September evidence update: 13,018 races, and what moved

This month's snapshot of the settled record: the top-pick comparison with Timeform, the straight-forecast study, and two corrections, one of them to this article.

Correction, 12 September 2026. This article first said the comparison covered races "since 1 January". It did not. The Timeform forecast only enters our capture on 18 May 2026, so the window is 18 May to 11 September. The figures themselves are unchanged. We have also separated two things that were run together: this comparison scores the current model over past races, which is a backtest, because retraining on 28 July rewrote the stored ratings for earlier races. The record of ratings exactly as published is below, and it is the smaller, more honest number.

Once a month the figures on this site are recomputed from the settled record, and the date goes beside every number. This is the first snapshot under that routine, dated 11 September 2026.

The top-pick comparison

On 13,018 races since 18 May, racing the model never trained on, the TrapMetrics top pick won 32.3% of the time. The Timeform forecast favourite won 28.7% of the same races. That is 3.6 percentage points, with a 95% interval of 2.7 to 4.5.

The Betfair favourite won 37.0%. The closing market beats every rating we have measured, ours included, and we show that number beside ours every time.

The record as published

The figures above score today's model over past races. Because the model was retrained on 28 July, the ratings stored against races before 1 August are not the ratings anyone saw at the time, so that comparison is a backtest.

The live record, using the ratings exactly as they were published each morning before racing, runs from 1 August to 11 September: 3,714 races over 41 racing days. The top pick won 32.0%, the Timeform forecast favourite 29.8%, the Betfair favourite 36.3%. The gap to the forecast is 2.2 percentage points, with a 95% interval of +0.6 to +3.7, resampled by racing day so a single good day cannot stand in for many good races.

Six weeks is a short record. It is the one we will keep publishing as it grows, beside the longer backtest, and we would rather show you both than pick the flattering one.

What changed in how we count

The August figure was 30.5% against 27.7%, on 8,012 races. The count is now stricter in two ways. A race only counts if every runner that ran has a rating, a Timeform forecast and a Betfair SP, so nothing is compared on a partial field. And when two dogs share top spot in a forecast, a winner among them scores a share of the credit rather than depending on which one happened to be listed first. The gap to Timeform widened slightly under the stricter rules, which is the direction you want a result to move when you tighten the test.

The straight-forecast study

Since 4 August, when the Racing Post forecast joined the morning capture, 3,887 races have settled with every source available. The model's most likely pair landed the straight forecast 10.9% of the time, Timeform's 9.2% and the Racing Post's 9.1%. A blend of all three landed 11.3%, which 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 grade, distance or agreement filter we have tested gets past that.

One number we corrected

Several pages said the System Builder tests rules against 1.56 million runs. That count included dogs declared but not run. The Builder actually tests against 1,512,527 runs with a result, in 267,046 races since June 2021, and the pages now say so.