2026 Big Data Challenge Announces Monthly Star Winners and Shares Their Winning Strategies

In the 2026 China University Computer Competition – Big Data Challenge online stage A, the organizing committee reviewed the submitted models and code, reproduced the results, and announced the list of student teams that earned the monthly star award, each receiving an 800 RMB prize.

The two best‑performing teams shared their practical experience. The first team highlighted their feature engineering: basic time‑series features such as moving averages, returns, volatility across 3, 5, 10, 20, 40‑day windows; cross‑sectional rank features like daily stock rank and excess return over market average; and environment features including market sentiment, rise‑fall ratio, and overall volatility. No external data were added, using only the competition’s provided training and test sets.

For modeling, they built baselines with LightGBM, HistGradientBoosting, and Random Forest, noting that LightGBM Ranker aligns with the ranking objective. Model fusion was central: multiple base models were trained, their out‑of‑fold predictions fed into a second‑level fusion model to reduce single‑model volatility, plus two auxiliary models predicting daily Top‑1 and short‑term explosive targets. Return prediction and up/down probability were modeled separately.

Training and validation employed time‑series cross‑validation with four folds, a five‑day gap between folds to avoid leakage, each validation set covering 20 days and training sets at least 120 days. The random seed was fixed at 20260416 for reproducibility; hyper‑parameters were largely left at defaults, with only the number of trees tuned (LightGBM 200‑300 trees, HistGradientBoosting ~300 rounds).

The combination strategy used a multi‑stage stock selection pipeline: generate a candidate pool, perform fine ranking, then apply a risk penalty. They introduced a target_precision_gate label requiring positive return, daily rank within the top 25 %, stable short‑ and mid‑term performance, and maximum drawdown not exceeding 3 %. Position sizing was not forced to full; positions were reduced or emptied when market conditions were unfavorable. The overall emphasis was on reliable features and validation before fine‑tuning parameters.

The first‑stage monthly star winner also reflected on data understanding and processing: recognizing the strong temporal and cross‑sectional nature of stock data, handling missing and outlier values, and stressing that a proper time‑series split is essential for realistic performance estimation. Their feature engineering philosophy favored stable, interpretable features over overly complex ones to avoid online bias. Model choice prioritized robustness over extreme fitting, and they warned against over‑optimizing a specific period, which can cause severe drawdowns elsewhere.

Both teams thanked the organizers for providing the platform and data, and encouraged other participants to continue sharing insights and improving their approaches.