A LambdaRank ranker, three filters, no magic.
The pick selection model is a LightGBM LambdaRank trained on multi-year North American data. Inputs: morning-line, post position, distance, surface, recent form figures, trainer + jockey rolling stats, sire-distance fit, class. Output: a per-runner ranking that orders the field by predicted finish.
A runner only surfaces on the published card when the qualified-plays filter also fires — top-tier trainer and jockey on a rolling 30-day window, fair-value odds (morning line 4–1 to 12–1), and a race-shape that’s not flat. Public favorites pay nothing on the signal; the band is where the math actually works.
The 12–15% WIN ROI in the 90-day backtest is descriptive of past performance, not predictive — and we say that everywhere it appears.
— Methodology, /performance