基于结构修剪神经网络的股票指数预测模型

摘要

Since stock market is a nonlinear system with internal structural complexity and external factors variability, proposed forecasting index system which involved Shanghai Composite Index' s price, volume, and macroeconomic indicators closely related to stock market. Then analyzed the long-run equilibrium and causal relationship among the variables. Based on Bayesian theory, in order to optimize structure and ensure generalization ability of network, added network complexity function to error function which could delete insensitive hidden layer neurons through dynamic adjusting penalty factor. The empirical results with different forecasting index systems indicate that the pruning structure neural network model based on BP algorithm can be an effective way to improve forecasting accuracy. Comparing with other neural network models, the proposed model can improve forecasting performance with higher forecasting accuracy and more concise structure.%股票市场是非线性系统,具有内部结构复杂性和外部因素多变性,在股市指数价格和成交量基础上,引入宏观经济指标共同构建模型预测指标体系,并分析各指标之间的长期均衡关系和因果关系.在贝叶斯分析的基础上,将代表网络复杂性的惩罚项引入模型误差函数中,并通过动态调整惩罚因子删减网络中对股票市场不敏感的隐层神经元,在保证模型泛化能力的同时实现网络结构精简.以上证指数为例,构建基于BP算法的结构修剪神经网络预测模型,在不同的预测指标体系下对股票市场运行规律进行学习,并对上证指数进行仿真预测.最后,通过与其他神经网络预测模型比较验证该模型的有效性.