4.12 Copulas & Dependence

Basic Prerequisites: Probability & Measure Theory (CDFs, the quantile transform) and Credit Risk & the Merton Model (the Vašíček one-factor model). (these are the folder-level prerequisites for pages 02–06; page 01 states its own, smaller, entry requirements)

1. Intuition & Practical Objective

Every risk model has two halves. The first is the marginals - how risky is each position on its own? The second is the dependence structure - when one position goes bad, what does the other one do? This folder is about the second half, and about the single most consequential fact in all of quantitative risk management:

The marginal distributions do not determine the joint distribution. Two books can have literally identical position-level loss distributions and VaRs that differ by a factor of three, because the way the positions move together is a separate modelling choice.

The word for that separate choice is a copula. A copula is the mathematical object that glues fixed margins into a joint distribution, and one theorem - Sklar’s (1959) - says this decomposition is always possible and (for continuous margins) unique:

This folder is the model topic-folder for the Kwant-Atlas build. It is a hub: it (a) gives the fast formula lookup below (job #1 of this pillar), covering Sklar’s theorem, the Gaussian, $t$, Gumbel and Clayton copulas, rank correlation, tail dependence, and the Vašíček one-factor / Gaussian-copula portfolio-credit layer, and (b) routes you to six sub-pages that walk from raw intuition through the theorem, the 2008 Gaussian-copula episode, tail dependence, the practice of fitting and its failure modes, and the Archimedean / portfolio-credit extensions.

The one-sentence essence. “A joint distribution is a copula acting on its margins; risk lives in the copula, the margins are ordinary, and the Gaussian copula’s zero tail dependence is the reason a portfolio model can be blind to the one event - a joint crash - it exists to survive.”

2. Mathematical Ground Truth & the Copula Lookup

Quick-Reference Lookup (job #1). All formulas below are transcribed from the verified corpus - McNeil, Frey & Embrechts (2015) Ch 7 (definitions 7.1, Sklar Thm 7.3, Fréchet bounds Thm 7.8, Gauss copula (7.10), $t$ copula (7.11), Gumbel (7.12), Clayton (7.13), Kendall (7.28), Spearman (7.33), tail dependence (7.34)–(7.38), Table 7.1) and Vašíček (1987/1991), cross-checked against Bluhm (2010) and Bielecki & Rutkowski (2002). Every number in the check column was re-executed and reproduced exactly (§3 and the sub-pages).

Notation. $C$ copula, $u_i=F_i(x_i)$ uniform margins, $\Phi$ standard-normal CDF, ${\Phi }_P$ joint normal CDF with correlation matrix $P$, $t_{\nu }$ Student-$t$ CDF with $\nu$ d.o.f., $\hat{C}$ survival copula, $p$ default probability, $\rho$ asset correlation, $\varrho$ generic copula parameter.

Critical caveat. The copula is not identified by the correlation alone. Linear (Pearson) correlation is not invariant under monotone transforms and can be $\pm 1$ only under elliptical dependence; rank correlations (Kendall, Spearman) are copula functionals and are the right calibration input. And a copula fitted to calm-regime data says nothing about the crisis regime - the single largest source of error in this folder (sub-page 05).

3. Computational Implementation - the hub engine

Standard library only (math.erf gives the exact normal CDF). It reproduces the headline numbers above: the Gaussian copula’s joint-tail excess over independence, and the Vašíček one-factor loss quantiles.

The joint-tail excess is the entire content of dependence risk: at the $99\%$ level the Gaussian copula puts $25\times$ more mass in the joint tail than independence would. The sub-pages show where even this is not enough (page 04: it is asymptotically $0$, not merely small).

4. Failure Modes & First-Principles Breakdowns

Hub signposts - the folder’s failure-mode analysis lives in 05 · Failure Modes & Practice. In one line each:

1. Zero tail dependence. The Gaussian copula is asymptotically independent for every $\varrho <1$; the $t$-copula (or an EVT tail) is required whenever the joint extreme is the risk being measured.
2. Static correlation in a crisis. $\rho$ estimated through the cycle is not $\rho$ in a crisis; a single copula calibrated on calm data understates joint extremes by an order of magnitude (page 05).
3. Model uncertainty is the dominant term. The copula family, the margins and the parameter are all estimated with wide error bands; the difference between two defensible copulas exceeds the difference between most portfolios (page 05, and see Model Risk & Validation).

5. References

• McNeil, Frey & Embrechts: Quantitative Risk Management: Concepts, Techniques and Tools (rev. ed. 2015, Princeton)
• Nelsen, Roger B.: An Introduction to Copulas (2nd ed., 2006, Springer)
• Joe, Harry: Multivariate Models and Dependence Concepts (1997)
• Bluhm, Overbeck & Wagner: Introduction to Credit Risk Modeling (2nd ed., 2010)
• Bielecki & Rutkowski: Credit Risk: Modeling, Valuation and Hedging (2002, Springer)
• Li, David X.: On Default Correlation: A Copula Function Approach, Journal of Fixed Income 9(4):43–54 (2000)
• Salmon, Felix: Recipe for Disaster: The Formula That Killed Wall Street, Wired (23 Feb 2009)

6. Connected Graph Bridges

• Foundational base: Probability & Measure Theory (CDFs, quantile transform, joint laws) · Statistics & Inference (estimators)
• Sibling topics: Credit Risk & the Merton Model (the one-factor case) · Extreme Value Theory & Fat Tails (multivariate EVT) · VaR & Expected Shortfall (the portfolio quantile that depends on the copula)
• Sub-pages (in-folder): 01 From Zero · 02 Sklar’s Theorem & Copulas · 03 The Gaussian Copula & 2008 · 04 Tail Dependence & the $t$-Copula · 05 Failure Modes & Practice · 06 Advanced Extensions (Archimedean, portfolio credit)

Beginner: start at 01 · Practitioner: start at 05