Why Compare Competing Copulas?
A copula describes how variables move together after their marginal distributions have been separated from their dependence structure. For an insurer modeling two lines of loss, the choice of copula can affect simulated joint extremes even when the fitted marginal loss distributions are unchanged. A sound selection therefore asks more than which model has the largest likelihood: it also asks whether diagnostics support the model and whether its tail behavior is plausible for the application.
This case study uses a hypothetical insurer’s paired annual losses from two business lines. The losses are assumed to be continuous, and each pair refers to the same year. The marginal models have already been fitted; the aim here is to compare Gaussian, Student-t, Clayton, and Gumbel copulas for the dependence between the two losses. All data and numerical results below are illustrative, not industry observations.
Preparing the Data and Fitting the Candidates
In practice, the unknown marginal distributions are often replaced for dependence estimation by rank-based pseudo-observations. With \(n\) paired observations and no ties, let \(R_{iX}\) and \(R_{iY}\) be the ranks of the two losses in pair \(i\). Define \(u_i=R_{iX}/(n+1)\) and \(v_i=R_{iY}/(n+1)\). These values lie strictly between zero and one. For tied observations, average ranks are a common approximation; substantial ties require care because the continuous-margin assumption no longer holds exactly.
For a candidate copula with density \(c(u,v;\theta)\), estimate its parameter or parameters by maximizing the pseudo-log-likelihood:
This is a likelihood for the transformed data, not a full likelihood for the original losses and their margins. The Gaussian and Student-t copulas use dependence parameters \(\rho\), with the Student-t copula also using degrees of freedom \(\nu\). Clayton and Gumbel are Archimedean copulas, with positive parameters \(\theta\). Their bivariate distribution functions are:
The Gaussian copula is the bivariate standard normal distribution function with correlation \(\rho\), evaluated at the standard normal quantiles of \(u\) and \(v\). The Student-t copula is the corresponding bivariate \(t\) distribution with correlation parameter \(\rho\) and degrees of freedom \(\nu\), evaluated at univariate \(t_\nu\) quantiles. In each case, the copula density \(c\) is the mixed partial derivative of its distribution function. Maximum pseudo-likelihood requires a valid parameter range and numerical optimization; rank-based estimates provide useful starting values.
Worked Example: From Loss Ranks to Tail Evidence
For 10 hypothetical annual loss pairs, the ranks of Line A’s losses are 1 through 10. In the two years with Line A rank 9 or 10, the corresponding Line B ranks are 8 and 10. Use the rank/(n+1) convention to calculate the empirical upper-tail co-exceedance proportion at threshold 0.8, and explain what this small sample can establish.
Method. First convert ranks to pseudo-observations using \(n+1=11\). Line A ranks 9 and 10 give \(9/11=0.8182\) and \(10/11=0.9091\), so both exceed 0.8. The paired Line B pseudo-observations are \(8/11=0.7273\) and \(10/11=0.9091\). Only the second exceeds 0.8.
There is one joint exceedance among two Line A exceedances. The empirical conditional proportion is therefore \(1/2=0.5000\). This is a descriptive estimate based on only two observations beyond the threshold; it is not a reliable estimate of asymptotic tail dependence. Treating it as precise evidence for a tail-dependent copula is the common mistake this calculation avoids.
Rank Calibration and Tail Behavior
Kendall’s tau, a rank-based measure of concordance, offers a transparent way to calibrate or check candidate parameters. For the Gaussian and Student-t copulas, \(\tau=(2/\pi)\arcsin(\rho)\). For Clayton, \(\tau=\theta/(\theta+2)\); for Gumbel, \(\tau=1-1/\theta\). These relationships give starting estimates under the corresponding copula models. They do not replace checking the fitted likelihood or goodness of fit.
The tail-dependence coefficients describe limiting conditional probabilities. For continuous uniform variables, the lower-tail coefficient is \(\lambda_L=\lim_{q\to 0+}P(V\le q\mid U\le q)\), and the upper-tail coefficient is \(\lambda_U=\lim_{q\to 1-}P(V>q\mid U>q)\). A positive coefficient indicates asymptotic clustering in that tail. Zero does not rule out dependence at finite thresholds.
- The Gaussian copula has \(\lambda_L=\lambda_U=0\) when \(|\rho|<1\).
- The Student-t copula has symmetric tail dependence: \(\lambda_L=\lambda_U=2t_{\nu+1}\left(-\sqrt{(\nu+1)(1-\rho)/(1+\rho)}\right)\), where \(t_{\nu+1}\) is the standard \(t\) distribution function.
