Stats Glossary

Box score possessions are estimated with the standard formula, FGA minus ORB plus TOV plus 0.44 times FTA. Play by play possessions are counted directly off the event stream instead of estimated, so the two figures differ slightly for the same game.

Points per game rewards playing fast, since a faster team simply gets more chances at the basket. Per 100 possessions does not, which is why every rating on this site is built on it.

Ratings are fit with ridge regression across every game at once. Each row is team offense, opponent defense and game location, so every team's rating depends on every other team's and the whole system resolves together. Home advantage is learned from the data rather than assumed, and blowout margins are capped so that a single lopsided result cannot drag a team's season rating around.

On the ratings page tempo is opponent adjusted with its own ridge model, where both teams in a game pull the pace toward their own norm. Playing a team that walks it up will drag your raw pace down, and the adjustment accounts for that.

Kills include garbage time so the figures line up with published kill counts. The transition and halfcourt splits exclude it, on the basis that garbage time defense is not really defense.

Men's thresholds by opponent rank are, at home, Q1 up to 30, Q2 up to 75 and Q3 up to 160. At neutral sites, Q1 up to 50, Q2 up to 100 and Q3 up to 200. On the road, Q1 up to 75, Q2 up to 135 and Q3 up to 240. Everything left over is Q4. Women's thresholds are different and are applied automatically on the women's pages. On the men's side Q1 and Q2 split further into A and B bands, which show up on hover.

AdjO minus AdjD. The number shown blends the current season ridge result with a preseason projection, weighted by games played, so that a two game sample in November does not swing a rating wildly. The weight on the prior decays as games accumulate, and by February the rating is almost entirely current season.

League average sits around 106. Elite offenses run 120 and up.

Elite defenses sit under 90.

Calculated as the residual between actual and expected margin per 100 possessions split by venue, then differenced as away plus neutral minus home. It is shrunk heavily toward zero for teams without many road games, since venue splits are noisy on small samples.

Built from the spread of per game performance versus expectation residuals, combining overall volatility with the size of the largest swings, then shrunk toward league average for teams with few games played.

Weighted with exponential decay so the most recent game counts most.

Counted greedily and without overlap, so a run of six stops is two kills and a run of five is one. Streaks carry across halftime and garbage time is included, which matches the convention used by published kill counts. It is worth knowing that kills correlate around -0.88 with points allowed per possession, so in practice they mostly restate stop rate rather than adding independent information. Stops also cluster only slightly more than chance would predict, which undercuts the momentum idea behind the stat.

Year over year correlation is 0.53 on offense and 0.50 on defense. Both columns are deliberately left without a color scale, since running more is a style choice rather than a virtue. A low Trn%D can come from getting back well, but it can equally come from a slow ball control offense that never gives up a run out, so it should not be read as a defensive rating on its own.

Shown uncolored next to the adjusted versions, so the gap between the two columns is the schedule effect for that team.

These deserve more caution than the halfcourt pair. Transition is only about 21 percent of possessions, so the samples behind them are roughly a quarter the size and the ratings repeat year over year at 0.35 on offense and 0.33 on defense, against 0.67 for the halfcourt versions. Adjustment does help, since the raw versions sit at 0.29, but it does not close the gap. A large edge here is more likely to be noise than the same edge in AdjHcO.

Schedule dominates these columns, which is why raw halfcourt correlates only 0.86 with AdjO on offense and 0.66 with AdjD on defense. They carry no color scale for that reason, since a gradient would largely be ranking teams by strength of schedule.

These correlate 0.97 and 0.96 with AdjO and AdjD, but they are not a replacement for them. Because they use only about 79 percent of possessions they repeat year over year slightly less well, at 0.67 against 0.73 for AdjO and 0.72 for AdjD. Read them as a description of where a team's efficiency comes from rather than as a better strength rating.

This is the mechanism behind the transition columns, since live ball giveaways are what actually produce opponent transition. It repeats year over year at 0.61, better than raw turnover rate at 0.50, and it is close to independent of everything else on the page, with a maximum correlation of 0.30 against any existing column. Opponent adjusted.

Repeats at 0.57 on offense but only 0.46 on defense, and the defensive side carries a moderate 0.42 correlation with adjusted halfcourt defense, so it is the weakest of the three profile stats. Opponent adjusted.

