MarketMoverIs the model's price right?
Every night the model is graded on a week of races it was never trained on. Here is how its prices held up, and what happened to every tip we put out.
Out-of-sample grades of the model in force, refreshed nightly on the latest seven days of settled races, and every published tip against the tipped bookmaker's own last price before the start.
Reading the live record…
Runners the model priced between $2 and $8 an hour out, against what its prices said would win.
Of the betting market's accuracy an hour before the jump, built from public prices and form.
Graded this week across thoroughbreds, greyhounds and harness, none of which the model had seen.
When the model says a runner is a 1 in 4 chance, it wins about 1 in 4 times
Each dot is a group of runners the model priced alike. Across is what its price said the chance of winning was; up is how often they actually won. On the dashed line is exactly right.
Reliability diagram. Runners are binned by the implied price of their win probability; x is the mean predicted probability in the bin, y the observed win rate. Axes are square-root scaled. The market is the de-vigged consensus of every book quoting (power method), graded on the same runners.
Calibration by price band Technical
| Band | Runners | Model expected | Won | Won / expected | z | Market expected | Market won / expected |
|---|
Expected is the sum of predicted win probabilities in the band; z is (won − expected) / √Σp(1−p), approximated as √(expected × (1 − expected / runners)). The model's and the market's bands hold different runners, because each is binned by its own price.
Close to the market's accuracy, and closer as the jump nears
How far the model's prices are from the betting market's at predicting the winner, at each point before the race. Smaller is closer.
Gap in mean log-loss of the winner, (model − market) / market, per checkpoint on the held-out week. The market is the best-known predictor at every checkpoint; the model reaches its accuracy from public prices and form.
Scores by checkpoint Technical
Log-loss is −ln(p) of the winner, averaged over races: lower is better. Uniform is a runner-count guess, the floor any model must beat. Information share is how much of the market's improvement over uniform the model recovers, (uniform − model) / (uniform − market). Brier is the multi-class Brier score over the field.
| Checkpoint | Races | Model | Market | Uniform | Gap | Information share | Brier |
|---|
By code
| Code | Races | Model | Market | Gap | Information share | Brier |
|---|
Night by night
Did the tips close shorter?
Every settled tip, placed by the price we tipped it at and how its price moved at the same bookmaker by the jump. Above the line it shortened; below it drifted.
Each tip against the tipped bookmaker's last recorded win price at or before the advertised start. Move is tipped / close − 1: positive means the book shortened after the tip, so the tipped price beat that book's close.
The record starts on 28 September 2026, when the tips moved to the rule in use today. Every tip since then appears here once its race is official, and none is left out.
The live tips, tested Technical
For the tips in the current filter. Unchanged prices are left out of the sign test.
How far each price moved by the close
The rule the tips run on Technical
As the bot runs it now, read from the live channel.
How the rule was chosen
Every runner of 11,237 settled races from 8 to 27 September was priced at seven points before the jump, each best bookmaker price set against the same book's price at the jump. The rule was chosen on 8 to 20 September and read on 21 to 27 September, which it never saw. Australia and New Zealand, read on the held-out week:
| Variant | Tips | Closed shorter | Drifted | Over the book's close | Level stakes |
|---|---|---|---|---|---|
| The rule | 99 | 61% | 25% | +9.4% ± 2.3 | −2% ± 22 |
| $3 to $13 | 179 | 59% | 28% | +6.7% ± 1.5 | +2% ± 14 |
| $4 to $8 | 82 | 65% | 24% | +8.1% ± 2.0 | +11% ± 23 |
| $4 to $21 | 113 | 61% | 27% | +11.7% ± 2.8 | +18% ± 29 |
| Within 15% of the market | 216 | 59% | 28% | +6.7% ± 1.5 | −6% ± 16 |
| Within 10% of the market | 54 | 72% | 19% | +11.1% ± 2.7 | +18% ± 28 |
| Without the market-drift filter | 167 | 43% | 43% | +2.1% ± 1.7 | +20% ± 20 |
| Every runner $4 to $13 an hour out | 5,037 | 30% | 57% | −6.9% ± 0.3 | −29% ± 3 |
± is one standard error. Figures from 27 September 2026 (tools/tip-rules.py). A bookmaker's close still carries its margin: on the same week the tipped prices sat 9.4% (± 1.1) under the de-vigged market five minutes before the jump. Beating a book's close says the tip was taken before that book moved; it is not by itself a claim that following the tips makes money, and none is made here.
