5D Library › Macro-Liquidity Regime Switching, Extreme Tail Dependence and Barbell Asset Allocation · 中文
Chapter 4 Correlation in Crises and Vine Copulas
“Correlations go to one in a crisis”
The author’s framework repeats a common saying: when a crisis comes, correlations between assets approach 1 and diversification fails. The saying contains a real phenomenon and a statistical trap. This chapter first sets the evidence on both sides next to each other and then discusses how to describe asymmetric tail dependence among several assets: vine copulas.
The trap: correlation conditional on high volatility
Suppose two markets’ returns satisfy , with independent of . The correlation is . If in a crisis only the variance of grows, to times its former value, while the transmission coefficient and the noise in stay the same, the crisis-period correlation rises mechanically:
Nothing about the link between the markets has changed, yet the measured correlation has risen.
Evidence
Forbes and Rigobon showed that correlation coefficients conditional on a crisis period are biased upwards by heteroskedasticity, and gave the adjustment above (Forbes and Rigobon 2002). They examined the 1997 East Asian crisis, the 1994 Mexican peso collapse and the 1987 US stock market crash: unadjusted correlations found “contagion” in several countries, adjusted correlations found “virtually no contagion”. They concluded that the high co-movement in these periods was a continuation of strong cross-market linkages that existed all along: “no contagion, only interdependence”. Corsetti, Pericoli and Sbracia pointed out that the adjustment depends on particular model assumptions (notably about the variance of each market’s own shocks), and that when these are relaxed the conclusion can become “some contagion, some interdependence” (Corsetti et al. 2005).
The first part of the program below simulates three “worlds”. Each time it compares the correlation over 500 calm days with that over 60 crisis days, repeated 2,000 times, and uses a Fisher z test (one-sided, 5 %) to judge whether the crisis correlation is significantly higher. (1) Interdependence: throughout, and the standard deviation of doubles in the crisis (). (2) Contagion: in the crisis rises from 0.4 to 0.8. (3) Common factor: both and are driven by a common factor whose variance becomes seven times larger in the crisis, while each market’s own shocks are unchanged.
Calculation
# (1) Forbes & Rigobon (2002): conditioning on a high-volatility period raises measured
# correlation even when the link between markets is unchanged. y = beta*x + e; in a
# 60-day "crisis" the variance of x is 4 times higher (delta = 3). Adjusted correlation:
# rho* = rho_c / sqrt(1 + delta (1 - rho_c^2)). Three SIMULATED worlds, 2,000 crises each.
# (2) Longin & Solnik (2001): exceedance correlation corr(x, y | both beyond theta sd) for
# a bivariate normal, a two-regime variance mixture and a Clayton copula (normal margins).
import math, random, statistics
def corr(a, b):
ma, mb = statistics.fmean(a), statistics.fmean(b)
sab = sum((x - ma) * (y - mb) for x, y in zip(a, b))
return sab / math.sqrt(sum((x - ma) ** 2 for x in a) * sum((y - mb) ** 2 for y in b))
def world(kind, n, crisis, rng):
xs, ys = [], []
for _ in range(n):
if kind == "common factor": # x = f + u, y = f + e; crisis: var(f) x 7
f = rng.gauss(0, math.sqrt(7) if crisis else 1)
xs.append(f + rng.gauss(0, math.sqrt(1.5))); ys.append(f + rng.gauss(0, 2))
continue
x = rng.gauss(0, 2 if crisis else 1)
beta = 0.8 if (kind == "contagion" and crisis) else 0.4
xs.append(x); ys.append(beta * x + rng.gauss(0, math.sqrt(0.84)))
