Glossary index › L

Law of Large Numbers

The law of large numbers is a statistical principle stating that as the number of trials or observations in a random process increases, the average of the results gets closer and closer to the theoretical expected value. Flip a fair coin 10 times and you might get 7 heads and 3 tails - a 70% heads rate purely from small-sample noise. Flip it 10,000 times and the heads rate will almost certainly land very close to 50%, because the large sample size smooths out the short-run randomness.

In trading, this law is the entire statistical justification for why a strategy needs a large enough sample of trades before its results mean anything. A trading system that wins 5 trades in a row could simply be lucky - five trades is nowhere near enough data to distinguish a genuinely profitable strategy (say, one with a real 55% win rate) from random noise. Professional quantitative traders typically want hundreds or thousands of historical trades in a backtest before trusting the resulting win rate or expectancy, precisely because the law of large numbers only kicks in with real force at scale; a 60% win rate over 20 trades is statistically almost meaningless, while the same 60% over 2,000 trades is a much more reliable signal of a genuine edge.

The concept also underlies risk and insurance-style thinking in markets. A market maker or options seller who takes the same side of thousands of small, independent bets - collecting a slightly-favorable spread or premium on each - relies on the law of large numbers to turn a modest statistical edge per trade into a highly predictable profit over the full volume, even though any single trade could go either way. This is structurally identical to how an insurance company prices policies: individually unpredictable, collectively very predictable.

The practical takeaway for retail traders is a caution against overreacting to small samples: a losing streak of 4-5 trades on a strategy with a real, tested edge doesn't disprove that edge any more than a coin landing heads 4 times in a row disproves that it's a fair coin, and the reasoning is about sample size rather than about any particular instrument, so a run of results proves no more in forex than it does in a portfolio of shares.

The most expensive misunderstanding of this law is the belief that it promises a correction. It does not. A run of bad results is never cancelled by a compensating run of good ones; it is simply diluted, becoming a smaller and smaller fraction of a growing total until it stops mattering as a proportion while remaining exactly as large in absolute terms. A sequence that is a hundred flips behind after a thousand tosses is still a hundred flips behind after ten thousand, which is why the average has moved back toward half while the shortfall itself has not moved at all. Reading the law the other way round, as an assurance that what is owed will be repaid, produces the belief that a loss makes a win due, and that belief is the engine behind every system that answers a losing trade by doubling the next one.

There is a second qualification that matters more in markets than in textbooks. The mathematics assumes trials are independent of each other and drawn from the same underlying distribution, and trading results satisfy neither condition. Conditions change, so the process generating this year's returns is not the one that generated the last decade's; and losses arrive in clusters because whatever caused the first one is usually still there for the second. A sample collected across several different market environments is therefore not a large sample of one experiment but a small sample of each of several, which is the reason a thousand historical trades can look statistically convincing and still say very little about the next hundred.

What is the law of large numbers? That as the number of trials rises, the average result converges on the theoretical expectation.
Can you show it? Ten coin flips can easily give seven heads. Ten thousand will land very close to half, because the short-run noise is swamped by the sample size.
Why does it matter in trading? It is the entire reason a strategy needs a large sample before its results mean anything. Five winning trades in a row is indistinguishable from luck.
How large is large enough? Professionals generally want hundreds or thousands of trades before trusting a win rate. Sixty percent over twenty trades says almost nothing; the same figure over two thousand says a great deal.
Where else does it appear? In market making and option selling, where thousands of small independent bets each carrying a slight edge produce a predictable total, exactly as an insurer prices policies.
What is the practical lesson? Not to overreact to small samples. Four or five losses in a row no more disproves a tested edge than four heads disproves a fair coin.
Does it mean losses get corrected? No, and this is the most expensive misreading of it. The law says deviations are diluted by later results, not reversed by them: a hundred flips behind stays a hundred flips behind, it simply stops mattering as a proportion of ten thousand. Believing that a losing run makes a winning run due is the gambler's fallacy, and it is the reasoning behind every position-doubling system that has ever destroyed an account.
Does it even apply to trading? Less cleanly than to coins, because it requires the trials to be independent and identically distributed, and trades are neither. Markets change regime, a strategy's edge in one environment can be absent in the next, and losses cluster because the conditions that cause them persist. A large sample drawn from several different regimes is not the same thing as a large sample of one experiment, which is why a thousand backtested trades still does not guarantee the next hundred behave like them.

By InsiderTicker Editorial Team · Last updated