基于相空间重构理论与递归神经网络相结合的股票短期预测方法

New Approach of Short-term Stock Prediction Based on Combination of Phase Space Reconstruction Theory and Recurrent Neural Network

根据股票指数时间序列复杂的非线性特性,提出以相空间重构理论与递归神经网络相结合的股票短期预测新方法。以相空间重构理论确定最佳延迟时间和最小嵌入维数,以最佳延迟时间为间隔的最小嵌入维数作为递归神经网络的输入维数,并按预测相点步进递归的生成训练数据进行短期预测,提高了预测精度和稳定性。该方法应用于沪市股票综合指数预测,其结果与传统的单纯用BP网络模型预测的结果相比较,精度大大提高,证明了该预测模型和方法在实际时间序列预测领域的有效性和实用性。

A new approach of short-term stock prediction using PSRT （Phase Space Reconstruction Theory） combined with RNN （ Recurrent Neural Network） was presented according to the complex nonlinear character of stock time series. The optimal delay time and minimal embedding dimension were determined by PSRT and the input dimension of RNN was decided by minimal embedding dimension. The training samples were generated by means of the stepping recursive phase points,which could improve precision and stability of prediction. The new method was applied to shot-term forecasting of Shanghai stock index. Compared to the traditional standard BP neural network, the results showed higher precision. So this research acquires effective progress in the practical prediction of time series.