18+ • Bet Responsibly • Racing Alpha does not accept betsBeGambleAware.org
RacingAlpha

How does AI predict horse races?

Updated

AI predicts horse races by scoring every runner on a learned combination of features — Official Rating, Racing Post Rating, top-speed figure, trainer 14-day form, days since last run, class, going preference, draw, and (for modern models) live market drift. Each horse receives a 0-100 score; the highest-scored horse in races that pass the model's eligibility filter becomes the prediction. The 'AI' part isn't fortune-telling — it's a function that's been calibrated against thousands of past results to find which combinations of signals correlate with winning. Different models weight these inputs differently, which is why running multiple strategies side by side (as Racing Alpha does) is the honest way to evaluate them.

Common questions

What data does an AI horse racing model use?

The core inputs across every credible model are: OR (Official Rating) and RPR (Racing Post Rating) for form; TS (Top Speed) for clock; trainer 14-day strike rate and run count for recent yard form; days-since-last-run for freshness; race class and going for context; draw for flat-race draw bias; and live odds + price drift for market signal. Advanced models add pace inferred from past comments, sire/distance preferences, and headgear changes. Racing Alpha's v1.1 model also includes Betfair Exchange drift signals.

What's the difference between AI rating and a tipster's hand-picking?

An AI rating model assigns a SCORE to every horse in every race, every day, using exactly the same rules every time. A hand-picker chooses a subset of races and applies judgement. The AI's strength is coverage and consistency — it never skips a race, never has an off day, and applies the same logic at 2pm and 7pm. The hand-picker's strength is selectivity — they can choose to pass on races where the data is thin or contradictory. Most modern services combine the two: AI screens; human filters.

Are AI models the same as 'machine learning' models?

Roughly yes, in the way the public uses the term. 'AI' is a marketing umbrella; under the hood you'll usually find one of three approaches: (a) heuristic weighted-sum (Racing Alpha v1.0 — explainable, every score breaks into its components); (b) gradient-boosted trees like XGBoost or CatBoost (Racing Alpha v1.1 — handles non-linear interactions, harder to interpret); (c) deep learning / neural networks (rare in horse racing because samples are small and noisy). Boosted trees currently produce the best results on UK/IRE racing data.

How does the AI handle small fields or weak races?

Most good models apply a 'segment filter' — they refuse to pick in race types where the model has historically not produced positive ROI. Common exclusions: very small fields (5 or fewer runners — narrow market, little edge), very young horses (2yo races where form is thin), very long-odds picks (>20/1 where variance overwhelms any edge). Racing Alpha's v1.3 filter is published in /methodology and is part of why the v1.0 advised tip is a subset of the field-wide top-rated runners.

Why don't all AI horse racing models agree?

Because they weight inputs differently. A model that heavily weights recent form will look at one horse and rate it high; a model that weights speed figures will pick a different horse. Both can be right on different races. The honest way to evaluate which is better is to run them side by side over a long sample. Racing Alpha publishes three (v1.0 heuristic, v1.1 ML, Furlong manual) on a public head-to-head dashboard so the punter can see which model's read is currently delivering.

Racing Alpha is for information and entertainment for adults aged 18 and over. We do not accept bets. Past performance does not predict future results. BeGambleAware.org