Abstract: Neural networks have the capability of approaching nonlinear function with any accuracy. Time-serial prediction model can perfectly show the nonlinear tendency of information. In this paper, the genetic algorithm (GA) with global search capability is adopted to improve traditional back propagation (BP) neural networks, based on analyzing defect of BP neural networks. At the same time, the theory of optimizing initial weights and threshold of neural networks by means of GA method is discussed in detail. In the course of explaining prediction method, two examples are taken to analyze GA-BP neural networks prediction performance and characteristics. The result of prediction shows that the prediction accuracy and adaptability of GA-BP neural networks was better than that of traditional BP neural networks.