Three languages, one race
Every serious approach to horse racing analysis speaks one of three dialects:
- Ratings (Elo-style): each horse carries a single number; beating a rival moves both numbers in opposite directions
- Speed figures: each past performance is converted into one time-based number, adjusted for track, distance and condition
- Market probability: the odds, cleaned of the bookmaker's margin, read as the crowd's probability estimate
They describe the same reality with different strengths. The craft is in the translation.
Language 1: Elo-style ratings
Borrowed from chess, the idea is simple:
- Every horse starts at a baseline rating
- After each race, ratings update based on who beat whom, scaled by margin and strength of opposition
- A rating difference maps to an expected win probability via a logistic curve
Strengths: self-correcting, handles fields of different sizes naturally, needs no track-specific normalization. Weaknesses: slow to react to rapidly improving young horses; says nothing about why (distance, going, draw).
A practical refinement is decay: weight recent form more heavily (our own system uses EMA-style adaptation — see the EMA article), and context flags: only update fully when the horse ran under comparable conditions.
Language 2: speed figures
Speed figures answer: "How fast did this horse actually run, once we strip out the context?"
The pipeline:
- Raw finishing time for the race
- Normalize against the course and distance (a 1200m time at Sha Tin means nothing next to a 1200m time at Happy Valley)
- Adjust for going — a good-to-firm day makes every raw time look better; the standard fix is a daily track variant estimated from all races that day
- Adjust for weight carried, on a researched (not guessed) pounds-per-length scale
- Optionally adjust for trouble in running (wide trips, blocked runs) — powerful but subjective
Strengths: directly comparable across horses, courses and seasons. Weaknesses: the adjustments ARE the model. Garbage variant estimates in, garbage figures out. This is why honest pipelines publish their variant methodology.
Language 3: market-implied probability
Decimal odds (O) imply probability (1/O). Summed across a field, this exceeds 1 — the excess is the overround (the operator's margin). To read the market honestly, remove it.
Two standard methods:
- Proportional (basic): divide each implied probability by the total. Simple, but known to distort — longshots carry disproportionate margin (the favourite–longshot bias)
- Power method: find an exponent (k) such that (\sum (1/O_i)^k = 1). Better behaved across odds ranges and easy to compute
What remains is the crowd's probability estimate — the strongest single predictor available, and the benchmark any model must be compared against.
The synthesis: calibration and divergence
A model that outputs "Horse A will win" is entertainment. A model that outputs "Horse A wins 22% of the time" is testable.
The workflow we follow at Mystique:
- Build probabilities from ratings + speed figures + conditions (gradient boosting on tabular features works well)
- Calibrate them (Platt scaling / isotonic regression) on a strictly future validation window, verified with Brier score and calibration curves
- Compare against market probabilities from step 3 above: the divergence between model and market is where research questions live
- Audit over time: does the model's divergence at race time predict anything about how odds move later? This is the same idea as closing line value — a process metric, not a profit promise
What this map is for
None of these three languages is "the secret". The market is extremely good; beating it consistently is genuinely hard, and anyone promising otherwise is selling something. The point of building your own ratings and figures is understanding: knowing when a horse's price reflects its form, when it reflects a narrative, and being able to say why in numbers.
That is the standard of every article in this series: method public, results reproducible.
This article is for research and education only and is not betting advice. Horse racing involves risk. 18+ (or the legal age in your jurisdiction).