Law of large numbers in gambling: why volume favours the house

Take any bet whose outcomes are independent and identically distributed: every spin of the same wheel, every flip of the same coin, every hand dealt from a freshly shuffled shoe with the same rules. Give that bet an expected value, call it μ. The law of large numbers says that the average of your first n results, (X₁ + X₂ + … + Xₙ) / n, converges to μ as n grows.

There are two formal versions.

• Weak law: for any small margin you choose, the probability that the sample average sits outside μ ± that margin shrinks towards zero as n grows.
• Strong law: with probability 1, the sequence of averages eventually settles at μ and stays there.

For a gambler the difference rarely matters. Both say the same practical thing: over enough repetitions, the average result per bet stops being a surprise.

Jacob Bernoulli worked on the idea for about two decades and it appeared in *Ars Conjectandi* in 1713, eight years after his death. He framed it as drawing coloured pebbles from an urn: draw often enough and the observed proportion approaches the true proportion. A roulette wheel is an urn you draw from with replacement, so his framing fits casino games almost exactly.

The expected value of a casino bet is negative by design. The size of that negative number is the house edge, and that page keeps the definition and the table of edges by game. This page is about what happens when you repeat a bet with that edge many times. If you want the arithmetic behind μ itself, expected value in gambling walks through it.

This is the part most people get backwards. The law says your average converges. It says nothing reassuring about your total.

Use a concrete bet: $1 on Purple on a 15-slot wheel where 7 Purple slots pay 2x, 7 Silver slots pay 2x and 1 Green slot pays 14x. A Purple bet wins $1 with probability 7/15 and loses $1 with probability 8/15.

• Expected value per bet: (7/15)(+1) + (8/15)(−1) = −1/15 ≈ −$0.0667.
• Variance per bet: 1 − (1/15)² ≈ 0.996, so the standard deviation is almost exactly $1.

After n bets the expected total is −n/15, and the standard deviation of the total is about √n dollars. The expected loss grows in proportion to n. The noise grows only with √n. That mismatch is the whole law in one line.

The probabilities come from the exact binomial distribution, not a rule of thumb. Read the table two ways.

From the player's seat

At 10 spins the edge is invisible. A loss of 67 cents is buried under a $3 swing. Nearly a third of short sessions finish ahead, so plenty of players have real memories of beating the wheel. At 1,000 spins the expected loss is more than twice the standard deviation and finishing ahead becomes rare.

From the average's seat

The standard deviation of the average is 1/√n: about 32 cents per bet at n = 10, 3 cents at n = 1,000, 1 cent at n = 10,000. That shrinking spread around −6.67 cents is the convergence Bernoulli proved.

A fair game shows the same effect with no drift. On a 50/50 coin flip the average result converges to zero, yet the typical distance of your total from zero after 10,000 flips is about 100 units. Being $100 up or down after 10,000 flips is ordinary, not a sign that anything is owed back. Variance in gambling goes deeper on that spread.

A single player might make a few hundred bets in a night. A casino, or any operator running a house-banked game, sees millions. Put the house on the other side of the same Purple bet.

• One million $1 bets.
• Expected hold: 1,000,000 × 1/15 ≈ $66,667.
• Standard deviation of the hold: about √1,000,000 ≈ $1,000.

So the house can say with about 95% confidence that its result lands between roughly $64,700 and $68,700. That is a spread of about ±3% around its forecast. The house is not gambling in any meaningful sense. It is running a business with a known margin and a small statistical wobble.

Why big bets still matter to a casino

The √n rule assumes equal stakes. One player betting $1 million on a single spin adds as much variance as a trillion $1 bets would add in total. That is why physical casinos set table maximums and why high-roller play is priced and managed separately. Volume only protects the house when no single bet dominates it.

Why the edge does not need to be large

At a 1% edge, the expected result after n bets is 0.01n and the noise is about √n. The expected loss overtakes one standard deviation at n = 10,000 bets and two standard deviations at 40,000. A slot machine resolves hundreds of spins an hour. A busy casino floor passes those counts every few minutes. Small edges are enough when volume is huge.

The same arithmetic explains why a player cannot turn a negative game into a positive one by playing more. Extra volume is exactly what lets the house's average show up.

The phrase "law of averages" is often used to mean something the mathematics does not say: that after a run of Silver, Purple is now due, or that a losing player is owed a comeback.

The real law works by dilution, not correction. Suppose the first 10 spins you watch contain 8 Purple. Over the next 990 spins you expect 462 more Purples, because each spin still has a 7/15 chance. Your running total is then about 470 out of 1,000, or 47.0%. The early excess of 3.3 Purples above expectation has not been cancelled. It is still there. It has simply become a small share of a large sample.

The belief that the wheel actively corrects itself is the gambler's fallacy, and that page covers the independence mistake in detail. The related effect where an extreme result tends to be followed by a less extreme one, without any balancing force, is regression to the mean.

A quick test of which idea you are using: if your reasoning changes the probability of the next spin, it is the fallacy. If it only changes what you expect the long-run percentage to look like, it is the law.

Two bets with the same edge can converge at very different speeds. On the 15-slot wheel, Green and Purple both return 14/15 of the stake on average, so both carry a 6.67% edge. Their spread is very different.

The last column solves n/15 = 2√n × σ, which gives n = (30σ)². Because Green's standard deviation is about 3.5 times larger, it needs about 12 times as many bets before the edge dominates the luck.

That is why long-shot bets, lottery tickets and high-volatility slots produce more "I won big" stories than even-money bets do, even at the same edge. They do not return more. They take longer for the law to catch up, so more players quit while ahead of their expected result.

When the law does not apply cleanly

The theorem has conditions. Real gambling sometimes breaks them.

• Changing odds. In blackjack the composition of the remaining shoe shifts the edge hand to hand. That dependence is what card counting exploits.
• Changing stakes. Progressions such as the Martingale strategy vary the stake with results. The law still applies per dollar wagered, so expected loss remains edge × total turnover.
• Infinite expected value. When a game's expected value is not finite, averages do not settle at all. The St Petersburg paradox is the classic example.
• Skill games. In poker or player-vs-player matches your expected value depends on opponents, so it can move as the field changes.

PVPspinArena runs three player-vs-player games in USDC or ETH on Base, and each one shows the law differently.

• [Roulette](/roulette) is the table above. Every bet returns 14/15 ≈ 93.33% on average, so over enough spins your average result per dollar drifts to −6.67 cents, whichever colour you pick. Picking Green changes how long that takes, not where it ends.
• [Coinflip](/coinflip) is two players on a 50/50. Over many flips your win rate converges to 50%, and your average result converges to minus whatever fee was shown before entry. Your total can still sit well above or below that line for a long time.
• [Jackpot](/) gives each player a win chance equal to their share of the pot. Over many rounds your share of wins converges to your average share of pots, so your average return converges to what you put in minus any fee shown before entry.

Every round comes from committed seeds, and any settled round can be checked on fairness. Verification tells you each draw was honest. It is the honesty of each draw that makes the law of large numbers a reliable forecast rather than a hope.

The practical use for a player is budgeting, not beating the game. Multiply your planned turnover by the edge and you have the expected cost of a session. Everything else is variance. If sessions are getting longer so that you can "let the averages work", that is chasing, and the tools on responsible gambling are there for it. Gambling on PVPspinArena is 18+. More guides on probability sit in the casino knowledge topic.