基于CNN-LSTM-Attention模型的股票预测及系统开发

Stock Prediction and System Development Based on the CNN-LSTM-Attention Model

摘要: 近年来，中国A股市场呈现出“高噪声、厚尾分布、政策扰动”等复杂特征，传统计量模型难以刻画其特征。深度学习技术凭借其强大的非线性建模能力，为金融时间序列预测提供了新范式。而现有研究多聚焦于美股或指数层面，对A股个股的研究较少，且缺乏系统化的模型部署与交互式展示。本文以京粮控股(000505)为研究对象，构建CNN-LSTM-Attention混合模型，融合CNN的局部特征提取能力、LSTM的长期依赖建模能力与Attention的动态权重分配机制，提出一种适用于A股市场的次日收盘价预测框架。实验结果表明，该模型在测试集上R²高达0.92，显著优于传统的LSTM模型，并通过FastAPI与ECharts实现了可部署、可交互的预测系统。

Abstract: In recent years, China’s A-share market presents complex features such as “high noise, thick-tailed distribution, and policy perturbations”, which are difficult to be characterized by traditional econometric models. Deep learning technology provides a new paradigm for financial time series forecasting by virtue of its powerful nonlinear modeling capability. Existing research focuses on the U.S. stock or index level, with less research on A-share stocks and a lack of systematic model deployment and interactive display. In this paper, we take Beijing Food Holdings (000505) as the research object, construct a hybrid CNN-LSTM-Attention model, integrate the local feature extraction capability of CNN, the long-term dependence modeling capability of LSTM and the dynamic weight allocation mechanism of Attention, and propose a next-day closing price prediction framework for A-share market. The experimental results show that the model has a high R² of 0.92 on the test set, which significantly outperforms the traditional LSTM models, and realizes a deployable and interactive prediction system through FastAPI and ECharts.