- Clayton has \(\lambda_L=2^{-1/\theta}\) and \(\lambda_U=0\).
- Gumbel has \(\lambda_L=0\) and \(\lambda_U=2-2^{1/\theta}\).
Worked Example: Calibrating Parameters and Comparing Tails
Suppose the sample Kendall’s tau for the hypothetical paired losses is 0.40. Calculate the rank-calibrated Gaussian correlation, Clayton parameter and lower-tail coefficient, and Gumbel parameter and upper-tail coefficient. Interpret the comparison without treating these calibrations as proof of model fit.
Method. For the Gaussian copula, invert \(\tau=(2/\pi)\arcsin(\rho)\):
For Clayton, solve \(0.40=\theta/(\theta+2)\). Thus \(0.40\theta+0.80=\theta\), so \(0.60\theta=0.80\) and \(\theta=1.3333\). Its lower-tail coefficient is \(2^{-1/\theta}=2^{-0.75}\approx0.5946\), while its upper-tail coefficient is zero.
For Gumbel, solve \(0.40=1-1/\theta\), giving \(1/\theta=0.60\) and \(\theta=1.6667\). Its upper-tail coefficient is \(2-2^{1/1.6667}=2-2^{0.60}\approx2-1.5157=0.4843\), while its lower-tail coefficient is zero.
The rank concordance is the same input for these calibrations, but their implied extreme behavior differs: Clayton concentrates asymptotic dependence in the lower tail, and Gumbel in the upper tail. The Gaussian calibration has no asymptotic tail dependence. Kendall’s tau alone cannot determine which of those tail patterns best describes the data.
Comparing Fit and Making a Defensible Choice
Use diagnostics that address different questions. Information criteria summarize the trade-off between fit and parameter count. AIC is \(2k-2\ell_{\mathrm{PL}}\), and BIC is \(k\log(n)-2\ell_{\mathrm{PL}}\), where \(k\) is the number of estimated copula parameters. When these are applied to pseudo-likelihoods, treat them as comparative criteria for the same data and estimation procedure, not as exact full-likelihood evidence. Smaller values favor a candidate relative to the alternatives being compared.
A goodness-of-fit diagnostic can compare the fitted copula with the empirical dependence pattern, with significance assessed by a parametric bootstrap. In that procedure, simulate data from the fitted candidate, refit it in each replicate, and compare the observed diagnostic statistic with the bootstrap distribution. Re-estimation matters: holding fitted parameters fixed can misrepresent the variability of the diagnostic. A large bootstrap p-value means the test did not detect lack of fit; it does not prove the model is true.
Consider the following illustrative results for \(n=240\) pseudo-observations. The bootstrap p-values and fit statistics are stipulated hypothetical outputs from a diagnostic procedure. The Student-t estimate has \(\hat\rho=0.60\) and \(\hat\nu=4\); the other estimates are represented by their fitted pseudo-likelihoods. The observed data also show more upper-tail than lower-tail co-exceedance at the chosen finite threshold.
| Copula | Estimated parameters | \(\ell_{\mathrm{PL}}\) | k | AIC | Bootstrap p-value | Tail implication |
|---|---|---|---|---|---|---|
| Gaussian | \(\hat\rho=0.59\) | 22.4 | 1 | -42.8 | 0.08 | No asymptotic tail dependence |
| Student-t | \(\hat\rho=0.60,\ \hat\nu=4\) | 28.1 | 2 | -52.2 | 0.31 | Symmetric tail dependence |
| Clayton | \(\hat\theta=1.25\) | 17.6 | 1 | -33.2 | 0.02 | Lower-tail dependence only |
| Gumbel | \(\hat\theta=1.65\) | 20.9 | 1 | -39.8 | 0.11 | Upper-tail dependence only |
The Student-t fit has the smallest AIC and the largest bootstrap p-value among these candidates. Its extra parameter is penalized in AIC, yet its improved pseudo-log-likelihood more than offsets that penalty. Its symmetric tail dependence is not an exact representation of the stated upper-versus-lower asymmetry, so model selection should disclose that limitation. The Clayton result is less attractive both by AIC and by the diagnostic, and its lower-tail-only structure conflicts with the observed direction of finite-threshold clustering.
Worked Example: Checking the Information-Criterion Choice
Use the hypothetical pseudo-log-likelihoods and parameter counts in the table to calculate AIC for the Student-t and Gaussian copulas. Then identify the winner and show how the wrong parameter count could distort the comparison.