Every tip in this view Technical
| Day | Race | Runner | Book | Before the start | Tipped | Close | Move | Finish |
|---|
Overseas races: did we get more than the fair price?
Australian bookmakers price overseas racing loosely, and now and then one offers a runner at more than it is really worth. These tips are those runners. Each is judged against the fair price just before the jump: every bookmaker's price with the margin taken out. Above the line we got more than it was worth.
Tips where a bookmaker's best price sat at or above the de-vigged consensus of every book quoting, on races outside Australia and New Zealand. Each is judged against that consensus six minutes before the start, the fair price nearest the jump the collector prices: over-fair is tipped / fair − 1, the expected return of a level stake if the fair price is right.
The value tips, tested Technical
If the fair price is right, the winners should match the number it expected, and the return should track the average over fair. Both are shown so neither is taken on trust.
The rule, and how it was chosen Technical
Tested before it went on, 8 to 27 September
Every overseas runner whose best bookmaker price met the rule an hour and a half to ten minutes out, the first time it did, checked against the raw prices: still on offer at that book's next read, no late scratching, and not a race whose key two races shared.
| Read | Tips | Won | Fair price expected | Over fair, median | Level stakes |
|---|---|---|---|---|---|
| The rule (within 15% of the other books) | 79 | 35 | 31.3 | +35% | +127% ± 46 |
| 8 to 20 September | 37 | 17 | 17.1 | +53% | +83% ± 41 |
| 21 to 27 September | 42 | 18 | 14.2 | +31% | +166% ± 78 |
| No cap on the other books | 157 | 48 | 45.5 | +17% | +48% ± 26 |
| more than 30% above the other books | 57 | 6 | 8.3 | n/a | −46% ± 23 |
| The same rule in Australia and New Zealand | 12 | 3 | 3.0 | −19% | not used |
± is one standard error. The 15% cap was set on the same three weeks, so the uncapped row is the conservative read. Resampling whole race days, the capped rule's return was below zero in none of 5,000 draws. The prices lasted a median of 10 minutes at the book and three in four were still there at the next read. Bookmakers take smaller bets on overseas racing than at home, and may restrict accounts that win.
The terms
Method
- Graded on unseen races
- Each night the model is fitted on older races and tested on the latest seven days, so these figures are what it did on races it had not seen.
- Shortened
- The tip's price at the same bookmaker was shorter at the jump than when we tipped it. More money came for it after we did.
- Price against the close
- How much better or worse the tipped price was than the closing price, on average. Above zero means the tips got better prices than the market settled on.
- Settled
- A tip appears here only once its race has an official result.
- Fair price
- Every bookmaker's price for the runner with the bookmaker's margin taken out, pooled. It is the best estimate of the runner's real chance; a price above it is worth more than it costs.
- Held-out grading
- At 03:00 Sydney time the model is refitted on up to 90 days of settled races that end seven days ago, and both the refit and the model in force are graded on the latest seven days. A refit is published only if it grades no worse; the grades shown are always of the model in force.
- The model's price
- A blend of a form prior (fitted temperature per code) and the de-vigged market consensus, the market's weight fitted by time to the jump and by how many books are quoting.
- The market
- Each book's prices de-vigged with the power method and fitted margin curves, pooled as a weighted geometric mean; the exchange counts at 0.7, more when well matched.
- Checkpoints
- The race as the collector held it at first prices, a day, three hours, an hour, thirty and five minutes before the advertised start, and at the start. A book read well behind the others is left out as stale.
- The fair price near the jump
- The de-vigged consensus six minutes before the advertised start, worked out once per tip from the stored prices and kept.
- The close
- The tipped bookmaker's last recorded win price at or before the advertised start, from the same price feed the tip was taken from.
- Refresh
- This page reads
api.marketmover.net/v1/public/recordon opening and every five minutes. Tips appear within minutes of an official result; grades change once a night.