Abstract: Long short-term memory (LSTM) neural networks have demonstrated superior performance in predicting complex nonlinear systems such as stock price fluctuations. However, the traditional LSTM models do not take into account the coupling relationships among the three gating mechanisms and the impact of long-term memory on the model’s input. This paper enhances the transmission of long-term memory peeking information and the stability of the model by incorporating long-term memory into the input gate and coupling the three gating mechanisms into a unique gate mechanism. It constructs an LSTM model based on feature selection with random forest (RF-MIP-LSTM) and derives the forward and backward propagation algorithms for the model. Through predictions and comparisons on stock prices of Agricultural Bank of China, Yantian Port, Gree Electric Appliances, and the Shanghai Stock Exchange Index, the RF-MIP-LSTM model exhibits superior convergence speed and predictive accuracy compared to the traditional LSTM model.

Key words: stock price prediction, random forest (RF), long short-term memory (LSTM) neural network, long peephole

摘要： 长短时记忆（LSTM）神经网络在预测股价波动这类复杂的非线性系统中展现了较好的性能，然而LSTM模型没有考虑三个门控机制的耦合关系和长时记忆对模型输入的影响。通过增加输入门控的长时记忆窥视和耦合了三个门控机制的唯一门机制，增强了长时记忆信息传递和模型的稳定性，构建了基于随机森林特征选择的RF-MIP-LSTM模型，并推导了模型的前向与反向传播算法。通过对中国农业银行、盐田港、格力电器三只股票价格和上证指数的预测和比较，表明RF-MIP-LSTM模型的收敛速度和预测精度均优于LSTM模型。

关键词: 股价预测, 随机森林（RF）, 长短时记忆（LSTM）神经网络, 长时窥视孔