Copula & Tail Dependence Portfolio Lab

1. Sklar's Theorem (1959) & Joint Decomposition

In standard portfolio theory, dependency is commonly simplified using Pearson linear correlation \(\rho\). However, linear correlation only completely describes dependency under elliptical, jointly Gaussian distributions. For arbitrary fat-tailed, skewed, or non-linear financial asset returns, linear correlation fails. Sklar's Theorem states that for any continuous multivariate distribution \(F(x_1, \dots, x_d)\) with marginals \(F_1, \dots, F_d\), there exists a unique copula \(C: [0, 1]^d \to [0, 1]\) such that: \[ F(x_1, x_2, \dots, x_d) = C\bigl(F_1(x_1), F_2(x_2), \dots, F_d(x_d)\bigr) \] This decoupling allows financial engineers to model individual asset marginal distributions (using generalized Pareto, Student-t, or empirical distributions) separately from their joint non-linear dependency copula.

2. The Gaussian Copula & The 2008 Blind Spot

The bivariate Gaussian copula is parameterized by the correlation matrix \(R\): \[ C_{\text{Gauss}}(u_1, u_2; \rho) = \Phi_\rho\bigl(\Phi^{-1}(u_1), \Phi^{-1}(u_2)\bigr) \] While mathematically convenient, the Gaussian copula possesses zero asymptotic tail dependence for any \(|\rho| < 1\): \[ \lambda_L = \lim_{u \to 0^+} \frac{C(u, u)}{u} = 0, \quad \lambda_U = \lim_{u \to 1^-} \frac{1 - 2u + C(u, u)}{1 - u} = 0 \] During the 2007–2008 subprime mortgage crisis, Collateralized Debt Obligation (CDO) tranches priced via David Li's Gaussian copula severely underestimated the probability of simultaneous defaults. In systemic market panics, defaults cluster aggressively, violating the Gaussian assumption of asymptotic independence.

3. Student-t Copula: Symmetric Extreme Dispersion

The Student-t copula introduces a scalar mixing variable \(W \sim \chi^2(\nu)/\nu\) that scales joint asset co-movements: \[ C_t(u_1, u_2; \rho, \nu) = t_{\nu, \rho}\bigl(t_\nu^{-1}(u_1), t_\nu^{-1}(u_2)\bigr) \] Unlike the Gaussian copula, the Student-t copula generates strictly positive symmetric tail dependence: \[ \lambda_L = \lambda_U = 2 t_{\nu+1}\left( -\sqrt{\nu + 1} \sqrt{\frac{1 - \rho}{1 + \rho}} \right) > 0 \] As degrees of freedom \(\nu \to \infty\), the Student-t copula converges asymptotically to the Gaussian copula.

4. Archimedean Copulas: Clayton & Gumbel Asymmetry

Archimedean copulas are generated via an additive generator function \(\phi(t)\):

• Clayton Copula: Generator \(\phi(t) = \frac{1}{\theta}(t^{-\theta} - 1)\) for \(\theta > 0\). Exhibits strong lower tail dependence: \[ \lambda_L = 2^{-1/\theta}, \quad \lambda_U = 0 \] Ideal for equity crash contagion, interbank liquidity freezes, and credit default co-movements.
• Gumbel Copula: Generator \(\phi(t) = (-\ln t)^\theta\) for \(\theta \ge 1\). Exhibits strong upper tail dependence: \[ \lambda_L = 0, \quad \lambda_U = 2 - 2^{1/\theta} \] Ideal for speculative market bubbles, momentum contagion, and multi-asset upside rallies.