How the NBA injury report works
What each number means, where the data comes from, how accurate it is and what it can't see.
The numbers, in one line each
| Number | What it means | Unit |
|---|---|---|
| Rating | How good a player has been, on a 0-99 scale anchored to league ranks. Mine, not 2K's. | 0-99 |
| Offense / Defense | How much better than an average player a team scores, and defends, with him on the floor, controlling for the other nine players. | points per 100 possessions |
| Worth | Offense plus defense over a replacement-level player, times his minutes. Offense and defense shares add up to it. | points of margin per game |
| Cost while out | What his absence takes off his team's margin each game, after the teammates who soak up his minutes. Usually less than his worth. | points of margin, ±2 |
| Talent missing | The sum of those costs for a team's players listed out or doubtful: from tonight's official report for teams that play tonight, otherwise from the current injury list. | points of margin |
| Wins over 82 | Worth over a full season. | wins |
| Team rating | The margin I'd expect against an average team in the next game, blowouts capped. Deliberately more modest than raw point differential. | points |
| Adjusted margin | A game's margin with home court removed and the opponent's strength that night added back, counting the opponent's own injuries. Team averages count a blowout as no more than 12 points. | points |
| Wins lost to injury | Every injured player's cost on every night he sat, added up and turned into wins. | wins |
The rating scale
| Rating | Level | Worth (pts/game) |
|---|---|---|
| 98 | Best in the league | 8.75 |
| 96 | Top 5 | 5.8 |
| 92 | Top 10 | 4.95 |
| 90 | Top 20 | 4.2 |
| 81 | Solid starter (top 60) | 3.2 |
| 78 | Top 100 | 2.4 |
| 75 | Median NBA player | 1.1 |
| 70 | Replacement level (the kind of player any team can sign) | 0 |
| <70 | Below replacement level so far | <0 |
The anchors were measured on Nov. 20, 2026 (406 players with five or more games) and are frozen, so a rating moves only when that player's own play does. They're re-measured only when the method changes (logged below). Ratings between anchors are straight lines. Where I show NBA 2K27 ratings, they're from 2K's official Top 100 list (the August 2026 launch list), labeled as 2K's.
Who sat, and why
The NBA's official injury report, posted every 15 minutes on game days. For past games I use the last report of the day and check it against the box score: a player counts as out only if he was listed and didn't play. Only injuries and illnesses count. G League assignments, rest and personal reasons don't, and "personal reasons" is shown exactly as listed. "Injury management" (load management for a known condition) counts as an injury when a player sits 3 or more games in a row and as rest when it's a single night off. Injury descriptions use the official report's wording. "Day-to-day" is ESPN's status, not the league's. Expected return dates are ESPN's estimates, labeled as such; I never set my own. About 37 report rows (mostly two-way players not on a roster) can't be matched to a player and are left out.
Player ratings: offense and defense
I rebuild who was on the floor for every stretch of every game from ESPN's play-by-play: about 8,000 stretches so far this season plus 40,000 from 2025-26. For each one I know the ten players and how many points each side scored per possession. A ridge regression credits each player while controlling for the other nine on the floor, with separate offense and defense numbers. This is the method behind RAPM, EPM and LEBRON, scaled down to public data.
Each player starts from his box score, and the on-court results move him only as far as the evidence supports. The box score part is learned, not hand-set: I fit how each stat per 100 possessions (points, two- and three-point attempts, threes made, free throws, assists, turnovers, rebounds, steals, blocks, fouls) relates to last season's on-court offense and defense ratings, which is how EPM and BPM build theirs. Steals (+0.8 per 100) and blocks (+0.5) carry defense; assists, made threes, free throws and offensive rebounds help offense, and turnovers hurt it. This season's box is blended with last season's, with an age curve; a player with no NBA history starts at replacement level. Garbage time is dropped (Cleaning the Glass's rule), free throws go to the lineup on the floor at the foul, team rebounds between free throws don't count as new possessions, and lineup changes between quarters are inferred from who appears in the play-by-play. Each player's rebuilt minutes are checked against his box score; 217 of 226 games matched within a minute at launch.
Worth, tonight's cost and talent missing
Worth is a player's offense plus defense over replacement level (−2 per 100 possessions, Basketball-Reference's convention) times his share of the game. It's the right measure of how good a player is.
Cost while out is narrower: it prices an absence against the teammates who absorb the minutes, weighted by how much each usually plays, and it lines up with how betting markets price star absences: Jokic about 5.3 points (markets: 5.5 to 6), Shai Gilgeous-Alexander 4.8 (about 4), Luka Doncic 4.3 (about 3.5), Cade Cunningham 2.5 (about 3). An earlier version needed a 0.65 fudge factor to get there; the learned box score fixed the cause, so it's gone. Treat single-game costs as ±2 points. Who takes the minutes matters, and I tested it: on 586 games from last season, spreading a missing player's minutes across the whole rotation, starters included, predicted both real minutes and final margins better than giving them to players at his position. That's why a good player on a deep team, like Dylan Harper on the Spurs, costs much less than he's worth. Wins lost to injury adds up the same cost, so the season total and the nightly number agree. The weekly Injury Luck Index ranks all 30 teams on it.
The game model and adjusted margins
Home margin = home court (0.5 points) + the gap between the two lineups that actually played + a small team term for what lineups miss. When rating teams, a blowout counts as no more than 12 points: in a rolling test that predicted each day's games from the ones before, capping cut the typical miss from 16.1 to 15.6 points, so a 38-point win over a bad team no longer outweighs a close loss to a good one. A team rating is the margin I'd expect against an average team in its next game, so it's more modest than a team's raw point differential on purpose.
