5D Library › Macro-Liquidity Regime Switching, Extreme Tail Dependence and Barbell Asset Allocation · 中文

Chapter 3 Copulas and Tail Dependence

A question correlation cannot answer

Two assets have a correlation of 0.5. How likely are they to fall together on the worst 1 % of days? The correlation coefficient cannot say. Pearson correlation measures linear association and depends on the shape of the marginal distributions; zero correlation is not independence; and the same correlation can belong to joint distributions with entirely different tails. Embrechts, McNeil and Straumann set out these “pitfalls” systematically before the crisis (Embrechts et al. 2002). To discuss “falling together” on its own, the part of a joint distribution that describes dependence must be separated from the parts that describe each asset’s own behaviour. That is what a copula does.

Sklar’s theorem

Evidence

Sklar’s theorem (Sklar 1959; Nelsen 2006): any joint distribution function F(x1,…,xd)F(x_1, \dots, x_d) can be written as

F(x1,…,xd)=C(F1(x1),…,Fd(xd)),F(x_1, \dots, x_d) = C\big(F_1(x_1), \dots, F_d(x_d)\big),

where the FiF_i are the marginal distributions and CC, a joint distribution function on [0,1]d[0,1]^d with uniform margins, is called the copula; CC is unique when the margins are continuous. Conversely, any copula combined with any margins defines a valid joint distribution through the same formula.

Modelling can therefore proceed in two steps: first choose a marginal distribution for each asset (for example, the fat-tailed distributions of Chapter 2), then choose a copula for the dependence structure separately. A copula is invariant under increasing transformations of each variable, so the matching measures of dependence are rank correlations such as Kendall’s τ. This chapter uses four copulas:

  • the Gaussian copula and the Student t copula, with correlation parameter ρ\rho (and for the t, degrees of freedom ν\nu), both with τ=2πarcsin⁡ρ\tau = \frac{2}{\pi}\arcsin\rho;
  • the Clayton copula C(u,v)=(u−θ+v−θ−1)−1/θC(u,v) = (u^{-\theta} + v^{-\theta} – 1)^{-1/\theta}, with τ=θ/(θ+2)\tau = \theta/(\theta + 2);
  • the Gumbel copula C(u,v)=exp⁡{−[(−ln⁡u)θ+(−ln⁡v)θ]1/θ}C(u,v) = \exp\{-[(-\ln u)^\theta + (-\ln v)^\theta]^{1/\theta}\}, with τ=1−1/θ\tau = 1 – 1/\theta.

Tail-dependence coefficients

The lower and upper tail-dependence coefficients are

λL=limu→0+P(V≤u∣U≤u)=limu→0+C(u,u)u,λU=limu→1−P(V>u∣U>u).\lambda_L = \lim_{u \to 0^+} P(V \le u \mid U \le u) = \lim_{u\to 0^+}\frac{C(u,u)}{u}, \qquad \lambda_U = \lim_{u \to 1^-} P(V > u \mid U > u).

They ask whether, as one asset falls into ever more extreme parts of its tail, the conditional probability that the other falls into an equally extreme part stays above zero in the limit.

Lower tail of the Clayton copula. C(u,u)=(2u−θ−1)−1/θ=u(2−uθ)−1/θC(u,u) = (2u^{-\theta} – 1)^{-1/\theta} = u\,(2 – u^{\theta})^{-1/\theta}, so C(u,u)/u→2−1/θC(u,u)/u \to 2^{-1/\theta}; that is, λL=2−1/θ\lambda_L = 2^{-1/\theta}. Its upper-tail coefficient is zero.

Upper tail of the Gumbel copula. On the diagonal C(u,u)=u21/θC(u,u) = u^{2^{1/\theta}}, so

λU=limu→1−1−2u+u21/θ1−u=2−21/θ\lambda_U = \lim_{u\to 1^-}\frac{1 – 2u + u^{2^{1/\theta}}}{1 – u} = 2 – 2^{1/\theta}

(by l’Hôpital’s rule). Its lower-tail coefficient is zero.

