Clayton Copulas and Lower-Tail Dependence

Introduction

A dependence model can have a substantial effect on the probability that two risks occupy an extreme region at the same time. Correlation alone does not describe how dependence behaves far into a tail. The Clayton copula is a useful model to study when the question concerns lower-tail clustering: it has a positive lower-tail dependence coefficient for positive parameter values.

This tutorial constructs the bivariate Clayton copula from its Archimedean generator, derives that coefficient, and calculates joint probabilities at specified marginal quantiles. The distinction between a tail limit and a probability at a finite threshold is central: the coefficient describes limiting behavior, not the joint probability at every quantile.

Clayton Copula and Its Parameter

For this tutorial, take two continuous random variables with marginal distribution functions \(F_X\) and \(F_Y\). Their probability integral transforms are \(U=F_X(X)\) and \(V=F_Y(Y)\), each uniform on \([0,1]\). We use the conventional bivariate Clayton model with parameter \(\theta>0\); this positive-parameter model represents positive dependence and has lower-tail dependence. The independence case is obtained as the limit \(\theta\to0\).

Definition: The Clayton Archimedean generator and its inverse, for \(\theta>0\), are \(\phi(t)=(1+t)^{-1/\theta}\), \(t\geq0\), and \(\phi^{-1}(u)=u^{-\theta}-1\), \(0<u\leq1\). The resulting bivariate copula is \(C_\theta(u,v)=\phi\big(\phi^{-1}(u)+\phi^{-1}(v)\big) =(u^{-\theta}+v^{-\theta}-1)^{-1/\theta}\) for \(u,v\in(0,1]\), with boundary values defined by continuity.

The generator is decreasing, satisfies \(\phi(0)=1\), and tends to zero as \(t\) tends to infinity. For \(\theta>0\), it is completely monotone, so it gives a valid Archimedean copula in two dimensions. The copula joins the specified marginal distributions without changing them: for example, \(P(U\leq u,V\leq v)=C_\theta(u,v)\). Thus, if \(x\) and \(y\) are marginal quantiles, the joint probability below both thresholds is \(C_\theta(F_X(x),F_Y(y))\).

Deriving Lower-Tail Dependence

The lower-tail dependence coefficient measures the limiting proportion of probability in one lower tail that also lies in the other lower tail, when both marginal thresholds approach zero together. For continuous uniform margins, it is defined by the diagonal limit below. The equality follows because \(P(U\leq q)=q\).

Definition: For a copula \(C\), its lower-tail dependence coefficient is \(\lambda_L=\lim_{q\downarrow0}P(V\leq q\mid U\leq q)=\lim_{q\downarrow0}C(q,q)/q\), when the limit exists. A positive value indicates asymptotic lower-tail dependence; a value of zero indicates lower-tail independence under this definition.

For the Clayton copula, set \(u=v=q\). Then \(C_\theta(q,q)=(2q^{-\theta}-1)^{-1/\theta}\). Dividing by \(q\) and factoring \(q^{-\theta}\) from the expression inside the power gives a limit that can be evaluated directly.

\begin{aligned} \lambda_L &=\lim_{q\downarrow0}\frac{(2q^{-\theta}-1)^{-1/\theta}}{q}\\ &=\lim_{q\downarrow0}\frac{\big[q^{-\theta}(2-q^\theta)\big]^{-1/\theta}}{q}\\ &=\lim_{q\downarrow0}\frac{q(2-q^\theta)^{-1/\theta}}{q}\\ &=2^{-1/\theta}. \end{aligned}

The cancellation in the third line is valid for \(q>0\), and \(q^\theta\to0\) as \(q\downarrow0\) when \(\theta>0\). This proves \(\lambda_L=2^{-1/\theta}\). The coefficient increases with \(\theta\): stronger positive parameter values correspond to greater asymptotic lower-tail dependence. It is not itself a finite-threshold joint probability.

Worked Example: A Finite Joint Probability and Its Tail Limit

A hypothetical model uses a Clayton copula with \(\theta=1\). Find the probability that both uniform variables are at most \(0.10\), and find the lower-tail dependence coefficient. Compare the two quantities.

Method. Use the diagonal copula formula for the finite-quantile probability, then use the derived limit for \(\lambda_L\). Here \(q=0.10\), so \(q^{-1}=10\).

