Posterized

Updated 2026-11-16 · 1996-97 through 2025-26 regular seasons, plus playoffs and play-in

Behind the Numbers

Every number on this site comes from play-by-play logs that were never designed for analysis, run through models with knobs that have to be set somehow. This page shows the work: what was wrong with the data and how we fixed it, what we couldn’t fix, how the main model (RAPM) works, and the tests that decided each modeling choice — including the ones that didn’t pan out.

The figures below cover every regular season from 1996-97 through 2025-26 (35,546 games), plus 2,440 playoff and 37 play-in games. The current season is processed too, but it’s left out of the tests until it’s complete.


Part 1: The data

Where it comes from

Everything starts from NBA.com, pulled once and saved raw so the source never has to be hit again:

  • Play-by-play (every event: shots, fouls, rebounds, substitutions) — 1996-97 onward
  • Rotation data (when each player checked in and out, to the tenth of a second)
  • Box scores and game logs (points, minutes and plus-minus per player per game)

From these we rebuild every possession and every stint — a stretch of play with the same ten players on the floor — and record who was on offense, who was on defense, and how many points were scored. Every plus-minus number on the site is built from those stints.

Problems we found and fixed

The play-by-play feed is a scorekeeper’s log, not a lineup database, and it has changed format several times over 30 years. Turning it into “who was on the floor for every possession” meant fixing a long list of problems:

Problem What we did How often
Substitutions only name the incoming player as text (“SUB: H. Williams FOR B. Williams”), not by ID Match the name against the game’s roster, handling initials, accents, suffixes (Jr., III), family-name-first names (Yao Ming), and players who changed names mid-career 41,755 substitutions needed more than a plain last-name match
Teammates who share a last name, where only one has a suffix (Robert Williams III and Grant Williams on the 2020-21 Celtics) Match on the full name including the suffix before anything looser concentrated in 2019-2024
The name doesn’t match anyone on the roster uniquely Fall back to the rotation data to see who actually checked in 1,678 substitutions in 795 games
The outgoing player’s ID is sometimes wrong Correct it by matching the name against the players actually on the floor 9 substitutions
Events at the same clock time recorded out of order in older games Re-order by the feed’s internal event ID, not its display order throughout 1996-2000
Rotation data sometimes corrupt — stints that end before they start, missing players Detect it and infer lineups from play-by-play alone 290 games (83 of them in 2008-09)
Rotation data sometimes well-formed but wrong — players 20+ minutes off their box-score minutes Cross-check every player’s rotation minutes against the box score; more than 3 minutes off means we don’t trust it included in the 290
A substitution missing from the play-by-play, so the lineup stays wrong until the quarter ends, while the rotation data is intact Rebuild that game’s lineups from the rotation data; the result must still match every player’s box-score minutes 31 games recovered
Quarter starters missing when the rotation data can’t be trusted Infer them from play-by-play, carrying players over from the previous period when needed 23 games
Free throws credited to the wrong lineup See below every game
Some missed free throws logged with no player name and no “MISS” — just “Free Throw 2 of 2” — so they counted as makes Count a free throw as made only when the log shows the shooter’s points (“(7 PTS)”). In 30 seasons, every other free-throw row says exactly one of the two about 100 games, 2008-09 to 2016-17, each one point off
Part of the 2000-01 feed pads event types with trailing spaces (“Made Shot ”), so nothing matched Strip whitespace from every text field 11 games had crashed the parser
NBA.com’s own rate stats are wrong in early seasons (e.g. rebound % is 0 in more than half of 1996-97 player-games with rebounds) Ignore NBA-computed rates; recompute everything from raw counts all early seasons
Neutral-site games (Mexico City, Paris, NBA Cup semifinals) list both teams as the road team from 2024-25 on Take the designated home team from the box score, and flag the game about 5 games a season
A cancelled game (Celtics–Pacers, April 16, 2013, after the Boston Marathon bombing) still sits in the schedule with 0 minutes Drop games that were never played 1 game

The free-throw fix was the biggest single correction. When a player is fouled and a substitution happens before the free throws, who gets credit for those points? The official NBA convention is the lineup on the floor when the foul happened, not when the free throws were shot. Our first version used the latter, and only about 40% of player-games matched the official box-score plus-minus. After crediting free throws to the lineup at the time of the foul, 99.75% of player-games match exactly — 733,966 player-games across 30 seasons, with every season between 99.5% and 99.9%. Minutes match in essentially every player-game.

