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 belief in Flat racing, and the one being re-argued this season: the draw.
The short version
- The draw bias is real, and bigger than chance. Of 242 British course-and-trip combinations with enough Flat racing to test, 46 showed a stall bias too large to be luck. Chance alone would have produced about 16. There are roughly three times as many biased draws as there should be.
- It persists. Take the biases found in 2008–16, freeze them, and follow them through 2017–26: the favoured third of the stalls still wins 10.80% of the time against 8.85% for the other two — 22% more often, across 117,000 runners.
- And it pays exactly nothing. Those same winners came in at an A/E of 1.015, inside one standard error of par. Picked properly — found in one four-year block, bet in the next, four times over — it is 1.005.
- The gap between those two numbers is the whole column. In the block where a bias is found it looks like an A/E of 1.15 to 1.21. In the block where you would actually have bet it, 1.00. Four blocks out of four.
- Chester is the picture. A low draw over six furlongs there won 16.97% against 6.91% in the years we picked it, and 16.37% against 7.27% in the nine years since — the bias did not fade at all. The horses still came in at roughly what their prices said.
- Has course management flattened the draw? Barely, and not since about 2011. In sprints the favoured third used to win 1.42 times as often as the rest; now it is 1.33. Over seven furlongs and beyond there is no change at all. What moved was the price, not the track.
- The jockey does not change the draw. Across 1,025 rider-by-course-by-stall records, how a rider did from a given third in 2008–16 tells you nothing about 2017–26 (correlation 0.03). The 30 best-looking records, at an A/E of 1.62, came back at 1.019. The same method finds real persistence in riders' overall records, so it is not the method failing.
- One hint and one small lean. In 16+ runner fields the market is slowest to price the draw (A/E 1.118, but only about one standard error). And a low draw in a soft-ground sprint has returned 1.054 for eighteen years, steady across both halves — about 5%, which commission mostly eats.
- Build it: System Builder — the Draw IV band breakdown shows a strike rate climbing and an A/E column that doesn't, on one screen. The Holdout view turns any rider-and-stall hunch into this column's test in one click.
This week's talking point: is the draw still worth knowing?
The draw is the one edge every racing fan believes in. You learn it early — low at Chester, the far side at Ascot over five — and you carry it for life. It is also the only bias you can see from the stands, which is probably why it has survived every attempt to price it away.
This season the belief has been under new argument. The case against it runs like this: racecourses water more evenly and move the running rail more often than they used to, the ground is better maintained than it was twenty years ago, and so the old draw biases have been flattened. The numbers you learned, on this argument, describe a sport that no longer exists.
That is testable, and it is really two claims. Has the bias weakened? And — the question that matters if you are betting rather than admiring — was it ever worth money in the first place?
We hold every British Flat result back to 2008 with a Betfair Starting Price attached: eighteen years, between 3,000 and 4,700 races a year with eight or more runners, a price on 99.6% or more of the runs in every single year. That is enough to answer both. (A few races a year carry no stall numbers at all because they are started by flag rather than from stalls — Salisbury's mile-and-six-furlong races account for 37 of the 47 since 2008. With no draw, they add nothing to any comparison here.)
Then there is a third question, which almost nobody asks out loud but which turns up in every trends piece and every pub argument: does the jockey change it? Some riders are supposed to be good from a wide stall, or hopeless from low at a particular track. We tested 1,025 of those records. That section is the one we would read first.
What A/E means, in one paragraph. A/E is short for "actual over expected". Take a group of horses, turn each one's starting price into the chance the market gave it, and add those chances up — that is how many winners the market expected from the group. Then count how many it actually had. An A/E of 1.00 means the group won exactly as often as its prices implied: correctly priced. 1.10 means it won 10% more often than the market allowed for.
A strike rate tells you how often something wins. A/E tells you whether you were paid properly for it. They are different questions, and the draw is the cleanest example in racing of the answer being "yes" to the first and "no" to the second.
First: there is no such thing as "a low draw"
Pool every British Flat sprint since 2008 and split each field into thirds by stall, and the draw vanishes. Over five and six furlongs on turf the low third wins 9.26% of the time and the high third 8.68%, and both return within a point of par (A/E 1.010 and 0.992). On the all-weather it is 10.87% against 9.13%, and the high third actually returns marginally the better of the two.
