Why Asymptotic Tail Dependence Matters
A dependence model can represent substantial association in ordinary observations yet still imply that extreme outcomes do not cluster in the limit. For risk management, this distinction matters: a model may give a positive probability of two losses exceeding high thresholds while implying that the conditional probability of one extreme, given another, approaches zero as the thresholds become more extreme.
The Gaussian copula is a standard way to join marginal distributions using a bivariate normal dependence structure. Its tail behavior follows from that structure, not from the particular marginal loss distributions. The key result is precise: for any nondegenerate bivariate Gaussian copula, both asymptotic upper and lower tail-dependence coefficients are zero. This does not mean the variables are independent, nor does it mean finite-threshold joint exceedances are negligible.
The Gaussian Copula and Its Tail Coefficients
Let \(Z_1,Z_2\) be standard normal random variables with correlation \(\rho\), where \(-1<\rho<1\). The Gaussian copula is \(C_\rho(u,v)=P(Z_1\le\Phi^{-1}(u),Z_2\le\Phi^{-1}(v))\), where \(\Phi\) is the standard normal distribution function. Equivalently, the copula variables are \(U=\Phi(Z_1)\) and \(V=\Phi(Z_2)\). The restriction \(-1<\rho<1\) makes the bivariate normal distribution nondegenerate.
For a common upper threshold \(q\), set \(t=\Phi^{-1}(q)\). Then \(U>q\) is equivalent to \(Z_1>t\), and \(V>q\) is equivalent to \(Z_2>t\). The conditional co-exceedance probability is therefore \(P(Z_1>t,Z_2>t)/\bar\Phi(t)\), where \(\bar\Phi(t)=1-\Phi(t)\).
Why the Upper-Tail Coefficient Is Zero
The proof uses a sum of the two latent normal variables. If both exceed \(t\), their sum must exceed \(2t\). The sum is normal with mean zero and variance \(2(1+\rho)\). This turns a two-variable event into a one-variable normal tail bound.
If \(Z_1>t\) and \(Z_2>t\), then \(Z_1+Z_2>2t\). Thus \(P(Z_1>t,Z_2>t)\le P(Z_1+Z_2>2t)\).
Since \(\operatorname{Var}(Z_1+Z_2)=1+1+2\rho=2(1+\rho)\), the standardized threshold is \(2t/\sqrt{2(1+\rho)}=at\), where \(a=\sqrt{2/(1+\rho)}\).
For \(\rho<1\), \(a>1\). Hence the conditional probability is bounded above by \(\bar\Phi(at)/\bar\Phi(t)\), which tends to zero as \(t\) tends to infinity.
For completeness, the last limit follows from the standard normal tail asymptotic \(\bar\Phi(x)\sim\phi(x)/x\) as \(x\to\infty\), where \(\phi\) is the standard normal density. Applying it to numerator and denominator gives
This proves that the upper-tail coefficient is zero: the conditional probability is nonnegative and bounded above by a quantity tending to zero. For the lower tail, use \((-Z_1,-Z_2)\). This pair is also standard bivariate normal with correlation \(\rho\), and a joint lower-tail event for \(Z_1,Z_2\) is a joint upper-tail event for \(-Z_1,-Z_2\). The same argument proves \(\lambda_L=0\).
The strict condition \(\rho<1\) is essential to the argument because it gives \(a>1\). At the degenerate endpoint \(\rho=1\), \(Z_1=Z_2\) almost surely and both coefficients equal one. That endpoint is not a nondegenerate Gaussian copula.
Worked Example: Bounding an Upper-Tail Co-Exceedance
A hypothetical portfolio model uses a Gaussian copula with latent standard normal variables \(Z_1,Z_2\) having correlation \(\rho=0.5\). Find an upper bound for the conditional probability that both copula variables exceed \(0.99\), given that the first exceeds \(0.99\). State what the bound does and does not establish.
Model and method. The latent vector is bivariate normal with means zero, variances one, and correlation \(0.5\). There is no interest-rate or mortality assumption in this copula probability calculation. Set \(q=0.99\), so \(t=\Phi^{-1}(0.99)\approx2.3263\), and \(\bar\Phi(t)=0.01\).
Apply the sum bound. Here \(a=\sqrt{2/(1+0.5)}=\sqrt{4/3}\approx1.1547\). Thus \(at\approx1.1547(2.3263)=2.6862\). The event that both latent variables exceed \(t\) implies that their sum exceeds \(2t\), so \(P(Z_1>t,Z_2>t)\le\bar\Phi(2.6862)\approx0.00361\).
Convert to the conditional probability. Divide by \(P(Z_1>t)=0.01\): \(P(Z_2>t\mid Z_1>t)\le0.00361/0.01\approx0.361\). The calculation gives an upper bound, not the exact conditional probability.
Check and interpretation. The bound is below one, as a conditional-probability bound must be. It shows that this method limits the conditional co-exceedance probability to approximately \(0.361\) at this threshold. It does not establish that the actual probability equals the bound, or that the probability is small enough for a particular capital decision.
Zero in the Limit Is Not Independence
Tail coefficients describe limiting behavior, not all dependence at high but finite thresholds. In a Gaussian copula, the latent variables remain dependent whenever \(\rho\ne0\). For any fixed \(q\) strictly between zero and one, the joint upper-tail probability can be positive even though its ratio to the marginal exceedance probability tends to zero as \(q\) approaches one.
A further practical point is that the convergence can be slow when \(\rho\) is close to one. The sum-bound multiplier \(a=\sqrt{2/(1+\rho)}\) is then only slightly greater than one. The bound still tends to zero eventually, but the theorem alone does not imply that co-exceedance probabilities at a business-relevant threshold are negligible.
