Law of Large Numbers
The Law of Large Numbers (LLN) is a fundamental theorem of probability stating that as the number of independent, identically distributed random trials increases, the sample mean of the observations converges to the true population (expected) mean. In finance, it underpins the statistical validity of using historical average returns as estimates of true expected returns and is central to the logic of diversification.
Key takeaways
- The Weak LLN states that the sample mean converges in probability to the population mean; the Strong LLN states that convergence occurs almost surely (with probability 1).
- The LLN justifies the use of historical average returns as estimates of expected returns—but only when returns are drawn from a stationary distribution, a condition that often fails in financial markets.
- In insurance and risk pooling, the LLN is the mathematical foundation for the principle that independent risks diversify away when combined in large numbers.
- The law does NOT apply to sequences of outcomes that are not independent or not identically distributed, which limits its applicability to financial time series exhibiting serial correlation or regime changes.
- The LLN is often confused with the Gambler's Fallacy: the law describes the long-run average, not that short-run deviations are 'corrected' by future outcomes.
Explanation
The Law of Large Numbers exists in two forms. The Weak Law of Large Numbers (WLLN), proven by Jakob Bernoulli in 1713, states that for independent, identically distributed (i.i.d.) random variables X₁, X₂, ..., Xₙ with finite mean μ, the sample mean X̄ₙ = (X₁ + ... + Xₙ)/n converges in probability to μ as n approaches infinity: for any ε > 0, P(|X̄ₙ − μ| > ε) → 0 as n → ∞. The Strong Law of Large Numbers (SLLN) provides a stronger guarantee: the sample mean converges to μ almost surely, meaning P(lim_{n→∞} X̄ₙ = μ) = 1.
In finance and economics, the LLN is invoked in numerous contexts. In portfolio theory, it provides the mathematical justification for diversification: if individual stock returns are approximately i.i.d. (or at least uncorrelated), then the variance of the portfolio average return decreases as the number of stocks increases, with the portfolio return converging to the expected return of the average stock. This is why a well-diversified portfolio eliminates idiosyncratic (firm-specific) risk while retaining systematic (market) risk.
In insurance and actuarial science, the LLN is foundational. An insurer cannot predict whether any specific policyholder will file a claim, but with tens of thousands of independent policyholders, the actual claim rate converges closely to the expected claim rate, allowing premiums to be priced accurately. The same logic applies to credit card issuers, mortgage lenders, and any business where risk is pooled across many independent counterparties.
The LLN also has important implications for empirical finance. Historical average returns are the standard estimator for expected returns in mean-variance optimization (e.g., the Markowitz framework). The LLN assures that these historical averages converge to true expected returns as the sample size grows—but only under the i.i.d. assumption. In practice, stock returns exhibit time-varying volatility (GARCH effects), fat tails, and structural breaks (the pre-2008 return distribution differs from the post-crisis period), which violate the i.i.d. assumption and limit the precision of historical return estimates even with large samples. Research by Merton (1980) showed that even a century of data may be insufficient to estimate expected equity returns precisely enough to be useful in portfolio optimization.
The LLN is often misapplied via the Gambler's Fallacy, which holds that after a sequence of bad outcomes, future outcomes must be good to 'balance out.' The LLN makes no such statement: it describes the long-run average of many independent trials, not a correction mechanism for short-run deviations. Each new observation is independent of the history, and the law simply states that the accumulating average will converge to μ.
Formula
X̄ₙ = (1/n) Σᵢ Xᵢ → μ as n → ∞ (where μ = E[X])
Example
A quantitative analyst estimates the expected annual return of the S&P 500 using historical data. With 10 years of annual returns, the standard error of the mean estimate is σ/√10, where σ ≈ 15% (historical standard deviation). This gives a standard error of about 4.7%, meaning the estimated expected return of, say, 7% has a 95% confidence interval of roughly 7% ± 9.4%. Extending the sample to 50 years narrows the standard error to 2.1%, and a 100-year sample narrows it further to 1.5%. The LLN tells us the estimate converges to the true mean—but the convergence is slow for high-variance series. This illustrates why estimating expected returns with precision requires either very long sample periods or additional structure such as economic models.
Related terms
Compound Interest Continuous Compounding Convergence Diversification Equity Fat Tails Future Value Jensens Inequality Mean Variance Optimization Normal Distribution Portfolio Optimization Standard Deviation