The Gambler's Fallacy and the Hot Hand

A trader on a five-trade losing streak feels a "correction is due" and sizes up the next trade, expecting the odds to swing back in their favor — that's the gambler's fallacy. Another trader on a five-trade winning streak feels "in the zone" and also sizes up, believing the streak is evidence of a real, temporarily stronger edge — that's the hot-hand belief. These two beliefs point in opposite directions, yet both can be simultaneously wrong about the same sequence, because both are really claims about the underlying process — is it independent and identically distributed, or not? — dressed up as claims about a short observed streak.

An analogy: coin flips versus a shooter's rhythm

If a fair coin lands heads five times in a row, the gambler's fallacy says tails is "due" — but a fair coin has no memory; the sixth flip is still 50/50. That's a real, well-understood error, because coin flips genuinely are independent. But now imagine a basketball player who's made five shots in a row — is a sixth make more or less likely than their season average? Unlike a coin, a human shooter's makes could plausibly be correlated (rhythm, confidence, fatigue), so dismissing the streak as pure noise is its own potential mistake if the process genuinely isn't independent. The key question is the same either way: is the process actually memoryless, or does it have real streak-dependence? — and that has to be checked, not assumed.

The logic, one piece at a time

If successive outcomes $X_1, X_2, \ldots$ are truly independent and identically distributed (i.i.d.) with $P(\text{win}) = p$ constant, then by definition

$$P(X_{n+1} = \text{win} \mid X_n, X_{n-1}, \ldots) = p,$$

regardless of the recent history. In plain English: under a genuinely i.i.d. process, no amount of streak history changes the probability of the next outcome — not toward reversal (gambler's fallacy) and not toward continuation (hot hand). Both beliefs are errors if and only if the process really is i.i.d. If the process instead has real regime persistence — a trading strategy's edge genuinely varying with a slow-moving market regime, so that recent wins are informative about the current regime — then some genuine streak-dependence can exist, and dismissing every observed streak as illusory is itself a mistake. The correct approach is to statistically test whether outcomes are actually independent, not to assume either a reversal or a continuation story by default.

Worked example 1: the gambler's fallacy, quantified

A trader has lost 5 trades in a row on a strategy with a true, stable 50% win probability, independent trade to trade. The probability the next trade wins is still exactly $p = 0.5$. If the trader wrongly believes a reversal is "due" and assumes, say, a 70% chance of winning next, they'll oversize the position on zero statistical support. Over many such five-loss streaks under a genuinely i.i.d. 50% process, the following trade wins 50% of the time, not 70% — the extra size earns no extra expected return, only extra variance.

Worked example 2: testing for a real hot hand

Suppose over 500 trades, a strategy's overall win rate is 55%, but its win rate immediately following a streak of 3+ prior wins is 62%, and immediately following a streak of 3+ prior losses is 51%. Is that a real hot hand, or noise? With roughly 80 trades falling in the "after 3 wins" bucket, the standard error of a 62% observed rate is $\sqrt{0.55 \times 0.45/80} \approx 5.6$ percentage points, so 62% is only about $(62-55)/5.6 \approx 1.25$ standard deviations from the baseline 55% — not close to statistically convincing on its own. This is exactly the sample-size trap from the law of small numbers: a streak-conditional win rate computed on a small subsample looks meaningfully different from the baseline, but the gap is well within normal sampling noise, and a rigorous test (comparing many such streak-conditional rates, correcting for multiple comparisons) is needed before concluding the strategy has genuine momentum.

Adjust n and p above and watch how often a fair, memoryless 50/50 process still produces a run of 5+ same-direction outcomes just by chance — streaks this long are common under pure randomness, not evidence of either a due correction or a hot hand.

What this means in practice

Both fallacies drive real trading mistakes: doubling down after losses expecting mean reversion on no guaranteed timescale, or scaling up size after a winning streak believing in a "hot hand" a proper test would reveal as noise. The discipline is symmetric: don't assume reversal after losses, don't assume persistence after wins — test whether the process is actually i.i.d. (runs tests, autocorrelation of outcomes) rather than trusting either gut narrative.