely applied to function approximation or classification issues. An autoregressive moving-average method has been utilized on the topic of time series. SVR-GM(1,1|τ) model employ the support vector machines (SVM) learning algorithm to improve the control and environment parameters in GM(1,1|τ) model, that is, enhancing generalization capability in the non-periodic short-term prediction. Therefore, this proposed method could smooth the overshooting problem, that often occurred in GM(1,1|τ) model or autoregressive moving-average (ARMA) method, so as to achieve better prediction accuracy. Finally, the comparison of performance on international stock price indexes forecasting between the various models has been done.