4.3.1 Extreme Value Theory & Fat Tails from Zero

Basic Prerequisites: Probability Theory.

---

1. Intuition & Practical Objective

This page builds the why of extreme value theory with no prior EVT knowledge needed. The objective is one idea: the distribution of the center of the data tells you almost nothing about the distribution of its extremes - and extreme-value theory is the branch of statistics that models extremes directly, so that we can price and bound rare losses honestly.

Start with the dumbest question: why can’t we just use a normal distribution? Because a normal’s tail probability decays as $e^{-x^2/2}$ - absurdly fast. Under a normal with daily vol $\sigma =1\%$, a single-day $-6\sigma$ move ($-6\%$) has probability

$$P(X<-6\sigma )=\Phi (-6)\approx 1.0\times 10^{-9},$$

roughly once in a billion trading days (single-sided) - on the order of 4 million years at 250 trading days/yr. Real equity markets saw multiple $-6\sigma$ days in the twentieth century alone. Something is structurally wrong with the normal as a tail model even if it is a fine center model.

Three steps:

1. The CLT governs averages, not extremes. Sums of i.i.d. finite-variance variables converge to a normal - that’s why portfolio means look Gaussian. But the maximum of $n$ variables converges to a completely different set of laws. Risk management is about the maximum drawdown, the worst day, the tail of the loss distribution - extremes, not averages. Different question, different mathematics.

2. A fat tail is a power law, not a wide bell. A distribution is heavy-tailed when its survival function decays like a power of $x$, not an exponential:

$$1-F(x)\sim x^{-\alpha },x\to \infty ,$$

for a tail index $\alpha >0$. A Student-t with $\nu$ degrees of freedom has $\alpha =\nu$. If $\alpha <4$ the kurtosis is infinite; if $\alpha <2$ even the variance doesn’t exist. Financial returns consistently show $\alpha \approx 3$–$4$ - the fourth moment is infinite or borderline, which is why sample kurtosis jumps all over the place (see 02 · Stylized Facts).

3. Extremes have their own universal laws. Just as the CLT says “every finite-variance sum is eventually Gaussian,” two theorems say the same for extremes: the Fisher–Tippett–Gnedenko theorem (normalized maxima are Generalized Extreme Value) and the Pickands–Balkema–de Haan theorem (high-threshold exceedances are Generalized Pareto). You don’t need to know the parent distribution $F$ at all - you only need its tail index $\xi$. That is the entire practical promise of EVT.

---

2. Mathematical Ground Truth & Derivations

Why maxima stabilize - the intuition behind Fisher–Tippett–Gnedenko. Take $M_n=\max (X_1,\dots ,X_n)$. As $n$ grows, the maximum of a heavy-tailed $X$ is essentially its largest order statistic, which “pulls” the right endpoint. If we can find location/scale sequences $a_n,b_n$ such that $(M_n-b_n)/a_n$ converges to a non-degenerate law, that law must be the Generalized Extreme Value family (de Haan §1.1.2):

$$G_{\xi }(x)=\exp \{-(1+\xi x{)}^{-1/\xi }\},1+\xi x>0,\xi \in \mathbb{R},$$

with $\xi =0$ interpreted as the limit $G_0(x)=e^{-e^{-x}}$. The single parameter $\xi$ sorts every “reasonable” distribution into exactly one of three domains of attraction:

$\xi$	Name	Tail	Examples	Financial relevance
$\xi >0$	Fréchet	power-law $x^{-1/\xi }$	Student-t, Pareto, Cauchy	financial returns
$\xi =0$	Gumbel	exponential	normal, lognormal, gamma	central-limit-land
$\xi <0$	Weibull	bounded	uniform, beta	losses with a hard cap

Heavy tails survive block maxima; light tails wash out. This is the cleanest way to see the Fréchet vs Gumbel split. Take block maxima (say, the max of every 100 observations). If the parent is normal (light tail), the block maxima have an even lighter, Gumbel-like tail. If the parent is Student-t (heavy tail), the block maxima keep the same tail index $\alpha$ - the fat tail is preserved in the extremes. We verify this numerically in §3.

Why $\xi$ (not the whole distribution) is what matters. The GEV/GPD convergence results hold for a huge class of parents; the only free quantity that survives is the tail index $\xi$. This is why EVT is called “the robust statistics of tails”: you estimate one parameter from the tail observations, and the entire extreme-quantile machinery follows (04 · Peaks Over Threshold).

---

3. Computational Implementation - block maxima: heavy tails survive, light tails don’t

Direct proof of the intuition in §2: compute block maxima of 100 from normal, t₃, and t₆ parents, then estimate the tail index of those maxima. The heavy-tailed parents keep their tail index ($\alpha =\nu$); the normal’s maxima are far lighter (huge $\alpha$). Stdlib only.

The t₃ block maxima estimate $\alpha \approx 3.0$ - the true tail index of the parent is preserved in the maxima. The normal’s maxima show a much larger (lighter) tail index. Fat tails are a property that survives aggregation to extremes; thin tails are not.

---

4. Failure Modes & First-Principles Breakdowns

1. Applying the CLT to the tail. The single most common risk-management error: treat the loss distribution’s mean as if it told you about its 99.9th percentile. The CLT governs sums; the extremes obey GEV/GPD. Using normality at the 99.9% tail of real returns understates VaR by a large factor (see 04 · Peaks Over Threshold for the number).
2. “Fat tail” ≠ “wide normal.” A heavy tail is a power-law survivor function, not merely a higher kurtosis. Kurtosis is itself undefined when $\alpha <4$, so “estimate the kurtosis then adjust” is chasing a quantity that may not exist.
3. Conflating the two tail-index conventions. $\xi$ (extreme value index / GPD shape) vs $\alpha =1/\xi$ (tail index). A textbook that says “alpha $=3$” and one that says “xi $=1/3$” mean the same thing. Get this wrong and you invert the whole analysis (see the index hub’s scaling caveat).

---

5. References

• de Haan & Ferreira, Extreme Value Theory: An Introduction (2006)
• Embrechts, Klüppelberg & Mikosch, Modelling Extremal Events for Insurance and Finance (1997)
• McNeil, Frey & Embrechts, Quantitative Risk Management (2015)

---

6. Connected Graph Bridges

• Back: Index Hub
• Forward: 02 · Stylized Facts · 03 · Extreme Value Theory
• Base: Probability Theory · VaR & Expected Shortfall