Hot Hand Fallacy

Interpreting a Hot Hand

The hot hand fallacy is an unwarranted inference that recent successes make another success more likely. For independent events, the earlier results do not change the next event’s probability. For human performance, independence needs to be investigated rather than assumed. This belief is especially prominent in sports, where players, coaches, and fans often think that a basketball player who has made several consecutive shots is “hot” and more likely to make the next one. The studies discussed here compare that belief with shooting data and examine limitations of the original tests. They do not establish that every apparent streak is an illusion, or that a genuine performance change must be negligible.

The Widespread Belief in Streak Shooting

A survey of 100 basketball fans (players and spectators from Cornell and Stanford, including intramural team captains) revealed a near‑universal belief in sequential dependencies among shots. Overall, 91% of respondents believed that a player has a better chance of making a shot after having just made his last two or three shots than after having just missed them. The same belief extended to free throws: 68% thought that a player has a better chance of making his second free throw after making the first than after missing the first. In numerical terms, when asked about a hypothetical 50% field‑goal shooter, the average estimate of his shooting percentage rose to 61% after a made shot and fell to 42% after a missed shot. For a hypothetical 70% free‑throw shooter, the estimated percentages for the second free throw were 74% after a make and 66% after a miss. The belief also guides strategic decisions: 84% of fans agreed that it is important to pass the ball to someone who has just made several shots in a row, and 96% thought that players who have made a series of shots tend to take more shots than they normally would.

Tests of Sequential Dependence

Field goals. The foundational study by Gilovich, Vallone, and Tversky (GVT) found almost no evidence of streak shooting. Conditional probabilities computed for Cornell University varsity basketball players showed that, for most players, the likelihood of making a shot was not higher after a hit than after a miss—in fact, serial correlations were often slightly negative. Only one player showed a statistically significant positive correlation, and the overall pattern was consistent with a binomial model that assumes a constant hit rate. The same researchers also tested whether players experienced more “hot” or “cold” nights than expected by chance. Using the Lexis ratio—the ratio of the observed standard error of per‑game percentages to the standard error expected under independence—on data from the Philadelphia 76ers, they found ratios ranging from 0.56 to 1.03 (mean 0.84) for the seven players who played at least 10 games with at least 10 shots. None of the ratios was significantly greater than 1, so this analysis did not detect game‑to‑game variation beyond its chance baseline. In a further analysis, GVT categorized every possible sequence of four shots from a 50% shooter as good (3–4 hits), moderate (2 hits), or bad (0–1 hits). Under pure chance, each sequence is equally likely; the actual distribution of sequences for the professional players did not deviate significantly from that baseline.

Free throws. Free‑throw data from the Boston Celtics over two seasons were examined to test for dependency between the first and second shots of a pair. Free throws are ideal for such a test because they are taken from the same location without defensive pressure, removing shot‑selection and defensive confounds. Across nine players, the correlation between the outcome of the first and second free throw was positive for four players and negative for five, and none were significantly different from zero. This analysis did not detect a reliable increase after a made first free throw. A nonsignificant result does not establish independence.

Predictability. A betting experiment conducted with the Cornell players tested whether they could predict their own hits and misses. Before each shot, the shooter and an observing teammate independently chose to bet high (risking more for a higher reward) when confident and low otherwise. The average correlation between shooters’ bets and their actual performance was only .02, and only 5 of 26 individual correlations reached significance—four of which were modestly positive (.20–.22) and one was negative (–.51). Observers fared no better (mean correlation .04). In contrast, both shooters’ and observers’ bets were strongly correlated with the outcome of the previous shot (mean r = .40 for shooters, .42 for observers). This pattern is consistent with the hot‑hand belief, but the bets showed little predictive value in this experiment. The agreement between shooters’ and observers’ bets disappeared once the influence of the previous shot was removed.

