How it works

Everything here comes from ESPN's play-by-play: a log of every shot, rebound, foul, turnover and substitution in every NBA game. From the substitutions we know which ten players were on the court at every moment, and that's what the ratings are built on.

Words you'll see

Possession

One turn with the ball. A possession ends when the team scores, misses and the other team gets the rebound, or turns the ball over. Each team gets about 99 possessions a game.

Points per 100 possessions

The unit every rating uses. Comparing per possession, not per game, is fairer: fast teams get more possessions, so they score more points per game without being any better. "+3 per 100" means about 3 more points over a full game's worth of possessions.

Plus-minus

How much a team outscored the other team while a player was on the court. A player whose team wins his minutes by 5 points is +5.

Adjusted plus-minus

Plain plus-minus is misleading: a bench player who always shares the court with the team's star looks great. Adjusted plus-minus works out each player's share by comparing thousands of different groups of ten players, so each player is credited only for what changes when he is out there.

Box score

The basic stats line: points, rebounds, assists, steals, blocks, turnovers, shots made and missed.

Rotation

The players a coach actually uses. A team has 240 minutes to share out (5 players for 48 minutes); most of them go to 8-10 players.

Back-to-back

Playing on two nights in a row. Teams are worse on the second night.

Play-in

A mini-tournament for the teams that finish 7th to 10th in each conference. Two of the four get the last two playoff spots.

The ratings at a glance

• Overall - the main ranking. It blends his Impact with the size of his role: scoring, usage (how much of the offense runs through him), creating his own shot, playmaking, rebounding, steals and blocks, shooting efficiency on his volume, and minutes. See "Overall: impact and role" below.
• Impact - how many points a game he adds to his team compared with a replacement-level player (the kind of player any team can sign at any time): how good he is on each trip down the court, times how many minutes he plays. This is what the game predictions use.
• Offense - the offensive side of Overall: how much more his team scores with him on the court, plus the size of his offensive role - scoring, usage, creating his own shot, playmaking and efficiency on his volume. Stars who carry an offense take harder shots and usually shoot a little worse than role players; that's the job, so the role counts here, not just efficiency.
• Defense - how many fewer points the other team scores with him on the court (from his on-court results, his box score and who he guarded), compared with other players at his position. Centers pile up blocks and rebounds that guards can't, so guards are rated against guards and centers against centers. Each player's "Why" panel also shows the raw evidence: what his team allowed with him on the court and off it, and for centers, how the defense did against small lineups with no big man (a listed center or anyone 6'10" or taller), where he has to guard smaller, quicker players. These are shown for context and aren't part of the rating. We tested counting how players defend small and two-big lineups in the rating itself, comparing everyone on the court: the real differences between players turned out to be tiny (about a tenth of a point per 100 possessions), so the swings you see in these raw splits are mostly chance, especially under about 1,500 possessions.
• If he's out - how many points per game his team's predicted result drops when he misses a game and his minutes go to his actual teammates.
• Minutes - how much he's expected to play in a normal game, from his last 10 regular-season games (before he's played this season, his last 40 - the last few games of a season are a poor guide). Playoff games don't count: rotations shrink in the playoffs. A team has 240 minutes to give out; the deepest bench players lose minutes first.
• Confidence - how much of his rating comes from his own results on the court, rather than his box score (below). Low means it's still mostly the box score.
• Positions - guards, forwards and centers, from each team's roster. The small "#3 centers" under Overall is his rank among players at his position.

Tap any player for "Why he's rated here": every step below, with his own numbers.

Overall: impact and role

Impact (below) is the most accurate measure of how much a player helps his team win - it's what predicts games best. But it can rate an efficient role player, like a backup center who finishes lobs and protects the rim in 20 minutes a night, almost as high as a star who carries an offense for 36. Coaches, scouts and fans also weigh how big a role a player can handle: being the one the defense is built to stop, creating shots when the play breaks down, running the offense night after night. So Overall blends the two.

