Inspiration
Points per shot is the number everyone quotes, and it answers the wrong question. Rudy Gobert scored 1.365 points a shot in 2025-26; Luka Dončić scored 1.127. Nobody thinks Gobert is the better shooter. He has a better diet, taking only the shots the offence hands him at the rim, while Dončić manufactures contested jumpers. Every attempt is two separate decisions welded together, and the box score refuses to pull them apart.
So we asked a question a front office actually has to answer: of the two, which one can a team change?
What it does
Shot Diet pulls 1,087,633 regular-season field goal attempts (2021-22 through 2025-26, every game of all five seasons) and splits each player's efficiency in two:
PPS − league PPS = (xPPS − league PPS) + (PPS − xPPS)
shot SELECTION shot MAKING
A gradient-boosted model estimates what a league-average shooter would score on each attempt given where it came from, how it was created, the clock and the venue. What the diet is worth in average hands is selection. What the shooter added on top is making. They sum exactly to efficiency above league average.
It ships as a reference tool: search any player, get his sheet — summary line, hex shot chart, zone-by-zone table, season history, league ranks, and a shot-diet prescription.
The finding
| Shot selection | Shot making | |
|---|---|---|
| Attempts before the metric is 50% signal | 11 | 311 |
| Within-season reliability (split-half, Spearman-Brown) | 0.97 | 0.62 |
| Season-to-season correlation | 0.90 | 0.58 |
| Spread across players (SD, pts/100) | 8.5 | 8.8 |
| Of which is repeatable | 8.4 | 6.9 |
In any one season the two look equally decisive, which is why they get conflated. Strip out measurement noise and 59% of the genuinely repeatable difference between players is simply which shots they take. Selection is knowable from eleven attempts. Shot making needs most of a season before half of it is even signal.
The two also pull against each other (r = −0.27): the players handed the easiest shots are the weaker shot-makers, and the best shot-makers are handed the hardest shots. That is precisely why raw efficiency flatters one group and buries the other.
2025-26. DeMar DeRozan has the worst shot diet of any high-volume player in the league (−17.1 pts/100, with 48.5% of his attempts from the mid-range) while being a clearly above-average shot-maker (+12.0). Gobert is the exact mirror (+35.5 / −8.2). Jokić takes a below-average diet and is the league's second-best maker of it (+21.3), for the fifth straight season.
The result we did not expect
The optimiser is a linear program: reallocate at most 5% of a player's attempts across six zones to maximise expected points, using empirical-Bayes estimates of his own zone-by-zone shooting. Graded against the following season it is worth +2.49 points per 100 shots and helps 99.5% of players. A nice headline, so we ran the controls that could kill it.
| Out-of-sample comparison | Mean pts/100 | 95% CI |
|---|---|---|
| Personalised prescription vs. do nothing | +2.49 | +2.43 to +2.55 |
| Generic league-average advice vs. do nothing | +2.50 | +2.44 to +2.56 |
| Personalised vs. generic league-average advice | −0.01 | −0.04 to +0.03 |
| Shrunk vs. unshrunk (raw hot-zone) advice | +0.33 | +0.26 to +0.40 |
Personalising a shot diet to the individual shooter is worth nothing. Every bit of the gain comes from the league-average structure of the floor: get out of the mid-range. This is not an artefact of a tight constraint, because the null holds from a 3% to a 20% move budget, and the two prescriptions genuinely disagree for about half the league. Meanwhile shrinking rates clearly beats trusting them, and the advantage widens the more volume the optimiser is allowed to move.
It is the same story the reliability table tells. Shot making needs 311 attempts to be half signal; a single zone inside a single season never gets close. We shipped the negative result rather than the feature.
Actionable impact
- Shot quality belongs to the scheme far more than to the roster. Selection shows up in eleven attempts and persists at r = 0.90. It is the high-leverage, low-variance lever a staff actually controls.
- Coach the diet with league-average zone values. Do not build a personalised shot-diet plan off one season of shooting splits; you will be coaching noise. This is a concrete warning about something analytics departments really do.
- Stop paying for shot-making outliers as if they were fixed. A third of the season-to-season variance in making does not carry over. The stabilisation thresholds say exactly how much of a hot season to believe.
- Read every efficiency leaderboard through the split. A 1.36 pts/shot centre and a 1.13 pts/shot guard can be the same player wearing different roles.
How we built it
Python: nba_api → pandas → scikit-learn → SciPy LP → Streamlit + Plotly.
HistGradientBoostingClassifier over shot geometry (x, y, distance, angle off-centre), clock state, venue, season and the recorded play type. Two models: geometry-only and geometry-plus-play-type; the latter drives the headline split, on the reasoning that whether a shot is a cut, a pull-up or a turnaround fadeaway is a property of the offence rather than the shooter's touch.
| Model | Log loss | Brier | AUC | Gain vs. base rate |
|---|---|---|---|---|
| League mean make rate | 0.6915 | 0.2492 | 0.500 | |
| Zone average make rate | 0.6587 | 0.2330 | 0.633 | 4.74% |
| Distance spline logistic | 0.6542 | 0.2309 | 0.642 | 5.39% |
| xPPS-loc (geometry) | 0.6492 | 0.2289 | 0.648 | 6.12% |
| xPPS-full (+ play type) | 0.6358 | 0.2236 | 0.663 | 8.06% |
Calibration matters more than discrimination here, because an uncalibrated model would not make the decomposition add up. Across all 1.09M shots the model's expected points sit 0.004 points per 100 from actual scoring.
Challenges
The API lies quietly. shotchartdetail truncates every response at exactly 102,400 rows. Request a season and you get roughly the first half of it, through mid-January, with no error and no warning. Our first pull looked perfectly healthy and was missing 53% of the data. The loader now pages by calendar month and asserts no chunk reaches the cap; all five seasons come back at the full 1,230 games.
Zones that were too clever. An initial ten-zone scheme split left from right. The empirical-Bayes prior strength then swung from k = 199 on one wing to k = 2018 on the other for the same shot. That is binomial noise rather than talent, and the optimiser started recommending left-side threes over identical right-side threes. Collapsed to six well-sampled zones.
Players grading themselves. A high-volume specialist takes enough shots to shift the very model that judges him. Every prediction used to evaluate a player is out-of-fold with folds grouped by player ID, so the model scoring a player has never seen one of his shots.
What we learned
That the honest version of a result is usually the more interesting one. We set out to build a personalised shot-diet recommender and finished with proof that personalisation does not work at this sample size, which is a far more useful thing to hand a coaching staff than a tool that quietly fits noise.
What's next
Defender distance (the public feed has none, so some of what we call "making" is really separation-creation), free
Log in or sign up for Devpost to join the conversation.