Basketball stats glossary

A player who has played enough, and shot enough, for a rate stat to mean something. Without a floor, a rookie who takes one 3-pointer and makes it tops the 3P% leaderboard at 100% — and, worse, shifts everybody else's percentile down a notch.

Two conditions, both required. Presence: at least half the games her team has played, never fewer than 5 — so the rule holds in June as well as in September, and in every league we cover. Volume: a per-game floor that depends on the stat. Nothing extra for counting stats like points or rebounds; 12 minutes for anything built on possessions (ratings, USG%, AST%); 5 field-goal attempts for FG% and TS%, 1.5 three-point attempts for 3P%, 1.5 free throws for FT%.

Those shooting floors are the official WNBA qualification minimums converted to a per-game basis — 100 made field goals over a season is almost exactly 5 attempts a game.

An unqualified player is never hidden. Her row stays in the table and her numbers are shown; what she loses is the ranking — no percentile badge, no place in the chart, and she sorts to the bottom of the column. The Qualified button in the chart toolbar turns the whole thing off if you want to see everyone.

Off is what the subject produced. Def and Opp are two names for the same thing : what the opponent produced against them. OffTS% is your true shooting ; DefTS% and oTS% are the true shooting you allowed.

Net is the signed gap, Off minus Def. Positive is always better, which means two of them are deliberately flipped : NetTOV% is Def minus Off, because forcing turnovers is good and committing them isn't, and NetRB% compares your offensive rebounding to the offensive rebounding you conceded.

One asymmetry is worth knowing on player pages. A player's Off stats are her own box score. Her Def stats are what the opposing team did while she was on the floor — a five-player result, not an individual one. So a player's Net figure mixes an individual number with a lineup number. It's still the best available read on two-way impact, but it isn't a like-for-like subtraction.

Wherever you see Freq%, it's a share of attempts, not a rate of success. Rim Freq% 32% means a third of her shots come at the rim — it says nothing about whether she makes them. The FG% column right next to it answers that.

Frequencies within one family always total 100%, which is what makes them readable as a diet : six shot zones, or catch-and-shoot versus pull-up on threes.

A single shooting-efficiency number that combines 2-pointers, 3-pointers, and free throws. Unlike FG%, it gives 3PT shots their proper 50% bonus and accounts for how many free throws the player draws.

TS% = PTS / (2 × (FGA + 0.44 × FTA))

The denominator of TS%: how many scoring possessions a player or team actually used, free throws included. Two free throws count as roughly 0.88 of a shot, because most trips to the line come from a single possession. Where TS% tells you the efficiency, TSA tells you the workload — two players at 58% TS% do not carry the same weight at 400 versus 900 TSA. Shown per game, alongside FGA and 3PA.

TSA = FGA + 0.44 × FTA

On the league Opp Shooting tab, DefTSA is the same thing seen from the other side: how many scoring possessions a defence hands over per game. Lower is better. Read next to DefTS%, it separates a defence that concedes few shots from one that concedes bad ones.

FG% adjusted for the fact a 3-pointer is worth 50% more than a 2-pointer. Lighter than TS% (it ignores free throws) but more honest than raw FG%.

eFG% = (FGM + 0.5 × 3PM) / FGA

The eFG% you would get with this player's (or team's) shot distribution if every shot were converted at the league average rate for the zone it came from. It measures shot selection, not shooting: a rim-heavy diet scores high because rim shots go in more often, whoever takes them.

LocEFG% = Σzones freq(zone) × league eFG%(zone)

The zones are the six we use everywhere on the site — Rim, Short Mid, Mid, Long Mid, Corner 3, Top 3 — and the league reference is scoped to the same season and season type you are looking at. A 2026 player is compared to the 2026 league, never to a blend.

What this number does not include: shooting fouls. Cleaning the Glass has had shot-location data on shooting fouls since 2010-11 and folds them into where a team shoots from. We do not have it. Measured on the WNBA 2026 regular season: of 4 422 shooting fouls, only 1 265 have a field goal attempt recorded alongside them — a missed shot on a foul is never logged as an FGA. The ones we can see are 55% rim shots against 27% for the league overall, so the attempts that go missing are overwhelmingly attempts at the basket. Our shot distribution therefore under-represents the rim, and LocEFG% is biased downward, systematically, for players who live on contact. The bias is the same for everyone, so comparing two players stays valid; the absolute value is less so.

