All scores →
Loading…
Loading…
Loading…

Small samples in basketball: when a hot streak is noise

By 4Basketball Staff · · 11 min read

A hot week or a cold month often says less than it seems. Learn how small samples fool us, how big a sample you need for common stats, and a simple way to shrink extreme numbers toward reality.

A player starts the season making nearly every three-pointer he tries. After eight games he is shooting 52 percent from deep, and the commentary writes itself: he has changed his routine, he has found his rhythm, he is a new kind of threat. A month later, his percentage is back near where it has always been, and the commentary quietly moves on. A team opens the season 9-2, and it is pronounced a contender. By the holidays, it is an ordinary team with a good start. A rookie scores 30 in his first start, and he is the next big thing, until he scores eight a night for the rest of the year.

These episodes have something in common. Each takes a number drawn from a small sample, a limited set of shots, games or minutes, and treats it as a measure of ability. In basketball, as in any sport, small samples are the largest single source of false conclusions. Randomness is built into the game: shots that are the same in every measurable way go in sometimes and miss sometimes, and over a few dozen attempts the luck has not had time to even out.

This article explains how small samples mislead, how to estimate how much a number should be trusted, which statistics need how much data, what the famous hot-hand debate teaches, and a simple technique for pulling extreme results back toward what is likely to be true. It is the foundation for several other pieces on this site, including regression to the mean and clutch statistics.

Why randomness takes so long to settle

Flip a fair coin ten times and you may well get seven heads. Flip it a thousand times and you will get close to 500. The difference is not that the coin changed; it is that the more flips you make, the more the random ups and downs cancel. This is the law of large numbers, and it applies to anything that has a random component, including a jump shot.

A shooter's true ability is the probability that a given shot goes in. If we could watch him take a million identical shots, we would learn it precisely. We cannot, so we see a few hundred shots a season and have to estimate. The key quantity is the standard error, which describes how far the observed percentage is likely to stray from the true one by chance alone. For a shooting percentage p taken over n attempts, the standard error is the square root of p times (1 minus p) divided by n, all under one root sign.

Take a player whose true three-point percentage is 36 percent. Here is how much his observed number would typically vary, depending on the number of attempts.

Attempts Standard error Typical range (about 95 percent of the time)
10 about 15 points roughly 6 to 66 percent
50 about 7 points roughly 22 to 50 percent
100 about 5 points roughly 26 to 46 percent
500 about 2 points roughly 32 to 40 percent

On ten attempts, a 36 percent shooter can look like a 66 percent sniper or a hopeless case, with no change in skill whatsoever. Even at 100 attempts, a span of twenty percentage points is within the realm of chance. Only at several hundred attempts does the picture sharpen, and even then it has a spread of a few points. When you hear that a player is shooting 45 percent from three after 20 games and have in mind that his career figure is 36, the sensible reading is that most of the gap is luck.

How much data different statistics need

Not all statistics are equally noisy. Some settle quickly because they depend on a player's role and physical traits, which are consistent from game to game. Others take most of a season, or more, to stabilize because they depend on rare, high-variance events.

  • Quick to stabilize: rebounding rate, blocks and steals per minute, usage, and free-throw attempt rate. These reflect position, role and style, and they show up reliably within a few weeks.
  • Slow to stabilize: three-point percentage, which takes hundreds of attempts, and opponent three-point percentage on defense, which can take more than a season. Shooting from a particular zone is even slower.
  • In between: overall field-goal percentage, free-throw percentage and assist rate. Free throws stabilize faster than threes because they are taken from the same spot with no defender, so there is less variation in what counts as a typical attempt.

The practical lesson is to match your trust to the statistic. Three weeks into a season, it is reasonable to believe a player's rebounding rate, but not his three-point percentage. A coach who sees a player getting a larger share of the shots is observing a real change in role; a fan who sees his shooting percentage climb is, probably, observing a short run of luck.