return xs, ys
def fisher_z_greater(r1, n1, r0, n0):
z = (math.atanh(r1) - math.atanh(r0)) / math.sqrt(1 / (n1 - 3) + 1 / (n0 - 3))
return z > 1.645
if __name__ == "__main__":
rng = random.Random(2002)
print("world calm rho crisis rho adjusted 'contagion' found by test")
print(" raw rho adjusted rho")
for kind in ("interdependence", "contagion", "common factor"):
rc, ra, raw_hit, adj_hit, r0s = [], [], 0, 0, []
for _ in range(2000):
x0, y0 = world(kind, 500, False, rng); x1, y1 = world(kind, 60, True, rng)
r0, r1 = corr(x0, y0), corr(x1, y1)
delta = statistics.pvariance(x1) / statistics.pvariance(x0) - 1
adj = r1 / math.sqrt(1 + delta * (1 - r1 * r1))
r0s.append(r0); rc.append(r1); ra.append(adj)
raw_hit += fisher_z_greater(r1, 60, r0, 500)
adj_hit += fisher_z_greater(adj, 60, r0, 500)
print(f"{kind:18s} {statistics.mean(r0s):6.3f} {statistics.mean(rc):6.3f}"
f" {statistics.mean(ra):6.3f} {raw_hit / 2000:5.1%}"
f" {adj_hit / 2000:5.1%}")
print("exceedance correlation: corr of (x, y) on days when both are beyond theta sd")
N = 400_000; nd = statistics.NormalDist(); models = {}
pts = []
for _ in range(N):
a = rng.gauss(0, 1); pts.append((a, 0.5 * a + math.sqrt(0.75) * rng.gauss(0, 1)))
models["normal (rho 0.5)"] = pts
pts = []
for _ in range(N):
s = 2.0 if rng.random() < 0.2 else 0.75
a = rng.gauss(0, 1); b = 0.5 * a + math.sqrt(0.75) * rng.gauss(0, 1)
pts.append((s * a, s * b))
models["variance mixture"] = pts
pts = []
for _ in range(N):
v = rng.gammavariate(1 / 1.0, 1) # Clayton theta 1 (same Kendall tau 1/3)
u1, u2 = ((1 + rng.expovariate(1) / v) ** -1.0 for _ in range(2))
pts.append((nd.inv_cdf(u1), nd.inv_cdf(u2)))
models["Clayton, normal margins"] = pts
th = (0.0, 0.5, 1.0, 1.5, 2.0)
print("model side theta = " + "".join(f"{t:6.1f}" for t in th))
for name, pts in models.items():
sd = statistics.pstdev([p[0] for p in pts])
for side, sg in (("down", -1), ("up", 1)):
row = ""
for t in th:
sel = [(x, y) for x, y in pts if sg * x > t * sd and sg * y > t * sd]
row += f"{corr([p[0] for p in sel], [p[1] for p in sel]):6.2f}"
print(f"{name:26s} {side:4s}{'':10s}{row}")
# world calm rho crisis rho adjusted 'contagion' found by test
# raw rho adjusted rho
# interdependence 0.399 0.655 0.407 85.4% 0.5%
# contagion 0.399 0.866 0.664 100.0% 96.1%
# common factor 0.282 0.720 0.502 99.7% 63.3%
# exceedance correlation: corr of (x, y) on days when both are beyond theta sd
# model side theta = 0.0 0.5 1.0 1.5 2.0
# normal (rho 0.5) down 0.27 0.22 0.18 0.12 0.13
# normal (rho 0.5) up 0.27 0.22 0.18 0.15 0.12
# variance mixture down 0.48 0.48 0.41 0.27 0.23
# variance mixture up 0.48 0.47 0.39 0.24 0.18
# Clayton, normal margins down 0.57 0.58 0.58 0.60 0.61
# Clayton, normal margins up 0.06 0.02 -0.00 -0.01 -0.01In the interdependence world the average correlation rises from 0.399 to 0.655 in the crisis, and the unadjusted test declares “contagion” in 85.4 % of simulations; the adjusted correlation is 0.407, almost back to its calm level, with a false-alarm rate of only 0.5 %. In the world with real contagion, the adjusted correlation still rises from 0.399 to 0.664 and the test detects it correctly in 96.1 % of simulations. The common-factor world exposes the problem: the transmission mechanism between the two markets has not changed, only the common shock has grown, yet the adjusted test declares “contagion” in 63.3 % of simulations, because the Forbes–Rigobon adjustment assumes the crisis originates in market alone. Whether a larger common shock counts as “contagion” is itself a matter of definition, and that is exactly why the test’s conclusion depends on the model.
