Introduction
A dependence model can describe meaningful association without predicting that extreme outcomes will cluster together asymptotically. The Frank copula is a useful example: its parameter can represent positive or negative association, yet its lower- and upper-tail dependence coefficients are both zero for every finite, nonzero parameter value.
For an actuarial model, this distinction matters. A copula determines how marginal loss percentiles combine into joint probabilities, and therefore affects aggregation even when it has no asymptotic tail dependence. The calculations below use uniform variables \(U\) and \(V\) to represent marginal percentiles, following the copula construction.
Interpreting the Parameter
The Frank family allows any real parameter. A positive \(\theta\) gives positive association: joint lower-percentile probabilities tend to exceed the corresponding independence probabilities. A negative \(\theta\) gives negative association: those probabilities tend to be smaller. At \(\theta=0\), the margins are independent. The magnitude of \(\theta\) controls the strength of association within this family; \(\theta\) is not itself a Pearson correlation.
The formula is symmetric in \(u\) and \(v\), so the model is exchangeable: swapping the two margins does not change the joint distribution. Its boundary values also have to match uniform margins: \(C_{\theta}(u,1)=u\), \(C_{\theta}(1,v)=v\), and \(C_{\theta}(u,0)=C_{\theta}(0,v)=0\). These properties make corner calculations straightforward.
Worked Example: Comparing a Frank Joint Probability With Independence
A hypothetical model uses uniform percentile variables with a Frank copula parameter \(\theta=2\). Calculate the probability that both variables are at or below their respective median, and compare it with independence.
Method. Substitute \(u=v=0.5\) and \(\theta=2\) into the copula formula. First, \(e^{-2(0.5)}=e^{-1}\approx0.367879\), so \(e^{-1}-1\approx-0.632121\). The product of the two numerator terms is \((-0.632121)^2\approx0.399576\). The denominator is \(e^{-2}-1\approx-0.864665\).
The ratio is \(0.399576/(-0.864665)\approx-0.462117\). Therefore, $$ C_2(0.5,0.5) = -\frac{1}{2}\ln(1-0.462117) = -\frac{1}{2}\ln(0.537883) \approx 0.3101. $$
Under independence, the probability is \(0.5(0.5)=0.25\). The Frank model therefore gives a median-by-median joint probability of about \(0.3101\), greater than \(0.25\), consistent with positive association. The calculation avoids the mistake of treating \(\theta=2\) as a correlation or substituting it directly into a correlation formula. As a check, the result lies between zero and each marginal probability, \(0.5\).
Calculating Joint and Rectangular Probabilities
For any thresholds \(u\) and \(v\), the joint lower-tail probability is \(C_{\theta}(u,v)\). A probability over intervals requires four copula evaluations. If \(0\le a\le b\le1\) and \(0\le c\le d\le1\), inclusion-exclusion gives $$ P(a<U\le b,\ c<V\le d) = C_{\theta}(b,d)-C_{\theta}(a,d)-C_{\theta}(b,c)+C_{\theta}(a,c). $$ The formula subtracts the two portions below the lower edges, then adds back the lower-left overlap that was subtracted twice. It works for continuous uniform margins; including or excluding a boundary does not change the probability.
Worked Example: Evaluating an Interior Frank Copula Value
For a hypothetical portfolio model with \(\theta=2\), calculate \(C_2(0.4,0.7)\). Interpret the result as a joint percentile probability.
Method. Evaluate each exponential term first. We have \(e^{-0.8}\approx0.449329\), so \(e^{-0.8}-1\approx-0.550671\), and \(e^{-1.4}\approx0.246597\), so \(e^{-1.4}-1\approx-0.753403\). Their product is approximately \(0.414877\), while the denominator \(e^{-2}-1\) is approximately \(-0.864665\).
Thus the fraction is \(0.414877/(-0.864665)\approx-0.47981\), and the logarithm argument is \(1-0.47981\approx0.52019\). Substitution gives $$ C_2(0.4,0.7) \approx -\frac{1}{2}\ln(0.52019) \approx 0.3268. $$
This means \(P(U\le0.4,V\le0.7)\approx0.3268\). The answer is no greater than the smaller marginal probability, \(0.4\), as required. A common arithmetic trap is to misdivide the negative fraction: rounding the ratio to \(-0.47985\) instead gives an inaccurate logarithm argument. Using \(-0.47981\) and \(0.52019\) to five decimal places gives the stated result, rounded to four decimal places.
Worked Example: Calculating a Rectangular Probability
A hypothetical model again uses \(\theta=2\). Calculate the probability that \(0.2<U\le0.5\) and \(0.3<V\le0.7\). Use six-decimal copula values in the inclusion-exclusion calculation, then round the probability to four decimal places.
