How to read APM

A player at +4.0 is estimated to add four points per 100 team possessions compared with a league-average player. Zero is league average, and negative values fall below that line. Minutes tell you how much evidence sits behind the estimate; they do not change the scale of the rating.

The model starts with the impact suggested by a player's box score, then adds the share of team performance that the recorded individual stats do not explain. That gives every player a final rating on the same points-per-100 scale.

Where the signal comes from

Training begins with RAPM. The play-by-play is split into possessions, with five offensive players, five defensive players, and a scoring result on each one. Weighted ridge regression works through those overlapping lineups to estimate individual effects, while shrinking unstable estimates toward zero.

Each training row is paired with a trailing RAPM window of up to five seasons. A player's window can contain fewer seasons when the full five-year history is unavailable, and lower-information estimates receive more help from a prior based on team quality and minutes per game. The site also shows three-year RAPM whenever a complete three-season window is available. Version 2.2 learned from the 2009–2025 seasons shown above, and the same fitted formula scores each newly finalized game.

How the box score becomes APM

Every box-score input is put on a per-100 team-possession basis. The model uses adjusted points, three-pointers, assists, turnovers, offensive and defensive rebounds, steals, blocks, fouls, field-goal attempts, and free-throw attempts. The adjusted-points step removes the team's shooting environment so individual scoring is credited in context.

Position and offensive role live on continuous 1-to-5 scales. Position comes from a player's share of team rebounds, steals, fouls, assists, and blocks, with 50 prior minutes at her listed position. Offensive role comes from assist share and her share of team scoring above an efficiency threshold. Both scales are centered so a team's minute-weighted average lands at 3.0.

Selected stats interact with those roles. In total APM, assists, rebounds, steals, and blocks vary with position. The offensive regression instead lets turnovers, rebounds, steals, and blocks vary with position. In both models, field-goal and free-throw attempts vary with offensive role. That lets the same event carry different information for a guard, wing, or center.

Putting the player back in her team

The raw box-score estimates rarely add up perfectly to what a team actually did. The model closes that gap with one shared adjustment. Minutes weight the raw team average, and the same correction is then added to every player in that team scoring group.

Rtg is the team's lead-adjusted net rating for that game; when a game-level value is unavailable, the schedule-adjusted season rating is used. The adjustment makes the minute-weighted player estimates line up with team strength. Season values remain weighted by the minutes a player supplied, so missed time and trades fall into place naturally.

Splitting offense and defense

OAPM comes from a separate offensive regression with its own coefficients and role terms. Defensive impact is the remainder: DAPM = APM - OAPM. Offense leaves a fuller trail in the box score, while defensive positioning, communication, screen work, and rotations are mostly invisible. DAPM therefore carries more uncertainty than OAPM.

Fit and how to use the ratings

Across - player-window training rows, total APM has a - correlation with its RAPM target. Mean absolute error is - points per 100 possessions, and root mean squared error is -. The offensive fit is tighter because scoring and creation are recorded more directly than most defensive work.

Start with APM, check the minutes behind it, then use on-court ratings, on/off, and three-year RAPM for context. RAPM itself can move when minutes are limited or the same players share most of their court time, and older play-by-play is not equally complete in every season. Current-season ratings will keep moving as new games arrive. The displayed tenths are useful for sorting players; they are not tenths of certainty.