That is not because the draw doesn't matter. It is because "low" means opposite things at different tracks. Low is everything at Chester and a positive disadvantage at Ascot over five. Pooled, they cancel out. Every honest draw question has to be asked one course and one distance at a time, which is how we asked it.
How we cut it. A "cell" is one course, one surface (turf or all-weather) and one trip rounded to the nearest furlong — Chester turf 6f, Kempton all-weather 7f, and so on. Within each race the stalls are ranked among the horses that actually ran, so a non-runner closes the gap rather than leaving a hole, and the field is split into a low, middle and high third. A cell had to carry at least 30 winners in the years we picked it from. That left 242 cells to test.
The bias is real: 46 cells out of 242
Take 2008 to 2016 and ask, for every cell, whether any third won a bigger share of the races than its share of the runners — by enough that luck becomes an implausible explanation (two standard errors or more). Forty-six cells qualified.
Chance alone would have produced about sixteen. When you take the best of three thirds in each of 242 cells, roughly one cell in fifteen clears that bar by accident. So the honest reading is not "46 biases", it is "about thirty more biased draws than randomness can account for". The draw is real, and racing's collective folklore about it is broadly correct.
And it keeps going. Freeze those 46 cells, never look at them again, and watch the following nine years:
| Period | Group | Runners | Wins | Strike rate | A/E |
|---|---|---|---|---|---|
| 2008–16 (where they were found) | favoured third | 37,639 | 4,363 | 11.59% | 1.124 |
| 2008–16 | the other two thirds | 79,887 | 6,633 | 8.30% | 0.930 |
| 2017–26 (the next nine years) | favoured third | 37,662 | 4,068 | 10.80% | 1.015 |
| 2017–26 | the other two thirds | 79,482 | 7,035 | 8.85% | 0.990 |
The strike rate survived. A horse in the favoured third of a biased cell won 22% more often than one in the other two thirds, across 117,000 runners and nine years of racing that played no part in finding the bias. If what you want is a filter that lifts a shortlist's hit rate, the draw is one of the best in the sport.
Now look at the last column. 1.015, against a standard error of 0.016. The market had already paid for it.
What a bias looks like when you find it, and what it pays when you bet it
The table above still flatters the idea, because it picks from one nine-year block and tests on another — a single split that happened to work. The harder version, and the only one we would act on, is a walk-forward: find the biases in a four-year block, bet them in the next four-year block, then move on and do it again. That is what betting a draw bias would actually have felt like, because at every point you only know what has already happened.
| Found in | Bet in | Cells | Favoured SR | Other SR | A/E when found | A/E when bet |
|---|---|---|---|---|---|---|
| 2008–11 | 2012–15 | 26 | 10.88% | 9.05% | 1.179 | 1.030 |
| 2012–15 | 2016–19 | 24 | 10.33% | 8.89% | 1.209 | 0.962 |
| 2016–19 | 2020–23 | 34 | 10.50% | 8.71% | 1.156 | 1.026 |
| 2020–23 | 2024–26 | 28 | 10.63% | 9.52% | 1.145 | 0.986 |
| All four blocks pooled | about 10.6% | about 9.0% | 1.15–1.21 | 1.005 | ||
Read the last two columns as a pair, because together they are the most useful thing in this article. In the years where you find a draw bias it looks like an A/E of 1.15 to 1.21 — a 15 to 21% edge, the kind of number that makes people build systems. In the years where you would have bet it, 1.00. Four blocks out of four, 4,517 winners against 4,492.6 expected.
The strike-rate gap, meanwhile, is intact in every block: about 10.6% against 9.0%. This is not a case of the effect disappearing. It is a case of the effect being fully paid for.
How much of a "found" edge is just the act of looking?
We can put a number on that, and it is the part of this exercise we would most want a reader to steal.
Go back to the cells where there was no real bias — the 196 that did not clear the bar. In each one, pick whichever third happened to do best in 2008–16 anyway. Those thirds show an A/E of 1.079 over the years they were picked from. Nothing is there. We know nothing is there, because we selected precisely the cells where the evidence was too weak to believe. And still, simply choosing the best of three after the fact manufactures an apparent 8% edge.
