Why Compare Copulas Under Stress?
A risk aggregation model can use identical marginal loss distributions and the same dependence parameter yet produce different probabilities of simultaneous extreme losses. The choice between a Gaussian copula and a Student-t copula matters because the Student-t construction allows a shared scale factor to make large values occur together more often. That difference can be particularly important when a risk assessment focuses on stressed thresholds rather than average outcomes.
A fair comparison holds the marginal distributions and the copula dependence parameter fixed. It then asks how much the modeled joint exceedance probability changes with the copula family and, for a Student-t copula, with its degrees of freedom. A larger joint exceedance probability means more modeled co-extremes at the particular thresholds being tested; it does not, by itself, establish that either model is appropriate.
The last expression follows by inclusion-exclusion. The event that both uniform variables exceed \(u\) is the complement of the union of the events \(U_1\le u\) and \(U_2\le u\). Each of those events has probability \(u\), and their intersection has probability \(C(u,u)\). Therefore,
This formula assumes continuous margins and a common probability threshold \(u\). If the thresholds differ, say \(u_1\) and \(u_2\), the corresponding expression is \(1-u_1-u_2+C(u_1,u_2)\). When losses have different marginal distributions, equal probability thresholds generally do not mean equal dollar thresholds: the loss cutoffs are \(F_1^{-1}(u)\) and \(F_2^{-1}(u)\).
Matched-Parameter Formulas
For a bivariate Gaussian copula with latent correlation \(\rho\), let \(\Phi\) be the standard normal distribution function and \(\Phi_2(\cdot,\cdot;\rho)\) the standard bivariate normal distribution function with correlation \(\rho\). For the bivariate Student-t copula, let \(T_\nu\) be the univariate Student-t distribution function with \(\nu\) degrees of freedom and \(T_{\nu,2}(\cdot,\cdot;\rho)\) the bivariate Student-t distribution function with the same degrees of freedom and latent correlation \(\rho\). The matched-threshold formulas are:
The Gaussian copula uses normal quantiles; the Student-t copula uses Student-t quantiles. Those are not interchangeable, even when the two copulas have the same \(\rho\). The value \(\rho\) is the correlation parameter of the underlying elliptical construction. With non-normal margins, it should not automatically be described as the Pearson correlation of the losses themselves.
A Student-t copula can be represented using a bivariate normal vector with correlation \(\rho\) and a single, shared positive scale factor derived from a chi-squared variable. When the shared scale is small, both standardized components can become large in magnitude together. This common scaling mechanism helps explain why Student-t copulas can have greater probability of joint extremes than Gaussian copulas with the same \(\rho\). The degrees of freedom \(\nu\) control the heaviness of the scale mixture: smaller \(\nu\) generally permits more pronounced clustering in the tails. It does not follow that every finite-threshold exceedance probability must increase monotonically in \(\nu\) across all choices of \(\rho\) and thresholds; calculate the probability for the specified model and threshold.
Worked Example: A Finite Stress Threshold With Zero Latent Correlation
Suppose two hypothetical losses have continuous margins, and stress is defined as exceeding each loss's 75th-percentile threshold. Compare (a) a Gaussian copula with \(\rho=0\) and (b) a Student-t copula with \(\nu=1\) and \(\rho=0\). Compute both joint exceedance probabilities. The loss margins are identical between the two models; their dollar thresholds need not be identical if the marginal distributions differ.
Gaussian method. With \(\rho=0\), the Gaussian copula gives independent uniform variables. Each loss exceeds its 75th-percentile threshold with probability \(1-0.75=0.25\), so
Student-t method. A Student-t distribution with one degree of freedom is standard Cauchy. Its 75th percentile is \(1\), so the event is \(Z_1>1,Z_2>1\) for a bivariate Student-t vector with \(\nu=1\) and \(\rho=0\). Its density is
Use polar coordinates \(x=r\cos\theta,\ y=r\sin\theta\). In the first quadrant, both coordinates exceed 1 when \(r\) exceeds the larger of \(1/\cos\theta\) and \(1/\sin\theta\). By symmetry about \(\theta=\pi/4\), the probability is
For the radial integral, \(\int_a^\infty r(1+r^2)^{-3/2}\,dr=1/\sqrt{1+a^2}\). For the angular integral, substitute \(z=\cos\theta\). The final probability is rounded to four decimal places. The check is that \(0.0833\) is below either marginal exceedance probability, \(0.25\), and above the independent-model value \(0.0625\). Thus, even with zero latent correlation, the shared Student-t scale produces more joint exceedances at this finite threshold. This example does not imply positive asymptotic upper-tail dependence when \(\rho=0\).
Common mistake avoided: setting \(\rho=0\) in a Student-t copula does not justify multiplying the marginal exceedance probabilities. That multiplication is valid for the Gaussian copula at \(\rho=0\), but the Student-t copula retains dependence through its shared scale.
What Changes at the Center of the Distribution?
The contrast between these copulas is not equally visible at every threshold. At the 50th percentile, the joint probability of two positive latent components in a bivariate elliptical model depends on \(\rho\), but not on the degrees of freedom. This makes a median-threshold calculation a useful check: it can confirm that two implementations share the same dependence parameter without testing how well they represent extreme co-movement.
Worked Example: A Median-Threshold Check Across Copulas
A hypothetical analyst sets the latent correlation to \(\rho=0.6\) and compares Gaussian and Student-t copulas with \(\nu=4\). Find the probability that both uniform variables exceed \(0.5\). State what the result does, and does not, show about the two models.
