4.3.6 Advanced Extensions

Basic Prerequisites: 04 · Peaks Over Threshold and 05 · Failure Modes.

---

1. Intuition & Practical Objective

The two real-world defects of textbook EVT - (a) extreme returns are dependent/regime-driven, and (b) risk is a portfolio problem in many dimensions - motivate the extensions on this page. The objective is the launchpad: show the two extensions that matter most in practice, filtered (conditional) EVT and multivariate EVT via copulas, and hand off the deeper machinery to the corpus monographs.

Why these two? Filtered EVT fixes the i.i.d. violation of the POT method in one step: fit a volatility model, standardize the residuals, apply EVT to the (near-i.i.d.) residual tail, and scale back by current volatility. This is the McNeil–Frey (2000) method, and it beat conditional normality and unconditional EVT in their backtests. Multivariate EVT matters because a bank’s risk is a book of correlated positions, and Gaussian-copula dependence understates the joint probability that several positions crash together - the exact failure behind 2008.

---

2. Mathematical Ground Truth & Derivations

2.1 Filtered / conditional EVT (McNeil & Frey 2000)

Model returns as a heteroscedastic process

$$X_t=\mu +{\sigma }_t{Z}_t,Z_t\text{ i.i.d. heavy-tailed},$$

where ${\sigma }_t$ follows GARCH(1,1) ${\sigma }_t^2=\omega +\alpha X_{t-1}^2+\beta {\sigma }_{t-1}^2$ (Tsay Ch 3). The two-step method:

1. Filter: estimate $\hat{\mu },{\hat{\sigma }}_t$ (GARCH or EWMA), form standardized residuals ${\hat{Z}}_t=(X_t-\hat{\mu })/{\hat{\sigma }}_t$.
2. Tail: fit a GPD to the residual tail and estimate the residual quantile ${\hat{z}}_q$ (the EVT machinery of 04 · POT).
3. Rescale: one-day conditional VaR $={\hat{\mu }}_{t+1}+{\hat{\sigma }}_{t+1}{\hat{z}}_q$; conditional ES $={\hat{\mu }}_{t+1}+{\hat{\sigma }}_{t+1}\mathbb{E}[Z\mid Z>{\hat{z}}_q]$, with $\mathbb{E}[Z\mid Z>z_q]=\frac{\hat{\beta }+\hat{\xi }{z}_q}{1-\hat{\xi }}$ (McNeil & Frey eq. 14).

Why it works. Standardized residuals are far closer to i.i.d. than raw returns, so the EVT limit theorem and the excesses-independence assumption hold (Embrechts–Klüppelberg–Mikosch Ch 5; McNeil & Frey §2.2 cite the AR(1) example where residual-based Hill estimates are far more stable than raw-data Hill). McNeil & Frey’s backtests (S&P, DAX, BMW, $/$£, Gold): conditional EVT was correct in 11/15 cases and never rejected, while conditional normal failed 11 times and unconditional EVT misread stress periods.

2.2 Multivariate EVT & copulas

Univariate EVT answers “how bad can one risk get?” Portfolio risk asks “how likely do several risks get bad together?” The dependence structure is captured by a copula - a function $C:[0,1{]}^d\to [0,1]$ with uniform margins that couples them. Sklar’s theorem: any joint CDF $F$ with margins $F_i$ writes as $F(x_1,\dots ,x_d)=C(F_1(x_1),\dots ,F_d(x_d))$.

The relevant quantity for joint extremes is the upper tail dependence coefficient

$${\lambda }_u=\underset{q\to 1}{\lim }\mathbb{P}(U_1>q\mid U_2>q),$$

the probability that one variable is extreme given another is. Key facts:

• Gaussian copula: ${\lambda }_u=0$ for $\rho <1$ - extremes are asymptotically independent even for strong correlation. This is the theoretical reason a Gaussian-copula risk model says “two assets never crash together” and why it failed in 2008.
• $t$-copula (with low dof): ${\lambda }_u>0$ - it has genuine tail dependence, which is why it’s preferred for portfolio tail risk.
• Archimedean copulas (Clayton, Gumbel): explicit tail dependence, used in credit (the one-factor models behind CreditMetrics/Vasicek, Credit Risk & Merton).

