AI that explains every rating
PuntPulse runs two machine-learning models — one for greyhounds, one for thoroughbreds. Every rating comes with the factors that drove it, so you're never asked to trust a black box.
The data
The models learn from a continuously-updated database of Australian and UK racing covering 2010–2026.
- 1,533,204 races across 137,291 meetings in the database (the models rate Australian races, plus UK greyhounds and UK and Irish gallops)
- Form for 249,231 greyhounds in Australia, New Zealand and the UK, and the horses running at Australian TAB, UK and Irish meetings
- Every signal is built as of race day: the model only ever sees what was known before the jump — no hindsight.
Fields, prices and results are picked up as they publish, so cards are live well before the jump and reviewed once the results land.
What the greyhounds AI looks at pulse-ml-7
Form
How each runner has been performing recently, compared with the rest of the field.
Class
The level a runner has been racing at, and how today's race compares.
Speed
How quickly a runner has run, and how it tends to race.
Draw & conditions
Where a runner starts from, and the track and distance.
Connections & breeding
The people and bloodlines behind a runner.
What the gallops AI looks at pulse-gallops-mkt-3
Form
How each runner has been performing recently, compared with the rest of the field.
Class & weight
The level a runner has been racing at, and what it carries today.
Speed
How quickly a runner has run, and how it tends to race.
Draw & conditions
Where a runner starts from, the track, the distance and the going.
Connections
The people behind a runner.
The market
What the betting market thinks, and how that view is changing.
These are the broad groups. The individual inputs and their weighting aren't published, and they change as the models are retrained.
From form to probabilities
- Every runner is described by a set of numbers, each built only from information available before the race.
- A gradient-boosted machine-learning model — trained on hundreds of thousands of past runs — learns how those signals combine, including interactions a hand-set formula would miss (a wide draw matters more over a short trip, for example).
- Its scores become win probabilities that sum to one across the field, with place chances derived from them.
- The market already knows a lot, so both models are built on top of it. Each starts from the overround-adjusted market price, corrected for the market's known lean (long shots win less often than their price implies, favourites more), and is trained to learn only what that price misses. The adjustment is small — the market is a strong guide — and a race has no rating until the market has priced every runner.
- Every model is tested walk-forward: retrained month by month on the past only, then scored on races it has never seen — the honest test of a model.
- Each win chance becomes a fair price and is compared with the price on offer. Where the price on offer is longer than the model's fair price, the card shows the gap as an estimated edge, and a Kelly calculator works out what that estimate would imply for a bank you enter. An estimated edge is not a proven one: the models haven't been shown to beat the market after bookmaker margins.
- The “why” on every runner comes from the model itself: how much each factor pushed that runner's chance above or below what its price implies.
What this is not
The ratings are statistical estimates — not certainties, not tips and not betting advice — and they don't make betting profitable. Racing has genuine race-to-race randomness — interference, a slow start, a wide run — that no model, AI or otherwise, removes. A 35% top pick means the model expects that runner to lose roughly two times in three. If you bet, only ever use money you can afford to lose.
Read tonight's card with the model beside you
Free to start: the model's top three runners in every race, with the reasoning — no card required. 18+ only.