Methodology

A per-100-possession estimate of a player's impact on offense and defense that controls for the other nine players on the floor. It answers “how much better is his team when he plays, after accounting for who he plays with and against?”

Caveat. Single-season RAPM is noisy. Read it as a descriptive signal, not a verdict. Prefer IPM (next) or the 3-year version (the same fit pooled over three seasons with recency weighting) when they disagree.

Validation. The lineup reconstruction is validated game-by-game: reconstructed per-player minutes match the official box score to within half a second across every game.

Our flagship impact metric. It is the same RAPM ridge, but instead of shrinking noisy players toward zero it shrinks them toward what their box-score profile predicts. That single change is most of what makes the brand-name metrics (EPM, LEBRON) stable, and every input here is our own.

Caveat. IPM inherits the box score's blind spots on defense: the defensive prior explains far less than the offensive one (r-squared roughly 0.10 to 0.25 vs 0.35 to 0.46), so defensive IPM leans more on the on/off evidence. It is also a season-level estimate; it does not yet decay within the season or use tracking data the way EPM does.

Validation. In held-out prediction tests across nine seasons, IPM had the lowest error in 17 of 17 splits: all 9 within-season tests and all 8 predict-the-next-season tests, beating pure RAPM, the SPM prior alone, and a league-average baseline. Single stints are extremely noisy, so the error differences are small; the unbroken 17-for-17 ordering is the signal. Full numbers below.

Two independent reads of team strength, shown side by side: Team IPM is what the roster is worth (each player's IPM, weighted by minutes, times the five players on court), and SRS is what actually happened (point margin adjusted for opponent and venue). Schedule strength (SOS) falls out of the same solve.

Caveat. Team IPM leads the sort because it earns it: in held-out tests it predicts future game margins better than SRS does. But single-season team samples are small, and Team IPM inherits IPM's defensive blind spots. Read the two columns together; big gaps between them are the story (a team out- or under-performing its talent).

Validation. Chronological three-way split per season (ratings fit on the first 60% of games, tested on the final 20%). Mean held-out margin RMSE below. A fitted blend of the two predicts best of all but takes a negative SRS weight (the two ratings are collinear), so we publish the two readable columns and keep the blend for the upcoming win-probability work.

The chance the home team wins, estimated before tip-off from two team ratings built only from games already completed: the season's venue-adjusted point margin so far, and a roster value built from last season's player IPM weighted by the minutes each player has actually logged this season.

Caveat. No injury or rest information, and every prediction uses only completed games: the roster side cuts off strictly before game day, while the results side can also include games played earlier the same evening (worth about 0.2 points of accuracy). The 2020 bubble and empty-arena 2020-21 games mute home court and ride through as noise. Sportsbooks, armed with injury news and market feedback, still grade slightly better; that gap is the honest cost of using only our own numbers.

Validation. Leave-one-season-out across 2018-19 to 2025-26: 9,509 held-out games, averaged over the eight held-out seasons. The blend posts the lowest Brier score in both windows, and its two ingredients trade places the way they should: the roster prior carries the early season before the results rating has a sample, then hands over. Eight of the ten reliability bins hold 150 games or more, and every one of those lands within 1.1 points of its prediction; the 0-10% and 90-100% tails hold 7 and 56 games and are too thin to read.

How much a specific five-man unit outperforms the sum of its parts, the part of a lineup's net rating the individual players don't already explain.

Caveat. Only shown for units with enough shared possessions, for seasons with published lineup data (2017-18 onward).

How much each made basket swung its team's chance of winning, summed over the season, a leverage-weighted measure of scoring. A buzzer-beating game-winner is worth ~0.9 wins; a garbage-time bucket ~0.

Caveat. This is scoring WPA: misses and turnovers are not debited (that needs a possession-aware model). Read it as leverage-weighted scoring, not all-in-one value.

Box Plus/Minus estimates a player's per-100 contribution from their box-score line and team context; VORP scales it by playing time against a replacement-level baseline.

Caveat. Our estimate is computed on NBA.com data and tracks Basketball-Reference closely (within ~0.2–0.3) but won't match to the decimal. Different pace/ratings sources, and no position labels in the feed.

Whether a player finishes better or worse than a league-average shooter would on his exact shot diet, pure shot-making, separated from shot selection.

Caveat. Location-based, so it rewards tough shot-making rather than easy looks. A rim-running big who dunks a lot won't top it just for being efficient.

Box Creation estimates the open shots a player generates for teammates per 100 possessions, the shot-creation half of offense that assists alone miss. Offensive Load is the total share of the offense he shoulders (creation + scoring load + turnovers).

Caveat. Box-score estimates: they capture creation volume, not pass quality. Ben Taylor's Passer Rating (which would add that) needs layup-assist tracking we don't carry, so we deliberately don't show it rather than approximate it.