A model that shows its working
No black box. Every rating decomposes into the same handful of signals, and you can see which ones moved each runner.
The data
PuntPulse runs on a continuously-updated database of Australian greyhound racing covering 2010–2026 — every meeting, race, runner, sectional split and past-start form line from public results.
- 736,041 races and 5,860,297 individual runner performances
- Box statistics, distance-band records and best times for 165,904 greyhounds
- Race-day expert-form snapshots, so a historical race is always rated on information available at the time
Fields and results are picked up as they publish, so cards are live well before the jump and results-reviewed the same night.
The signals pulse-0.6
Most influential first — exact weighting isn't published since it shifts as the model is retrained.
Recent form
Recency-weighted placings over the last six starts (from the dog's own race history), normalised to field size, with a softener for beaten margins under two lengths.
Class & strike rate
Career win and place rates — the race-day expert-form snapshot where we have it, otherwise built from the dog's full result history.
Early speed
Average lead position at the first marker across recent runs. Box-to-first-turn speed is decisive over sprint trips.
Best-time credentials
The dog's best time at the distance versus the rest of the field, scaled across the range on offer.
Box suitability
The dog's box-by-box win and place record, blended with a track rail-bias prior — always counts, so a lightly-raced field never collapses to a flat line.
Market read
The overround-adjusted market price as a relative signal. When our own form is thin the final number leans further toward the market; with a full set of form it keeps its independent read.
From signals to probabilities
- Each signal becomes a 0–1 sub-score. Where a runner is missing data for a signal, its weight is redistributed across the signals we do have — a lightly-raced dog isn't punished for an absent career record, it's just rated on less.
- The weighted blend is the runner's rating (shown 20–100 within a race).
- A softmax over the field turns ratings into win probabilities that sum to one. Place probability is a calibrated inflation of the win number.
- The result is then blended toward the overround-adjusted market price, weighted by how much form we had: with little to go on we mostly follow the market (no false edge); with a full set of form the model keeps its own read.
- Each final win probability is turned into a fair price and compared to the price on offer. Expected value above break-even after the bookmaker's margin gets a value flag and a fractional-Kelly stake.
- Field entropy and the gap between the top two picks set the confidence read for the race.
What this is not
The ratings are statistical estimates, not predictions of certainty and not financial advice. Greyhound racing has genuine race-to-race randomness — interference, boxing, track pattern — that no model removes. A 35% top pick means the model expects that runner to lose roughly two times in three. Stake accordingly, only ever with money you can afford to lose.
Read tonight's card with the model beside you
Free to start. Three races a day, full probabilities and the top pick's reasoning — no card required.