Then follow the same picks forward into 2017–26: 1.002.
An edge of 1.08 can be produced out of thin air by choosing the best of three options with hindsight. So when any trends piece shows you a draw angle returning 12% over the period it was discovered in, the first question is not "is the sample big enough?" — it is "how many stalls, courses and distances were looked at before this one was shown to me?"
Has the draw bias weakened?
This is the topical claim, and the answer is: a little in sprints, not at all beyond seven furlongs, and not recently in either.
Measured as the ratio of the favoured third's strike rate to the other two, sprints averaged 1.42 across 2008–16 and 1.33 across 2017–26. But the decline is neither gradual nor modern: 2008, 2009 and 2010 are the outliers (1.58, 1.85, 1.45), and from 2011 onward the series is flat, with last year at 1.21 and this year, part-run, at 1.37. Over seven furlongs and further the ratio went from 1.26 to 1.30 — if anything a shade stronger.
| Bet in | Sprint 5–6f: strike-rate ratio | Sprint A/E | 7f+: strike-rate ratio | 7f+ A/E |
|---|---|---|---|---|
| 2012–15 | 1.23 | 1.017 | 1.18 | 1.042 |
| 2016–19 | 1.09 | 0.907 | 1.20 | 0.992 |
| 2020–23 | 1.32 | 1.060 | 1.12 | 1.000 |
| 2024–26 | 1.16 | 0.953 | 1.10 | 1.004 |
| Pooled A/E | — | 1.001 | — | 1.008 |
What did change is the price. In sprints, the favoured third's A/E ran at 1.109 across 2008–16 and 1.026 across 2017–26. The bias barely moved. The money attached to it halved, and then went.
So the fair summary of this season's argument is not that course management has killed the draw. It is that the draw still matters about as much as it did a decade ago, and the market stopped giving it away. Those are very different statements, and only the second one costs you anything.
Chester is the picture
If you want one cell that tells the whole story, it is the Chester sprint. Everybody knows about it, the knowledge is correct, and it is worthless.
| Cell | Favoured third | 2008–16 SR: favoured v rest | 2017–26 SR: favoured v rest | 2017–26 A/E |
|---|---|---|---|---|
| Chester 6f | low | 16.97% v 6.91% | 16.37% v 7.27% | 1.134 |
| Chester 5f | low | 17.75% v 6.62% | 16.04% v 8.11% | 0.945 |
| Chester 7f | low | 15.61% v 7.51% | 12.32% v 8.40% | 0.941 |
| Pontefract 1m | low | 14.08% v 7.31% | 14.11% v 7.52% | 1.044 |
| Sandown 5f | low | 15.53% v 7.02% | 12.00% v 9.82% | 0.994 |
| Kempton (AW) 7f | low | 10.56% v 8.20% | 10.61% v 8.42% | 0.961 |
| Wolverhampton (AW) 7f | middle | 11.33% v 8.96% | 10.89% v 9.28% | 1.035 |
| Goodwood 6f | low | 11.28% v 7.01% | 8.19% v 8.42% | 0.866 |
| Ascot 5f | high | 9.80% v 5.42% | 9.57% v 5.22% | 1.265 |
Two warnings about that table, and we mean both of them.
The first: do not bet a single line of it. When you carry 46 cells forward, two or three of them will come back at 1.2 or better for no reason other than arithmetic. The Ascot 5f figure of 1.265 is exactly the number this column exists to be suspicious of — 62 winners where 49 were expected, which is a perfectly ordinary run of luck at that sample size. Shown here as an example of the spread, not as a tip.
The second: look at Goodwood. A real, statistically strong low-draw bias over six furlongs in 2008–16 (11.28% against 7.01%) has simply gone in the years since — 8.19% against 8.42%, which is nothing at all. A draw bias is a fact about a piece of grass, and grass gets re-laid, re-watered and re-railed. Some of them do die.
But then Chester, Pontefract and Kempton did not fade in the slightest. The strike rates in the right-hand columns are almost identical to the left-hand ones, nine years apart. The low draw at Chester over six furlongs wins two and a quarter times as often as the rest of the field, and has for eighteen years. It is one of the largest, most reliable and most useless pieces of information in British racing.
