Every signal we've built so far has been about the horse. Does this horse have the form, the draw, the right ground, the ceiling clearance. But before any of that matters, a human being made a series of decisions: this race, this trip, this class, this day, this jockey. The trainer controls almost everything about where a horse turns up — everything except the draw.
So we asked a simple question: are some trainers systematically better at those decisions than others — and does the market notice?
We called it TrainerPlacement, and we validated it the hard way.
163,245 GB runs from January 2019 to June 2026. We trained a win-probability model on everything before 2025 and tested it — out of time, no peeking — on the 72,000 runs since. The model saw only placement features: how far today's trip and class sit from the horse's proven profile, days off versus the horse's best freshness band, field size, recent form. Crucially, it never saw the starting price and never saw the trainer's name.
Then three questions:
Q1: Yes. The placement-only model scored an AUC of 0.691 against a coin-flip baseline of 0.5. Where and how a horse is placed tells you a real amount about whether it wins. Trainers' decisions are not noise.
Q2: No. And this is the finding that matters. Adding every placement feature to a starting-price model moved the AUC from 0.7714 to 0.7710 — nothing. The market has already read the racecard the same way we did. Placement skill is real, and it is fully priced in. There is no standalone bet here, and we're telling you that plainly because the alternative — dressing a null result up as an edge — is how tipping services die.
Q3: This is where it gets interesting. Once you know what a trainer's placement patterns should produce, you can measure who beats that expectation and who falls short of it.
Top of the table: William Haggas — 183 wins in the test period against 142 expected from placement patterns alone. That's 29% above expectation, over three standard deviations, across 667 runs. Not luck. Charlie Appleby, Aidan O'Brien and Andrew Balding sit in the same quadrant: yards that extract more wins than their entries "deserve."
Bottom of the table is just as informative. A cluster of yards — mostly high-volume, big-field handicap operations — deliver 40–55% fewer wins than their placement patterns imply, at three sigma or worse. When one of their runners shows up with attractive-looking form, that form has a habit of not converting.
The catch, and we'll say it again: even Haggas's runners returned a market A/E of 0.44. The public knows he's brilliant and prices him accordingly. You cannot blindly back the top of this table and profit.
Elimination and adjustment — the same philosophy as every signal on the dashboard. You don't use TrainerPlacement to pick a winner. You use it to answer: when this yard's runner looks good on paper, should I believe the paper?
For a TP↑↑ yard: yes, and maybe a bit more than the form shows — especially with unexposed types where the placement decision is most of the information. For a TP↓↓ yard: discount the form-based positives. The pattern says they underdeliver, systematically, year after year.
The TP badge is now live on dre.cogtrix.eu, in the Trainer column of every racecard:
Hover the badge for the full numbers: wins versus expected, run count, sigma, and percentile. Context signal, not a tip — but it's context the market gives you no discount for ignoring.