The real phenomenon: stronger dependence in extreme declines
Evidence
Longin and Solnik modelled monthly returns of the five largest stock markets from 1959 to 1996 with extreme value theory and studied “exceedance correlation” (the correlation between two markets when both returns exceed a threshold). They found that the correlation of large negative returns is much greater than multivariate normality would imply, while the correlation of large positive returns is not inconsistent with normality; they also warned that correlation conditional on volatile periods rises even when the underlying correlation is constant, so it must be interpreted with care (Longin and Solnik 2001). Ang and Chen found for US equity portfolios that correlations with the market in falling markets are clearly higher than a normal model implies, while in rising markets they are not (Ang and Chen 2002).
The second part of the program shows why asymmetry is the key evidence. For a bivariate normal distribution with correlation 0.5, exceedance correlation falls as the threshold rises: 0.27 when both exceed 0, only 0.12 to 0.13 when both exceed 2 standard deviations, and symmetric in the two tails. This is a mathematical property of the normal distribution (the asymptotic independence of Chapter 3). A two-regime variance mixture (on 20 % of days volatility is 2.7 times its usual level, with the same correlation in both regimes) raises exceedance correlation across the board (0.48 at threshold 0), but it stays symmetric and still falls with the threshold. Only the Clayton copula (with normal margins) gives an asymmetric pattern: the exceedance correlation in the lower tail is about 0.57 to 0.61 at every threshold, while in the upper tail it is close to zero.
Longin and Solnik’s finding carries weight precisely because the normal model predicts the opposite direction: the more extreme, the less correlated. The two bodies of evidence can be placed side by side as follows. Much of the rise in correlation conditional on volatility is a mechanical statistical effect and cannot be taken directly as evidence of “contagion” (Forbes and Rigobon); but for extreme negative returns, dependence in historical data really is stronger than the normal model allows, and asymmetric (Longin and Solnik). Neither says that correlation “goes to one”, and so the book does not write it as a law.
Vine copulas: building high-dimensional dependence from bivariate blocks
The four copulas of Chapter 3 are not flexible enough in high dimensions. The Gaussian and t copulas describe all dependence with a single correlation matrix (plus, for the t, one degree-of-freedom parameter), so the tail behaviour of every pair of assets is fixed by the same structure; Archimedean copulas (Clayton, Gumbel) usually have a single parameter in high dimensions, so every pair has the same dependence. The idea of a vine copula is to factor a -dimensional density into a product of bivariate copulas (“pair copulas”), each of which can have its own family and parameters.
Evidence
Bedford and Cooke proposed organising this factorisation with a set of nested trees, a “vine” (Bedford and Cooke 2002); Aas, Czado, Frigessi and Bakken gave pair-copula constructions for two widely used structures, C-vines and D-vines, together with maximum-likelihood estimation and simulation algorithms, and applied them to financial returns (Aas et al. 2009).
In three dimensions, for example,
with and conditional distributions given by “h-functions”: . The first tree links adjacent variables in pairs, the second links variables one apart conditional on the variable between them, and so on. In a D-vine every tree is a path (1–2–3–4), suited to variables with a natural order; in a C-vine every tree is a star rooted at a “central” variable (a market index, say), suited to a dominant factor. It is usually assumed in addition that the parameters of the conditional copulas do not vary with the values of the conditioning variables (the “simplifying assumption”), a modelling choice made for estimability.
The program below builds a D-vine for four assets: the three adjacent pairs in the first tree use Clayton copulas (, Kendall’s τ = 0.429, lower-tail dependence ), the second tree uses Gaussian copulas with correlation 0.3, and the third tree is conditional independence. It then builds a Gaussian copula from the six pairwise Kendall’s τ of the simulated vine (), so that the two models have exactly the same pairwise rank correlations, and compares joint tail events and portfolio risk (margins Student t with 4 degrees of freedom and standard deviation 1, equal weights).
