Back propagation(BP) neural network has strong nonlinear mapping ability and can handle nonlinear classification and prediction problems well, but BP neural network converges slowly and easily falls into local minima, which affects the prediction accuracy. In order to improve the prediction accuracy, this paper took BP neural network as the research object, reduced the dimensionality of the data by Principal Component Analysis(PCA), and optimized the BP neural network using Genetic Algorithm(GA), in which, the roulette selection method is used to select the individuals with better adaptability when performing the selection operation, so as to establish a PCA-GA-BP neural network prediction model and compare with a single BP neural network model. The experimental results has shown that the accuracy of the optimized PCA-GA-BP neural network model has been improved by 9.8066 % , which has provided a constructive basis for the application and improvement of BP neural networks.