The Gaussian copula. For ρ<1\rho < 1, λL=λU=0\lambda_L = \lambda_U = 0, a property called “asymptotic independence”: the conditional probability P(V≤u∣U≤u)P(V \le u \mid U \le u) tends to zero as u→0u \to 0, but slowly. The Student t copula has symmetric, positive tail dependence λ=2tν+1(−(ν+1)(1−ρ)/(1+ρ))\lambda = 2\,t_{\nu+1}\!\left(-\sqrt{(\nu+1)(1-\rho)/(1+\rho)}\right) (Embrechts et al. 2002; McNeil et al. 2015).

The three formulas in the author’s framework (0 for the Gaussian, 2−1/θ2^{-1/\theta} for the Clayton, 2−21/θ2 – 2^{1/\theta} for the Gumbel) are all correct. What must be added is direction: the Clayton copula has only lower-tail dependence and the Gumbel copula only upper-tail dependence. Using a Gumbel copula for joint crashes requires its version rotated by 180 degrees (the “survival” copula); otherwise the model describes joint booms.

The program below gives the four copulas the same Kendall’s τ = 1/3 (Gaussian and t with ρ=0.5\rho = 0.5, the t with 4 degrees of freedom; Clayton θ=1\theta = 1; Gumbel θ=1.5\theta = 1.5), simulates 400,000 pairs from each, and computes the conditional probabilities at the 5 %, 1 % and 0.2 % quantiles for comparison with the theoretical limits.

Calculation

# Tail dependence of four copulas with the SAME Kendall's tau = 1/3: Gaussian (rho 0.5),
# Student t (rho 0.5, 4 d.f.), Clayton (theta 1) and Gumbel (theta 1.5). Empirical
# lambda_L(u) = P(V < u | U < u) and lambda_U(u) = P(V > 1-u | U > 1-u) from 400,000
# simulated pairs, against the theoretical limits as u -> 0. Margins do not matter: the
# copula is invariant to increasing transformations, so empirical quantiles suffice.
import math, random

N = 400_000

def gaussian(rng, rho=0.5, df=None):
    out = []
    for _ in range(N):
        a = rng.gauss(0, 1); b = rho * a + math.sqrt(1 - rho * rho) * rng.gauss(0, 1)
        if df:
            w = math.sqrt(rng.gammavariate(df / 2, 2) / df); a, b = a / w, b / w
        out.append((a, b))
    return out

def clayton(rng, th=1.0):                     # Marshall-Olkin: gamma frailty
    out = []
    for _ in range(N):
        v = rng.gammavariate(1 / th, 1)
        out.append(tuple((1 + rng.expovariate(1) / v) ** (-1 / th) for _ in range(2)))
    return out

def gumbel(rng, th=1.5):                      # Marshall-Olkin: positive stable frailty
    a, out = 1 / th, []
    for _ in range(N):
        w, e = rng.uniform(0, math.pi), rng.expovariate(1)
        s = (math.sin(a * w) / math.sin(w) ** (1 / a)
             * (math.sin((1 - a) * w) / e) ** ((1 - a) / a))
        out.append(tuple(math.exp(-(rng.expovariate(1) / s) ** a) for _ in range(2)))
    return out

def lam(pairs, u):
    xs = sorted(p[0] for p in pairs); ys = sorted(p[1] for p in pairs); n = len(pairs)
    k = int(u * n); xl, yl, xh, yh = xs[k], ys[k], xs[n - k - 1], ys[n - k - 1]
    lo = sum(1 for x, y in pairs if x < xl and y < yl) / k
    hi = sum(1 for x, y in pairs if x > xh and y > yh) / k
    return lo, hi

def t_cdf(x, df, steps=4000):                 # Simpson's rule on the density, x < 0
    c = math.exp(math.lgamma((df + 1) / 2) - math.lgamma(df / 2))
    c /= math.sqrt(df * math.pi)
    f = lambda t: c * (1 + t * t / df) ** (-(df + 1) / 2)
    a, h = x - 400.0, 400.0 / steps
    s = f(a) + f(x) + sum((4 if i % 2 else 2) * f(a + i * h) for i in range(1, steps))
    return s * h / 3

