The Thursday Take is our weekly column: one racing talking point, put to the test against the Proform database — every British and Irish runner, every result. This week, the oldest system in the game, and the one nearly every punter has tried: back the favourite that got beaten.
The short version
- The belief is true. A horse that started favourite last time and got beaten wins its next race 18.51% of the time, against 9.75% for a runner that was neither favourite nor winner. Nearly twice as often, across 48,795 British runners since 2018.
- And the price already knows, to three decimal places. Those horses returned an A/E of 1.0071 — two-thirds of a standard error from dead par. A/E is simply what actually won divided by what the prices said should win; 1.00 means the market got it exactly right.
- The refinement everyone reaches for does nothing. Beaten a neck or beaten twelve lengths changes the strike rate a lot — 23.47% down to 14.44%, falling in a straight line — and changes the value not at all. Every band from 0 to 20+ lengths sits between 0.976 and 1.066, and not one clears one and a half standard errors.
- The best number in the study is a trap, and we show the whole trap. Beaten favourites that had been odds-on returned +22.51% at level stakes in Britain, stable across two eras. The same rule in Ireland returned −7.62%. The British figure was +1.27 standard errors all along: a big return wearing a suit.
- Then we hunted properly, and found nothing. We cut the cohort into 39 slices by layoff, race type and handicap status, picked them on 2018–21 and bet them on 2022–26. Not one slice reached two standard errors in-sample, the spread of results was 0.815 where chance alone gives 1.000, and all 39 pooled out-of-sample at 1.0023.
- One thing survived everything: the long break. A beaten favourite returning after 120 days or more came in at an A/E of 1.0697 across Britain and Ireland — +2.56 standard errors, and agreeing to within half a percentage point in two separate countries (1.0686 and 1.0732). Everyone else off the same break managed 1.0161, so this is about five points more than a general layoff effect.
- It is an ingredient, not a system. Level stakes on it is −0.94% — break-even. The honest claim is that the market's estimate is roughly 7% too low, which is enough to wipe out its own margin and no more. Against −14.80% for everyone else off the same break, that is standing still instead of bleeding. Four of the nine years are below par.
- Build it: System Builder — Beaten Favourite LTO is a single on/off tick that reproduces the headline in one click, and adding Days Since Last Run of 120 or more gives you the one finding that survived.
This week's talking point: the oldest system in the game
Every punter meets this one early. A short-priced favourite runs below itself — bad trip, wrong ground, went too fast — and next time it turns up at a bigger price against a weaker field. The market, so the belief runs, over-reacts to one bad afternoon. Back it.
It is not a new idea and we are not presenting it as one. "Beaten favourite" has its own entry in the public betting guides, it has been a named angle in tipping columns for decades, and it is one of the first filters anyone builds when they get hold of a database. That is exactly why it is worth the week: an angle this old has had a very long time to be either proven or priced away, and almost nobody publishes which of the two happened.
We hold every British and Irish runner since 2018 with a Betfair Starting Price attached — 708,873 British starters in this cohort, of which 48,795 were beaten favourites last time out. That is a large enough sample to answer the question properly, and to keep Ireland back as a separate country to test any finding against.
The one piece of jargon in this column, in plain words.
Every figure below is an A/E: actual over expected. Count the winners a group of horses actually produced, then add up what their starting prices said they should have produced. Divide one by the other. 1.00 means the market priced the group exactly right. Above 1.00 means those horses won more often than their prices implied — the market was too low on them. Below means the opposite. We use it rather than profit because profit can be moved a long way by two or three big winners, and because it tells you why a system wins or loses instead of just that it did.
The belief is true, and it is worth nothing
Here is every British runner since 2018, split by what it did last time out.
| Last time out | Runners | Wins | Strike rate | A/E | Return |
|---|---|---|---|---|---|
| Favourite and won | 28,171 | 6,890 | 24.46% | 0.9973 | −1.76% |
| Favourite and beaten | 48,795 | 9,034 | 18.51% | 1.0071 | −2.95% |
| Won but not favourite | 56,885 | 9,993 | 17.57% | 0.9903 | −5.05% |
| Neither | 575,022 | 56,038 | 9.75% | 1.0002 | −12.43% |
The first thing to say is that the folk belief is right about the horses. A beaten favourite wins nearly twice as often as an ordinary runner. Anyone who tells you the angle is nonsense has not counted.
The second thing is that the market has it to three decimal places. 1.0071 is two-thirds of a standard error above par — on a sample this size that is noise, and it is the single most precisely-priced group of horses we have measured in this column.
