Optimizing Stock Trend Prediction in the Chinese Market: A Comparative Study of Machine Learning Models with Bayesian Hyperparameter Tuning

Predicting stock market trends has become increasingly vital in today’s volatile global economy. This paper addresses the challenge of accurate stock market prediction, particularly within the Chinese stock market, which has been subject to unique regulatory and economic shifts. This paper explores the application of five machine learning models to predict stock trends, comparing their performances and enhancing accuracy through hyperparameter tuning. This paper utilized a dataset from the CSI 300 Index, applying various feature engineering techniques and standardizing inputs. After training and backtesting the models, Bayesian optimization is employed to fine-tune the top-performing models: LightGBM, XGBoost, and GRU. This optimization process focuses on improving the models’ annualized returns, Sharpe ratios, and drawdowns. Each model’s predictions were tested on a backtest dataset from 2024, and key performance metrics were recorded. The initial results indicated that LightGBM, XGBoost, and GRU outperformed the other models, with LightGBM achieving an annualized return of 16.44%, XGBoost 11.44%, and GRU 16.11%. After Bayesian optimization, LightGBM improved to 20.97%, XGBoost to 17.20%, and GRU reached 27.22%, though with a higher drawdown. These results demonstrate that GRU’s risk-adjusted performance can offer significant returns, while LightGBM strikes a balance between risk and return. Future work could extend the dataset beyond the CSI 300 Index for better generalizability. While Bayesian optimization improved gradient boosting models, further tuning of neural networks and exploring advanced architectures like Transformers may enhance performance. Integrating real-world factors, such as transaction costs and macroeconomic indicators may strengthen their practical applications.