How We Conquered the 2025 Chinese University Big Data Challenge: Financial Time‑Series Lessons

Our team, named "Stay Overnight" (team members: Chen Keyan, Liu Yifan, and Tan Junwen, all from Chongqing University of Posts and Telecommunications), participated in the 2025 China University Computer Competition – Big Data Challenge and achieved the second place nationwide.

The competition task centered on financial market data characterized by high volatility and strong non‑linearity, which repeatedly challenged our problem‑solving process. Early on we were torn between time‑series methods and supervised learning, and we closely examined the baseline and state‑of‑the‑art (SOTA) models provided by the organizers.

During the initial phase we lacked a clear direction; we oscillated between model families, drew inspiration from community discussions, and realized that focusing solely on theoretical performance metrics was insufficient. Instead, we needed to balance model complexity with the practical characteristics of the data.

After extensive trial‑and‑error, we shifted our strategy toward supervised learning and devoted significant effort to feature engineering. We built a pipeline that combined fundamental time‑series features with domain‑specific financial technical indicators, which ultimately improved the model's predictive power.

The competition’s evaluation period introduced additional difficulties: the high volatility of stock indicators made model generalization a critical concern, and leaderboard rankings fluctuated dramatically across the A and C phases. Despite these setbacks, we persisted with incremental optimizations and maintained a steady improvement in scores.

In conclusion, we are grateful for the platform that allowed us to confront real‑world financial data, learn from many outstanding teams, and acquire valuable experience that will benefit future competitions and professional work. We look forward to seeing more excellent teams emerge in future editions of the contest.