if __name__ == "__main__":
    rng = random.Random(2002)
    lt = 2 * t_cdf(-math.sqrt(5 * 0.5 / 1.5), 5)
    models = (("Gaussian", gaussian(rng), 0.0, 0.0),
              ("t (4 d.f.)", gaussian(rng, df=4), lt, lt),
              ("Clayton", clayton(rng), 2 ** -1.0, 0.0),
              ("Gumbel", gumbel(rng), 0.0, 2 - 2 ** (1 / 1.5)))
    us = (0.05, 0.01, 0.002)
    res = {name: [lam(pairs, u) for u in us] for name, pairs, a, b in models}
    for side, j in (("lower", 0), ("upper", 1)):
        print(f"{side} tail   u = 0.05    0.01   0.002    limit as u -> 0")
        for name, pairs, l_lo, l_hi in models:
            print(f"  {name:12s}" + "".join(f"{r[j]:8.3f}" for r in res[name])
                  + f"   {(l_lo, l_hi)[j]:6.3f}")
    print("P(both in their worst 1 %) per model (independence: 0.0001):")
    print("   " + "   ".join(f"{name} {res[name][1][0] * 0.01:.4f}"
                             for name, p, a, b in models))
# lower tail   u = 0.05    0.01   0.002    limit as u -> 0
#   Gaussian       0.249   0.125   0.065    0.000
#   t (4 d.f.)     0.333   0.282   0.265    0.253
#   Clayton        0.510   0.505   0.506    0.500
#   Gumbel         0.170   0.063   0.026    0.000
# upper tail   u = 0.05    0.01   0.002    limit as u -> 0
#   Gaussian       0.245   0.125   0.071    0.000
#   t (4 d.f.)     0.334   0.275   0.242    0.253
#   Clayton        0.096   0.022   0.003    0.000
#   Gumbel         0.432   0.417   0.376    0.413
# P(both in their worst 1 %) per model (independence: 0.0001):
#    Gaussian 0.0013   t (4 d.f.) 0.0028   Clayton 0.0050   Gumbel 0.0006

The four copulas share the same Kendall’s τ but have entirely different tails. The Clayton copula’s lower-tail conditional probability is about 0.51 at all three quantiles, in line with its limit of 0.500, while its upper tail quickly goes to zero. The Gumbel copula is the opposite: 0.43, 0.42 and 0.38 in the upper tail, heading for 0.413, and its lower tail goes to zero. The t copula is symmetric, about 0.24 to 0.27 at the 0.2 % quantile, near its limit of 0.253. The Gaussian copula’s conditional probability falls from 0.25 to 0.125 and then to about 0.07: it does go to zero, but at the 1 % quantile there is still a one-in-eight chance of both assets falling into their worst 1 %. The last line turns this into the probability that both assets fall into their worst 1 % together: 0.0001 under independence, 0.0013 for the Gaussian copula, 0.0028 for the t and 0.0050 for the Clayton; the Gumbel copula gives only 0.0006, because its dependence sits in the upper tail. Figure 3 plots the four copulas in normal scores; the points in the lower-left and upper-right corners are the days on which both assets fall into their lowest and highest 5 %.

Four copulas, each with Kendall’s τ = 1/3 (1,500 simulated points each, converted by rank to normal scores). Purple: both assets in their lowest 5 %; teal: both in their highest 5 %. Dotted lines mark the 5 % and 95 % quantiles.

So “the Gaussian copula has zero tail dependence” is an accurate statement about a limit; it cannot be read as “the Gaussian copula holds that extreme losses never come together”. Its problem is that the further into the tail one goes, the faster the probability it gives for joint extremes falls relative to the t or Clayton copulas: at the 1 % quantile they differ by a factor of about 2 to 4, and at more extreme quantiles by more.