How much each part counts isn't our guess. We looked at every season from 2020-21 to 2025-26 and found the mix that best picks out the players who were chosen as All-Stars - the league's own verdict (coaches, media and fans) on who its best players are. On seasons we left out while working it out, the blend's top players included 107 of 147 All-Stars; Impact on its own found 79. In order of how much they count:

Each is compared with the average NBA rotation player, over the same games as the rest of the ratings (this season in full, earlier seasons counted down). Team winning is his team's record in the games he played - All-Star voters reward players on winning teams, and adding it picked out more All-Stars. Tap a player to see his own numbers for each.

Regular season only. These are regular-season ratings: every number here - Overall, Impact, Offense and Defense - comes from regular-season games. Play-in and playoff games aren't counted (they'll get their own ratings). Preseason games never count.

Time missed. When a player sat out much of last season, his numbers come from older seasons, which can make him look as good as he was before he got hurt. In our data, players who missed a few games came back about the same, but those who missed most or all of a season came back clearly worse - and more so the older they were (over 30). So Overall takes off points for the share of last season he missed, more for a whole season out and more again past 30. Age on its own doesn't count: Overall is about how good he is now. Checked on seasons left out while working it out, this made our preseason Overall closer to how players actually did. It fades away as he plays games this season. It only counts on Overall: Offense and Defense say how well he plays when he's on the court, so a great defender who was hurt for most of last season still shows as a great defender.

Rookies. A player with no NBA games starts where newcomers have been on average: about 5 points per 100 possessions worse than an average player, below a typical bench player. That includes top draft picks. We tested starting them by draft pick, and expecting top picks to play more minutes. Both made predictions worse, because even No. 1 picks usually struggle in their first season. Rookies are listed by draft pick until they play, and their ratings move quickly once they do.

Stats and the season. A player's stats box switches to this season after his first game. The numbers behind Overall (points, assists and the rest) blend seasons the same way the ratings do: this season's games count fully and last season's about a third as much, so this season takes over gradually. After about 25 games it's half of the number. That way one big night doesn't jump a player up the list.

How Impact is worked out

1. What his box score says. We worked out what each part of a player's game is worth by comparing players' stats per possession with their on-court results (adjusted plus-minus) over several seasons:

• Shooting, by where it's from - threes, shots at the rim and mid-range shots count separately, makes and misses alike. A made three per 100 possessions is worth about +1.8 points; a missed mid-range shot costs more than a missed layup.
• Creating his own shot - baskets he made without an assist. Players who can score on their own make life easier for everyone else.
• Passing, rebounding, turnovers, fouls, steals and blocks.
• Position - the same stat can mean more for one position than another. A steal or block by a guard is a better sign than the same play by a center (the idea comes from Basketball-Reference's Box Plus/Minus).
• Minutes per game - coaches give minutes to the players who help them win, so it's evidence too. It also stops a bench player with big numbers in a few garbage-time minutes from looking like a star.
• Passing counts on offense only. Left to itself, the math also gave great passers credit on defense, because passers tend to play on good teams. That made Nikola Jokić a top-20 defender. Now passing only counts toward Offense. Turnovers and carrying the offense still count on defense: a lost ball often turns into an easy basket the other way, and players who carry the offense tend to save energy on defense. This made predictions slightly better, not worse.
• Who he guarded - from the NBA's arena cameras, which track which defender is on which player all game. Three things count: how many points the players he guarded scored on him compared with what they usually score, how good those players are (coaches put their best defenders on the other team's best scorers), and how often he guarded someone 3 or more inches taller or shorter than himself. That last one is versatility: defenders who can switch onto bigger and smaller players help more. Adding these made game predictions better in every season we tested. We also tried turnovers forced against the man he guards, but it made predictions worse, so it's left out.

2. Team adjustment. If a team did better than its players' box scores add up to - usually because of things box scores can't see, like defense away from the ball - each player gets a share of the difference, by playing time. If it did worse, each takes a share of that.

