Jason Caffey

NBA impact metrics, explained

Impact metrics all try to answer one question: how much better is a team when a given player is on the floor?

The box score records what a player did — points, rebounds, assists — but not what those actions were actually worth, and it misses whole categories of value entirely: screening, spacing, defensive communication, shot deterrence, making the right pass one beat earlier. Impact metrics start from the scoreboard instead. They track how the score moves while a player is on the court, then use regression to adjust for the quality of every teammate and opponent sharing the floor with them.

Almost every metric on this page is expressed the same way: points per 100 possessions, relative to an average NBA player. A +4 means the team plays about 4 points per 100 possessions better than it would with an average player in that spot — an All-Star-level number. 0 is average, and most rotation players land between -2 and +2. Offense and defense are usually reported separately and summed for the total.

One reading convention for this site: green is always good for the player's team, red is always bad — on both ends. A positive defensive number means the player helps his defense, even when the underlying stat (like opponent shooting) is something the defense wants to push down.

RAPM regularized adjusted plus-minus

The foundation of every impact metric on this site. RAPM estimates how many points per 100 possessions a player adds, using nothing but the scoreboard and who was on the floor — no box score at all. If a player makes his team better, RAPM sees it, whether or not it shows up in his stats.

the idea, step by step

• Plus-minus is just the score margin while a player is on the court. Simple, but hopelessly polluted: play next to a superstar and your plus-minus looks great no matter what you do.
• Adjusted plus-minus (APM) fixes that with regression. Every stretch of game time becomes a data point — these five offensive players against these five defenders produced this many points — and the regression solves for the individual contribution of each player, controlling for everyone else on the floor.
• Regularized APM (RAPM) adds ridge regression, which pulls every estimate toward zero (league average) unless the data provides real evidence otherwise. This is the fix for teammates who share the floor constantly and would otherwise be statistically impossible to tell apart — and it stops small samples from producing absurd numbers.

why it matters

• Everything counts. Screens that free a shooter, rotations that kill a possession, passing up a good shot for a great one — none of it appears in the box score, and all of it moves RAPM.
• It can't be gamed by stat-hunting. Empty scoring on bad shots, chased steals, and stolen rebounds don't help a team's scoring margin, so they don't help RAPM.
• It's honest about role players. Low-usage players who make lineups work — elite spacers, connectors, defensive communicators — routinely grade far better in RAPM than in box-score metrics.
• It improves with sample. Over multiple seasons of possessions, RAPM converges on a genuinely reliable estimate of a player's two-way impact. Long-sample RAPM is the standard other impact metrics are judged against.

limitations

• Noisy in small samples — single-season RAPM for a low-minute player should be taken lightly.
• By itself it says nothing about how a player creates impact — that's what our Six-Factor decomposition is for.
• It measures what happened with real lineups, not a lab-isolated skill. A player's number always carries some context from how he was used.

variations on this site

Time Decay RAPM: Uses a player's full history but exponentially down-weights older possessions, so the estimate reflects the current player while still borrowing stability from the past. This is the best single answer to "how good is he right now?"

3-Year RAPM: Restricted to the last three seasons, with recent possessions weighted more heavily. A middle ground between single-season noise and career-long inertia.

Lifetime RAPM: A 28-year calculation with an age curve built in, so a player is compared against expectations for his age at every point in his career. Built for career-arc and historical comparisons rather than current-season evaluation.

Six-Factor RAPM

Our signature metric. Regular RAPM tells you how much a player helps; Six-Factor RAPM tells you where the impact comes from. It splits total impact into the three ways a possession can be won or lost — shooting efficiency, turnovers, and rebounds — on both offense and defense.

