What's the difference between Gaussian and Student-t copulas, and why does tail dependence matter?

Copulas are one of the most important (and most misunderstood) tools in quantitative risk management. Let's build this from the ground up.

What a Copula Does:

A copula separates the marginal distributions of individual variables from their dependence structure. Sklar's theorem guarantees that any joint distribution can be decomposed into:

F(x, y) = C(F_X(x), F_Y(y))

Where C is the copula function that captures how the variables move together.

Why This Matters:

You might have two assets where each has a skewed, fat-tailed return distribution. The copula lets you model their joint behavior without assuming they're jointly normal. You can pick any marginals and any copula.

Gaussian Copula:

• Assumes the dependence structure is that of a multivariate normal distribution
• Parameterized by a correlation matrix
• Zero tail dependence — as you move toward the extremes, the variables become independent
• Simple and tractable — this is why it was so popular for pricing CDOs

Student-t Copula:

• Assumes the dependence structure of a multivariate Student-t distribution
• Parameterized by a correlation matrix AND degrees of freedom (nu)
• Positive tail dependence — extreme events tend to cluster together
• As nu approaches infinity, it converges to the Gaussian copula

Tail Dependence — The Critical Distinction:

Tail dependence measures the probability that one variable has an extreme outcome given that the other already has. Formally:

lambda_L = lim P(Y < F_Y^{-1}(u) | X < F_X^{-1}(u)) as u -> 0

For the Gaussian copula: lambda = 0 (no tail dependence, regardless of correlation)

For the Student-t copula: lambda > 0 (increases with correlation and lower degrees of freedom)

The 2008 Lesson:

Banks modeled CDO default correlations using Gaussian copulas. During normal times, defaults were mildly correlated. But the Gaussian copula assumed that even if correlation was 0.3, the probability of many simultaneous defaults was essentially zero. When the housing crisis hit, defaults clustered catastrophically — exactly the tail dependence the Gaussian copula couldn't capture.

Harbor Point Capital's CDO desk, for instance, might have priced senior tranches as nearly risk-free because the Gaussian copula showed vanishing joint default probability. A Student-t copula with 4-5 degrees of freedom would have shown substantially higher prices for tail risk.

For the FRM exam, remember: Gaussian copula + high correlation still means zero tail dependence. That's the trap.

Dive deeper into dependence modeling in our FRM Part I Quantitative Analysis course.