THE AI LAB

THE AI LAB · TESTED LIVE

Is the StableBet Rating any good?

The StableBet Rating is the gold number beside every runner on our racecards: today's official rating, less 2lb for each length the horse was beaten on its last finished run, up to 10 lengths. In handicaps it has done useful work: in our tests, the top-rated horse won more often than the horse with the top official rating. It is a quick way to read a race, not a tip.

Tests on past races only go so far. So before its first live race, we wrote down what it has to do from here and how many races it takes to judge. This page keeps that count from the live data, rebuilt every night, and the answer goes up here whichever way it lands.

Research, not tips. 18+ · please gamble responsibly.

Racing to 4 October 2026 · updated 4 October 2026 · after racing each evening

The prediction, written in advance

The prediction being tested

+3.30 to +5.98 points

How many percentage points more often the top StableBet Rating should win a handicap than the top official rating.

We registered this on 24 September 2026, before the first race of the first live card went off. It covers version 1.0 of the rating, on every live card from 24 September 2026. The range runs from the lowest to the highest end of the intervals our tests on past races produced, so it says what those tests expect to happen on races run after it was written down.

No verdict is read before 1,500 qualifying handicap races. That first read decides. After it, the record is read again every 500 races as a check, each time on the first races in date order, so a verdict can never move between reads.

The races that count: Races with four or more starters, one winner, and an SP and a published StableBet Rating for every starter; primary: handicaps.
What is measured: Win credit to the top StableBet Rating minus win credit to the top official rating, ties sharing credit, in percentage points.

What we publish for every runner follows a public data contract, the rating's JSON Schema, so anyone can check the fields and the rules behind them.

The count so far

Still collecting

279 of 1,500 handicap races scored1,221 to go before the first verdict

19.0%

Top StableBet Rating won

15.8%

Top official rating won

+3.23

Difference, points

−1.55 to +7.95

95% interval

0−5+10

Live difference +3.23, 95% interval −1.55 to +7.95Predicted range, being tested +3.30 to +5.98

The live test is running, and it is early. After 279 handicaps the top StableBet Rating has won 19.0% of them and the top official rating 15.8%, a difference of +3.23 points. The 95% interval runs from −1.55 to +7.95, wide enough to take in zero and the whole predicted range, so on this sample luck and a real difference look the same. The first verdict comes at 1,500 races, 1,221 from now, roughly seven weeks of racing at the pace so far, and it will be published here whichever way it lands.

How the difference has moved

The chart starts when there is enough racing to draw it.

Beside the model and the favourite

The same 279 handicaps, picked two more ways. The model's pick is the horse our AI race predictor rated most likely to win that morning. The favourite is picked at the off, so it knows the final market and this is not a fair fight: it is here as a yardstick.

Top StableBet Rating19.0%
The highest StableBet Rating in each race.
Top official rating15.8%
The highest official rating in each race.
The model's top pick19.7%
The horse our AI race predictor rated most likely to win, from the morning card.
The SP favourite28.3%
The shortest price at the off. It knows the final market.

Priced in

The market reads the same form. Backing the top-rated horse in every handicap so far found 58 winners from 306 selections, where their starting prices implied about 57.7. That is an actual over expected of 1.005, and the 95% interval runs from 0.801 to 1.216. A ratio of one means the price already carries everything the rating knows. We registered a band of 0.95 to 1.06 for it, and this sample is still far too small to read either way. In the frozen test year before it went live (1 September 2025 to 31 August 2026), the same ratio in handicaps was 0.980 (95% interval 0.938 to 1.020) over 8,744 selections: what priced in looks like.

So the rating helps you read a race. It does not find prices that are wrong, and backing its top-rated horse blind would lose roughly the bookmaker's margin.

Every split

Only handicaps decide the test. The other splits are reported beside it, as registered, and the smaller ones swing a long way on a few races.

Top rated is the top StableBet Rating; OR is the official rating. Differences are in percentage points.

Split	Races	Top rated won	Top OR won	Difference	95% interval
All races	307	21.3%	17.1%	+4.23	−0.30 to +8.82
HandicapsDecides	279	19.0%	15.8%	+3.23	−1.55 to +7.95
Non-handicaps	28	44.6%	30.4%	+14.29	−1.79 to +28.57
Great Britain	229	22.9%	19.0%	+3.93	−1.35 to +9.68
Ireland	78	16.7%	11.5%	+5.13	−2.56 to +14.10
Flat	223	23.5%	18.9%	+4.63	−0.79 to +9.87
Jumps	84	15.5%	12.3%	+3.17	−5.16 to +11.91

The latest races

The most recent scored handicaps from the live record, with the top-rated horse, the top official rating and the winner of each.

Newest first. A name marked won is the horse that won the race.

