Skip to content

opportunity-adjusted-impact-v3 · 2026-27

Full impact is live.

The regular-season leaderboard and player profiles use a validated estimate of player strength in points per 48 minutes. Box scores, play-by-play, tracking, hustle and defended shots enter one regularized model, with teammate and opponent lineup context. Short samples borrow from prior seasons and the league. No position receives an automatic bonus.

The number in each game log is Game impact, a separate measure of one game in points of margin, described below. It is not a slice of Full impact and the two are never added.

Read the historical validation, source coverage and remaining limits ↗

Research / Publication contract · Revised November 1, 2026

What goes into
the number.

There are two numbers. Game impact credits the points of one game’s final margin to the players in it. Full impact estimates how strong a player is over a season, per 48 minutes. They answer different questions and are never added together.

Start with the label.

Game impact · one game

impact-ledger-v1 is in points of scoreboard margin. It uses the final box score, the play-by-play and the lineups on the floor, and a team’s players add up to its final margin. How it is calculated ↓

Unavailable

A game that cannot be scored has no number and the page says why. That covers players who did not play, games whose lineups cannot be confirmed, and competitions the model is not calculated for. Missing evidence is never a zero, and an older model’s number is never shown in its place.

Until November 2026 the per-game number was a box-score formula with fixed weights, shown as box contribution. It rated big men about three and a half points per 36 minutes above guards for reasons of role, and it rewarded shot volume. Game impact replaced it on every page. Games it cannot score show no number.

The full-impact calculation.

Each game is split into two-minute clock windows. Exact substitution intervals determine each player’s share of those minutes. Points and exposure are conserved, including overtime and free throws scored without elapsed time. This uses minutes, not a possession estimate.

Player strength = jointly fitted feature estimate + regularized player-season residual

Expected window margin = home advantage × minutes + sum(home player strength × minutes) − sum(away player strength × minutes)

Displayed full impact = 48 × player strength, centered against the season’s minutes-weighted league average

Features pool toward their league mean with a 120-minute prior before standardization. Missing observations contribute no invented counts. The fit minimizes duration-weighted scoring error, with ridge penalties of 7,680 for player residuals and 10,000 for feature coefficients. Earlier-season windows receive half weight; residual continuity across seasons adds a penalty half as strong as the player penalty. These settings were chosen on the tuning games, before the final test.

Box, play-by-play and ten tracking families are fitted together, so correlated assists, passing and scoring counts are not simply added as separate bonuses. The shared coefficients and residuals are statistical estimates; their decomposition does not identify independent causal skill values. Direct gravity and per-game defensive matchup feeds stay contextual until their own history and attribution checks pass.

Player uncertainty comes from 100 game-cluster bootstrap fits with the feature snapshots held fixed. It measures conditional sampling variation, not every source error or uncertainty about a player’s true ability. The public artifact includes the inputs’ coverage, coefficients, model settings and validation results.

Game impact.

Game impact answers one question about one game: how many points of the final margin belong to each player? It is measured in points. A team’s players always add up to that team’s final margin, so in a 110–100 game the winners total +10 and the losers −10.

Each player’s number comes from two places.

Team result = one fifth of the team’s scoring margin in every stretch the player was on the floor

Counted plays = for each kind of play: share kept × (points the player’s plays were worth − what a typical player at the same position produces in those minutes − what the five on the floor averaged above typical)

Game impact = Team result + Counted plays

The team result is the only place where spacing, screens and defense that leaves no statistic can register. The counted plays then move credit inside each lineup toward the players who made the shots, passes, rebounds and stops. The counted plays of the five on the floor cancel exactly, so they never change a team’s total. Every player card shows each line of this sum.

What each play is worth

Values are measured from the season before the one being scored, so nothing is fitted on the games it describes. For this season that is 2025-26 (1,204 games), when a possession was worth 1.139 points and the offense rebounded 26% of its misses.

