How we rate a race

The unique elements behind a TrapMetrics race page, in plain English: point-in-time features, a calibrated chance, the TMR scale, measured going, track geometry, selection classes and the consensus forecast.

Open any race on the site and you will see a win chance and a TrapMetrics rating for each dog, a consensus forecast, and a handful of badges. This is what each one is, where it comes from, and what it has been measured to do.

1. Fifty-two things knowable before the off

For every runner the model computes 52 measurements from the dog's previous runs, never from the race itself. Times against the track and trip par over the last one, three and six runs. Split times against par. Whether those runs were clear or troubled, and the dog's record in each case. Early pace, slow starts, running wide, running on. Days since the last run, class moves, trainer form over the last hundred runners, the sire's and dam's progeny, the dog's record over this course and distance.

The rule is simple and absolute: if a measurement could not have been known at 08:45 on race day, it is not in the model. The evening before, when the card is published, the same measurements are computed for tomorrow, and the morning run recomputes them with anything that settled overnight.

2. A chance, not a rank

The model is a boosted conditional logit, which is a mouthful for a simple idea: within each race the six dogs' scores are turned into six chances that add up to one. The output is a win chance per dog, and the whole product rests on that chance meaning what it says.

It does. On the held-out record, dogs the model gives 25% win about a quarter of the time. The rank order is clean too: the model's top-rated dog wins 26.4% of its races and its bottom-rated 8.5%, with every step between in order.

3. The TMR scale

A 30% chance in a five-dog race and a 30% chance in a six-dog race are not the same achievement, so the site also shows the chance on a handicap scale. TMR 100 is an average runner in an average race. On held-out 2026 racing, dogs rated 115 or better won 40.4% of the time, and dogs rated 125 or better won 59.4%. The second group is rare, about one runner every two days, and the site says so rather than pretending selectivity is free.

4. Going measured from the dogs

The official going allowance is one number per meeting. We measure ours per race, from how every dog in the field ran against its own recent times, over 265,607 races. It correlates with the official figure, at +0.42, and it changes through a meeting in a way a single allowance cannot. We also tested whether weather adds anything on top: heavy rain in the 24 hours to the off matters a little, about a twentieth of a second, and nothing else does. That test is on the methodology page with the rest.

5. Track geometry we measured ourselves

Circumference and run-up to the first beam for all 24 tracks, worked out from our own split times with a standing-start model and checked against the published figures. Three published figures were wrong; ours agree with the timings. Those measures feed the speed ratings and they are on each track's page, because a bettor who knows a track has a 55-metre run-up reads a trap draw differently.

6. Selection classes, with their records

The site names a few recorded conditions and keeps their running records:

• Synergy: our model and the Timeform capture model both make a dog at least ten points likelier than the forecast price does.
• TMR only and TF only: one of the two does.
• Conflict: the capture model likes it and ours does not.
• RP Synergy: the same idea against the Racing Post forecast.

Each class is a published record, not a tip. When both models' top pick agree the dog wins 31.6% of the time, on 1,131 races, nearly double a random pick. The Conflict class is followed as a lay. The records are on the forward-test page, positive and negative alike, and none of it is advice.

7. The consensus forecast

At the top of each race is one forecast line: the most likely first and second, the reverse, and a trio. It is not our model's ranking. It blends three win-chance opinions, ours, the Timeform forecast price and the Racing Post forecast, and expands the blend into finishing positions. We measured why. On 4,466 settled races since 1 August the blend landed 11.7% of straight forecasts, against 11.0% for our model alone, 9.5% for Timeform alone and 9.0% for the Racing Post alone. The version that performs is the version shown, and it can leave our top pick out when both forecasts rank it lower. The tooltip says so.

8. The verdict

After the race, the page says what happened and how each measure did: whether the top pick won, whether the forecast landed, what the sectionals and the official comment say about how the race was run. A rating that cannot be wrong in public is not worth reading.