Where the Model Disagrees With the Market at the St Leger

Settled. The St Leger Festival ran from 10 to 13 September. All four days are below, each with the model's card for the day's feature as it published that morning and the result underneath it. Across the meeting's 31 races the model's first choice won three and the favourite on the same morning card won fourteen. The closing section has the full count.

The model's top-rated runner at Doncaster was, in most of the races, not the horse the market had made favourite. The two also disagreed about how likely the whole thing was. The model's figure for its own first choice sat below the market's figure for its favourite in almost every race we published.

The record is clear about what that is worth. Across 11,314 races the market's prices have been the sharper forecast, a Brier score of 0.089 against the model's 0.096, on a measure that scores a stated probability against what actually happened and where the lower number is the better one. Backing the model's top pick flat at starting price across all of those races returns −14.1%. Where the two parted company at Doncaster, the market was usually the one closer to the result.

So the disagreement is worth reading. It is not a shortlist. A betting market is a large number of people with money at stake, and money concentrates. Our model sees ratings and form figures, and it has no access to the things the money knows, so it spreads the chance more evenly across a field than the market does.

That is why its first choice is so often a horse at a longer price, and why that fact on its own means very little. A flatter forecast will name more outsiders than a sharper one whether or not it has found anything.

What follows is the model's card for each day of the festival, with the market alongside it and the result of each feature race, and a closing note on what four days did and did not show.

Thursday: the May Hill Stakes

The Group 2 May Hill is the two-year-old fillies' championship of the meeting and the feature of the opening day.

Read it as two columns rather than a ranking. Model chance is what our model makes of each filly. Market is the chance implied by the price on the card with the bookmaker's charge taken back out, so that column sums to a hundred and the two are comparable. The vs Market column states the difference in percentage points and nothing more, which is deliberate: a gap is a disagreement, not a valuation.

The table is sorted by the model's number, so the top row is the model's first choice. Look across that row to the Market column and the pattern from the opening section usually shows up on its own. The model's leading filly is rated lower by the model than the favourite is rated by the market, even when they are the same horse.

The wide gaps are worth a moment because of where they land. They tend to fall on horses in the middle and at the back of the market, the ones with the least money and the least attention on them, and those are exactly the runners a form-only model has the least information about. A large disagreement is generally a sign that the model is short of evidence rather than long on insight.

Result: Musical Times won at 7/4 for William Buick and Charlie Appleby, from a field of eleven. She headed the morning market and the model did not have her on top, which is the ordinary shape of the meeting rather than the exception.

Friday: the Doncaster Cup

Cup Day's feature is the Doncaster Cup (Group 2), the closing leg of the Stayers' Triple Crown after the Ascot Gold Cup and the Goodwood Cup. Eight were declared and the card below is the morning's eight. Queenstown came out, so seven ran.

Read it the same way as Thursday's card. The Model chance column is our model's estimate, Market is the chance implied by the price on the card with the bookmaker's charge removed, and vs Market is the difference in percentage points between the two.

A small field changes what the gap means. With eight runners rather than eleven, the market's chance is spread across fewer horses, so both columns carry larger numbers and a disagreement of a given size is proportionally smaller than the same number would be on Thursday. Comparing the raw vs Market figures across the two days is not comparing like with like.

Staying races are also the part of the card where a form-only model has least to go on. Stamina at two miles and beyond tends to be demonstrated in a handful of races a season, so the record behind each runner is thinner than it looks. Where the model parts company with the market here, short evidence is the more likely explanation.

Result: Caballo De Mar won at 5/6 for Oisin Murphy and George Scott. He was second in the morning market and the model rated him fourth. Queenstown, the model's own top-rated runner on this card, was the horse that came out, which is a reminder of what a morning estimate is: a read of the declared field, not of the race that eventually runs.

Saturday: the St Leger

The oldest Classic in the world, the fifth and last of the British season, over a mile and three quarters at Doncaster. Eleven ran.

This is the one race of the meeting where the model and the market agreed on the name and still disagreed on almost everything else. Both put Enceladus on top. The market gave him a shade over a quarter of the race; the model gave him an eighth. Look down the Model chance column and the eleven figures sit between roughly six and thirteen per cent, which is a field the model can barely separate. That is the known weakness at this level rather than a discovery: a Classic gathers horses with short, high-class form lines, and short form lines are what a form-only model has least to read.

The gap on the card is therefore not a case for anything. It is the shape of a flat forecast sitting next to a sharp one.

Result: Highwayman won at 11/1 for Joseph O'Brien, ridden by Saffie Osborne, who became the first woman to ride the winner of a British Classic. The model had him fifth of eleven on its own figures; the morning market had him fourth on price. Neither side's first choice won, and neither side had the winner anywhere near the top of its list.