The Gaussian copula and the criticism of 2008

Evidence

David X. Li proposed in 2000 to describe the dependence between the default times of several obligors with a copula, taking the Gaussian copula as his main example, with the correlation parameter derived from the correlation of asset returns (Li 2000). The one-factor Gaussian copula then became the market standard for pricing and quoting the tranches of collateralised debt obligations (CDOs). A 2009 article in Wired called it “the formula that killed Wall Street” (Salmon 2009). After interviewing practitioners in investment banks, MacKenzie and Spears concluded that practitioners had deep misgivings about the standard models before the crisis (for example, unstable implied tranche correlations that needed patches such as “base correlation”), and that the crisis’s central mechanism was associated with simpler uses of the Gaussian copula (developed in a companion article), not with the more sophisticated versions used in banks’ derivatives departments (MacKenzie and Spears 2014). Embrechts and co-authors had pointed out well before the crisis that the Gaussian copula has no tail dependence and therefore understates joint extremes (Embrechts et al. 2002).

An accurate statement of the 2008 criticism is therefore this: the Gaussian copula has zero tail dependence and at extreme quantiles gives much lower probabilities of joint default than models with tail dependence; its single correlation parameter was usually calibrated on calm periods; and practitioners knew these limits, yet the model was widely used for quoting and rating. The crisis also had causes in housing, leverage, incentives and regulation that this book does not discuss, so the book does not say that “a formula caused the financial crisis”.

Note

Estimating tail dependence is hard. Any finite sample estimates only the conditional probability at some quantile, not the limit: in ten years of daily data the worst 1 % is just 25 days. In practice one compares the likelihoods of several copula families or models the joint tail with the methods of Chapter 2 (McNeil et al. 2015).

Chapter summary

Sklar’s theorem splits a joint distribution into margins and a copula, so that the dependence structure can be modelled separately; rank correlations and tail-dependence coefficients are measures at the level of the copula. The Gaussian copula (ρ<1\rho < 1) has zero tail dependence, but only in the limit: with correlation 0.5, the conditional probability of a joint crash at the 1 % quantile is still 0.125. The Clayton copula has only lower-tail dependence, 2−1/θ2^{-1/\theta}, the Gumbel copula only upper-tail dependence, 2−21/θ2 – 2^{1/\theta}, and the t copula symmetric dependence in both tails. With the same Kendall’s τ, the four copulas give probabilities of both assets falling into their worst 1 % that differ by a factor of more than eight. The accurate core of the 2008 criticism of the Gaussian copula is zero tail dependence and calibration on calm periods, not that “a formula killed Wall Street”.

References

Embrechts, Paul, Alexander J. McNeil, and Daniel Straumann. 2002. “Correlation and Dependence in Risk Management: Properties and Pitfalls.” In Risk Management: Value at Risk and Beyond, edited by M. A. H. Dempster. Cambridge University Press.
Li, David X. 2000. “On Default Correlation: A Copula Function Approach.” The Journal of Fixed Income 9 (4): 43–54. https://doi.org/10.3905/jfi.2000.319253.
MacKenzie, Donald, and Taylor Spears. 2014. “‘The Formula That Killed Wall Street’: The Gaussian Copula and Modelling Practices in Investment Banking.” Social Studies of Science 44 (3): 393–417. https://doi.org/10.1177/0306312713517157.
McNeil, Alexander J., Rüdiger Frey, and Paul Embrechts. 2015. Quantitative Risk Management: Concepts, Techniques and Tools. Revised. Princeton University Press.
Nelsen, Roger B. 2006. An Introduction to Copulas. 2nd ed. Springer.
Salmon, Felix. 2009. “Recipe for Disaster: The Formula That Killed Wall Street.” Wired 17 (3).
Sklar, Abe. 1959. “Fonctions de répartition à nn Dimensions Et Leurs Marges.” Publications de l’Institut de Statistique de l’Université de Paris 8: 229–31.

Fan, Y. (2026). Macro-Liquidity Regime Switching, Extreme Tail Dependence and Barbell Asset Allocation: Vine Copulas, Hidden Markov Models and Sovereign Buffer Dynamics [宏观流动性体制转换、极值尾部相依与杠铃资产配置] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23828049 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).