College is a special case, twice over. On the NCAAW and NCAAM pages the column is served for 2026 only: ESPN supplies shot coordinates on 99.7% of 2026 women's shots and 99.6% of the men's, but on 14% to 42% of earlier seasons — and coverage is a property of the game, not the shot, so those seasons are a few hundred fully-covered games picked by ESPN rather than a random sample. Earlier seasons show "—". And the six zones behind the college number are read from coordinates, while the Rim% and Mid% columns beside it are read from the play description ("layup", "jumper"). Same page, two different splits: you cannot recompute the college LocEFG% by hand from the columns next to it.

One more deliberate choice: heaves are kept (end-of-period desperation shots). Cleaning the Glass removes them because they inflate 3-point frequency and drag LocEFG% down. Our play-by-play feed carries no heave marker, so excluding them would mean inventing a clock-and-distance rule, and it would put this column at odds with the Freq% columns next to it. Measured on the WNBA 2026 regular season, shots from 30+ feet in the last three seconds of a period are 171 out of 12 908 three-point attempts — 1.3% of 3PA, 0.5% of all attempts.

Real eFG% minus LocEFG%, in percentage points. This is the actionable half: it isolates shot-making from shot selection. Positive means the player converts better than their shot diet predicts; negative means the shots are good and the conversion is not.

eFG +/- = eFG% − LocEFG%

A player at 52% real eFG% with a 54% LocEFG% is not shooting well — they are shooting from good places. The reverse (48% real, 44% expected) is a shooter carrying a poor shot profile. The same shooting-foul caveat as LocEFG% applies, and in the same direction.

How often a player gets to the line relative to their field goal attempts. A proxy for foul-drawing skill, paint pressure, and physicality.

FTr = FTA / FGA

Share of field goal attempts taken from beyond the arc. Tells you the shot diet — a guard at 0.65 lives behind the line, a center at 0.05 stays inside.

3PAr = 3PA / FGA

Every field goal attempt is filed into one of six locations, from the play-by-play shot coordinates :

Rim        2PT, under 3 ft
Short Mid  2PT, 3 ft to under 10 ft    (SMR)
Mid        2PT, 10 ft to under 16 ft
Long Mid   2PT, 16 ft to the arc       (LMR)
Corner 3   3PT the league files as a corner attempt
Top 3      3PT everywhere else (above the break)

Distances are measured from the shot's court coordinates, not read from the feed's own distance field — the two disagree by half a foot or more on the same shot. The bounds are half-open and the distance is not rounded : a 2.9 ft shot is a rim shot, a 3.0 ft shot is not. Rounding to the nearest foot, which we did until September 2026, silently moved every shot within half a foot of a boundary into the wrong zone.

The corner is the league's own call, not ours and not Basketball-Reference's. Each league labels its own shots, and we publish that label. On the NBA this puts us about 1.5 points above Basketball-Reference's corner share ; on the WNBA, where the league marks as « corner » only a narrow band along the baseline, we sit around 9% of three-point attempts where Basketball-Reference reads about 19%. Both numbers are correct measurements of two different definitions. We measured the WNBA three-point line to settle it : its minimum distance rises continuously from 21.80 ft at the baseline to 22.68 ft at the top, with no straight corner segment at all — so a WNBA « corner » is a convention, and a convention belongs to the league that sets it. The NBA line, by contrast, does break : straight at ~22 ft up to 3 ft of depth, arc at ~23.75 ft from 9.5 ft on.

Each zone reports three things. FG% is accuracy inside the zone. Freq% is the share of all the player's attempts taken there — the six add up to 100%, so it describes a shot diet. Astd% is the share of the makes in that zone that came off a teammate's pass : high at the rim means cuts and rolls, low means self-created.

The tables labelled 0-3 / 3-10 / 10-16 / 16-3P are the same six zones under Basketball-Reference's naming, not a second measurement — the distance buckets match theirs, the corner split does not (above). Zone AST% is only computed for a player's own shots — the opponent version always reads zero.

Splits three-point attempts into shots taken off the pass and shots the player created herself. Freq% is the split of her 3PA between the two ; FG% is accuracy within each.

Read this one with care. It is not optical tracking. The feed tags a shot with a descriptor — pullup, step back, fadeaway, turnaround — on roughly 23% of threes. A shot carrying one of those tags is filed as a pull-up ; everything else, tag or no tag, defaults to catch-and-shoot. So catch-and-shoot is the residual bucket and will be slightly overstated.