Worked Example: A Lower-Tail Bound at a Finite Threshold
A hypothetical model of two related claim-count indicators uses a Gaussian copula with latent standard normal variables having correlation \(\rho=-0.5\). Calculate an upper bound for the probability that both copula variables fall below \(0.01\). Use the bound to assess the scale of the event probability; do not treat it as an exact probability.
Model and method. The latent vector is bivariate normal with means zero, variances one, and correlation \(-0.5\). Interest and mortality assumptions are not applicable. Since \(\Phi^{-1}(0.01)\approx-2.3263\), the lower-tail event is \(Z_1<-2.3263\) and \(Z_2<-2.3263\).
Reflect to an upper-tail event. Define \(W_1=-Z_1\) and \(W_2=-Z_2\). These are standard normal with correlation \(-0.5\), and the event becomes \(W_1>2.3263\), \(W_2>2.3263\). Their sum has variance \(2(1-0.5)=1\). If both exceed \(2.3263\), their sum exceeds \(4.6526\).
Bound the probability. Therefore \(P(Z_1<-2.3263,Z_2<-2.3263)\le\bar\Phi(4.6526)\approx0.00000164\). The calculation uses the standard normal upper tail at the displayed threshold; the bound is approximately \(1.64\times10^{-6}\).
Check and interpretation. The bound is smaller than the marginal lower-tail probability \(0.01\), as required. It is a conservative upper bound, not an estimate of the exact joint probability. The negative latent correlation makes this same-direction joint event particularly constrained, but the conclusion remains specific to this finite threshold and this model.
Worked Example: High Correlation and a Finite Loss Threshold
Consider two hypothetical continuous loss variables \(X\) and \(Y\) with unspecified continuous marginal distributions, joined by a Gaussian copula with latent correlation \(\rho=0.8\). Find a bound for the probability that \(Y\) exceeds its 99.5th percentile conditional on \(X\) exceeding its 99.5th percentile. Explain why this is not a tail-dependence coefficient.
Model and assumptions. The probability transforms of \(X\) and \(Y\) follow a bivariate Gaussian copula with standard normal latent variables and correlation \(0.8\). The loss margins are continuous, so exceeding each margin's 99.5th percentile corresponds to a copula variable exceeding \(q=0.995\). No interest or mortality assumptions are needed.
Find the latent threshold and sum-bound multiplier. The threshold is \(t=\Phi^{-1}(0.995)\approx2.5758\), and \(P(U>0.995)=0.005\). Also, \(a=\sqrt{2/(1+0.8)}=\sqrt{10/9}\approx1.0541\), so \(at\approx1.0541(2.5758)=2.7152\).
Calculate the bound. The joint exceedance event implies that the sum of the latent normals exceeds \(2t\). Therefore \(P(U>0.995,V>0.995)\le\bar\Phi(2.7152)\approx0.00331\). Dividing by the conditioning probability gives \(P(V>0.995\mid U>0.995)\le0.00331/0.005\approx0.662\).
Check and interpretation. The conditional bound is less than one. It may still be substantial at this finite threshold, despite the model's zero asymptotic upper-tail coefficient. The number \(0.662\) is neither the exact conditional probability nor a limiting coefficient; it is an upper bound obtained at \(q=0.995\).
Common Exam Trap: Correlation Is Not the Tail Coefficient
A common error is to take the latent normal correlation as the Gaussian copula's asymptotic tail-dependence coefficient. For example, with \(\rho=0.5\), the incorrect calculation is “\(\lambda_U=\rho=0.5\).” This confuses a dependence parameter with a limiting conditional probability.
The correction is to use the tail definition and the Gaussian sum argument: for \(-1<\rho<1\), \(a=\sqrt{2/(1+\rho)}>1\), and the upper bound on the conditional exceedance probability tends to zero. Thus \(\lambda_U=0\), even when \(\rho=0.5\). The finite-threshold conditional probability need not be zero; it simply cannot be identified with \(\rho\).
Check Your Understanding
For each problem, distinguish an asymptotic coefficient from a finite-threshold probability or bound. Unless stated otherwise, latent variables are standard bivariate normal with the specified correlation.
- A Gaussian copula has latent correlation \(\rho=0.25\). State its upper- and lower-tail-dependence coefficients and give the condition on \(\rho\) that supports the result. (Answer: \(\lambda_U=\lambda_L=0\), since \(-1<0.25<1\).)
- For a Gaussian copula with \(\rho=0\), find \(P(U>0.98,V>0.98)\). Explain why this case is independent. (Answer: \(0.02(0.02)=0.0004\); \(\rho=0\) makes the bivariate normal latent variables independent, so the copula variables are independent.)
- A Gaussian copula has \(\rho=-0.5\). At \(q=0.01\), use the normal-sum method to bound \(P(U\le q,V\le q)\). Use \(\Phi^{-1}(0.99)\approx2.3263\) and \(\bar\Phi(4.6526)\approx0.00000164\). (Answer: \(P(U\le0.01,V\le0.01)\le0.00000164\).)
- Two continuous hypothetical losses are joined by a Gaussian copula with \(\rho=0.8\). At \(q=0.995\), use \(t\approx2.5758\), \(a=\sqrt{10/9}\approx1.0541\), and \(\bar\Phi(2.7152)\approx0.00331\) to bound the conditional upper co-exceedance probability. Is the result the tail coefficient? (Answer: the conditional probability is at most \(0.00331/0.005\approx0.662\); no, it is a finite-threshold bound, while the upper-tail coefficient is zero.)