PartPlayPoints eachShare kept
ScoringShots at the rim+0.28100%
ScoringMidrange shots−0.11100%
ScoringThree-point shots+0.13100%
ScoringMakes at the rim, above the league rate+1.70100%
ScoringMidrange makes, above the league rate+1.70100%
ScoringThree-point makes, above the league rate+2.70100%
ScoringFree throws made+0.5243%
ScoringFree throws missed−0.4843%
ScoringBaskets a teammate assisted−0.22100%
PlaymakingAssists+0.22100%
PlaymakingLive-ball turnovers−1.3147%
PlaymakingDead-ball turnovers−1.1147%
ReboundingOffensive rebounds+0.8476%
ReboundingDefensive rebounds+0.3043%
Defensive playsSteals+1.3192%
Defensive playsBlocks+0.9071%
Defensive playsShooting fouls−0.3862%
Defensive playsOther defensive fouls−0.0762%

A shot is scored twice. Shot selection is what an attempt from that spot is worth at the league’s accuracy. Shot making is the makes above that accuracy, each worth the basket less the offensive rebounds a miss would have kept. An assisted basket moves 0.22 points from the scorer to the passer, which is how much more a shot a pass creates is worth than one a player creates alone. A turnover costs the possession; a live-ball turnover costs a little more because the other team scores more easily off a steal.

The share kept is how much of a play a team was measured to gain when one player does more of it. A defensive rebound keeps 43% because most of them would have been collected by a teammate anyway. A steal keeps 92% because almost every steal is a possession the team would not otherwise have had.

Why each player is compared with their own position

Centers collect more rebounds and blocks than guards, and guards more assists, because of where they stand. Comparing every player with a typical guard, wing, forward, big forward or center in the same minutes keeps a role from being scored as a skill. Over a season each position group averages close to zero, which matches what lineup-based measures and ESPN’s Net Points show.

How it was tested

Game impact adds up to each game’s margin by construction, so agreeing with the games it describes proves nothing. The test uses games it has not seen. Each player is rated only on games before the one being predicted. The ratings of the ten players on the floor are then compared with the margin of that game. The table shows the correlation between the predicted and the actual margin. Zero means no information and higher is better.

SeasonGamesGame impactTeam plus-minus aloneEarlier box formula
2024-25
Design season. Credit shares were measured here.
9480.4480.4500.403
2025-26
Test season. Nothing was fitted on it.
1,2040.4750.4500.473
2026-27
Games played after the design was fixed.
1540.4870.4700.404

On the test season Game impact beat plus-minus alone by +0.012 to +0.040 (95% range from resampling games). Against the earlier box formula the difference was −0.023 to +0.027, which is a tie. Game impact replaced that formula because the formula rated big men well above guards and rewarded shot volume, and Game impact does neither while predicting at least as well.

Computed 2026-11-11 · current-season games through 2026-11-10 · impact-ledger-v1 · Download the figures ↗

What it does not see

A defender who takes away a shot without a steal, block or foul is credited only through the team result, shared equally with the four teammates on the floor. The same is true of a screen or of the space a shooter creates. We tested whether the NBA’s per-game tracking counts could tell those five apart: contested shots, deflections, shots defended at the rim, box-outs and matchup results, over two full seasons. None of them improved the prediction of games the model had not seen, so none is in the formula. One game is also a small sample: a team’s margin in a player’s minutes depends on teammates, opponents and luck. Game impact describes what happened in a game. It is not a rating of how good a player is. That is what Full impact estimates.

When a game has no number

Game impact needs the checked play-by-play and the exact lineups on the floor, which exist only after the final. A game appears without a number until that record is processed. A few games never get one because the official substitution record does not reconcile with the box score. Preseason games are not scored. In each case the page says so.

Publication reconciles player points with team scores, team minutes with game length, made shots with attempts, the play-by-play counts with the box score, and every lineup stretch with the final margin. Source checksums, retrieval times, calculation times and model versions are retained.

The live estimate during a game.

While a game is in progress the page shows a live box estimate. It is a box-score calculation with fixed weights, measured against the game’s own average production per minute, and its baseline moves as the game develops. It is not Game impact, which needs lineups that are confirmed only after the final. When the game ends the estimate is replaced, and the two numbers will differ.

How hard a defender’s assignments are.

Player profiles show who a defender is asked to guard. The NBA’s matchup tracking records which opponent each defender was assigned to and for how many partial possessions. We take the scoring rate of each of those opponents, in points per 36 minutes, and average it by the possessions spent on each one. A scorer’s rate uses this season and half of last season, pulled toward the league rate when he has played little.

