TrapMetrics2,836 winners from 6,887 runners
The record behind the productStronger claims.
Stronger claims.
The numbers beside them.
Every figure here comes from our own settled record, with its race count, its window and its date. The closing market beats every rating, ours included, and that result is shown too. Snapshot 11 September 2026, refreshed monthly.
Straight forecast / since 4 August 2026Measured against
Measured against
Timeform.
And the Racing Post.
The model’s most likely pair landed the straight forecast more often than either published forecast, on the same races.
10.9%TrapMetrics model
straight-forecast hit rate
straight-forecast hit rate
3,887 settled races from 4 August to 10 September 2026. A straight forecast is the first two home in the right order.
Straight forecast / same races
Combine all three opinions11.3%
The three-source blend landed more than any source alone. That is why the racecard shows a consensus forecast.
A hit rate is not a return. The forecast pool keeps about 28% of every pound staked, and no filter we have tested beats it.
Held-out test / Top-pick accuracy
32.3% versus 28.7%.
The TrapMetrics top pick against the Timeform forecast favourite, on the same held-out races.
That is 3.6 percentage points, with a 95% interval of +2.7 to +4.5, on 13,018 races and 71,230 runners.
A separate comparison from the straight-forecast study above, on a larger set of races.
The Betfair favourite scored higher than both. BSP is only known at the off, so it is a benchmark, not a price you could have taken in the morning.
Probability quality
How the whole set of chances scored.
Log-loss rewards putting probability on the winner and punishes confident mistakes. Lower is better.
| Source | Log-loss | Top pick won |
|---|---|---|
| Betfair SP | 1.491 | 37.0% |
| TrapMetrics Rating | 1.576 | 32.3% |
| Timeform forecast | 1.629 | 28.7% |
Sources & definitions
How each figure is measured.
Top pick and log-loss
Races from 1 January to 10 September 2026, which the model never trained on. Complete fields only: every runner that ran has a rating, a Timeform forecast and a Betfair SP. Each source’s chances are normalised within the race. When joint favourites share top spot, a winner among them scores a fair share. 13,018 races, 71,230 runners.
Straight forecasts
The 08:45 ledger from 4 August, when the Racing Post forecast joined the capture, to 10 September 2026: 3,887 settled races. Each source’s win chances are expanded into ordered pairs with the same correction the racecard uses, and its most likely pair is scored against the first two home. The blend averages the model, Timeform and the Racing Post.
Rating thresholds
Every settled runner with a TMR at or above the threshold, from 1 January to 10 September 2026. Not one runner per race, and the higher groups sit inside the lower ones: 115+ won 41.2% of 6,887, 120+ won 49.9% of 1,654, 125+ won 62.7% of 386. 115 is the scale’s built-in “twice the average chance” mark; 120 and 125 were looked at after the fact, so read them as description, not as a rule.
The System Builder record
1,512,527 runs with a result, in 267,046 races since June 2021: the table every Builder rule is tested against, settled at Betfair SP after commission.
Refreshed monthly
This snapshot is dated 11 September 2026. The figures are recomputed at the start of every month by the same queries, so if a number moves, this page moves with it.
What these figures do not say
A hit rate or a strike rate is not a return. The closing market is more accurate than every rating here, ours included. Nothing on this page is a promise about the next race.
Your next step
Find your place in the research.
Try the tools, read the evidence and choose the access that fits your questions.
View access optionsOpen a real race
Register before accounts open and Research is £19.99 a month, locked in.