Can we beat the bookies?
We have run every popular betting system through the Lab and watched the bookmaker's margin beat them all. So we asked the harder question: knowing everything the data has taught us, what is the best attempt we can actually make?
This page is the answer. We studied the model's calibration to find the prices it reads sharply, and we studied how the rare professionals who win actually do it: not by out-picking the market, but by consistently beating the closing price. The strategy below is built from both, it is unproven, and we are testing it live. Every bet is logged before the off and published here, win or lose, as the sample grows. Follow it.
Research, not tips. Paper-settled honestly, wins and losses alike. 18+ · please gamble responsibly.
The running record
1,468 bets · 76 days on the recordThe same bets, settled two ways: the Starting Price line shows what the margin does to everyone, the morning-price line shows what taking the early price with BOG changes. What the two lines mean. An early sample still sits within the range luck produces; the closing-line value is the figure that settles first, measured across the whole record, and it is not settled yet. Research, not a signal to stake.
Month by month
The same bets as the chart, grouped by month, so a good or bad stretch is visible on its own rather than buried in a running total. A month is still a small sample: read the direction, not the number.
| Month | Bets | Strike | At BOG | At SP | P&L at BOG |
|---|---|---|---|---|---|
| June 2026 | 342 | 14.3% | +27.2% | -5.2% | +£930 |
| July 2026 | 647 | 15.8% | +28.9% | -11.6% | +£1,867 |
| August 2026 | 479 | 12.7% | -3.8% | -25.5% | −£183 |
How wide are the error bars?
The figures with their 95% intervals, and how far they still are from meaning much.
The reliability meter
How close we are to a precise reading.
How reliable is the read?
1,468 bets settled · around 3,531 for a tight read
About 42% of the way to a reading you could actually trust. The honest headline is that we do not know yet. The closing-line value looks clear of zero on the bets so far, but so far is a small, roughly in-sample sample, exactly the kind of early signal that melts away as often as it holds. It would have to survive thousands of bets to mean anything, and the profit would take far longer again, which is why an early win is the last thing to trust.
The bets, as they were logged
Every pick was written to the record before the off, at the morning price shown. Newest first. These are settled results for the record, not tips: by the time a bet shows here, it has already run.
| Date | Race | Pick | Morning | SP | CLV | Result |
|---|---|---|---|---|---|---|
| 2026-08-23 | Brighton16:50 | War Chant | 9.00 | — | — | pending |
| 2026-08-23 | Worcester16:38 | Dream Diamond | 6.50 | — | — | pending |
| 2026-08-23 | Naas16:00 | Soul Love | 5.50 | — | — | pending |
| 2026-08-23 | Worcester15:38 | Network Gold | 8.00 | — | — | pending |
| 2026-08-23 | Worcester15:08 | Spanish Darling | 9.00 | — | — | pending |
| 2026-08-23 | Brighton14:50 | Last Dandelion | 4.50 | — | — | pending |
| 2026-08-23 | Worcester14:38 | Little Ledgend | 7.00 | — | — | pending |
| 2026-08-23 | Brighton14:20 | Sea Of Charm | 7.00 | — | — | pending |
| 2026-08-23 | Worcester14:08 | Ras Kassar | 7.00 | — | — | pending |
| 2026-08-22 | Lingfield19:45 | Harry Brown | 8.00 | 2.88 | +178.3% | Lost · −£10 |
| 2026-08-22 | Redcar19:30 | Bonnie's Boy | 7.00 | 4.00 | +75.0% | Lost · −£10 |
| 2026-08-22 | Lingfield19:15 | Madame De Sevigne | 8.50 | 3.25 | +161.5% | Lost · −£10 |
| 2026-08-22 | Redcar18:30 | Yafaarr | 8.00 | 9.00 | −11.1% | Lost · −£10 |
| 2026-08-22 | Lingfield18:15 | Mountbatten | 3.50 | 1.80 | +94.4% | Lost · −£10 |
| 2026-08-22 | Redcar18:00 | Woodleigh | 9.00 | 8.00 | +12.5% | Won · +£80.00 |
| 2026-08-22 | Chester17:35 | Dino Bellagio | 8.00 | 7.50 | +6.7% | Lost · −£10 |
| 2026-08-22 | Redcar17:25 | Rajapour | 5.00 | 5.00 | +0.0% | Won · +£40.00 |
| 2026-08-22 | Lingfield17:10 | Orchard | 8.50 | 5.50 | +54.5% | Won · +£75.00 |
| 2026-08-22 | Newmarket17:04 | Maelstrom | 9.00 | 6.00 | +50.0% | Won · +£80.00 |