Race	Top rated	Top official rating	Winner
Curragh 17:254 Oct	Flanker Jet	Flanker Jet	Blue Blue Moon
Killarney 17:054 Oct	Keep Em Quite	Keep Em Quite	Joe's Turn
Uttoxeter 16:564 Oct	Better Off Alonewon	Better Off Alonewon	Better Off Alone
Curragh 16:504 Oct	Sinbad My Dad	Ben Vrackie	Whimsy
Kelso 16:404 Oct	Helsgetaway	Helsgetaway	Luminaries
Uttoxeter 16:204 Oct	Holy Legend	Inedit D'amour, Ask A Sainte (joint top)	Prospect House
Curragh 16:154 Oct	Cool Azul	Nakamura	Nouvel Espoir
Killarney 15:554 Oct	Kilbrin Roccowon	Bythesametoken	Kilbrin Rocco
Uttoxeter 15:454 Oct	Prettylady	Lucky Sevens	Berry Edge
Kelso 15:304 Oct	Tommy's Oscar	Tommy's Oscar	Breaking Cover
Killarney 15:204 Oct	James The Bravewon	Good World	James The Brave
Uttoxeter 15:154 Oct	Jubilant	Divilabotherwon	Divilabother
Curragh 15:054 Oct	Cristal Clere	Eastwatch	Iamimmaculate
Kelso 14:204 Oct	Klitschko	Klitschko	Interchangeable
Killarney 14:104 Oct	Colin's Friend	Regards To Rose, Big Dee (joint top)	Footstepstoglory
Southwell 20:303 Oct	Shopaholic	Shopaholic	Blue Tulip
Southwell 20:003 Oct	South Parade	South Parade	Toptime
Southwell 19:303 Oct	New Imagewon	New Imagewon	New Image
Southwell 17:553 Oct	Nostrumwon	Nostrumwon, Wid, Cristo (joint top)	Nostrum
Redcar 17:403 Oct	Ziggy's Triton	Ziggy's Triton	Emperor Caradoc

What we've tried, to make it better

The rating only changes through a test written down before it runs. Each idea is registered first, with the rule it will be judged by, then scored once on a frozen year of races it was never tuned on (1 September 2025 to 31 August 2026), against version 1.0. To pass, it has to find the winner more often by a margin that still holds once every idea in its batch is allowed for, stay ahead in at least three of the four quarters of that year, and lose no more than 2 points of coverage, the share of runners it can rate.

Batch B2, run on 4 October 2026

The check on the test itself (SBR-012)

Version 1.0 against the top official rating on the same frozen year: the top-rated horse won 18.7% of 7,954 handicaps, against 14.4%, a difference of +4.38 points (95% interval +3.50 to +5.28). Version 1.0 was chosen with this year in view, so this figure checks that the test still finds what we found. The first clean test is the live count at the top of this page.

SBR-008Did not pass
Moving version 1.0 by how the horse's last run went against its starting price changes how often the top-rated horse wins a handicap.
Against version 1.0: −0.22 points (95% interval −0.55 to +0.12). Under this rule the top-rated horse won 18.5% of 7,954 handicaps; under version 1.0 it won 18.7%.
It found the winner less often than version 1.0.
The test as registered
Moving version 1.0 by how the horse's last run went against its starting price changes how often the top-rated horse wins a handicap. The surprise is the result last time (1 for a win, 0 otherwise) less the market's chance last time: 1 over the horse's starting price divided by the race's total of 1 over the starting price (the overround taken out), which exists only where every starter in that race had a starting price above 1.0. So a beaten horse the market gave a 40% chance reads -0.4, a beaten outsider it gave 5% reads -0.05, and a winner it gave 5% reads +0.95. The run is the one version 1.0 reads, the last finished run, and the figure is version 1.0 plus a weight times the surprise, rounded as version 1.0 is. The weight, in pounds for a whole unit of surprise, is chosen on the development window alone (1 October 2023 to 31 August 2025): the value in the rule's grid, from -16 to +16 in steps of 2, whose top-rated horse won development handicaps most often, the one nearest 0 on a tie. A negative weight credits a horse the market fancied and a positive one a horse that beat the market's view; a weight of 0 is version 1.0 itself. A runner whose last run has no market chance gets no figure. Direction unknown.
The runner's own reasons: the difference is not positive (-0.22 points); p 0.3758 is not under 0.05.
SBR-009Did not pass
Moving version 1.0 by the horse's form on today's going against its form on other going changes how often the top-rated horse wins a handicap.
Against version 1.0: +0.07 points (95% interval −0.48 to +0.62). Under this rule the top-rated horse won 18.8% of 7,954 handicaps; under version 1.0 it won 18.7%.
Its lead was too small to tell apart from luck once all two ideas tested together were allowed for.
The test as registered
Moving version 1.0 by the horse's form on today's going against its form on other going changes how often the top-rated horse wins a handicap. It tests Brett's point that ground is often the whole story (IDEA-011) as a change to the number. Today's going is folded into the racecard's eight bands (firm, good to firm, good, good to soft, soft, heavy, standard and slow, with the Irish yielding scale folded as the racecard folds it). Over the horse's finished runs before today whose band and distance behind the winner are both known, with lengths capped at 10 as in version 1.0, the rule averages the lengths beaten on today's band and on every other band. The adjustment is a weight times (the average on other bands less the average on today's band) times n / (n + 2), where n is the number of runs on today's band, so one run counts a third and the adjustment grows with the evidence. It is 0 when today's going is unknown, or the horse has no run on today's band or none elsewhere, so the rule rates exactly the runners version 1.0 rates. The weight, in pounds per length, is chosen on the development window alone (1 October 2023 to 31 August 2025): the value in the rule's grid, from -4 to +4 in steps of 0.5, whose top-rated horse won development handicaps most often, the one nearest 0 on a tie. A weight of 0 is version 1.0 itself. Direction unknown.
The runner's own reasons: p 0.8046 is not under 0.05.

