Method

Hoop Impact Bench (HIB) is an open-source benchmark for NBA player impact metrics. It computes how well each metric’s ratings predict NBA games, on the same games with the same code: github.com/lumahoops/hoop-impact-bench.

A game

A team’s rating is 5 × the minutes-weighted average of its players’ preseason ratings. Only the players who played count, at the minutes they played, so the test is the ratings alone. Predicted margin = home rating − away rating. The pick is the higher-rated team. No home-court points.

W, L, PCT

Picks right, picks wrong, share right.

GB

Games behind the leader.

L10

Record over the last 10 games.

STRK

Current streak.

LAST

Record on the latest game day.

▲ ▼

Places moved since the previous game day. Equal records share a place (T5), ordered by margin r².

Margin r²

Share of the variance in final margins that predicted margins explain, fit with a home-court term. Small numbers: the 90% bootstrap interval.

Win % r

Correlation of team rating with team win % (Narsu, 2024).

Net err

RMSE of team rating against team margin adjusted for opponent and home court (Snarr, 2020). Grades scale as well as order.

Avg miss

Mean points between predicted and final margin.

Two tracks

Preseason

Each metric’s rating from before opening night, never updated, on every game. Under 250 minutes last season (rookies included) or not rated: −2.0. ARC 2-year is not on this board: a two-year window ending 2025–26, not a forecast.

Dynamic

Each metric’s rating as published the morning of the game, saved before tip-off. Every player counts at the metric’s value, rookies included; not rated: −2.0. A metric that publishes nothing newer keeps its last value. A game counts only when every metric has a value saved before tip-off.

Pre

On the dynamic standings: the preseason record on the same games.

Historical

Each metric’s results on the 2020–21 to 2025–26 regular seasons (7,230 games), computed using its ratings from before each season.

Established

Under 250 minutes the season before, rookies included: −2.0 for every metric (Snarr, 2020).

Full forecast

Every player at the metric’s own rating.

Missing

A player missing from a file: −2.0.

Rank

Places in the middle 90% of 2,000 resamples of team-seasons. Overlapping ranges: the data do not separate the metrics.

Unranked

References. Same-season RAPM is fit on the same games (a ceiling). Everyone 0 and last season’s team are floors.

Past ratings come from each metric’s current model, which may fit past seasons better than the model of the time did. Live games come after the ratings, so live is the stricter test.

Saved files

Archived

Public before the games, with a dated copy.

Saved by HIB

Saved by the bench at the snapshot.

Backfill

Taken later from a site’s date view: the current model, for a past date.

Each file’s sha256 is in the repository. The files belong to their makers and are not published here.

harness v1.1.0 · fingerprint cfde3f922e882983… · snapshot Oct 31

Run it

pip install hoop-impact-bench
hib --ratings my_metric.csv

One CSV per metric: season, nba_id, rating. To join the live standings, use the submission form.