The joint probability is \(C_1(0.10,0.10)=(10+10-1)^{-1}=1/19=0.0526316\ldots\). To four decimal places, this is \(0.0526\).

The lower-tail coefficient is \(\lambda_L=2^{-1/1}=1/2=0.5000\). The finite-threshold ratio is \(C_1(q,q)/q=0.0526316\ldots/0.10=0.526316\ldots\), or \(0.5263\) to four decimal places. It is close to, but not equal to, the limiting value \(0.5000\).

Check. The joint event is a subset of either marginal event, so its probability must be no greater than \(0.10\); \(0.0526\) satisfies this bound. The common mistake is to report the limiting coefficient \(0.5000\) as the joint probability. The coefficient is a conditional-probability limit, not \(P(U\leq0.10,V\leq0.10)\).

Calculating Joint Probabilities at Selected Quantiles

At finite thresholds, use the copula itself. If both marginal events are at their \(q\)-quantiles, the joint probability is \(C_\theta(q,q)=(2q^{-\theta}-1)^{-1/\theta}\). This probability depends on both \(q\) and \(\theta\); the tail coefficient depends only on \(\theta\). At unequal quantiles \(u\) and \(v\), use the full expression \(C_\theta(u,v)=(u^{-\theta}+v^{-\theta}-1)^{-1/\theta}\), not the diagonal formula.

Worked Example: Comparing Parameters at the Same Quantile

An analyst compares hypothetical Clayton models with \(\theta=0.5\), \(1\), and \(2\). For each model, calculate the probability that both uniform variables are at most \(0.10\), and calculate the corresponding lower-tail coefficient. Give probabilities to four decimal places.

Method. Keep the quantile \(q=0.10\) fixed, substitute each parameter into the diagonal formula, and then calculate \(2^{-1/\theta}\). Since \(q^{-1}=10\), the substitutions can be checked directly.

For \(\theta=0.5\), \(q^{-\theta}=10^{0.5}=\sqrt{10}\), so \(C_{0.5}(q,q)=(2\sqrt{10}-1)^{-2}=1/(2\sqrt{10}-1)^2\). Using \(2\sqrt{10}-1=5.324555\ldots\), the probability is approximately \(0.0353\). Its coefficient is \(2^{-2}=0.2500\).

For \(\theta=1\), \(C_1(q,q)=(20-1)^{-1}=1/19=0.0526316\ldots\), or \(0.0526\). Its coefficient is \(2^{-1}=0.5000\).

For \(\theta=2\), \(q^{-2}=100\), so \(C_2(q,q)=(200-1)^{-1/2}=1/\sqrt{199}=0.070888\ldots\), or \(0.0709\). Its coefficient is \(2^{-1/2}=0.7071\) to four decimal places.

Check. The probabilities increase from \(0.0353\) to \(0.0709\) as \(\theta\) increases, as do the limiting coefficients. All three joint probabilities are at most the marginal probability \(0.10\). The finite-quantile figures are distinct from their respective tail coefficients.

Worked Example: Unequal Marginal Quantiles for Hypothetical Losses

Suppose two continuous hypothetical loss variables have marginal thresholds \(x\) and \(y\) such that \(F_X(x)=0.10\) and \(F_Y(y)=0.20\). Their dependence is modeled with a Clayton copula having \(\theta=1\). Find the probability that both losses are below their respective thresholds.

Method. Transform each loss threshold to its marginal probability, then substitute those two probabilities into the full, non-diagonal copula formula. The loss amounts themselves are not needed once their marginal probabilities are known.

The required probability is \(C_1(0.10,0.20)=(0.10^{-1}+0.20^{-1}-1)^{-1}\). The inverse probabilities are \(0.10^{-1}=10\) and \(0.20^{-1}=5\), so the quantity inside parentheses is \(10+5-1=14\). Therefore the joint probability is \(1/14=0.0714286\ldots\), or \(0.0714\) to four decimal places.

Check. The joint event is contained in the event with probability \(0.10\), the smaller marginal probability, and \(0.0714\leq0.10\). Using the diagonal formula here would be incorrect because the two marginal quantiles are not equal.

