Methodology

The unit of observation is one team's performance in one game. Two kinds of measurement go into it: counting statistics taken straight from the box score, and the per-possession ratings published by the league, which express scoring relative to how many possessions a team used. The second matters because teams differ in pace — a fast team scores more points without being more efficient, and per-possession figures remove that difference.

Points allowed are not a separate measurement: they are the opposing team's points in the same game. So the two rows describing a game are consistent with each other by construction, and no game can contribute a scoring margin to one team that it does not contribute, with the opposite sign, to the other.

Preseason is out of scope. Exhibition minutes are not drawn from the same distribution as competitive ones — rotations are experimental and effort is not comparable — so including them would contaminate an early-season average precisely when it is least stable.

Courtrend's derived estimates, charts, and rankings may be cited and shared freely with attribution to Courtrend. Courtrend is an independent analytics project and is not affiliated with, endorsed by, or sponsored by the National Basketball Association (NBA) or its teams. All NBA team names, logos, and statistics are the property of their respective owners.

unit of observation  one team's performance in one game
coverage             all 30 franchises, seasons 2016-17 to 2025-26
phases               Regular Season, PlayIn, Playoffs (preseason excluded)
measurements         box-score counts; per-possession ratings as published
                     by the NBA's public stats API
refresh              nightly, so a corrected box score propagates within a day
missing data         a game with no recorded score is treated as missing,
                     never as a zero
invalid data         a box score with no shot attempts of any kind is
                     impossible; such rows are rejected rather than smoothed
                     over, and nothing is published until they are resolved

A single game is a poor measurement of a team. Shooting variance alone moves a game's result several points in either direction, and garbage-time minutes distort the rest. Averaging the most recent 10 games keeps the part that persists and discards much of the part that does not.

Choosing the window length is the one real decision, and it is a trade. A short window tracks a genuine change quickly but is dominated by noise; a long window is precise about a level the team may no longer be playing at. Ten games is about three weeks of schedule: long enough that sampling noise is roughly a third of what a single game carries, short enough that a real change in a team surfaces within a month.

Three properties are worth being explicit about. The average is trailing, so no point uses information from a game played after it — the chart at any date is what could have been known at that date. It is unweighted: within the window every game counts equally, and outside it none counts at all. And it is indexed by games, not days, so a window may span two weeks or six.

x̄(t) = (1/k) · Σ x_i        over the k most recent games up to and including t
                                k = min(10, games played)

Simple rectangular moving average of the raw game values, per team.
Before a team has 10 games the window is the games that exist — a shorter
average, never a padded or missing one.

We use a flat window rather than an exponentially weighted one deliberately. An EWMA responds faster and never fully forgets, but its effective sample size is not something a reader can hold in their head, and its value cannot be checked by hand. "The last 10 games" can be recomputed by anyone with the box scores, and being checkable is worth more here than being optimally responsive.

One team scores 108 to 116 every night; another alternates 95 and 130. Their averages are identical and their seasons are not. We measure the difference as the dispersion of the individual game results over a window — the sample variance of the raw game values, not of the smoothed line.

That distinction is the whole point. The variance of an already-smoothed series would be tiny and would mostly reflect how much the average drifted, which is a different question. On a team alternating 80 and 140 points, the dispersion of games is 1000 and the dispersion of the moving average is near zero. It is the first number that tells you what watching them is like.

s²(t) = (1/(k−1)) · Σ (x_i − x̄(t))²        sample variance, ddof = 1

Reported for points, assists, rebounds, turnovers, True Shooting % and net rating.
k = 10 for a team's own series, k = games played for the League Overview.
A single game has no dispersion to measure: the published series reports zero,
and the League Overview leaves that team unranked rather than ranking a zero.

Two windows, for two questions. A team’s own charts carry the 10-game figure, which is what “erratic lately” means. The League Overview ranks teams on the dispersion of every regular-season game they have played to the date you are viewing, because “the most erratic team this season” should not be decided by ten games and thrown away by the eleventh. Same estimator, same games, wider window — and roughly three times the precision, since the standard error falls with the square root of the sample.

The featured teams are compared at the same date and within the regular season only, so the six panels answer the same question about the same population. Where two teams tie, the order is fixed rather than arbitrary, so the selection does not change between reloads.

Field-goal percentage treats a three-pointer like a layup and ignores free throws entirely, so it ranks teams by shot selection as much as by shooting. True Shooting divides points by the scoring chances used to produce them, counting a free throw as 0.44 of a chance. As a rough scale: 0.50 is poor, 0.57 is around typical, 0.60 is very good.

