Horse racing as an ML ranking problem: 1.18M runners, model AUC 0.729 still trails the 0.790 market baseline
gcampb41 · reddit · 2026-09-14
Inspired by Bill Benter's Hong Kong racing models, the author built Hoofs, a British & Irish horse racing ML project:
- Data: 1.18M historical runner records over ten years, with a unified feature bank of 1,700 candidate signals per runner, plus fairly novel proprietary data
- Modeling: runner-level win/place probability models ranked within races, plus a race-level confidence model (field size, probability concentration, entropy); public Top 1–3 rankings are deliberately market-agnostic
- Validation: strict chronological walk-forward, out-of-fold calibration, explicit future-leak checks; tracked via AUC, log loss, Brier and racing-style strike rates
- Results: on a 2018–2025 benchmark (886K runners, 94K races), model-only win AUC ≈0.729 / place ≈0.708 vs market-only ≈0.790 / ≈0.762. The strong market baseline was the hardest part — extracting information not already in the price; the author finds positive EV mainly before the market fully forms
- Engineering story: live strike-rate degradation exposed data gaps, triggering a full rebuild of raw datasets, feature bank, chronological lineage and model families. First live day after rebuild: Top 1 hit 10 of 23 races (43.5%), Top 1–3 covered 16 of 24
In practice the author uses daily reports as the first analysis layer and market data as the second layer for actual betting decisions.
More from Research
- Prof. Tom Yeh Explains Two-Tower MLP by Hand, From Recommenders to CLIP — ProfTomYeh · 2026-09-14
- Agentic paper reproduction is great, but doesn't solve the authorship credit problem, says Dietterich — tdietterich · 2026-09-14
- Hugging Face agents reproduced 2,226 ICML papers — about a third of the conference — mmitchell_ai · 2026-09-14
- Dead fly connectome sim, 25.56M synapses, surfs 41 seconds with balance control — Scobleizer · 2026-09-14
- Szepesvari: knowing NS's limits tells you when simulations stop being reliable — CsabaSzepesvari · 2026-09-14
- SenseNova-U1.5: 8B encoder-free unified model does visual understanding and generation in one — KyeGomezB · 2026-09-14