That number alone would flatter anyone who plays against starters. So it is set next to the scoring rate of every opponent who was on the floor with him, weighted by the minutes they shared. The difference is the part that belongs to him: among the five defenders on the floor, he took the bigger scorers. A rank needs 150 tracked partial possessions.

This describes the job, not how well it was done, and it is not part of Game impact or Full impact. We tested it on two full seasons, with values set on 2024-25 and scored on 2025-26. Crediting the assignment did not make predictions of lineups or games better, and a larger credit made the ratings agree less with independent measures. Tracking assignments are also the league’s estimates and miss help defense and switches. See the league ranking.

Why the previous blend was retired.

The previous blend is retired from publication. It rescaled a season RAPM estimate to box-score averages and added the same player term to every appearance, regardless of game minutes. It also removed team averages without establishing that this improved attribution. More data does not fix incompatible units.

The model separates three concepts: total recorded contribution, an exposure-weighted rate, and a stabilized estimate of ability with uncertainty. A minute-adjusted total is not a per-minute rate. Season rates must divide summed contribution by summed exposure; they must not average individual game rates.

QuestionEvidence to testWhat cannot be assumed
Spacing and creationOn/off-ball gravity, potential assists, secondary assists, drives and touchesMade threes alone measure spacing, or a provider score already has point units
Perimeter and help defenseAssignments, contest samples, deflections, opponent and lineup contextLow opponent FG% proves defense, or pressure faced by a shooter measures his own defense
Rebounding and screeningContested chances, deferred rebounds, screen assists and teammate opportunitiesEvery rebound has equal marginal value, or a screen assist is extra credit on top of all the same scoring credit
Low-minute playersVerified exposure, uncertainty, role and opponent qualityA short hot stretch establishes stable ability, or missing tracking is zero performance

Any lineup model needs verified five-player units, clocks, reconciled scores and player minutes. Completed possessions must group free throws and offensive-rebound continuations correctly. The existing scoring-event cache does not meet that definition.

Publication requires a new version, common units, source timestamps, a held-out chronological evaluation and feature-ablation results. We must compare against the earlier box formula and lineup-only baselines, check errors by role and exposure, and test missing-source behavior. Correlated signals need joint calibration. Position averages will not be forced to match: equal means would not prove fair attribution.

Separate views of the game.

NBA tracking, matchup statistics, gravity, leverage, shot quality and play-by-play features have their own coverage and units. Season aggregates are not individual-game observations. Tracking updates cannot silently change the impact formula.

A low opponent shooting percentage over a small matchup sample is not proof of defensive ability; assignments, help and shot selection all matter. Expected shooting percentages are source estimates. Public statistics alone do not establish player intent, a coach's plan or a film observation.

ESPN Analytics adds a separately attributed Net Points view on player pages. Its offensive and defensive totals and overlapping skill slices use ESPN's own possession denominator. These are descriptive model estimates, not an independent ground truth or additional points to add to RTG impact. Current snapshots are joined by NBA player ID and checked against official regular-season games and minutes.

What validation establishes.

Game impact adds up to each game’s margin by construction, so agreement with the same game proves nothing. Its test is on games it has not seen: rate each player only on earlier games, then check how well the players on the floor predict the margin of later games and lineup stretches. Split-half stability is reported separately, and early-season samples may be too small to estimate it.

Forecast evaluation uses probabilities saved before tipoff, with source cutoffs no later than archive times. The original preseason forecast stays unchanged; later results create new snapshots. Checks show their actual computation date. Passing software tests does not prove predictive quality.

Official historical box statistics can pass verification while historical impact remains held. Corrected inputs must be reconciled before dependent claims return to publication. Prior versions are preserved for audit and rollback.

The sources.

Implementation: Game impact in models/impact_ledger_v1.py, its publication in webapp/impact_ledger_publisher.py, final-source checks in webapp/core_collection.py, and the live estimate in webapp/live_studio.py. A dataset must identify the model and sources it actually used.

The morning read

Last night's NBA games in your inbox at 8 a.m.

Every game in one email: quarter scores, how long each team led and the night's top players by Game impact. Pick a team and its game comes first.

Free. One email on mornings after games. Unsubscribe in one click.

Join 672 people who have already signed up.

The top of the Nov. 10 morning read: Miami beat Utah 115 to 109, with scores by quarter and a chart showing Miami led for 43 of 48 minutes.
The morning read for Nov. 10