A cross-check on 200 games supports the split : 87% of the shots we call pull-ups have no assist attached, and 75% of the ones we call catch-and-shoot do. Directionally sound, not shot-by-shot exact.

Mean distance of all field goal attempts, in feet. A blunt one-number summary of shot selection : useful for spotting a big who has moved out to the arc, or a guard who has stopped settling.

AvgD = sum of shot distances / FGA

Attempts arriving without a distance still count in the denominator, so the figure runs slightly low. Opp AvgD is the same measure for the shots an opponent took — a defence that pushes it up is keeping people out of the paint.

« Layup » here means a shot from 0 to 3 feet — it is the Rim zone under another label, not a detection of the shot's mechanics. A dunk, a tip-in and a floater from three feet all count.

Layup %FGA is the share of attempts taken there. Layups Md is the number made : a season total on player pages, a per-game figure on team pages.

F2P% and F3P% are the share of attempts taken from two and from three — they add to 100%. F3P% is the same number as 3PAr, written as a percentage. 2P% is accuracy inside the arc, the counterpart to 3P%.

The share of a team's two-pointers and three-pointers, teammates only, that this player assisted. It separates the guard who feeds shooters from the one who feeds cutters.

The numerator only counts passes thrown while she was on the floor, while the denominator covers every game the team played. A player who missed games will read low.

Points scored in the paint, on the fast break, on second chances after an offensive rebound, and by players off the bench. Same points, sorted by how they were created — two teams can score 85 and get there in opposite ways.

Opp PtsOffTOV is the mirror : points the opponent scored off your turnovers. The three categories overlap by design — a fast-break layup off a steal counts in all three.

Points scored per 100 possessions. Pace-neutral, so a slow team can still have an elite ORtg. Team-level ORtg above 115 is excellent in the modern WNBA / NBA / NCAA contexts. At the player level, Metrics shows the individual Dean Oliver rating : points produced per 100 possessions the player uses (not the team's score while they're on court).

ORtg = 100 × PTS / Possessions

Points allowed per 100 possessions. Same logic as ORtg, lower is better. Combined with ORtg, gives the cleanest snapshot of a team's two-way quality. Player-level DRtg is the individual Dean Oliver estimate of points allowed per 100 defensive possessions.

DRtg = 100 × PTS allowed / Possessions

Simply ORtg − DRtg. A positive number means the team scores more than it allows per 100 possessions. The single best one-number team summary.

Net Rtg = ORtg − DRtg

Same idea as ORtg, but expressed per single possession (so ORtg / 100). Often used in lineup and play-type contexts because individual possessions are the natural unit there.

Estimated possessions a team plays per 40 / 48 minutes. High pace = lots of possessions, fast tempo. Pace doesn't make a team better — it just inflates the box score. ORtg / DRtg strip pace out.

The denominator is minutes actually played, not games, so an overtime night doesn't inflate the number.

Each of these is one of the Four Factors, measured for you and for your opponent, then subtracted. They answer « did we win this particular battle ? » rather than « how good were we at it ? ». A team can shoot poorly and still post a positive NetTS% if it made the other side shoot worse.

NetTS%  = OffTS%  − DefTS%
NetAST% = TmAST% − OppAST%
NetTOV% = DefTOV% − OffTOV%   (flipped : forcing them is good)
NetRB%  = OffORB% − the ORB% you conceded
NetFTR  = OffFTR  − DefFTR

Zero is average, positive is better, and the unit is percentage points — a NetTS% of +4 means four points of true shooting above what you allow. NetEff is the same number as NetTS%, kept under a second name on the league tables.

Combines the three ways a team wins extra trips — offensive rebounding, forcing turnovers, defensive rebounding — against the one way it loses them, its own turnovers. Positive means you finish the night having played more possessions than the opponent.

NetPoss = OffORB% + DefTOV% + DefDRB% − OffTOV% − 100

NetPoss and NetTSA answer two different questions. NetPoss asks how many trips you win ; NetTSA asks how many shots you get out of them. A team can win the possession battle and still lose the shot battle.

The −100 is what puts average at zero : a rebound always goes to one side or the other, so OffORB% and DefDRB% add up to about 100 before anything else is counted.

The figure is centred on zero everywhere : +5 means you played five percentage points more possessions than the opponent, −3 means the reverse.

The gap between the shots a team creates for itself and the shots it concedes, per 100 possessions. Positive means that over an equal number of trips, you get the ball up more often than the opponent does.

