What Is the Hot-Hand Belief?
The hot-hand belief is the expectation that a person who has recently succeeded is more likely to succeed again because the successes indicate a temporarily elevated state of performance. In basketball, observers may judge that a shooter who has made several consecutive shots has become more likely to make the next one. Similar judgments appear in discussions of athletes, traders, performers, students, and teams. The central scientific question is whether a run of successful outcomes provides information about the conditional probability of future success after accounting for ordinary chance, stable ability, task difficulty, strategic adjustments, and selection effects.
The term hot-hand fallacy refers specifically to an unwarranted inference from a streak to a subsequent success. The expression is sometimes used as though any belief in streak-dependent performance were mistaken. Contemporary evidence makes that blanket interpretation untenable. Some environments show statistically detectable, heterogeneous short-term dependence, while others show weak, null, or negative associations. The existence of a real hot hand in a given task and the accuracy of a particular observer's belief about that hot hand are two separate empirical questions. Both must be tested against an explicit statistical model rather than decided by the mere appearance of consecutive wins or misses. Gilovich, Vallone, and Tversky (1985)90010-6); Miller and Sanjurjo (2024).
Streaks, performance states, and beliefs
A streak is an observed sequence of outcomes, such as four successful shots in a row. A performance state is a hypothesized, potentially changing characteristic of the process generating those outcomes: concentration, motor coordination, fatigue, confidence, a favorable tactical matchup, or some combination. A hot-hand belief is a judgment about that state and its consequences. These concepts can diverge. A long winning streak can occur with constant success probability, a genuine improvement in success probability can produce no visibly long streak, and a person can correctly notice an improvement but overestimate its size or duration.
In a simple independent Bernoulli model, each trial has success probability p and outcomes are independent conditional on p. The probability of three specified consecutive successes is p cubed. If p equals 0.50, that probability is 0.125; if p equals 0.70, it is 0.343. Such sequences are normal consequences of independent trials. The observation of three successes does not by itself show that p increased. Yet knowledge about the performer's ability, changing circumstances, and how the sequence was selected can make recent outcomes informative. An investigator therefore asks not simply whether a streak occurred, but what competing data-generating processes predict about the next trial.
Scientific and clinical status
Hot-hand belief is a topic in cognitive psychology, behavioral economics, statistical inference, and performance science. It is not a psychiatric diagnosis or a clinical symptom by definition. It becomes a cognitive bias only when the observer's judgment systematically departs from an appropriate benchmark in a specified task. A positive conditional dependence in the data is itself a statistical property, while confidence that such dependence exists is a psychological judgment. Neither a statistical test result nor a personal feeling of momentum automatically identifies an enduring trait or mental disorder.
Why the Debate Matters
The dispute over the hot hand has an unusually instructive history because an influential demonstration of apparent irrationality helped establish the heuristics-and-biases approach, but subsequent statistical work showed that an important conventional test for the effect was itself biased. The development illustrates how cognitive psychology depends on sound measurement, explicit null models, and separation of intuitive explanations from mathematical inference. It also demonstrates the difficulty of moving from an observed sequence to a claim about a psychological mechanism.
The historical progression cannot be summarized as a simple reversal from false to true. The 1985 study provided evidence that observers expected continuation after success and did not find corresponding positive shot-to-shot dependence in its data. The 2018 correction identified a finite-sample problem that can hide genuine continuation effects. Subsequent studies sometimes detect hot-hand performance after correcting statistical bias, while others find that the effect varies with shooting context, location, defensive behavior, or individual player. Each conclusion concerns a particular sample, estimator, and definition of success. Gilovich et al. (1985)90010-6); Miller and Sanjurjo (2018); Pelechrinis and Winston (2022).
The Original Basketball Experiments
Gilovich, Vallone, and Tversky in 1985
Thomas Gilovich, Robert Vallone, and Amos Tversky investigated common beliefs about streak shooting in a landmark study published in Cognitive Psychology. Their evidence included analyses of Philadelphia 76ers field goals, Boston Celtics free throws, and a controlled shooting experiment involving Cornell varsity basketball players. The observers and players they studied generally expected recent makes to improve the prospect of another make. In the data analyses reported by the authors, success on a previous shot did not reliably predict success on the next shot; the controlled experiment similarly separated players' expectations from measured performance. The authors proposed that observers misperceive the clustering that naturally occurs in random sequences. Gilovich et al. (1985)90010-6).
This result shaped an influential interpretation: people expect random sequences to display too much local alternation and therefore regard naturally occurring clusters as extraordinary. The hypothesis is psychologically meaningful, especially when a person treats a few successes as definitive evidence of a newly changed ability. The study nevertheless estimated sequential dependence using finite observed records, and the absence of a positive conventional estimate is not equivalent to proof that every underlying performance process is independent.
