Research & education only — not betting advice. 18+.
Most "prediction" content in horse racing is storytelling. We wanted a system that answers a narrow, testable question instead:
When our statistical model disagrees with the market about a horse, and the market then moves toward our model's view before the race — can we detect that in near real time?
That detection layer is what we call the cold-horse radar (冷門雷達). This post is the engineering breakdown.
Signal 1: steam moves (落飛)
A steam move is a significant odds drop in a short window — e.g. a horse opening at 29.0 and steaming to 19.0 (a -34.5% move). The implied probability moves from ~3.4% to ~5.3%.
Implementation notes:
- Poll the public odds board at fixed intervals; snapshot everything
- Compute
delta = (odds_now - odds_open) / odds_open - Threshold: flag at |delta| >= 20% — below that is noise in the pools we watch
- Direction matters: a drop (落飛) means money arrived; a drift (升飛) means the market is cooling
Signal 2: model-vs-market divergence
Our model (gradient boosting over ratings, speed figures, going, draw, weight, jockey/trainer features) outputs a calibrated win probability. The market's implied probability comes from overround-removed odds.
divergence = model_probability - market_implied_probability
We flag when divergence >= 6 percentage points and implied probability <= 22% (roughly 7/1 or longer). The 6pp threshold was tuned down from 8pp after backtesting showed the stricter filter starved the radar of signals on ordinary race days.
Why calibration is the whole game
A divergence signal is only meaningful if your model's "22%" actually means 22%. We verify with:
- Brier score per meeting and per season (public on our blog)
- Calibration curves — predicted probability buckets vs realized frequency
- Time-series split validation — always train on the past, validate on the future
If calibration drifts, we widen the divergence threshold automatically. A miscalibrated radar is worse than no radar — it manufactures false confidence.
What we publish
Every signal, hit or miss, goes into an open scorecard on the site. Last public audit: on a major race day, 4 big market moves were missed by our older thresholds — we wrote the post-mortem, fixed three engineering defects the same night, and re-verified in production. Honest losses are the only kind of track record worth having.
Full methodology series (Cantonese, Japanese and English) lives at mystique-racing.com/blog — newest entry: Ratings, Speed Figures and Market Odds: A Map for Racing Data Science.
Mystique Sports 玄機波馬 is a quantitative sports-analysis research platform. Everything above is reproducible methodology, not tipping.
Top comments (0)