How every game is checked

Every game is rebuilt independently and checked against the official box score:

  • Does our final score match the box score?
  • Does every player’s plus-minus match? Every player’s minutes?
  • Does every stint have exactly five players per side?
  • Do both teams get a sensible number of possessions in each quarter?
  • Did every substitution resolve to a real player on the floor?

Each game gets a grade:

  • Clean — every check passes.
  • Minor — cosmetic feed problems (a running score out of order, a possession with an odd sequence of events) or lineups we had to recover from a fallback. Points and minutes still land on the right players.
  • Suspect — anything that could put points or minutes on the wrong players. Suspect games are excluded from every metric on the site.
Season Games Clean Minor Suspect
1996-97 1,189 1,081 82 26
1997-98 1,189 1,119 65 5
1998-99 725 702 22 1
1999-00 1,189 1,155 30 4
2000-01 1,189 1,150 39 0
2001-02 1,189 1,154 33 2
2002-03 1,189 1,136 51 2
2003-04 1,189 1,145 39 5
2004-05 1,230 1,194 30 6
2005-06 1,230 1,169 58 3
2006-07 1,230 1,185 44 1
2007-08 1,230 1,186 43 1
2008-09 1,230 1,114 116 0
2009-10 1,230 1,151 76 3
2010-11 1,230 1,182 46 2
2011-12 990 892 97 1
2012-13 1,229 1,164 65 0
2013-14 1,230 1,118 110 2
2014-15 1,230 1,128 99 3
2015-16 1,230 1,109 118 3
2016-17 1,230 1,077 151 2
2017-18 1,230 910 318 2
2018-19 1,230 1,116 110 4
2019-20 1,059 970 87 2
2020-21 1,080 978 100 2
2021-22 1,230 1,174 55 1
2022-23 1,230 1,181 47 2
2023-24 1,230 1,187 43 0
2024-25 1,230 1,182 48 0
2025-26 1,230 1,189 40 1
Regular season 35,546 33,198 2,262 86 (0.2%)
Playoffs 2,440 2,297 141 2
Play-in 37 33 4 0

A few seasons stand out:

  • 1996-97, the first season of play-by-play, is the messiest.
  • 2008-09 has many minor games because its rotation data is broken in 83 games, so lineups there come from play-by-play alone.
  • 2017-18 has three times the usual minor games because the feed’s running score is out of order in many of them. It’s cosmetic: final scores and plus-minus still match.
  • Short seasons: 1998-99 and 2011-12 were lockouts; 2019-20 and 2020-21 were shortened by the pandemic.

What we can’t fix

Some problems can’t be corrected — only measured and disclosed:

  • 88 suspect games are dropped (86 regular season, 2 playoffs). Players in those games lose those possessions from their samples. The most common reasons: a substitution names someone we can’t identify (50 games), a player is subbed out who wasn’t on the floor (44), a player’s minutes don’t reconcile with the box score (42). In about half of them (43) the rotation data is unusable too, so there’s nothing independent to rebuild the lineups from.
  • Possessions are inferred, not recorded. The feed has no possession column, so we rebuild possessions from the sequence of events. Our counts land within about 1.5% of the standard box-score possession estimate in a typical game. When a lineup changes mid-possession (say, a substitution between free throws), the possession is split between the lineups by their share of its events — a reasonable rule, but a rule.
  • 2003-04 and 2004-05 box scores list whole minutes only, so the minutes cross-check is coarser for those seasons.
  • Name-based substitution matching is checked every way we can think of, but a substitution that resolves to a plausible-but-wrong player and still passes every check would go unnoticed.
  • Scorekeeper errors that are internally consistent are invisible. If an assist or a block was recorded for the wrong player, nothing in the data contradicts it.
  • Some data simply doesn’t exist for older seasons. Player tracking starts in 2013-14, hustle stats partway through 2015-16 (and are missing for 126 games that season), and defensive matchup data in 2017-18. Anything built on those can’t go back further.