Where the edge hides — mostly, it doesn't
If the market prices the draw on average, there might still be a corner where it doesn't. We looked in the two obvious places.
By price: nothing. Out of sample, the favoured third returns 1.032 among horses shorter than 3.0, 1.026 between 3 and 7, 1.005 between 7 and 15, and 0.977 at 15 and bigger. If anything the bias is best paid in the shortest prices, which is the opposite of the usual "the market ignores it in the outsiders" story.
By field size: a gradient, and this is the one genuinely interesting thing in the section.
| Runners | Favoured third A/E (walk-forward) | Other two thirds |
|---|---|---|
| 8–11 | 0.993 | 1.001 |
| 12–15 | 1.029 | 0.981 |
| 16 or more | 1.118 (114 v 102.0) | 0.924 |
In the biggest fields the favoured third won 7.04% against 4.60% — a bigger ratio than anywhere else — and returned 1.118. That is the shape you would expect if the market prices the draw well in a ten-runner race, where everyone can see it, and less well in a twenty-runner cavalry charge where there is more to think about.
Now the honesty. That 1.118 carries a standard error of about 0.105, which puts it roughly one standard error from par: the kind of number that is right about as often as a coin. It rests on 114 winners across four blocks. And in the simpler single-split version the same group came back at 1.366, of which three individual cells supplied more than half. It is a hint that the market is slowest to price the draw in the biggest fields. It is not a bet, and we are not dressing it up as one. If it is still there in five years we will say so then.
The question nobody tests: does the jockey change the draw?
This is the section we would read first, because it is where the money is lost.
Every week, somewhere, you will see a line like "this rider is 3 from 102 when drawn wide at this track" or "he has an outstanding record from low stalls here". These records are easy to produce and impossible to argue with, because they are true. The question is whether they mean anything about next season.
We built every rider-by-course-by-stall-third record we could: a jockey, a course, a third of the draw, at least 50 rides in 2008–16 and at least 30 in 2017–26. That is 1,025 records. Then we asked what the first half predicted about the second.
| How the record looked in 2008–16 | Records | A/E then | A/E in 2017–26 | Above par next time |
|---|---|---|---|---|
| Outstanding (two standard errors better than expected) | 30 | 1.621 | 1.019 | 17 of 30 |
| Good | 481 | 1.208 | 1.005 | 223 of 481 |
| Poor | 508 | 0.809 | 0.968 | 223 of 508 |
| Dreadful (two standard errors worse) | 7 | 0.447 | 1.001 | 4 of 7 |
The thirty best rider-and-stall records in British Flat racing returned an A/E of 1.62 in the years that made them famous — 655 winners where 404 were expected — and 1.019 afterwards. The seven that looked unbackable at 0.447 returned 1.001. Both tails went to par, and roughly half of each group finished above par next time, which is what you would get from tossing coins.
It gets more emphatic than that. If rider-and-stall skill existed, the spread of these records would be wider than chance allows. It is narrower: the standard deviation of the 1,025 records is 0.953 where pure randomness predicts 1.000. And the correlation between a record's first half and its second half is 0.030. Zero, to the nearest useful number.
The control, which is why we believe the null. A test that finds nothing is worthless unless you can show it finds something when something is there. So we ran the identical method on riders' overall records — 79 jockeys with 500-plus rides in both halves, no course and no stall involved.
Correlation between halves: 0.274, with a spread of 1.309 against the 1.000 chance predicts. Riding ability is real, it shows up in the prices, and this method sees it. What it cannot see, because there is nothing to see, is a rider who is specifically good from stall four at Beverley.
So: slice rider by course by stall and you will always, always find someone who is 3 from 102 somewhere. There are tens of thousands of such slices. What this table says is that when you have finished finding them, they go on to do precisely what everyone else does.
What last Saturday looked like
Doncaster's straight course, 12 September, two big-field handicaps on the same afternoon. The Portland, 21 runners, was won from stall 16 — the high third. The Pertemps, 17 runners, was won from stall 6 — the middle. Same strip of grass, an hour apart, opposite answers.
Doncaster over five and six furlongs in 16-plus runner fields is not a biased cell by our test: across 137 races since 2008 the low third wins 4.88%, the middle 5.09% and the high 6.12%, with A/E figures of 0.895, 0.951 and 1.107 on 55 winners. There is a lean in there if you squint, and it is well within the range that 137 races produces by accident.
