In this study, the author focus on the exchange rate forecasting. Exchange rates fluctuation is extremely complex, not only contains the linear part but also includes non-linear elements, In this paper, Simulated Annealing Algorithm is introduced to overcome the neural network easy fall into local minimum defects in BP neural network basis, in order to optimize the network weights and thresholds, and thus improve the prediction accuracy. Through several forecast experiments about the major currencies against, the result show that compare to the single use of BP neural network, after introduced Simulated Annealing Algorithm, the prediction accuracy and stability has been further improved, meanwhile time-consuming less than genetic algorithms and other optimization algorithms.