Part 2: Building RAPM

The problem with plus-minus

Plus-minus is the most direct measure of impact there is: when you were on the floor, did your team outscore the other team? But it’s badly confounded. A bench player on a great team posts a great plus-minus; a star carrying a terrible bench posts a mediocre one. Plus-minus measures the lineup, not the player.

We can’t run an experiment where we swap one player and hold everyone else fixed. But across a season, the NBA runs thousands of small, messy versions of that experiment for us: every substitution changes one player and keeps the rest. Adjusted plus-minus uses all of them at once.

The model

Every stint becomes one row of a regression:

100 (points scored−points allowed)possessions=μ+h⋅HCA+∑offenseoj−∑defensedk+ε\frac{100\,(\text{points scored} - \text{points allowed})}{\text{possessions}} = \mu + h \cdot \text{HCA} + \sum_{\text{offense}} o_j - \sum_{\text{defense}} d_k + \varepsilon
  • ojo_j is player jj‘s offensive impact, dkd_k is player kk‘s defensive impact, both in points per 100 possessions relative to an average player (positive is good on both ends).
  • μ\mu is the league-average offense and HCA is home-court advantage (h=+12h = +\tfrac12 at home, −12-\tfrac12 on the road).
  • Each stint is weighted by its possessions, so a 12-possession stint counts 12 times as much as a 1-possession one.

Every player gets two numbers, and their sum is RAPM — the points per 100 possessions a team gains with that player on the floor instead of an average player.

Games at a neutral site (the Mexico City and Paris games, NBA Cup semifinals in Las Vegas) get h=0h = 0: no home-court advantage for either team.

One season, several seasons, all-time

The site has RAPM over a single season, three- and five-season windows, and one all-time fit over every season since 1996-97. Longer windows mean more data per player and steadier ratings; shorter ones follow how a player changes.

Multi-season fits need one adjustment. League scoring has climbed from about 102 points per 100 possessions in 1997 to about 116 today. With a single league-average term μ\mu across 30 years, everyone who played in a high-scoring era would look like a better offensive and worse defensive player than they were — their total RAPM barely moves, but the split does. So each season gets its own μ\mu. The box-score prior gets the same treatment: a player’s box line is compared with his own season’s league average before seasons are combined, so a 1998 player’s three-point rate is judged against 1998, not against today.

All-time ratings only count seasons from 1996-97 on, so players whose careers started earlier — Stockton, Robinson, Olajuwon — are rated on the back half of their careers.

A side result: home-court advantage has roughly halved. It ran about +3.5 to +4 points per 100 possessions around 2000, has been around +1.5 to +2 in recent seasons, and bottomed out at +1.0 in 2020-21, when most games were played without fans.

Why plain regression fails, and why ridge works

Solve that regression by ordinary least squares and you get garbage. Teammates who almost always share the floor are nearly impossible to tell apart, and players with a few hundred possessions get extreme ratings that are mostly noise.

Ridge regression (Sill, 2010) fixes this by adding a penalty that pulls every rating toward zero (league average):

min⁡o, d  ∑stintsw (y−y^)2  +  λ(∑joj2+∑kdk2)\min_{o,\,d}\;\sum_{\text{stints}} w\,(y - \hat y)^2 \;+\; \lambda\Big(\sum_j o_j^2 + \sum_k d_k^2\Big)

λ has a concrete meaning: it’s measured in possessions. Ignoring teammates, a player with nn possessions keeps about nn+λ\frac{n}{n + \lambda} of what his raw on-court numbers suggest. At λ = 2,000, a player with 2,000 possessions — about a half-season for a rotation player — keeps half his signal; one with 6,000 keeps three quarters. Low-minute players get pulled hard toward average; heavy-minute players speak for themselves. That’s exactly what we want.