Which is the point. A straight with no reliable bias still produces "the high side came home" on any given Saturday, every Saturday, and somebody will tell you it was the draw.
The one lean that has held for eighteen years
We looked at a great many slices this week, and one of them refuses to go away. In British turf sprints (seven furlongs and shorter, eight or more runners), on soft ground or worse, a low draw has returned an A/E of 1.053 in 2008–16 and 1.055 in 2017–26. Two separate eras, the same answer, 2,188 winners against 2,074 expected across the whole period — 1.054, with a standard error of 0.023.
| British turf sprints, 8+ runners | Low third A/E | Middle | High |
|---|---|---|---|
| Good to soft or softer | 1.054 | 0.966 | 0.982 |
| Good | 1.001 | 0.996 | 0.997 |
| Good to firm or faster | 0.982 | 1.015 | 0.994 |
The gap between the soft-ground and fast-ground low draw is 0.072, with a standard error of 0.030 — a shade over two standard errors. Stability across two independent nine-year blocks is worth more than that significance test, and it is the reason we are printing it at all.
But be clear about the size of it. A 5% lean is not a system. Commission on a winning book takes a large bite of 5%, the effect is one of many slices we examined this week (which is exactly the multiple-comparisons trap we spent two sections warning about), and horses in the same race are not independent of one another, so the honest error bars are a little wider than the ones printed. It is a thumb on the scale when you are already deciding between two horses in a soft-ground sprint. It is not a reason to bet one.
And the obvious counter-trade is not a trade
If the favoured third returns 1.00, and the unfavoured thirds also return roughly 1.00, somebody will ask the sensible question: is there money in opposing the fashionable stalls, since everyone is piling into them?
No. The other two thirds came back at 0.990 out of sample and 0.99 to 1.00 in the walk-forward blocks. That is not an edge, it is the same par with a different label. Backing the horses everyone discards because their A/E is near 1.00 would be committing precisely the error this column exists to point out: an A/E of about 1.00 is a statement that a group is correctly priced, not an invitation to bet it.
Build it yourself
Three ways to reproduce and extend this week's work in the System Builder, which is included with Proform Premium.
1. See the bias and the price on one screen. Set Course to Chester, Race Type to Turf, and Distance to Nearest Furlong between 5 and 6. Then open the Draw IV breakdown. Draw IV measures how well a stall has done at today's course, race type and trip over the previous five years — so it is calculated only from racing that had already happened, which is the walk-forward discipline built into the product. Watch the strike-rate column climb across the bands and the A/E column sit still. That contrast is this entire article in one screen.
2. The field-size hint, labelled as a hint. Add Draw IV of 1.25 or more and Field Size of 16 or more, with Race Type set to Turf. This is the 1.118 group from the table above. Run it, then run it on a different date range, and treat the difference between the two answers as the honest measure of what you have found.
3. Test a rider-and-stall angle properly. Take any "this jockey is superb from a low draw here" claim you like. Set Jockey, Course and Draw Group, and then — before you look at the headline number — open the Holdout view, which splits the record into the years used to build it and the most recent twelve months, and tells you whether the two agree. Section five of this article is that button, run 1,025 times.
One note on the third recipe: the product's Draw Group setting splits a field into halves, low and high, while this article used thirds. The halves are the better tool for a single-race question and the thirds were the better tool for comparing cells of wildly different field sizes; the conclusions do not depend on which you pick, but the numbers will not match line for line.
Or take the raw rows away with you. Data Export, which launched on Sunday, will hand you the underlying rows as a spreadsheet: Draw, Draw IV, Draw Group, Field Size, Distance to Nearest Furlong, Jockey and Official Going are all tickable columns, and date, course, race type, finishing position, starting price and Betfair Starting Price come as standard on every file.
Everything in this article — the cells, the thirds, the walk-forward blocks — is a pivot table away from that file. If you would rather check our arithmetic than take it on trust, that is the door.
What we would take away from this
The draw is one of the best-documented real effects in racing, and racing's folklore about it is right. It is also the single clearest demonstration we have run of the difference between a fact about horses and an edge over a market. The favoured third at Chester wins twice as often as the rest. Every bookmaker, every trader and every algorithm in the sport knows it wins twice as often, and the price says so.
