SSCL-GBM: A Semi-Supervised Stock Prediction Approach With Custom Loss Function

Accurately identifying effective market features and making reasonable model predictions are crucial for investment decision making in financial market prediction. This study proposes a LightGBM model, semi-supervised classification and learning with gradient boosting machine (SSCL-GBM), that combines semi-supervised learning and a custom loss function for classification tasks in stock price prediction. In typical return prediction tasks, labels are derived from future price movements, which can create a temporal mismatch between historical features and future-dependent labels. We adopt a pseudo-labeling mechanism to approximate labels in the final prediction window to address this issue and avoid relying on unavailable future outcomes during training. This enables the model to utilize the most recent data without violating temporal causality. Additionally, we design a custom loss function that balances return and risk, significantly reducing false positives. Experimental results show that our SSCL-GBM model significantly outperforms key indicators, such as cumulative returns and the Sharpe ratio, validating the effectiveness of this method in financial market prediction.