Every game is split into stretches between substitutions, when the same ten players are on the floor, about 40,000 a season. Because each player shows up in hundreds of those stretches, with different teammates and opponents, his part of the score can be separated from theirs. Example: if the Spurs are +17 whenever Wembanyama plays, but most of those minutes are with good starters who are also good without him, part of the +17 belongs to them. Regular season and playoff games count, preseason games don’t.

This season’s ratings count this season’s games only. Each one starts from what his box score says so far, read through what each box-score number was worth in ten seasons of these ratings, together with his minutes a game, the coaches’ own read. His shooting is pulled toward the league’s percentages until he has taken enough shots to trust them, and each team’s players together are moved halfway toward the team’s own margin so far. Then the rating moves with every possession of lineup data, the start weighing about as much as 7,000 possessions, so a few games can’t swing it. Nothing carries over from last season, so every player started the season even.

Tested on every season from 2018-19 to 2025-26, each rated after 5, 10, 20 and 41 games from that season’s games alone: these ratings predicted the rest of the season better than this season’s first version in all 8 seasons at every point (after 10 games, an error of 14.1 points a game against 14.5, and 65% of winners against 62%), and by midseason as well as the ratings that count last season. With last season is the rating the way every earlier season is rated, described next.

With last season: Each season’s rating starts from a starting point and moves with every possession, with the start weighing about as much as 5,000 possessions, so a few games can’t swing a rating. The start is 0.9 times his latest rating, with a quarter of it from what his box score that season says, then moved by how players his age usually change from one season to the next. A player in his first season starts a little below average, since new players usually are, adjusted for where he was drafted, and his own box score counts more as his minutes add up. Tested on every season from 2018-19 to 2025-26, these starts predicted the rest of the season better than starting from last season’s rating alone, before the season and after 5, 10 and 20 games, in all 8 seasons. Feeding this season’s box score into the ratings of players rated before, or letting a few games move a rating faster, predicted worse.

Likely rank: every rating has an uncertainty, wider for players with fewer possessions. Drawing every player’s rating 2,000 times from its own uncertainty and ranking each draw gives the range his rank falls in 8 times out of 10. Two players whose ranges overlap a lot are close to a coin flip apart. As a check, ratings built from one half of a season’s games and from the other half differed about as much as the uncertainty said, a little less, so the ranges lean cautious.

Change is this season’s rating against last season’s. Early in the season every rating sits nearer average, so changes go unmarked until the season is over; then a change is marked in color only when it is bigger than the uncertainty in the two ratings together.

On/off is simpler and noisier: his team’s margin per 100 possessions with him on the floor, and in the same stretch of the season with him off it (games he missed while on the roster count as off, as on NBA.com). It includes everyone he played with and against, so it can disagree with the rating.

Clutch is what happened in the last 5 minutes within 5 points, from the Clutch tab: how much his shots, free throws, turnovers and passes moved his team’s chance to win. Research hasn’t found clutch play to be a repeatable skill, so it is shown as a record, not as part of the rating.

Total: (rating + 2.0) × regular-season possessions ÷ 100 ÷ 35.2. A replacement-level player, the kind a team can sign at any time, rates about −2.0 per 100. Over the last 10 regular seasons, 35.2 points of scoring margin was worth one win. Added up by team, the players’ totals track team records closely, with a typical miss of about 3.5 wins.

Tested on every season since 2017-18 with ratings from before that season: told who actually played, they picked the winner of 65.1% of games, against 56.1% for always picking the home team.

Per 100 ranks the 271 players with 317 or more possessions, regular season and playoffs. Total ranks every player who played in the regular season.

The card’s “next season” line: every pair of back to back seasons with 2,000 or more possessions from 2016-17 to 2025-26, grouped by age and by how good the player was, with the average change in rating and the range 8 in 10 changes fell in. Only players who kept a big role both seasons count, so it says how ratings move for players who stay in a rotation.