Interpretation for Loss Variables

The words “lower tail” refer to small values of the variables on their original scale, equivalently to small values of their CDF transforms. If \(X\) and \(Y\) are losses, events \(F_X(X)\leq q\) and \(F_Y(Y)\leq q\) describe losses below their lower \(q\)-quantiles. They do not describe simultaneous large losses. For conventional nonnegative loss variables, large-loss concerns are usually in the upper tail, so a lower-tail coefficient should not be interpreted as a measure of joint adverse-loss behavior.

This distinction also matters when variables have been transformed, signed, or defined with a direction other than “larger means more loss.” The copula coefficient concerns the lower ends of the variables as modeled, not an informal notion of bad outcomes. State the event in terms of the modeled variables and marginal quantiles before assigning a risk interpretation.

For a Clayton model with \(\theta>0\), the derived coefficient is positive. This does not mean that every pair of low observations has probability \(\lambda_L\), nor does it determine the probability at a chosen finite quantile without evaluating the copula. Likewise, this coefficient alone does not describe the model’s upper-tail behavior; that is a different tail question.

Common Mistakes and a Calculation Checklist

The following errors can change either the numerical answer or its meaning:

Confusing the limit with a finite probability. \(\lambda_L\) is \(\lim_{q\downarrow0}C(q,q)/q\). For a stated \(q\), calculate \(C(q,q)\) directly.
Dropping the minus one. The bivariate Clayton expression includes \(u^{-\theta}+v^{-\theta}-1\). Omitting \(-1\) changes the result, including at moderate quantiles.
Using a diagonal formula for unequal quantiles. When \(u\ne v\), substitute both values into \(C_\theta(u,v)\).
Reversing the tail interpretation. Lower CDF quantiles of a loss are small losses, not large losses. Identify which event the modeled tail actually represents.
Applying the positive-parameter result outside its stated setting. The derivation here assumes the conventional Clayton model with \(\theta>0\); do not use its formula as a general result for other parameter ranges or other copulas.

Identify the marginal events.
Write their CDF probabilities \(u=F_X(x)\) and \(v=F_Y(y)\), checking whether the thresholds are equal.

Evaluate the finite joint probability.
Use \(C_\theta(u,v)=(u^{-\theta}+v^{-\theta}-1)^{-1/\theta}\) for the specified \(\theta>0\).

Calculate the tail coefficient separately.
For the positive-parameter Clayton model, use \(\lambda_L=2^{-1/\theta}\); do not substitute it for the finite joint probability.

Check the interpretation and bounds.
The joint probability cannot exceed either marginal probability. For loss variables, say explicitly whether the event concerns small or large losses.

Key takeaway: The Clayton generator \(\phi(t)=(1+t)^{-1/\theta}\) yields \(C_\theta(u,v)=(u^{-\theta}+v^{-\theta}-1)^{-1/\theta}\), and its lower-tail dependence coefficient is \(2^{-1/\theta}\) for \(\theta>0\). Calculate finite joint probabilities from the copula, and interpret “lower tail” on the modeled variables’ scale.

Check Your Understanding

For each problem, use the stated positive-parameter Clayton model and give probabilities to four decimal places where needed.

A Clayton model has \(\theta=1.5\). Find its lower-tail coefficient and the joint probability \(P(U\leq0.10,V\leq0.10)\). (Answer: \(\lambda_L\approx0.6300\); joint probability \(\approx0.0637\))
For \(\theta=2\), calculate \(P(U\leq0.05,V\leq0.05)\) and compare it with the marginal probability \(0.05\). (Answer: \(1/\sqrt{799}\approx0.0354\), which is less than \(0.05\))
A Clayton model has lower-tail coefficient \(0.2500\). Find its parameter \(\theta\). (Answer: \(\theta=0.5\))
With \(\theta=1\), find the joint probability at equal marginal quantiles \(q=0.20\). Compare it with the independence value. (Answer: Clayton probability \(0.20/(2-0.20)=0.1111\); independence value \(0.20^2=0.0400\))
Two hypothetical loss thresholds have marginal probabilities \(0.10\) and \(0.20\). Under a Clayton model with \(\theta=1\), find the probability both losses are below their respective thresholds, and state whether this is a joint large-loss probability. (Answer: \(1/(10+5-1)=1/14\approx0.0714\); no, these are lower-quantile, small-loss events)

What You'll Learn