A systematic machine learning framework for next-day stock return direction prediction: multi-dimensional feature fusion, ensemble selection, and data balancing

This study proposes a comprehensive machine learning framework for predicting the next-day price direction of individual stocks, using China Merchants Securities (600999.SH) from 2017 to 2024 as the empirical case. We construct a multi-dimensional heterogeneous feature set by integrating historical price data, technical indicators, five-factor model risk exposures (MRP, SMB, HML, RMW, CMA), and a composite Investor Sentiment Index (ISI). An innovative ensemble feature selection strategy combining nine methods with a voting mechanism reduces the original feature pool by 73.2%, retaining 15 core features, among which BIAS_5, PVT, and ROC_5 show the strongest predictive power. To handle the prevalent class imbalance in financial returns, we systematically evaluate nine data balancing techniques across nine classification models, generating 81 combinations. Results demonstrate that Random Oversampling paired with LightGBM delivers the best performance (AUC = 0.916, F1-score = 0.834), while the baseline model without balancing also achieves highly competitive results. Ablation experiments highlight price momentum indicators as the most critical predictive component. This work provides a replicable end-to-end paradigm for robust, scientific application of machine learning in stock directional forecasting, offering valuable guidance for quantitative investment practices in the A-share market.