"Trainer in great form", "top jockey booked" — connections are quoted in every preview. They do matter. But how much is genuine predictive signal, and how much is small-sample noise dressed up as insight?
The signal that's real
Over large samples, the best trainers and jockeys genuinely outperform — better horses, better placement, better in-running decisions. A yard's long-run strike rate and a jockey's class are real, persistent edges that a model should absolutely use.
The noise that fools people
"Trainer is 5 from 7 this week!" sounds compelling and means almost nothing — seven runs is far too small to separate skill from luck. Hot and cold streaks are mostly variance (the same variance that gives every process losing runs). Betting a horse purely because its trainer had a good Tuesday is chasing randomness.
How to tell them apart
- Trust long-run strike rates and course/condition records over recent streaks.
- Be wary of any stat quoted over a tiny sample.
- Ask whether the "form" reflects better horses, or just a lucky few days.
How a model uses connections
Our model includes trainer and jockey factors built from long histories, weighted by how much the data says they actually predict — not by this week's headline. That's the advantage of letting decades of results decide importance rather than a hunch. It sits alongside ratings, going and draw. See the data behind the model.