Method. For Student-t, \(k=2\) and \(\ell_{\mathrm{PL}}=28.1\), so \(\mathrm{AIC}=2(2)-2(28.1)=4-56.2=-52.2\). For Gaussian, \(k=1\) and \(\ell_{\mathrm{PL}}=22.4\), so \(\mathrm{AIC}=2(1)-2(22.4)=2-44.8=-42.8\). Since \(-52.2<-42.8\), AIC favors Student-t.
A common mistake is to count only the Student-t correlation and omit its estimated degrees of freedom. That incorrect calculation gives \(2(1)-56.2=-54.2\), making its AIC appear too favorable by 2. The correct count is two fitted copula parameters. The comparison is also meaningful only because both candidates were fitted to the same pseudo-observations using the same criterion.
Justifying the Selection Process
The parameter penalty in AIC has a practical purpose: a more flexible model can improve its fitted likelihood simply by having additional parameters. AIC balances that improvement against model complexity. For the two-parameter Student-t candidate versus a one-parameter Gaussian candidate, the Student-t is favored exactly when its log pseudo-likelihood gain exceeds one:
Here the gain is \(28.1-22.4=5.7\), which exceeds 1, so the reduction in AIC is \(2(5.7)-2=9.4\), consistent with \(-52.2-(-42.8)=-9.4\). This algebra explains why the additional flexibility is supported by these illustrative results; it does not establish that the Student-t copula is the true dependence model.
For this hypothetical case, Student-t is a defensible working choice because it has the best AIC and the strongest non-rejection in the bootstrap diagnostic. Its estimated \(\nu=4\) implies heavier joint tails than a Gaussian copula, but its symmetric tail dependence may not capture observed upper-tail asymmetry. An actuary could retain it as a practical candidate while reporting that limitation and assessing sensitivity of downstream results to plausible alternatives. The selection is a modeling judgment, not a declaration that one statistic settles the question.
Common Selection Traps
- Choosing by Kendall’s tau alone. Equal rank concordance can coexist with different tail structures, as the Clayton and Gumbel calibrations demonstrate.
- Reading a non-significant diagnostic as proof. A p-value such as 0.31 indicates that the test did not reject at common significance levels; it does not establish correctness or guarantee accurate extreme-loss estimates.
- Ignoring parameter counts. Information criteria must penalize all estimated parameters, including Student-t degrees of freedom when estimated.
- Confusing finite-threshold clustering with asymptotic tail dependence. A finite sample can show joint exceedances even when a copula’s limiting coefficient is zero, and a small number of extremes gives uncertain evidence.
- Selecting a copula without considering its intended use. A model adequate for ordinary dependence summaries may be inadequate for a decision driven by one tail. Report the relevant tail assumptions and, where warranted, test sensitivity to competing models.
Check Your Understanding
Work each problem using the definitions and methods in this tutorial. Numerical answers are rounded to four decimals where appropriate.
- For \(n=19\), a paired loss has ranks 14 and 18 in the two margins. Calculate its pseudo-observations using rank/(n+1). (Answer: \(14/20=0.7000\) and \(18/20=0.9000\).)
- A hypothetical sample has Kendall’s tau \(0.25\). Find the Clayton parameter from rank calibration and state which tail has asymptotic dependence. (Answer: \(\theta=2(0.25)/(1-0.25)=0.6667\); lower tail only.)
- A Gumbel copula has \(\theta=2\). Calculate its upper-tail dependence coefficient and identify its lower-tail coefficient. (Answer: \(\lambda_U=2-\sqrt{2}=0.5858\); \(\lambda_L=0\).)
- A Gaussian candidate has \(\ell_{\mathrm{PL}}=12.0\) with one parameter. A Student-t candidate has \(\ell_{\mathrm{PL}}=15.0\) with two parameters. Calculate both AIC values and select by AIC. (Answer: Gaussian \(2-24=-22.0000\); Student-t \(4-30=-26.0000\); Student-t has lower AIC.)
- A fitted Student-t copula has \(\hat\rho=0.50\) and \(\hat\nu=4\). Using the tail-coefficient formula, simplify the square-root argument and state whether tail dependence is symmetric. (Answer: \((5)(1-0.50)/(1+0.50)=5/3\), so \(\lambda_L=\lambda_U=2t_5(-\sqrt{5/3})\); the tail dependence is symmetric.)