The Gambler's Fallacy and the Law of Large Numbers

In 1913 at the Casino de Monte-Carlo, a roulette ball landed on black 26 times in a row. Crowds packed around the table and bet heavily on red, convinced a correction was overdue. The casino made a fortune. The gamblers had confused two different statistical ideas, and that confusion cost them.

What the Law of Large Numbers Actually Says

The law of large numbers states that as the number of trials increases, the sample mean converges to the population mean. Flip a fair coin 10 times and you might get 7 heads. Flip it 10,000 times and the proportion of heads will be very close to 0.5.

This is a real and important result. It is why casinos make reliable profits. The house edge on roulette is small on any single spin, but across millions of spins the actual payout proportion converges to the theoretical one and the house collects its margin with certainty.

What the Gambler's Fallacy Gets Wrong

The gambler's fallacy is believing that the law of large numbers operates by correcting past deviations. After 26 blacks, the gamblers believed red was "owed" a run to restore the 50/50 balance. This is wrong. The roulette wheel has no memory. Each spin is an independent event with P(red) = 18/38 regardless of what has come before.

The law of large numbers does not work through correction. It works through dilution. If you have seen 26 blacks in a row and then spin the wheel 100,000 more times, the eventual proportion of reds will be close to 50% because 100,000 spins swamp the 26 initial results. The early imbalance becomes irrelevant, not corrected.

This distinction matters. Correction would imply future dependence on past results. Dilution implies the past is simply overwhelmed by future volume. The math requires dilution; dependence would violate independence.

Formal Statement of Independence

For the law of large numbers to apply, the trials must be independent and identically distributed (iid). Under iid, the expected value of each trial is the same, and the trials carry no information about each other. The convergence of the sample mean to the population mean follows from these conditions, not from any self-correcting mechanism in the process.

If you could verify that a process is iid, you can be confident the long-run average will converge. But you cannot infer anything about the next individual trial from this.

Sequences and Runs

A related misconception is that long runs are improbable events that signal non-randomness. In a sequence of 200 fair coin flips, the probability of seeing a run of at least 7 heads in a row somewhere in the sequence is over 95%. Long runs are expected in random sequences, not evidence against them. When people see a streak and call it a pattern, they are often seeing exactly what randomness produces.

This matters in quality control, sports statistics, and financial analysis. Before concluding that a run of failures or successes is evidence of a systematic effect, calculate how probable such a run would be in a purely random process. The answer is often more probable than intuition suggests.