What the original analysis established—and what it left open
The classic evidence remains valuable for documenting beliefs and for raising the possibility that observers overinterpret runs. Its broader implication was constrained by the statistical properties of the chosen comparisons. Real-game shots may differ in difficulty, position, fatigue, defensive attention, and teammate behavior. Even a controlled task can generate limited sequences that make rare conditioning events difficult to estimate. A null result may reflect genuinely absent persistence, offsetting forces, inadequate power, a flawed estimator, or some combination. Evaluating these possibilities requires alternative designs and estimators rather than a categorical inference from a single study.
The original authors addressed an important descriptive question: how do people reason about apparent streaks? The later literature added a separate inferential question: how should researchers estimate the difference between success after a streak and success under other conditions? Confusing the two questions allows a biased statistical procedure to masquerade as a demonstration of cognitive bias.
The Finite-Sequence Selection Bias
Why the ordinary conditional comparison can mislead
A common hot-hand statistic subtracts the hit rate after misses from the hit rate after hits within a shooter's observed sequence. An intuitive analysis assumes that, for independent and identically distributed trials, the expected value of this difference is zero. Miller and Sanjurjo demonstrated that this expectation can fail in finite sequences when the calculation conditions on the appearance of particular previous outcomes and reuses the same observed sequence for estimation. Among eligible positions, the selection rule induces a systematic downward bias. The effect can be substantial when sequences are short or when the conditioning streak is long. Miller and Sanjurjo (2018).
The issue is not the ordinary fact that a fair coin has a 50 percent probability of heads on its next independent toss. It arises when an analyst selects positions inside finite sequences using observed preceding hits and then computes the average of within-sequence ratios. The denominator—the number of eligible shots—varies across sequences and is related to the same outcomes that determined eligibility. Averaging the resulting statistics does not automatically reproduce the naive independence benchmark. The relevant expectation is conditional on the estimator's selection and averaging procedure.
A concrete demonstration of the selection problem
Consider the sequence H, H, T, T, where H denotes a hit and T a miss. In this particular sequence, one trial follows an H and is an H, while the trial following the second H is a T. The hit fraction after H is therefore one-half. The trial after the first T is another T, so the hit fraction after T is zero; the observed difference equals one-half. Another independently generated sequence can yield the opposite result. The essential lesson emerges only when all admissible sequences and the analyst's conditioning rule are considered together, particularly the sequences in which an eligible denominator is zero and the behavior of ratios conditional on eligibility.
Miller and Sanjurjo's derivation establishes that in many commonly used finite-sample designs the average conventional hit-after-hit minus hit-after-miss estimator is negatively biased under independence. Accordingly, a measured difference near zero may conceal genuine positive dependence relative to the correctly computed null distribution. The direction and size of correction depend on sample length, hit probability, streak length, and the exact statistic. A researcher should use the appropriate exact or simulation-based finite-sample benchmark rather than mechanically adding an arbitrary correction. Miller and Sanjurjo (2018); Miller and Sanjurjo (2019).
Why this was a methodological turning point
The correction concerns selection in statistical analysis, not a proof that every athlete enters an enhanced physiological state after scoring. It changed what constitutes a meaningful null hypothesis for certain observational estimators. It also supplied an example of a broader scientific risk: apparently elementary conditional frequencies can behave unexpectedly when observations are selected by their own histories. A replication that repeats the same biased estimator without investigating its null expectation cannot decisively settle whether genuine streak dependence is absent.
Formal Models of Hot-Hand Performance
Independence and conditional success probability
Let Y(t) equal one for a success on trial t and zero for a failure. Under an independent model with fixed success probability p, P[Y(t)=1 given Y(t-1)=1] equals P[Y(t)=1 given Y(t-1)=0], and both equal p. An elementary descriptive contrast is Δ1 = P(hit next | hit now) − P(hit next | miss now). A positive population Δ1 means positive first-order serial dependence in observed outcomes. It does not uniquely identify a psychological performance state, because observed dependence can arise through changing shot selection, time trends, or other common causes.
For k preceding makes, researchers may define Δk as the probability of success after k consecutive hits minus an explicitly specified baseline, such as success after k consecutive misses, after non-hit histories, or the shooter's expected probability under a fitted model. These baselines are not interchangeable. Two studies may report different effect magnitudes because they use different conditioning sets, even when analyzing the same shots. The selected horizon—one next trial, one minute, a game, or multiple games—also changes the scientific question.
A latent-state model of performance
A plausible alternative to constant p is a latent-state model in which an unobserved performance variable fluctuates over time and influences success probability. The state might reflect concentration, automaticity, physiological arousal, fatigue, or an opponent's adjustment. A hidden Markov model, dynamic logistic model, or hierarchical state-space model can represent time-varying probabilities. These models distinguish stable differences across performers from short-term within-performer change, although they do not by themselves prove which psychological process causes the change.
