Bayesian regularization neural network model for stock time series prediction

With strong nonlinear characterization ability, a BP neural network can effectively describe the characteristics of nonlinear time series. However, there are still some limitations, such as the ease of falling into a local optimum. Aiming at this problem, the Bayesian regularization optimization algorithm was used to improve the BP neural network. Under the premise of minimizing the objective function, the algorithm adjusts the weight update function through the conditional probability density and the prior probability of the historical data. Thus, the generalization capability of BP neural network will be enhanced. After an empirical study on stock time series prediction, we found that the improved network could prominently increase the prediction ability, while the ability of volatility prediction was better than that of other traditional algorithms.