摘要
Abstract
Addressing the issues of unidirectional modeling bias,noise sensitivity,and vanishing gradients in LSTM for stock prediction,this paper proposes an improved model that integrates bidirectional LSTM,multi-head attention mechanism,and residual connections.This model utilizes bidirectional LSTM to extract historical and future information,enhances key time steps through the attention mechanism,and ensures the stability of deep network training through residual connections.Experiments based on data from 30 A-shares from 2018 to 2025 show that the mean squared error of the improved model is 0.00016,a reduction of 91%compared to the traditional LSTM(0.00185);the annualized return rate of real-time backtesting reaches 32%,significantly outperforming the benchmark strategy.This study provides an effective solution for financial time series prediction.关键词
长短期记忆模型/注意力机制/残差连接/股票预测Key words
Long Short-Term Memory/Attention Mechanism/Residual Connection/Stock Prediction分类
信息技术与安全科学
