A forecasting method of time series called wavelet-domain predictor is proposed.Some non-stationary time series can be decomposed into several approximate stationary time series with wavelet decomposition.Decomposed time series are forecasted with auto-regression model

to obtain forecasting results of the original time series.Experiments with sunspot activity data show that the method is better than traditional forecasting approaches and neural network approaches

and can be applied to forecasting of some non-stationary time series effectively.