Clayton, Gumbel, and the Smile of Tail Dependence

The $t$-copula handles symmetric tail dependence: two assets are as likely to reach extremes up together as down together. Real financial markets do not oblige. Downside moves cluster more than upside — on the S&P 500’s worst 1% of days, almost every other equity index is having its own worst 1% at the same time; on the S&P’s best 1% of days, the co-movement is much weaker. That is the empirical regularity nailed down independently by Longin & Solnik (2001) and Ang & Chen (2002).

For risk management this asymmetry is not one you can afford to get wrong: a model with symmetric tail dependence errs in exactly the direction that hurts. Underestimating downside clustering is the same thing as overestimating diversification benefit at the one moment diversification should not be trusted. What we need is a family of copulas whose mathematical structure lets the lower and upper tails be governed separately — and that is the Archimedean copulas: Clayton has only lower-tail dependence, Gumbel and Joe only upper, Frank neither. This article builds the family up from its generator, derives each member’s tail dependence formula, and ends with how to recognize which one your data is asking for.

One Function Generates the Whole Family

The entire structure of an Archimedean copula is determined by a single function $\varphi$, the generator:

where $\theta$ is a parameter and $\varphi: [0,1] \to [0, \infty]$ is continuous, decreasing, convex, with $\varphi(1) = 0$ and $\varphi(0) = \infty$. Any $\varphi$ meeting these conditions yields a valid copula. The elegance of the construction is that multivariate dependence collapses into “transform each coordinate, add, transform back” — so every property of the copula, including the behavior of each tail, is encoded in the shape of the generator near 0 and 1. Choosing a different $\varphi$ means choosing a different tail temperament; that is the lever this article pulls repeatedly.

Three structural features of the family are worth laying on the table up front. It is exchangeable: the copula is symmetric in its arguments, so the dependence structure is the same however you label the assets — a strong assumption, convenient now, and a liability in high dimensions (we settle that bill later). It typically carries one parameter: $\theta$ has a clean interpretation and calibrates directly. And its real value is that tails can be asymmetric: different members put dependence in the upper tail, the lower, both, or neither — something the Gaussian and $t$-copulas cannot offer at any parameter value. For the multivariate form to be computable, $\varphi$ needs a closed-form inverse; Clayton, Gumbel, and Frank all deliver one.

Clayton: Lower Tail Only

The Clayton copula has generator

with inverse $\varphi^{-1}(s) = (1 + \theta s)^{-1/\theta}$; substituting into the definition gives

Two endpoints anchor the intuition: as $\theta \to 0$ this degenerates to the independence copula $C(u,v) = uv$, and as $\theta \to \infty$ to the comonotonic copula $C(u,v) = \min(u,v)$ (perfect dependence). Every $\theta$ in between carries an exact lower tail dependence coefficient:

As $\theta \to 0$, $\lambda_L \to 0$; as $\theta \to \infty$, $\lambda_L \to 1$; monotone increasing in $\theta$. To get a feel for the magnitudes:

On the upper side:

The asymmetry is structural: as $u, v \to 0$ the joint density blows up, while as $u, v \to 1$ it vanishes and the joint probability falls back to the independence prediction $u \cdot v$. Translated: Clayton says “these assets huddle together on the way down more than on the way up” — precisely the empirical shape of daily equity data, with $\theta$ setting how tight the huddle is.

One sign convention has to be watched: the $u, v$ here are uniformized returns, so crashes live in the lower tail. If your variable is defined as a loss, crashes move to the upper tail and the roles of Clayton and Gumbel swap entirely (equivalently, flip via the survival copula). Confirm which way your tail points before picking a family.

Gumbel: Upper Tail Only

Clayton locks down the lower tail; the mirror-image need — clustering only in the upper tail — is met by the Gumbel copula. Its generator is

with inverse $\varphi^{-1}(s) = \exp(-s^{1/\theta})$, giving

At $\theta = 1$ this is exactly the independence copula; as $\theta \to \infty$ it approaches comonotonicity. The upper tail dependence is

zero at $\theta = 1$ and approaching 1 as $\theta \to \infty$:

And the lower tail gets nothing:

Under the returns convention, Gumbel models co-movement in benign states — everything rallying together; under the loss convention it becomes the model of crash clustering. Gumbel also holds a singular credential: it is the only copula that belongs to both the Archimedean family and the extreme-value copula family (Gumbel, 1960, was born in multivariate extremes), which makes it particularly convenient wherever the modeling connects to EVT.