Over a window we take the ratio of the totals — all the points divided by all the chances — rather than averaging the ten single-game percentages. The two are not the same estimator. Averaging percentages gives a 70-shot night the same influence as a 95-shot night, which understates exactly the games where a team did most of its scoring.

per game     TS% = PTS / (2 · (FGA + 0.44 · FTA))

over the window
             TS% = ΣPTS / (2 · (ΣFGA + 0.44 · ΣFTA))

i.e. a ratio of sums (volume-weighted), not a mean of per-game ratios.
The two coincide only when volume is constant across the window.

Most numbers here only mean something relative to a baseline: 112 points is good or bad depending on the year and the pace of play. For every date we form that baseline by carrying each team's most recent rolling line forward and averaging across them (widening to all teams in that phase as they play — all 30 during the regular season), scoped to each season and phase. It is the comparison line on charts, and during the regular season (once all teams have played) it coincides with the reference the schedule adjustment in section 10 measures opponents against.

L(d, p) = (1/|T|) · Σ  x̄_j(d)        j ∈ T, all teams with a line as of date d in phase p

A cross-sectional mean of team-level rolling averages carried forward: an average of averages,
holding composition within each season and phase.

At game 4 an average of 10 games cannot exist; there are four. We use those four rather than withholding the number, which keeps early figures available but makes them noticeably less precise — the standard error of a mean falls with the square root of the sample, so a 4-game average carries about 1.6 times the sampling noise of a full window.

Because that precision differs by view, the treatment does too:

• Charts stay closed until a team reaches 10 games. A line whose early points average four games and whose later points average ten is not one comparable series, and drawing it invites exactly the wrong reading.
• Power Rankings publish from the first game, shrinking a team's own noisy figure toward the league average in proportion to how little is known about it (section 11).
• The data table shows the partial averages unshrunk, because its purpose is to expose the underlying values rather than to interpret them.

standard error of a k-game mean  ∝  1/√k

    k = 3   →  1.8 × the sampling noise of a full window
    k = 5   →  1.4 ×
    k = 8   →  1.1 ×
    k = 10  →  1.0 ×   (calibrated)

Group-play games count in the standings like any other, but turning on the Cup view recalculates every rolling metric strictly across tournament games. Because an NBA Cup run spans roughly four to seven games, these averages cover fewer games than the standard window used elsewhere, and early points reflect only the tournament games played up to that date.

True Shooting % is recalculated over the tournament subset using the same possession-weighted formula applied elsewhere. Schedule adjustments are omitted under this view because opponent strength cannot be estimated reliably from the tournament games alone.

With the view off, the tournament is still marked on charts: group play is shaded and the knockout dates are drawn as vertical lines, alongside the trade deadline and the All-Star break. Those are context for reading the line, and change nothing about it.

Playoff basketball is not a continuation of the regular season: rotations shorten, the opponent repeats, and effort is not comparable. Treating the two as one population would quietly break every cross-team comparison — a team still playing in June would be measured on postseason form against an eliminated team's months-old regular-season figure. So rankings and volatility use regular-season games only, all season.

A team's own charts answer a narrower question — how is this team playing now — and there the window runs continuously. The first playoff game is averaged with the nine regular-season games before it, and only from the tenth playoff game is the line purely postseason. Against those points we draw the team's own regular-season level, so the change in environment is visible rather than implied.

The same division decides what each page draws once a season is over. The League Overview is a comparison — six teams picked out of thirty and set side by side — so its panels stop at the last regular-season game, even for a team that went on a deep run. Plotting the postseason there would put games on the line that the ranking behind it never saw, and would give a finalist twenty-odd more of them than a lottery team in a grid meant to be read across. A team's own page has no such constraint, so it shows the whole year through the last game actually played, with the boundary marked and the regular-season level drawn against what followed. Same rule, read twice.

within a team's own series   window is continuous across the phase change
across teams                 regular-season games only, always
reference level              the team's own average regular-season rolling
                             level for the metric on screen
end of season                after 14 days without a game, no line is drawn
                             rather than a stale one extended

Ten games may span fifteen days or thirty-five. We do not discount older games by elapsed time and we do not restart anything at the All-Star break or the trade deadline. That keeps every published figure the same kind of estimate, and it means a rolling line in late February can be describing a roster that no longer exists.

