How we score every horse, every race.
A flagship model and a set of benchmark AI variants run on every UK & Irish race, each with a different scoring hypothesis. Pick the one you trust. Track records are public. Everything is free.
Our stance
What we are
- An independent racing research site for UK & Irish racing
- Open methodology — we publish how every score is built
- Independent P&L tracking for every variant, every day
- Free to use — no paywall, no ads, no affiliate links
What we aren't
- A tipping service — we don't tell you what to back
- A guaranteed-profit promise — the variants lose money some months
- A black box — what goes into each score is set out below, and every model's record is public
- Selling anything — Telegram is free, the site is free
Every AI scoring variant we run
Three of these can power the AI Score column on your racecards. The rest are scored on every race, every day, as tracked benchmarks — published here so you can see what we compared against and how it did.
How the default AI Score is built
The AI Score you see by default comes from our flagship model, The Racing Formula: three gradient-boosted models — one each for flat turf, all-weather and jumps — retrained every morning on years of UK and Irish results. Its output is a ranking of each horse's chance in this race, not a promise: a high score means "well fancied on our read," not "will win."
What goes in
Well over a hundred inputs per model: ratings and handicap marks, recent form, class, course, distance and going records, speed figures from past races, trainer and jockey form, age, weight and freshness, field strength and, on the flat, the draw, early pace and pedigree. The models trained on 25 October 2026 used 141 inputs for flat turf, 133 for all-weather and 124 for jumps; the count moves a little as inputs are added or retired.
What comes out
An estimate of each runner's chance of winning, calibrated against how often horses given similar estimates actually won.
The 0–100 score
That estimate is scaled so the field spreads out and reads as a ranking. The score is not a percentage: a typical race's top pick scores in the high 60s.
Where prices come in
The runner's own price is not an input to the model. Two inputs do draw on past prices: how often each trainer's and jockey's earlier runners won compared with what the market expected of them. After the model has scored the field, the morning price is used to rein in long-priced runners: the estimate for any runner priced longer than 19/1 is capped; one longer than 49/1 without strong recent form, a high rating or a high-strike-rate trainer has its estimate cut to a fifth; and a runner longer than 19/1 that tops its race is moved just below the best-rated runner shorter than 7/1 unless it is rated clearly above it. The Top pick is chosen only from runners priced between 1/2 and 99/1. Live price moves are never used.
When it appears
The flagship's scores are added later in the morning. Before that, and in any race it has not scored in full, the card uses the Baseline score below — so every runner in a race is always ranked by the same model.
How well it ranks runners, band by band, is on the performance page — along with the honest finding that it tracks the market rather than beating it. For a plain-English overview of what the scores can and can't tell you, see our AI horse racing predictor.
How the Baseline score is built
Nothing here is a black box. The Baseline is a transparent formula: every runner's 0–100 score is built from these signals, weighted and combined. It is the benchmark every lettered variant modifies, and the score the card falls back to when the flagship model has not scored a race. Like the default score it ranks each horse's chance in this race, and it reads the morning price only — never live exchange moves. See our honest calibration curve for why ranking isn't the same as profit.
Recent form
Finishing positions over the last five runs, weighted so more recent runs count for more.
Course & distance record
Strike rate at this course, at this distance, and on today's going.
Class move
Whether the horse is dropping or rising in class versus its last run.
Trainer form
The trainer's win percentage over the last 14 days — hot yards get a boost.
Rating trend
Whether Racing Post Ratings are climbing over recent starts (an improving profile).
Handicap angle
The gap between the horse's best recent rating and its official mark — a "well-in" signal.
Market read
The morning price, as a sanity check on the fundamentals. Never the live exchange — that would let the market write our score.
We publish this because we think you should be able to see exactly what drives a score. It also keeps us honest: a score you can inspect is a score we can be held to.
What kind of number is it?
Not every figure on the site is the same kind of thing. Some are rankings from a model, some are simple calculations, some are historical records and some come straight from the betting market. Knowing which is which tells you how far to trust it.
| Number | What it is | Built from | Updated | Don't lean on it alone when… |
|---|---|---|---|---|
| AI Score (0–100) | Model-derived ranking of the runners in one race. Not a probability. | By default, the flagship model: form, ratings, class, conditions records, speed figures, trainer and jockey form and, on the flat, the draw, early pace and pedigree. The runner's own price is not a model input; the morning price reins in long-priced runners afterwards, and live price moves are never used. How it is built | Each morning — flagship scores land later in the morning, with the Baseline shown until then | …comparing runners in different races, or reading a score as a chance of winning. See the reliability curve. |
| Pace verdict and pace map | Calculated from past running lines | Each runner's running style over its last ten races | With the racecard | …the field is full of lightly raced horses, or tactics are likely to change. Pace maps guide |
| Draw statistics | Historical win rates | Results since 2015, grouped by the low, middle and high third of each field, by course and distance band | Weekly | …the runs behind a figure are few, or the ground is unusual. Raw percentages carry a small field-size tilt. Draw bias guide |
| Steamers and drifters | Market-derived fact | Morning price against the live Betfair Exchange price | Every 30 seconds on the racecards | …the market is thin, or the whole race has moved. A move is not a verdict. Market movers guide |
| Value board "fair" price | Market-derived calculation | The Betfair Exchange price with the margin stripped out, compared with bookmaker prices | Through the day | …you treat it as a model's view. It is arithmetic between two markets, and unvalidated. Fair odds guide |
| Trainer and jockey strike rates | Historical record | Winners from runners over 14 days, 30 days and a career, with minimum samples for the longer splits | Daily | …the figure rests on a handful of runners. Trainer form guide |
| OR, RPR and TS | Not calculated by us — the official handicap mark and independent performance and speed ratings | As assessed by the handicapper and the ratings compilers | With the racecard | …a single big figure is doing all the work. Form guide |
Every one of these is information to weigh, not a reason to bet on its own. Our research guides explain each in detail — what it means, how it is built and where it falls short.
The angles we flag
Alongside the score, we tag each runner with the specific angles it triggers — the plain-English reasons it's interesting. These are angles to research, not tips. Tap any one for the full explanation.
Want to see how every variant is testing — win rate, P&L and ROI per model, losing months included?
📊 Open the Transparency pageThe Racing Formula is a horse-racing analysis and statistics website. We don't give betting advice. All performance numbers above are historical and shown only so you can judge the methodology — they're not a forecast of future returns. Gamble responsibly. 18+. BeGambleAware · GamStop.