Observed clustering can also result from exogenous conditions. Repeated easy attempts can produce runs of hits even if no success causes a later success. In sports, location, distance, shot type, defensive coverage, and playing time may affect success. A model that omits these variables may attribute environmental persistence to the performer's hot state. Conversely, a model that treats defensive responses caused by previous makes as ordinary independent confounders may remove part of the real strategic process by which a streak alters subsequent success.
Prediction is different from explanation
A variable is predictively useful when it improves out-of-sample probability forecasts. A causal interpretation additionally requires specifying how a prior success might change the subsequent probability, controlling appropriate alternative explanations, and examining experimental or quasi-experimental evidence. A streak may predict future success because it reveals a temporarily favorable condition without itself causing that condition. An observer can be statistically well calibrated about elevated performance while holding an incorrect causal story about confidence or momentum.
The practical benchmark for a predictive claim is a proper out-of-sample scoring rule or a predeclared difference in predictive performance, compared with a model that already includes base rates and relevant contextual predictors. For a causal claim, the benchmark depends on identification assumptions and interventions. Calling a pattern a hot hand should never replace this distinction.
Meta-Analytic Evidence and the Changing Research Picture
The 2013 sports meta-analysis
Avugos and colleagues reviewed sports studies published between 1985 and 2012. Their meta-analysis included 22 publications, 30 studies, and 56 independent effect sizes. They reported a small mean effect size of approximately 0.02 with p = .49 using a random-effects model and concluded that their evidence did not support a general hot-hand effect in sports. This was an important systematic summary of the earlier literature, not a demonstration that all forms of state dependence are impossible. The included work predated later attention to finite-sequence selection bias and combined multiple sports, designs, and estimands. Avugos et al. (2013).
The number 0.02 should not be interpreted as a universal two-percentage-point gain on a basketball shot. It is a mean effect size within that synthesis's reported measures. Differences between standardized effects, probabilities, and percentage-point changes matter. Statistical nonsignificance in a heterogeneous synthesis also is not the same as evidence that every relevant true effect equals exactly zero. The inferential weight of a synthesis depends on measurement compatibility, study quality, statistical power, and the theoretical plausibility of one common underlying effect.
Reanalyses and controlled shooting evidence
In their 2018 Econometrica paper, Miller and Sanjurjo derived the finite-sequence bias and reconsidered evidence from the canonical experiment. In a later analysis published in The Review of Economics and Statistics, they examined four controlled shooting experiments, multiple measurement sessions separated by months for some individuals, and improved approaches to testing sequential dependence. They reported strong evidence for hot-hand performance across datasets and within some individuals, along with evidence that expert observers could predict which shooters would become hottest. Miller and Sanjurjo (2018); Miller and Sanjurjo (2024).
Controlled shooting designs reduce variation in some factors that complicate game records. They also differ from real competition in defensive pressure, shot choice, incentives, feedback, and social interaction. Their results support the existence of persistent performance variation under specified controlled conditions and undermine the blanket assertion that belief in streak-dependent shooting is necessarily irrational. They do not quantify a single universal effect for all players, settings, or tasks.
Evidence from Elite Basketball
The NBA Three-Point Contest
Miller and Sanjurjo analyzed 34 years of televised NBA Three-Point Contest data from 1986 through 2020. Using approaches that addressed earlier statistical limitations, they found positive evidence of hot-hand shooting for some competitors and an average pooled increase on the order of five to nine percentage points while on a streak of hits, depending on the analysis. They also found considerable heterogeneity in both magnitude and direction across individuals and examined miss streaks separately. Miller and Sanjurjo (2021).
The contest is a valuable naturalistic-control compromise: shooters operate under meaningful performance pressure, but shots are taken in relatively consistent conditions without in-game defenders determining the next shot's difficulty. It remains important to account for the shooter's movement across shooting locations, the structure of the rack, and changes in timing. A percentage-point conditional difference observed in this event cannot be assumed to apply to ordinary possessions in basketball games.
Real games: both positive and negative results
Pelechrinis and Winston examined NBA player-tracking data from two seasons and applied a framework that considered expectations associated with shot quality. Among 153 players with at least 1,000 shots, they found statistically detectable positive streakiness for subsets of players, with reported positive adjusted differences among those classified as showing hot-hand evidence. At the same time, they reported that the league-wide average tended toward regression after consecutive makes. Their estimate among selected hot-hand players and the aggregate population estimate answer different questions and should not be collapsed into one result. Pelechrinis and Winston (2022).
The authors reported average adjusted hot-hand differences among the identified players ranging approximately from 1.5 percentage points for a one-hit conditioning window to 5.8 percentage points for four consecutive hits. They cautioned that longer-streak samples were smaller and selected-effect estimates could be inflated. This caveat matters whenever analysts highlight only performers whose estimated effects meet a significance threshold. Selection on a noisy positive estimate tends to exaggerate its size, especially when few opportunities follow rare streaks.
