Before a race, a bookmaker turns form into a price. We built models that do the same job from form alone, looked at what they lean on, and measured how close they get to the final market. They get close. They do not get past it.
Ben Dunlop · Published 10 November 2026
A price starts as a chance for every runner, worked out from form: who rides it, who trains it, how fast it has run, what class it has been racing in. The chances add to 100%. This is the step our form-only models do.
The bookmaker shortens every price so the book adds to more than 100%. Before the 2026 Melbourne Cup, a major bookmaker's fixed-odds book on our card added to 119.7%. The extra 19.7% is the margin.
Then the bets come in. Backed horses shorten, the rest drift, and bookmakers keep an eye on the Betfair exchange, where punters bet against each other. Point Of Law was $8 on our card and started at $7. Betfair's starting price, at the jump, is the market's last word.
Our newest form-only price comes from a neural network that never sees the market. It reads a few hundred facts about each runner, keeps a memory of each horse's last ten runs, and weighs every runner against its rivals in the race.
To see what it relies on, we took one kind of information at a time and shuffled it between the horses in each race, so one runner got another's jockey figures, say. Then we measured how much worse the price got at picking winners. The bigger the drop, the more the network leans on that information. Tested on 10,875 races it was not trained on.
How much worse the price gets without it. Biggest = 100.
Three things to read into it, and one not to.
Another way to look inside: take one horse in each race, change its last run in the network's memory, leave its rivals as they were, and see how far its win chance moves. We did this to 10,882 horses.
| If its last run had been | All horses | Favourites | Outsiders |
|---|---|---|---|
| Won its last start | +2.2 | +3.4 | +1.2 |
| Beaten 3 lengths, mid-field | −0.3 | −2.4 | +0.2 |
| Ran last, beaten 10 lengths | −2.0 | −6.7 | −0.4 |
| Last start at a city track | +0.3 | +0.8 | +0.1 |
| Last start at a country track | −0.2 | −0.8 | −0.1 |
| Led at the 800 m last start | +0.1 | +0.3 | 0.0 |
| Last run 250 days ago | −0.1 | −0.2 | 0.0 |
Change in win chance, in percentage points: +2.2 takes a 10% chance to 12.2%.
A win last start is worth about +2.2 points, roughly a third more chance than the horse had. A favourite that ran last, beaten ten lengths, loses 6.7 points. Where it raced counts a little. How long ago it raced barely counts at all.
Only the memory of that one run changed. The horse's summary figures, such as its average finishing spot over five runs, stayed as they were, so these numbers are what the run adds on top of them.
The Melbourne Cup at Flemington on Tuesday 3 November 2026: 23 runners over 3,200 m on a Soft 5 track. The prices are a major bookmaker's fixed odds on our card before the race, and the book added to 119.7%. The Cup is free on RaceAlpha before and after it runs, so every runner is shown with our numbers; the full card and our read are on the Melbourne Cup page.
The bookmaker's price with the margin taken out, so the field adds to 100%.
Our model with no market input, the Form only sort on the card.
The card's number: form, calibrated against results and the market.
| Runner | Market | Form only | Win chance | Finished |
|---|---|---|---|---|
| 12Hopewell Rock$7.50 | 11.1% | 3.6% | 11.4% | 9th |
| 20Point Of Law$8 | 10.4% | 9.7% | 11.2% | 1st |
| 2Trustyourinstinct$8.50 | 9.8% | 3.1% | 9.4% | 11th |
| 24Shockletz$10 | 8.4% | 6.2% | 8.6% | 6th |
| 16So You Are$12 | 7.0% | 11.2% | 7.6% | 8th |
| 11Juja Kibo$13 | 6.4% | 10.7% | 6.7% | 3rd |
| 4Middle Earth$17 | 4.9% | 3.6% | 5.5% | 10th |
| 10Goodie Two Shoes$19 | 4.4% | 1.7% | 4.6% | 22nd |
| 1Sons And Lovers$21 | 4.0% | 1.7% | 3.4% | 4th |
| 18James J Braddock$21 | 4.0% | 3.1% | 4.3% | 7th |
| 6Enfjaar$26 | 3.2% | 3.6% | 2.9% | 15th |
| 14Soul Of Spain$26 | 3.2% | 2.0% | 3.3% | 13th |
| 22Mr Monaco$26 | 3.2% | 8.3% | 2.9% | 17th |
| 23Zakouma$27 | 3.1% | 4.8% | 2.3% | 14th |
| 15Zahrann$34 | 2.5% | 4.8% | 2.3% | 2nd |
| 3Knight's Choice$35 | 2.4% | 2.9% | 2.1% | 12th |
| 8Wine Dark Sea$35 | 2.4% | 3.1% | 2.3% | 19th |
| 19Kizlyar$41 | 2.0% | 2.0% | 1.9% | 16th |
| 7Royal Supremacy$46 | 1.8% | 4.8% | 1.7% | 5th |
| 5Vauban$51 | 1.6% | 3.6% | 1.6% | 23rd |
| 17Future History$51 | 1.6% | 3.6% | 1.7% | 20th |
| 13Onesmoothoperator$67 | 1.2% | <0.1% | 1.1% | 21st |
| 21Strictly Business$67 | 1.2% | 2.0% | 1.2% | 18th |
| Field | 100% | 100% | 100% |
The price under each name is the bookmaker's fixed price.
