4.3 Extreme Value Theory & Fat Tails

Basic Prerequisites: Probability Theory and Value at Risk & Expected Shortfall. (these are the folder-level prerequisites for pages 02–06; page 01 states its own, smaller, entry requirements)

1. Intuition & Practical Objective

Financial returns are not normal. The Central Limit Theorem governs sums; risk management is about extremes. A Gaussian model says a 5-sigma daily move happens once in roughly every 1.7 million days (~6,500 years); real markets delivered exactly such moves on 1987-10-19, 1998 (LTCM), 2008-09-15, and repeatedly through 2020. Extreme value theory (EVT) is the branch of probability that studies the distribution of maxima and of threshold exceedances - and it is the principled replacement for “fit a normal to everything” when the quantity you care about lives in the tail.

This folder is the model topic-folder for the quantitative-risk build. It is a hub: it (a) gives you the fast formula lookup below (job #1), and (b) routes you to six sub-pages that walk from raw intuition through the stylized facts, the extreme-value theorems, the peaks-over-threshold method, the failure modes, and the extensions.

The one-sentence essence. “For the purposes of tail risk, every ‘well-behaved’ distribution looks like one of three limiting extreme-value laws, and every high-threshold exceedance looks Generalized Pareto - so estimate the single tail index $\xi$ and price extreme quantiles from it, instead of pretending the center of the distribution tells you about the 99.9th percentile.”

2. Mathematical Ground Truth & Derivations

Quick-Reference Lookup (job #1). All formulas below are transcribed from de Haan & Ferreira (2006) and McNeil, Frey & Embrechts (2015), and cross-checked against McNeil (1997, 2000); the numbers in the check column were re-executed and reproduced exactly from the verified corpus (see §3).

Notation: $X$ a loss/return variable, $F$ its CDF, $u$ a high threshold, $N_u$ the number of exceedances of $u$ out of $n$ observations, $\xi$ the shape parameter (extreme value index, tail index $\alpha =1/\xi$), $\beta >0$ a scale, $\mu$ a location.

The three extreme-value laws (Fisher–Tippett–Gnedenko, de Haan §1.1). For i.i.d. $X_i$ with normalized maxima converging to a non-degenerate $G$, $G$ must be the Generalized Extreme Value (GEV) family:

$\xi >0$ Fréchet (heavy/power-law tails) · $\xi =0$ Gumbel (light/exponential tails) · $\xi <0$ Weibull (bounded support). Financial losses sit in $\xi >0$.

Critical scaling caveat. There are two “tail index” conventions in the literature. de Haan writes the extreme value index $\gamma$ (= $\xi$ here, the GPD shape, can be any real); Hill (1975), Embrechts–Klüppelberg–Mikosch, and most econometrics (Tsay, McNeil–Frey 2000) write the tail index $\alpha =1/\gamma$, so a Student-t with $\nu$ dof has $\alpha =\nu$, $\xi =1/\nu$. A tail “alpha” near 3 means the 4th moment is infinite - kurtosis doesn’t exist. Never quote a tail exponent without the convention.

3. Computational Implementation - the EVT engine

This runs on the standard library only. It fits a GPD to the excesses above a threshold by maximum likelihood (profile likelihood: grid over $\xi$, scale solved by bisection on the score equation), then inverts the tail estimator into VaR and ES. It reproduces every verified number above.

EVT tracks the empirical extreme quantile on data with true $\xi =1/3$ - and does it while smoothing the single noisiest observation, which is exactly what a parametric tail model is for.

4. Failure Modes & First-Principles Breakdowns

Hub signposts - the folder’s full analysis lives in 05 · Failure Modes & Practice. In one line each:

1. Threshold choice = the whole method. Too low $u$ and central data contaminate the GPD (bias); too high $u$ and $N_u$ is tiny (variance explodes). This bias–variance tradeoff is the implementation issue in EVT (McNeil 1997 §4.4).
2. Small-sample tail estimation is intrinsically hard. Estimating the 99.9th percentile when you’ve seen ~2 losses that big is model-based extrapolation, not measurement; different thresholds can double the answer.
3. Regime dependence / dependence in the tail. Real extremes cluster (volatility clustering, contagion); the i.i.d. excesses assumption is violated, so unconditional EVT overreacts in stress periods - the motivation for McNeil–Frey’s GARCH-filtered EVT.

5. References

• de Haan, Laurens & Ana Ferreira: Extreme Value Theory: An Introduction (2006, Springer)
• McNeil, Alexander J.: Estimating the Tails of Loss Severity Distributions Using Extreme Value Theory, ASTIN Bulletin 27(1):117–137 (1997)
• McNeil, Alexander J. & Rüdiger Frey: Estimation of Tail-Related Risk Measures for Heteroscedastic Financial Time Series: An Extreme Value Approach, Journal of Empirical Finance 7(3–4):271–300 (2000)
• Embrechts, Klüppelberg & Mikosch: Modelling Extremal Events for Insurance and Finance (1997, Springer)
• McNeil, Frey & Embrechts: Quantitative Risk Management (2015, Princeton)
• Hill, Bruce M.: A Simple General Approach to Inference About the Tail of a Distribution, Annals of Statistics 3(5):1163–1174 (1975); Pickands, James III: Statistical Inference Using Extreme Order Statistics, Annals of Statistics 3(1):119–131 (1975); Balkema & de Haan: Residual Life Time at Great Age (1974). (PDFs in corpus; scanned/image-only.)
• Tsay, Ruey S.: Analysis of Financial Time Series

6. Connected Graph Bridges

• Foundational base: Probability Theory · VaR & Expected Shortfall (the risk measures EVT feeds)
• Sibling topic: Parametric, Historical & Monte Carlo VaR (the estimation methods EVT improves on)
• Sub-pages (in-folder): 01 From Zero · 02 Stylized Facts of Fat Tails · 03 Extreme Value Theory · 04 Peaks Over Threshold · 05 Failure Modes · 06 Advanced Extensions

Beginner: start at 01 · Practitioner: start at 05