NBA projections · Accuracy ledger

NBA Projection Accuracy

Every projection we publish is recorded before the game and graded the next morning. Record as of Nov 22, 2:31 AM ET.

This record is the real test. The models were tested on two past seasons before launch, and that test is shown at the bottom of this page. But it was not a blind test: the 2025-26 season, and 2024-25 with it, were in view while the models were designed, and early checks on parts of both shaped design choices. Results on seasons that informed the design are an optimistic guide, not proof. The record below is kept on games nobody had seen when the projection was made. It is the out-of-sample test, and it is the one to judge these projections by.

It is an error ledger, not a win-loss record: no sportsbook line or price is involved, so there is nothing to win or lose. What can be measured is how far each projection landed from what happened, and whether the plainest alternative, the player's average over his last ten games, would have been closer. Both are shown, including when the plain average wins.

The record since launch

Projections are published twice on a game day and the two runs are scored separately, never pooled, because they do not know the same things. 8 graded slates, Nov 14, 2026 to Nov 21, 2026.

Overnight run

Published after the night's games. It knows who sat out each team's last game and nothing newer. This is the case the test before launch scored.

7 graded slates: 1,063 player projections, of which 89 (8.4%) were for players who then did not play. A projection assumes he plays, so those are counted here and not scored.

Overnight run, to date. Error is the mean absolute miss per game. "vs last ten" is how much smaller the projection's error is than the last-ten average's (negative: the plain average was closer). The 80% range should hold the result about 80% of the time; for stats that are usually 0, 1 or 2 it holds more, because a whole-number range cannot be exactly 80%. A stat marked "No" is still recorded, unpublished.
Stat Player-games Projection error Last-ten error vs last ten In 80% range Published
Minutes 974 4.530 4.862 +6.8% 84.4% Yes
Points 974 4.795 4.941 +3.0% 78.3% Yes
Rebounds 974 1.974 2.050 +3.7% 87.5% Yes
Assists 974 1.390 1.409 +1.3% 91.3% Yes
Threes 974 0.994 1.029 +3.4% 93.0% Yes
Steals 974 0.741 0.760 +2.5% 93.7% Yes
Blocks 974 0.520 0.520 +0.0% 95.8% No (test)
Turnovers 974 0.946 0.967 +2.2% 93.0% Yes
Pts + Reb + Ast 974 6.076 6.265 +3.0% 79.7% Yes
Pts + Reb 974 5.709 5.880 +2.9% 78.6% Yes
Pts + Ast 974 5.144 5.306 +3.0% 79.0% Yes
Reb + Ast 974 2.630 2.695 +2.4% 85.5% Yes
Overnight run, day by day: the last 7 graded slates. One night is a small sample: the projection loses to the last-ten average on some nights and that is shown as it is.
Slate Players Minutes error Last ten Minutes in range Points error Last ten PRA error Last ten
Nov 21, 2026 109 4.47 5.02 81.7% 4.69 4.83 6.03 6.21
Nov 20, 2026 170 4.00 4.26 89.4% 4.90 4.99 6.16 6.33
Nov 19, 2026 70 4.45 4.79 88.6% 4.54 4.78 5.80 5.83
Nov 18, 2026 225 4.64 4.85 84.4% 4.58 4.73 5.92 6.06
Nov 17, 2026 88 4.51 5.13 81.8% 4.77 4.96 6.20 6.65
Nov 16, 2026 148 4.19 4.54 85.8% 4.58 4.62 6.10 6.12
Nov 15, 2026 164 5.33 5.57 79.3% 5.37 5.60 6.27 6.63

Pregame run

Published before the evening tips. It also reads the listed injury report. It was never tested before launch, because no history of injury reports exists: this table is its only test.

8 graded slates: 1,011 player projections, of which 64 (6.3%) were for players who then did not play. A projection assumes he plays, so those are counted here and not scored.

