Shot quality

The method

Every shot gets an expected make chance from a logistic regression fit on 219,152 2025-26 regular-season shots. It uses only what the play-by-play records: distance (2 ft steps), two or three, corner three, paint or midrange, shot type (layup or dunk, jumper, floater or hook), fourth quarter or overtime, the last 24 seconds of a period, and home or away.

Expected points = make chance x 2 or 3. Shot selection is a player's expected points per shot minus the league's (1.091). Shot making is actual minus expected points, per 100 shots. Players need 200+ shots to be ranked. Free throws are not included.

The test

• Fit on 175,364 shots, tested on 43,788 shots from 246 games it never saw (a random 20% of games).
• AUC 0.644 (0.5 is a coin flip, 1.0 is perfect). A zone-only average scores 0.632.
• Log loss 0.653 vs 0.659 for zone only and 0.692 for one league-wide FG% (lower is better). Brier score 0.231.
• Test games: league FG% 47.3%, mean predicted 47.0%.
• All shots: 239,129 actual points from field goals vs 239,153 expected (+0.01%).

What it cannot see

• How close the nearest defender was. A wide-open three and a contested one look the same.
• Play type (catch and shoot, pull-up, post-up, putback) and the shot clock.
• Who the shooter is. That is on purpose: the gap between actual and expected is the shot making.
• Fouls and and-ones, so drawing contact is not credited here.
• Playoff defense: on 2025-26 playoff shots it predicted about 4.5% more points than were scored.

Shot making is noisy: over 300 shots, luck alone moves it by about 6 points per 100 either way.