3. His on-court results. From every substitution we know who was on the court for every possession. Adjusted plus-minus works out each player's effect on the score after accounting for the other nine players. Two kinds of luck are taken out: free throws and threes are counted at each shooter's usual percentage, not whether they happened to drop - whether an opponent makes his free throws isn't the defense's doing. Minutes at the end of blowouts (a 20-point lead in the 4th quarter) don't count.

4. The blend. On-court results are the truest measure, but a few games of them can be luck, so they count for more the more he plays:

• Low confidence - under about 2,000 possessions of evidence (20 full games): mostly the box score.
• Medium - 2,000 to 6,000: both count.
• High - over 6,000: mostly his own results.

Defense leans on on-court results sooner than offense, because box scores see much less of it - a guard who stays in front of his man and contests shots may never get a steal or a block. The ratings use the last three seasons: this season counts in full, last season about a third as much, the season before about a tenth. Only real games count - never preseason games. A player with no NBA minutes at all (most rookies) starts at about 5 points per 100 below average, which is how such players have played on average.

5. Impact. His rating per 100 possessions, minus replacement level (about −1.9 per 100 - the average of the players just outside the league's 300 or so rotation spots), times his minutes per game, is the points a game he adds.

Every setting was chosen by testing: we rebuilt the ratings week by week through past seasons, using only games played before each week, and kept whatever predicted the next week's games best. Ideas that sounded right but made predictions worse were left out - counting recent games more than older ones, counting five seasons instead of three, and adding on-court team stats (rim shots allowed, threes allowed) to the box score.

The 0-99 scale

Every rating is on a 0-99 scale whose marks come from where real players ranked at the end of the last four seasons. A 99 is as good as the season's best player:

Offense and Defense have their own scales built the same way, so a 90 Defense means one of the best defenders in the league for his position. The per-100-possession numbers behind them are in each player's "Why" panel.

Game predictions

Each game's predicted result adds up:

• Lineup strength - each team's players' ratings, weighted by the minutes each is expected to play, leaving out injured players (players listed as day-to-day count at half their minutes). This is most of the prediction.
• Home court - about 2 points for the home team.
• Elo - a running team score that goes up with wins and down with losses (the system FiveThirtyEight published for the NBA). It adds a little on top.

We also tested rest (teams playing the second night of a back-to-back): it mattered with the first version of the ratings, but once the ratings improved it no longer made predictions better, so it's left out.

The predicted margin becomes a win chance: in past seasons, results landed within about 13 points of the prediction two times out of three. We tested it on the whole 2025-26 season without letting the model see any of it: our picks won 69.0% of games; picking the betting favourite every time won 68.7%. Our predicted margins missed the final margin by 14.02 points on average, the same as the betting line (14.02) - level with the market, not ahead of it. When the model said 80-90%, the team won 88% of the time.

Projected wins and playoff chances

We play out every remaining game 20,000 times, using each game's win chance. In each run every team also gets a random boost or drop, because over a season teams really do drift from their rating - trades, injuries, slumps. How big that drift was in past seasons sets its size (about 4 points a game). Then the play-in and playoffs are played out, best of seven, with home court to the team with the better record. Playoff series use the same team strengths as the regular season. We tested two playoff-specific ideas on every playoff game since 2021-22: giving stars more minutes (coaches shorten the bench in the playoffs), and counting players' playoff games in their ratings. Neither predicted playoff games any better - playoff games are close to a coin flip around the prediction even for the betting line - so neither is used. Projected wins is the average over all the runs; each chance is the share of runs it happened in. Averages bunch toward the middle: in a single season a few teams get hot and others get hurt, but no one knows which in advance. So no East team may average more than 51 wins, even though a typical simulated season has about four East teams at 50 or more - in line with recent seasons (two to four a year). The standings show each conference's typical season under its name.