NetTSA = 100 × TSA / possessions − 100 × opponent TSA / opponent possessions

NetPoss asks how many trips you win ; NetTSA asks how many shots you get out of them. The two come apart : turning the ball over, or settling for a trip that ends at the line, costs shots without costing a possession.

It is a volume figure, not an efficiency one — it says nothing about whether those shots went in. Read it next to NetTS%, which says the opposite thing. A team can lead the league in shots created and still lose by ten.

Averaged across a league it comes to zero, by construction : every attempt one team creates is an attempt another team concedes.

How much better or worse than league average opponents have shot from three while this player or team was on the floor. Positive means they have missed more than expected.

3PLuck = league average opponent 3P% − opponent 3P% against you

Contesting a three-pointer moves its odds a little ; whether it drops is mostly out of anyone's hands. Treat a large 3PLuck as a warning that a defensive rating is about to regress, not as a skill.

Dean Oliver's four levers of winning : shooting (eFG%), turnovers (TOV%), offensive rebounding (ORB%), and getting to the line (FT/FGA). Each side of the ball has the same four — Metrics shows both for every team.

Dean Oliver's "how often a possession the player uses ends in at least one point". It rewards efficiency and ball security in one number : a possession that ends in a turnover, or in a miss with no offensive rebound / free throw, does not count. Higher = fewer wasted possessions.

Floor% = Scoring Possessions / Total Possessions

Share of a player's points that came off a teammate's assist — i.e. scored without creating the shot herself. High for cutters, spot-up shooters and screen-runners ; low for isolation / pick-and-roll creators. Complements Astd% (which is about made shots, not points).

OffBallScoring% = 100 × (PTS via assist) / PTS

Share of team possessions a player "uses" by shooting, getting fouled, or turning the ball over while on the floor. League average is 20% (five players → 100%). Stars sit around 28–32%, role players 12–15%.

Estimated share of teammate field goals a player assisted while on the floor (Dean Oliver). Better than raw AST/G because it adjusts for pace and minutes. Do not confuse with Astd% below : AST% measures what she creates for others, Astd% measures how often she is set up. Also distinct from TmAST%, which is a team-wide rate, not a personal one.

Share of made field goals that came off an assist — for your team (TmAST%) and for the other side (OppAST%). On a team row it is the whole season; on a player row it is measured only while that player is on the floor.

TmAST%  = team AST / team FGM × 100
OppAST% = opponent AST / opponent FGM × 100

These are collective measures, not individual ones. They describe how much the ball moves while she plays — not how much she moves it. That is AST%, and the two sit in very different ranges : across the 2025 WNBA season TmAST% and OppAST% both cluster near 68 %, because they track the league's overall assisted share, while an individual AST% spans roughly 5–45 %. Seeing 30 % next to 68 % on the same row is expected — they are not the same statistic.

Because both sides use the same formula over the same minutes, their difference is meaningful : that is NetAST%.

Share of a player's own made field goals that came off a teammate's assist. A shot-creation signal, not a playmaking one : self-creating guards run low (they make their own looks), spot-up shooters and rim- runners run high. Shown on the Offense tab — distinct from AST% (Dean Oliver) on the Playmaking tab.

Astd% = 100 × (assisted FGM) / FGM

The shot-zone tables carry the same ratio zone by zone (Rim Astd%, 3PT Astd%…) : assisted makes in the zone over makes in the zone. On a team row it is computed over the team's makes.

AST% divided by USG%. A quick playmaker-vs-scorer profile : above 1.0 the player creates more for teammates than she uses possessions herself (pure point guards sit well above 1) ; well below 1 signals a finisher / scorer profile.

AST:USG = AST% / USG%

Assists per turnover. The oldest ball-security number there is, and still the most readable : above 2 is careful, below 1 is loose. It says nothing about volume — a player who never passes posts a fine ratio.

AST/TO = AST / TOV

A player with zero turnovers shows a dash rather than infinity.

Of everything a player does with the ball — shooting, drawing fouls, passing, losing it — the share that ends in an assist. Unlike AST/TO it accounts for volume, so a high-usage guard and a bench passer can be compared directly.

AstRatio = 100 × AST / (FGA + 0.44 × FTA + AST + TOV)

The share of her team's made field goals that she assisted. Close cousin of AST%, with one difference that matters : AST% adjusts for minutes played, this one doesn't. Read AST% for a rate, read this for a raw share of the team's offence.