See the Asymmetry Yourself

The formulas say Clayton has a lower tail and no upper tail; simulate five thousand points and check. Sampling uses the gamma-frailty construction (the Marshall–Olkin method): a shared shock $G \sim \text{Gamma}(1/\theta)$ divides independent exponentials, and the transform lands exactly on Clayton:

import numpy as np

# Simulate from Clayton copula (Marshall-Olkin / gamma frailty method)
np.random.seed(11)
theta = 2.0
n = 5000

def clayton_sim(theta, n):
    g = np.random.gamma(1/theta, 1, size=n)
    e1 = np.random.exponential(1, size=n)
    e2 = np.random.exponential(1, size=n)
    u = (1 + e1/g)**(-1/theta)
    v = (1 + e2/g)**(-1/theta)
    return u, v

U, V = clayton_sim(theta, n)

# Diagnose asymmetric tail dependence
q = 0.05
empirical_lower = ((U < q) & (V < q)).mean() / q
empirical_upper = ((U > 1-q) & (V > 1-q)).mean() / q

print(f"Empirical lower-tail cond dependence: {empirical_lower:.3f}")
print(f"Empirical upper-tail cond dependence: {empirical_upper:.3f}")
print(f"Theoretical lower tail dependence: {2**(-1/theta):.3f}")
print(f"Theoretical upper tail dependence: 0 (Clayton has none)")

The output shows lower-tail conditional dependence far above upper — the shape Clayton was built to capture. The Gaussian and $t$-copulas cannot produce it in principle: both force $\lambda_L = \lambda_U$.

Joe: Gumbel’s Other Upper-Tail Relative

Gumbel has no monopoly on the upper tail. The Joe copula has the explicit form

for $\theta \ge 1$, with generator

At $\theta = 1$ it degenerates to independence; as $\theta \to \infty$ it approaches comonotonicity. Its tail profile matches Gumbel’s:

Same upper-tail formula, but Joe is a different family: a different generator, a different dependence shape in the body — at equal $\lambda_U$, Joe pushes more mass toward the extreme and sits closer to independence in the body. So choosing between Gumbel and Joe is not a disagreement about tail coefficients; it is a choice about what you want the body to look like. In low dimensions Joe is sometimes preferred for its simpler likelihood surface; the price is the same as Gumbel’s — a total absence of lower-tail dependence.

With all three asymmetric members assembled, fold Frank into the reference table:

Frank: Neither Tail

The last row of the table deserves its own explanation. The Frank copula has generator

inverse $\varphi^{-1}(s) = -\log(1 + (e^{-\theta} - 1) e^{-s}) / \theta$, and copula

As $\theta \to 0$ it is independence; $\theta > 0$ gives positive dependence, $\theta < 0$ negative — it is one of the few family members that can express negative dependence at all. But for any $\theta \ne 0$, $\lambda_L = \lambda_U = 0$: correlation in the body, tails clean on both sides. That is precisely its niche — when the data has genuinely weak overall dependence and no tail clustering, Frank fits the body without smuggling in spurious tail dependence. And the way to misuse it is equally clear: put it on data with visible tail clustering and its good body fit will paper over the tail mismatch.

Estimation: Backing $\theta$ Out of Kendall’s $\tau$

The one-parameter design pays off at calibration time: a single rank statistic suffices. For Archimedean copulas, Kendall’s $\tau$ relates to the generator through one elegant general formula (Nelsen, 2006, §5.1):

Substituting each family’s $\varphi$ gives its specific relationship. Clayton comes out in closed form:

So does Gumbel:

Frank’s answer cannot avoid the Debye function $D_1$:

The estimation recipe is then: compute the sample $\hat\tau$, invert to $\hat\theta$. Clayton and Gumbel invert in simple closed form:

Frank and Joe have no elementary closed form for $\tau(\theta)$ (Joe’s integral from the general formula must be evaluated numerically), so invert with a one-dimensional root-finder or go straight to MLE on the likelihood. A sanity check on independence comes free: at $\hat\tau = 0$, the Clayton inversion gives $\hat\theta = 0$, Gumbel gives $\hat\theta = 1$, Frank gives $\hat\theta = 0$ — each landing exactly on its own independence point. This rank-correlation calibration (Genest & Favre, 2007) is often near-optimal for one-parameter Archimedeans; to refine, run full copula log-likelihood MLE over $\theta$.