For a bivariate elliptical distribution with continuous symmetric margins, the probability that both components are positive is
At \(u=0.5\), both copula quantiles are zero. With \(\rho=0.6\), \(\arcsin(0.6)\approx0.6435\) radians, so
This answer applies to both the Gaussian copula and the Student-t copula with \(\nu=4\). It checks their central-quadrant association at the same \(\rho\), but it does not show that their probabilities agree at high thresholds. The calculation is checked by noting that it exceeds \(0.25\), the independent quadrant probability, because \(\rho\) is positive. The result is rounded to four decimal places.
Common mistake avoided: concluding that equal median-quadrant probabilities imply equal stress results. The formula only evaluates the central quadrant; extreme-threshold probabilities require the appropriate Gaussian or Student-t copula calculation.
Finite Exceedance and Tail Dependence
A joint exceedance probability at a selected level, such as \(u=0.99\), is a finite-threshold quantity. Upper-tail dependence instead describes the limiting fraction of exceedances of one variable that are accompanied by exceedances of the other as the threshold approaches the upper endpoint. For a continuous copula \(C\), its upper-tail dependence coefficient is
when the limit exists. The Gaussian copula has \(\lambda_U=0\) for \(-1<\rho<1\). For the Student-t copula with finite \(\nu\), the coefficient is
The coefficient is a limit, not a substitute for a finite-threshold calculation. For example, \(\lambda_U=0\) says that the conditional exceedance fraction tends to zero as \(u\) approaches 1; it does not say that joint exceedance at \(u=0.95\) is zero. Likewise, a positive Student-t coefficient does not provide the exact probability at a particular stress level.
Worked Example: Comparing Upper-Tail Dependence
Compare the upper-tail dependence coefficients of a Gaussian copula and a Student-t copula with \(\nu=1\), both using \(\rho=0.5\). Interpret the comparison without treating the limiting coefficient as a finite-threshold probability.
For the Gaussian copula, \(\lambda_U=0\) for this correlation, since \(-1<0.5<1\). For the Student-t copula, substitute \(\nu=1\) and \(\rho=0.5\):
For two degrees of freedom, \(T_2(x)=\frac12+\frac{x}{2\sqrt{x^2+2}}\). Hence \(T_2(-\sqrt{2/3})=0.25\), and the Student-t coefficient is \(2(0.25)=0.5\). The comparison is therefore \(0\) for the Gaussian copula and \(0.5\) for the Student-t copula. The latter value means that the limiting conditional exceedance fraction is one-half; it is not the probability that both losses exceed a particular dollar amount. The calculation is checked by observing that a coefficient must lie between 0 and 1. No rounding is needed for these exact results.
Common mistake avoided: reporting \(0.5\) as the joint probability of an event at a specified stress threshold. To obtain that probability, use \(1-2u+C(u,u)\) at the stated \(u\); the tail coefficient answers a different, limiting question.
Practical Comparison and Exam Traps
A disciplined comparison follows the same sequence for each model. First specify the marginal distributions and the marginal probability thresholds. Next specify the copula family and its parameters. Then evaluate the corresponding copula distribution function and apply inclusion-exclusion. If the question concerns tail dependence, calculate or interpret the limit separately.
Translate each loss cutoff into a probability threshold using its own marginal distribution. Check whether thresholds are equal in dollars or only equal in percentile.
Record the latent correlation \(\rho\), and, for the Student-t model, the degrees of freedom \(\nu\). Do not silently change one while comparing families.
Use the Gaussian or Student-t copula distribution function with its own marginal quantiles, followed by inclusion-exclusion.
A finite-threshold difference concerns the specified stress level; a tail-dependence coefficient concerns the limit as that level tends to one.
Bivariate normal and Student-t distribution functions are often evaluated numerically. In a numerical implementation, report the threshold and parameter values alongside the result, use the correct marginal quantiles for each copula, and retain enough precision until the final probability is reported. When comparing multiple values of \(\nu\), keep the margins, \(\rho\), and stress thresholds unchanged so that the comparison isolates the intended modeling change.
A further trap is to compare models at the same raw dollar cutoff when their margins differ, then attribute the entire difference to the copula. A higher joint exceedance probability may instead reflect a change in one or both marginal exceedance probabilities. To isolate dependence, compare at matched marginal probability thresholds, or explicitly report the different marginal probabilities. In either case, state the event precisely.
Check Your Understanding
For each problem, distinguish the specified finite-threshold event from any limiting tail statement.
- A Gaussian copula has \(\rho=0\), and both losses are tested at their 90th-percentile thresholds. Find the joint exceedance probability. (Answer: 0.0100)
- A bivariate Student-t copula has \(\nu=1,\rho=0\), with both uniform thresholds set to \(0.75\). Use the Cauchy polar-coordinate result to find the joint exceedance probability. (Answer: \(1/12\approx0.0833\))
- For either a Gaussian or Student-t copula with \(\rho=0.5\), find the probability that both uniform variables exceed \(0.5\). (Answer: \(\frac14+\frac{\arcsin(0.5)}{2\pi}\approx0.3333\))
- Find the Gaussian copula upper-tail dependence coefficient when \(\rho=0.7\). (Answer: 0)
- For a Student-t copula with \(\nu=1,\rho=0.5\), find \(\lambda_U\). (Answer: 0.5000)