experiment design and evaluation protocol; an independent script that re-verifies factor RankIC, coverage and the leaderboard regressions; writing. Model code was built with AI tools and checked by me.
Two questions, one empirical and one about the competition itself. First: on the CSI 300, do price–volume signals (trade-weighted-price deviation, similar-stock co-movement) improve a LightGBM cross-sectional baseline out of sample? Second: when 85 teams optimise against a public leaderboard, how much of the public advantage is real and how much is selection on noise?
Organiser data for 2017–2022 (three tables). Six-year walk-forward evaluation with the number of models and the random seeds held equal between the baseline and the augmented model, so that any RankIC difference cannot be explained by extra tuning. Portfolio returns are computed through the official constrained optimiser; confidence intervals use 21-day block resampling.
The experimental design and the evaluation protocol (equal model counts, placebo seeds, block resampling). I wrote an independent verification script that recomputes factor RankIC, coverage, and the leaderboard regressions from the raw tables, and used it to check every number in the draft. I wrote the paper. The model training code was produced with AI tools; it was reproduced on Windows and Linux and the 58 submitted models match bit-for-bit.
Manuscript in preparation for a Chinese CSCD journal. The private test period is not disclosed by the organiser, so the paper describes it as an unknown future window. I have not yet rewritten the model code by hand; that is on my list before the paper is submitted.