Every possession ends one of three ways: the offense turns the ball over, it gets a shot off and scores (or doesn't), and if the shot misses, someone controls the rebound. Six-Factor RAPM measures how a player bends each of those stages in his team's favor:

the six factors

• oTS — Offensive scoring efficiency. How much the player raises the points his team generates per scoring attempt while he's on the floor. Driven by his own scoring and foul-drawing, but also by spacing, gravity, playmaking, and transition play that improve his teammates' shots.
• oTOV — Offensive turnover economy. How the player affects his team's rate of losing the ball before a shot goes up. High-load creators carry the most turnover risk, so protecting the ball under real responsibility scores well here.
• oREB — Offensive rebounds. How much the player turns his team's misses into extra value: securing offensive rebounds, keeping balls alive, and converting the points that follow.
• dTS — Defensive efficiency suppression. How much the player lowers opponents' points per scoring attempt — contests, rim protection and deterrence, screen navigation, and foul discipline.
• dTOV — Defensive turnover creation. How much the player forces opponents into losing the ball: steals, deflections, ball pressure, and offensive fouls drawn.
• dREB — Defensive rebounds. How much the player ensures opponent misses actually end the possession, mainly through defensive rebounding and boxing out.

The six pieces sum back to the player's total: RAPM = oTS + oTOV + oREB + dTS + dTOV + dREB. This is a zero-residual decomposition — the factors account for total RAPM exactly (R² = 1), so every point of a player's impact on the scoreboard flows through one of these six channels.

how it's built

Same engine as RAPM, run three times. Instead of one regression predicting the score, three separate ridge regressions on the same lineup data predict shooting efficiency, turnover rate, and rebounding rate — each producing an offensive and defensive estimate per player. A second stage converts each factor into points per 100 possessions, which is what the tables display.

how to read it

• These are team-level effects, not personal stats. A player's oTS is not his own true-shooting percentage — it's how much more efficiently his team scores with him out there. Great passers and elite spacers post strong oTS without high scoring volume.
• Positive is always good, on both ends. A positive dTS means opponents score less efficiently against his lineups.
• Profiles matter more than totals. Two players can both be +3 with completely different shapes — one an efficiency engine, another a possession hoarder who wins the turnover and rebounding battles.

limitations

• Inherits RAPM's noise in small samples — trust multi-year numbers most.
• It shows where impact appears, not the exact mechanism. A big dTS could be rim protection, scheme, or communication; the factor surfaces the pattern, film explains it.

The 6Factor tab breaks these down further — shot-location value (rim, mid-range, three), free throws, and turnover types — and the Decomp tab shows the full offense/defense mirror.

EPM estimated plus-minus

EPM, created by Taylor Snarr at Dunks & Threes, is a hybrid impact metric: RAPM stabilized with a statistical prior built from box-score and player-tracking data. It trades some of pure RAPM's neutrality for much more stability and predictive power.

The core problem with pure RAPM is noise: it needs years of possessions to settle. EPM's answer is to give the regression a smart starting point. Instead of assuming every player is average until the lineup data says otherwise, EPM assumes each player is what his stats suggest he is — and then lets the lineup data push that estimate up or down.

how it works

1. A statistical plus-minus (SPM) model estimates each player's impact from box-score and tracking stats — "estimated skills" tuned specifically to predict impact rather than to describe production.
2. Those estimates become the Bayesian prior in a RAPM calculation, which captures whatever impact the stats miss.

strengths

• Far more stable than pure RAPM in a single season — usable within weeks, not years
• Among the most predictive public impact metrics, using data from players' entire careers and updating nightly
• Separate offensive (OEPM) and defensive (DEPM) components

limitations

• The prior is only as good as the stats behind it — value the box score can't see (screening, spacing, communication) can be pulled toward what the stats expect
• That same bias tends to favor high-usage players relative to pure RAPM
• The model is proprietary, so its inner workings can't be independently verified

DARKO daily adjusted and regressed kalman optimized

DARKO, created by Kostya Medvedovsky, is a projection system rather than a rating: its headline number, DPM (Daily Plus-Minus), is an estimate of how good a player is right now, updated after every game. It's consistently among the most predictive public impact metrics.

how it works

DARKO is built around a Kalman filter — the same technique used to track moving objects from noisy sensor readings. Each game is a new noisy reading of a player's true skill. The filter blends every new reading with everything it already believes about the player, weighting by how reliable each source is. A rookie's estimate moves fast because the model knows little; a ten-year veteran's estimate barely budges on one hot week.