So the dates where that is most likely are marked: the All-Star break is shaded and the trade deadline is drawn as a line. A window that straddles the deadline is best read as two different teams' games averaged together — the estimate is not wrong, but the thing it estimates changed underneath it.

games are exchangeable within a window, regardless of

    · rest — no back-to-back or schedule-density term
    · travel and altitude
    · venue — no home/away split
    · availability — an absent rotation player is invisible
    · roster continuity — a trade does not restart a window
    · time — a game 5 weeks ago counts like a game last night

Beating a contender and beating a rebuilding team are not the same achievement, and the raw numbers cannot tell them apart. The adjustment restates a team's figure as if every opponent behind it had been exactly league-average, by adding the average gap between the opponents faced and the league baseline on the dates they were faced.

The gap is averaged over the same 10 games as the figure it adjusts. That matters more than it sounds: a single opponent can sit twelve points either side of average, which is further than a 10-game figure typically moves in a month. Applying one opponent's gap to a 10-game average would make the adjusted line swing with whoever happened to be on the schedule that night rather than with how the team is playing.

• Net rating rises when the opponent is stronger than average.
• Offensive rating rises when the opponent's defence is better than average — scoring against a stingy defence is worth more.
• Defensive rating falls when the opponent's offence is better than average — conceding to a great offence is forgiven.
• Points scored rise when the opponent's defence is better than average — 110 against a stingy defence is worth more than 110 against a sieve.
• Points allowed fall when the opponent's offence is better than average — conceding to a great offence is forgiven.
• True Shooting % rises when the opponent's defence is better than average — shooting efficiently against a stingy defence is worth more.
• Defensive rebounds rise when the opponent is strong on the offensive glass — securing defensive boards against a great offensive rebounding team is worth more.
• Offensive rebounds rise when the opponent is strong on the defensive glass — winning offensive boards against a great defensive rebounding team is worth more.
• Total rebounds are the sum of adjusted offensive and defensive rebounds.
• Assists rise when the opponent's defence allows fewer assists per made basket — creating assists against a stingy defence that suppresses ball movement is worth more.
• Turnovers fall when the opponent forces turnovers at an above-average rate. Because turnovers are a metric where fewer is better, facing a ball-hawking defence that inflates your turnovers reduces your adjusted figure to credit you for schedule difficulty.

For each game g the team played, with opponent o(g) and league baseline L(g)
on that date, writing N for rolling net rating, O for offensive rating, D for
defensive rating, P for points scored, A for points allowed, S for True Shooting %,
S_opp for opponent True Shooting % allowed, r_D for defensive rebound rate (DREB / pool_D),
r_O for offensive rebound rate (OREB / pool_O), pool_D for defensive rebound pool per game,
pool_O for offensive rebound pool per game, AST_opp for opponent assist rate allowed (AST / FGM),
TOV_opp for opponent turnover rate forced (TOV / POSS), FGM_t for trailing field goals made per game,
and POSS_t for trailing possessions per game, the game's gaps are:

    gN(g)   =  N_o(g)  −  N_L(g)
    gO(g)   =  D_L(g)  −  D_o(g)
    gD(g)   =  O_L(g)  −  O_o(g)
    gP(g)   =  A_L(g)  −  A_o(g)
    gA(g)   =  P_L(g)  −  P_o(g)
    gS(g)   =  S_opp,L(g)  −  S_opp,o(g)
    gr_D(g) =  r_O,o(g)  −  r_O,L(g)
    gr_O(g) =  r_D,o(g)  −  r_D,L(g)
    gAST(g) =  AST_opp,L(g)  −  AST_opp,o(g)
    gTOV(g) =  TOV_opp,L(g)  −  TOV_opp,o(g)

Each gap is signed so that adding it credits the team for schedule difficulty.
Possessions are estimated using the two-team averaged estimate:
    POSS = 0.5 × ( (FGA − OREB + TOV + 0.44 × FTA) + (opp_FGA − opp_OREB + opp_TOV + 0.44 × opp_FTA) )

At game t the published figures add the mean gap over the same 10-game
window the statistic itself covers:

    Ñ_t    =  N_t  +  mean( gN(g) : g in the last 10 games )
    Õ_t    =  O_t  +  mean( gO(g) : g in the last 10 games )
    D̃_t    =  D_t  +  mean( gD(g) : g in the last 10 games )
    P̃_t    =  P_t  +  mean( gP(g) : g in the last 10 games )
    Ã_t    =  A_t  +  mean( gA(g) : g in the last 10 games )
    S̃_t    =  S_t  +  mean( gS(g) : g in the last 10 games )
    R̃_D,t  =  R_D,t  +  mean( gr_D(g) : g in the last 10 games ) × pool_D,t
    R̃_O,t  =  R_O,t  +  mean( gr_O(g) : g in the last 10 games ) × pool_O,t
    R̃_t    =  R̃_D,t  +  R̃_O,t
    ÃST_t  =  AST_t  +  mean( gAST(g) : g in the last 10 games ) × FGM_t
    T̃OV_t  =  TOV_t  +  mean( gTOV(g) : g in the last 10 games ) × POSS_t

Ratings, points and shooting are additive in their own units. Rebounding, assist
and turnover rates are adjusted for opponent strength and re-expressed at count
scale using the team's own trailing denominator (missed-shot pool, made field goals,
and possessions respectively).