Choosing the Family: Let the Tails of the Data Speak

Parameters are easy; the real decision is which family, and the answer lives in the asymmetry this article opened with. Step one is always to compute the empirical upper and lower tail dependence coefficients. If $\hat\lambda_L > \hat\lambda_U$, Clayton enters the shortlist; if $\hat\lambda_U > \hat\lambda_L$, Gumbel or Joe; if the two are comparable and both positive, no single Archimedean member can produce that shape — go back to the $t$-copula. With the direction settled, let information criteria arbitrate the details: fit the candidates by MLE and compare AIC or BIC (AIC leans permissive, BIC conservative). For harder evidence, run a goodness-of-fit test — the Cramér–von Mises statistic $T_n = \sum_{i} \left(C_n(\mathbf{u}_i) - C_{\hat\theta}(\mathbf{u}_i)\right)^2$, with $C_n$ the empirical copula evaluated at the pseudo-observations and the null distribution obtained by bootstrap; reject when $T_n$ lands in the upper tail. This is the strictest gate, and the most compute-hungry.

Once chosen and fitted, risk measures go through simulation: two-stage MLE for marginals and copula, ten thousand or more joint scenarios, portfolio loss per scenario, then the 99% (or 97.5%) quantile and tail mean. How much the answer moves depends on $\theta$ and the portfolio: with pronounced lower-tail dependence, the joint probability of several components sitting in their worst 1% simultaneously runs well above the Gaussian-copula figure, lifting portfolio VaR accordingly and ES more — anywhere from a few percent to tens of percent depending on sample and portfolio, which is why reports should carry the estimate’s sensitivity to $\theta$ rather than a bare point estimate. Clayton $\theta$ fitted on daily equity data commonly lands around 1 to 2 (i.e. $\lambda_L$ roughly 0.5–0.7), but the number drifts visibly with sample period and market regime — do not treat it as a constant.

The Family’s Ceiling

Three structural limits mark how far Archimedeans go. First, high dimensions. When $\varphi^{-1}$ is completely monotone (true for Clayton with $\theta > 0$, Gumbel, and Frank with $\theta > 0$), the exchangeable multivariate version is well defined — but exchangeability turns from convenience into shackles as $n$ grows: every pair of assets is forced to share the same $\theta$, and thirty stocks with identical pairwise dependence is plainly false. The standard escape at scale is the vine copula: build the joint from pair-copulas, each pair free to take its own family and parameter (Aas et al., 2009). Second, exchangeability again from another angle: Archimedeans cannot express directional dependence — “asset 1 drags down asset 2 harder than the reverse” — and when that matters you need non-exchangeable constructions, such as the Marshall–Olkin copula or Khoudraji-type asymmetrizations of an Archimedean. Third, time: every copula here assumes static dependence, while real market dependence rides volatility regimes; the common engineering answer is DCC-GARCH plus copula — let the correlation move, let the copula govern shape.

Compressed into one actionable sentence, the article’s judgment reads: compute empirical tail dependence first and let the direction of the asymmetry choose the family — most equity portfolios will point you at Clayton; start from $\tau$-calibration, refine with MLE, gate with Cramér–von Mises. The most popular copula is rarely the right answer; for most equity portfolios, Clayton for downside stress testing and the $t$-copula for general portfolio risk is a far more honest starting point than “default Gaussian.”

References

Aas, K., Czado, C., Frigessi, A., & Bakken, H. (2009). Pair-copula constructions of multiple dependence. Insurance: Mathematics and Economics, 44(2), 182-198.

Ang, A., & Chen, J. (2002). Asymmetric correlations of equity portfolios. Journal of Financial Economics, 63(3), 443-494.

Clayton, D. G. (1978). A model for association in bivariate life tables and its application in epidemiological studies of familial tendency in chronic disease incidence. Biometrika, 65(1), 141-151.

Genest, C., & Favre, A.-C. (2007). Everything you always wanted to know about copula modeling but were afraid to ask. Journal of Hydrologic Engineering, 12(4), 347-368.

Gumbel, E. J. (1960). Distributions des valeurs extremes en plusieurs dimensions. Publications de l’Institut de Statistique de l’Université de Paris, 9, 171-173.

Longin, F., & Solnik, B. (2001). Extreme correlation of international equity markets. Journal of Finance, 56(2), 649-676.

McNeil, A. J., Frey, R., & Embrechts, P. (2015). Quantitative Risk Management: Concepts, Techniques and Tools (rev. ed.). Princeton University Press. Ch. 7.

Nelsen, R. B. (2006). An Introduction to Copulas (2nd ed.). Springer.