Calculation
# A four-asset D-vine (order 1-2-3-4) built from pair copulas (Aas et al. 2009):
# tree 1: Clayton theta 1.5 for (1,2), (2,3), (3,4); tree 2: Gaussian rho 0.3 for
# (1,3|2) and (2,4|3); tree 3: independence for (1,4|2,3). ILLUSTRATIVE parameters.
# Compared with a Gaussian copula whose six pairwise Kendall's taus match the vine's.
# Margins: Student t, 4 d.f., scaled to sd 1. 100,000 simulated days per model.
import math, random, statistics
ND = statistics.NormalDist()
TH, RHO = 1.5, 0.3
def h_clayton(u, v, th=TH): # h(u|v) = dC(u, v)/dv
return v ** (-th - 1) * (u ** -th + v ** -th - 1) ** (-1 - 1 / th)
def hinv_clayton(w, v, th=TH):
return ((w * v ** (th + 1)) ** (-th / (1 + th)) + 1 - v ** -th) ** (-1 / th)
def h_gauss(u, v, r=RHO):
return ND.cdf((ND.inv_cdf(u) - r * ND.inv_cdf(v)) / math.sqrt(1 - r * r))
def hinv_gauss(w, v, r=RHO):
return ND.cdf(ND.inv_cdf(w) * math.sqrt(1 - r * r) + r * ND.inv_cdf(v))
def clip(u):
return min(max(u, 1e-12), 1 - 1e-12)
def dvine(rng):
w = [clip(rng.random()) for _ in range(4)]
x1 = w[0]
x2 = clip(hinv_clayton(w[1], x1))
f1_2 = h_clayton(x1, x2) # F(x1 | x2)
x3 = clip(hinv_clayton(clip(hinv_gauss(w[2], f1_2)), x2))
f2_3, f3_2 = h_clayton(x2, x3), h_clayton(x3, x2)
f1_23 = clip(h_gauss(f1_2, f3_2)) # F(x1 | x2, x3)
a = w[3] # (1,4|2,3) independent
f4_3 = clip(hinv_gauss(a, f2_3)) # F(x4 | x3)
return x1, x2, x3, clip(hinv_clayton(f4_3, x3))
def kendall(a, b):
n, s = len(a), 0
for i in range(n):
ai, bi = a[i], b[i]
for j in range(i + 1, n):
s += 1 if (a[j] - ai) * (b[j] - bi) > 0 else -1
return 2 * s / (n * (n - 1))
def gauss_copula(rng, R):
L = [[0.0] * 4 for _ in range(4)] # Cholesky factor
for i in range(4):
for j in range(i + 1):
s = R[i][j] - sum(L[i][k] * L[j][k] for k in range(j))
L[i][j] = math.sqrt(s) if i == j else s / L[j][j]
z = [rng.gauss(0, 1) for _ in range(4)]
return tuple(ND.cdf(sum(L[i][k] * z[k] for k in range(i + 1))) for i in range(4))
def t4(u): # t, 4 d.f., sd 1
lo, hi = -60.0, 60.0
for _ in range(60):
m = (lo + hi) / 2; q = m / math.sqrt(m * m + 4)
lo, hi = (m, hi) if 0.5 + 0.75 * (q - q ** 3 / 3) < u else (lo, m)
return lo / math.sqrt(2)
def report(name, us):
n = len(us)
for q in (0.05, 0.01):
both = sum(1 for u in us if u[0] < q and u[3] < q) / n
allf = sum(1 for u in us if max(u) < q) / n
print(f" {name:9s} all four below their {q:.0%} quantile: {allf:.5f}"
f" assets 1 and 4 both: {both:.5f}")
loss = sorted(-sum(t4(v) for v in u) / 4 for u in us)
k = int(0.99 * n); var = loss[k]; es = statistics.fmean(loss[k:])
print(f" {name:9s} equal-weight portfolio (t4 margins): 99 % VaR {var:.2f},"
f" ES {es:.2f}")
if __name__ == "__main__":
rng = random.Random(2009)
V = [dvine(rng) for _ in range(100_000)]
sub = V[:1500]; tau = [[1.0] * 4 for _ in range(4)]
for i in range(4):
for j in range(i + 1, 4):
tau[i][j] = tau[j][i] = kendall([u[i] for u in sub], [u[j] for u in sub])
print("Kendall's tau of the vine (1,500 days); Clayton theta 1.5 implies tau"
f" {TH / (TH + 2):.3f}")
for i in range(4):
print(" " + "".join(f"{tau[i][j]:7.3f}" for j in range(4)))
R = [[math.sin(math.pi * tau[i][j] / 2) for j in range(4)] for i in range(4)]
G = [gauss_copula(rng, R) for _ in range(100_000)]
print("joint tail events, 100,000 days each (independence: 0.05^4 = 0.0000063)")
report("D-vine", V); report("Gaussian", G)