Pregame run, to date. Error is the mean absolute miss per game. "vs last ten" is how much smaller the projection's error is than the last-ten average's (negative: the plain average was closer). The 80% range should hold the result about 80% of the time; for stats that are usually 0, 1 or 2 it holds more, because a whole-number range cannot be exactly 80%. A stat marked "No" is still recorded, unpublished.
Stat Player-games Projection error Last-ten error vs last ten In 80% range Published
Minutes 947 4.488 4.846 +7.4% 84.8% Yes
Points 947 4.683 4.865 +3.8% 79.6% Yes
Rebounds 947 1.955 2.031 +3.7% 88.5% Yes
Assists 947 1.385 1.403 +1.3% 90.8% Yes
Threes 947 0.968 1.006 +3.7% 93.7% Yes
Steals 947 0.751 0.772 +2.7% 93.9% Yes
Blocks 947 0.535 0.539 +0.8% 95.0% No (test)
Turnovers 947 0.952 0.988 +3.6% 92.9% Yes
Pts + Reb + Ast 947 5.991 6.222 +3.7% 80.8% Yes
Pts + Reb 947 5.594 5.809 +3.7% 79.3% Yes
Pts + Ast 947 5.045 5.244 +3.8% 79.7% Yes
Reb + Ast 947 2.582 2.671 +3.3% 86.1% Yes
Pregame run, day by day: the last 8 graded slates. One night is a small sample: the projection loses to the last-ten average on some nights and that is shown as it is.
Slate Players Minutes error Last ten Minutes in range Points error Last ten PRA error Last ten
Nov 21, 2026 89 4.38 4.95 84.3% 4.52 4.65 5.68 5.84
Nov 20, 2026 166 4.13 4.25 88.5% 4.92 5.07 6.22 6.42
Nov 19, 2026 74 4.43 4.64 90.5% 4.38 4.65 5.66 5.67
Nov 18, 2026 216 4.41 4.67 84.7% 4.59 4.81 5.82 6.07
Nov 17, 2026 85 4.36 5.04 83.5% 4.83 5.05 6.21 6.71
Nov 16, 2026 150 4.17 4.53 86.7% 4.55 4.60 5.98 6.07
Nov 15, 2026 83 5.50 5.81 75.9% 5.31 5.63 6.09 6.52
Nov 14, 2026 84 5.27 5.96 79.8% 4.37 4.55 6.30 6.58

The rule that takes a number down

  • Taken down on evidence, not on a bad week. Each morning, for each stat, the run's most recent 300 graded player-games are compared game by game: how far the projection missed, against how far the last-ten average missed. The stat is withheld only if those 300 differences show, at 90% confidence, that the projection is the worse of the two. Losing on average error is not enough by itself; it has to be a loss that chance alone would rarely produce. The minutes projection is held to the same rule.
  • The test. A one-sided paired t-test on the per-game differences in absolute error. A t-test was chosen over a bootstrap because it uses no random numbers: anyone can recompute the verdict from the record and get the same answer. With 300 games the two methods agree closely.
  • Coming back takes a fresh 300 games. A withheld stat keeps being projected and recorded, unpublished. It returns only after 300 graded player-games, all played after it was taken down, in which its average error beat the last-ten average. If that block falls short it is thrown away and the count starts again from zero. A few good nights inside a longer stretch do not bring a stat back.
  • Automatic, and separate for each run. The rule is applied by the same job that grades the slate. Nobody decides case by case. The overnight run and the pregame run each have their own switch for every stat, judged on their own records: one can be withheld while the other is published.
Where every number stands today. "Counting": fewer than 300 graded player-games so far, published on the strength of the test before launch. For a published stat the note gives its error and the last-ten average's over the most recent 300 games.
Stat Overnight runPregame run
Minutes Published (last 300: 4.15 v 4.54) Published (last 300: 4.25 v 4.56)
Points Published (last 300: 4.84 v 4.96) Published (last 300: 4.76 v 4.96)
Rebounds Published (last 300: 1.95 v 2.07) Published (last 300: 1.96 v 2.05)
Assists Published (last 300: 1.36 v 1.37) Published (last 300: 1.34 v 1.34)
Threes Published (last 300: 0.97 v 1.00) Published (last 300: 0.96 v 1.01)
Steals Published (last 300: 0.75 v 0.75) Published (last 300: 0.77 v 0.78)
Turnovers Published (last 300: 0.94 v 0.93) Published (last 300: 0.95 v 0.95)
Pts + Reb + Ast Published (last 300: 6.10 v 6.26) Published (last 300: 6.01 v 6.20)
Pts + Reb Published (last 300: 5.77 v 5.90) Published (last 300: 5.68 v 5.87)
Pts + Ast Published (last 300: 5.12 v 5.26) Published (last 300: 5.00 v 5.21)
Reb + Ast Published (last 300: 2.58 v 2.68) Published (last 300: 2.57 v 2.67)

