When Does a Shooting Streak Become a Skill?
Split-half reliability across 3.1 million shot attempts, and why a single season cannot measure three-point shooting.
July 29, 2026 · 7 min read · 2015-16 to 2025-26
A player goes 14 for 26 from three over five games and the conclusion arrives before the next tip-off: he has found his stroke. The honest question underneath it is a measurement question. How many attempts does a shooting percentage need before it tells you about the shooter rather than about the sample?
There is a clean way to answer it. Take every attempt a player took in a season, shuffle them, and deal them into two piles of equal size. If a percentage measures something real, the two piles should agree with each other. If it does not, they will not. Do that for every player, correlate pile A against pile B, and repeat at different pile sizes. The size where the correlation reaches 0.50 is the point where half the spread you see between players is real and half is still noise — the conventional bar for calling a stat stable.
Across 3,090,988 attempts from 2015-16 to 2025-26, the three shot types answer very differently.
~24
Free throw attempts to stabilize
r reaches 0.50
~71
Two-point attempts to stabilize
r reaches 0.50
Never
Three-point attempts to stabilize
within a single season
Split-half reliability by attempts, single seasons
Why free throws are easy and threes are hard
The ordering is not about difficulty, it is about how much genuine difference there is between players relative to the coin-flip noise in any small batch of shots. Free throws are uncontested and identical every time, and the gap between a good free throw shooter and a poor one is enormous — roughly 90 percent against 60. That signal is large enough to cut through the randomness of a few dozen attempts.
Three-pointers are the opposite. Almost every rotation player in the league lives between about 33 and 40 percent, so the real differences are small, while each individual attempt is close to a coin flip. Add defensive pressure, shot difficulty, and who is passing, and the true signal is buried under variance that a single season of attempts cannot dig out.
Two-pointers land in between, and for a reason worth noting: much of what looks like two-point shooting skill is really shot selection. A player who takes most of his twos at the rim will post a high percentage reliably, because where he shoots from is a stable trait even when his touch is not.
So is three-point shooting a skill at all?
It is — the single-season answer was a sample size problem, not a verdict. Pool each player's attempts across all eleven seasons instead of within one, and the curve climbs exactly the way a real skill should.
Three-point reliability: single season against pooled careers
That is the useful version of the finding. Three-point shooting is real and measurable, and it takes on the order of 964 attempts to measure. A rotation player takes maybe 300 to 500 in a season. The number you want simply does not exist yet at the moment you most want to use it.
What to do with a hot streak
Treat a single-season three-point percentage as a weak update to what you already believed about the shooter, not as a new fact that replaces it. The practical form of this is regression toward a prior: a player shooting 44 percent on 60 attempts should be projected far closer to his career rate than to 44.
It is also why the leaderboard qualification thresholds on this site exist. They are not bureaucracy. A three-point percentage leaderboard without a volume floor is a list of players who took very few threes.
Method and limitations
2015-16 to 2025-26 · figures as of July 29, 2026
Data. Two-point and three-point attempts come from shot-level records, so makes and misses are exact: 913,577 three-point and 1,545,071 two-point attempts. Free throws are not shot-chart events, so 632,340 attempts were reconstructed from per-game attempts and percentage; that reconstruction lands on an integer make count within 0.01 for all but 0.01 percent of games.
Procedure. For a target size n, player-seasons with at least 2n attempts are eligible. Each player's attempts are shuffled, 2n are drawn without replacement and split into two disjoint halves of n, and the percentage in each half is computed. The reported r is the Pearson correlation of half A against half B across all eligible player-seasons, averaged over 25 replicates. Replicate-to-replicate standard deviation is under 0.02 everywhere except the highest-n three-point bins, where it reaches 0.06.
Range restriction is a real caveat. Raising n shrinks the eligible pool to higher-volume shooters, who are more alike than the league as a whole. Narrower true spread mechanically lowers a correlation, so the high-n end of each curve is biased downward rather than upward. This affects the three-point curve most: only 124 player-seasons reach 600 attempts. The career-pooled curve partly relieves this, which is one reason it sits above the single-season curve at the same n.
What this does not show. Reliability is not accuracy or value. A stat can be highly reliable and still be a poor description of a player's contribution. Nothing here adjusts for shot difficulty, defender proximity, or whether an attempt was assisted, so “skill” here means the stable part of a player's own percentage, not isolated shooting talent. The 0.50 threshold is a convention, not a natural boundary. Splitting attempts at random also treats a season as one population, which deliberately ignores real within-season changes such as injury or a mechanical adjustment.