| 2026-08-22 | Newton Abbot16:52 | Mojito Des Mottes | 6.00 | 7.50 | −20.0% | Won · +£65.00 |
| 2026-08-22 | Redcar16:50 | Maleny | 8.00 | 12.00 | −33.3% | Lost · −£10 |
| 2026-08-22 | York16:45 | El Floridita | 9.00 | 34.00 | −73.5% | Lost · −£10 |
| 2026-08-22 | Curragh16:35 | Asakir | 9.00 | — | — | Void · £10 back |
| 2026-08-22 | Newmarket16:29 | Pearl River | 9.00 | 6.00 | +50.0% | Lost · −£10 |
| 2026-08-22 | Killarney16:20 | Yoradreamer | 5.00 | 8.00 | −37.5% | Lost · −£10 |
| 2026-08-22 | Redcar16:18 | Orfilia | 9.00 | 17.00 | −47.1% | Lost · −£10 |
| 2026-08-22 | Newton Abbot16:15 | Morning Mayhem | 9.00 | 2.75 | +227.3% | Lost · −£10 |
| 2026-08-22 | Curragh16:00 | Vanir | 9.00 | 6.50 | +38.5% | Lost · −£10 |
| 2026-08-22 | Chester15:50 | Angel Ceecee | 7.00 | 9.00 | −22.2% | Lost · −£10 |
| 2026-08-22 | Killarney15:45 | Kilgame | 5.50 | — | — | Void · £10 back |
One £10 win single per qualifying race, on the model's biggest positive edge whose morning price sits in its reliable 3 to 9 band, taken at the morning price with Best Odds Guaranteed. Paper-settled to Starting Price, wins and losses alike; a non-runner is void with the stake returned. The £10 is a research unit, not a stake we are telling anyone to place.
Two prices, two very different stories
Every bet on the board is settled twice, because a horse does not have one price. There is the price at breakfast time and the price at the off, and the gap between them is what this whole experiment runs on.
At Starting Price: −14.6%
The SP is the final price at the off: the sharpest price of the day, with the bookmaker's margin baked in. Settled there, these picks lose, exactly as every system we have ever tested loses. This line is the honest floor, and it is why nobody beats the bookies by simply picking horses.
Morning price + BOG: +17.8%
The same bets at the price we published in the morning, with Best Odds Guaranteed: if the SP ends up bigger than the price we took, BOG pays the bigger one; if it shortens, we keep the price we locked in. An early price plus that one-way promise is where the entire edge, if it holds, lives.
A worked example. Take 6.0 in the morning about a horse the market later agrees with, and it starts at 5.0: the morning taker locked in a bigger price than the horse's final assessed chance. If the market disagrees instead and it drifts out to 7.0, BOG pays the 7.0. Averaged over every settled bet, our morning prices have beaten the SP by +2.9%: that is the closing-line value tile above, and consistently beating the closing price is the one mechanism behind every long-term winner the industry has produced. Whether ours persists at scale is the open question this test exists to answer.
The full story of morning prices, the SP and closing-line value: Morning price vs SP, the gap that decides everything.
What we think has the best chance, and why
You do not beat a bookmaker by naming more winners; the market already prices winners better than we do. You beat one by paying less than the true chance, over and over. That is a single measurable thing, closing-line value, and it is the only mechanism that survives contact with the margin. Our strategy is built to chase exactly it. Even so, our best guess at an edge is still a guess, and only a very large number of bets will ever settle whether it was real.
Value, in the model's good range
We back the model's biggest edge over the market in each race, but only in its reliable mid-range (roughly 2/1 to 8/1), where its calibration is sharp. The longshots, where it over-rates outsiders, are left out on purpose.
The morning price, plus BOG
We take the price we publish in the morning, with Best Odds Guaranteed, rather than the Starting Price. At SP the same picks lose, and the page shows that line too. That reliance on the early price is also why a real punter could not keep this up: bookmakers cut back accounts that win, so those prices do not last.
Judged on closing-line value
The headline is the closing-line value, not the profit. It has low variance, so it settles in months where a profit line takes years. Even so, on this sample it is still an early reading, and the returns come with a wide error bar.