The importance of shot location and task similarity
A working-paper analysis by Lantis and Nesson of NBA Three-Point Contests suggested that apparent streak benefits were concentrated within the same shooting locations rather than carrying across locations. Their other NBA analysis distinguished some positive patterns in free throws from mixed or declining hit rates following field-goal streaks. These analyses provide specific hypotheses about motor continuity, contextual stability, and strategic response; as working papers they should be weighted according to subsequent peer-reviewed verification and methods rather than treated as universal conclusions. Lantis and Nesson (2021, working paper); Lantis and Nesson (2019, revised 2021 working paper).
A transition from an open catch-and-shoot opportunity to a defended off-balance attempt changes the task. The model must ask whether consecutive makes predict success at comparable opportunities or whether the appearance of momentum changes what opportunities are offered. In a competitive environment both processes can operate simultaneously. The sign of observed serial dependence can thus differ from the sign of underlying motor-performance persistence.
Defensive responses and social feedback
A 2026 analysis by Kondur and Shen examined how professional basketball defenders respond spatially to recent offensive shooting history. Their focus on defensive adaptation highlights an important causal pathway: past makes can alter defenders' attention and position, changing the difficulty of subsequent shots. This pathway makes it difficult to infer a shooter's intrinsic performance state from shot outcomes alone. Kondur and Shen (2026).
If defenders increase pressure on a perceived hot shooter, raw hit rates can remain flat even when the shooter has improved, or can fall because higher-quality defenders or harder attempts offset a temporary advantage. If teammates route more passes toward the same player, the number and difficulty of attempts can change at the same time. Statistical adjustment should be justified against a causal diagram: controlling for a variable that lies downstream of previous makes may answer a different question from estimating the overall effect of the streak on the next outcome.
Psychological Mechanisms Behind Hot-Hand Beliefs
Pattern perception and the law of small numbers
People are capable of detecting temporal patterns in noisy environments. That capacity supports learning but can produce false positives when a short sequence is considered unusually meaningful. Tversky and Kahneman's work on the law of small numbers described how people expect small samples to resemble population properties too closely. In a sequence that happens to contain several successes, an observer may infer a stable performance state from fewer observations than the relevant base rate and variance justify. Tversky and Kahneman (1971).
A belief about the streak can remain reasonable when the environment contains persistent states. The question is whether the observer's assumed degree of state persistence matches the actual data. Treating every cluster as evidence for a hidden cause encourages overinterpretation; treating every cluster as meaningless randomness risks ignoring genuine information. The appropriate response is calibrated updating, with uncertainty that reflects the amount of data and the prior prevalence of meaningful fluctuations.
Selective attention, memory, and narrative formation
Spectacular streaks are memorable, frequently replayed, and easy to summarize as coherent narratives. Ordinary mixed sequences attract less attention. Attention and memory may therefore enrich the subjective sample of dramatic performances, making rare runs feel common. The observer who recalls a celebrated consecutive scoring sequence may fail to count comparable opportunities that ended in misses or to consider how many players and games produced opportunities for unusually long runs.
Narrative explanations can be persuasive because they give an observed run a causal structure: confidence builds, rhythm improves, the environment responds, and success follows. Some of these mechanisms are plausible and in particular tasks may operate. The presence of a vivid narrative is not evidence that a specific mechanism has been isolated. A stronger inference uses prospective measurement of confidence, arousal, fatigue, movement, and task difficulty rather than reconstructing those variables from the sequence after the fact.
Confidence, motivation, and motor processes
Repeated success may influence confidence and perceived control, and changes in confidence can alter attention, risk taking, movement, or persistence. Improved automatic motor coordination is another potential mechanism; conversely, conscious monitoring, pressure, or fatigue may interrupt a favorable run. These mechanisms predict different patterns across task types and timescales. A short positive run in a repetitive motor skill may be more informative than the same count of successes spread across varied tasks and opponents.
Researchers should avoid identifying confidence as the cause merely because a successful performer later reports feeling confident. Confidence can be a consequence of earlier hits, a predictor of subsequent behavior, a correlate of a favorable hidden state, or a retrospective reconstruction. Tests of mediation require temporally ordered measures and assumptions about confounding; rigorous experiments manipulating feedback or task conditions can provide stronger causal evidence than stories about performance momentum.
Social inference and strategic adaptation
A hot-hand belief is often distributed across a team or audience. Teammates may pass more frequently to a player viewed as hot, opponents may guard that player more closely, and observers may update predictions rapidly after salient outcomes. These actions can change the system whose behavior is being forecast. A judgment that was calibrated before tactical adaptation may become less calibrated afterward because the prediction itself alters opportunities, shot choice, and defensive pressure.
The feedback loop illustrates an important distinction between individual cognitive errors and equilibrium behavior in groups. Apparent absence of an offensive hot-hand effect in a game could reflect an effective defensive response, while persistent individual hot-hand performance can coexist with mistaken audience estimates of its magnitude. Studies across sports and competitive contexts differ in how much strategic adjustment is available; those structural differences constrain generalization.
