Gambler's Fallacy

011 min

Gambler's Fallacy at a Glance

• What it is — expecting an independent random event to correct itself after a run, even though each trial carries the same odds.
• Origin — named from Amos Tversky and Daniel Kahneman's 1971 work on the "law of small numbers."
• Where it bites — a PM who keeps a losing A/B variant running because it feels "due" to turn around.
• Guard against it — check whether the process can actually remember its last outcome before betting on a correction.

021 min

Where Gambler's Fallacy Shows Up

A data analyst is monitoring two versions of a checkout page in an A/B test. Version B has converted worse than version A for four days running. On day five, the analyst tells the team version B is "due for a good day" and argues against calling the test early, expecting the losing streak to reverse on its own. Nothing about the underlying conversion rate has changed; four independent days are being treated as though they owe a correction to the average.

The same analyst runs into the pattern again a month later, on a different test. This time version B has converted better for four days straight, and the analyst treats that streak as proof the variant is genuinely stronger, planning to ship it without waiting for statistical significance. In the first case, a losing streak read as "must average out." In the second, a winning streak read as "must be real." Both readings skip the same question: whether four days, on its own, is enough evidence to say anything about the process generating the numbers.

The pattern generalizes past A/B tests. Whenever an outcome is genuinely independent from one trial to the next — a coin flip, a shuffled deck, a roulette wheel — people tend to expect the string of results to self-correct toward the average faster than probability actually delivers it.

032 min

Why Gambler's Fallacy Happens

By definition, the gambler's fallacy is the belief that an independent and equally probable outcome which happened less frequently than expected is more likely to happen in the future, or the reverse when an outcome has happened more often than expected. The clearest account of why people believe this comes from what Amos Tversky and Daniel Kahneman called the representativeness heuristic: people judge how likely an outcome is by how well it matches, or "represents," their mental picture of a random process, rather than by calculating the actual odds. A fair coin's mental picture is a roughly even mix of heads and tails with no long runs. A sequence like six heads in a row does not match that picture, so people expect the next flip to correct the imbalance, even though a coin has no memory of what it did last.

Tversky and Kahneman named the broader pattern the "law of small numbers": people expect a small sample to reflect the same proportions as a very large one, when only large samples reliably do that. It sits alongside a related shortcut from the same research program, the availability heuristic, which judges likelihood by how easily examples come to mind rather than by how representative a sample looks. Which mechanism explains more of the gambler's fallacy is not fully resolved — a competing account treats the fallacy as a byproduct of a genuinely useful skill, since humans are unusually good at detecting real, non-random patterns in the world, and that same pattern-detection machinery misfires on sequences that are actually random.

041 min

Where Gambler's Fallacy Comes From

Tversky and Kahneman gave the mechanism its name in a 1971 paper titled "Belief in the Law of Small Numbers," published in Psychological Bulletin. The paper did not run a new gambling experiment; it surveyed how professional researchers themselves misjudge random sampling, arguing that the same faulty intuition about small samples that misleads a roulette player also misleads a scientist reasoning from a small pilot study.

The event most often used to illustrate the fallacy happened decades earlier, independent of that research. An example of the gambler's fallacy occurred in a game of roulette at the Monte Carlo Casino on August 18, 1913, when the ball fell in black 26 times in a row. As the streak grew, gamblers increasingly bet on red, certain a correction was overdue, and lost millions of francs betting against a sequence in which each spin was exactly as independent as the first.

051 min

Individual Effects

For one person, the fallacy shows up as a decision that treats an unrelated past outcome as evidence: a trader who closes a position because a currency pair has risen for five straight sessions and feels "overdue" for a pullback, or an interviewer who rates the sixth strong candidate of the day more harshly, assuming the next one must be weaker to balance things out.

The direction of the error always runs against whatever streak is showing. A run of losses reads as license to keep betting because a win feels close; a run of wins in a genuinely independent process reads as a reason to expect a reversal, rather than a reason to update toward "the win rate really is this high." Either way, the person spends attention correcting for a trend the process itself was never tracking.

061 min

Systemic Effects

The fallacy compounds once it operates inside a market rather than inside one person's head, because thousands of bettors making the same correction at once becomes a measurable pattern in the data. Economists Charles Clotfelter and Philip Cook studied Maryland's daily numbers lottery and found a clear and consistent tendency for the amount of money bet on a particular number to fall sharply immediately after it is drawn, and then gradually to recover to its former level over the course of several months — even though each daily drawing is independent of the last. They concluded the pattern is consistent with lottery players, as a group, being subject to the gambler's fallacy.

The systemic version is more expensive than the individual one because it moves real money at scale: a lottery operator's return on a given number shifts measurably in the weeks after that number is drawn, purely because bettors are reasoning about odds that never changed.