How often the rule takes down a number that is fine

Measured on the 2025-26 test projections, overnight case. Every stat in this table beat the last-ten average over that season, so each time the rule would have taken one down, it would have been wrong. Across all of them, a raw loss on average error happened in 94 of 1,023 stretches of 300 player-games (9.2%); the confidence test would have withheld in 10 (1.0%).

The rule is read every morning, not once per stretch, so over a whole season it gets many chances to be wrong. Run night by night through 2025-26, it took down 7 of the 11 numbers at least once, 9 times in all, and each came back after its fresh block: 29 of 2,310 stat-nights spent withheld (1.3%). Stats that share an ingredient go down together: in that replay a poor stretch for points took combined stats that contain points down on the same morning. The rule errs the other way too. Blocks, which failed the test before launch (-1.4% against the last-ten average), would have been withheld on only 48.6% of nights: a stat that is slightly worse can stay up for a while before the evidence is strong enough. Like the rest of the test before launch, this is a guide and not proof.

False withholdings by stat, 2025-26. "Stretches": the season's graded player-games in date order, cut into consecutive blocks of 300; the false-withhold rate is the share of blocks in which one reading of the rule would withhold. "Lost on raw error" is what a rule without the confidence test would have withheld. "Season replay": the rule run as it runs live, read after each of 210 slates, with the 300-game wait to come back.
Stat Stretches Lost on raw error False-withhold rate Season replay: times taken down Slates spent withheld
Minutes 93 2 (2.1%) 0 (0.0%) 1 (back 1) 2 of 210 (0.9%)
Points 93 11 (11.8%) 1 (1.1%) 1 (back 1) 4 of 210 (1.9%)
Rebounds 93 4 (4.3%) 0 (0.0%) 0 0 of 210 (0.0%)
Assists 93 13 (14.0%) 0 (0.0%) 0 0 of 210 (0.0%)
Threes 93 18 (19.4%) 2 (2.1%) 3 (back 3) 8 of 210 (3.8%)
Steals 93 12 (12.9%) 0 (0.0%) 1 (back 1) 3 of 210 (1.4%)
Turnovers 93 7 (7.5%) 0 (0.0%) 0 0 of 210 (0.0%)
Pts + Reb + Ast 93 8 (8.6%) 2 (2.1%) 1 (back 1) 4 of 210 (1.9%)
Pts + Reb 93 5 (5.4%) 2 (2.1%) 1 (back 1) 4 of 210 (1.9%)
Pts + Ast 93 10 (10.8%) 3 (3.2%) 1 (back 1) 4 of 210 (1.9%)
Reb + Ast 93 4 (4.3%) 0 (0.0%) 0 0 of 210 (0.0%)

The test before launch

Read this as a guide, not as proof. The 2025-26 season was meant to be the untouched season the publish decision was read on. It was not untouched: a check on two months of it, and one on two months of 2024-25, came before the full test and led to three design choices. The numbers below are therefore somewhat better than the models deserve, by an amount nobody can state. The record above has no such problem.