• Every box-score stat is projected separately — shooting, rebounding, assists, turnovers, and more, each with its own decay rate (three-point shooting takes far longer to trust than rebounding, and the model knows it).
• Those projected stats feed a box estimate (Box DPM), which is then combined with on/off and RAPM-style lineup data to produce full DPM, with offensive (O-DPM) and defensive (D-DPM) components.
• Age curves are built in, so projections account for typical development and decline.

strengths

• DARKO's projected stats beat season averages badly as predictors of what a player does next — a player's season averages are a worse guess about tomorrow than DARKO's estimate
• Reacts quickly and sensibly to role changes, breakouts, and decline
• Daily updates make it uniquely useful in-season

limitations — our take

• The box component seems to undervalue playmaking — Box O-DPM rated Jaylen Brown and Nikola Jokić as similar offensive players in 2023, which no full-impact measure supports.
• Like most box-prior metrics, it can flatter high-usage scorers relative to pure lineup data.
• It projects skill; it isn't a record of what actually happened in a given season. For "who impacted winning last year," RAPM-family metrics are the better tool.

LEBRON luck-adjusted estimate using a box prior regularized on-off

LEBRON is Basketball Index's flagship impact metric. Like EPM, it's RAPM with a box-score prior — its two distinguishing ideas are a "luck adjustment" that strips variance out of the on/off data and a prior tailored to each player's offensive role.

the two big ideas

• Luck adjustment. Raw on/off data is contaminated by things no player controls — teammates running hot or cold from three or the free-throw line during his minutes. LEBRON replaces some of this observed shooting with expected shooting before running the regression, so a player isn't credited or punished for teammate shooting variance.
• Role-aware box prior. The box-score starting point (built on PIPM weightings) is stabilized differently for each offensive archetype. The same box-score line means different things from a post hub than from a spot-up wing, and padding by role keeps thin samples from being read the same way for both.

how it works

1. Build a box-score prior for each player from PIPM weightings
2. Stabilize that prior using the player's offensive archetype
3. Run luck-adjusted RAPM on the on/off data with that prior
4. Output O-LEBRON, D-LEBRON, and total LEBRON in points per 100 possessions above average, plus Wins Added for season-level totals

strengths

• The luck adjustment removes a genuine source of noise that most impact metrics carry
• Role-aware priors handle unusual player types better than one-size-fits-all box models
• Clean offensive/defensive split and an intuitive wins conversion

limitations — our take

• The luck adjustment is a modeling choice, not a free lunch — deciding what counts as "luck" (e.g., how much of opponent three-point shooting a defense controls) shapes the results.
• Misclassify a player's archetype and his prior is padded with the wrong comparisons.
• The box prior appears to skip advanced play-by-play stats like offensive fouls drawn, which shortchanges elite guards who generate extra possessions that never hit the box score.

Contesting Metrics

These measure the two halves of shot defense: how often a player challenges shots, and what happens to those shots when he does. They also track the team-level question that matters most for rim protectors — does the rim get attacked less, and less successfully, with him out there?

individual contest metrics

• Total Contests per 100: How many shots the player contests per 100 defensive possessions — a volume measure of defensive activity.
• Contest Dif%: How much worse shooters do against this player's contests than they normally shoot. A -5% means shooters lose five percentage points off their usual accuracy when he's the one contesting.
• Rim Contests per 100 and Rim Contest Dif%: The same two ideas restricted to shots at the basket, where contests are most valuable.

team rim-protection metrics

• rRim Acc On: Opponent shooting accuracy at the rim with the player on the floor, relative to that season's league average.
• Rim Acc On/Off: How opponent rim accuracy changes when the player steps on the court versus sits.
• rRim Freq On and Rim Freq On/Off: The same pair for how often opponents get to the rim at all. This is where deterrence shows up: elite rim protectors make teams stop trying rim attempts, which never appears in blocks or contest stats.

