Records that outrun their shot quality
Splitting every team’s shooting edge into the looks it generates and the shots it makes, across eleven seasons.
July 29, 2026 · 6 min read · 2015-16 to 2025-26
Two teams shoot the same effective field goal percentage. One gets there by manufacturing layups and corner threes. The other gets there by making contested mid-range jumpers at an unusual rate. One number cannot tell them apart, and the difference is supposed to matter enormously: the first team is doing something repeatable, the second is riding variance that will regress.
That claim is testable. Put every shot into a bin defined by its zone, distance, and whether it was a two or a three, and assign it the league-wide make probability for that bin. Add it up and you get the eFG% a team's shot locations imply — and the same for the shots it allowed. The gap between those two is generation: are you creating better looks than you give up? Whatever is left over, the difference between what a team actually shot and what its locations implied, is conversion.
Across 330 team-seasons from 2015-16 to 2025-26, the conventional story does not survive.
0.84
Conversion correlation with record
explains 71 percent of the variance
0.34
Generation correlation with record
explains 11 percent
0.54
Conversion, year over year
against 0.58 for generation
Conversion is where the standings are decided
Conversion correlates with win percentage at 0.84. Generation manages 0.34. The part of shooting that is supposed to be noise tracks the standings roughly two and a half times as tightly as the part that is supposed to be skill.
The first objection is a scale objection, and it is a fair one. Conversion simply varies more between teams: a standard deviation of 2.37 points of eFG% against 0.90 for generation, or 2.7 times as wide. A component with more spread will move the standings more even if each unit of it matters equally.
So standardize both and regress win percentage on the pair together. That puts them on the same footing, one standard deviation against one standard deviation:
- Conversion: 0.81
- Generation: 0.21
- Together they account for 75 percent of the variance in record, and they are nearly independent of each other (r = 0.15).
The gap survives standardization. Per standard deviation, conversion moves a team's record close to four times as much as generation does. This is not a spread artifact.
And it is not luck either
If conversion were variance, it would not repeat. A team that overshot its shot quality one year would be a coin flip the next, and the year-over-year correlation would sit near zero. It does not.
Year-over-year persistence, same franchise in consecutive seasons
Generation
r = 0.585
Conversion
r = 0.539
Generation persists slightly better, 0.58 against 0.54, which is the one place the conventional story holds up. But “slightly better” is not the same as “skill against luck.” Both are real, repeatable team properties. One of them just matters far more.
Why conversion behaves like a skill
The explanation is in the construction. Conversion is a residual: it is whatever the location model failed to account for. And the location model is deliberately simple — it knows the zone, the distance, and whether a shot was a two or a three. It does not know how open the shooter was, who was closing out, how much time was left on the clock, or whether the shot came off a pass or a dribble.
Everything the model cannot see lands in conversion. That includes genuine randomness, but it also includes having shooters who are better than league average from the same spots, rim protection that makes an eight-footer harder than an average eight-footer, and closeouts that turn an open three into a rushed one. Those are skills. Calling the residual “luck” assumes the location model captured everything that mattered, and a model built from zone and distance plainly did not.
Where each team stood in 2025-26
The upper-left corner is the interesting place to be: below-average shot generation carried by strong conversion. The lower-right is the frustrating one, good process without the finishing. Portland generated the second-best looks in the league and finished around .500 because it converted 2.2 points below expectation.
What this changes
If you are projecting a team forward, do not treat a conversion edge as something that automatically evaporates. Regress it, because 0.54 is meaningfully short of 1.0, but regress generation by a similar amount rather than assuming one is signal and the other is mirage.
The sharper version: a residual is only as trustworthy as the model it is left over from. Improve the shot-quality model — add defender distance, add clock context — and some of what currently reads as conversion skill will move into generation where it belongs. You can see the same components per team on the team comparison page.
Method and limitations
2015-16 to 2025-26 · figures as of July 29, 2026
Data. 330 team-seasons, 2015-16 to 2025-26, regular season only. Shot locations come from shot-level records; expected values use a per-season league baseline binned by zone, distance capped at 32 feet, and shot type. Records and Pythagorean win percentage come from team game logs.
Definitions. Conversion is actual eFG% minus expected eFG%, taken on offense minus the same quantity on defense, where a team's defensive shots are the attempts its opponents took in its games. Generation is expected eFG% created minus expected eFG% allowed. Both center near zero across the league by construction, since every attempt is one team's offense and another team's defense. Units throughout are percentage points of eFG%.
Persistence. 300 pairs of consecutive seasons for the same franchise. This is a correlation, not a causal claim, and roster continuity drives both components — a team that keeps its personnel will persist on both axes. It does not isolate scheme from talent.
What this does not show. The central caveat is the one in the section above: conversion is a residual from a location-only model, so it absorbs shooter quality, defensive contest quality, and true randomness together, and this analysis cannot separate them. A richer shot-quality model would shift the balance between the two components, so the specific magnitudes here are properties of this decomposition rather than facts about basketball. Eleven seasons is also a modest panel; the persistence estimates carry real sampling error, and no era adjustment was applied even though league-wide three-point rates moved substantially over the window.