A stretch of exactly average opponents leaves every quantity unchanged, so the
adjustment is a no-op by construction. Games whose opponent cannot be identified
are left out of the mean rather than counted as average; only where no game in
the window has an identifiable opponent does the unadjusted value stand.

The adjustment exists across all nine categories (eleven quantities). Every metric is adjusted for strength of schedule when the SOS toggle is active.

The ranking estimates one thing: how much better than its opponents a team has been, per 100 possessions, over its most recent 10 regular-season games. No win-loss record, no conference weighting, no bonus for a marquee win, no human vote. With the schedule adjustment on, the opponent-adjusted net rating from section 10 is used instead.

Using record instead would be a worse estimator of the same thing — a one-point win and a thirty-point win are the same in the standings, and margin is the better predictor of how a team has actually played. Using margin alone, however, is unstable early, so a team's own figure is shrunk toward the league average with a weight that grows with games played: at 3 games the score is 30% the team and 70% the league, at 9 games it is 90% the team, and from 10 games it is entirely the team.

S_i   team i's rolling net rating in its most recent regular-season game
      on or before the viewed date (opponent-adjusted when that view is on)
G_i   team i's regular-season games played by that date
B     mean of S_j over the teams that have played at least one game

    score_i  =  w·S_i  +  (1 − w)·B ,        w = min(G_i, 10) / 10

w is continuous at G = 10, so no team moves in the table merely by reaching
its 10th game. Equal scores keep a fixed order, so the table is stable
across reloads rather than reshuffling.

    trend_i  =  rank_i(t − 7 days)  −  rank_i(t)          positive = climbed

Every one of these came out of a review of our own work, and each is tracked with a decision attached — fix it, or accept it and say so here. Nothing on this list is a surprise to us, which is the point of publishing it.

Rolling average

The average of the most recent 10 games, recomputed after every game.

Possession

One team's turn with the ball. Teams differ in pace, so per-possession rates compare them fairly.

Offensive rating

Points scored per 100 possessions.

Defensive rating

Points allowed per 100 possessions. Lower is better.

Net rating

Offensive minus defensive rating — roughly, points better than the opponent per 100 possessions.

True Shooting %

Points per scoring chance, counting free throws at 0.44 of a chance. Section 4.

Variance

How widely a team's individual games scatter around its own level. Section 3.

Standard error

How much an estimate would move if you could rerun the same stretch of games. Smaller means more trustworthy.

Shrinkage

Pulling a small-sample estimate toward a baseline, because extreme values from few games are usually partly luck. Section 11.

League baseline (LGA)

The average team as of a given date, carrying every team's most recent line forward within each season and phase. Section 5.

Strength of Schedule

Restating a figure as if every opponent behind it had been league-average. Section 10.

Calibration

The stretch before a team has played 10 games and a full window exists. Section 6.

Phase

Regular season, play-in, or playoffs. Preseason is out of scope.

Time Machine

The date control. Every view is rebuilt as of that date, using only games played on or before it.

To verify any point on a chart: take the team's last 10 games up to that date, average the statistic, and compare. That is the whole of it — no weighting to reconstruct, no coefficients to guess. For a schedule-adjusted figure you also need each of those 10 opponents' averages, and the baseline on the date each was faced.

Every method described on this page is computed in exactly one place and checked against hand-worked examples, so the arithmetic is verified independently of whatever the live data happens to be. The figures quoted here — window length, calibration threshold, free-throw weight, comparison window, Cup dates — are not typed into this page. Where the value lives in code this page can read, it is rendered straight from it; where it lives in the pipeline that builds the data, an automated check compares the two and fails on any disagreement. That check runs alongside the rest of our tests, so the page and the numbers are verified together rather than assumed to agree.

Two further checks examine the published data rather than the arithmetic. One re-derives every stored figure from the original games and reports any disagreement, which is how we catch a change in method quietly rewriting history. The other audits what is published — date formats, duplicate team-dates, values outside plausible ranges, and whether the league baseline exists for every date that needs one.

Both are run deliberately, around changes to how the numbers are built, rather than continuously. So they are how a mistake gets found, not a guarantee that none exists — which is the honest version, and the reason section 12 is written the way it is.