4.3.3 Extreme Value Theory

Basic Prerequisites: Probability Theory and 02 · Stylized Facts.

---

1. Intuition & Practical Objective

This page is the mathematical spine of extreme value theory: the Fisher–Tippett–Gnedenko theorem (which law governs sample maxima), the Generalized Extreme Value distribution, and the two workhorse estimators of the tail index - the Hill estimator and the Pickands estimator. The practical objective: estimate the single parameter $\xi$ (the extreme value index) that determines the entire extreme-quantile machinery, and understand which estimator to trust when.

The intellectual move mirrors the CLT: the CLT says sums of i.i.d. finite-variance variables are always eventually normal, whatever the parent. Fisher–Tippett–Gnedenko says maxima of i.i.d. variables are always eventually GEV, whatever the parent - with the single shape parameter $\xi$ inherited from the parent’s tail. You never need to know $F$; you need only $\xi$. That is the power and the economy of EVT.

---

2. Mathematical Ground Truth & Derivations

2.1 Fisher–Tippett–Gnedenko (de Haan §1.1)

Theorem. Let $X_1,\dots ,X_n$ be i.i.d. and $M_n=\max (X_1,\dots ,X_n)$. If there exist sequences $a_n>0$, $b_n$ such that $(M_n-b_n)/a_n$ converges in distribution to a non-degenerate $G$, then

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

with $G_0(x)=e^{-e^{-x}}$ for $\xi =0$. This is the GEV family. The parent $F$ is said to be in the maximum domain of attraction of $G_{\xi }$, written $F\in \text{MDA}(G_{\xi })$.

• $\xi >0$: Fréchet (heavy tails, $1-F(x)=x^{-1/\xi }L(x)$ slowly varying) - the financial case.
• $\xi =0$: Gumbel (light tails, exponential decay).
• $\xi <0$: Weibull (bounded support).

With location $\mu$ and scale $\sigma$ inserted: $G_{\xi ,\mu ,\sigma }(x)=G_{\xi }((x-\mu )/\sigma )$.

2.2 Tail-index estimation - why the order statistics

For $\xi >0$, $1-F(x)=x^{-1/\xi }L(x)$; the quantity $\alpha =1/\xi$ is the tail index. Estimating $\xi$ means using the largest order statistics. Let $X_{(1)}\ge X_{(2)}\ge \cdots \ge X_{(k)}\ge X_{(k+1)}$ be the $k$ largest observations above an intermediate threshold.

Hill estimator (de Haan §3.2.2, eq. 3.2.2; Hill 1975). Estimates $\gamma =\xi$ from the mean of the log-spacings in the tail:

$${\hat{\gamma }}_H=\frac{1}{k}\sum _{i=1}^k\log \frac{X_{(i)}}{X_{(k+1)}},$$

so $\hat{\alpha }=\frac{k}{{\sum }_{i=1}^k\log (X_{(i)}/X_{(k+1)})}$. Hill is consistent only for $\xi >0$ (de Haan Thm 3.2.2) - it is the estimator of choice for fat-tailed financial data.

Pickands estimator (de Haan §3.3.1, eq. 3.3.1; Pickands 1975). Works for any real $\xi$:

$${\hat{\gamma }}_P=\frac{1}{\log 2}\log \frac{X_{(n-k)}-X_{(n-2k)}}{X_{(n-2k)}-X_{(n-4k)}},$$

where $X_{(n-k)}$ is the $(k+1)$-th largest, etc. It estimates $\gamma$ via quantiles of the limiting GEV, in contrast to Hill’s moment-of-the-tail approach (de Haan Remark 3.3.4). Its virtue is validity for all $\xi$; its cost is high variance for $\xi >0$ (de Haan §3.3, compare of asymptotic variances §3.4) - precisely the regime of interest.

Rule of thumb from de Haan §3.4. For heavy tails ($\xi >0$) the Hill estimator has the lowest asymptotic variance among these; the MLE of the GPD (04 · Peaks Over Threshold) and the “negative Hill” estimator of Falk are even more efficient, while the Pickands estimator is the most variable. Use Hill for $\xi >0$; use Pickands only when you cannot assume heavy tails.

Threshold / $k$ choice. Both estimators require choosing $k$ (how many upper order statistics to use). Too small $k$: high variance. Too large $k$: bias, because the second-order term of $F$ pollutes the fit. McNeil & Frey (2000) simulation: for n=1000 t₄ samples, Hill is efficient only for $k\lesssim 70$, and deteriorates rapidly beyond; the GPD-based estimator is far more robust to $k$ (see 05 · Failure Modes).

---

3. Computational Implementation - Hill & Pickands on a known heavy tail

Fit the tail index of a Student-t₄ (true $\alpha =4$, so $\xi =0.25$) using both estimators. Stdlib only.

Read the output honestly: Hill is accurate and stable ($\hat{\alpha }\approx 3.3$–$3.8$ around the true $4$, gently degrading as $k$ grows - the bias-variance tradeoff in action). Pickands is erratic ($\hat{\xi }=-0.11$ to $+0.12$ against a true $0.25$) - exactly the high variance for $\xi >0$ that de Haan §3.4 warns about. On a single sample Pickands can even give the wrong sign of the tail; use it only when you cannot assume heavy tails.

---

4. Failure Modes & First-Principles Breakdowns

1. Hill is only consistent for $\xi >0$. Its defining identity (de Haan eq. 3.2.1) assumes a power-law tail. Applied to Gumbel (light) or Weibull (bounded) data it misbehaves (de Haan §3.5); if you don’t know the tail is heavy, use Pickands or the GPD MLE.
2. $k$-sensitivity is structural, not a bug. Hill’s estimate must move with $k$: small $k$ = variance, large $k$ = bias from the second-order term. The McNeil–Frey (2000) simulation shows Hill’s efficient $k$-range is narrow ($k\lesssim 70$ for n=1000), so “just take more order statistics” is wrong. Use a Hill plot (estimate vs $k$) and read the stable plateau.
3. Convention error: $\xi$ vs $\alpha$. The Pickands/Hill estimators estimate $\gamma =\xi$ directly; reporting $\hat{\gamma }$ as a “tail index” without noting $\alpha =1/\gamma$ inverts the intuition ($\alpha =3$ and $\xi =1/3$ are the same tail).
4. Pickands’ variance. Its quantile-based construction (de Haan Remark 3.3.4) makes it the least efficient estimator in the heavy-tail regime - it is not a good default for fat-tailed financial returns.

---

5. References

• de Haan & Ferreira, Extreme Value Theory: An Introduction (2006)
• Hill, Bruce M., A Simple General Approach to Inference About the Tail of a Distribution, Annals of Statistics 3(5):1163–1174 (1975). (Corpus PDF, scanned.)
• Pickands, James III, Statistical Inference Using Extreme Order Statistics, Annals of Statistics 3(1):119–131 (1975). (Corpus PDF, scanned.)
• McNeil & Frey (2000)

---

6. Connected Graph Bridges

• Back: 02 · Stylized Facts · Index Hub
• Forward: 04 · Peaks Over Threshold (the GPD/POT method that builds on this tail index)
• Base: Probability Theory · Stochastic Calculus