14 November 2026
The NBA's expected field goal percentage model looks at only one second of each shot: the 60 frames before the ball leaves the shooter's hand. It does not know the shot clock, the score, or who is shooting. That is a deliberate design choice, and it shapes how every number the model produces should be read.
Tracking data and the inputs behind shot-quality models
Shot-quality metrics rest on optical tracking. According to the NBA's introduction to expected field goal percentage, the league's cameras capture 29 points on each player's body and record movement 60 times per second. That is enough resolution to describe not just where a shooter stood, but how he was standing: whether his body was squared to the rim, whether he was balanced, and how a defender's body was positioned relative to him.
From that raw feed, the NBA's model builds three families of inputs for every attempt: contest posture, shooter orientation and balance, and court location. Contest posture goes beyond a single distance reading. A defender two feet away with a hand up and a defender two feet away who is trailing the play are different situations, and the model is built to separate them.
The older commercial framing from Second Spectrum, the company behind much of the league's tracking analysis, leaned more on geometry. ESPN's account of that work described quantified shot quality as a prediction built from the positions, movement, distances, and velocities of the shooter, the closest defender, the next closest defender, and the ball. The inclusion of the second-closest defender matters. It means the model was never only asking "how tight was the contest" but also "how crowded was the area."
The NBA's version trains separate models for jump shots and for finishes around the rim. The reason is intuitive. A pull-up jumper and a layup through traffic fail for different reasons, and a single model would blur the signals that matter for each: balance and footwork on the perimeter, contact and angle at the basket.
How expected effective field goal percentage is calculated
Start with the actual-results side. The NBA's stats glossary defines effective field goal percentage as (FGM + 0.5 × 3PM) / FGA. The adjustment gives a made three-pointer extra credit for being worth more than a two. Take a hypothetical player who hits 5 of 10 shots, two of them threes: the calculation is (5 + 1) / 10, or 60 percent, even though his plain field goal percentage is 50.
Expected efficiency asks what that same set of shots should have produced. The NBA describes expected field goal percentage as the field goal percentage an average NBA player would be expected to shoot given the player-tracking details of each shot taken. The Second Spectrum concept, as laid out in The Athletic's introduction to shot quality, is the same idea expressed in effective terms: an average NBA player's expected effective field goal percentage on a shot, based on factors such as shot location, defender distance, and other tracking details.
The two measures are cousins, not twins. The NBA's published model works in field goal percentage, so a corner three and a long two with identical contest profiles are judged on whether they go in, not on how many points they are worth. Second Spectrum's formulation folds the extra point in. When you see a number labeled "expected eFG%" or "xeFG%," check which of the two you are looking at before comparing it with another source.
The comparison is the useful part. Suppose a hypothetical player's shots carry an expected eFG% of 52 percent and he actually posts 56. He has outperformed the average-shooter baseline on those looks by four points. Because the baseline is an average player, that gap folds together several things: shooting skill, shot-making difficulty the model cannot see, and plain luck over a small sample. A shooter who beats expectation across thousands of attempts is telling you something. One who does it over a hot month is mostly telling you about the month.
Shot bins by location and contest
Simplified shot-quality work, the kind a fan can do with public tables, starts by sorting attempts into location buckets and then layering contest on top. Location comes first because it carries the most information. Cleaning the Glass publishes league shooting frequency by zone, including rim, short mid-range, long mid-range, corner three, non-corner three, and all threes combined. Its league-wide splits, as of its early 2026 listing, look like this:
| Shot location | Share of league attempts |
|---|---|
| Rim | 31.3% |
| Short mid-range | 22.7% |
| Long mid-range | 8.0% |
| Corner three | 10.0% |
| Non-corner three | 28.1% |
Two things stand out. Rim attempts and non-corner threes together account for roughly six of every ten shots in the league, which is the modern shot diet in a single glance. And the long mid-range, at 8.0 percent, is the smallest bucket on the board, well below the short mid-range at 22.7. Teams and players have largely stopped taking the shot that pays the least per attempt, while the short mid-range, which includes floaters and closer pull-ups, survives at a much higher rate.
The second layer is contest. In a simplified model, analysts typically split each location into bands of defender distance, such as tightly contested, contested, and open. The NBA and Second Spectrum models do something richer, because they treat the defender as a position, a posture, and a velocity rather than a single number. But the principle is the same: for any location, the expected value of a shot moves with how much the defender interferes.
A practical caution for readers building their own bins: location and contest are correlated. Rim attempts tend to come with defenders nearby, while corner threes often arrive with more space. If you compare a corner three to a rim finish without accounting for contest, you are comparing two shots that differ in several ways at once.
What the metrics can and cannot tell you
The NBA is explicit about what its model leaves out. It does not use the shot clock, player identity, or score differential as inputs. Excluding the shooter's identity is what makes the output a baseline. If the model knew it was watching a great shooter, the expected number would rise, and the gap between expectation and result, the whole point of the exercise, would shrink toward nothing.
That choice has consequences. A model that does not know the shooter cannot tell you whether a given look was good for that particular player. A heavily contested step-back from an elite shot creator and the same shot from a bench big man get the same expected value. Coaches who design offenses around specific players will often want to override the baseline for exactly that reason.
The one-second window is another boundary. The model sees the 60 frames before release, which captures the gather, the footwork, and the contest at the moment of the shot. It does not see the possession that led there. A shot that comes after three passes and a defensive rotation may look identical, in that final second, to one that comes after an isolation dribble sequence. Off-the-dribble creation and the work done to generate space can show up only through their end result, the defender's position and the shooter's balance.
Because the shot clock is excluded, late-clock heaves are not graded as a separate category. They are scored on posture, orientation, and location like any other attempt, which can make a forced shot with eight seconds of context behind it look better than it was.
In practice, shot-quality metrics are best used as a filter. They help separate a team that is shooting poorly from a team that is getting poor shots, and a defense that is allowing good looks from one that is merely unlucky on the makes. They are weaker as verdicts on individual shooters.
Why floor spacing and shot gravity matter
Spacing shows up in these models through the defender inputs. When ESPN described Second Spectrum's approach, it noted that the prediction used both the closest defender and the next closest defender. A shooter's look gets better when the defense is stretched, because the nearest help is farther away and the closest defender has more ground to cover on a closeout. The location of a shot also changes with spacing: more room means more corner threes and cleaner rim attempts, the two most efficient places on the floor.
This is where the distinction between shot quality and shooter quality becomes useful in the other direction. A player who draws extra attention, what analysts call gravity, does not necessarily score more efficiently himself. His effect appears in the shots of his teammates, whose defender distances grow and whose expected values rise. The models do not measure gravity directly. They record its footprint: the contest profile and location of the shots a lineup produces.
For readers, the takeaway is to look at the shot profile before the shooting percentage. If a team's expected number is high and its actual number lags, the process is probably working and the makes may follow. If both are low, the problem is upstream, in the spacing and the shot selection that produce those looks.