2026 Big Data Challenge: Award Winners Revealed with In‑Depth Competition Experience Shares (Phase 2)

The 2026 China University Computer Competition Big Data Challenge (online stage A) has completed the "Monthly Star" award selection; the winning student teams, each receiving 800 CNY, are announced. The competition focuses on a cross‑sectional ranking problem for CSI 300 component stocks, requiring participants to select the top‑K stocks each day and assign reasonable position weights.

Team 1 Experience

They first ran the official baseline to fully understand the data format, prediction target, and scoring metric. Recognizing that the task is a ranking problem rather than single‑stock price prediction, they avoided early pitfalls such as over‑emphasizing single‑stock return errors, which harmed ranking stability.

In data processing they emphasized time‑window selection, construction of training samples, and chronological validation splits to prevent leakage. Features were built from returns, volatility, volume changes, turnover rate, and other derived signals, aligning the training target with the competition’s ranking metric.

For modeling they added time‑series modules: multi‑scale temporal convolutions captured short‑term fluctuations and mid‑term trends, while additional components modeled cross‑stock relationships to encode competition among stocks on the same day. Feature‑crossing was used to combine price‑volume and trend‑volatility signals.

The training strategy avoided excessive model complexity. Instead of deepening the network indiscriminately, they iterated in small steps, modifying one structure or loss function at a time and observing validation changes. Loss functions incorporated ranking‑related objectives that emphasized top‑ranked stocks, and auxiliary tasks such as direction prediction and volatility modeling provided extra supervision.

Post‑processing was treated as a critical stage because the final submission contains only a few stocks. They experimented with different score‑fusion and weight‑adjustment schemes and calibrated the final rankings using historical performance, which noticeably improved result stability.

Team 2 (Yuzhi) Experience

They identified the core difficulty as the temporal nature and noise of stock data. After finding that absolute return prediction was highly sensitive to market sentiment, they transformed the target to a relative excess‑return ranking, converting the problem from regression to sorting, which immediately boosted model stability.

Feature engineering followed a "multi‑scale price‑volume" approach: four time windows (5, 10, 20, 60 days) were used to derive features, supplemented by classic indicators such as RSI, MACD, and Bollinger Bands. The authors observed that the combination of short‑term momentum and long‑term trend information was essential for filtering false signals.

Model selection started with LightGBM regression, which yielded modest results. Switching to XGBRanker for learning to rank dramatically improved performance, as ranking models naturally fit the stock‑selection scenario by focusing on relative ordering rather than absolute price changes. The ranking scores were then passed through a mean‑variance optimization layer to produce risk‑aware weights for the final portfolio.

Both teams stressed that robust data handling and feature design outweigh the pursuit of overly complex models. They avoided over‑fitting to specific market phases by prioritizing cross‑period stability and incremental improvements.

Additional competition details: more than 2,500 teams (over 4,000 participants) have registered, with the registration deadline on July 15. The official website, email, and QQ groups provide further information.