基于双通道输入的CNN-BiLSTM模型的股价预测

Stock Price Prediction Using a Dual-Channel Input CNN-BiLSTM Model

摘要: 针对现有混合神经网络模型在处理多源时序数据时输入结构单一、难以有效融合不同频率特征的问题，提出一种基于双通道输入的CNN-BiLSTM股价预测模型。该模型设计双通道输入结构，分别处理5分钟级高频数据(开盘价、最高价、最低价、收盘价、成交量)和日级低频数据(日均价、日成交量、调整后收盘价)：高频通道通过两层卷积神经网络(CNN)提取日内局部波动特征，低频通道直接输入双向长短期记忆网络(BiLSTM)捕捉长期趋势依赖，最后将双通道特征进行门控融合以完成股价预测。以贵州茅台(600519.SH) 2020年1月至2025年12月的交易数据为实证对象，采用50日滑动窗口预测未来5日收盘价。实验结果表明，所提模型在测试集上取得预测准确率为98.07%，显著优于仅使用日级数据的RNN、GRU、LSTM，仅使用分钟级数据的CNN-LSTM、单通道CNN-BiLSTM以及LSTM-Transformer等基准模型，验证了双通道输入结构在捕捉多尺度时序特征方面的有效性。研究为高频金融时序预测提供了新的模型输入端架构思路，对投资者的量化决策具有积极的参考价值。

Abstract: To address the problem that existing hybrid neural network models have a single input structure and cannot effectively fuse features of different frequencies when processing multi-source time series data, this paper proposes a CNN-BiLSTM stock price prediction model based on dual-channel input. The model designs a dual-channel input structure to separately process 5-minute high-frequency data (open price, high price, low price, close price, volume) and daily low-frequency data (daily average price, daily volume, adjusted close price). The high-frequency channel uses a two-layer convolutional neural network (CNN) to extract intraday local fluctuation features, while the low-frequency channel directly inputs data into a bidirectional long short-term memory network (BiLSTM) to capture long-term trend dependencies. Finally, the dual-channel features are fused via a gating mechanism to complete stock price prediction. Taking the trading data of Kweichow Moutai (600519.SH) from January 2020 to December 2025 as the empirical object, a 50-day sliding window is used to predict the closing price of the next five days. Experimental results show that the proposed model achieves a prediction accuracy of 98.07% on the test set, significantly outperforming baseline models such as RNN, GRU, LSTM using only daily data, CNN-LSTM using only minute-level data, single-channel CNN-BiLSTM, and LSTM-Transformer, verifying the effectiveness of the dual-channel input structure in capturing multi-scale temporal features. This study provides a novel model input architecture for high-frequency financial time series prediction and has positive reference value for investors’ quantitative decision-making.