The models were tested on the 2024-25 and 2025-26 seasons, month by month: for each month, the models were fitted only on games played before that month and then scored on it, so no single result comes from a game the fitted model had seen. 5,461 games and 116,979 player-games, from Oct 18, 2022. The numbers are for the overnight run's case: the model knew who sat out each team's previous game and nothing newer.

Minutes. Mean absolute miss per game. "Absences known" assumes the model knew exactly who would sit, the best a pregame run could do.
Season Player-games Projection Last-ten average vs last ten In 80% range Absences known
2024-25 27,731 4.862 5.251 +7.4% 79.7% 4.662 (+11.2%)
2025-26 28,095 4.807 5.232 +8.1% 80.2% 4.631 (+11.5%)
2024-25 by stat. "Largest calibration gap": the stated chances of clearing the research line were sorted into ten groups; this is the widest distance in any group between the chance stated and how often it happened.
Stat Projection Last-ten average vs last ten In 80% range Largest calibration gap Passed
Points 4.594 4.747 +3.2% 79.4% 1.5% Yes
Rebounds 1.907 1.971 +3.2% 87.6% 1.3% Yes
Assists 1.329 1.374 +3.3% 90.9% 1.6% Yes
Threes 0.893 0.911 +2.0% 92.9% 1.7% Yes
Steals 0.713 0.733 +2.7% 94.3% 1.3% Yes
Blocks 0.525 0.520 -1.1% 95.5% 1.4% No
Turnovers 0.880 0.911 +3.5% 93.6% 2.3% Yes
Pts + Reb + Ast 6.026 6.264 +3.8% 78.6% 1.7% Yes
Pts + Reb 5.545 5.748 +3.5% 78.7% 2.4% Yes
Pts + Ast 5.046 5.233 +3.6% 79.2% 1.7% Yes
Reb + Ast 2.616 2.717 +3.7% 85.3% 1.7% Yes
2025-26 by stat (the season the publish decision was read on). "Largest calibration gap": the stated chances of clearing the research line were sorted into ten groups; this is the widest distance in any group between the chance stated and how often it happened.
Stat Projection Last-ten average vs last ten In 80% range Largest calibration gap Passed
Points 4.529 4.681 +3.3% 80.1% 2.1% Yes
Rebounds 1.871 1.940 +3.5% 88.4% 2.6% Yes
Assists 1.331 1.370 +2.9% 91.1% 1.6% Yes
Threes 0.886 0.902 +1.7% 93.5% 2.0% Yes
Steals 0.719 0.736 +2.3% 94.5% 2.6% Yes
Blocks 0.517 0.510 -1.4% 95.4% 1.3% No
Turnovers 0.886 0.916 +3.3% 93.6% 2.2% Yes
Pts + Reb + Ast 5.942 6.207 +4.3% 79.8% 2.6% Yes
Pts + Reb 5.452 5.675 +3.9% 79.6% 2.2% Yes
Pts + Ast 4.984 5.186 +3.9% 80.0% 1.8% Yes
Reb + Ast 2.576 2.682 +4.0% 86.0% 1.3% Yes

What the test decided

  • Minutes: passed. On 2025-26 the error was +8.1% against the last-ten average (the bar was 5%) and the 80% range held the result 80.2% of the time (the bar was 77% to 83%).
  • 10 of 11 stats beat the last-ten average. A stat passed only if it beat that average and every calibration group was within 4 points.
  • Did not pass: Blocks. It did not beat the last-ten average, so it is not published. It is still recorded in the tables above, unpublished.
  • What the test cannot show: there is no free history of sportsbook lines for NBA player stats, so nothing here was tested against a market, and none of it says anything about betting results. There is also no history of injury reports, which is why the pregame run had no test at all before launch.