The return column says the same thing in money, and it says something useful about betting generally. The four figures run −1.8%, −3.0%, −5.1%, −12.4%. You lose least backing the horses the market already respects, and most backing the ones it has dismissed — not because the respected ones are value, but because the margin is spread unevenly and it sits heaviest on the outsiders. That is a fact about bookmaking, not about form.
The refinement that should work, and doesn't
The obvious next move is to grade the defeat. A favourite beaten a neck feels like a completely different horse from one beaten twelve lengths. On strike rate, that instinct is exactly right.
| Beaten last time by | Runners | Strike rate | A/E | Standard errors |
|---|---|---|---|---|
| Under a length | 8,899 | 23.47% | 1.0211 | +0.95 |
| 1 to 2½ lengths | 9,617 | 19.83% | 0.9757 | −1.07 |
| 2½ to 5 lengths | 10,531 | 18.24% | 1.0049 | +0.22 |
| 5 to 10 lengths | 9,071 | 16.25% | 1.0059 | +0.23 |
| 10 to 20 lengths | 5,183 | 15.45% | 1.0171 | +0.48 |
| 20 lengths or more | 3,462 | 14.44% | 1.0656 | +1.42 |
| Did not complete | 2,032 | 16.83% | 1.0165 | +0.30 |
The strike rate falls in an almost perfect line, 23.47% down to 14.44%. The closer the horse got last time, the more often it wins next time. That is real and it is large.
And the value column does not follow it anywhere. The whole table sits between 0.976 and 1.066, and not one row clears one and a half standard errors. The market has already read the distance column. Everything the beaten margin tells you about the horse is in the price before you get there.
The row that will tempt a writer is the bottom one: horses beaten twenty lengths or more returned +9.72% at level stakes. It is +1.42 standard errors on 3,462 runners. Which brings us to the part of this column we think is worth more than any of the findings.
The best-looking number in the study, and how it dies
If you are hunting for the angle inside this cohort, the place your eye goes is the biggest mistakes: favourites that were odds-on and still got beaten. The market was maximally wrong about them last time. Surely it over-corrects.
| Beaten favourite, by its price last time | Runners | A/E | Standard errors | Return |
|---|---|---|---|---|
| Britain — odds-on (1/2 or shorter) | 726 | 1.0915 | +1.27 | +22.51% |
| Britain — 1/2 to evens | 3,799 | 1.0407 | +1.23 | −1.69% |
| Britain — evens to 2/1 | 15,247 | 0.9835 | −0.92 | −5.53% |
| Britain — 2/1 to 7/2 | 20,174 | 1.0197 | +1.15 | −1.21% |
| Britain — bigger than 7/2 | 8,832 | 0.9935 | −0.24 | −5.17% |
| Ireland — odds-on, same rule | 213 | 1.0245 | +0.20 | −7.62% |
There it is: +22.5% at level stakes, on a rule a child could follow, and when we split it across two eras it held up in both. This is the precise point at which an angle gets a name, a landing page and a monthly subscription.
Then you ask Ireland. Same rule, same years, same definition, a country that was never involved in finding it: an A/E of 1.0245 and a return of −7.62%. Gone.
And the thing is, the number told you before Ireland did. +1.27 standard errors on 726 bets was never significant. It was a big return wearing a suit — and 726 bets feels like a lot right up until you work out that it is about eighty a year, which is one moderate losing run away from looking like nothing at all.
This is why this column quotes A/E and standard errors rather than profit. A return of +22.51% is a fact about what happened. It is not a claim about what happens next, and the only thing that separates the two is the sample size sitting underneath it. Any angle you are shown with a profit figure and no sample size attached is being sold to you, not shown to you.
And then we hunted properly
The honest way to search a cohort like this is to fix the method before you look. Cut the beaten favourites into cells — days since their last run, race type, handicap or not — measure each cell on the first four years, and then bet it blind on the next five. If there is a real slice in here, it shows up twice. If we are just finding shapes in noise, the good-looking cells collapse.
39 cells had enough runners in both halves to test.
| Cells tested | 39 |
| Cells reaching two standard errors in the first half | 0 |
| Spread of the results (chance alone gives 1.000) | 0.815 |
| Best cell in the first half | 1.1551 (+1.57) |
| … the same cell in the second half | 1.0470 (+0.59) |
| All 39 cells pooled, second half | 1.0023 |
Read the third row twice, because it is the most interesting number in the article. If you slice a cohort 39 ways and every cell is genuinely priced correctly, random variation alone should still scatter the results with a spread of about 1.000 — you would expect one or two cells to look excellent purely by luck. We got 0.815. Less scatter than chance. Not only is there no angle hiding in these 39 slices, there is slightly less apparent angle than pure randomness would have manufactured for us.
