Hoop Impact Bench

After 88 games, xPM leads on game R² at .201, ahead of EPM projection (Danny Harlan) at .184. 8 of 45 pairs of metrics are separated on game R² at the 90% level; across 45 comparisons, about 4 would be separated by chance alone, so these are early signals.

Hoop Impact Bench tests NBA player impact metrics (EPM, DARKO, LEBRON, BPM and others) on games played after their ratings were saved. In each game, every player's rating is weighted by the minutes he played, which gives each team a rating in points per 100 possessions. The home team's rating minus the away team's is the predicted margin, and the tables measure how closely the predicted margins follow the final margins.

2026-27 is the current NBA season, which opened on October 20, 2026. The retro tables cover the six seasons before it.

The preseason board uses the ratings each metric had published before opening night of the 2026-27 NBA season (October 20, 2026). These ratings stay fixed, and every 2026-27 regular-season game is computed using them.

Columns

• Game R²: the share of the variation in final game margins that the predicted margins explain. Higher is better. The small numbers are its 90% range.
• Win % r: the correlation between team ratings and team winning percentage. Higher is better.
• Net error: the average miss, in points per 100 possessions, between team ratings and opponent-adjusted team margins. Lower is better. It appears once every team has played 20 games; before that, a few games decide each team's margin.
• W-L: each game counts as a win when the team the metric rates higher won (no home-court edge, the same for every metric).
• Rank: the range of places a metric reaches across 2,000 resamples of the games (the middle 90%). Metrics whose ranges overlap are not separated by the games so far.
• Minutes rule: A player with fewer than 250 minutes in 2025-26, including every rookie, is rated -2.0 for every metric.
• Italic rows are reference points and are not ranked: RAPM fit on these same games (an upper reference), the average of all metrics, last season's RAPM, and two floors.

Labels

• archived: published by the maker before opening night, with a public date (archive snapshot, release or commit)
• saved: copied before opening night, with no public record of the date
• backfill: taken later from the maker's view of a past date, which shows its current model

Sources for the tests: game margins and team net error follow Taylor Snarr's 2020 metric comparison; team win % r is Krishna Narsu's 2024 test; the 250-minute cutoff is Neil Paine's. Questions or errors: open an issue.

Data: frozen.json, every game, history. Computed with eval.py v1.1.0 (sha256 cfde3f922e88...). Rules: live protocol.

After 6 games, ARC Live leads on game R² at .423, ahead of Pulse at .379. Every metric's rank range still overlaps the others, so the games so far do not separate any two metrics.

Hoop Impact Bench tests NBA player impact metrics (EPM, DARKO, LEBRON, BPM and others) on games played after their ratings were saved. In each game, every player's rating is weighted by the minutes he played, which gives each team a rating in points per 100 possessions. The home team's rating minus the away team's is the predicted margin, and the tables measure how closely the predicted margins follow the final margins.

2026-27 is the current NBA season, which opened on October 20, 2026. The retro tables cover the six seasons before it.

The dynamic board uses each metric's ratings as published on the morning of each game day, saved before that day's first game. Metrics that do not update daily, or have published nothing new that morning, are computed using their most recent ratings.

1 game is left out because it started before that day's ratings were saved.

Columns

• Game R²: the share of the variation in final game margins that the predicted margins explain. Higher is better. The small numbers are its 90% range.
• Win % r: the correlation between team ratings and team winning percentage. Higher is better.
• Net error: the average miss, in points per 100 possessions, between team ratings and opponent-adjusted team margins. Lower is better. It appears once every team has played 20 games; before that, a few games decide each team's margin.
• W-L: each game counts as a win when the team the metric rates higher won (no home-court edge, the same for every metric).
• Rank: the range of places a metric reaches across 2,000 resamples of the games (the middle 90%). Metrics whose ranges overlap are not separated by the games so far.
• Minutes rule: Every player is rated at the metric's own value. A player the metric does not rate is -2.0.
• Italic rows are reference points and are not ranked: RAPM fit on these same games (an upper reference), the average of all metrics, last season's RAPM, and two floors.

Labels

• pulled: new ratings saved from the maker's site that morning
• carried: no new ratings published, so the most recent saved ratings are used
• empty: the ratings could not be saved that morning, so that day's games are left out for every metric

Sources for the tests: game margins and team net error follow Taylor Snarr's 2020 metric comparison; team win % r is Krishna Narsu's 2024 test; the 250-minute cutoff is Neil Paine's. Questions or errors: open an issue.

Data: daily.json, every game, history. Computed with eval.py v1.1.0 (sha256 cfde3f922e88...). Rules: live protocol.

Hoop Impact Bench tests NBA player impact metrics (EPM, DARKO, LEBRON, BPM and others) on games played after their ratings were saved. In each game, every player's rating is weighted by the minutes he played, which gives each team a rating in points per 100 possessions. The home team's rating minus the away team's is the predicted margin, and the tables measure how closely the predicted margins follow the final margins.

2026-27 is the current NBA season, which opened on October 20, 2026. The retro tables cover the six seasons before it.

The retro tables compute each season from 2020-21 to 2025-26 (7,230 games, 180 team-seasons) using the ratings each metric had before that season (HIB v1.1.0).

The past ratings come from each metric's current model rather than from what it published at the time, so a model may fit seasons it was built on. The 2026-27 tables do not have that problem.

Columns: game R² is the share of the variation in final game margins that the predicted margins explain; win % r is the correlation between team ratings and team winning percentage; net error is the average miss, in points per 100 possessions, between team ratings and opponent-adjusted team margins. Rank is the range of places across 2,000 resamples (the middle 90%). Italic rows are reference points and are not ranked.

Method, sources and every file: the README and provenance.