The two controls

Every batch carries two controls, and they decide whether its results count at all. The negative control (SBR-010) swaps the lengths beaten between the horses in each race, which keeps the shape of the data and throws away what the lengths say, so it has to fail. It did, at −5.05 points. The positive control (SBR-011) cheats on purpose by reading each race's result, so it has to pass. It did, at +81.27 points. Both behaved, which shows the test can see a real effect without inventing one, so the batch counts.

None of the two ideas passed, so version 1.0 stays on the racecard. A failed idea is still a result, and each one is published here with its numbers.

Batch B1, run on 27 September 2026

The check on the test itself (SBR-000)

SBR-001Did not pass
Turning lengths into pounds at the rate the official handicapper uses for the trip, instead of a flat 2lb, changes how often the top-rated horse wins a handicap.
Against version 1.0: +0.14 points (95% interval −0.14 to +0.44). Under this rule the top-rated horse won 18.9% of 7,954 handicaps; under version 1.0 it won 18.7%.
Its lead was too small to tell apart from luck once all five ideas tested together were allowed for. It was ahead in two of the four quarters, and at least three are needed.
The test as registered
Turning lengths into pounds at the rate the official handicapper uses for the trip, instead of a flat 2lb, changes how often the top-rated horse wins a handicap. On the Flat the rate is the BHA's published scale: 3.41lb a length at five furlongs and 1.22lb at a mile and six furlongs (BHA, Performance figures, Poundage scales), joined by a smooth curve that comes out at about 17lb divided by the trip in furlongs, so more pounds a length over sprints and fewer over staying trips. The BHA publishes its jumps scale only as a graph, with no figures in the text, so over jumps the rate is k divided by the trip in furlongs, with k chosen on the development window alone: the value in the rule's grid whose top-rated horse won development handicaps most often (the lowest on a tie). The trip is the one the lengths were recorded over, and lengths are still capped at 10. Direction unknown.
The runner's own reasons: p 0.9295 is not under 0.05; 2 of 4 folds agree in sign (at least 3 needed).
SBR-002Did not pass
Averaging the figure over the last three finished runs, instead of reading the last run alone, changes how often the top-rated horse wins a handicap.
Against version 1.0: −1.09 points (95% interval −1.89 to −0.23). Under this rule the top-rated horse won 17.6% of 7,954 handicaps; under version 1.0 it won 18.7%.
It found the winner less often than version 1.0.
The test as registered
Averaging the figure over the last three finished runs, instead of reading the last run alone, changes how often the top-rated horse wins a handicap. Each run gives today's official rating less 2lb a length beaten (capped at 10). A run among the three whose distance is unknown is left out, at least one run is needed, and the result reports how many runs each figure used. Direction unknown.
The runner's own reasons: the difference is not positive (-1.09 points); p 0.0550 is not under 0.05.
SBR-003Did not pass
Crediting a last-time winner 2lb for each length it won by (capped at 10 lengths), instead of holding it at its official rating, changes how often the top-rated horse wins a handicap.
Against version 1.0: +0.19 points (95% interval −0.24 to +0.65). Under this rule the top-rated horse won 18.9% of 7,944 handicaps; under version 1.0 it won 18.7%.
Its lead was too small to tell apart from luck once all five ideas tested together were allowed for.
The test as registered
Crediting a last-time winner 2lb for each length it won by (capped at 10 lengths), instead of holding it at its official rating, changes how often the top-rated horse wins a handicap. A dead heat for first counts as a margin of nothing, and a winner whose margin is unknown gets no figure. Everything else is version 1.0. Direction unknown.
The runner's own reasons: p 0.9295 is not under 0.05.
SBR-004Did not pass
Counting a fall, pull-up or other non-completion last time as beaten 10 lengths, instead of reading the finished run before it, changes how often the top-rated horse wins a handicap.
Against version 1.0: +0.17 points (95% interval −0.05 to +0.39). Under this rule the top-rated horse won 18.9% of 7,954 handicaps; under version 1.0 it won 18.7%.
Its lead was too small to tell apart from luck once all five ideas tested together were allowed for.
The test as registered
Counting a fall, pull-up or other non-completion last time as beaten 10 lengths, instead of reading the finished run before it, changes how often the top-rated horse wins a handicap. A horse whose only runs were non-completions gets a figure too. Everything else is version 1.0. Direction unknown.
The runner's own reasons: p 0.5597 is not under 0.05.
SBR-005Did not pass
Scaling the lengths beaten by the size of that race's field changes how often the top-rated horse wins a handicap.
Against version 1.0: +0.14 points (95% interval −0.18 to +0.46). Under this rule the top-rated horse won 18.9% of 7,954 handicaps; under version 1.0 it won 18.7%.
Its lead was too small to tell apart from luck once all five ideas tested together were allowed for.
The test as registered
Scaling the lengths beaten by the size of that race's field changes how often the top-rated horse wins a handicap. The formula is lengths x 10 / starters in that race (finishers plus non-completions), then capped at 10 and taken off at 2lb a length, so a length counts double in a five-runner race and half in a 20-runner one. Everything else is version 1.0. Direction unknown.
The runner's own reasons: p 0.9295 is not under 0.05.

