Stock Return Prediction Based on Ensemble and CNN Algorithms

Although traditional rolling training methods are widely used for model performance evaluation and optimization, they often lead to data leakage during the training process, which affects the model's reliability and effectiveness. To address this issue, this paper proposes the use of ensemble models based on XGBoost and LightGBM, along with a CNN model, to predict future stock returns while avoiding the use of future data during rolling training, thereby effectively preventing data leakage. Using daily trading data from the constituent stocks of the CSI 300 Index, a multi-factor model is constructed based on various technical indicators (such as exponential moving average, standard deviation, linear regression slope, relative strength index, and Williams %R) for stock return prediction. Since return calculation involves the closing prices of the next five trading days, the traditional LightGBM multi-factor model experienced data leakage during rolling training. To solve this, ensemble models and the CNN model were introduced to predict the returns for the last five days of each rolling training window using the return data from the previous 252 days, effectively preventing data leakage. Backtesting results show that the cumulative return of the ensemble model increased to 193.54%, demonstrating higher stability and predictive power across different backtesting periods, outperforming the traditional and CNN models in key metrics such as annualized return, Sharpe ratio, and maximum drawdown.