You can read more about role-driven numbers in usage rate and rebounding rates, and about how to evaluate extreme figures on leaderboards in how to read a leaderboard: whose numbers are meaningful.

The hot hand, and what it teaches

The most famous small-sample argument in basketball is about the hot hand: the belief that a player who has made several shots in a row is more likely than usual to make the next one. Nearly every player and fan believes it. In 1985 a team of psychologists, Thomas Gilovich, Robert Vallone and Amos Tversky, examined shot sequences from professional games and found no evidence of it; a player's chance of making a shot did not appear to depend on whether the previous shot went in. The finding became a standard example of how people see patterns in randomness.

The story did not end there. Decades later, other researchers, most prominently Joshua Miller and Adam Sanjurjo, identified a subtle statistical bias in the way the original analysis was done, which had tended to understate any real streakiness. After correcting for it, they argued that the original data in fact showed modest evidence of a hot hand. Subsequent studies, with larger data sets, have found small effects in some settings and little in others. The current state of the debate is that a hot hand may exist but is much smaller than players and fans believe, and it is hard to detect in a typical stretch of games.

For the purposes of reading statistics, the debate offers two lessons. First, streaks happen in random sequences far more often than intuition expects. In 100 coin flips, the longest run of heads or tails is typically six or seven. A 40 percent three-point shooter who takes 500 attempts in a season will, by chance alone, produce several runs of five straight makes; the expected number is about three. Those runs feel significant when they happen, but they are what randomness looks like.

Second, even if a hot hand exists, the evidence that shows it is hidden in large data sets. A single streak in a single game does not demonstrate anything. The correct stance is not to deny that players feel hot, but to avoid the leap from a streak to a change in ability.

Teams, records and the early season

Small samples also distort team evaluation. Over a short stretch, the standings reflect schedule, injuries and close-game luck as much as quality. A team that starts 9-2 has played eleven games, a number so small that even a team with a true winning percentage of .550 will reach 9-2 by chance more often than people expect over a league of thirty teams. Ten or fifteen games into the season, margin and net rating are better guides than record, but they too carry large uncertainty. As we explain in margin of victory, the uncertainty in an average margin after ten games is nearly four points, which is the difference between an average team and a contender.

The shorter the season, the bigger the problem. In leagues with fewer games, such as the WNBA or international tournaments, a team's record is drawn from a smaller pool of outcomes, and the randomness occupies a larger share of the standings. A single injury or a handful of tight games can alter a team's finish more than in an 82-game schedule. Sensible readers of those leagues give extra weight to margin and to the quality of the roster, and less to the win column.

A simple way to correct for small samples

You do not need advanced mathematics to treat small samples sensibly. The core technique is to shrink an observed number toward a sensible baseline, in proportion to how little data you have.

The easiest version is to add "ghost" attempts at the average rate. Suppose a player has made 5 of his first 6 three-point attempts, an 83 percent rate. The league average is about 36 percent. Imagine adding 100 attempts at that average: 36 makes out of 100. His adjusted record becomes 41 makes in 106 attempts, or about 39 percent. That is a far more reasonable guess about his true ability than 83 percent, and it moves toward the raw number as he accumulates real attempts. After 300 real attempts, the ghost 100 carry much less weight.

The number of ghost attempts you add depends on how noisy the statistic is. For three-point shooting, adding a hundred or so is sensible. For a statistic that stabilizes faster, you would add fewer. For something even noisier, you would add more. The choice is a judgment, but any reasonable choice gets you closer to the truth than taking the raw number at face value.

You can also shrink toward something better than the league average. A player who has shot 38 percent on threes for five seasons and starts the year 5 for 20 should be pulled toward 38 percent, not toward the league mean, since his past record is information. The approach is the same one that underlies most good forecasting: combine what you knew before with what you have just seen, weighting each by how reliable it is.

Small samples hide in other places

Shooting percentages are the obvious case, but small samples turn up in many less obvious corners of basketball.

