Clustering Illusion: Why Random Data Looks Like a Pattern

Randomness is clumpier than it looks from the inside. If each event is independent, nothing prevents five of the same outcome in a row, and over a long enough record such runs are certain to appear. What people expect instead is that a fair process will keep its proportions visible in every stretch of the record, so a sequence that holds 50-50 locally feels random while a sequence containing a run of seven feels designed.

Kahneman and Tversky named this expectation local representativeness in 1972. People treat a small sample as if it should mirror the process that generated it, which makes genuinely random data look patterned and makes smoothed, alternating data look random. They demonstrated the same error with a distribution rather than a sequence: asked whether 20 marbles handed out among five children would more often come out 4, 4, 5, 4, 3 or 4, 4, 4, 4, 4, a significant majority, 36 of 52 participants, chose the uneven split, although the even one is the more probable single outcome.

The arithmetic of runs is where the intuition goes wrong by the largest margin. In 40 flips of a fair coin the longest run of one side averages 5.7, and some run of five or more appears in roughly three quarters of all sequences. A sequence whose longest run is three or less, which is what most people produce when asked to fake randomness, occurs in about 4 percent of real sequences. So the usual test, looking for a stretch that seems too long to be chance, flags the real data and clears the fake.

The same mistake runs through spatial data. Hits scattered at random over a grid leave some cells empty and some with several, and the empty-and-crowded appearance is what prompts the search for a cause: a target, a contaminated site, a lucky shooter. Clarke's 1946 count of flying bomb hits across south London is the standard example because the tally matched the Poisson prediction closely. Where a cluster does have a cause, the way to find out is to state the hypothesis and the boundaries before looking at the map, since boundaries drawn around a cluster that has already been spotted will always confirm it.

Thomas Gilovich, Robert Vallone and Amos Tversky published "The hot hand in basketball: On the misperception of random sequences" in Cognitive Psychology in 1985, and Tversky and Gilovich summarized it for CHANCE in 1989. The paper has two halves: what fans believe about streaks, and what the shooting records show.

For the belief half, 100 knowledgeable basketball fans were surveyed and then asked to judge sequences. In the survey, 91 percent agreed that a player has a better chance of making a shot after having just made two or three in a row. In the judgment task, each respondent saw sequences of 21 hits and misses, said to come from a player's record, and classified each as streak shooting, chance shooting or alternating shooting. All the sequences contained 11 hits and 10 misses, and they differed only in how often the outcome switched: from 40 percent of the time up to 90 percent. The sequence that switched half the time, which is what independence produces, was classified as streak shooting by 62 percent of respondents in the 1985 paper and by 65 percent in the authors' 1989 summary of the same study, and the sequences chosen as the best examples of chance shooting were those switching 70 and 80 percent of the time.

For the record half, the authors analyzed field goals by the Philadelphia 76ers, free throws by the Boston Celtics, and a controlled experiment in which Cornell varsity players each took 100 shots. In the 76ers data, for eight of the nine players the probability of a hit was slightly higher after a miss, averaging .54, than after a hit, averaging .51, the reverse of what streak shooting predicts.

That second half has since been overturned, and a page about this illusion should say so. Miller and Sanjurjo showed in 2018 that the measure the paper used carries a selection bias: counting what follows a streak of hits in a finite sequence understates the true probability. In the controlled Cornell data the bias flips the result: the +3 percentage point difference the original authors read as nothing becomes, once the bias is computed from each player's own shooting rate, about +13 percentage points in favor of hot hand shooting. The belief half stands untouched: people really do classify random sequences as streaky. What no longer stands is the conclusion that basketball streaks are nothing but that illusion.

The illusion needs sparse data and a reason to look for a cause. The rows below are illustrations, not findings from the studies cited on this page.

The countermeasures are arithmetic, because the intuition is the thing that is wrong.

• Work out what chance predicts before judging the data. For runs, the longest stretch in n flips lands near the logarithm of n in base two; in 40 flips that is about five or six, not two.
• Ask for the denominator. A cluster of four in one place is only surprising next to how many places there were, how long the window was and how many comparable clusters would be expected somewhere.
• State the boundaries first. A region drawn around a cluster already noticed will always look significant, which is why cluster investigations set the area and the period in advance.
• Count the misses. Streak stories are built from the hits that got attention; the record has to include the stretches nobody mentioned.
• Simulate when you can. A spreadsheet that generates a few hundred random sequences shows directly how often the pattern you are worried about appears by chance.
• Keep the opposite error in view. Miller and Sanjurjo's correction shows that dismissing every streak as an illusion is its own mistake; the point is to test streaks, not to assume them in either direction.

Three nearby biases all concern streaks, and they differ in what the observer concludes next.

• It is not the gambler's fallacy, which takes a run as a sign that the other outcome is now due. The clustering illusion is the earlier step: deciding that the run means something at all.
• It is not the hot hand fallacy, which takes a run as a sign that the same outcome will continue. It draws on the same misreading of chance, and since 2018 the evidence that basketball streaks are purely illusory has weakened.
• It is not illusory correlation, the perception of a relationship between two variables that are unrelated. The clustering illusion concerns the shape of a single stream of data.
• It is not pareidolia, in which the perceptual system finds a face or a figure in noise. That one happens in perception; the clustering illusion is a judgment about frequency and cause.