Gumbel Copulas and Upper-Tail Dependence

Why Upper-Tail Dependence Matters

When an insurer models two lines of business, the central concern may be whether unusually large losses can occur together. Correlation summarizes a form of average association, but it does not by itself describe how dependence behaves in the extremes. A copula gives a way to model dependence separately from the loss distributions’ marginal behavior. The Gumbel copula is particularly relevant when the model should allow stronger association in the upper tail than in the lower tail.

The distinction is consequential for risk aggregation. If two losses are each at a high percentile, their joint exceedance probability affects how often a portfolio faces simultaneous severe losses. The Gumbel copula’s upper-tail dependence coefficient describes a limiting conditional probability as the percentile approaches one. It is not itself the joint probability at any particular finite percentile.

Definition: For continuous loss variables $X$ and $Y$, define $U=F_X(X)$ and $V=F_Y(Y)$. Both transformed variables are uniform on $(0,1)$. The Gumbel copula with parameter $\theta\ge1$ is $$ C_\theta(u,v)=\exp\left\{-\left[(-\ln u)^\theta+(-\ln v)^\theta\right]^{1/\theta}\right\},\qquad 0<u,v\le1. $$ It specifies $P(U\le u,V\le v)$. At $\theta=1$, $C_1(u,v)=uv$, the independence copula. Larger finite values of $\theta$ represent stronger upper-tail dependence.

Tail Dependence Coefficients

For uniform variables $U$ and $V$, the lower-tail dependence coefficient is the limiting probability that one variable is in the lower tail given that the other is there. The upper-tail coefficient is the analogous limiting probability for the upper tail:

\lambda_L=\lim_{q\downarrow0}P(U\le q\mid V\le q) =\lim_{q\downarrow0}\frac{C(q,q)}{q}, \qquad \lambda_U=\lim_{q\uparrow1}P(U>q\mid V>q).

These are limits along equal marginal percentiles. They describe tail behavior as the percentile becomes more extreme, not dependence at every threshold. For a Gumbel copula, set $a=2^{1/\theta}$. On the diagonal,

C_\theta(q,q)=\exp\left\{-\left[2(-\ln q)^\theta\right]^{1/\theta}\right\} =\exp\{-a(-\ln q)\}=q^a.

This diagonal expression yields both tail coefficients. In the lower tail, $C(q,q)/q=q^{a-1}$. For finite $\theta\ge1$, $a\ge1$, so this ratio tends to zero as $q\downarrow0$. In the upper tail, the conditional probability can be written using inclusion-exclusion:

P(U>q\mid V>q) =\frac{P(U>q,V>q)}{1-q} =\frac{1-2q+C_\theta(q,q)}{1-q} =\frac{1-2q+q^a}{1-q}.

Let $t=1-q$, so $t\downarrow0$ as $q\uparrow1$. Since $(1-t)^a=1-at+o(t)$, the numerator is $1-2(1-t)+(1-t)^a=(2-a)t+o(t)$, while the denominator is $t$. Therefore the ratio tends to $2-a$.

Theorem: For the Gumbel copula with finite $\theta\ge1$, $$ \lambda_L=0,\qquad \lambda_U=2-2^{1/\theta}. $$ At $\theta=1$, both coefficients are zero and the copula represents independence. For finite $\theta>1$, upper-tail dependence is positive while lower-tail dependence is zero. As $\theta$ increases without bound, $\lambda_U$ approaches one.

This establishes the model’s asymmetry precisely: for finite $\theta>1$, the probability of joint extreme upper-tail behavior remains positive in the limit, but the corresponding lower-tail coefficient is zero. A zero lower-tail coefficient does not mean that lower-tail events can never occur together; it means their conditional probability vanishes in the limiting definition. At $\theta=1$, there is no such upper-tail clustering either.

From Loss Percentiles to Joint Exceedance

Suppose the loss margins are continuous and the copula linking them is $C_\theta$. For thresholds $x$ and $y$, let $u=F_X(x)$ and $v=F_Y(y)$. Inclusion-exclusion gives the joint probability that both losses exceed their thresholds:

P(X>x,Y>y)=1-u-v+C_\theta(u,v).

The copula value $C_\theta(u,v)$ is a joint probability of being at or below both thresholds, not a joint upper-tail probability. The subtraction and addition in the formula account for the two marginal exceedances and their overlap. If the desired quantity is conditional on $X>x$, divide the joint exceedance probability by $1-u$, provided $u<1$.