Hot Hand Versus Gambler's Fallacy
The gambler's fallacy is the expectation that an outcome is due to reverse after a short run in a sequence that is actually independent: for example, thinking that tails has become more likely because a fair coin landed heads several times. A hot-hand expectation anticipates continuation of success. In the same independent and fixed-probability task, an unjustified expectation of continuation and an unjustified expectation of reversal both conflict with the independence benchmark. Yet in a process with a persistent hidden state, continuation can be evidence-based, and in a process with negative serial dependence, reversal can be evidence-based. The words fallacy and belief must therefore be attached to a specified model and task.
Rabin and Vayanos analyzed how errors about short-run randomness can generate both gamblers'-fallacy judgments and hot-hand inferences under different informational conditions. Their theoretical treatment helps explain why apparently opposite beliefs can arise in related contexts, without implying that every empirical report of hot-hand performance is a cognitive error. Rabin and Vayanos (2010).
Hot Hand Versus Regression to the Mean
Regression to the mean describes a statistical tendency for an extreme measurement to be followed by a less extreme measurement when repeated observations are imperfectly correlated and contain variability. A player scoring unusually well in one brief interval will often perform closer to that player's typical level in the next interval. This tendency does not show that success causes failure. It also does not rule out temporarily elevated performance: a genuinely hot state may persist while unusually extreme observed scores regress relative to their own extreme starting point.
Suppose a performer's typical success rate is 60 percent. In one short session the performer hits 90 percent of attempts. The next session can be closer to 60 percent because the first session contained favorable randomness, even if underlying skill never changed. Alternatively, a latent increase from 60 percent to 70 percent might raise expected success in subsequent attempts, while the observed 90 percent still regresses. Distinguishing these possibilities requires estimates of stable skill, transient state variability, and measurement noise; the pattern of scores alone is insufficient.
Distinguishing Heuristics, Biases, Preferences, and Statistical Errors
The intuitive rule 'feed the player who is scoring' can be a heuristic: a fast policy using recent success to guide allocation of opportunities. Its rationality depends on how strongly recent scores predict future outcomes compared with other available information, and on the consequences of the resulting actions. Calling the rule a bias requires showing a systematic departure from a specified decision criterion, such as expected team points given shot quality and possession strategy.
The finite-sequence selection bias is an error in a statistical estimator; it can occur even when the analyst's psychological judgment is correct. A logical fallacy is a defect in reasoning structure relative to logic, while hot-hand prediction is generally an uncertain empirical inference requiring probability and data. Preferences concern how outcomes are valued: a coach may knowingly choose a popular player to sustain morale despite a small reduction in expected scoring. Such a decision can be explained by broader utilities rather than incorrect probability estimation. Clinical symptoms are distinct again, and ordinary streak-based reasoning does not constitute one.
For broader conceptual context, see Cognitive Biases: Definition, Types, and Psychological Research, Heuristics in Psychology: How Mental Shortcuts Guide Judgment, and Availability Heuristic: How Recall Shapes Probability Judgments.
How Scientists Measure Hot-Hand Effects
Defining events before examining outcomes
A credible analysis specifies which trial counts as a hit, which events count as successive trials, what constitutes a streak, and how the target outcome is defined. In basketball, a made field goal, a successful free throw, and a shot from a fixed contest location are different tasks. A researcher should specify whether intervening possessions, substitutions, or elapsed minutes interrupt the sequence and whether a three-hit streak is overlapping with subsequent four-hit streaks. Altering such choices after inspecting the results creates researcher degrees of freedom and can magnify apparent effects.
The denominator deserves explicit attention. How many opportunities follow three consecutive makes? How many occur under the comparator condition? Are some players excluded because they never produce a qualifying streak? Conditioning on eligibility changes the population represented by the estimate. An informative article or research report states the number of eligible trials and sequences along with the estimated difference.
Exact null distributions and simulation
For short records, researchers can simulate independent sequences using each performer's estimated or prespecified probability and the same number of attempts. The simulation must apply the exact selection rule and test statistic used on real data. The resulting distribution shows whether the observed statistic exceeds what the finite-sample estimator would ordinarily produce even without serial dependence. Exact combinatorial calculations can accomplish the same task when feasible. These methods directly address the bias that arises from selecting streak-conditioned positions. Miller and Sanjurjo (2018).
There is an additional difficulty when baseline success rates are estimated from the same short sequences: uncertainty in those rates can affect inference. Hierarchical models can partially pool performers, while prespecified permutation or randomization tests can preserve appropriate aspects of the experimental design. A confidence interval or posterior interval should reflect uncertainty introduced by both sparse streak opportunities and between-person heterogeneity.