And all 39 pooled, bet blind in the second half, came back at 1.0023. Par, to two decimal places, across five years.
What survived: the horse that went away
One cut was specified before any of this started, because a previous episode of this column had already found it somewhere else. In August we looked at big-field Flat handicaps and found that the horse returning after a long absence was consistently better priced than the horse coming back in a fortnight. So we asked the same question of the beaten favourite — in two countries and two eras, four independent cells.
| Days since last run | Britain 2018–21 | Britain 2022–26 | Ireland 2018–21 | Ireland 2022–26 |
|---|---|---|---|---|
| 0 to 13 | 1.0359 | 1.0078 | 0.9938 | 0.9455 |
| 14 to 27 | 0.9846 | 0.9862 | 0.9655 | 1.0260 |
| 28 to 59 | 1.0189 | 0.9948 | 0.9847 | 0.9485 |
| 60 to 119 | 0.9609 | 1.0247 | 0.9082 | 0.9687 |
| 120 or more | 1.0630 | 1.0737 | 1.1159 | 1.0410 |
The bottom row is the highest figure in its column four times out of four. Pooled, and with a control group run alongside it:
| Returning after 120 days or more | Runners | Strike rate | A/E | Standard errors | Return |
|---|---|---|---|---|---|
| Beaten favourite — Britain | 6,755 | 16.14% | 1.0686 | +2.19 | −1.08% |
| Beaten favourite — Ireland | 2,211 | 15.92% | 1.0732 | +1.33 | −0.48% |
| Beaten favourite — both | 8,966 | 16.08% | 1.0697 | +2.56 | −0.94% |
| Everyone else — both | 113,151 | 8.85% | 1.0161 | +1.60 | −14.80% |
Two countries that share almost no races agreeing to within half a percentage point — 1.0686 and 1.0732 — is about as good as this kind of evidence gets outside a laboratory.
It is also not merely a layoff effect, and we ran the control to check. There is a small general under-pricing of the long-absent horse: everyone else off the same break comes in at 1.0161. The beaten favourite sits about five points above that. The interaction — the extra bit that belongs to the combination rather than either part — measures +1.85 standard errors. Suggestive. Not conclusive. We would rather print that number than hide it.
What it is actually worth, which is less than it looks
Level stakes on the pooled group returns −0.94%. Break-even. This is the part that gets left out when an angle like this is sold, so we will be blunt about it: a 7% under-estimate is enough to cancel the market's own margin and nothing more. You are not being paid to bet these horses. You are, roughly, no longer paying to.
That still matters, because the comparison is the point. The same break, the same years, every other horse: −14.80%. The difference between standing still and bleeding is a real difference, and it is the correct way to use a finding this size — as one ingredient in a selection you were making anyway, not as a system you bet on its own. There are about 775 British qualifiers a year, which is two or three a week.
The caveats, on the page rather than in a footnote
- Four of the nine years are below par. Year by year in Britain the figures run 0.973, 1.253, 1.107, 0.933, 1.101, 0.920, 0.957, 1.147 and 1.370. A +2.56 standard error result across nine years is not a promise about the next nine months, and anyone who shows you only the good years is showing you a sales page.
- This year is running hot and we are deliberately not leading with it. 2026 to date sits at 1.370 on 423 runners. That is one part-year on a small sample, and it is exactly the shape the odds-on section above exists to warn you about. If this finding is real, 2026 will look much more ordinary in hindsight.
- It is uneven by code, and we are not going to tidy that up for you. All-weather 1.1215, hurdles 1.0828, turf Flat 1.0801, bumpers 1.0547 — and chases 0.9896. The tempting move is to drop chases and quote the other four. That is precisely the multiple-comparisons mistake the 39-cell hunt above was built to measure, and doing it here would undo the whole column. One code in five landing below par is what four real effects and one ordinary bit of variation look like.
- It is not the survivor of a search. That matters more than it sounds. The 120-day cut was specified in advance because a different cohort had already produced it — it is not the best-looking cell we could find after the fact. Inside the 39-cell grid its signal is spread thinly across about ten cells, where nothing could have cleared two standard errors even if it were real.
