Abstract: In view of the problem of being apt to fall into local optimum caused by stochastic acquisition of initial weights of BP neural network, Bayesian Regularization（BR） algorithm is used to improve the traditional BP neural network model. The algorithm can adjust the fitness function of the BP model through the prior probability of historical data on the premise of ensuring the minimum network error, so that the generalization ability of the network is improved. The empirical study on stock time series forecasting shows that the prediction accuracy is 42.81% higher than that of the traditional BP model.

Key words: Bayesian Regularization（BR）, neural network, stock time series prediction

摘要： 针对BP神经网络初始权值随机获取所导致的易陷入局部最优的问题，采用贝叶斯正则化（Bayesian Regularization，BR）算法改进传统BP神经网络模型。该算法在保证网络误差最小的前提下，通过历史数据的先验概率调整BP模型的适应度函数，使网络的泛化能力得到提升。在股票时间序列预测的实证性研究中表明，比传统BP模型在预测精度上提高42.81%。

关键词: 贝叶斯正则化, 神经网络, 股票时间序列预测