Updated 12 September 2026

Gambler’s fallacy

Five heads in a row. Surely tails is overdue — the coin needs to “balance out.” This intuition is wrong, and it is wrong in a precise way: a fair coin has no memory.

Each flip is independent. The probability of tails on the next flip is exactly $\tfrac{1}{2}$, regardless of what came before. The coin cannot “know” it owes you tails.

What to notice

• Convergence with probability one. Under the independent, fixed-p model, the running proportion converges toward the true probability p, not because the coin corrects itself, but simply because new flips dilute old history. With 1000 flips, one outlier streak is a tiny fraction of the total.
• Short runs diverge wildly. In the first 20 flips, the running proportion swings dramatically. This is where gamblers are tempted to “see patterns.”
• Biased coins (p ≠ 0.5). Drag p away from 0.5. The running proportion still converges with probability one — now to the true p, not 0.5. The gambler’s error remains the same.

Law of Large Numbers vs the fallacy

The Law of Large Numbers says the running average converges to the true mean. It does not say past outcomes must be corrected. Long-run balance happens through dilution, not compensation.

The gambler’s fallacy incorrectly adds a compensation mechanism to an independent fixed-probability model. Real financial or sporting sequences need not satisfy that model; independence must be justified rather than assumed from this coin example.

Before a streak and after observing it

For an independent fair coin, a specified initial sequence of five heads has probability 1/32. Conditional on those five heads already occurring, a sixth head still has probability 1/2. The probability of six initial heads, computed before any toss, is 1/64. Those are three distinct events and information sets.

The table concerns an initial block starting at a specified position. The probability of finding a streak anywhere in a long sequence is different because many overlapping starting positions are possible; multiplying one streak probability by the number of positions generally double-counts outcomes.

Change the model: no replacement

Shuffle 10 H tokens and 10 T tokens, then draw without replacement. After five heads, five H and ten T remain, so the next-head probability is 5/15 = 1/3. After all ten H tokens have appeared, another H is impossible. The change follows from removing objects from a known finite population, not from a coin remembering its history.

The original probability of five initial heads in this urn is $(10/20)(9/19)(8/18)(7/17)(6/16)\approx0.016254$, distinct from the fair independent coin’s 0.03125. Both the prior streak probability and the conditional next draw depend on the protocol.

Work the dilution calculation

Given five initial heads and 100 future fair independent flips, expected future heads are 50. The expected final proportion is 55/105 ≈ 0.523810. The expected excess heads above half the total remains 2.5; dividing that fixed expected excess by 105 makes the proportional discrepancy small. No compensating bias toward tails has been introduced.

References

Random Services: independence formalizes what conditioning leaves unchanged. Its urn sampling chapter distinguishes replacement protocols. Compare geometric waiting and beta learning for two different ways to reason about the next observation.