The bit we find most convincing, and it isn't a number
In August this column tested the big-field Flat handicap — a completely different population, different race conditions, different sort of horse — and found that 14 to 27 days was the worst-priced layoff band and the long break the best. This week, across every code and every kind of race in two countries, 14 to 27 is the worst band in three of the four cells and 120 days or more is the best in all four.
Two independent cohorts, arrived at by different routes, producing the same shape. That is the strongest evidence this series has produced for anything — stronger, honestly, than the standard errors, because a result that replicates in a population you did not tune it on is much harder to manufacture by accident than a result that merely looks big.
The plausible story is simple enough. A horse off 120 days has a stale form line and an absence that reads as a worry, and the market discounts the worry a little too heavily. A horse back in a fortnight has fresh evidence, and fresh evidence is what markets price best.
Build it yourself
This is the cheapest recipe the column has published, and every number above is reproducible in the product today — there are no gaps this week.
- Beaten Favourite LTO — a single on/off tick in the System Builder. That one click reproduces the headline table at the top of this article. It uses the strict definition we used: outright favourite, not joint-favourite, and beaten.
- Add Days Since Last Run set to 120 or more, and you have the only finding here that survived everything. Run it and you will see the strike rate and the A/E move in different directions, which is the lesson of the whole piece on one screen.
- LTO Distance Beaten is there if you want to watch the section-two table build itself — the strike rate marching down and the value column refusing to follow.
- The Holdout view is the one that matters most here. It splits any system you have built into an early period and a recent one and shows you both. Every angle in this article that died, died there first.
- In Data Export the same four columns are available to tick — Beaten Favourite LTO, Days Since Last Run, LTO Distance Beaten and Times Beaten Favourite — if you would rather rebuild this week's tables in your own spreadsheet and check our arithmetic. We would encourage that.
The take
The oldest system in racing is real and worthless at the same time, which is the most common thing an honest test finds. Beaten favourites win nearly twice as often as ordinary horses, and the market charges you for every bit of it. Grading the defeat adds nothing. The biggest-looking number in the whole study — a 22% return on odds-on losers — was a standard error and a quarter, and Ireland killed it.
What is left is small, specific and replicated: the beaten favourite that has been off for four months is priced about seven per cent too short, in two countries, across two eras, for the same reason the same thing was true of Ebor runners in August. It will not pay your rent. It will stop a bet costing you money, which is a different and more useful thing.
None of which says that systems do not work, and it is worth being exact about why. The reason we can tell you the long-break figure is worth about seven per cent, and that the 22% one was worth nothing at all, is that both went through the same test — an out-of-sample split, a second country, a control group, and a sample size printed next to every claim. An angle that survives all of that is worth building on. One that arrives as a big return with nothing underneath it is worth what this column keeps finding it is worth. The work was never finding the numbers. It is knowing which ones to throw away, and that is a job you can only do on your own ideas if you can see both columns at once.
Last week we asked whether draw bias is still worth betting — the bias turned out to be real in 46 of 242 British course-and-trip combinations, and to return exactly par once you pick it without hindsight. Next week we put a question to the test that we have wanted to answer for a long time, and have only recently had the data to answer properly.
Questions we get asked about this
Does backing beaten favourites make money?
No. Across 48,795 British runners since 2018 they won 18.51% of the time against 9.75% for an ordinary runner, but they returned an A/E of 1.0071 — par — and a level-stakes loss of 2.95%. They win far more often than average and they are priced to do exactly that.
Is a favourite beaten a short head a better bet than one beaten twenty lengths?
It wins more often — 23.47% against 14.44% — and it is not a better bet. Every beaten-distance band from under a length to twenty-plus returned between 0.976 and 1.066, and none of them reached one and a half standard errors. The market has already priced the margin of defeat.
What about favourites that were odds-on and got beaten?
In Britain those 726 runners returned +22.51% at level stakes, which is the most attractive figure in this study and also a false one. It was +1.27 standard errors, and the identical rule applied to Ireland returned −7.62%. A large return on a small sample is not an edge.
Is there any version of this angle that works?
One. A beaten favourite returning from a break of 120 days or more returned an A/E of 1.0697 across Britain and Ireland, +2.56 standard errors, with both countries agreeing to within half a percentage point and a control group confirming it is about five points stronger than the general long-layoff effect. At level stakes it is break-even (−0.94%) against −14.80% for everyone else off the same break — an ingredient in a selection, not a system on its own.
How many of these run in a year?
About 775 in Britain, which is roughly two or three a week across all codes.