%TmFGM AST = 100 × AST / team FGM

Of the offensive possessions her team played while she was on the floor, the share in which she appears in the box score — a shot, a free throw, an assist, a turnover, an offensive rebound, or a foul drawn. One possession counts once, however many events she registers in it.

Involvement% = 100 × on-court offensive possessions with ≥ 1 event
                / on-court offensive possessions

This is an exact count, not an estimate: every event it reads is written in the play-by-play. What it measures is how often she leaves a trace — not how often she touches the ball. A player who catches it, swings it and gets no assist leaves no trace, and that possession counts as uninvolved. Dribbles, post entries and simple passes are invisible to the play-by-play, so Involvement% is a floor, never a touch count.

We previously published this as OnBall%, alongside OffBall% and Touch/Poss. The name promised ball-handling the play-by-play cannot see; OffBall% was its exact complement, and Touch/Poss borrowed a tracking term no public WNBA source can verify — it also correlated at r = 0.96 with this column. Both were removed in September 2026.

Turnovers per 100 plays (shooting plus FT trips plus turnovers). A high-usage guard at 12% is excellent; over 18% is a problem.

TOV% = 100 × TOV / (FGA + 0.44 × FTA + TOV)

Share of available offensive rebounds a player grabs while on the floor. More signal than raw ORB/G because it controls for minutes and missed shots.

An offensive rebound a player grabs off their own missed shot — as opposed to the usual offensive rebound, which follows a teammate's miss. Three columns: the raw count (MeB), the per-game average (MeB/G), and the share of that player's offensive rebounds (MeB%).

Not inferred — read straight from the feed. Every rebound event carries a pointer to the shot it followed, on 100% of offensive rebounds. We never guess from event order, so a blocked shot followed by a scramble is never miscounted.

Read the count and the rate together. The rate is mostly a position marker, not a skill ranking. A guard grabs few offensive rebounds but takes most of them off her own shot — she is already standing where the ball comes back. A center grabs many, most of them off other players' misses. In the NBA the highest rates belong almost entirely to perimeter creators (45–56%), the lowest to pure centers (12–15%).

Denominator: the player's own offensive rebounds, never the team total — 37 to 39% of offensive rebound events belong to no player (dead balls, out of bounds) and would flatten the rate by a third. Same reasoning as DRB% below.

Free throws are excluded. A rebound off your own missed free throw is technically your shot, but the shooter cannot enter the lane before the ball hits the rim — a different skill, and only 31 such cases across the 2025 WNBA season.

League average: 23.2% in the WNBA (2025), 19.8% in the NBA (2025-26). Stable within a player from season to season (r = 0.62 to 0.80 across eleven season pairs), far less so within a team — a team's figure reflects who is on the roster, not a system.

WNBA and NBA only. The NCAA feed does not link rebounds to shots, and we would rather show nothing than a number built on a different method.

Rebound rate split by where the missed shot came from: rim, short mid, mid, long mid, corner three, above the break and free throw, plus a 3PT total. For each zone, the share of the available rebounds that were grabbed — offensive rebounds on your own team's misses from that zone (ORB%), defensive rebounds on the opponent's misses from it (DRB%). Found in the Rebound tab of the League, Team, Players and Player pages.

Origin, not location. The public play-by-play has coordinates for shots, never for rebounds. This is where the miss came from — not where the ball was caught.

Read straight from the feed. Every rebound points to the shot it followed, on 100% of player rebounds (WNBA 2022 onward, NBA 2021 onward). Nothing is inferred from event order, and the zones add up to the player's rebounds.

Players: chances while on the floor. A player's rate only counts the rebounds available while they were on court. A team's counts every rebound available in its games. Five players share each chance, so a player's rates run at about a fifth of a team's — compare players with players.

Blank under 50 chances — below that, the rate is noise. Corner threes are rare in the WNBA: most players never reach 50 corner-three chances in a season, so read the 3PT total there. Hover a cell to see the count behind the rate.

Team rebounds (dead balls, out of bounds) belong to no player and are left out, as everywhere on Metrics.

WNBA and NBA only. Same reason as MeBounds: the NCAA feed does not link rebounds to shots.

Share of available defensive rebounds grabbed. Closing out possessions on the glass is one of the four winning factors (Four Factors).

