The Law of Large Numbers: Why Casinos Never Lose Long-Term

On the night of August 18, 1913, the roulette wheel at the Monte Carlo Casino landed on black. Then it landed on black again. And again. By the time the ball had fallen black for the twenty-sixth consecutive spin, the betting floor was in chaos. Gamblers, certain that red was now "due," piled their money onto red with mounting desperation — surely, after such an absurd run, the universe owed them a correction. It never came. Players lost millions of francs chasing a red that was never any more likely to appear than it had been on spin one. They had stumbled, in real time, into one of the most persistent misunderstandings in all of mathematics — and into a vivid, expensive demonstration of a theorem that actually says something quite different from what they believed.

That theorem is the Law of Large Numbers, and it is the reason casinos can offer games that are individually unpredictable and yet, as a business, never lose money over time. It is also the reason insurance companies can price policies for events they can't foresee, why pollsters trust a few thousand phone calls to estimate the opinions of millions, and why a badly shuffled deck of intuitions about randomness gets corrected, slowly, by data. It is one of the oldest results in probability theory, and it is still routinely misunderstood by the very people — gamblers, mostly — who most want it to say something it doesn't.

The Concept

The Law of Large Numbers says something almost embarrassingly simple: if you repeat a random process many times and average the results, that average will settle down and converge toward the true theoretical average — the "expected value" — of the process. Flip a fair coin once, and you might get heads. Flip it ten times, and you might get seven heads, a proportion of 70%, wildly off from the expected 50%. But flip it ten thousand times, and the proportion of heads will almost certainly land very close to 50%. Flip it ten million times, and it will be closer still.

Crucially, the theorem does not say that the number of heads and tails will "even out" or that some corrective force nudges the tally back toward balance. It doesn't. If you flip 10,000 more coins after an unlucky run of tails, the absolute gap between your heads-count and your tails-count can easily grow larger in raw terms. What shrinks is the gap as a proportion of the total. Ten extra tails out of ten flips is catastrophic; ten extra tails out of ten million flips is a rounding error. The theorem is a statement about ratios and averages, not about the universe balancing its books.

This distinction is exactly what wrecked the gamblers in Monte Carlo. Each spin of a roulette wheel is statistically independent — the wheel has no memory. The odds of black on the twenty-seventh spin were the same as the odds on the first, regardless of what came before. The 1-in-66.6-million odds applied to the whole sequence of 26 blacks in a row, not to what would happen next. Believing that red was "due" is a related but distinct error known as the gambler's fallacy, and it is precisely the mistake the Law of Large Numbers is most often — and wrongly — invoked to justify.

Why It Matters

The Law of Large Numbers is quietly one of the most economically important theorems ever proven, because it is the mathematical foundation of the entire insurance industry. An insurer cannot know whether any individual policyholder's house will burn down this year. But across a large enough pool of policyholders, the proportion who file claims settles into a stable, predictable pattern. Actuaries use decades of claims data to estimate that stable long-run rate, then price premiums so that the pooled payouts, averaged over enough customers, land comfortably below the pooled revenue. No individual prediction is required — only the confidence that large numbers behave predictably even when small numbers don't.

Casinos run on the identical principle, just with better odds baked in from the start. Every casino game has a built-in house edge — the small percentage of each bet that the casino expects to keep, on average, over the long run. European roulette's house edge is about 2.7%; American roulette, with its extra double-zero pocket, runs about 5.26%; blackjack played with perfect basic strategy can be pushed down near 0.5%. On any single spin or hand, the house can lose — and does, constantly. But across the millions of bets placed in a casino each year, the Law of Large Numbers guarantees that the actual proportion of the casino's take converges toward that built-in percentage. The casino doesn't need to know who wins tonight. It only needs enough bets placed, over enough time, for the averages to do their work. This is also why a casino is far more nervous about a single high-stakes whale placing one enormous bet than about ten thousand tourists at the low-limit tables — the whale's bet hasn't been smoothed by large numbers yet.

The same logic underwrites political and consumer polling. A well-designed poll of a few thousand randomly selected people can estimate the preferences of tens of millions with a small, quantifiable margin of error — not because a few thousand people are inherently representative of a nation, but because random sampling combined with the Law of Large Numbers guarantees that the sample average converges toward the population average as the sample grows, and statisticians can calculate exactly how tight that convergence should be. It is why pharmaceutical trials enroll thousands of patients rather than ten, why A/B tests at tech companies need large sample sizes before a "winning" version of a webpage can be trusted, and why a single glowing (or scathing) product review means far less than a rating averaged across ten thousand buyers.