The following examples use hypothetical one-year losses with continuous exponential margins. In each, the exponential distribution is specified by its mean parameter $\beta$, so $F(x)=1-\exp(-x/\beta)$ for $x\ge0$. The losses are nominal one-year amounts; no interest or mortality model is involved, and no discounting is applied.

Worked Example: Joint Exceedance at the 95th Percentile

A hypothetical insurer models two one-year losses, $X$ and $Y$, with exponential margins having means $\beta_X=\$2{,}000{,}000$ and $\beta_Y=\$1{,}200{,}000$. Their dependence is represented by a Gumbel copula with $\theta=2$. Find the thresholds at the 95th percentile for each margin, the probability both losses exceed their respective thresholds, and the probability $Y$ exceeds its threshold given that $X$ does.

Method and thresholds. For an exponential margin, the percentile $u$ threshold is $x_u=-\beta\ln(1-u)$. Thus the thresholds are

x_{0.95}=-\$2{,}000{,}000\ln(0.05)=\$5{,}991{,}464.55, \qquad y_{0.95}=-\$1{,}200{,}000\ln(0.05)=\$3{,}594{,}878.73.

At both thresholds, $u=v=0.95$, and $-\ln(0.95)=0.0512933$. The copula value is

C_2(0.95,0.95) =\exp\left\{-\sqrt{2(0.0512933)^2}\right\} =\exp(-0.0725397) \approx0.9300289.

Using inclusion-exclusion, the joint exceedance probability is

P(X>x_{0.95},Y>y_{0.95}) =1-0.95-0.95+0.9300289 =0.0300289 \approx0.0300.

The marginal probability that $X$ exceeds its threshold is $1-0.95=0.05$. Therefore,

P(Y>y_{0.95}\mid X>x_{0.95}) =\frac{0.0300289}{0.05} =0.600578 \approx0.6006.

The joint probability is about 3.00%, and the conditional probability is about 60.06%. The method avoids the common mistake of treating $C_2(0.95,0.95)$, which is a lower-rectangle probability, as the joint exceedance probability.

Worked Example: Unequal Percentiles and a Conditional Exceedance

In a second hypothetical portfolio, $X$ and $Y$ are one-year exponential losses with means $\beta_X=\$1{,}500{,}000$ and $\beta_Y=\$800{,}000$. A Gumbel copula with $\theta=2$ links them. Calculate the joint probability that $X$ exceeds its 98th-percentile threshold and $Y$ exceeds its 95th-percentile threshold. Then find the probability of the latter event conditional on the former.

Method and thresholds. The exponential quantile formula gives

x_{0.98}=-\$1{,}500{,}000\ln(0.02)=\$5{,}868{,}034.51, \qquad y_{0.95}=-\$800{,}000\ln(0.05)=\$2{,}396{,}585.86.

The corresponding marginal CDF values are $u=0.98$ and $v=0.95$. The required logarithms are $-\ln(0.98)=0.0202027$ and $-\ln(0.95)=0.0512933$. Therefore,

\begin{aligned} C_2(0.98,0.95) &=\exp\left\{-\sqrt{(0.0202027)^2+(0.0512933)^2}\right\}\\ &=\exp\left\{-\sqrt{0.0004081494+0.0026310021}\right\}\\ &=\exp(-0.0551285) \approx0.9463635. \end{aligned}

The joint upper-tail probability and the conditional probability are

\begin{aligned} P(X>x_{0.98},Y>y_{0.95}) &=1-0.98-0.95+0.9463635\\ &=0.0163635\approx0.0164,\\[4pt] P(Y>y_{0.95}\mid X>x_{0.98}) &=\frac{0.0163635}{1-0.98} =\frac{0.0163635}{0.02} \approx0.8182. \end{aligned}

The denominator is 0.02 because the condition is that $X$ exceeds its 98th-percentile threshold. Dividing by 0.05 instead would condition on $Y$’s exceedance, reversing the requested conditioning event.

Worked Example: A 90th-Percentile and 80th-Percentile Pair

A hypothetical insurer models $X$ and $Y$ as one-year exponential losses with means $\beta_X=\$1{,}000{,}000$ and $\beta_Y=\$2{,}000{,}000$, respectively, linked by a Gumbel copula with $\theta=2$. Find the probability that $X$ exceeds its 90th-percentile threshold and $Y$ exceeds its 80th-percentile threshold, and then find the conditional probability of the $Y$ exceedance given the $X$ exceedance.