Team rebounds are excluded. The play-by-play feed logs a rebound event whenever the ball changes hands off a miss — including dead balls, the end of a quarter, or a missed free throw going out of bounds. Those belong to no player, and the official box score leaves them out of player totals, so we do too. It matters more than it sounds : 41% of the offensive rebound events in the feed are team rebounds, against 10% of the defensive ones. Counting them would tilt the ratio, not just blur it. Corrected 12 August 2026 — our totals now match the official box score game for game.

Combined ORB% + DRB%, weighted by opportunities. A clean one-number rebounding rate.

Share of opponent possessions ended by the player's steal. Pace-neutral equivalent of STL/G.

Share of opponent 2-point attempts blocked while on the floor. The cleanest individual rim-protection signal in the box score.

Blocks divided by personal fouls. A discipline read on rim protectors : two players can block the same number of shots, but the one doing it without fouling stays on the floor in the fourth quarter.

BLK/PF = BLK / PF

Charges counts offensive fouls she drew — the willingness to step in front of a driver, which shows up nowhere else in a box score. PFD counts every foul drawn, on either end : a scorer's ability to get to the line, mostly.

Both are per game. Charges depend on the feed tagging the fouled player, so treat a zero as « none recorded » rather than a certainty.

Field goal attempts the opponent took per game while she was on the floor. A volume, not a rate — a starter will always show more than a reserve. Read it next to the opponent shooting percentages, where it tells you how big the sample behind them is.

Dean Oliver's method for splitting a team's wins among its players. Offensive Win Shares compare what a player produced to what a marginal player would have produced on the same possessions; Defensive Win Shares do the same on the other end. WS is their sum.

Sanity check built in: a league's Win Shares should add up to roughly its number of wins. On Metrics that lands at 98% for the WNBA and 97% for the NBA — a useful reminder that WS is an allocation, not a measurement.

Hidden below a games threshold. WS builds on individual offensive and defensive ratings, which we hide for very small samples. A blank cell means "not enough games", never "zero contribution".

Win Shares at a common playing-time scale, so a starter and a reserve can be compared. The scale follows the league: WS/48 in the NBA, WS/40 in the WNBA and NCAA, where games are 40 minutes long. Using 48 everywhere would inflate the shorter leagues by a fifth and put us at odds with every published reference.

An estimate of a player's impact in points per 100 possessions, above a league-average player. Unlike Win Shares, BPM weights each statistic by the player's estimated position and offensive role — a block from a guard and a block from a centre are not worth the same, and neither are their assists. Both are estimated from the box score itself.

The team's total is then anchored to its efficiency adjusted for strength of schedule, which redistributes the credit the box score leaves unassigned. DBPM has no regression of its own: it is BPM minus OBPM, by construction.

Good on offence, indicative on defence. The box score records blocks, steals and rebounds — not positioning, communication or rotations. Treat DBPM as a hint, not a verdict.

Coefficients come from twenty years of NBA data. Applying them to the WNBA and to college is a deliberate choice, not a neutral one: no equivalent regression has been published for those leagues.

Not comparable across leagues. BPM measures a player against the average of their own league. Adjusted team ratings span roughly −12 to +12 in the NBA and the WNBA, but −78 to +68 in women's college basketball, where the best programmes play very small schools. A college +24 and Jokić's +13 are not the same quantity. The same caveat applies to VORP, whose replacement level is NBA-calibrated.

BPM converted into cumulative value: how far above a replacement-level player (fixed at −2.0) someone performed, multiplied by their share of playing time and scaled to a full season. It rewards being good and being available — a high BPM over few minutes stays small.

Every other metric on this tab is built from the box score. RAPM is not. It reads only who was on the floor and what the scoreboard did, then solves for the value of each player that best explains every possession of the season at once — with teammates and opponents accounted for. That is how a player who neither scores nor rebounds can still rank high: RAPM sees the effect, not the act.

Heavily regularised, on purpose. Five players share every possession, so the raw problem has no single answer. We pull every estimate toward zero until the model predicts games it has never seen — the penalty was chosen by cross-validation across whole games, never across possessions of the same game. The price is compression: the best player in the league lands near +3.5, not +10. Read the ranking, not the gap.

Where this comes from. The estimator is the one Dan Rosenbaum published in 2004 and Joseph Sill regularised at Sloan in 2010, and nothing about it is ours: one signed column per player, each stretch of play weighted by its possessions, the ridge penalty applied to players and never to home-court advantage. What every implementation has to choose for itself — the penalty, the window, which players are estimated individually — we choose by validation and publish below, following the WNBA-specific study Dan Falkenheim released in 2026. There is no single "official" RAPM to copy; there is a method, and there are the choices you make inside it. Ours are on this page.