Methodological Critiques and Evidence for a Subtle Hot Hand

Later studies discussed here question whether the original analyses could reliably detect a hot hand. Korb and Stillwell (2003) simulated a player with a genuine hot hand—bursts during which the shooting percentage temporarily spiked to 90%—and found that the significance test used by GVT failed to detect the hot hand in most simulations. The test was underpowered, meaning that even a real streak‑based boost might not have been flagged as statistically significant by the original methods. Huizinga and Weil (2009) identified a behavioral confound: after making a shot, players are more likely to attempt a more difficult shot (e.g., a longer jumper or a contested attempt), which could wash out a genuine small improvement in shooting ability. If being “hot” causes both a slight boost in accuracy and a tendency to take riskier shots, the observable hit rate might remain unchanged. Attali (2013) found that after making a three‑pointer, players were indeed more likely to attempt a longer shot, a shift not observed after easier baskets like layups, consistent with overconfidence triggered by perceived hotness rather than a true performance lift.

These critiques show why an unchanged observed hit rate may not settle the question. Test power, shot difficulty, and defensive responses affect what an analysis can detect. The size and duration of any performance change must be estimated from suitable data, rather than inferred from a few memorable baskets.

The Cognitive Roots of the Illusion

Why do short streaks feel so persuasive? One possibility is a memory bias: long runs of hits or misses are more memorable than alternating sequences, leading observers to overestimate sequential dependencies when recalling events. However, the phenomenon can be produced even when the data are fully displayed and do not rely on memory.

The study account describes 100 basketball fans judging six sequences of 11 hits and 11 misses with different probabilities of alternation—the probability that a shot differs from the previous one. The short sequence example below reports a different composition, so the exact sequence length is inconsistent within this article. A sequence with a .5 alternation rate—matching the expected rate for independent fair coin tosses—was classified as “chance shooting” by only 32% of fans; 62% identified it as an example of “streak shooting.” Instead, fans overwhelmingly selected sequences with higher alternation probabilities (.7 or .8) as the best examples of chance. In other words, people expect random sequences to contain far more alternations than they actually do, and they perceive the genuine null pattern as streaky. This also means that negatively correlated (alternating) sequences are often misperceived as random.

This misperception stems from a broader cognitive bias: the representativeness heuristic. People believe that even short sequences should be highly representative of the generating process, and therefore expect local balance—many alternations—in a random sequence. When a sequence instead shows the clustering that naturally emerges from independence, it is seen as patterned and non‑random. Passive exposure alone may not correct that expectation. Comparing observed runs with a suitable chance model gives viewers a more useful baseline.

Why the Illusion Is Especially Compelling in Basketball

Although the misperception of randomness is a domain‑general phenomenon, basketball provides several additional layers of reinforcement. Plausible causal mechanisms for streakiness—such as confidence, fatigue, or being “in the zone”—readily come to mind, making the belief feel explanatory. Observations of a player’s hustle, defense, or passing may be overgeneralized into a perception that the player’s shooting is also hot. Even the ambiguous outcome of a ball that rolls around the rim before dropping or popping out can be coded in line with prior expectations: if the player made the previous shot it is seen as a continuation of being hot; if the previous shot was missed, it confirms a cold spell.

These factors may strengthen the illusion, but they are not necessary for it to occur. Expectations about randomness can influence both casual viewers and experts.

Consequences of the Hot‑Hand Belief

An exaggerated belief in a hot hand can affect strategy. Passing the ball to a player perceived as “hot” is a standard tactic endorsed by players and coaches. Opposing defenses, sharing the same belief, will concentrate on guarding that player. If another equally skilled player is left less guarded as a result, that player may actually have a better chance of scoring. Thus, the hot‑hand fallacy can lead directly to suboptimal decisions on the court. A similar inference can arise when a mutual fund that beats the market five years in a row is perceived as skillfully “hot,” or a gambler raises a bet after several wins. A streak alone does not identify skill or persistence. Ask whether underlying conditions have changed, whether comparable streaks occur under a suitable baseline, and whether the pattern predicts new outcomes.

A constant-probability model is a useful baseline, but actual shots differ in distance, defense, fatigue, and other conditions. Assess whether a streak adds predictive information after accounting for those differences, and how uncertain the estimate remains.