The Silicon Tipster League has no valid call on this race at all. Every League pick at Doncaster that Saturday was logged after its race had been run, because that morning's collection failed and was re-run in the afternoon, so the picks are excluded from the season board and from the meeting's record. The model's own card above is unaffected: it published at 05:03 that morning, hours before the first race. The full account is in how we measure a festival.

Sunday: the Sceptre Fillies' Stakes

The closing card's feature is the Group 3 Sceptre over seven furlongs, and it is the opposite kind of race to Saturday's. Twenty were declared, Sukanya came out, nineteen ran.

A big field does two things at once, and they pull in opposite directions. It gives a form-only model more to read, because most of these fillies have run often enough to leave a record. It also spreads the market's own chance thinner, and the bookmaker takes a slice on every extra runner: this was the dearest of the meeting's four feature races to bet in, at close to twice the charge on Saturday's eleven-runner Classic. The same disagreement in percentage points means something different in a race that costs that much more.

The model and the market agreed on the name again, both putting Forty Years On on top, and again they disagreed on the size of it. The model's figures across the field run from about three to ten per cent.

Result: Pina Sonata won at 12/1 for Daniel Muscutt and James Fanshawe. She was neither side's pick. That is two feature races in two days where the model and the market named the same horse and the same horse was beaten, which is worth saying plainly because agreement between a flat forecast and a sharp one is not evidence about either of them.

What the four days showed, and what they did not

We said at the start of the meeting that we would count how often the model's first choice won and how often the market's favourite did, and publish both, including the case where ours came out ahead. Ours did not come out ahead.

Across the meeting's 31 races, the model's first choice won three and the favourite on the same morning card won fourteen. They named the same horse in six of the 31. The model's first choice was withdrawn in three races and the market's in one, so on the races where each actually ran the count is three from 28 against fourteen from 30. Both figures are read from the model's own published cards for 10 to 13 September and the meeting's frozen record of what won; the two feeds disagree about race identifiers, so the races are matched on their runners rather than their ids.

That is the result, and it should not be dressed up in either direction. Four days is a few dozen races, which is nowhere near enough to prove anything about a model. What it does do is sit exactly where the long record already sits: the all-time record runs to 11,314 races and settles at −14.1%, and the market's prices score the better Brier over every window we publish. A meeting that came out the same way is one more ordinary week, not a finding.

The interesting part is not the score. It is that in both of the meeting's feature races the model and the market picked the same horse and were both beaten, at 11/1 and 12/1. Agreement between a flat forecast and a sharp one is not a stronger signal than either of them alone, and this meeting is a small, clean demonstration of why.

Nothing here was presented as a selection and nothing here becomes one now. This page carries the model's arithmetic. The form angles, the going and the draw for each of the four days are on our St Leger Festival tips page, written as research, and the day-by-day running is in the St Leger Festival results hub. The model card does not become a tip because it puts a number beside a horse's name.

If the Market column is the part you want to understand, what chance is that price really giving my horse shows how a price converts into a percentage and how often each price actually wins. For why a Classic generally costs less to bet in than the handicap after it, which race at the St Leger costs the most measures the charge race by race across the card.

Every all-time figure on this page is read live from the published record and moves when the record moves. The meeting's own figures come from its frozen record and will not.

Common questions

Did the model pick the favourite?

Usually not. Its top-rated runner is more often a horse further down the market, which is a consequence of it spreading chance more evenly across a field than the market does rather than a judgement that the favourite is wrong.

If the model disagrees with the market, who is right?

The market, more often. Over 11,314 races its Brier score is 0.089 against the model's 0.096, and lower is sharper.

Should I back the horses the model rates higher than the market?

No. Backing the model's top pick flat at starting price across every race it has called returns −14.1%, and the wide disagreements sit on the runners the model knows least about.

What does the vs Market column actually mean?

The difference in percentage points between the model's estimated chance and the chance implied by the starting price once the bookmaker's charge is removed. It is a gap between two opinions and carries no claim that either is correct.

Were the cards published before the race?

Yes. Every card above was generated on the morning of its own race, from the declared field, and the card labels itself a pre-race estimate once the off time has passed. Nothing is added or amended after a result, which is why the Doncaster Cup card still shows a runner who was later withdrawn.

Did you tip the St Leger?

No. Everything in the Lab is published research, the losing figures included, and there is nothing to subscribe to.

A word on all of this

None of this describes a way to win, and no approach we have tested across 30,701 races has made money. If your betting has stopped being fun, BeGambleAware has free, confidential help and the National Gambling Helpline is on 0808 8020 133. 18+.