Blank, not zero, for the smallest workloads. The least-used 30% of players are not estimated one by one. Too few possessions and the model cannot tell them apart from the teammates they shared the floor with, so they share a single reference value instead — which in the WNBA sits near −6.8 per 100, not near zero. We show them nothing rather than that shared figure, because it describes the group and not the player. We used to estimate them individually, which pulled each of them to roughly 0.0 and quietly placed a player with twelve minutes in the middle of the league.

A number near zero means "we don't know" as often as it means "average". The same caution applies well above the cutoff: a player just over it has been separated from her teammates barely better than the ones just under.

Two windows, side by side. The first three columns cover the selected season alone. The three marked 3Y are fitted over that season and the two before it, every game counting the same. We used to fade older games out on a 700-day half-life; when we finally tested that choice against everything else, it was the single most costly thing in the model, and we removed it. Reading the two windows together is the point: a player well above their own 3Y figure is having a better year than their recent past, and the gap is usually more informative than either number on its own.

Why the longer window exists. One season is often not enough to separate a player from their teammates. Split-half reliability — refit the model on odd-numbered games, then on even-numbered ones, and correlate — rises from 0.54 to 0.61 in the WNBA as the window goes from one season to three. The cost is that a player who changed sharply last summer is described partly by the player they used to be.

How reliable, exactly. WNBA, three-year window, players above 1,800 possessions on the floor: 0.61 for RAPM, 0.55 for ORAPM, 0.44 for DRAPM. Year-to-year, the single-season ranking correlates 0.41 with the next season's. Those are the numbers, unrounded. Defence is the weakest of the three and always has been — a lineup-only model sees points prevented, but it cannot see who prevented them.

We used to set a bar at 0.50 and we no longer do. Split-half reliability rewards timidity: a model that shrinks every player to zero reproduces itself perfectly and predicts nothing. We measured this directly — the settings that predict unseen games markedly better are the ones that score about 0.04 lower here. So reliability is now published rather than used as a gate, and what a column has to clear before it ships is out-of-sample prediction: it must beat simply guessing the average margin, on games the model never saw. Every column here does. Some candidates did not, and we wrote up why instead of shipping them.

Blank means the window does not exist. The earliest seasons we hold have no three-year history behind them, so the 3Y columns are empty there rather than quietly computed on one or two seasons under a label promising three.

ORAPM and DRAPM do not add up to RAPM, and that is not a rounding error. They come from a different regression. RAPM solves for one number per player against the scoring margin — the classical formulation. ORAPM and DRAPM come from a second fit that gives every player two coefficients instead of one, so that offence and defence can be read apart. Twice the unknowns from the same possessions means the two models land in slightly different places, and adding the halves of one to compare against the other is not a check that should pass. Use RAPM for where a player ranks, and the split for where their value comes from.

The split is less reliable than the total. Noise that cancels when the two halves are added does not cancel within each half — see the figures above. Read the total as a measurement and the offence/defence split as a strong hint, DRAPM most of all.

Prior-informed. The single-season column comes in two forms. The plain one starts every player from scratch each year: on 1 October, a ten-year veteran and a rookie are equally unknown to it. The prior-informed one starts each player from what the floor said about them over the previous three seasons — a separate window that shares no games with the one being fitted — and lets this season move them from there. It predicts unseen games better, which is why it exists.

What the prior-informed column is really telling you. It correlates 0.94 with the three-year column beside it, so most of the time it says something you can already read there. Its value is the disagreements: a player well above their prior-informed figure is outrunning their own recent history, and one well below is coasting on it. There is no prior-informed three-year column, because we measured one and it changed nothing — three seasons of evidence already contain what a prior would have added.

Season-long only. Filtering by date or opponent leaves these columns empty. RAPM is a regression over a full season; restricting it to a handful of games does not produce "RAPM over those games", it produces noise. Both signs read the same way here — a positive DRAPM means points prevented.

Not available for college. College play-by-play did not record substitutions before 2026, so the five players on the floor cannot be reconstructed at all. Those pages will stay empty for several more seasons — no window length fixes missing data.