Form alone was close to the market on the winner, Point Of Law: 9.7% against 10.4%. It was well short of the market on the favourite, Hopewell Rock, at 3.6% against 11.1%, and Hopewell Rock ran 9th. The runner it liked most against the market was Mr Monaco, 8.3% against 3.2%, and it ran 17th. One race proves nothing either way. Across thousands of races, the market is right more often.
The card's win chance starts from what the market knows, so it stays close to it.
Above the line, form liked a horse more than the market did; below it, less. Both scales are stretched at the bottom so long shots can be read.
Average distance from the market's chance, every runner in the Cup.
We scored four prices on 79,964 Australian races from January 2022 to October 2026: every race where all the runners had a Betfair price. Each form model was trained only on races before the six-week block it was tested on, the way it would have been used at the time.
The test is simple. How much chance did each price give the horse that went on to win? A price that gives winners more of its 100% is the better forecast.
Chance given to the winner on a typical race. The average punishes confident misses, the standard way forecasts are scored.
Betfair's starting price gave the winner 17.0% on a typical race. Our newest form-only price gave it 14.8%. Put another way: for every 100 our price gave the winners, the market gave 115.
| Year | Races | Market, for every 100 of ours |
|---|---|---|
| 2022 | 16,125 | 114 |
| 2023 | 16,959 | 115 |
| 2024 | 16,763 | 115 |
| 2025 | 16,662 | 115 |
| 2026 to Oct | 13,455 | 115 |
The gap barely moves from year to year. Across 42 six-week blocks it ran from 112 to 117, and the market was ahead in every one. The newest price did beat our earlier one in all 42, giving winners about 1% more chance. Progress, but on our side of the line.
If our price knew something the market missed, mixing a little of it into Betfair's price would help. So we let a fitting procedure choose the mix, one year at a time: fitted on the other years, checked on the year held out.
| Year held out | Weight on our price | Chance on winners |
|---|---|---|
| 2022 | −0.05 | −0.04% |
| 2023 | −0.03 | +0.01% |
| 2024 | −0.02 | +0.03% |
| 2025 | −0.04 | no change |
| 2026 | −0.03 | +0.02% |
A weight of 0 means ignore our price. A model that knew something extra would earn a clearly positive weight.
Every year it put a weight of just under zero on our price: in effect, none. The mix moved the market's chance on winners by about a hundredth of one percent. The final market already holds what these inputs know.
The other way round is to start with the market's chance and let form nudge it. Our all-inputs model does that. It is being tested on paper and is not on the card, and we make no claim for it until it has a record.
Its weight is spread thin. No single input carries more than 1.7% of its decisions, and the top ten together carry 14%. By kind of information:
Share of the model's decisions. Adds to 100%.
Past-run speed and sectionals carry a third of its weight. It reads an older list of inputs than the network above, without prize money, the form curve or stewards' reports, so the two orders are not a like-for-like comparison. And its nudges are small. Across all 152 runners at Mornington and Queanbeyan on 10 November, it sat 0.4 points from the market's chance on average. The card's win chance sat 0.6 points away, and form only 3.7.
Betfair's starting price beat every form-only price we built, in every year and every six-week block.
It shows why a horse is the price it is, and where form and money part company. That is what the Form only sort on the card is for.
None of this is a reason to bet. Plays are chosen by a separate rule, and every one is on the results page, losers included. The model's own record is on ratings accuracy.
If gambling is causing you or someone close to you a problem, call Gambling Help on 1800 858 858 (Australia) or visit gamblinghelponline.org.au.
Broadly, yes: a rated price from form first, a margin on top, then the money and the exchange move it.
Each bookmaker does it differently. Some frame their own opening market from ratings, some buy ratings in, and many follow a few leading markets closely once betting starts. The order holds either way: form sets the first price, the margin is added, and by the jump the price is mostly the money.
The extra over 100% is the bookmaker's margin.
Every price is shortened a little, so that over time more is taken in bets than is paid out. Before the 2026 Melbourne Cup, the fixed-odds book on our card added to 119.7%, and the starting prices added to 117.6%. Dividing each runner's share by the book's total takes the margin out and gives a market chance that adds to 100%.
No. Our best form-only price is behind the final market every year, and mixing it into Betfair's price adds nothing measurable.
This page is about how a price is made, not a reason to bet. Plays on RaceAlpha are chosen by a separate rule, logged before the race, and graded after it on the public results page, losers included.
The price Betfair sets when the race starts, by matching the starting-price bets placed on its exchange.
On an exchange punters bet against each other rather than against a bookmaker, and the starting price is set at the jump, after all the money is in. That is why we score every price against it: it is the closest thing to the market's final word.
The page shows the kinds of information and how much each one counts, not the recipe.
The families on this page cover what the models lean on. The full list of inputs and how each is built is the work behind the ratings, so it stays in-house.