There’s also a Bayesian reading: ridge is what you get if you believe, before seeing any games, that player ratings follow a bell curve centered on zero. A single possession carries about 110 points per 100 of noise (120 in recent seasons), so λ = 2,000 corresponds to believing a typical player’s offense or defense sits about 2.5 points per 100 from average (that’s the bell curve’s standard deviation).

How we test everything

A metric that describes the past well isn’t necessarily right — it might just be fitting noise. So every modeling decision on this site is judged by out-of-sample prediction:

  1. Fit ratings on one set of games.
  2. For games the model has never seen, take the lineups that actually played and how long each played, and predict the final margin from the ratings alone.
  3. Compare predicted margins with actual ones (root mean squared error, RMSE, in points).

We test two ways: same season (fit on the first half of a season, predict the second half) and next season (fit on one season or a window of seasons, predict the next one). The second is the harder, more useful test.

Differences between models are often small, so every comparison comes with a 95% interval from a paired bootstrap: resample games thousands of times and see how often model A beats model B on the same games. If the interval excludes zero, the difference is real.

When the model picks its own settings by cross-validation, folds are split by game so the possessions of one game never appear on both sides.

Does it work?

Predicting each season’s game margins from the season before, pooled over 29 seasons (34,268 games):

Model RMSE Correlation Picks winner
Home court only 13.87 — 57.9%
Raw on-court plus-minus 13.23 0.26 61.5%
RAPM, λ = 2,000 13.14 0.31 63.1%
RAPM, λ = 4,000 13.08 0.30 62.7%
RAPM + box-score prior, λ = 8,000 12.96 0.32 63.7%
  • RAPM beats raw plus-minus by 0.15 points of RMSE (95% interval −0.17 to −0.13).
  • The box-score prior beats plain RAPM by another 0.12 (−0.14 to −0.10).

These look like small numbers because NBA games are noisy — the actual margins have a standard deviation of 13.7 points, and most of that is unpredictable from who played. Games have also gotten harder to predict: that spread has grown from 12.7 points (1996-2005) to 14.7 (2015 on), with more threes and higher scoring. The box-score prior is the best model in every era. One caveat for the literature’s λ: in 2005-2015, RAPM at λ = 2,000 predicted no better than raw plus-minus (12.72 vs 12.72), while λ = 4,000 did (12.62).


Part 3: Iterations — what we tried, and what we learned

1. Single-season RAPM barely beats raw plus-minus — at first

Our first test (fit on 1996-97, predict 1997-98) was humbling: single-season RAPM essentially tied raw on-court plus-minus (12.9 vs 12.8 RMSE). One season is not much data. With 29 seasons of tests pooled, RAPM’s advantage is clear — but it’s modest, and a single season’s rating should be read with that in mind.

2. More seasons help — for plain RAPM

In our first multi-year test (predicting 1999-00), RAPM fit on a three-season window beat single-season raw plus-minus by 0.21 points (95% interval −0.39 to −0.03). We also tried weighting older seasons less (each prior season at half the weight of the next); it made no measurable difference, so multi-year ratings weight seasons equally. As it turned out, the box-score prior (next) gets most of the benefit of extra seasons from a single one.

3. The box-score prior: first it hurt, then it helped

Plain RAPM shrinks every player toward average. But we know more than that: a player who scores efficiently, rebounds, and blocks shots is probably above average. The box-score prior shrinks each player toward what his box score predicts instead:

oj=zj⋅γo+ajo_j = z_j \cdot \gamma_o + a_j

where zjz_j is the player’s per-100-possession box line (2s and 3s made and attempted, free throws, rebounds, assists, turnovers, steals, blocks, fouls, minutes per game), γo\gamma_o is a set of weights learned jointly with RAPM from the same possessions, and aja_j is the part of his impact the box score doesn’t explain — which is what ridge shrinks.

The first version raised the correlation of next-season predictions but made same-season predictions worse. Two causes, found by testing:

  • The box-score weights were unpenalized and overfit. Makes, attempts and minutes move together, so on half a season of data the weights swung wildly. A light penalty on the weights (scaled to each feature’s spread) fixed it.
  • λ was set too high for half-season data.

