How accurate is it?

We fit the model on one season and grade it on a different season it has never seen, which is the same test December runs on everyone. Everything below is measured on 30,373 players the model was not shown.

Did we spot the people who moved?

This is the number that matters and the one nobody publishes. Most players do not change level in a year, so a model that simply says “nobody moves” is right about 85% of the time and helps nobody. The question is whether we caught the players who did move.

Of 6,376 players who actually changed level, we called 3,709 of them correctly — 58%.

	We spotted	When we said so, we were right
Players who went up	62%	52%
Players who went down	54%	45%
Players who stayed put	84%	88%

Demotions are harder than bumps, and that is not a shortcoming of the model so much as of the data: a promoted player cannot keep playing their old level, so a bump is forced into the record, while a demoted player often stays on the same team and nothing marks it.

The technical version

We fit the model on one season and grade it on a season it has never seen — the same test December runs on everyone. On 30,373 held-out players it beat a “nobody moves” guess on both measures we hold ourselves to, cutting log loss by 34%.

Graded on year-end 2025, which the model never saw during fitting. Log loss 0.4367 against 0.6593 for a base-rate guess; Brier 0.2559 against 0.3537. We do not quote a headline percentage: most players do not move in a year, so a naive guess is already right about 85% of the time and an “86% accurate” claim would imply far more than it shows. These are our numbers, on our model — USTA has never published its algorithm.

These are our numbers, on our model. USTA has never published its algorithm, and we are not affiliated with them.

Back to search