The Central Limit Theorem

The central limit theorem is the reason the normal distribution is everywhere: whenever an outcome is the sum of many small, independent, comparably-sized shocks, its distribution is approximately Gaussian regardless of what the individual shocks look like. This is simultaneously the most useful and the most abused theorem in quantitative finance, useful because it justifies every $t$-statistic and confidence interval, abused because financial shocks are neither independent nor comparably sized, and the theorem's failure is exactly where tail risk hides.

The classical statement

Let $X_1, X_2, \dots$ be i.i.d. with mean $\mu$ and finite variance $\sigma^2 > 0$. The The Law of Large Numbers says $\bar X_n \to \mu$; the CLT describes the fluctuation around that limit. Standardising the sum, $Z_n = \frac{\sum_{i=1}^n X_i - n\mu}{\sigma\sqrt{n}} = \frac{\bar X_n - \mu}{\sigma/\sqrt n} \;\xrightarrow{d}\; \mathcal{N}(0,1),$ where $\xrightarrow{d}$ is convergence in distribution: $\mathbb{P}(Z_n \le z) \to \Phi(z)$ for every $z$. The $\sqrt n$ scaling is forced, it is the exact rate at which the LLN error shrinks, and the limit is universal, depending on the summands only through their variance. That universality is the miracle: the shape of the individual $X_i$ is completely forgotten.

Proof sketch via the MGF

Assume the Moment Generating Functions exists. Centre and scale to $Y_i = (X_i - \mu)/\sigma$, so $\mathbb{E}[Y_i]=0$, $\operatorname{Var}(Y_i)=1$, and $Z_n = \frac{1}{\sqrt n}\sum Y_i$. By independence the MGF multiplies: $M_{Z_n}(t) = \left[M_Y\!\left(\frac{t}{\sqrt n}\right)\right]^n.$ Expand the inner MGF around $0$ using $M_Y(0)=1$, $M_Y'(0)=0$, $M_Y''(0)=1$: $M_Y\!\left(\frac{t}{\sqrt n}\right) = 1 + \frac{t^2}{2n} + o\!\left(\frac1n\right).$ Therefore $M_{Z_n}(t) = \left(1 + \frac{t^2}{2n} + o(1/n)\right)^n \xrightarrow{n\to\infty} e^{t^2/2},$ which is exactly the standard-normal MGF. By the uniqueness of MGFs, $Z_n \Rightarrow \mathcal{N}(0,1)$. The cleanest way to see what is happening is through cumulants: standardising kills the mean, fixes the variance at $1$, and the $\sqrt n$ scaling drives every cumulant of order $\ge 3$ to zero (see Moment Generating Functions), leaving only the two cumulants a Gaussian has.

When it applies: Lindeberg and Lyapunov

The i.i.d. version is a special case. Real applications sum non-identical independent variables (heteroskedastic returns, varying position sizes), and the general CLT needs a condition ensuring no single term dominates the sum.

Lindeberg condition. For independent (not necessarily identical) $X_k$ with variances $\sigma_k^2$ and $s_n^2 = \sum_{k=1}^n \sigma_k^2$, if for every $\varepsilon > 0$ $\frac{1}{s_n^2}\sum_{k=1}^n \mathbb{E}\big[(X_k-\mu_k)^2\,\mathbf{1}\{|X_k-\mu_k| > \varepsilon s_n\}\big] \to 0,$ then the standardised sum is asymptotically normal. In words: the total variance contributed by "large" deviations is asymptotically negligible.

Lyapunov condition is a simpler sufficient version: for some $\delta > 0$, $\frac{1}{s_n^{2+\delta}}\sum_{k=1}^n \mathbb{E}\big[|X_k-\mu_k|^{2+\delta}\big] \to 0.$ Both encode the same moral: the CLT holds when the sum is genuinely a blend of many comparable contributions and fails when one term can dominate, precisely the fat-tailed regime where a single jump carries the day.

The rate: Berry–Esseen

Convergence in distribution says nothing about how close $Z_n$ is to normal at finite $n$. The Berry–Esseen theorem bounds it: if $\rho = \mathbb{E}|X_i - \mu|^3 < \infty$, $\sup_z \big|\mathbb{P}(Z_n \le z) - \Phi(z)\big| \le \frac{C\,\rho}{\sigma^3 \sqrt n}, \qquad C < 0.48.$ The error decays like $1/\sqrt n$ and, critically, scales with the skewness $\rho/\sigma^3$: the more skewed the summands, the slower the convergence. This is why the "$n \ge 30$" rule of thumb is a fiction for asymmetric data; convergence in the centre is quick but the tails converge slowest of all, which is catastrophic for finance because risk lives in the tails, not the centre.

Worked example, a portfolio of independent bets

You hold $n = 100$ independent positions, each with expected P&L \mu = \50$and standard deviation$\sigma = $400. Individual outcomes may be wildly non-normal (a binary event, a skewed payoff), but the *aggregate* book P&L is, by the CLT, approximately $$\text{Total} \approx \mathcal{N}\big(n\mu,\, n\sigma^2\big) = \mathcal{N}(5000,\ 100\cdot 400^2) = \mathcal{N}(5000,\ 4000^2).$$ So a 95% interval is roughly 5000 \pm 1.96\cdot 4000 = [-2840,\ 12840]$, and the probability of a losing day is$\Phi(-5000/4000) = \Phi(-1.25) \approx 10.6%$. Note the aggregate standard deviation ($$4{,}000$) is only$10\times$a single position's, not$100\times$, diversification's$\sqrt n$benefit, the same$\sqrt n$ as the CLT. This is exactly how desks reason about book-level risk from position-level statistics.

Failure modes and subtleties

• Finite variance is mandatory. For heavy tails with $\operatorname{Var} = \infty$ (α-stable with index $< 2$), normalised sums converge to a stable law with power tails, not a Gaussian. The generalised CLT (Gnedenko–Kolmogorov) covers this regime, and it is arguably the right one for financial returns. Assuming Gaussianity here understates crash probability by orders of magnitude.
• Dependence breaks it. Volatility clustering and autocorrelation violate independence; returns arrive in correlated bursts, so a "many independent shocks" story is false at short horizons. Mixing/martingale CLTs recover normality under weak dependence, but a market crash is one common shock hitting everything, the opposite of Lindeberg's "no dominant term".
• Tails converge last. Even when the CLT ultimately applies, at finite $n$ the approximation is good in the middle and poor 3+ standard deviations out, the exact region VaR and stress tests care about. Berry–Esseen quantifies this; skew makes it worse.
• The theorem is about sums, not products. Multiplicative processes (compounded returns) are normal in the log, that is the CLT applied to log-returns, producing lognormal prices (see Common Distributions), not normal prices.

In interviews

The single most common CLT question is a sanity check that you know it needs finite variance and gives a $\sqrt n$ scaling, candidates who say "averages are always normal" fail. Be able to sketch the MGF proof and to state the limit as $\mathcal{N}(0,1)$ for the standardised sum. A strong differentiator is naming when the CLT fails: infinite variance (stable laws), dependence (clustering), and slow tail convergence (Berry–Esseen, skewness). If the desk is quantitative-risk-focused, connect it to why Gaussian VaR under-reserves, the CLT's centre is fine but its tails are a lie for real returns. This is the theorem that makes Ordinary Least Squares (OLS) inference asymptotically valid, so expect it to surface whenever standard errors and $t$-stats do.