Accounting for shot difficulty and time-varying confounding
In game data, researchers may model pre-shot distance, location, shot type, defender separation, time remaining, game state, and identity of the shooter and defender. These factors determine the expected chance of making the next attempt. A successful analysis asks whether recent hits add predictive information beyond these variables and whether any estimated state dependence is robust across reasonable model specifications.
This adjustment must have a declared causal purpose. Some shot-difficulty variables are determined after the previous streak because defenders and teammates react to it. Conditioning on those variables estimates a direct effect under certain assumptions; omitting them can estimate a total observed process that includes strategic reactions. Neither choice is universally correct. A study should explain what estimand it targets, which variables lie before or after the prior successes, and what assumptions make its conclusions credible.
Power, multiple comparisons, and individual differences
Longer streaks provide fewer eligible observations than single-hit windows. Testing many players, streak thresholds, game segments, locations, and model specifications increases the probability of apparently notable findings by chance. Multiple-testing corrections, hierarchical estimation, preregistration, and independent validation protect against selectively reporting unusually positive results. Estimates for performers selected precisely because they scored as statistical outliers are likely to be exaggerated; this is a version of selection-induced estimation bias distinct from the specific finite-sequence bias in the conventional conditional-frequency statistic.
Individual effects can be positive, negligible, or negative, so a single population average can conceal practically important variation. Reliable claims about who has a hot hand require repeatability across sessions or held-out data. Evidence that a group-level average differs from zero does not establish that any given shooter is predictably hot at a particular moment.
What Counts as a Meaningful Effect Size?
A change from 45 percent to 50 percent hit probability is a five-percentage-point difference, a relative increase of approximately 11.1 percent, and a change of five expected made shots per 100 comparable attempts. These descriptions are mathematically different. A reported standardized mean effect of 0.02 is likewise not directly interchangeable with a five-percentage-point change in shooting percentage. Report the estimator, denominator, confidence interval, baseline probability, and relevant setting whenever a magnitude is cited.
Practical importance also depends on opportunity cost. If a player believed to be hot takes a difficult attempt rather than passing to a teammate with a substantially higher expected value, even a small genuine performance increase might not justify the choice. Conversely, if recent success meaningfully raises success probability under otherwise comparable circumstances, ignoring that information may lower expected performance. The relevant benchmark is the value of a strategy relative to the best feasible alternative, not a headline streak statistic viewed in isolation.
Cultural Differences, Expertise, and Generalizability
The hot-hand literature contains laboratory tasks, controlled shooting experiments, professional sports records, and studies of observers' beliefs. These contexts vary in training, incentives, feedback, cultural narratives, and the degree to which outcomes are genuinely independent. The available evidence does not justify a single estimate for prevalence, effect size, or belief accuracy across cultures and activities. When a study samples elite athletes, its results concern a selected population with unusual experience and motor expertise; they cannot simply be transferred to classroom tasks, workplace evaluations, or financial markets.
Expertise can help observers learn whether streaks predict performance, but experts can also hold salient conventional beliefs. The controlled-evidence finding that some expert observers anticipated which shooters would become hottest supports predictive discrimination in that design, not perfect calibration or general expertise across every environment. Cultural comparisons would require comparable tasks, language measures, baseline probabilities, and sampling strategies before apparent differences could be attributed to cultural learning rather than measurement. Miller and Sanjurjo (2024).
Beyond Sports: Work, Education, Finance, and Technology
Repeated performance in everyday life
A salesperson closing several deals, a musician performing exceptionally during a concert, or a student answering multiple questions correctly may be experiencing a favorable performance state or merely an ordinary run of successes. Recent outcomes can reveal information about preparation, task fit, or changing conditions. Their predictive value depends on whether the causes persist into subsequent decisions. Counting raw streaks without the number of attempts, changing task difficulty, and personal base rates invites mistakes.
Organizations sometimes reward or redirect resources after highly visible runs. Evidence-based evaluation compares each person's performance with task-adjusted expectations and considers uncertainty around small samples. Treating consecutive wins as independent confirmatory evidence overstates information when the observations share common causes, while rejecting all short-term signals can miss genuine improvements in skill or circumstances.
Financial markets and prediction
Financial prices and trading outcomes are influenced by strategic agents, changing volatility, heterogeneous information, and selection of visible success stories. A trader's recent wins may reflect skill, favorable risk exposure, or luck; a run alone does not separate them. The relevant analysis uses risk-adjusted out-of-sample returns, appropriate benchmarks, fees, drawdowns, and enough observations to distinguish persistent ability from variability. Evidence from basketball cannot establish a universal momentum or reversal strategy in markets. An intuitive belief about a winning streak should never be used as a substitute for independent financial risk assessment.
Human–AI decision support in the Artificial Era
AI systems can produce sequences of correct answers that persuade users to trust later outputs more strongly. A streak of accurate answers, however, can occur within a narrow easy domain and reveal little about future performance on a different task. Evaluating an AI assistant's reliability requires stratifying by task type and difficulty, accounting for correlated outputs, estimating uncertainty, and measuring prospective calibration. The same finite-sample and selection principles apply whenever a user highlights consecutive successful examples rather than a representative sample.
