The Central Limit Theorem — and Why Markets Break It

Almost every formula in this library leans on the normal distribution somewhere. This article is about where that reliance comes from — and exactly where it fails.

What the CLT says

Take independent, identically distributed random variables $X_1, X_2, \ldots$ with mean $\mu$ and finite variance $\sigma^2$. The Central Limit Theorem says their standardized sum converges to a standard normal:

$$\frac{\sum_{i=1}^{n} X_i - n\mu}{\sigma\sqrt{n}} \;\xrightarrow{d}\; N(0, 1)$$

Three things deserve emphasis, because interviews probe all three:

1. The individual $X_i$ can have any distribution — coin flips, die rolls, trade P&Ls — as long as the variance is finite. Normality is emergent, not assumed.
2. The scaling is $\sqrt{n}$. The sum's mean grows like $n$ but its noise grows like $\sqrt{n}$. This single asymmetry underlies diversification, why casinos win, and why market-making edge compounds.
3. Convergence is fastest in the middle of the distribution and slowest in the tails. The CLT is a statement about the center; it makes no promise about extreme events at any finite $n$.

The useful numbers

For a standard normal $Z$, memorize the tail ladder:

Threshold	$P(Z > z)$	"One in…"
$1\sigma$	15.9%	6
$2\sigma$	2.3%	44
$3\sigma$	0.13%	740
$4\sigma$	0.003%	32,000
$5\sigma$	0.00003%	3.5 million

Under normality, a $5\sigma$ daily move should happen about once in 14,000 years of trading. Equity markets produce them every few years.

Why returns are fat-tailed

The CLT's assumptions fail in markets in specific, nameable ways:

• Volatility is not constant. Returns are closer to a normal with randomly varying $\sigma$ — quiet regimes and violent regimes. A mixture of normals with different variances is itself fat-tailed: the big-$\sigma$ days dominate the tails. This is the single most important mechanism, and it's why GARCH-type models exist.
• Returns are not independent. Volatility clusters ("large moves follow large moves"), so extreme days arrive in bunches rather than diluting away.
• Feedback. Margin calls, stop-losses, and deleveraging make selling cause more selling. The system that generates returns changes its own parameters under stress — precisely when you're relying on the tail estimate.

The standard summary statistic is excess kurtosis: the normal has kurtosis 3; daily equity-index returns typically show 5–30. In plain terms, extreme outcomes are far more common than the normal predicts, and they matter more, because risk lives in the tails.

The interview version

Expect some form of: "Your risk model assumes normal returns. What's wrong with that, and does it matter?" A strong answer has three layers:

1. Where normality is fine: pricing and hedging decisions dominated by the distribution's center — small moves, short horizons, diversified books — because the CLT genuinely does its job there.
2. Where it's dangerous: anything that depends on tail probabilities — VaR at high confidence, out-of-the-money option pricing, leverage decisions, "how much can I lose in a crash." The normal doesn't just miss here; it misses by orders of magnitude.
3. What practitioners do about it: fatter-tailed distributions (Student-$t$), volatility models that let $\sigma$ move, stress tests that bypass distributional assumptions entirely, and — in options markets — the volatility smile, which is the market itself repricing the tails that Black–Scholes' normality understates.

That last point is worth savoring: the smile is empirical proof that the entire market agrees the normal model is wrong in the tails, and by how much.