Who Won the 2026 Big Data Challenge Monthly Star Awards? Winners Share Their Competition Insights

The 2026 China University Computer Competition – Big Data Challenge online A stage entered its second Monthly Star selection, and the committee released the list of winning student teams, each awarded 800 RMB.

Team 1 Experience – After running the official baseline and fully understanding the data format, prediction target, and scoring method, they recognized the task as a cross‑sectional ranking problem over the CSI 300 constituents, aiming to select the top‑K stocks each day. In data processing they emphasized chronological validation splits to avoid leakage, constructing training samples with careful time windows, and enriching features beyond raw price‑volume data with returns, volatility, volume change, and turnover. Labels were aligned with the competition’s ranking metric to teach the model relative ordering rather than absolute returns.

For modeling they introduced multi‑scale temporal convolutions to capture short‑term fluctuations and medium‑term trends, and also modeled inter‑stock competition information to reflect the horizontal relationship among stocks. Feature crossing highlighted combinations such as "volume‑price coordination" and "trend‑volatility fusion". Their training strategy avoided overly deep networks; instead they iterated small changes, using ranking‑oriented loss functions that emphasized the top positions and adding auxiliary tasks (direction prediction, volatility modeling) to provide richer supervision. Post‑processing involved score fusion, weight adjustments, and historical calibration, which significantly stabilized the final limited‑stock output.

Team 2 ("柚子") Experience – They stressed that the core difficulty lies in the time‑series nature and noise of stock data. Initially predicting absolute returns proved unstable, so they transformed the target into a relative excess‑return ranking, converting the problem to a sorting task. Validation was performed strictly by time order. Feature engineering used multi‑scale windows (5, 10, 20, 60 days) and classic technical indicators (RSI, MACD, Bollinger Bands). Model-wise they switched from LightGBM regression to XGBRanker ranking, then applied a mean‑variance optimization layer to translate scores into risk‑aware weight allocations.

Team 3 Experience – Their short‑term trading rule (pick on day T, buy at T+1 open, sell at T+5 open) required rigorous time‑series data cleaning and standardization to prevent leakage. They built a multi‑dimensional factor system covering price, volume, market sentiment, and risk constraints, avoiding over‑specialized factors. The core advantage was a machine‑learning ranking model tailored to the T+1/T+5 scenario, combined with multi‑layer risk filters to mitigate single‑model market bias. Throughout they avoided over‑fitting to training‑set metrics, focusing on cross‑period stability and adaptability to various market regimes.

All three teams concluded that modest, well‑validated improvements—through solid data handling, thoughtful feature design, appropriate ranking loss, and careful post‑processing—outperform chasing complex or “magical” modules.