Method and thresholds. The thresholds are

x_{0.90}=-\$1{,}000{,}000\ln(0.10)=\$2{,}302{,}585.09, \qquad y_{0.80}=-\$2{,}000{,}000\ln(0.20)=\$3{,}218{,}875.82.

Here $u=0.90$, $v=0.80$, $-\ln(0.90)=0.1053605$, and $-\ln(0.80)=0.2231436$. The copula value is

\begin{aligned} C_2(0.90,0.80) &=\exp\left\{-\sqrt{(0.1053605)^2+(0.2231436)^2}\right\}\\ &=\exp\left\{-\sqrt{0.01110084+0.04979304}\right\}\\ &=\exp(-0.2467669) \approx0.7813229. \end{aligned}

Thus the joint exceedance probability is

P(X>x_{0.90},Y>y_{0.80}) =1-0.90-0.80+0.7813229 =0.0813229 \approx0.0813.

Since $P(X>x_{0.90})=0.10$, the conditional probability is

P(Y>y_{0.80}\mid X>x_{0.90}) =\frac{0.0813229}{0.10} =0.813229 \approx0.8132.

The joint probability is checked against the conditional result: $0.813229\times0.10=0.0813229$, as required. The calculation also illustrates why a conditional exceedance probability need not equal the upper-tail dependence coefficient: these thresholds are fixed at the 90th and 80th percentiles, not taken to a common limiting percentile of one.

Interpretation and Common Exam Traps

The upper-tail coefficient can be computed directly for selected parameters. For example, when $\theta=2$, $\lambda_U=2-\sqrt{2}\approx0.5858$, while $\lambda_L=0$. This is a limiting conditional probability, whereas a finite-threshold calculation such as the examples above requires evaluating the copula at the actual marginal percentiles.

A frequent error is to use $C_\theta(u,v)$ as $P(X>x,Y>y)$. That is the wrong rectangle: $C_\theta(u,v)=P(X\le x,Y\le y)$. The correct joint exceedance calculation is $1-u-v+C_\theta(u,v)$. For instance, at $u=0.90$, $v=0.80$, and $\theta=2$, the incorrect calculation would report $0.7813$; the correct joint exceedance probability is $0.0813$. Another error is to describe upper-tail dependence as positive for every admissible parameter. At $\theta=1$, the model is independence and both tail coefficients are zero.

These distinctions matter when a dependence model is used to assess simultaneous severe losses. The coefficient communicates a limiting tail property; the copula evaluation and inclusion-exclusion calculation produce probabilities at specified thresholds. Neither substitutes for the other.

Key takeaway: For a Gumbel copula, $\lambda_U=2-2^{1/\theta}$ and $\lambda_L=0$. For actual loss thresholds, first map each threshold to its marginal percentile, evaluate $C_\theta(u,v)$, and use $1-u-v+C_\theta(u,v)$ for the joint upper-tail probability.

Check Your Understanding

For each problem, use the stated continuous margins and dependence model. Loss amounts are hypothetical one-year nominal losses, with no discounting or mortality assumption.

A pair of risks is linked by a Gumbel copula with $\theta=4$. Find its upper- and lower-tail dependence coefficients. (Answer: $\lambda_U=2-2^{1/4}\approx0.8108$, $\lambda_L=0$.)
Two continuous loss variables have a Gumbel copula with $\theta=1$. Their thresholds are the 90th and 70th marginal percentiles. Find the joint exceedance probability and the probability of the second exceedance conditional on the first. (Answer: joint probability $0.10(0.30)=0.0300$; conditional probability $0.3000$.)
For a Gumbel copula with $\theta=2$, both loss thresholds are the median. Find the joint exceedance probability and the conditional probability that one loss exceeds its median given the other does. (Answer: joint probability $2^{-\sqrt{2}}\approx0.3752$; conditional probability $2^{1-\sqrt{2}}\approx0.7504$.)
Two one-year losses have exponential margins with means $\$1{,}000{,}000$ and $\$2{,}000{,}000$, and a Gumbel copula with $\theta=2$. Find the joint exceedance probability when $F_X(x)=e^{-1}$ and $F_Y(y)=e^{-2}$, then give $P(Y>y\mid X>x)$. (Answer: $1-e^{-1}-e^{-2}+e^{-\sqrt{5}}\approx0.6037$; conditional probability $\approx0.9550$.)
A Gumbel copula has $\theta=3$. State the lower-tail coefficient and calculate the upper-tail coefficient to four decimal places. (Answer: $\lambda_L=0$; $\lambda_U=2-2^{1/3}\approx0.7401$.)

What You'll Learn