Research

Do horse racing systems actually work?

Short answer: almost none. Of 71 systems and hypotheses with a documented research record here, 0 have cleared our validation bar.

Why this question is harder than it sounds

Most systems are tested against the wrong opponent. A system that picks the winner more often than chance is easy to build and proves very little, because the horses it likes are usually the horses everyone else likes too. The market has already priced them.

The only test that matters is whether a rule beats the price that was actually available. That is a much less flattering question, and it is the one almost every promising idea fails.

What counts as a system here

A system is any rule that takes pre-race information and produces a repeatable selection or ranking. Ratings-based rules, draw rules, pace rules, market-movement rules, trainer and jockey patterns, and combinations of those. The register counts a system as evidenced when its identifier appears in at least two separate files in the research repository, so an idea that was thought about but never written down is not counted.

71 is therefore a lower bound on what this repository can evidence, not a claim about everything ever examined. Work predating the repository, or explored outside it, is not in that number.

Betfair Starting Price is the benchmark

Every result here is measured against Betfair Starting Price, the price the exchange settled at. It is used because it is the closest available thing to what a person could really have got, it is recorded independently of us, and it cannot be argued with after the fact.

We have measured how good that benchmark is: on 1,997 paired races the exchange price was statistically indistinguishable from perfectly calibrated, and the bookmakers' price on the same races was not.

Beating a horse rating is not the same as beating BSP. A rating can order a field beautifully and still add nothing, because the ordering was already in the price. That distinction is the single most common reason a system that looks profitable on paper is not.

What repeatedly failed

These are our own ideas, tested and abandoned. They are listed because the failures are the part of this work with any evidential value.

Ideas tested against the market benchmark, and what happened
IdeaWhat we expected What happened
Top-rated horseBetter ability should convert to better returnsRanked well. The advantage was already in the price.
Draw position aloneA structural bias exploitable on its ownA clean null on held-out races. The idea was killed.
Market steamersMoney moving late knows somethingThey win more, and the advantage vanishes once you compare them with drifters at the same final price.
Favourite fragility signalsCertain races should expose short pricesNot established. Accruing prospectively, unread until the pre-registered sample completes.

Each of these has its own page with the sample and method. Links below.

Why a backtest is not evidence

A backtest tells you what a rule would have done on data you already had. If you tried several variants and kept the one that worked, the number you are looking at is partly a measure of how many variants you tried. That is why nothing here is treated as evidence until it has been frozen in advance and settled forward, against prices recorded before the race.

Strike rate on its own proves almost nothing: a rule selecting short prices will have a high strike rate and can still lose money. Return on its own is almost as weak, because a handful of long-priced winners can carry a losing rule for months. The comparison that survives is between what happened and what the price implied should happen, on the same races.

What is still unresolved

The open question is whether the structure of a race — the pace it is likely to be run at, the roles available, how much the field is separated on ability — carries information the price has not already absorbed. That question is being accrued prospectively under a pre-registered specification, with a fixed sample size and a fixed review point. No claim is made about it before that sample completes, and the record is published either way.

What would count as evidence

A frozen rule, declared before the races it is tested on. A pre-declared sample size. Settlement against Betfair Starting Price. A result that survives being compared with the price rather than with chance. And a published record of the attempts that failed, so the surviving one cannot be the product of quiet selection.

A system that beats chance is common. A system that beats the price is the only kind worth having, and we have not found one.

Read next