Abstract: This paper presents a method for predicting multivariable time series with neural networks. First, we implement wavelet decomposition for Shanghai Stock Exchange (SSE) index time series of open, lowest, highest and close prices which are correlated one another and obtain the approximation series at lower resolution, considered as the trend of SSE index time series, by reconstruction with setting coefficients representing details zero. Then four attractors were reconstructed with these delayed approximation series, respectively, thus getting points on attractors in reconstructed phase-spaces or four sets of vector series. Then, combined these four vector series as one vector series of higher dimension, which was taken as the inputs to the neural network for training. Finally, the trained network was used to predict trend of SSE index time series at beginning of 2000 when the strongest bull market in history had just started.