4.3.5 Failure Modes & Real-World Practice in EVT

Basic Prerequisites: 04 · Peaks Over Threshold.

---

1. Intuition & Practical Objective

EVT is mathematically beautiful and empirically delicate in three specific ways. This page names them precisely, so a practitioner knows which choice to distrust and how the failures show up in the numbers. The objective is not cynicism - it is the discipline of knowing exactly where an EVT estimate is an approximation, so the residual tail risk can be managed.

The three failures, in one line each:

1. Threshold choice is the whole method - a bias–variance tradeoff with no free lunch (McNeil 1997 §4.4).
2. Small-sample tail estimation is intrinsically hard - distant quantiles are extrapolation, and different thresholds can double the answer.
3. Financial extremes are dependent and regime-dependent - the i.i.d. excesses assumption fails because volatility clusters, so unconditional EVT misbehaves in stress periods (McNeil & Frey 2000).

---

2. Mathematical Ground Truth & Derivations

Where the assumptions live. Every POT result rests on four interlocking assumptions:

• (A1) A high enough threshold $u$ that the GPD limit theorem applies - the convergence $F_u\to G_{\xi }$ is only asymptotic in $u\to x_0$.
• (A2) Enough exceedances $N_u$ that $(\hat{\xi },\hat{\beta })$ can be estimated - the MLE is asymptotically normal only as $N_u\to \infty$ (McNeil 1997 §3.6, for $\xi >-1/2$).
• (A3) i.i.d. excesses - the excess distribution is written for independent draws; clustering breaks it.
• (A4) A stable tail regime - $\xi$ is assumed constant in the tail; real regimes (calm vs crisis) can carry different $\xi$.

The bias–variance identity that drives everything. For a fitted GPD, MSE $={\text{Bias}}^2+\text{Var}$. Raising $u$ shrinks bias (better validity of the limit theorem) but inflates variance (fewer exceedances). Lowering $u$ does the reverse. There is a true optimum in $k=N_u$ that depends on the second-order behaviour of $F$ - which is unknown. This is not an implementation bug; it is the intrinsic price of a limit theorem.

Why distant quantiles amplify threshold error. The EVT quantile is ${\hat{x}}_q=u+\frac{\hat{\beta }}{\hat{\xi }}\left[\right(\frac{n}{N_u}(1-q){)}^{-\hat{\xi }}-1]$. The factor $(\frac{n}{N_u}(1-q){)}^{-\hat{\xi }}$ is raised to a power involving $\hat{\xi }$, so small errors in $(\hat{\xi },u,N_u)$ compound geometrically as $q\to 1$. McNeil 1997 Table 1: the Danish .999-quantile varies 95→147 across thresholds 10→4, while the .995-quantile barely moves (40→46). The farther out the quantile, the more threshold-sensitive the answer.

---

3. Computational Implementation - the failures in numbers

Experiment 1 - threshold sensitivity (bias–variance). Fit the GPD shape to the same heavy-tailed sample at five thresholds. True $\xi =1/3$.

The signature shape: bias at low thresholds (85% pulls $\hat{\xi }$ down to 0.26, away from the true 0.333), a stable plateau (90–98%: 0.30–0.36), and variance at sparse thresholds (99%: only 199 exceedances, $\hat{\xi }$ jumps to 0.54). The practitioner reads the plateau and picks a threshold there - this is why the mean-excess plot / Hill-plot stability guidance exists (McNeil 1997 §4.1).

Experiment 2 - unconditional vs conditional EVT (the clustering failure). The McNeil–Frey (2000) lesson: fit EVT to raw returns (unconditional) and to volatility-filtered residuals (conditional); the conditional version responds to changing vol while the unconditional one is violated several times in a row in stress periods. We reproduce the direction of that result with EWMA filtering:

Two lessons at once. First, on filtered residuals the tail is thinner ($\hat{\xi }=0.08$, because volatility has absorbed the clustering) yet still heavy enough that EVT’s $z_{0.99}=3.10$ beats the normal’s $2.33$ - the conditional EVT VaR is 33% higher, matching McNeil & Frey’s finding that conditional normality systematically understates the tail. Second, this is why you filter first: the residuals are much closer to i.i.d. than raw returns, so the EVT limit theorem applies properly (06 · Advanced Extensions).

---

4. Failure Modes & First-Principles Breakdowns (numbered)

1. Threshold choice - the no-free-lunch tradeoff (Experiment 1). There is no objective “right” $u$; the true optimum depends on $F$‘s unobservable second-order term. In practice: read the stability plateau of $\hat{\xi }$ vs $u$ (or the mean-excess plot) and pick inside it; report the sensitivity, don’t hide it.
2. Small-sample, distant-quantile estimation. Estimating the 0.9999-quantile from a few thousand losses is extrapolation. McNeil 1997: the answer is threshold-dependent precisely because the target lies beyond the data. Treat distant-quantile EVT outputs as a range, not a point.
3. Dependence / clustering (Experiment 2). i.i.d. excesses fail under volatility clustering and contagion. Unconditional EVT is violated several times in a row in stress periods (McNeil & Frey 2000 backtests). Fix: filter volatility (GARCH/EWMA) before EVT, or model the extremal index / point-process formulation (06 · Advanced Extensions).
4. Regime dependence of $\xi$ itself. The tail index is not a law of nature - it shifts between calm and crisis regimes (and across assets). Fitting one $\xi$ to 20 years of returns mixes regimes; conditional/rolling EVT and GARCH-filtering are the standard responses.
5. Tail-index convention errors. $\xi$ vs $\alpha =1/\xi$ inversion, and $\hat{\xi }$ near/below $-\frac{1}{2}$ (MLE regularity) or $\ge 1$ (infinite mean, infinite ES) all silently corrupt the answer.

---

5. References

• McNeil (1997), Estimating the Tails of Loss Severity Distributions Using EVT, ASTIN 27:117–137
• McNeil & Frey (2000), Estimation of Tail-Related Risk Measures for Heteroscedastic Financial Time Series, JEF 7:271–300
• Embrechts, Klüppelberg & Mikosch (1997)
• Tsay, Analysis of Financial Time Series

---

6. Connected Graph Bridges

• Back: 04 · Peaks Over Threshold · Index Hub
• Forward: 06 · Advanced Extensions
• Sibling: Parametric, Historical & Monte Carlo VaR · VaR & Expected Shortfall · Stress Testing