What TrapMetrics is

A calibrated probability engine for UK and Irish greyhound racing, published every morning with its own record attached.

TrapMetrics is a machine that rates every runner at every UK and Irish meeting by 08:45, tells you how sure it is, and keeps the receipts.

That last part is the whole point. Anyone can publish ratings. The question a bettor should ask is whether the ratings have been tested against what happened, by someone who was prepared to publish the answer either way. Ours have. The record is on the site, it updates every morning, and it includes the days we were wrong.

The one-sentence version

We turn five years of official results, sectional times, going and market prices into a win chance for every dog, we compare that chance with the published forecasts and the exchange, and we show you where they disagree.

What is underneath

The model is built on the official record: 1,512,527 runs across 267,046 races and 36,817 dogs at all 24 tracks, from June 2021 to this morning. For each runner it computes 52 measurements that were knowable before the off: how the dog's recent times compare with the track's par, its splits, its running style, its record when clear and when in trouble, its class moves, its trainer's form, its breeding. None of those measurements uses anything from the race itself. That rule is what makes the numbers honest, and it is enforced in code rather than promised in prose.

The output is a win chance per dog. We express it two ways. Model percent is the chance itself. The TrapMetrics rating puts it on a handicap scale, where 100 is an average runner in an average race, so a dog can be compared across fields of different sizes.

What makes it different

It is calibrated, and we can show you. When the model says 25%, dogs like that win about 25% of the time. That is a testable claim, and the evidence page tests it on every settled race.

It is measured against the professionals. On 8,012 held-out races the rating out-scored the Timeform forecast on both measures we use: a log-loss of 1.610 against 1.657, and a top pick that won 30.5% of the time against 27.7%. The difference is statistically significant. It also out-scored the GreyhoundStats rating on the same races.

It is honest about the market. The Betfair closing price is better than every rating we have tested, ours included: a log-loss of 1.451 and a top pick that wins 39.9%. We publish that number too. A model that claims to beat the market on public information is either lying or about to be.

It measures the things others assume. We work out the going from the dogs' own times rather than the official allowance, over 265,607 races. We measured every track's circumference and run-up from our own split times and found three published figures that were wrong. Those measurements feed the ratings, and they are on the track pages for anyone to read.

It explains itself after the race. Every settled race gets a verdict: what the model said, what happened, and which of our own measures were right and wrong about it.

What it is not

It is not a tipping service. The site publishes selection classes, which are named, recorded conditions such as both models agreeing the market is at least ten points wrong. Each class carries its own running record, good or bad, and nothing on the site tells you what to back.

It is not a black box. The methodology page lists what we tried and killed as well as what we kept. The approaches that did not survive testing are on the same page as the ones that did.

Who it is for

People who already bet on greyhounds with some method, and want a second opinion they can check. People who build their own systems and want a back-tester that refuses to flatter them. And anyone who has paid for ratings before and wondered whether they had ever been tested.

Where to start

The evidence page shows the record. Today's card shows the ratings for every meeting. A race page shows one race in full, with the consensus forecast, the form, and after the off, the verdict. The System Builder lets you write a rule and test it against five years of results with the holdout scored separately.

Everything is free while accounts are built. The price, when it comes, is published on the pricing page now, so nobody has to wonder what the catch is.