4.12.4 Tail Dependence & the t-Copula

Basic Prerequisites: 02 · Sklar’s Theorem & Copulas (dependence measures) and Extreme Value Theory & Fat Tails (why the tail is different).

---

1. Intuition & Practical Objective

Correlation measures the average dependence across the whole distribution. But a risk manager does not care about the average - she cares about the joint tail, the probability that two things go bad at the same time. This has its own measure, the coefficient of tail dependence $\lambda$, and it is not determined by correlation. The objective of this page is to make the central contrast concrete:

Two copulas with the same correlation can have completely different joint-tail behaviour. The Gaussian copula has $\lambda =0$ - it says joint extremes are impossible. The $t$-copula has $\lambda >0$ - it says joint extremes happen, and the strength is a function of the degrees of freedom.

This is the theoretical heart of the 2008 failure. The Gaussian copula does not slightly understate joint tails; it says they are asymptotically impossible, no matter how strong the correlation. As you push further into the tail - exactly where the bank’s survival is decided - the Gaussian copula’s joint-extreme probability decays to zero.

The replacement is the Student-$t$ copula. It keeps the elliptical, easy-to-simulate, correlation-parameterised structure everyone likes, but adds one parameter - the degrees of freedom $\nu$ - that controls tail dependence. Low $\nu$ means heavy tails and strong tail dependence. The $t$ copula is the standard repair for portfolio tail risk, and its calibration (correlation from Kendall’s $\tau$, $\nu$ by maximum likelihood) is routine.

---

2. Mathematical Ground Truth & Derivations

2.1 The tail-dependence coefficients

For random variables $X_1,X_2$ with CDFs $F_1,F_2$, the upper and lower tail-dependence coefficients are

$${\lambda }_u=\underset{q\to 1^{-}}{\lim }\mathrm{Pr}(X_2>F_2^{\leftarrow }(q)|X_1>F_1^{\leftarrow }(q)),{\lambda }_l=\underset{q\to 0^{+}}{\lim }\mathrm{Pr}(X_2\le F_2^{\leftarrow }(q)|X_1\le F_1^{\leftarrow }(q)).$$

They depend only on the copula (for continuous margins):

$${\lambda }_l=\underset{q\to 0^{+}}{\lim }\frac{C(q,q)}{q},{\lambda }_u=\underset{q\to 0^{+}}{\lim }\frac{\hat{C}(q,q)}{q},$$

where $\hat{C}$ is the survival copula, $\hat{C}(u,v)=u+v-1+C(1-u,1-v)$. For radially symmetric copulas ${\lambda }_l={\lambda }_u$.

2.2 The Gaussian copula is asymptotically independent ($\lambda =0$)

Let $(X_1,X_2)\sim N_2(0,P)$ with correlation $\rho$, and write $U_i=\Phi (X_i)$. Conditional on $X_1=x$, $X_2\mid X_1=x\sim N(\rho x,1-{\rho }^2)$, so

$$\lambda =2\underset{x\to -\infty }{\lim }\mathrm{Pr}(X_2\le x\mid X_1=x)=2\underset{x\to -\infty }{\lim }\Phi \left(x\sqrt{\frac{1-\rho }{1+\rho }}\right)=0(\rho <1).$$

Result. For every $\rho <1$, ${\lambda }_u={\lambda }_l=0$. However high the correlation, going far enough into the tail makes the two extremes independently rare.

2.3 The $t$-copula has genuine tail dependence ($\lambda >0$)

The bivariate $t$ copula is $C_{\nu ,\rho }^t(u,v)=t_{\nu ,P}(t_{\nu }^{-1}(u),t_{\nu }^{-1}(v))$. Conditional on $X_1=x$,

$$\sqrt{\frac{\nu +1}{\nu +x^2}}\frac{X_2-\rho x}{\sqrt{1-{\rho }^2}}\sim t_{\nu +1},$$

which yields the closed form

$$\boxed{\lambda =2t_{\nu +1}\left(-\sqrt{\frac{(\nu +1)(1-\rho )}{1+\rho }}\right)}$$

for $\rho >-1$. For fixed $\rho$, $\lambda$ increases as $\nu$ decreases; the $t$ copula is asymptotically dependent in both tails (radial symmetry). McNeil’s Table 7.1:

$\nu$	$\rho =0$	$\rho =0.5$	$\rho =0.9$
2	0.18	0.39	0.72
4	0.08	0.25	0.63
10	0.01	0.08	0.46
$\infty$ (Gauss)	0	0	0

Archimedean copulas give explicit, asymmetric tail dependence: Gumbel ${\lambda }_u=2-2^{1/\theta }$ (upper only), Clayton ${\lambda }_l=2^{-1/\theta }$ (lower only) - the subject of 06 · Advanced Extensions.