071 min

Examples

081 min

How Gambler's Fallacy Shows Up in Product, Design, and AI

In product and growth work, the fallacy usually shows up in how a team reads a funnel. A PM tracking daily signups notices three bad days in a row and delays a launch, waiting for the numbers to turn around on their own, while the actual cause, a broken referral link, sits unfixed the whole time. Treating a string of bad numbers as something that self-corrects can hide the actual, fixable cause of a metric moving.

Roulette tables post an electronic display of the last twenty spins, and the display is not neutral: showing a streak pulls bettors toward the "opposite" side, increasing total money wagered without changing the house edge at all. The information shown is genuinely true; the display exists because the house knows how people will misread it.

In AI and data work, the same error shows up in how a team reads evaluation runs. When an automated eval pipeline fails four times in a row because of flaky test infrastructure and then passes, an engineer might conclude the underlying system just got more reliable, when nothing about the failure rate actually changed. The run was independent of the last four, and a fifth failure remains exactly as likely as it always was.

091 min

How to Guard Against Gambler's Fallacy

Knowing the gambler's fallacy exists does little on its own to stop it. The same failure that lets a statistically literate gambler still feel a hot streak coming lets a data-literate analyst feel a metric is "due." Two techniques change the decision process instead of relying on willpower.

The first is checking independence explicitly before reacting to a streak: write down, in one sentence, whether the mechanism generating the outcomes can "remember" the last result. A coin, a shuffled deck, and daily traffic driven by unrelated causes all pass this test as independent; a process with real momentum, such as compounding interest or a snowballing PR crisis, does not, and treating it as independent would be the opposite mistake.

The second is pre-registering a stopping rule before watching results come in — standard practice in running an A/B test: decide in advance how many trials or what sample size ends the test, rather than deciding in the moment that a losing streak "feels" long enough to matter. This removes the exact point in the process where a betting streak turns into a business decision.

Awareness alone does not reliably fix this bias or most others — the fix has to change what information a decision is allowed to use, not just remind the decision-maker to be careful.

101 min

Common Misunderstandings

The most common misunderstanding is applying the fallacy to processes that are not actually independent. If a sales rep has closed four deals this week using a new pitch, expecting a fifth is not the gambler's fallacy: skill, timing, and a live pipeline can genuinely create real streaks. The fallacy only applies to mechanisms with no memory: dice, coins, roulette wheels, and lotteries with replacement, where the true odds of the next event never change based on the last one.

A second misunderstanding treats the fallacy as proof that intuitions about randomness are always wrong. The bias is specific: it is the belief that independent events correct for past deviations. It says nothing about genuine trends, momentum, or skill, and treating a real, non-random signal as though it must reverse would be a mistake in the opposite direction.

111 min

Gambler's Fallacy vs. Nearby Concepts

The nearest neighbor is the hot hand fallacy, and the two are near-opposites built from the same misreading of randomness. The gambler's fallacy expects a streak to reverse: a coin feels "due" for tails after a run of heads. The hot hand fallacy expects a streak to continue. A 1985 paper by Thomas Gilovich, Amos Tversky, and Robert Vallone questioned the hypothesis that basketball players have "hot hands," which the paper defined as the claim that players are more likely to make a successful shot if their previous shot was successful, and tested it against the individual records of players from the 1980–81 Philadelphia 76ers. A 2018 reanalysis, building on a correction identified by Joshua Miller and Adam Sanjurjo, found a statistical bias in how the original study counted streaks, reopening the question of whether some hot-hand effect is real after all.

A second neighbor is the narrative fallacy: the tendency to build a coherent story out of what is actually noise. The gambler's fallacy is one specific narrative fallacy applied to sequences of independent random events; the broader term covers any invented causal story imposed on data that has none.

121 min

Where the Evidence Is Contested

Not every researcher treats the fallacy as a clean case of irrationality with a clean cost. Juemin Xu and Nigel Harvey analyzed 565,915 sports bets made by 776 online gamblers in 2010 and found a twist: bettors who had just won switched to safer odds, and bettors who had just lost switched to riskier ones, in both cases because they expected their luck to reverse. That switching pattern paid off for the winners and backfired on the losers, so acting on the fallacy ended up manufacturing real winning and losing streaks rather than avoiding them.

A second complication is that not every expectation of reversal is a fallacy. In games without replacement, such as a deck of cards dealt without reshuffling or a lottery drawn from a shrinking pool of balls, the true odds genuinely shift after each draw, so expecting a change is correct reasoning rather than a bias. Everyday experience with these non-independent games may be exactly what primes people to misapply the same intuition to roulette and coin flips, which carry no memory of the last spin at all.

131 min

Related Concepts

• Availability Heuristic — judges likelihood by ease of recall, from the same research program.
• Anchoring — distorts a number toward a starting point rather than a streak.
• Sunk Cost Bias — keeps a decision going because of past investment, not a misread streak.
• Narrative Fallacy — the broader tendency this bias is one instance of.
• Confirmation Bias — favors evidence for an existing belief rather than misreading independent odds.

?6 questions

Questions people ask

§7 sources

Sources