Right on average, or right every time?

The objection, stated properly

We have written before that favourites win about 35% of the time and that this is roughly what their price implies. Hold the final exchange price constant and the apparent edges in our own work thin out to nothing.

A sharp reader does not accept that, and they are right not to. Suppose the market overrates the favourite by five points in big fields and underrates it by five points in small ones. Average the two together and you get a perfectly calibrated market. The errors have not gone anywhere. They have cancelled, and no average test will ever see them.

That is a real possibility and it is not answered by any of the work we have published so far. Every favourite study on this site makes an argument about the average. This one asks whether the average is hiding something.

What we did

We took 5,778 races across 165 race-days, from May to October 2026, and for each one compared the favourite's actual result with what its own Betfair starting price implied. Across the whole set that ratio came out at 0.998 — close enough to 1.000 that the difference is noise.

Then we stopped averaging. We split the same races nine ways, by states we named in advance: price band, field size, race type and others. If the market were overrating favourites in one kind of race and underrating them in another, a split like that should pull the two apart and show it.

We also carried three random partitions through the same machinery as controls. If a method finds patterns in meaningless splits, the patterns it finds in meaningful ones are worth nothing either.

Nothing survived. Divided by price into ten bands, the largest any band strayed from its own implied win rate was 1.28 standard deviations — well inside what chance produces. Across the nine named states, once you correct for having looked nine different ways, not one deviation stood up.

The limit, which belongs here and not in a footnote

This does not mean the market is perfect, and we are not going to let that sentence stand unqualified. A study of this size would reliably catch a cancelling error of about 6% — at four percent its chances are worse than a coin toss. So the honest statement is narrow and we would rather give you the narrow one: on the nine states we named, the favourite's price does not drift by more than roughly six percent. Smaller errors than that are not ruled out. They are simply beyond what 5,778 races can see.

There is a second limit and it matters more. We split the races one dimension at a time. A market could be well calibrated on every one of those dimensions separately and still be wrong where several of them line up at once — a particular field size and a particular going and a particular class, together. A test that slices one way at a time is blind to that by construction. We have not ruled it out. We have not looked for it.

Why we are not publishing the detail

The study was re-run from scratch by a second implementation before any of this was written. The headline figures reproduced exactly — the race count, the race-days, the ratio, the decile spread, the finding that nothing survives.

The individual splits did not. The original did not record where it drew its bin edges, and on the same data a different defensible choice moved one partition's result substantially. The conclusion is unchanged either way, which is why we are comfortable publishing it. But a number we cannot reproduce from our own record is a number you cannot check, so those numbers are not here. If that is frustrating to read, it was considerably more frustrating to find.

Where this leaves the question

The usual lament in this field is that you never have enough data. For this problem that is the wrong way round.

A situation that comes up in one race in a hundred, and in which the price is wrong by half, needs only about sixty of those races to prove. Sixty. The data is not the obstacle. The obstacle is that you have to name the situation before you look — because if you go hunting through hundreds of combinations for one that looks wrong, you will find several, and all of them will be noise. The multiplicity defeats you long before the sample size does.

And whatever you name has to be worth more than it costs. Exchange commission runs at roughly two to five percent, so a small error in a common situation is not worth finding even if it is real. Only a large error in a rare situation pays for the hunt.

That is a narrower target than most people looking at this market imagine they are aiming at, and it is the one we think is left. We have no selection to offer you off the back of it, and no position of our own: this is measurement, not a tip.

Related research

• How often does the favourite win?
• What is Betfair Starting Price?

More notes

• The backtest that already knew the answerThe most flattering results in racing analysis come from tests that were handed the answer before they started. How it happens, and the one habit that stops it.
• Big fields and what they do to a favouriteMore runners means more ways to be beaten. That part is true. The awkward part is that the market worked it out before we did.
• What is the favourite-longshot bias in horse racing?Bettors have long taken too little of the favourite and too much of the outsider. Economists have studied it since 1949. What our 60,671-race study found.

Elite Pass publishes research, not guaranteed outcomes. Findings are measured against the market and remain subject to replication and sample size. The evidence room and the public ledger carry the full record. All insights.