# Kendall's tau of the vine (1,500 days); Clayton theta 1.5 implies tau 0.429
# 1.000 0.446 0.413 0.327
# 0.446 1.000 0.441 0.412
# 0.413 0.441 1.000 0.414
# 0.327 0.412 0.414 1.000
# joint tail events, 100,000 days each (independence: 0.05^4 = 0.0000063)
# D-vine all four below their 5% quantile: 0.01906 assets 1 and 4 both: 0.02407
# D-vine all four below their 1% quantile: 0.00387 assets 1 and 4 both: 0.00485
# D-vine equal-weight portfolio (t4 margins): 99 % VaR 2.48, ES 3.49
# Gaussian all four below their 5% quantile: 0.00455 assets 1 and 4 both: 0.01163
# Gaussian all four below their 1% quantile: 0.00039 assets 1 and 4 both: 0.00129
# Gaussian equal-weight portfolio (t4 margins): 99 % VaR 2.13, ES 2.80The sample τ of the three first-tree pairs are 0.446, 0.441 and 0.414, in line with the theoretical 0.429; pairs one and two apart have τ of 0.41 and 0.33. The two models share their pairwise rank correlations, but the probability that all four assets fall below their own 5 % quantiles is 0.019 for the vine and 0.0046 for the Gaussian copula, about four times as high; below their 1 % quantiles, 0.0039 against 0.00039, about ten times. Even for the most distant pair, assets 1 and 4, the probability of both falling below their 1 % quantiles differs by a factor of nearly four (0.0049 against 0.0013). The equal-weight portfolio’s 99 % VaR is 2.48 against 2.13, 16 % higher, and its 99 % ES 3.49 against 2.80, 25 % higher. Equal pairwise correlations, in other words, do not mean equal risk; with these parameters the Gaussian copula systematically understates the risk of “falling together”.
Note
Vine copulas are not a cure-all. Four assets need 6 pair copulas and twenty need 190, each with a family and parameters to choose, and the choices themselves bring a risk of overfitting; the tree structure (which variables are adjacent) must be chosen too. The simplifying assumption may fail. Most important, tail parameters estimated on calm-period data may not apply in a crisis, which is the regime problem of Chapter 1. Vine copulas make it possible for a model to express asymmetric tail dependence; they cannot guarantee that the estimated dependence is the right one.
Hypothesis · not established
The author’s hypothesis (not established): lower-tail dependence between assets (for example, the Clayton parameters in the first tree of a vine) is significantly higher in tight global-liquidity regimes than in ample ones, and the difference persists at equal levels of volatility (that is, it is not merely the mechanical effect described by Forbes and Rigobon). What would show it wrong: with assets, window and regime-classification rule fixed in advance (using only information available in real time), a confidence interval for the difference in lower-tail dependence between regimes, controlling for volatility, that includes zero.
Chapter summary
“Correlations go to one in a crisis” is not a law. Correlation conditional on high volatility rises mechanically, and after their adjustment Forbes and Rigobon found almost no contagion in three crises; but the adjustment itself depends on the model and gives false alarms in a common-factor world. Longin and Solnik found that extreme negative returns are far more correlated than a normal model allows, and since the normal model predicts that more extreme means less correlated, this asymmetry is weighty evidence. Vine copulas build high-dimensional dependence out of bivariate copulas and allow different tail behaviour for different pairs; with exactly the same pairwise rank correlations, a D-vine with Clayton copulas in its first tree makes the probability of all four assets falling below their 1 % quantiles about ten times higher than a Gaussian copula does.
References
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).