We also tested whether shooting luck was the problem (a hot shooting half inflating a player’s box score). Regressing shooting percentages toward league average made no difference, so we dropped it.

With the fix, the prior beat plain RAPM in all six tuning splits, then on a season held out entirely (2000-01): −0.15 points within the season (−0.24 to −0.07). Across all 30 seasons it holds up (see the tables in section 5).

4. Lineup synergy: not there (yet)

Do some five-man units play better than the sum of their parts? We fit lineup ratings as the sum of the five players plus a synergy term for each unique lineup, with its own penalty. Setting that penalty to infinity reduces the model to plain player sums, which makes this a clean test.

Within a season, the best predictions came with synergy turned completely off. Next season, the gain was 0.02 points with an interval crossing zero. Most possessions next season are played by lineups that didn’t exist last season, and the ones that did had too few possessions to say anything. Lineup ratings on this site, including the ones the lineup optimizer uses, are the sum of the players’ ratings. Structured synergy (pairs of players, or lineup types like “four shooters”) is an open question.

5. Tuning λ: why we didn’t just use 2,000

The standard value in the RAPM literature is λ = 2,000, from Sill’s original work, and it’s what my DS785 capstone used. Our setup is on the same scale, so we could have adopted it directly. We decided to test it instead, scoring 14 values of λ from 300 to 40,000 three ways, in every one of the 30 seasons:

  • Cross-validation within the season — what explains this season’s possessions?
  • First half → second half — what predicts the rest of the season?
  • Next season — what predicts next year, from one season or from three?

The best λ by era:

How λ is judged 1996–2005 2005–2015 2015–now
Within-season cross-validation, plain 2,000 2,000 3,000
Within-season cross-validation, box prior 4,000 4,000 6,000
First half → second half, plain / box prior 2,000 / 3,000 2,000 / 3,000 2,000 / 3,000
Next season from 1 season, plain / box prior 4,000 / 8,000 4,000 / 8,000 3,000 / 6,000
Next season from 3 seasons, plain / box prior 6,000 / 16,000 6,000 / 16,000 4,000 / 8,000

We reproduce 2,000 — within-season cross-validation picks it for plain RAPM through the mid-2010s, as does every first-half → second-half test. The literature’s number answers the question it was asking.

But the right λ depends on the question:

  • Describing a season (“what happened?”) wants about 2,000–3,000.
  • Forecasting has wanted more shrinkage for most of this period. Players change from year to year — they age, change roles, change teammates — so this season’s evidence is a weaker guide to next season than to itself, and should be trusted less. Three-season windows want more still, because the oldest season is stale.
  • The box-score prior wants more shrinkage, because shrinking toward a box-score estimate costs much less than shrinking toward zero. A λ tuned for plain RAPM doesn’t carry over.

To make sure we weren’t tuning on the same games we score, we picked λ for each season using only the other seasons, then scored it on the held-out one:

Test Box prior vs plain RAPM Box prior vs plain RAPM at λ = 2,000
Same season −0.10 (−0.11 to −0.08) −0.10 (−0.11 to −0.08)
Next season, 1-year window −0.12 (−0.14 to −0.10) −0.19 (−0.21 to −0.16)
Next season, 3-year window −0.03 (−0.05 to −0.00) −0.12 (−0.15 to −0.10)

With three seasons of data, the prior’s edge over plain RAPM mostly disappears after 2005 — there’s enough plus-minus data that the box score adds little. And since 2005, a one-season box-prior model forecasts as well as any three-season model.

One subtlety: λ = 2,000 actually has slightly higher correlation with next season’s results than λ = 4,000 (0.31 vs 0.30) — it ranks teams about as well — but its predictions are too spread out, so its errors are bigger. Part of what extra shrinkage buys in forecasting is calibration.