That leaves the draw doing what it is genuinely good for: building a shortlist, explaining a result, telling you which horses had an excuse. Those are worth having. The one thing it will not do is beat the price — and the honest version of a strong bias is not "here is the angle", it is "here is a fact everyone has already paid for".
Last week's column asked what would happen to British racing if the Irish stopped coming. Next Thursday: the beaten favourite — racing's most reliable-sounding tip, and whether it survives the same treatment.
Common questions
Is draw bias real in British Flat racing?
Yes. Of 242 course-and-trip combinations tested on British Flat racing from 2008, 46 showed a stall bias larger than chance can explain, against the roughly 16 that pure luck would have produced. Those biases also persisted: over the following nine years the favoured third of the stalls won 10.80% of its races against 8.85% for the other two thirds.
Can you make money betting draw bias?
Not on this evidence. Picked without hindsight — found in one four-year block and bet in the next, repeated four times — biased draws returned an A/E of 1.005 at Betfair Starting Price, which is par. In the blocks where the biases were discovered they looked like 1.15 to 1.21, and that gap is the cost of hindsight rather than a real edge.
Has draw bias got weaker?
Slightly in sprints and not at all over longer trips, and the change is not recent. The favoured third's strike-rate advantage in sprints averaged 1.42 times the rest of the field in 2008–16 and 1.33 in 2017–26, with almost all of the decline coming from unusually strong 2008 to 2010 figures. From 2011 onwards the series is flat. What changed materially was the price: the sprint A/E fell from 1.109 to 1.026.
Which British courses have the strongest draw bias?
Chester is the clearest: over six furlongs the low third of the draw has won about 16.5% of the time against about 7% for the rest, and that has held from 2008 to today. Pontefract over a mile and Ascot over five furlongs (favouring the high side) also show large, persistent effects. Goodwood over six furlongs is the cautionary case — a strong low-draw bias in 2008–16 that has completely disappeared since.
Do some jockeys ride better from certain draws?
We could not find any evidence that they do. Across 1,025 rider-by-course-by-stall records, the correlation between a record's first nine years and its second nine years was 0.030. The 30 best-looking records returned an A/E of 1.62 over the years that produced them and 1.019 afterwards. The same method applied to riders' overall records found a correlation of 0.274, so it detects real jockey ability perfectly well — there simply appears to be no such thing as a stall specialist.
Does the draw matter more in big fields?
The strike-rate advantage is largest in fields of 16 or more, where the favoured third won 7.04% against 4.60%, and the return there was an A/E of 1.118 — the only group in this study that finished meaningfully above par. But the standard error on that figure is about 0.105, so it sits around one standard error from par and rests on 114 winners. It is a hint that big fields are priced less carefully, not a betting angle.
What is A/E?
A/E stands for actual over expected. Turn every horse's starting price into the chance the market gave it, add those chances up to get the number of winners the market expected, and divide the actual winners by that figure. An A/E of 1.00 means a group of horses won exactly as often as their prices implied. Above 1.00 means they were underestimated; below means overestimated. Unlike profit, it is not distorted by one big-priced winner, which is why this column uses it.
Method
British Flat racing, turf and all-weather, 1 January 2008 to 14 September 2026, restricted to races with eight or more runners and to horses that took part — non-finishers included, since leaving fallers and unseats out inflates every figure. 2008 is the floor because that is where complete Betfair Starting Price coverage begins in our data; a price is present on 99.6% or more of runs in every year since.
Draw positions are ranked among the horses that actually ran, so withdrawals close the gaps, and each field is split into thirds. Because thirds of a field are unequal by construction, every comparison here is between thirds of the same races and never between raw win counts. A cell is one course, one race type and one trip rounded to the nearest furlong, and needed at least 30 winners in the period it was selected from. Statistical significance for a bias is measured against the share of runners that third contributed to that cell. A/E is calculated at Betfair Starting Price across all starters.
One caveat worth stating: horses in the same race are not independent of one another — they share a draw structure, a going and a pace — so the standard errors quoted here are a little optimistic. That makes the null findings in this article stronger rather than weaker, and the one lean at the end correspondingly more fragile.