Two teams at +3.0 are not equal if one played the top of the table and the other the bottom. We solve every game at once for the team ratings that best explain the margins, home advantage included. In 2025 the swing was ±0.8 points per 100 possessions across the NBA and ±1.4 across the WNBA, whose schedule is less balanced.

College games are treated as neutral-court. Our college feed labels the two teams by their position in the payload, not by who hosted — more than half of college teams would otherwise appear to never play at home. Rather than infer a home edge from a field that doesn't carry one, we assume none. It is the honest reading: we don't have the information.

A lineup's offensive PPP minus defensive PPP. Positive = the 5-player combination outscores the opponent per possession when together. Sample size matters — Metrics shows minutes played alongside the rate.

Team's Net Rtg when player X is on the court minus Net Rtg when off. Strong positive signal that the player elevates the team — though it conflates lineup quality with the player themselves.

Point differential while the player is on the floor. Noisy at small samples but useful as a context check on box-score stats.

Each axis blends several stats : the axis value is the average of the player's league percentiles for those stats, computed within her position group (guards vs guards, etc.). Composition — Scoring : USG%, Off PPP, FGA, PTS. Playmaking : AST%, AST:USG, AST/TO, %TmFGM AST, Involvement%, TOV% (inverted), Assist Ratio. Rebounding : DRB%, ORB. Defense : STL, BLK, DREB, Charges/G, BLK/PF. Efficiency : USG%, TS%, eFG%, Off PPP. Volume : USG%, AST%, TOV% (inverted). TOV% enters inverted so that a high value always reads as "good" (protects the ball).

axis = mean(percentile(stat₁) … percentile(statₙ))

Shown when a player has 2+ seasons. Same six axes, but each underlying percentile is averaged across seasons, weighted by playing time — so a 60-minute rookie year barely moves the career shape.

career percentile = Σ (pctᵢ × MINᵢ × GPᵢ) / Σ (MINᵢ × GPᵢ)

One offensive trip ending with a made shot, defensive rebound, or turnover. Offensive rebounds extend the same possession. Estimated possessions = FGA + 0.44 × FTA − ORB + TOV.

Points per possession on a specific play type (transition, pick-and-roll ball-handler, post-up, spot-up, etc.). Surfaces which actions a player or team actually wins on.

Why you won't find it here: play types are hand- or model-tagged per possession by a proprietary provider. Our event log doesn't label them. The closest thing on Metrics is the possession-level efficiency split (OffRtg / DefRtg / PPP on-court) plus the fast-break and second-chance context flags.

Expected eFG% on a shot given location, defender distance, shot clock, and shooter movement. Tells you whether a shot was open or contested — not just whether it went in.

Why you won't find it here: defender distance and shot clock come from optical tracking (every player's position, 25× per second) — a proprietary feed we don't have. Our play-by-play records events, not positions, so it carries the shot location but never the defender's. What Metrics does give you is shot selection: attempt frequency and accuracy by zone (corner 3, above the break, mid-range, at the rim).

Average announced crowd for the team's home regular-season games, taken from the official box score of each game. Playoffs are excluded : the number of home playoff games varies from zero to four or more depending on how far a team goes, and playoff crowds run higher, so including them would make two teams incomparable and make the in-season figure jump the moment the postseason starts.

Only games played at the team's usual venue count. A team that stages a showcase game in the larger arena across town, or plays a regular-season game abroad, has that game listed separately rather than folded into the average — Chicago drew 8,025 at Wintrust Arena in 2025 and 19,549 for two games at the United Center, and one number covering both would describe neither.

Fill rate is that average against the arena's capacity in its basketball configuration, which is the only capacity that means anything here : many arenas curtain off an upper bowl and sell a smaller house on some nights.

The fill rate is left blank whenever capacity can't be proven. No feed publishes arena capacity, so we take it from the games themselves : the largest crowd at a game the league flagged as a sellout. Where an arena is reconfigured night to night that number isn't a capacity at all — Phoenix recorded sellouts from 9,876 to 17,071 in 2025 — and where a non-sellout crowd exceeds every sellout, the flag itself can't be trusted. In both cases you get the attendance and a dash, never a percentage we can't stand behind. About two thirds of teams have a fill rate in a given season.

An empty season is not a zero. Out of season the current year reads Not started, not 0. A genuine zero — the 2020 Orlando bubble, the closed-door games of 2021 — is real and is kept out of the average and counted separately, because the feed writes the same 0 for a game played behind closed doors and for one whose crowd was never announced.