How a player impact rating works
Singular gives every player one number: how many points per 100 possessions his team gains on its opponent while he plays, compared with the league's average player. Here is how that number is built, from box scores and lineup data to a forecast.
The unit: points per 100 possessions
Teams win by outscoring the other team, so the rating is measured in points of margin. It has an offense part (points his team scores) and a defense part (points the opponent doesn't score), and the two add up. A player at +2 on offense and +1 on defense is +3.
A possession is one trip with the ball. It ends when the other team gets the ball back; an offensive rebound keeps the same trip going. Because teams take turns, they finish a game with nearly the same count: in 2025-26, 99% of games ended with the two teams within two possessions of each other.
Some teams take more trips than others. In 2025-26 the Boston Celtics, the slowest team, averaged 95.7 possessions a game and the Miami Heat, the fastest, 104.3, counted from play-by-play. Counting per 100 possessions puts players on fast and slow teams on the same footing. What a rating is worth in a game still depends on pace:
Pace has changed across eras too. Teams in 1961-62 averaged about 126 possessions per 48 minutes, a box-score estimate (turnovers weren't recorded then). Play-by-play, a different method, counts 100.7 a game for 2025-26. Wilt Chamberlain's 50.4 points a game that season come to about 38 points per 100 of his team's possessions. Shai Gilgeous-Alexander scored about 44 per 100 in 2025-26.
More possessions help the better team
On one possession, anything can happen. Over many, skill shows. A better team's expected margin grows in step with the number of possessions, while the random swing grows only with its square root. So the stronger team gains from a faster game, and the underdog from a slower one. The effect is real but small:
A team 5 points per 100 better than its opponent wins about 64% of games played at 90 possessions and 66% at 110. Over a best-of-seven series that becomes 79% and 81%.
Plus-minus
The simplest impact stat is plus-minus: the team's margin while a player is on the floor. In 2025-26 Denver outscored opponents by 10.8 points per 100 possessions with Nikola Jokić on the floor. With Jonas Valančiūnas, the big man who played the most while Jokić sat, Denver was outscored by 1.7.
That gap mixes three things: the player, the four teammates beside him and the five opponents across from him. Backups tend to play next to other backups and against other backups. Starters face starters.
Players who averaged 28 or more minutes faced opposing lineups about 3.3 points per 100 stronger than players who averaged under 15.
Solving for all ten players at once
Play-by-play data split every game into stints: stretches when the same ten players are on the floor. For each stint we know who played, how many possessions each side had and how many points each side scored.
The model gives every player an offense rating and a defense rating. For any stint, start from the league's average scoring rate (plus a small edge for the home team), add the offense ratings of the five players with the ball and subtract the defense ratings of the five without it. That gives a predicted scoring rate for the stint.
Players move between lineups all season, so the stints can tell them apart. A backup who did well against weak lineups gets less credit for it; a starter who held his own against strong ones gets more. The method is called adjusted plus-minus.
One more step takes out some luck. Whether a three or a free throw goes in is partly chance, so each one counts at the shooter's expected percentage (with a small adjustment for the teammates around him) rather than at whether it dropped. One cost: a defense gets no credit for how well opponents shoot threes, which the model treats as luck.
Why each player starts from a guess
Stints are short and noisy. One hot shooting stretch can outweigh real differences between players, and one season of stints can't pin down every player on its own.
So each player starts at a guess, and the lineup data move him away from it. The more he plays, the further they can move him. Free throws work the same way: a player who makes 9 of his first 10 isn't a 90% shooter yet. The best estimate blends those ten shots with what we know about shooters like him.
In Singular the guess counts as much as 8,000 possessions of lineup data. A simplified picture, for a player whose guess is +2 and whose lineups alone would say +6:
That's how the model can rate a rookie or a deep bench player sensibly, and why one wild season doesn't swing a veteran's number too far.
Where the guess comes from
The guess comes from the box score. We take each player's points, shots, free throws, rebounds, assists, steals, blocks, turnovers and fouls per 36 minutes, with recent games counting more, plus his shooting percentages, minutes, height and age. A model trained on every earlier season turns these into a predicted lineup rating (its target is the lineup-based rating over three seasons): the box score's job is to predict what the lineups measure.
On offense, made shots, assists and offensive rebounds push the guess up; missed field goals and turnovers push it down. On defense, steals, blocks and height help. The defensive guess also uses one measure from play-by-play: how many more or fewer shots at the rim opponents took per 100 possessions with him on the floor, after adjusting for the other players on the court. More rim shots count against him. Each stat is also adjusted for the opponent, so ten rebounds against a team that rarely gives up rebounds count for more.
Last season carries forward too. If his lineup rating beat his box-score guess last season, this season's guess starts three quarters of that gap above his box score; if it fell short, three quarters of the gap below. A gap that repeats builds up season after season.
From this season to the next
The rating at the top of a player page is a forecast: what he should add over the rest of this season. A last step blends the lineup rating with the box-score rating (the guess without last season's carried gap), with weights chosen by how well each predicted the following season's lineup results. For a player with 4,000 or more minutes over the last three seasons, the offense forecast is about 0.81 times the lineup rating plus 0.26 times the box-score rating. The defense forecast is about 0.87 times the lineup rating plus 0.15 times the box-score rating. The weights differ for players with fewer minutes, and further terms adjust for role.
It also adjusts for age. Compared with a player of 27 or 28, the age step adds about 0.4 on offense for a player 22 or younger and takes off about 1.1 on offense at 33 or older. The defense step is mostly small, but at 21 or younger it takes about 0.3 off, so his net change is close to zero (at 22 it is +0.3); at 35 or older it takes another 0.2 off, for −1.4 in all. The exact size also depends on his role and how good he is. During the season, each game's box score nudges the forecast.
Each finished season also gets its own rating, as played: that season's lineup rating, which starts from his guess, with no forecast step. The season ratings below are as played.
Reading the number
Zero is the league's average player, weighted by minutes: in every completed season, pick a minute of play and any player on the floor, and on average he rates zero. Among the 279 players with 1,000 or more minutes in 2025-26, the median was 0.0. About one in five (53) were +2 or better. Six were +5 or better, led by Victor Wembanyama (+9.8), Nikola Jokić (+9.6) and Shai Gilgeous-Alexander (+8.5).
Ratings add up to teams. Weight each player by his share of the team's minutes and add up five players' worth, and you get an estimate of the team's net rating. Oklahoma City's 2025-26 players add up to +10.1; the Thunder's actual net rating was +11.1. New Orleans: −4.5 and about −4. The sums spread out less than real net ratings, in every season since 1996-97 (about 80% as much on average), likely because real results include luck that the ratings leave out, and each rating is pulled toward its guess.
How we check it
A rating earns trust by predicting games it hasn't seen. We freeze each model's ratings on a date, using only games played before it, and score them on the lineups and margins that follow. Frozen on January 6, 2026 and scored on the rest of that regular season, Singular's ratings did at least as well as those of DARKO and EPM, two widely used public models. The test has limits: it covers one season, its main scoring fits each model's scale to the games being scored, and the current version of the model was adopted after the test had been run.
Defense is the hardest part. The box score records little of it, so the defense forecast leans more on lineup data, which are noisy. The method page has the details and the full test results.