Big Data Challenge 2026: Monthly Star Winners Announced with Winning Teams’ Experience Shares (Third Edition)

The 2026 China University Computer Competition Big Data Challenge (online stage A) opened its Monthly Star award selection, publishing the winning team list and awarding each team 800 CNY.

Team "无糖可乐" experience – Leveraging prior mathematical modeling competition background, the team split the problem into data construction, validation system, ranking model, and post‑processing. They gathered historical HS300 data from 2015‑01‑05 to 2026‑03‑13 (605 859 rows, 300 stocks, 2 718 trading days) and built training (604 359 rows) and test (1 500 rows) sets, auditing sources such as Tencent, Baostock, and AKShare. Their core model, StockTransformer, processes 60‑day sequences per stock with a TransformerEncoder, then applies Cross‑Stock Attention to capture inter‑stock relationships, outputting ranking scores focused on Top‑5 performance. Feature engineering produced a 158 + 39 dimensional set (alpha factors, technical indicators, residual momentum, EWMA risk, cross‑sectional ranks) and used a WeightedRankingLoss combining listwise, pairwise, and top‑k weighting. Validation employed rolling‑window checks, reporting Top‑5 returns of 0.03739, 0.04065, 0.03173, with the best model achieving 0.05808 and a selection score of 0.18685. The final solution fused the Transformer, LightGBM LambdaRank, rule‑based factors, and covariance portfolio optimization via an ensemble strategy implemented in ensemble_predict.py and audited by competition_workflow.py.

Team "绫袅" experience – Using LightGBM to predict one‑week stock returns, the team initially engineered 97 price‑volume features, later expanding to 142 before encountering over‑fitting. Through information‑coefficient analysis and ablation, they pruned unstable or highly correlated features (>0.9) down to 20 high‑impact variables (volatility, volume, price position ranks). Training employed NDCG initially, then switched to average Top‑5 equal‑weight weekly return as the primary metric, performing 50 rolling‑window validations with a 5‑day hold‑out to avoid label leakage. Hyper‑parameter tuning used a two‑stage process: single‑parameter scans (learning rate, num_leaves, lambda_l1/l2, feature_fraction) followed by grid search on top candidates, selecting configurations based on validation returns. Reflections highlighted limited gains from pure price‑volume signals, the challenge of capturing external factors, and the need for richer features; future work will treat the regression target as a ranking problem, predicting percentile ranks and optimizing the Top‑5 average percentile, which already yields >10 % weekly returns in the best windows.

The competition attracted over 3 000 teams and 4 600 participants, with registration closing on July 15 12:00. Participants can download data for local development and submit code and models via the platform.