The psychology of treating a machine as an authority belongs to human judgment and trust calibration. It can be studied without assuming that apparent success or subjective fluency proves an AI system's consciousness. For a specialized treatment of trust and delegated judgment, see AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making.
Improving Streak-Based Judgments
Compare with the correct base rate
Begin with the probability of success under comparable conditions. A player who makes 70 percent of uncontested practice shots and a player who makes 35 percent of closely defended game shots will naturally produce different run lengths. A visible sequence becomes informative only after considering the ordinary frequency of such sequences for that performer and setting. Record attempts and conditions alongside successes, and resist interpreting a performance peak without the surrounding denominator.
Use a declared predictive model
Before making a decision, define what outcome will be predicted, what information may be used, how long a performance state could plausibly persist, and which baseline model must be beaten. Compare a recent-streak model with a model using stable ability and task difficulty alone. Evaluate both out of sample using a proper score such as log loss or the Brier score. A streak variable that fails to improve forecasts should receive little practical weight even when it makes a compelling story.
Keep uncertainty visible
After three or four consecutive successes, the effective sample for estimating an individual streak effect is often small. Present ranges of plausible success probabilities, explain when the model has encountered an unusual or poorly represented state, and make room for tactical adjustments. When performance estimates were chosen because they appeared unusually positive, account for selection bias and assess the effect on independent data.
Test strategies rather than slogans
In coaching or resource allocation, the appropriate decision may involve expected value, opportunity cost, and broader team goals. One can experimentally compare strategies that use recent success with alternatives based on stable ability and contextual information, using predetermined outcomes and sufficient sample size. A decision strategy that seems intuitive in a vivid anecdote is not established as optimal without comparative evidence. The general literature on cognitive debiasing can inform probability education and feedback, but demonstrations of improved task scores do not automatically establish long-term real-world transfer. See Cognitive Debiasing Strategies: Which Interventions Work?.
Research Limitations and Open Scientific Questions
A continuing difficulty is that studies do not share one definition of the hot hand. Some measure raw serial correlation; others estimate deviations from shot-quality models, latent state persistence, conditional probabilities after specified streak lengths, or individuals who pass selected significance thresholds. Their findings cannot be combined casually. Published results may also be sensitive to handling missing opportunities, nonstationarity, player exclusions, and multiple specifications. Earlier meta-analyses may therefore be informative about historical designs while needing reconsideration under newer estimators.
Causal mechanisms remain difficult to identify. Positive temporal dependence can arise from motor processes, psychological confidence, changing environmental conditions, tactical decisions, or data collection. A model that estimates serial dependence does not automatically distinguish these pathways. Better work combines preregistered sampling plans, repeated within-person measurement, context-rich observations, experimental manipulation where ethical and feasible, transparent code, and independent replications.
The most defensible scientific synthesis is conditional. Genuine streak-related performance variation has empirical support in some controlled shooting contexts and among subsets of athletes; broad extrapolations from hot-hand beliefs to every performer or setting are unsupported. The accuracy of an observer's belief is evaluated by whether their probability judgments track the relevant conditional probabilities, including their uncertainty. Future comparisons should prioritize shared estimands and measurement transparency, rather than attaching a universal yes-or-no label to a heterogeneous family of phenomena.
Frequently Asked Questions
Is the hot-hand fallacy real?
People sometimes infer too much from a short run of successes, which can constitute a judgment error under an appropriate independence model. Yet some controlled and field studies detect genuine positive performance dependence. The existence of a hot-hand belief, an error in that belief, and a real hot-hand performance effect must be assessed separately. Miller and Sanjurjo (2024).
Does scoring three shots in a row make the fourth more likely?
Under independent trials with constant success probability, no. In a task where ability or circumstances persist, the streak might provide evidence about the next success probability. The size of any prediction improvement depends on the performer, setting, and way the next shot is selected. Three makes alone do not establish the effect.
What is the difference between the hot hand and the gambler's fallacy?
The hot-hand expectation predicts continuation; the gambler's fallacy predicts reversal. In a truly independent process, both can be erroneous if they imply that recent outcomes changed fixed probabilities. In processes with time-varying conditions, either positive or negative dependence may be real, so the correct conclusion rests on evidence about the data-generating process.
Did the 1985 study prove there is no hot hand?
No. It reported no positive dependence under its analyses and offered an influential explanation of streak perception. Later work identified downward finite-sequence selection bias in a conventional statistic and reported positive dependence in several corrected or controlled analyses. The original findings remain historically important but do not settle all subsequent contexts. Gilovich et al. (1985)90010-6); Miller and Sanjurjo (2018).
Is a hot hand the same as momentum?