The rules, in full
- Value picks, in the reliable band.One £10 win single per race, on the model's biggest positive edge whose morning price sits between 3 and 9(decimal). The band is fixed from the model's calibration, chosen before the results, never to flatter them.
- Morning price, Best Odds Guaranteed. Logged at the price we publish in the morning; if the SP is bigger, BOG pays the bigger one. That early price is the whole edge.
- Settled honestly at SP.Wins and losses alike, dead heats at the industry sp÷n, fallers counted as the losses they are. Closing-line value is read off every settled bet, win or lose.
- Non-runners are void. The £10 comes back and stays back. Picks are immutable once logged.
What the early numbers tell us, and what they cannot yet
Four things frame the record so far. The sample. 1,468 bets is a real start and still a small one: the shaded cone on the chart shows the range a strategy with no edge at all wanders into by luck, and early stretches this good sit inside it. The vintage. Most of the record is backfill, replayed under exactly the rules the live picks follow, so the sharpest test is every bet logged from here on. The two intervals. The return at BOG carries a wide error bar (+17.8%, 95% interval +1.6% to +34.0%), while the closing-line value (+2.9%, interval −0.2% to +6.0%) is already much tighter: that is why CLV is the headline, it settles in months where a profit line needs years. The real world. Bookmakers restrict accounts that keep beating the morning price, so this record measures the model, not a method anyone could simply copy at scale.
What we expect to happen
If the edge is real, the closing-line value stays positive as the sample grows, its interval narrows around it, and the BOG line keeps climbing clear of the luck cone. If it is not, the CLV sinks toward zero, and we will publish that outcome just as prominently: either answer is the reason the Lab exists. For a precise read on the closing-line value the board needs roughly 3,531 bets; it is 42% of the way there, updated daily.
Where this idea comes from
The seed of this strategy runs in our Silicon Tipster League as Stablebet Edge, which backs the model's biggest market disagreement in every race and is scored, honestly, at Starting Price. This page is that same idea sharpened into our best shot: the reliable price band, the morning price, Best Odds Guaranteed, and closing-line value as the measure.
Questions
Can an AI model beat the bookies at horse racing?
That is the question this page exists to test, live and in public. The market prices races formidably well: at the Starting Price nothing we have tested comes out ahead, and this strategy loses at SP too. But the professionals who do win share one measurable habit, beating the closing price, and our model shows exactly that signal. So we built our best version of it and we are running it in the open, still unproven, every bet published, until the sample is big enough to settle the question either way.
What is closing-line value, and why is it the headline?
It is how much bigger the price you took is than the price the horse finally starts at. If you keep taking 6.0 about horses that go off at 5.0, you are consistently getting more than the true chance, and that is the mechanism behind every long-term winner. It does not follow that we have it: an early, in-sample reading cannot show that. We measure it because it has low variance, so it firms up in months where a profit line takes years. It is the number worth watching here, but on this sample it is still an early reading, not a settled one.
Why take the morning price rather than the Starting Price?
Because that is where the edge lives. At SP the strategy loses, like everything else, so settling at SP would remove the whole point. The model's value picks are systematically priced bigger in the morning than they start, so taking the early price, with Best Odds Guaranteed on top, is the only version with a chance. That gap, morning versus SP, is the closing-line value.
How are the picks chosen?
From each race the model's single biggest positive edge, its widest gap over the market, but only where the morning price sits in the model's reliable mid-range (roughly 2/1 to 8/1). The band is set from the model's calibration, not from whatever scored best: outside it the model over-rates rank outsiders and the edge is a mirage. One £10 win single per qualifying race, logged before the off.
Is this real money, and will following it make me money?
No, and we do not claim it will. The bets are paper-settled and the £10 is a research unit, not a stake we are telling anyone to place. The return still carries a wide error bar on this sample, and bookmakers restrict accounts that keep winning, so a real punter would not hold these prices for long. The closing-line value is the figure worth following, and whether it holds over thousands of bets is precisely what the experiment is running to find out. Nothing here is betting advice.
What happens with non-runners and short cards?
A non-runner is void: the £10 comes back and is never moved to another horse, exactly as a locked morning bet would be. On a thin card the strategy backs whatever qualifies and simply stakes less, recorded honestly. Picks are immutable once logged.
Lab-wide questions (how the model works, how good it is) live on the Lab FAQ.
Gamble responsibly. This page is research and entertainment, not betting advice, and betting should never be a way to make money. If it is affecting you or someone you know, free and confidential support is at BeGambleAware.org. 18+.