Interestingly the defensive side is the more stable one, repeating at 0.56 against 0.43 for offense. Opponent adjusted.

Non D1 opponents are included at a rating slightly below the worst D1 team rather than being dropped, so scheduling cupcakes shows up in the number instead of being invisible.

Each team's schedule is split at its own average opponent, and performance versus expectation is compared on either side of that line. A negative value describes a team that feasts on weak opponents and disappoints against good ones.

Every game is scored against what a bubble quality team would have been expected to do against that opponent at that venue, using a Pythagenport win expectation model. Beating a top team on the road is worth far more than beating a bad one at home, and a genuinely bad loss can be worth less than nothing.

See Quadrants under general concepts for the rank and venue thresholds.

The weights live in model_weights.json and combine AdjEM with RQ and the related resume components.

Use AdjT on the ratings page for the version that accounts for who a team played.

Dean Oliver's framework. Shooting matters more than the other three by a wide margin, which is worth remembering before reading too much into a rebounding edge.

Below a full season sample the displayed value is shrunk toward zero in proportion to games played and marked with an asterisk, and hovering shows the raw figure alongside the weight applied. Without that shrinkage a four game cameo would sit on top of the leaderboard.

Calculated as (BPM plus 2) times (MPG divided by 40) times (games divided by 33). Unlike BPM this is a counting stat, so it rewards availability as well as quality.

Both are box score derived. DRtg in particular is heavily shaped by team defense rather than by the individual, so it is worth reading with some caution at the player level.

Adj ORtg is ORtg minus the difference between average opponent AdjD and league average, so facing stingy defenses earns a boost. Adj DRtg applies the mirror correction using opponent AdjO.

Points divided by shooting possessions, where the denominator is FGA plus 0.44 times FTA.

The reference group is configurable. Clicking the gear beside the column changes what league average means for the comparison.

Calculated as (FGM plus 0.5 times 3PM) divided by FGA.

Built from play by play by carrying the most recent missed shooter forward within a possession and checking whether the rebounder is the same player, which keeps the link intact when a blocked shot event lands between the miss and the board. Misses at the free throw line are excluded. ZB% is only reported at 20 or more missed field goals, since below that a single putback reads as an enormous rate. Covers 2020-21 onward. Earlier seasons are blank because the events inside a possession are not stored in chronological order in that data, so a rebound cannot reliably be tied to the shot it followed.

Taken straight from the play-by-play event type, so no shot coordinates or tagging is needed and it works back to 2010-11. The distinction between a dunk, a layup and a tip is a scorer's-table judgement rather than a real difference in shot type, so the three are grouped into one rim rate.

Covers 2021-22 onward. The blocked tag does not exist in the earlier data, and recovering it from block events instead would need the within-possession event ordering that is unreliable before then.

Available for every season back to 2010-11, since the assister is recorded on roughly half of all made field goals throughout. See the Assist Network page for the full passer to scorer breakdown.

Counted per possession rather than per event, so a scrappy sequence with several whistles does not count more than a clean one.

Points per game times TS% divided by league average TS%.

Threes made per game times 3P% divided by league average.

Assists per game times assist to turnover ratio divided by league average.

Free throws made per game times FT% divided by league average.

Steals plus blocks per game times the sum of STL% and BLK% divided by league average.

Rebounds per game times total rebound percentage divided by league average.

Two separate limits. First, the upstream transition flag changes definition at 2019-20, with the share of events flagged jumping from about 8.5 percent to 15.6 percent without any matching change in how the game was played. Second, the event description tags behind Live%, 2nd% and Paint% do not exist at all before 2019 and are not applied completely for some years after: the share of layups and dunks carrying the paint tag runs 0.86 in 2020 and does not reach 1.00 until 2025. Publishing across either boundary would show trends that are tagging artifacts rather than basketball, so the earlier seasons are left blank.

Team ratings blend toward a preseason prior early in the year, player BPM is shrunk toward zero below a full season, and both Consistency and Road± are shrunk toward league average. All of it stabilizes as games accumulate.

Style columns such as Tempo, transition rate and the raw splits are deliberately left uncolored, since neither direction is good or bad and a gradient would imply otherwise.