OptiNod Academy
Gambler's fallacy and the hot hand — Mistaking independent trades for a streak with meaning
Raising bets after losses and becoming overconfident after wins both assume that a sequence predicts the next independent outcome. Test the sequence with a runs test.
Trade outcomes are often independent trials. Feeling “a win is due” after losses or “I am on a roll” after wins comes from related illusions. A runs test on trade records can check whether the sequence carries information.
The gambler's fallacy is the mistaken belief that, after one result repeats in independent trials, the opposite becomes more likely. Its name is associated with a roulette event at Monte Carlo in August 1913, when black appeared 26 times in a row. Gamblers kept betting on red because it was “due,” while black continued through the 26th spin. In psychology, Amos Tversky and Daniel Kahneman described the mechanism as the “law of small numbers” in 1971. People expect short samples to mirror overall probabilities and feel that an imbalance must soon correct itself.
The hot-hand belief points the other way: repeated success is expected to continue. Thomas Gilovich, Robert Vallone and Amos Tversky gave the idea its name in a 1985 analysis of basketball shots, reporting that success on previous shots did not raise the chance of the next one. Players and fans felt a scorer on a streak would score again, even though the recorded shots were close to independent.
Both beliefs can affect the same trader. After a losing streak, “it is time to win” prompts a larger next bet (gambler's fallacy). After a winning streak, “I am in rhythm” prompts a larger bet (hot hand). Both adjust size according to the order of recent outcomes. If outcomes are independent, that order contains no information about the next trade. Changing bet size in response leaves expectancy unchanged while increasing variance and risk of ruin.
Streaks are common enough to estimate. Across 100 trades with a 50% win rate, a logarithmic estimate puts the longest losing run around 6.6. Six losses do not prove a system is broken. They fall within a normal distribution of outcomes, and a lower-win-rate trend system can normally produce longer streaks.
Sequences make nonexistent patterns feel real
People expect random results to be evenly mixed. Six black roulette results in a row do not look random, while alternating red and black does. Real random sequences, however, naturally contain runs. Toss a coin 100 times and a run of at least six heads will often appear. This is the small-numbers illusion: people expect the overall 50/50 proportions to appear even in short stretches and assume an opposite outcome is needed to restore balance.
The gambler's fallacy and hot hand split that same mistake in two directions. “The streak will end” is the gambler's fallacy; “it will continue” is the hot-hand belief. Monte Carlo gamblers chose the first and kept betting on red. Basketball fans chose the second and wanted the ball to go to the player who had just scored repeatedly. The directions differ, but both misread runs in random outcomes.
Use a runs test to check whether outcomes are independent
Test actual trade records with a runs test. Arrange wins and losses in time order. Each uninterrupted block of the same result is a run; WWLLLWL has four runs. Under independence, the expected number of runs depends on the numbers of winning and losing trades.
- Expected runs: Multiply the count of wins by the count of losses, double it, divide by the total count, then add one.
With 50 wins and 50 losses across 100 trades, expected runs are 51, giving an average length of about two. Standardize the difference between observed and expected runs as a Z-score by dividing it by the appropriate standard deviation. An absolute Z above roughly 1.96 rejects independence at a conventional level. Fewer observed runs than expected (negative Z) means outcomes cluster, or positive serial correlation, which could support a hot-hand effect. More runs means outcomes alternate more often.
You can also estimate the longest losing run in advance using loss probability and trade count in a logarithmic approximation. At a 50% win rate over 100 trades, it is around 6.6; for a trend system winning 40%, around nine. If your longest run falls within the expected range, it is not by itself evidence of a broken system. Many individual trade records show no statistically significant serial correlation, so their outcome order is not distinguishable from independence.
Increasing bets after losses can concentrate ruin in one streak
Raising bets after losses turns the gambler's fallacy into a sizing rule. Martingale is the classic example. Double the bet after every loss to recover the previous losses at once, and after seven losses the next bet is 128 times the first. As above, six or seven losses are a normal maximum across 100 trades at a 50% win rate. An account unable to fund a 128-fold bet can face ruin during one ordinary streak. Martingale concentrates failure there because it assumes the run must end soon.
An early-2026 decline illustrates how this can build. Bitcoin was $97,924 on January 14 and fell to $60,000 by February 6, about 39% in a little over three weeks. A trader who bought larger at each stage because “a rebound is due” accumulated losses and placed the largest buy deepest in the decline. Each entry was justified by the fall already seen, but no falling bar guaranteed that the next one would rebound.
Overconfidence after wins mistakes a small sample for skill
The hot-hand belief raises size after a streak of profits. The trader assumes momentum in personal results will continue. But five consecutive wins also occur naturally in independent trials. With a 50% win rate, any specified five-trade sequence is all wins about 3% of the time; across dozens of trades, such stretches become unsurprising. Treating a short run as proof of skill or a favorable market state and increasing size puts a large amount behind a small-sample inference. This overlaps with overconfidence and the illusion of control.
The advance beginning in November 2024 created a run of profitable long trades into Bitcoin's January 20, 2025 high near $109,588. A trader winning repeatedly might increase entry size on the feeling of being “in the zone.” Price then fell about 32% to $74,508 by April 7. If the last entry was the largest, the greatest loss could arrive exactly where the assumption of continued success failed.
Size positions from edge and account balance
If outcome order does not inform the next result, remove it from sizing. Both the Kelly criterion and fixed-fraction sizing calculate bet size from edge (win rate and reward-to-risk) and current account balance, not the latest win or loss streak. Separating position sizing from outcome order denies both illusions a way into bet size.
- Ignore outcome order for sizing: Do not feed the current win or loss streak into the next position size.
- Keep a fixed risk share: Risk a fixed fraction of balance on each entry, such as 0.5–1%, without changing it after streaks.
- Run the test: Calculate expected runs and a Z-score from trade outcomes to see whether serial correlation is statistically meaningful.
- Know a normal losing run: Estimate the longest expected streak from win rate and sample size, and avoid changing size because of a run inside that range.
- Record the link: Note the preceding streak and bet size at each entry to see whether sequence has influenced sizing.
The point is not to ignore streaks blindly. Test whether they contain information; if they do not, remove them from the sizing formula.
The caveats are serial correlation and small samples
Some trend-following or momentum systems may show meaningful positive serial correlation: fewer runs than expected because outcomes cluster. Then a sequence-aware sizing rule might have evidence, but that evidence must come from the Z-score of the trader's own record. If Z stays within an ordinary range, roughly −2 to +2, the outcomes are not distinguishable from independence and a streak is a poor signal.
The normal approximation in a runs test needs enough trades. Under roughly 30 observations, the Z-score itself can be unstable, so a significance decision is unreliable. With a small sample, avoid attaching a new sizing rule to the sequence and keep fixed sizing while collecting more data.
Two numbers decide whether a streak is just chance
Winning and losing streaks occur naturally even in positive-expectancy systems. Across 100 trades with a 50% win rate, six losses in a row fall within normal variation, as do five wins. The problem is not the run itself; it begins when the run is treated as a signal about the next trade and used to change bet size. The gambler's fallacy concentrates ruin in a losing streak; hot-hand thinking can put the largest position near the end of a winning one. Two calculated figures—not willpower—help separate signal from chance: the runs-test Z-score and the expected longest losing run. If results are consistent with independence, remove recent outcome order from sizing.