This is the model, not the plays. Every race RaceAlpha has rated since 1 January 2026, scored against the result after it ran, whether we played the race or not. The plays are a small selection from these ratings and their record is on the results page.
The top-rated runner won 30% of its races and the market favourite 34%; they were the same horse 70% of the time. The winner came from our top three 65% of the time in fields averaging 10.2 runners. Backing every top-rated runner at starting price would have returned −14% of turnover; backing every favourite −10%. That gap is the market's margin plus the model's shortfall against it, and it is why a RaceAlpha play is never simply the top pick.
Runners are grouped by the win chance the model gave them before the race. In a well-calibrated model the black and green bars match. They do up to about 30%: runners given 10 to 15% won 13.4%, runners given 20 to 30% won 24.2%. Above that the model runs hot, and at the very top, 527 runners given 55% or more won 53.1%. The grey bars sit higher because a bookmaker's percentages add to more than 100.
| Stated chance | Runners | Model said | They won |
|---|---|---|---|
| under 3% | 16,884 | 1.6% | 1.8% |
| 3–6% | 13,431 | 4.4% | 4.6% |
| 6–10% | 9,506 | 7.7% | 8.5% |
| 10–15% | 7,211 | 12.4% | 13.4% |
| 15–20% | 4,897 | 17.3% | 18.6% |
| 20–30% | 4,152 | 24.2% | 24.2% |
| 30–40% | 1,661 | 34.4% | 30.8% |
| 40–55% | 1,062 | 45.7% | 36.7% |
| 55%+ | 527 | 67.6% | 53.1% |
When the model and the market name different horses, the market has been right more often. That is the honest state of the ratings on their own, and the reason the plays are chosen by patterns that beat the market on unseen races rather than by rank.
| Top pick given | Races | It won | Favourite won | At SP |
|---|---|---|---|---|
| Under 20% | 596 | 18% | 24% | −22% |
| 20 to 30% | 2,172 | 25% | 30% | −15% |
| 30 to 40% | 1,446 | 31% | 34% | −14% |
| 40 to 50% | 866 | 35% | 34% | −14% |
| 50% and over | 716 | 50% | 51% | −7% |
Confidence is earned at the top: when the model gives its pick 50% and over, it wins 50% of the time. Those are short-priced horses the market already likes, so the return is still under water at SP.
| Model rank | Runners | Won | Model said | At SP |
|---|---|---|---|---|
| 1st | 5,796 | 30.2% | 34.3% | −14% |
| 2nd | 5,722 | 19.7% | 18.8% | −16% |
| 3rd | 5,722 | 15.4% | 13.2% | −13% |
| 4th | 5,713 | 10.4% | 9.7% | −24% |
| 5th | 5,678 | 7.1% | 7.1% | −33% |
| 6th | 5,613 | 6.1% | 5.3% | −25% |
| 7th | 5,377 | 4.2% | 4.1% | −37% |
| 8th | 4,991 | 3.7% | 3.2% | −27% |
| 9th and below | 14,719 | 2.0% | 2% | −37% |
The order holds all the way down: each rank wins less often than the one above it. That is what a rating is for. Whether any rank is worth backing is a separate question, answered by the price, and the SP column says the answer is no for every rank on its own.
| Month | Races | Top-rated won | Favourite won |
|---|---|---|---|
| October 2026 | 996 | 34% | 34% |
| September 2026 | 604 | 29% | 33% |
| August 2026 | 557 | 29% | 32% |
| July 2026 | 519 | 28% | 32% |
| June 2026 | 482 | 26% | 31% |
| May 2026 | 589 | 28% | 31% |
| April 2026 | 592 | 30% | 33% |
| March 2026 | 579 | 32% | 38% |
| February 2026 | 498 | 32% | 37% |
| January 2026 | 380 | 30% | 35% |
| Races | Top-rated won | Winner in our top 3 | Favourite won | |
|---|---|---|---|---|
| Australia | 5,125 | 31% | 66% | 34% |
| Hong Kong | 671 | 26% | 56% | 29% |
Rows under 50 races are faded. Small fields are easier for everyone, the model included; the gap to the favourite is what to watch.
The results page grades the plays: the handful of runners we told you to back, logged with a price before the race. This page grades the ratings: every runner in every race the model rated, 59,331 of them since 1 January 2026, whether we played the race or not. Most races have no play, so the two records answer different questions. Results asks whether the plays made money. This page asks whether the model knows what it is talking about.
Two reasons. The market keeps a margin, so backing every favourite loses too (−10% of turnover here). And the model's top pick wins a little less often than the favourite. A rating is a chance, not a bet. A play needs more than a high rating: it needs a price that is wrong, and a pattern that stayed profitable on races the model had never seen. That filter is what the results page records.
73 races whose ratings were written or changed after race day, because a rating that could have seen the result proves nothing, and 71 races where the top-rated runner was scratched. Ratings from 2025 are left out as a block: a February 2026 backfill rewrote most of them, so they cannot be shown to be pre-race.
When the model gives a runner a 10% chance, about one in ten of those runners should win. The chart checks that band by band. It holds well up to about 30% and runs hot above it: runners given 55% or more won 53.1%. The market's implied chances sit higher than both because a bookmaker's percentages add to more than 100.
Because the alternative is asking you to take the word "AI" on trust. The model is the engine under every play, and the honest way to sell it is to show its error rate and where the market still beats it. The numbers update with every build; nothing is cherry-picked by date or track.
Computed 1 November 2026 from ratings stored before each race. 73 races with ratings touched after race day and 71 with a scratched top pick are excluded.
