By using the model of neural network, this paper selects monthly data of New shanghai Composite Index from beginning of its publication (January 2006) to June 2011,a total of 66. Then it divides data into two groups, previous 62 are training group, the rest 4 are prediction group. By comparing error sum of squares in different models of the time lag window, it selects the appropriate one. In order to improve the simulation, this paper optimizes the initial data before prediction. It shows a good fitting result that the other fitting error less than 3% except the data in June with the reason of larger fluctuation in the stock market that time, and explain the stock market's predictability in short-term.