Law of Large Numbers

The law of large numbers says that the average of many comparable random trials tends toward their expected value as the sample grows.

The law of large numbers says that the average result from many comparable random trials tends to move toward the process’s expected value as the number of trials grows.

It explains why a casino can have a losing table, shift, or even month while still operating games with a mathematical advantage. It does not say that a particular player must recover, that a color is due, or that a short session will resemble the published house edge.

The average that converges

Let (X_1,X_2,\ldots,X_n) represent the net result of successive wagers under the same probability and payout rules. The sample average is:

[ \bar X_n=\frac{X_1+X_2+\cdots+X_n}{n} ]

If the trials satisfy the relevant mathematical conditions and have expected value (\mu), the law of large numbers says that:

[ \bar X_n\rightarrow\mu\quad\text{as }n\text{ becomes large} ]

In plain English, the average result per wager becomes more likely to sit close to the theoretical average when the sample is large.

The statement is about averages, not about the order of wins and losses. A sequence can contain long runs, sudden jackpots, and deep drawdowns while its average slowly becomes more stable.

Roulette frequency example

On a fair European roulette wheel, red occupies 18 of 37 pockets, so:

[ p=\frac{18}{37}\approx0.4865 ]

If (R_n) is the number of red results in (n) spins, the observed red frequency is:

[ \hat p_n=\frac{R_n}{n} ]

The law of large numbers says that (\hat p_n) tends toward (18/37) as the number of comparable spins grows.

That does not mean 100 spins must produce 49 reds. It does not even mean the percentage moves closer on every additional spin. A result at spin 501 can push the percentage farther from 48.65% than it was at spin 500. Convergence is not a smooth correction mechanism.

The scale of ordinary fluctuation can be described by the binomial standard deviation:

[ \sigma=\sqrt{np(1-p)} ]

For 1,000 spins, the standard deviation of the red count is approximately 15.81 spins. For 100,000 spins, it is approximately 158.06 spins.

The absolute fluctuation is larger in the bigger sample, but relative to the sample size it is much smaller:

(15.81/1{,}000\approx1.58%);
(158.06/100{,}000\approx0.16%).

This is the useful intuition: the count does not freeze near its expectation, but the percentage error tends to shrink.

Why this is not the gambler’s fallacy

The gambler’s fallacy turns a long-run statement into a next-event prediction.

Suppose red has appeared only 40 times in 100 spins. The long-run red frequency may eventually move closer to 48.65%, but that does not require the 101st spin to be red. Assuming independent wheel spins, the next-spin probability remains 18/37.

Future results can move the cumulative percentage toward expectation without any single result being “owed.” The correction, when it appears in the average, comes from adding many ordinary trials—not from the wheel remembering a deficit.

Average loss can stabilize while total loss grows

A European roulette even-money wager loses on zero as well as on the opposite color. Its house edge is:

[ h=\frac{1}{37}\approx2.70% ]

For a constant $10 wager, expected loss per spin is:

[ 10\times\frac{1}{37}\approx$0.27 ]

Across (n) spins:

[ \text{Expected total loss}=n\times$10\times\frac{1}{37} ]

After 100 spins, total action is $1,000 and expected loss is about $27.03. After 10,000 spins, total action is $100,000 and expected loss is about $2,702.70.

The average loss per spin becomes more stable around $0.27, but the expected total loss grows with the number of wagers. This is why “playing long enough for the math to work” benefits the operator of a negative-expectation game, not the player seeking recovery.

Why casinos reach the long run more easily

A single player has limited time, money, and number of decisions. A casino combines activity across many:

players;
tables and machines;
shifts and gaming days;
wager sizes;
game types;
locations or channels.

That scale creates far more observations than one patron can produce. The property can also hold reserves and manage limits so that ordinary variance does not threaten operations after one large result.

The law still does not guarantee a fixed daily win. High-limit baccarat, major progressives, promotional awards, or an unusual concentration of large wagers can create substantial volatility. The casino’s edge becomes dependable through repeated properly priced action, not through immunity from bad periods.

The trials must be meaningfully comparable

A common reporting mistake is to combine unlike observations and then invoke the law of large numbers.

The average of a mixed dataset may be hard to interpret when:

game rules changed during the period;
different paytables were pooled together;
promotional free play was treated like cash wagering;
wager sizes changed sharply;
a machine configuration was altered;
manual ratings were inconsistent;
one jackpot game dominated the result;
data from different jurisdictions used different definitions.

More data do not repair a bad denominator or a shifting process. Before comparing actual performance with theoretical expected value, the casino must define the population, rules, period, and measurement method consistently.