For improving the prediction accuracy of the stock market datas by neural network, the stock market data is decomposed into a series wavelet components by harmonic wavelet, which are shift invariant and have different scale. Then, constructing the recursion neural network based on the charaters of the stock market, and each harmonic wavelet componet of stock market data is predicted by using of the constructed neural network. At last, the final stock market forecast data is obtained by harmonic wavelet reconstruction for forecasting results of different scale wavelet componets. The experimental results shows that the price fluctuations of different investmetn time horizon in stock market data can be well seperated, after the stock time series is decomposed by harmonic wavelet, the predicting accuracy of stock market is improved efficiently.