The wavelet transform and the polynomial regression were combined with the GARCH model to analyze and forecast China's foreign exchange reserve.The data were de-noised by the db4 wavelet to establish polynomial regression model.Seeing that between the residuals of the de-noised data and the regression model there exist the auto regressive conditional heteroskedasticity (ARCH) effects, the GARCH model was created.Linear superposing the regression model and GARCH model, a regression-GARCH based on wavelet analysis was proposed.Comparing the predicted value and the actual value, it is found that the result in the paper is quite well.In other words, the method presented has obvious advantages and it's a useful predictive analysis tool in dealing with the non-stationary time series which has obvious growth trend just like foreign exchange reserves.