interpretation

For the Dif%, accuracy, and frequency metrics, negative is good — opponents shooting worse or attacking the rim less. For the contest-volume metrics (per 100), higher means more defensive activity.

keep in mind

• Contest numbers reflect role and scheme as much as skill — a drop-coverage center racks up rim contests by assignment.
• On/off numbers carry the usual noise: backup quality and lineup context leak in.
• None of this captures communication or early help positioning — the stuff that prevents a shot from ever being attempted. That value shows up in dTS, not here.

Passing Metrics

Assists undercount playmaking: they require a teammate to make the shot, and they say nothing about how much time on the ball the passer needed. These metrics separate the three things that actually matter — how much a player creates, how efficiently he creates it, and what his passes cost in turnovers.

key metrics

• On-Ball Percent: The share of team time the player spends holding or dribbling the ball — the "cost" side of playmaking, since on-ball time is a finite team resource.
• Potential Assists per 100: Passes that lead directly to a shot attempt, per 100 possessions — an assist that doesn't punish the passer when the shooter misses. This is the cleanest measure of creation volume.
• Pot Ast per 1 min On-Ball: Creation per minute of ball possession. This is the efficiency ratio — it separates quick decision-makers who create instantly from players who need long stretches of dribbling to generate the same look.
• Assists per 100: Traditional assists, volume-adjusted.
• BadPass%: Of the player's creation attempts, how many end as bad-pass turnovers instead. The risk side of the ledger — lower is better.
• TOV Adj. Assist EFG%: How well teammates shoot off this player's passes, discounted for his bad-pass turnovers. High values mean his passes arrive as genuinely good shots — on time, on target, in rhythm.

how to read them together

Volume (potential assists), efficiency (per minute on-ball), quality (assist eFG%), and risk (BadPass%) are four different skills. The best playmakers score on all four; a high-volume creator with a high BadPass% and low assist eFG% is producing quantity, not quality.

keep in mind

• Passing to great shooters inflates assist eFG% — some of the credit belongs to the finisher.
• These don't measure pass difficulty or defensive pressure — a hockey-assist machine in a motion offense and an isolation creator face different degrees of difficulty.
• Scheme matters: low on-ball time can reflect role, not inability.

Defensive Playmaking Metrics

These measure the plays that end opponent possessions outright — steals, charges, and blocks the defense actually recovers. The theme throughout: a defensive play only counts fully if your team ends up with the ball.

key metrics

• STOP%: The headline number — the share of defensive possessions where the player personally ends the possession via a steal, a drawn offensive foul, or a block his team recovers. It's a strict standard: only clean, possession-ending plays count.
• Deflections per 100: Times the player gets a hand on the ball per 100 possessions. The best leading indicator of active hands and positioning — deflections predict future steals better than steals do.
• STL per 100 and Off Foul Drawn per 100: The two clean-turnover components of STOP%, shown separately. Drawn charges are the most underrated play in basketball: a guaranteed turnover plus a foul on the opponent.
• rFTOV: The player's forced-turnover rate relative to league average.
• BLKs per 100, BLKs Recov per 100, and BLK Recov %: Not all blocks are equal — a swat into the third row gives the ball right back. These separate total blocks from blocks the defense recovers, and the recovery rate tells you who blocks to end possessions versus who blocks for highlights.

interpretation

Higher is better across the board. STOP% is the best single summary; the components tell you how a player generates his stops.

keep in mind

• This is one slice of defense — a player can be an elite defender with modest playmaking numbers (contests, positioning, and communication live in the contesting metrics and dTS instead).
• Big steal and deflection numbers can come from gambling that also gives up easy baskets. Cross-check against dTS and dTOV to see whether the aggression nets out positive.
• Scheme inflates or suppresses these — aggressive, trapping defenses generate more turnover chances for everyone in them.