🐻⬇️🏀

Model Card

Same paired-comparison form as Bradley-Terry, but newer games receive more weight. This is the simplest bridge between static BT and fully dynamic team models.

Early in a season this model's ratings are blended with each team's preseason prior by games played: a team with n games keeps 6/(n+6) of its prior, so the order settles as games come in. In the pro leagues, whose prior is the team's Elo carried over from last season, it keeps 4/(n+4) in the NBA and WNBA, 8/(n+8) in G League and Euroleague and 2/(n+2) in Eurocup.

Validation here is out of sample: each week of a season is predicted by models refit on the season's earlier games, and every model is scored on the same games. Weekly fits use a preseason-prior blend for eligible families where available; ensembles use those components. Remaining families are scored as fit. Current scope: All Games. All scored games in the selected league and season. AP Poll is excluded here.

Held-out game predictions

Weekly out-of-sample games in the comparable full-season cohort. Prediction is the probability that team 1 wins; actual is 1 for a win and 0 for a loss. Plot: 400 of 4742 games (up to 25 largest errors and an even sample of the rest); calibration uses all 4742 games. Select a point to inspect its game and prediction; a verified game link appears when available. The table lists up to 25 largest prediction errors.

Calibration by predicted probability

Largest errors

Peers in NCAAW 2025-2026

Every model on the same games, ordered by full-season AUC.

League × Season Summary

Bradley-Terry Recency in every evaluated league-season; rank among the models on the same games.

Output Coverage

Stored site-output rows for this model.