市场状态与股价可预测性：基于深度学习模型的实证

Market Conditions and Predictability of Stock Prices: An Empirical Analysis Based on Deep Learning Models

摘要: 金融市场研究表明，股票价格的可预测性具有显著的时变性与状态依赖特征，市场运行机制会随着趋势转折、波动结构变化及外生冲击而发生阶段性调整；然而，现有基于深度学习的价格预测研究更多关注模型整体预测精度，对预测表现在不同市场阶段下的系统性差异缺乏专门分析，从而难以准确刻画预测可靠性的环境约束。基于此，文章以沪深300指数为研究对象，构建融合小波变换的LSTM-CNN混合预测模型，并在样本外框架下比较模型在不同市场阶段中的预测表现。实证结果显示，股票价格预测表现具有显著的阶段依赖性：在趋势相对明确、价格结构较为稳定的阶段预测误差显著较低，而在趋势转折或震荡特征突出的阶段预测可靠性明显下降，且上述差异在不同参数设定下具有稳健性。研究表明，股票价格预测模型的有效性高度依赖市场环境，对预测性能的评估有必要引入市场阶段视角，以更准确界定模型的适用边界与现实价值。

Abstract: Studies in financial markets suggest that the predictability of stock prices exhibits significant time variation and state dependence. The operating mechanism of financial markets will undergo stage-wise adjustments with bull-bear transitions, changes in volatility structures, and exogenous shocks. However, existing literature on stock price forecasting using deep learning focuses primarily on the overall forecasting accuracy of models, while lacking dedicated analyses of the systematic differences in forecasting performance across distinct market regimes. Consequently, it is difficult to accurately characterize the environmental constraints that affect forecasting reliability. Against this background, this paper takes the CSI 300 Index as the research object and constructs a hybrid LSTM-CNN prediction model integrated with wavelet transform. We then compare the out-of-sample forecasting performance of the model across different market stages. The empirical results show that stock price forecasting performance presents significant regime dependence: the prediction error is significantly lower in periods with relatively clear trends and stable price structures, whereas the forecasting reliability decreases substantially during trend reversals or high-volatility periods. These findings remain robust under different parameter settings. This study indicates that the effectiveness of stock price forecasting models is highly dependent on market conditions. It is necessary to incorporate the perspective of market stages into the evaluation of forecasting performance, so as to better identify the applicable boundaries and practical value of forecasting models.