The two controls

Every batch carries two controls, and they decide whether its results count at all. The negative control (SBR-006) swaps the lengths beaten between the horses in each race, which keeps the shape of the data and throws away what the lengths say, so it has to fail. It did, at −5.29 points. The positive control (SBR-007) cheats on purpose by reading each race's result, so it has to pass. It did, at +81.27 points. Both behaved, which shows the test can see a real effect without inventing one, so the batch counts.

None of the five ideas passed, so version 1.0 stays on the racecard. A failed idea is still a result, and each one is published here with its numbers.

Every registered test, its lock and its full result sit in the public research register, so anyone can check this page against it.

How it's checked

Each morning card is matched to its result by the runners the two share, since a horse runs once a day. The off time only breaks a tie, because meetings run late. Every race that does not qualify is left out and counted, with its reason.

Racing covered: 11 days, to 4 October 2026
Card races: 443
Matched to a result: 443 (1 after the off time moved)
Card races with no result: 0
Results with no card race: 13
Races scored: 307 (279 of them handicaps)
Races left out: 136
Starters carrying a rating: 79.8% of the 4,198 matched to a card

Why races were left out

A starter has no StableBet Rating121
Fewer than four starters11
Void or abandoned4

Every card scored so far ran version 1.0. Cards up to 24 September 2026 were published before each card carried a version stamp, and are counted as version 1.0, as the record's rules set out.

The rules in full

As registered on 24 September 2026, and read by the page exactly as written.

Registered: 24 September 2026, for version 1.0, scoring every live card from 24 September 2026.
The races that count: Races with four or more starters, one winner, and an SP and a published StableBet Rating for every starter; primary: handicaps.
The measure: Win credit to the top StableBet Rating minus win credit to the top official rating, ties sharing credit, in percentage points. Its interval comes from a paired race bootstrap of 2,000 resamples, seed 20260924, percentile 95%.
The prediction: Between +3.30 and +5.98 points.
When it is read: Not before 1,500 races. The first read decides; further reads every 500 races are monitoring, each on the first races in date order.
Confirmed: The interval's lower end is above zero and the interval overlaps the predicted range.
Weaker than predicted: Above zero, but the whole interval sits below +3.30.
Not confirmed: The lower end is at or below zero. The record keeps collecting, and if that still holds at 3,000 races, the racecard's evidence sentence changes to the live finding and a replacement is registered.
Stronger than predicted: The whole interval sits above +5.98. The pipeline is audited for leakage before the result is reported as a finding.
Priced in: Actual over expected at the starting price should sit between 0.95 and 1.06. If its whole interval falls outside that band, the pipeline is audited.
The results: Every race is settled from the published results files at the true starting price, never the morning price.

Where the rating lives

Today's racecards →The gold number beside every runner, with the reason whenever a runner has none.The rating in the Tipster League →Its top-rated horse in every fully rated handicap, logged before the off and settled at the starting price, beside the AIs and the favourite.

Questions

What is the StableBet Rating?

Our own figure for every runner on the racecard, in pounds on the official rating's scale: today's official rating, less 2lb for each length the horse was beaten on its last finished run, up to 10 lengths. A horse that won last time keeps its full mark. It is a quick way to read a race, not a tip, and it is published beside our AI race predictor's output without being part of it.