What's simplified: tied records are split at random instead of by the NBA's tiebreakers, and each team's strength assumes its current roster, with only players ruled out for the season left out. The two NBA Cup games added to each team's schedule in December are treated as games against an average team until they're set.

The draft game

Sign in with your TOPICAL account and draft a team: two point guards, two shooting guards, two small forwards, two power forwards and two centers. Your team's rating is the average Overall of your 10 players, and it moves every day as their ratings do. Play it one of three ways: open bid (pick anyone, no trades), hard cap ($55 to spend, buy and sell all season) or extra hard cap (the same with $30).

ESPN only lists players as guards, forwards or centers, so we give everyone one of the five spots from how he plays. Guards who set others up (6 or more assists per 36 minutes, or 5 for guards 6'5" and under) are point guards; the rest are shooting guards. Forwards 6'9" and up, or who grab 9 or more rebounds per 36 minutes, are power forwards; the rest are small forwards. Centers are the players listed at center, plus forwards 6'10" and up who rebound like one (11 or more per 36) without running the offense. A player's price, $1 to $10, is his rank at his spot split into ten equal groups: the top 10% cost $10 and the bottom 10% cost $1. Players can't be bought or sold from tip-off until the ratings include that game.

Awards odds

The Awards page gives each player's chance of winning MVP, Defensive Player of the Year, Rookie of the Year, Sixth Man of the Year and Most Improved Player. It works in two steps.

What voters reward. We took the full voting from 2020-21 to 2025-26 (every player who got a vote, not just the winners, from Basketball Reference) and fitted which numbers from a finished season explain it. MVP: points per game, his team's record and Impact (the points he adds per game). Defensive Player: steals plus blocks, his team's record and our Defense rating. Rookie of the Year: Overall and minutes. Sixth Man: points and his team's record. Most Improved: how much his Overall, scoring and minutes went up from last season (a bigger jump in minutes counts slightly against him - voters want players who got better, not just more playing time). Each set of numbers was picked from several we tried, by how well it predicted seasons left out of the fit. The rules count too: MVP, Defensive Player and Most Improved need 65 games; Sixth Man needs more games off the bench than starting; Rookie of the Year is a player's first NBA season.

The rest of the season. We play it out 10,000 times. Each player keeps his numbers so far and plays his team's remaining games at his current level, plus a real story drawn at random from a player in the same spot in a past season: how many of the games left he actually played (injuries included), and how much better or worse he was than his level at the time. Rookies who haven't played yet get the real first season of a rookie drafted around the same pick. Team records come from the same runs as the standings on the League page. Then voters pick from each finished season, and a player's chance is the share of seasons he wins. Players ruled out for the season, or out until a set date, miss those games.

How we tested it. For each season from 2020-21 to 2025-26 we fitted everything without that season, predicted it from opening night and every 4 weeks after, and checked the chance we gave the real winner. Every part had to earn its place. Drawing real stories (instead of assuming everyone plays at his level) helped a lot. Matching stories by draft pick for rookies helped. Matching stories by years in the league, and adding random swings to team records on top, didn't help, so they're not used. The MVP winner was in our top 3 at 32 of 34 checkpoints. Defensive Player and Sixth Man are harder (voters follow stories as much as numbers). Before the season, Most Improved and Sixth Man are close to guesswork: we gave their real winners under 1% on average. So until our tested chances reach 5% at that point in the season, those awards show only the betting market's odds.

The betting market column is DraftKings' odds (via ESPN), with the bookmaker's margin taken out so the chances add up to 100%. Our odds are our chance written the same way: +400 is a 1-in-5 chance. Coach of the Year and Clutch Player aren't covered: we don't rate coaches, and clutch voting has too few seasons.

Where the data comes from

Play-by-play, box scores, schedules, rosters and injury reports: ESPN. Who guarded whom: NBA.com's matchup data, from 2020-21 on. Past seasons: the sportsdataverse project's copies of ESPN's data. Everything is updated through the day as games finish.