Elite Athletics · case study
A national rowing team.
Selection is decided on a handful of races between crews that mostly never meet, in conditions that never repeat, with line-ups that change between every observation.
The model
An ensemble that predicts which boats win at practice.
A seat race is the one experiment every coach trusts: swap two athletes between crews, race again, see which crew got faster. It costs an afternoon and it settles exactly one question.
A squad of 28 gives 3,108,105 possible line-ups. Every pair of them that differs by a single seat is a seat race, and there are 248,648,400 — a quarter of a billion afternoons. The fastest eight the search found is twelve seconds up on the crew that had raced the most.
After nine practice races, it picked the winning boat 97% of the time.
It starts out guessing.
Every race is predicted using only the sessions that came before it. Nothing is ever scored on a race it was fitted on — which is the difference between a result and a demonstration.
29 of 30 races in the eights, across the 24 practices that followed the ninth. Measured over the entire record from a standing start, including the early races where it genuinely was guessing, it is 86%.
Margins compress.
In every class with enough races to measure it, the exponent is about a third. A four-length win and a two-length win are far closer in underlying ability than the water makes them look, which is why margin read literally is a bad selection signal.
Nothing here was a general model
It was built on one squad's own results, and it works for that squad.