🐻⬇️🏀

Model Card

Unregularized weighted lineup regression can be unstable, especially for players who share lineups. The fit uses a minimum-norm solution, then centers on published players by minutes. Segment outcomes are per 100 possessions, weighted by the square root of possessions, and scaled by 0.70 to per 40 minutes.

Validation here is out of sample. From two weeks into each season, every week's lineups are predicted by a fit on the season's earlier play, with the served settings; each player's point is the predicted and actual margin per 100 possessions of the held-out play the player was on the floor for.

Selected League / Season

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

Held-out player predictions

Plot Adjusted Plus/Minus against any counting stat (one point per player, every season with a served list).

NBA 2026-2027: 167 players with at least 200 held-out possessions. Across players, predicted and actual correlate at r = 0.37 (weighted by held-out possessions). Across these player averages, weighted R² = -0.76 (relative to the weighted mean of actual held-out outcomes). Per held-out segment, squared error is 34.5% above predicting the league average. Each dot is a player (167 with the most held-out possessions 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 margin per 100 possessions with the player on the floor against their held-out possessions (log scale). Orange points are the possessions-weighted mean miss of eight equal groups of players by playing time: a model that shrinks its low-possessions players too far, or not far enough, shows a slope.

Every held-out segment: implied against actual margin

2,271 held-out segments (stretches with one set of ten players on the floor), each predicted by the fit on earlier weeks. The shading counts possessions in each cell of predicted margin (the model's points for the segment, in 0.25-point bins; values beyond ±5 sit in the edge bins) against actual margin (points, integer rows). Orange points are the possession-weighted mean actual margin per predicted bin: on the dashed line, the model is calibrated.

Segments the model missed most

Ranked by |actual − predicted| / √possessions, among segments of at least two possessions, so short lucky runs do not crowd out real misses.

Output Coverage

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