Has it shifted over time? Yes — in opposite directions

  • Describing a season wants more shrinkage than it used to. The within-season best λ has crept up steadily (rank correlation with year +0.67), and plain RAPM flips from 2,000 to 3,000 around 2015. Part of the reason is noisier possessions: three-pointers went from 18% of shots to 37%, and the noise on a single possession rose from about 111 to 120 points per 100. Since λ behaves like noise² divided by the spread of player talent, that accounts for about a third of the increase. We don’t have a confirmed explanation for the rest. One candidate: the league now uses about 580 players a season instead of 440, so the pool has many more fringe players whose true impact is close to average.
  • Forecasting wants less shrinkage than it used to. In the modern era the forecasting λ (3,000 for one season) matches the within-season one. Before, it was twice as large. Put simply, a modern season’s ratings predict next season about as well as they predict the rest of the same season. One reading is that player impact carries over more steadily from year to year now — but that’s a reading of the numbers, not something we’ve confirmed.
  • How much it matters: for describing a season, barely — using 2,000 instead of 3,000 in the modern era costs almost nothing. For forecasting, a fixed 2,000 costs 0.04–0.14 points of game RMSE depending on the era.

What we use:

  • Leaderboards (describing what happened): one fixed λ for every season and every window — 2,500 for pure RAPM and 6,300 with the box-score prior. See below for why.
  • Predictions (game predictor, lineup optimizer, player comps): the box-score prior, with λ picked from recent seasons rather than a fixed constant. For the modern era that’s about 6,000 for a one-season model and 8,000 for three seasons.

Why the leaderboards use one λ for every season

Our first leaderboards let cross-validation pick λ separately for each season. That fits each season slightly better, but it quietly breaks comparisons between seasons: λ sets how far every rating is pulled toward zero, so a season that happened to get a lower λ has a wider spread of ratings, and its best players look better than an equally good player in a season that got a higher one.

The effect wasn’t small. For single-season pure RAPM, cross-validation picked 1,585 in six seasons, 2,512 in 21, and 3,981 in three. Among players with at least 1,000 possessions:

λ cross-validation picked Seasons Spread of ratings (SD) Same seasons at λ = 2,500
1,585 6 2.50 2.01
2,512 21 2.00 2.00
3,981 3 1.56 1.98

The 2002-03 leader rated +10.0 at that season’s λ and +7.6 at the common one; some individual players moved by as much as 2.5 points purely because of the λ choice. With one λ, the spread of ratings varies about a third as much from season to season.

So we fixed it. Averaging the cross-validation error over all 30 seasons, the best λ for pure RAPM is 2,500 for one-season, three-season, five-season and all-time fits alike, and 6,300 with the box-score prior (all-time very slightly prefers 10,000, but its curve is nearly flat, and one value keeps every window on the same scale). The cost is tiny — ignoring the modern drift toward 3,000 costs about 0.004 in fit — and in exchange a +5 means the same thing in 1998 as in 2024.

6. Still overconfident

Even the best model is overconfident about next season: regress actual margins on predicted margins and the slope is about 0.85 (0.81 for plain RAPM, 0.86 with the box prior; 0.87 and 0.91 in the modern era). A perfectly calibrated forecast would have a slope of 1. Predictions should be shrunk 10–20% more before they’re used to forecast games. The game predictor will estimate that factor from past seasons.


Open questions

  • Why has the within-season λ drifted up by more than the extra noise explains?
  • Why has home-court advantage halved?
  • Can structured synergy (player pairs, lineup archetypes) find real lineup effects where unique five-man units couldn’t?
  • Playoff RAPM: much less data, very different rotations. How much should it borrow from the regular season?
  • Combining RAPM with a box-score metric after the fact, as in my capstone’s C-RAPM, vs. fitting them jointly, as we do now.

References

  • Sill, J. (2010). Improved NBA adjusted +/− using regularization and out-of-sample testing. MIT Sloan Sports Analytics Conference.
  • Engelmann, J. (2011). A new player evaluation technique for players of the National Basketball Association (NBA). MIT Sloan Sports Analytics Conference.
  • Myers, D. (2020). Box Plus/Minus (BPM).