Per-minute and per-36 numbers. Scaling a reserve's production up to a full starter's workload makes a handful of minutes look like a season. A player who scores 12 points in 14 minutes in a few appearances may project to 30 per 36 minutes, and the projection is nearly worthless, because those minutes came against particular opponents, in particular game situations, and in a sample too small to separate skill from circumstance. The projection also assumes that his efficiency would survive a bigger role and tougher defenders, which is rarely the case. Our look at usage rate explains why volume and efficiency trade off.

Rookies and prospects. A college freshman's shooting percentage after 15 games, or an overseas prospect's numbers in a short tournament, carries wide uncertainty. Scouts lean on mechanics, free-throw shooting and the quality of the shots as supporting evidence precisely because the raw percentages are noisy. We cover this approach in how draft prospects are evaluated with data.

Matchup and head-to-head records. A player who scored 35 in two games against a particular defender has told you very little about the matchup. The same goes for a team's record against a particular opponent over a few games, or a coach's record in game sevens. These figures make memorable talking points and weak predictions.

Defensive statistics. Opponent shooting on a defender's contests is built on far fewer shots than the defender's own shooting is, which makes it noisy. The difficulty is a major theme of our article on defensive metrics.

In each case the remedy is the same: widen the range around the number, look for supporting evidence, and let the sample grow before you decide.

Signs that a change is real

It would be wrong to conclude that nothing in a short sample is ever meaningful. Some changes are real, and a few clues help to separate them from noise.

  1. There is a visible reason. A new role, a changed shot, a new position, a coaching change or a healthier body can all alter performance. A change with a plausible cause is more credible than one without.
  2. Other statistics move too. A player whose three-point percentage rises along with his free-throw percentage, and whose shots look cleaner, is more believable than one whose only change is the raw three-point number. Free-throw percentage is, in fact, often a better indicator of shooting touch than a short run of threes.
  3. The inputs changed. If a player is taking more open shots, or getting more shots at the rim, the improvement in his scoring has a mechanism. Our piece on shot charts shows how to check.
  4. The sample keeps growing. A change that persists across hundreds of attempts and several months is far more convincing than one that appears in a single week.
  5. The change is in a stable statistic. A jump in a statistic like rebound rate or usage in the first few weeks is likely to be real, whereas a jump in three-point percentage is likely to be noise.

What to do as a fan

A few habits help in everyday viewing.

  • Ask how many attempts or games a number rests on, and mentally widen your range accordingly.
  • Be skeptical of leaderboards that include players with few minutes; the top and bottom of a leaderboard are where small samples live. The leaders page is more informative when you look at players who have played a meaningful amount.
  • When a player or team has an extreme stretch, ask whether the cause is a change in inputs or a change in outcomes. Inputs are more durable.
  • Use the language of probability. "He is probably a good shooter who has been hot" is more accurate than "he has become a great shooter."
  • Give streaks their due as entertainment. The fact that a hot run is mostly randomness does not make it less fun to watch.

What to take away

Small samples are the main reason basketball numbers fool us. Randomness is large over dozens of shots and games and only fades over hundreds. Statistics differ widely in how fast they settle, so match your trust to the stat and the sample. Streaks are a normal feature of random sequences, and the hot-hand debate suggests any real effect is much smaller than perception. When a number looks extreme and the sample is small, shrink it toward a reasonable baseline, look for a cause, and wait for more data. The patient fan is wrong less often, and is better prepared when a change turns out to be real. For a look at the players behind the numbers, browse the players page.

Search more on this

Keep reading

Elo ratings in basketball: the idea and its uses

Elo began as a chess rating and now powers many basketball forecasts. This article explains the formula with a worked example, the adjustments basketball needs, and what an Elo rating can and cannot tell you.

4Basketball Staff · Jun 4, 2026 · 12 min read

Expected points and shot quality: tracking-era basketball

Tracking cameras let analysts estimate how likely each shot was to go in. This guide explains expected points, shot quality versus shot making, and the limits of the models behind the numbers.

4Basketball Staff · Jun 27, 2026 · 11 min read