Multivariate EVT generalizes the univariate limit: the joint tail converges to a multivariate GEV/Pareto whose dependence is summarized by a stable tail dependence function or Pickands dependence function (de Haan Ch 6–7; McNeil–Frey–Embrechts Ch 7). The practical message: fit univariate EVT tails on each margin, choose a copula with the right tail-dependence sign, and glue them together - never assume independence or Gaussianity in the joint tail.

---

3. Computational Implementation - the Gaussian copula understates joint extremes

Direct, stdlib demonstration of §2.2: sample the bivariate Gaussian copula at two correlations and measure the joint probability that both ranks exceed 95%, and the conditional probability one is extreme given the other. (In the doc we read the copula’s own ranks; any fat-tailed margins can be attached.)

At the 95% level the high-correlation copula shows subasymptotic tail dependence ($0.39$ conditional vs $0.05$ independent). But here is the trap the theory warns about: this conditional probability decays to zero as the threshold $\to 1$ for the Gaussian copula - it is asymptotically tail-independent. A risk model using it will report “joint crashes are impossible” exactly at the extreme level where a bank must not hear that. This is why the $t$-copula (positive ${\lambda }_u$) and EVT-based tail dependence replace it for portfolio tail risk.

Filtered EVT in one line (the McNeil–Frey method, implementation and numbers in 05 · Failure Modes, Experiment 2): GARCH/EWMA-filter the returns, fit the GPD to the residual tail, rescale the residual quantile by current volatility - reproducing McNeil & Frey’s finding that conditional EVT VaR (0.710) beats conditional normal (0.532) on heavy-tailed GARCH data.

---

4. Failure Modes & First-Principles Breakdowns

1. Gaussian-copula tail blindness. For any $\rho <1$ the Gaussian copula is asymptotically tail-independent (${\lambda }_u=0$) - it guarantees “no joint crashes” in the extreme tail. If your model uses it, the fat left tail of a portfolio is systematically understated. Use a $t$-copula or EVT dependence for tail risk.
2. Copula + marginal mismatch. A copula carries no information about the margins - you must model the heavy-tailed margins separately (univariate EVT), or the joint tail is wrong even with a “correct” copula.
3. Filtering model risk. The GARCH/EWMA filter must be right; a misspecified volatility model contaminates the residuals and hence the EVT tail. Validate the residuals are near-i.i.d. (ACF of squares) before fitting.
4. Estimation stability in multivariate EVT. Multivariate dependence functions and the associated parameters are even harder to estimate than univariate $\xi$ - small samples and high dimensions make them fragile (de Haan Ch 6–7). Report sensitivity, especially for ${\lambda }_u$.
5. Regime shifts in $\xi$ and dependence. Both the tail index and the dependence can change across calm/crisis regimes; a single static copula or tail index mixes regimes (see 05 · Failure Modes).

---

5. References

• McNeil & Frey (2000), Estimation of Tail-Related Risk Measures for Heteroscedastic Financial Time Series, JEF 7:271–300
• de Haan & Ferreira (2006)
• McNeil, Frey & Embrechts (2015)
• Embrechts, Klüppelberg & Mikosch (1997)
• Tsay, Analysis of Financial Time Series

---

6. Connected Graph Bridges

• Back: 05 · Failure Modes · Index Hub
• Portfolio risk: VaR & Expected Shortfall · Parametric, Historical & Monte Carlo VaR
• Credit (fat-tailed joint default): Credit Risk & Merton · Stress Testing
• Base: Stochastic Calculus (GARCH dynamics) · Probability Theory (copulas)