Every race. Every result. Published.
Historical walk-forward backtest across 15,048 Australian flat races, measured against the market favourite. Price-provenance revalidation is pending.
Model Top Pick
31.1%
15,048 races
Market Favourite
34.7%
Same races, same period
Gap vs Favourite
−3.6
Points. We are behind.
Flat-Stake ROI
-17.1%
Backing every top pick blind
Read this before you read anything else on the site.
The published historical test does not show the model beating the market. Over 15,048 back-tested races the starting-price favourite won more often than our top-rated runner, and flat staking lost money. A missing-odds fallback in that test could have used post-race prices, so the figures need a clean point-in-time rerun before being treated as validated proof.
What FormRace is actually for: reading 94 model inputs per runner across AU flat meetings, surfacing races where price and probability disagree, and settling calls in public afterwards. No betting edge is established here. The judgement stays yours.
We previously published a “+24.1% ROI over 2,798 races” figure. An internal audit found it was produced by a broken backtest and it has been removed from the product. The numbers on this page replace it.
Live — settled races
Verified rolling statistics are temporarily unavailable.
Top-pick strike rate vs the market favourite
% of races won · same races both seriesWhat backing every top pick would have returned
Flat $1 win bet · % return on turnoverWalk-forward test, Nov 2025 – Jul 2026. The favourite baseline is measured on the same races. Every value is also in the table below, so nothing here is gated behind a hover.
Month by month
Nov 2025 – Jul 2026| Month | Races | Model Top Pick | Favourite | Top-3 Place | Flat ROI |
|---|---|---|---|---|---|
| Nov 2025 | 1,996 | 32.8% | 35.5% | 66.8% | -11.1% |
| Dec 2025 | 1,738 | 33.4% | 33.4% | 67.2% | -15.0% |
| Jan 2026 | 1,567 | 35.2% | 36.1% | 68.8% | -8.7% |
| Feb 2026 | 1,558 | 35.2% | 36.9% | 68.7% | -11.9% |
| Mar 2026 | 1,788 | 32.1% | 36.0% | 62.5% | -14.2% |
| Apr 2026 | 1,677 | 25.3% | 34.6% | 55.3% | -35.2% |
| May 2026 | 1,736 | 26.4% | 32.3% | 57.0% | -27.3% |
| Jun 2026 | 1,612 | 26.7% | 33.8% | 54.5% | -18.1% |
| Jul 2026 | 1,376 | 33.1% | 34.1% | 65.0% | -10.2% |
Model top-pick and favourite strike rates are computed on the same races. Flat ROI is a $1 win bet on the model's top pick in every race at the settled price. Green means the model beat the favourite that month.
Top-3 place rate
62.9%
How often the model's top-rated runner finishes in the first three. Useful for narrowing a field before you form your own view.
Top-2 contains winner
50.2%
The winner is inside the model's top two runners in about half of all races — the strongest thing the ratings do.
Information vs market
-0.0109
Δη²: how much the model's probabilities explain race outcomes beyond the public odds. Negative means the market alone is still better informed. We track it publicly.
Methodology
The default XGBoost ranker and classifier use 94 inputs across 20 groups, including market price, pace projection, class, barrier, weight, ratings, jockey and trainer signals, and race context. Probabilities are isotonic-calibrated and blended with the public market before ratings are served.
Test months were held out chronologically and the favourite baseline uses the same race set. That protects against training on a race result directly, but the historical dataset could fill a missing pre-race price from starting or closing odds. This fallback is now disabled for new training data; these older figures still need to be rerun.
Every figure on this page is generated directly from the backtest report file — there are no hand-typed performance numbers on the site. Historical racing performance does not predict future results, and nothing here is betting advice.
See today's board — free
Start with the free daily briefing and the public receipts. Pro unlocks the full AU thoroughbred board — A$49/month, cancel anytime.