Simulation of the $t$ copula. Generate $\mathbf{Z}\sim N_d(0,P)$ and $W\sim {\chi }_{\nu }^2$ independently; set $\mathbf{X}=\mathbf{Z}\sqrt{\nu /W}$ and return $U_i=t_{\nu }(X_i)$. The chi-square mixing is what creates the joint extremes: a single small $W$ inflates all components together.

---

3. Computational Implementation - measuring tail dependence

Stdlib only. We (a) evaluate the closed-form $t$-copula $\lambda$ against McNeil’s Table 7.1, and (b) estimate the conditional tail probability directly from simulated Gaussian and $t$ copulas at increasing thresholds, showing the Gaussian decays toward $0$ while the $t$ copula stabilises at its analytic $\lambda$ (requires a Student-$t$ CDF, implemented via the regularized incomplete beta).

The analytic formula reproduces McNeil’s Table 7.1 to the published precision ($0.2532$ vs $0.25$; $0.6298$ vs $0.63$; $0.0819$ vs $0.08$). The simulation is the crucial picture: the Gaussian’s conditional tail probability falls as the threshold rises ($0.326\to 0.135\to 0.059$) - heading for its analytic limit of $0$ - while the $t$-copula’s stabilises near its analytic $\lambda =0.2532$. At the $99.9\%$ level the $t$ copula says an extreme in one name raises the chance of an extreme in the other to $26\%$; the Gaussian says $5.9\%$ and falling. That gap is the senior tranche.

Read the numbers as a warning. The Gaussian value at $q=0.99$ ($13.5\%$) is not small - so a Gaussian model does show meaningful tail co-movement at moderate quantiles. The failure is asymptotic: whatever threshold you pick, push further and the Gaussian’s joint-extreme probability keeps decaying. A risk system at a $99\%$ threshold looks fine; the same system at the capital-relevant $99.9\%$–$99.99\%$ is already diverging from reality.

---

4. Failure Modes & First-Principles Breakdowns

1. Subasymptotic $\ne$ asymptotic. The Gaussian copula can look tail-dependent at $95$–$99\%$ and still have $\lambda =0$. Backtests at ordinary quantiles cannot reveal the defect; only the extreme tail or a stress test can.
2. $\nu$ is as hard to estimate as the tail itself. Tail dependence is a limit at infinity; $\nu$ is inferred from finite data, usually from the same tail you are trying to measure. Small samples give wildly unstable $\hat{\nu }$ and $\hat{\lambda }$ - report intervals, not points.
3. Radial symmetry is a restriction. The $t$ copula has ${\lambda }_u={\lambda }_l$. Real markets are asymmetric: crashes co-move more than rallies (the lower-tail cluster that matters for a long book). Asymmetric copulas (skewed-$t$, Clayton for lower tail, mixture copulas) are required when the two tails differ.
4. One $\rho$, two tails. Calibrating a $t$ copula by Kendall’s $\tau$ fixes the average dependence; the tails are then set by $\nu$. If both are wrong, the joint tail is wrong for two independent reasons - which is why McNeil’s mixed/grouped copulas (Ch 7.3.3–7.3.4) and factor copulas exist.
5. Dependence is not the only tail input. The $t$-copula’s $\lambda$ assumes $t$ margins; with the wrong margins the joint tail is wrong even with the right copula. Fit the margins (EVT) and the copula.
6. $\lambda$ explains, it does not price. A single $\lambda$ summarises pairwise extremal dependence; a $d$-dimensional portfolio needs the full multivariate extreme-value structure (stable tail-dependence function), which is estimated with even greater difficulty (see EVT · 06).

---

5. References

• McNeil, Frey & Embrechts (2015)
• Joe, H. (1993, 1997)
• Coles, Heffernan & Tawn (1999), Dependence measures for extreme value analyses
• Embrechts, Klüppelberg & Mikosch (1997), Modelling Extremal Events
• de Haan & Ferreira (2006)
• Demarta & McNeil (2005), The t copula and related copulas

---

6. Connected Graph Bridges

• Back: 03 · The Gaussian Copula & 2008 · 02 · Sklar’s Theorem & Copulas · Index Hub
• Continue: 05 · Failure Modes & Practice · 06 · Advanced Extensions
• Sibling: EVT · 06 · Multivariate EVT & Copulas (the same tail-dependence machinery in the market-risk setting) · EVT & Fat Tails
• Application: Credit Risk · 06 · Portfolio Credit (the $t$-copula as the Gaussian repair in CDOs)