Method. The four required values, evaluated from the Frank formula, are \(C_2(0.5,0.7)\approx0.400123\), \(C_2(0.2,0.7)\approx0.169313\), \(C_2(0.5,0.3)\approx0.200124\), and \(C_2(0.2,0.3)\approx0.094389\). For example, the last value uses numerator product \((e^{-0.4}-1)(e^{-0.6}-1)\approx0.148748\), denominator \(e^{-2}-1\approx-0.864665\), and hence logarithm argument approximately \(0.827970\).
Apply the rectangle formula, with upper-upper corner first and lower-lower corner added back: $$ \begin{aligned} P(0.2<U\le0.5,\ 0.3<V\le0.7) &\approx 0.400123-0.169313-0.200124+0.094389\\ &=0.125075\\ &\approx 0.1251. \end{aligned} $$
The probability is positive and is less than either interval’s marginal probability: those are \(0.5-0.2=0.3\) and \(0.7-0.3=0.4\). The common mistake is to subtract only the two lower-edge terms and omit \(C_2(0.2,0.3)\); that omission would incorrectly produce approximately \(0.0307\).
Why the Frank Copula Has No Asymptotic Tail Dependence
For uniform margins, the lower-tail dependence coefficient is \(\lambda_L=\lim_{t\downarrow0}C_{\theta}(t,t)/t\), when the limit exists. It measures the limiting conditional probability that one variable is in an increasingly extreme lower tail, given that the other is. The upper-tail coefficient is \(\lambda_U=\lim_{t\downarrow0}[1-2(1-t)+C_{\theta}(1-t,1-t)]/t\).
To evaluate the lower-tail limit for fixed finite \(\theta\ne0\), expand the exponential near zero: \(e^{-\theta t}-1=-\theta t+O(t^2)\). The product of the two numerator terms is \(\theta^2t^2+O(t^3)\). Expanding the logarithm, whose argument is one plus a term of order \(t^2\), then gives $$ C_{\theta}(t,t) = \frac{\theta}{1-e^{-\theta}}t^2+O(t^3). $$ The coefficient is finite for fixed \(\theta\ne0\). Consequently, $$ \lambda_L = \lim_{t\downarrow0} \frac{\frac{\theta}{1-e^{-\theta}}t^2+O(t^3)}{t} =0. $$ At \(\theta=0\), independence gives \(C_0(t,t)=t^2\), so the lower-tail coefficient is zero there as well.
The Frank copula is radially symmetric, meaning \(C_{\theta}(u,v)=u+v-1+C_{\theta}(1-u,1-v)\). Taking \(u=v=1-t\) shows that the joint upper-tail probability is \(1-2(1-t)+C_{\theta}(1-t,1-t)=C_{\theta}(t,t)\). Its order is therefore \(t^2\), and division by \(t\) again tends to zero. Thus both tail coefficients are zero for every finite Frank parameter, including the independence limit.
Negative Association and the Meaning of Tail Independence
The sign of the parameter can be seen directly in a median calculation. For \(\theta=-2\), the Frank formula at \(u=v=0.5\) simplifies to $$ C_{-2}(0.5,0.5) = \frac{1}{2}\ln\left(1+\frac{(e-1)^2}{e^2-1}\right) = \frac{1}{2}\ln\left(1+\frac{e-1}{e+1}\right) \approx0.1899. $$ The simplification uses \(e^2-1=(e-1)(e+1)\). This is below the independence probability \(0.25\), as expected for negative association.
A zero tail-dependence coefficient concerns a limit as the percentile threshold approaches zero. At thresholds that are extreme but not arbitrarily close to zero, the joint probability can still differ materially from independence. Risk aggregation should therefore use the fitted copula’s probabilities at relevant thresholds, rather than treating “tail independent” as a synonym for “independent.”
This distinction is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus. In a model, the copula describes how marginal percentiles are paired; the marginal loss distributions still determine the dollar amounts associated with those percentiles. A central or tail joint probability from the copula is not, by itself, an aggregate loss amount.
Translate the event into thresholds or intervals for \(U\) and \(V\).
Use \(C_{\theta}(u,v)\) for a lower-rectangle event, or evaluate all four corners for an interval rectangle.
Compare with independence where useful, verify the probability is within valid bounds, and distinguish finite-threshold association from asymptotic tail dependence.
Check Your Understanding
Use the Frank copula definition and show enough intermediate work to check each calculation.
- For \(\theta=0\), calculate \(P(U\le0.4,V\le0.6)\). (Answer: \(0.2400\))
- For \(\theta=2\), calculate \(P(U\le0.2,V\le0.3)\). (Answer: approximately \(0.0944\))
- For \(\theta=2\), calculate \(P(0.2<U\le0.5,\ 0.3<V\le0.7)\), using the four-corner method. (Answer: approximately \(0.1251\))
- For \(\theta=-2\), calculate \(P(U\le0.5,V\le0.5)\) and compare it with the independence value. (Answer: approximately \(0.1899\), below \(0.2500\))
- For a finite Frank parameter, state the lower- and upper-tail dependence coefficients and explain why this does not imply independence. (Answer: both are zero; the copula can still produce non-independence at finite thresholds.)