Ranges and small samples
With/without splits show what happened and a best estimate that blends the games with what the model expected from the lineups that played (half and half at 25 games, because a single game's margin swings by about 15 points), with an 80% range around it. One or two games without a player says very little, and the pages say so. Win chances use the typical out-of-sample miss, 15.2 points; 80% of games land within 19.5 of the projection. When teammates missed the same games, the page says the split can't separate them.
How accurate is it?
I pick settings by backtest: ratings from games before a cutoff predict the margins of games after it, from the lineups that actually played, at three cutoffs (Nov. 6, 11 and 15). Typical misses on final margin:
| Model | Typical miss (pts) |
|---|---|
| Guess "even" every game | 16.2 |
| Hand-set box score (Game Score) only | 15.6 |
| Learned box score only | 15.5 |
| Game Score + both seasons' lineups (the first version) | 15.2 |
| Learned box score + both seasons' lineups (what I use) | 15.0 |
That's about 4% better than a box score alone and 8% better than guessing. Published metrics like EPM and LEBRON report 5 to 10% over box scores, so there's still room to improve. Two honest caveats: the three test sets overlap (it's really about 100 games), and the settings were tuned on the same cutoffs. Recency weighting (a hot streak counting extra) made predictions worse, so "last 10 games" is shown as context only.
Against Vegas
The toughest test: forecast games using only what was known before tip-off (each team's players minus whoever that day's official injury report listed out, at their usual minutes), in three back-to-back blocks of games, each predicted only from the games before it, and compare with the closing DraftKings spread.
| Forecast (103 games) | Typical miss (pts) | Picks the winner |
|---|---|---|
| Guess "even" | 17.5 | 48% |
| My model | 15.9 | 70% |
| Vegas closing spread | 15.2 | 73% |
| Half my model, half Vegas | 15.5 | 74% |
Vegas is better, as it should be: it sees everything, including news that never reaches a box score. My lines agree with Vegas closely (correlation 0.95, about 3.5 points apart on a typical game). This test also changed the model: my first team ratings swung too far from even (their spread was 10.9 points to Vegas's 8.1), so ratings are no longer scaled back up after capping blowouts. Tonight's game cards show the Vegas line next to mine. I'll keep score as the season goes.
What it can't see
Coaching and scheme changes, how well players fit together, minutes limits for players returning from injury, and effort in games that don't matter. Several players out at once probably cost more than the sum. Early in the season, ratings lean on last season and on box scores; on-court data takes about two seasons to stabilize for a starter. Defense is where it's least certain.
Known issues I'm working on: ratings describe this season, not reputations. LeBron James, at 41, rates around 80th: he's averaging 15.7 points and 6.8 assists on below-average efficiency, and Philadelphia has been about 5 points per 100 possessions better with him on the bench. A stricter backtest (separate test sets, a holdout, closing spreads as a baseline) is next.
Data and sources
NBA official injury report; ESPN box scores, play-by-play, rosters, injury list and return estimates; NBA 2K27 Top 100 ratings. Not affiliated with the NBA, ESPN or 2K. I never republish ESPN's or Rotowire's written injury notes and make no medical judgments.
Downloads, updated hourly on game days: injury list with ratings (CSV) · team ratings, injury luck and the top-10 split (CSV). Frozen weekly copies for citing a fixed number: /assets/nba/archive/<year>-W<week>-teams.csv and -injuries.csv (this week: 2026-W47 teams, injuries).
How to cite: "NBA Injury Report, Jay Egger, jayegger.com/nba/injuries/, data through Nov. 21, accessed [date]." Numbers update hourly, so include the data date.
Changelog and corrections
| Nov. 20, 2026 (night) | Pregame test against Vegas closing lines (my model 15.9 typical miss, Vegas 15.2). Team ratings no longer rescaled after capping blowouts (the rescale made forecasts too extreme), so they now read as expected margin vs. an average team. Vegas lines added to tonight's games. |
| Nov. 20, 2026 (late) | Learned box score replaces Game Score (backtest 15.0 vs 15.2; fixes the bias toward efficient bench scorers), the 0.65 absence factor is removed (costs now match betting markets on their own), blowouts capped at 12 points when rating teams (rolling test 15.6 vs 16.1), "vs. top 10" split on team pages, rating anchors re-measured. |
| Nov. 20, 2026 (evening) | Audit fixes. Wins lost to injury now use the same cost as the nightly numbers (league total about 16 wins, was 51). "Injury management" counts as injury only in runs of 3+ games. Possession counts fixed (team rebounds between free throws). With/without best estimates shrink at 25 games (was 8), ranges centered on the estimate, win odds use the out-of-sample miss (15.2). Rookies with no history start at replacement level. "Talent missing tonight" now uses tonight's official report and only teams that play tonight. |
| Nov. 20, 2026 | Launch. Offense/defense ratings (two-season, prior-informed lineup regression), rank-anchored 0-99 rating, cost at 65% of the lineup math, ranges on splits, win chances. |
Corrections:
- Nov. 20: the first Injury Luck Index used each player's full worth for wins lost (Warriors 6, league 49.5). On the corrected basis it's Warriors 2.5, league about 16. The post was updated and says so.
- Nov. 20: "the model expected" on with/without splits was computed from the with-him record, which can be two games (Kawhi Leonard's showed +20.2). It now comes from the lineups that actually played.
- Nov. 20: a team's "not yet submitted" note on the official report leaked into one player's injury description (Austin Reaves). Fixed in the parser.
Found an error? Tell me and it'll be fixed and logged here.