🐻⬇️🏀

Model Card

Box: game totals spread by minutes; cannot separate teammates; for seasons without play-by-play. Neutral-context sum of its offense and defense sides. Useful for separating orchestrators from players who mainly score without driving team assist volume.

Validation here is out of sample. From two weeks into each season, a model trained on earlier games predicts each following week without learning from that week. Predictions use the players' observed minutes; each point is the minutes-weighted predicted and actual rate in that player's games. The source inputs were not verified as identical to the published model revision.

Selected League / Season

Change league with the League selector at the top of the page.

Held-out player predictions

Plot Assists Box Elo against any counting stat (one point per player; choose any season with a served list).

Not enough held-out play to show a validation plot. NCAAM 2026-2027: 0 eligible players. A player needs at least 100 held-out minutes. Computed 2026-11-14 13:21:13.755478+00:00.

NCAAM 2025-2026: 3767 players with at least 100 held-out minutes. Across players, predicted and actual correlate at r = 0.93 (weighted by held-out minutes). Across these player averages, weighted R² = 0.76 (relative to the weighted mean of actual held-out outcomes). Per held-out team game, squared error is 18.7% below predicting the league average. Each dot is a player (500 with the most held-out minutes are plotted; larger dots had more); on the dashed line, actual equals predicted. Select a dot for the player.

Held-out miss against playing time

Each dot is a player's actual minus predicted team minus opponent assists per 100 possessions in the player's games against their held-out minutes (log scale). Orange points are the minutes-weighted mean miss of eight equal groups of players by playing time: a model that shrinks its low-minutes players too far, or not far enough, shows a slope.

Lineup RAPM vs Box Elo, per season

Both forms of this family are shown side by side wherever a season publishes them; this table is information, not a switch. Each column is the held-out correlation of that form's predictions with the actual per-segment margin (fit before a date, scored on later segments); within two paired standard errors the two predict equally well. Correlations are small in absolute terms: a segment holds about one such event. Held-out correlation does not penalize a Box form for not telling teammates apart; the season links show each form's team share.

Output Coverage

Rows in each model's current published run (or legacy rows where no run is selected).