Our Winning Strategies for the 2025 China University Big Data Stock Prediction Contest

Team Name: Chef Team

Team Members: Su Zhihan (Tsinghua University), Liu Shuangshi (Tsinghua University), Xie Linfeng (Tsinghua University)

Ranking: 4th place nationwide

The 2025 China University Computer Competition – Big Data Challenge required participants to forecast the prices of the 300 constituent stocks of the Shanghai and Shenzhen stock exchanges. The dataset offered high quality, clear objectives, and well‑defined features, providing a solid foundation for modeling.

Our approach combined several innovations:

We introduced a GRU‑based deep time‑series model to capture temporal patterns and used its predictions as enhanced features for traditional tree models such as XGBoost.

Stocks were grouped into low, medium, and high risk categories based on historical volatility, and separate models were trained for each cluster to improve generalization across market conditions.

A custom asymmetric loss function was designed to give higher weight to extreme price movements, increasing sensitivity to large fluctuations.

Finally, we built a two‑stage ensemble that merged the GRU model with XGBoost, leveraging the strengths of deep learning and gradient‑boosted trees.

This pipeline secured a fourth‑place finish in the competition, confirming the effectiveness of multi‑source feature fusion and risk‑aware modeling.

Beyond the ranking, the competition deepened our understanding of end‑to‑end data‑science workflows, from data preprocessing and feature engineering to model selection and hyper‑parameter tuning. We also recognized that relying solely on historical price data is insufficient; incorporating fundamentals, financial reports, sentiment, and market‑wide indicators would likely boost model robustness.

Future work includes expanding the feature set, experimenting with more diverse neural architectures, refining ensemble strategies, and further optimizing the asymmetric loss to better capture extreme market events.