Abstract: A projection pursuit autoregression model based on wavelet decomposition (PPARWD) has been developed to reveal the characteristics of mid-and-long term runoffs and resolve the problem of low prediction accuracy. This model adopts a new idea, processing then forecasting, and makes use of the multi-resolving power of wavelet analysis and the high-dimensional approaching capacity of projection pursuit autoregression (PPAR). It decomposes a time series of annual runoff into one approximate signal and several detailed signals by wavelet decomposition, and then uses the PPAR model to predict each of these signal series and reconstructs the final results. This PPARWD model is applied to the annual runoff at the Yichang hydrological station, and compared it with the PPAR model, a BP neural networks model, and an autoregressive moving average (ARMA) model. The results show that it has better prediction accuracy and stability and its predictions are insensitive to the decomposed scale coefficients.