Momentum is a broad descriptive term for apparent continuation across time; a hot-hand hypothesis specifies measurable conditional performance dependence. The same pattern of successes may have different psychological, physiological, or strategic explanations. A convincing scientific account needs a testable mechanism and clearly defined time window, not the label momentum alone.
Can coaches reliably identify a hot player?
Some research has found that expert observers can predict which controlled-task shooters will become hottest. That is evidence of discriminatory information in those conditions, not a universal guarantee of correct forecasts in competitive games. Coaching decisions additionally depend on shot difficulty, defense, teammates, and opportunity costs. Miller and Sanjurjo (2024).
Do long streaks necessarily imply extraordinary skill?
No. Long sequences can arise from high stable probabilities and large numbers of opportunities. A rare event becomes less surprising when many players, attempts, teams, or games could have produced it. Skill should be estimated using all relevant attempts, comparable conditions, and uncertainty, rather than one selected run.
Does regression to the mean disprove performance streaks?
No. Regression of unusually extreme observed scores toward a person's typical performance can coexist with positive but less extreme temporary state dependence. The two claims concern different quantities. Researchers must distinguish a noisy observed peak from changes in the underlying probability of success.
How should an AI system's streak of correct answers affect trust?
Treat the successes as evidence only to the extent that they come from representative, sufficiently independent, relevant tasks. Confidence in future accuracy should be based on calibration and held-out testing across task categories, not an anecdotal chain of correct outputs. The empirical issue is predictive reliability; a sequence of accurate answers does not establish a subjective mental state.
Related Articles
Availability Heuristic: How Recall Shapes Probability Judgments — How memorable examples influence subjective frequency and risk estimates.
Representativeness Heuristic: Similarity and Probability Errors — The role of perceived typicality in probabilistic reasoning.
Base Rate Neglect: Why Prior Probabilities Matter — Integrating prior information into predictions under uncertainty.
Cognitive Biases: Definition, Types, and Psychological Research — Normative standards and methods for identifying judgment errors.
Heuristics in Psychology: How Mental Shortcuts Guide Judgment — How efficient inference strategies interact with task conditions.
Cognitive Debiasing Strategies: Which Interventions Work? — Experimental approaches to probability calibration and improved judgment.
AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making — Trust calibration when predictions and recommendations come from AI systems.
References
Avugos, S., Köppen, J., Czienskowski, U., Raab, M., and Bar-Eli, M. (2013). The “hot hand” reconsidered: A meta-analytic approach. Psychology of Sport and Exercise, 14(1), 21–27. https://doi.org/10.1016/j.psychsport.2012.07.005
Gilovich, T., Vallone, R., and Tversky, A. (1985). The hot hand in basketball: On the misperception of random sequences. Cognitive Psychology, 17(3), 295–314. https://doi.org/10.1016/0010-0285(85)90010-6
Kondur, A., and Shen, W. (2026). Statistical analysis of NBA defensive responses to hot-hand streaks. https://doi.org/10.1177/22150218261458601
Lantis, R. M., and Nesson, E. T. (2019; revised 2021). Hot shots: An analysis of the “hot hand” in NBA field goal and free throw shooting. NBER Working Paper 26510. https://www.nber.org/papers/w26510
Lantis, R. M., and Nesson, E. T. (2021). The hot hand in the NBA 3-point contest: The importance of location, location, location. NBER Working Paper 29468. https://www.nber.org/papers/w29468
Miller, J. B., and Sanjurjo, A. (2018). Surprised by the hot hand fallacy? A truth in the law of small numbers. Econometrica, 86(6), 2019–2047. https://doi.org/10.3982/ECTA14943
Miller, J. B., and Sanjurjo, A. (2019). A bridge from Monty Hall to the hot hand: The principle of restricted choice. Journal of Economic Perspectives, 33(3), 144–162. https://doi.org/10.1257/jep.33.3.144
Miller, J. B., and Sanjurjo, A. (2021). Is it a fallacy to believe in the hot hand in the NBA three-point contest? European Economic Review, 138, 103771. https://doi.org/10.1016/j.euroecorev.2021.103771
Miller, J. B., and Sanjurjo, A. (2024). A cold shower for the hot hand fallacy: Robust evidence from controlled settings. The Review of Economics and Statistics, 106(6), 1607–1619. https://doi.org/10.1162/rest_a_01280
Pelechrinis, K., and Winston, W. (2022). The hot hand in the wild. PLOS ONE, 17(1), e0261890. https://doi.org/10.1371/journal.pone.0261890
Rabin, M., and Vayanos, D. (2010). The gambler's and hot-hand fallacies: Theory and applications. The Review of Economic Studies, 77(2), 730–778. https://doi.org/10.1111/j.1467-937X.2009.00582.x
Tversky, A., and Kahneman, D. (1971). Belief in the law of small numbers. Psychological Bulletin, 76(2), 105–110. https://doi.org/10.1037/h0031322