In theory, both GRNN and AR*-GRNN models can map any nonlinear system. However, the error they cause is inevitable, because a nonlinear time series such as financial series is in fluenced by many factors, including wild market fluctuations, crashes and subsequent rebounds. From Table 3, we can see that the forecasting results from GRNN are satisfying and are more accurate than those from ARMA and GARCH models. However, a better technique is employed to examine its feasibility in forecasting nonlinear time series. Experimental results indicate that the AR*-GRNN out performs the former two approaches in terms of forecasting accuracy. Also a conclusion can be drawn that innovation series inferred from AR* models can capture volatility and statistical information of the financial time series. The results indicate that it is effective method to combine the time series models with