Law of Large Numbers
Why expected value materialises over the long run, how many trades you need for your edge to show, and the relationship with the Central Limit Theorem.
The Law: Sample Mean → Population Mean
The Law of Large Numbers (LLN) is one of the fundamental theorems of probability. In its intuitive form it says: the empirical average of a large number of independent observations converges to the theoretical (expected) mean. Formally: if X₁, X₂, X₃, …, Xₙ are i.i.d. random variables with finite mean μ, then as n → ∞, the sample mean X̄ₙ = (X₁ + X₂ + … + Xₙ) / n converges to μ. In trading terms, if your strategy has a true expected value of $15 per trade, then across many independent trades, the observed average P&L per trade will converge to $15. That is the mathematical reason why a sustained positive edge generates gains over the long run, even though any individual trade can win or lose. The LLN is the underlying justification for systematic trading and long-term investing. It was rigorously proved in the seventeenth century by Jacob Bernoulli and extended by Chebyshev and Kolmogorov. It is a cornerstone of probability theory, inferential statistics and modern quantitative finance. Without the LLN there would be no way to build statistical edges that reliably yield money: it would all be pure chance.
Weak Law vs Strong Law
Mathematically there are two versions of the LLN, with technical importance but similar practical implications. (1) The Weak Law of Large Numbers (Bernoulli 1713, Khinchin) asserts convergence in probability. For any ε > 0, Pr(|X̄ₙ − μ| > ε) → 0 as n → ∞. Intuitively: "the probability that the sample mean is far from μ becomes arbitrarily small for large n". (2) The Strong Law of Large Numbers (Borel 1909, Kolmogorov) asserts almost sure convergence. Pr(X̄ₙ → μ) = 1 as n → ∞. Intuitively: "with probability 1, the sample mean converges to μ". The Strong Law is strictly stronger than the Weak Law: it implies the Weak, but not vice versa. For practical trading applications the distinction is academic — both guarantee your long-run average converges to the true EV. But both versions carry a critical caveat: they assume the variables have a finite mean. In distributions with extremely fat tails (Cauchy, for example, with infinite variance), the LLN can fail: the sample mean does not converge. In finance, as long as returns have finite variance (which is typically the case), the LLN applies, but convergence can be far slower than normal models suggest.
Direct Application in Trading
The LLN has practical daily implications for traders. (1) Patience with a statistical edge: if your strategy has an EV of +$20 per trade and you make 200 trades a year, annual EV is +$4,000, but the actual result in any specific year can land anywhere within a wide range. Convergence to EV requires many trades — typically hundreds or thousands — and the process is noisy. A good trader with a +$20 edge per trade can easily have three consecutive losing months from pure statistical variance. Understanding that is crucial to not abandoning correct strategies during drawdowns. (2) Diversification accelerates convergence: running multiple independent parallel trades raises the effective "n" and speeds convergence to the portfolio EV. Trading 1 trade a week with a +$20 edge is very different from 20 trades a week with the same edge — the second converges 20× faster to its aggregate EV. (3) Time needed for the edge to show: if the variance per trade σ² is $500, the standard error of the EV estimate after n trades is σ/√n. To have 95% confidence that your observed EV is within $5 of the true one with σ=500, you need n ≈ (1.96 × 500 / 5)² ≈ 38,416 trades. For $10 of precision: n ≈ 9,604. That is a surprisingly high number. It implies that most retail traders do not accumulate enough trades to reach statistical significance about their strategy; the one that "worked for 50 trades" proves nothing rigorously.
The Central Limit Theorem
Related to but distinct from the LLN is the Central Limit Theorem (CLT), which is arguably even more powerful. While the LLN describes where the sample mean converges (to μ), the CLT describes how it is distributed around μ for large n. The CLT says: for X₁, X₂, …, Xₙ i.i.d. with mean μ and variance σ² < ∞, the variable √n × (X̄ₙ − μ) / σ converges in distribution to a standard normal N(0, 1) as n → ∞. This is extraordinary: regardless of the shape of the original distribution (normal, uniform, exponential, bimodal, whatever), the sample average is approximately normally distributed for large n. Implications: (1) Confidence intervals: we can build them for μ based on X̄ₙ ± z × σ/√n; it is the basis of practically all inferential statistics. (2) Hypothesis tests: under the null hypothesis H₀: μ = μ₀, the statistic (X̄ₙ − μ₀) / (σ/√n) follows N(0,1) approximately. (3) Why the normal appears everywhere: averages of many independent variables (measurements, errors, returns across many trades) naturally converge to normal, which is why the normal distribution is so common. In trading, the CLT implies that the annual returns of a strategy with many trades tend to be approximately normal even if individual trade returns are highly non-normal. That makes modelling performance at the portfolio level much easier.
How Many Samples You Actually Need
A frequent but rarely rigorously answered question: how many trades do I need to prove I have an edge? The answer depends on the size of the edge relative to the variance. Using the CLT, the number of trades needed to statistically distinguish an edge μ from zero at 95% confidence and 80% power is approximately n ≈ 8 × (σ/μ)². Examples: (1) An edge of $15 per trade, σ=$200: n = 8 × (200/15)² = 1,422 trades. (2) An edge of $5 per trade, σ=$200: n = 8 × (200/5)² = 12,800 trades. (3) An edge of $20 per trade, σ=$100: n = 8 × (100/20)² = 200 trades. The practical conclusion is sobering: most traders do not have enough trades to demonstrate their edge with statistical significance. That has several implications: (a) do not trust track records of fewer than 200-500 trades; they are very probably noise; (b) systematic high-frequency traders — market makers, algorithmic desks — accumulate evidence quickly because they make thousands of trades a day; (c) long-term fundamental investors, with one to five decisions a year, will never reach statistical significance about their method; they depend on a qualitative thesis and on entry timing; (d) backtests can mislead: a 1,000-trade backtest looks robust but may cover a single market regime; you need data spanning several cycles.
LLN, Casinos and the Gambler’s Fallacy
The LLN explains perfectly why casinos always win. Every game has negative EV for the player by design: American roulette gives the house a 5.26% edge, blackjack without counting around 1%, slot machines between 5% and 15%. For each individual bet, the outcome is uncertain, but across thousands of bets a day and millions a year, the LLN guarantees the aggregate result converges to the house edge. A mid-sized casino generates over $100M in house edge annually through pure mathematics. Interestingly, the principle inverts exactly for the trader with a genuine edge: they make a casino of themselves, playing with statistical advantage on their side. Successful hedge funds (Medallion, SAC, Citadel) run millions of trades a year; the law guarantees that if they genuinely have an edge, it eventually materialises in money. A related paradox is the Gambler’s Fallacy: the cognitive bias of believing that after several losses, a win becomes "more likely" to "balance things out". That is false for i.i.d. variables — each event is independent of the last. If you have lost 10 coin flips in a row, the probability on the 11th is still 50/50 (assuming a fair coin). The LLN only guarantees the convergence of averages across large samples, not "compensation" in individual events. This bias destroys traders who increase size after losses (martingale) — eventually a large bad run wipes them out.