The Details

The theorem's first rigorous proof belongs to the Swiss mathematician Jacob Bernoulli, who spent roughly two decades wrestling with it before it appeared in his book Ars Conjectandi ("The Art of Conjecturing"), published posthumously in 1713 by his nephew Nicolaus Bernoulli. Bernoulli called it his "Golden Theorem," and for good reason: he had shown, for the first time, that empirical frequencies — the messy, observed proportions of real-world trials — must mathematically converge toward the theoretical probabilities underlying them. This was a genuinely new kind of claim. Before Bernoulli, probability was largely a tool for analyzing idealized games of chance with known, symmetric odds — a fair die, a fair deck of cards. Bernoulli's theorem justified extending probability to situations where the "true" odds were unknown and had to be estimated from accumulated observation, opening the door to applying mathematical reasoning to demography, insurance, and what he called "civil, moral, and economic affairs." The name "Law of Large Numbers" itself wasn't coined until more than a century later, in 1837, by the French mathematician Siméon Denis Poisson.

Modern probability theory actually recognizes two versions of the law, and the difference is subtle but real. The Weak Law of Large Numbers says that as the number of trials grows, the probability that the sample average differs from the true average by more than some tiny amount shrinks toward zero — convergence "in probability." The Strong Law of Large Numbers makes a bolder claim: that the sample average doesn't just probably land near the true average, but converges to it with certainty (technically, "almost surely") as the number of trials goes to infinity. The strong law took much longer to prove rigorously — that work fell largely to the Russian mathematician Andrey Kolmogorov in the 1930s, more than two centuries after Bernoulli's original insight — and it remains one of the conceptual pillars on which all of modern statistics rests.

It's worth sitting with a visual image of what convergence actually looks like, because it's not a smooth glide into place. Picture a graph tracking the running proportion of heads in a long sequence of coin flips, with the x-axis as the number of flips and the y-axis as the cumulative proportion. Early on — the first few dozen flips — the line lurches wildly, spiking up toward 80% or crashing down toward 20% as a lucky or unlucky streak happens to dominate a small sample. But as the flip count climbs into the hundreds, then thousands, then millions, the lurches shrink and the line settles into an ever-narrower band hugging the 50% mark, like a wave whose amplitude decays even as it keeps rippling. The wildness never fully disappears at any finite point — there is no flip number after which the line becomes perfectly flat — but its scale shrinks toward zero, in exactly the ratio sense the theorem promises.

That mental image also exposes the theorem's most important limit: it's an asymptotic promise about the very long run, and it says nothing precise about how quickly convergence happens or what to expect from any specific short run. A gambler at the Monte Carlo table who reasoned "the wheel must correct itself soon" was applying a short-run expectation to a theorem that only makes claims about the infinite long run — and the wheel, having no memory of its own history, was never obligated to correct anything.

Takeaways

The Law of Large Numbers guarantees that averages of many independent trials converge toward the true expected value — it says nothing about individual trials "balancing out" or streaks being "due" to end.
Jacob Bernoulli proved the theorem around 1689 and published it in 1713's Ars Conjectandi; the name "Law of Large Numbers" came later, from Siméon Denis Poisson in 1837.
Casinos, insurers, and pollsters all depend on the same mathematical fact: unpredictable individual outcomes become statistically stable in aggregate, which is why a casino's house edge (roughly 0.5% to 5.26% depending on the game) reliably turns into long-run profit even though any single bet is a coin toss.
The gambler's fallacy — believing red is "due" after 26 blacks in a row, as happened at Monte Carlo in 1913 — is a misapplication of this theorem, not a consequence of it; each independent trial has no memory of previous outcomes.
There are two formal versions, the Weak and Strong Laws, differing in how strictly they define "convergence" — a distinction that took mathematicians from Bernoulli's era through Andrey Kolmogorov's 1930s work to fully resolve.

Resources: Jacob Bernoulli's Ars Conjectandi (1713) is the theorem's original source; for a more modern treatment, the Strong Law of Large Numbers is covered in most introductory probability theory textbooks, including Kolmogorov's foundational 1933 work on the axiomatic foundations of probability.