Extreme Value Inference for CoVaR and

Systemic Risk

Xiaoting Li∗
Department of Statistics, University of Manitoba
and
arXiv:2603.27458v1 [[Link]] 29 Mar 2026

Harry Joe
Department of Statistics, University of British Columbia

March 31, 2026

Abstract

We develop an extreme value framework for CoVaR centered on v(q | p; C), the
copula-adjusted probability level, or equivalently, the CoVaR on the uniform (0,1)
scale. We characterize the possible tail regimes of v(q | p; C) through the limit behavior
of the copula conditional distribution and show that these regimes are determined
by the joint tail expansions of the copula. This leads to tractable conditions for
identifying the tail regime and deriving the asymptotic behavior of v(q|p; C). Building
on this characterization, we propose a minimum-distance estimation approach for
CoVaR that accommodates multiple tail regimes. The methodology links CoVaR
and ∆ CoVaR to the underlying joint tail behavior, thereby providing a clear inter-
pretation of these measures in systemic risk analysis. An empirical analysis across
U.S. sectors demonstrates the practical value of the approach for assessing systemic
risk contributions and exposures with important implications for macroprudential
surveillance and risk management.

Keywords: conditional quantiles, copulas, tail order, tail dependence, extreme value theory

∗
The authors gratefully acknowledge support from the Natural Sciences and Engineering Research
Council of Canada.

1
1 Introduction

From a macroprudential view (International Monetary Fund 2011, Borio & Zhu 2012),

systemic risk arises from the gradual buildup of financial imbalances, such as credit expansion,

rising leverage, and concentrated asset exposures. While these imbalances do not immediately

generate instability, they fundamentally alter how the system responds to severe shocks.

In probabilistic terms, the buildup of systemic vulnerability manifests as changes in the

extremal dependence structure of the system.

The literature has developed a range of measures to quantify systemic risk by examining

the system’s response under stress (Nolde & Zhou 2021). Among them, conditional Value-

at-Risk (CoVaR) (Adrian & Brunnermeier 2016) is defined as the system’s Value-at-Risk

conditional on the distress of a given institution. It has been extensively used in the

literature to study systemic risk; see, e.g., Fan et al. (2018), Ascorbebeitia et al. (2022).

By its probabilistic nature, CoVaR is an extreme conditional quantile. The copula repre-

sentation of CoVaR has been proven useful to study its theoretical properties; see, e.g.,

Bernard & Czado (2015), Li & Joe (2026). It makes explicit how the extremal dependence

structure, captured by the tail behavior of the joint copula, determines CoVaR through an

adjusted probability level. This copula-adjusted level, denoted by v(q | p; C), represents

CoVaR on the uniform scale and lies at the core of our methodology. Earlier works along

this direction include Jaworski (2017) and Nolde et al. (2025), which derived the limit of

v(q | p; C) as p ↓ 0 under asymptotic dependence.

Previous work in the copula literature, e.g., Hua & Joe (2011), Li & Joe (2023), has

developed a tail expansion framework to describe the extremal dependence structure

of multivariate random vectors. Specifically, a bivariate copula C with a well behaved

joint lower tail is said to admit a lower tail expansion if there exist a tail order function

2
a(w1 , w2 ) : (0, ∞)2 → (0, ∞), a tail order κ ≥ 1, and a slowly varying function ℓ(u) such

that, as u → 0+ ,
C(uw1 , uw2 )
a(w1 , w2 ) = lim .
u↓0 uκ ℓ(u)

The tail behavior is further refined based on the tail order: C has strong tail dependence if

κ = 1, intermediate tail dependence if κ ∈ (1, 2], tail orthant independence if κ = 2, and

super tail orthant independence if κ > 2. The recent work Li & Joe (2026) has established

the value of tail expansions in analyzing the limiting behavior of v(q | p; C) at the extreme

quantile level q = p for κ ∈ [1, 2].

In this work, we propose an extreme value inference framework for CoVaR centered on this

copula-adjusted probability level. One main contribution is to characterize the possible

tail regimes, corresponding to tail attraction, tail repulsion, and tail balance between firm-

level distress and system-wide loss, through the limiting behavior of the copula conditional

distribution. We then formally establish how these regimes are determined by tail expansions

of the copula, thereby providing tractable conditions for regime identification and for deriving

the behavior of v(q | p; C).

Building on this characterization, we propose a minimum-distance estimation approach

for CoVaR(q | p) that accommodates multiple tail regimes and establishes its asymptotic

consistency. A similar approach was used in Nolde et al. (2025) to estimate CoVaR(q | p)

for the case q = p under the regime that we refer to as strong tail attraction.

We also introduce a ∆ CoVaR and derive its limiting behavior as p ↓ 0. Unlike the ∆ CoVaR

in earlier works such as Adrian & Brunnermeier (2016), Girardi & Ergün (2013), our

copula-based formulation isolates the effect of extremal dependence from that of marginal

volatility. It aligns with the IMF’s macroprudential view (International Monetary Fund

2011) of systemic risk and supports monitoring tools that distinguish structural fragility

from high-frequency market turbulence.

3
The rest of the paper is organized as follows. Section 2 introduces the basic definitions and

some properties of CoVaR and ∆ CoVaR. Section 3 and Section 4 present our main results

on the behavior of CoVaR and ∆ CoVaR, respectively. Section 5 develops the estimation

procedure and establishes its asymptotic properties. Section 6 demonstrates the empirical

value of the proposed methodology through a comprehensive analysis of systemic risk in the

U.S. market, highlighting how extremal dependence within the system shapes CoVaR and

∆ CoVaR and reveals differences in the systemic roles of assets and institutions. Section 7

concludes with a brief discussion of the implications for macroprudential monitoring and

risk management. Proofs and technical details are provided in the supplementary material.

2 Basic definitions and properties of CoVaR

Let (Xi , XS ) denote, respectively, the return of institution i and the system-wide return.

Their bivariate distribution is FXi ,XS (x, s) = Pr(Xi ≤ x, XS ≤ s), for (x, s) ∈ R2 , with

continuous marginals FXi and FS . Define the generalized inverse g ← (u) := inf{x ∈ R :

g(x) ≥ u}, where g ← = g −1 = {x : g(x) = u} if g is strictly monotone.

Definition 2.1. Fix a small p ∈ (0, 1) and VaRi (p) = FX←i (p) be the individual p-level VaR.

Define the conditional distribution of XS given the stress event {Xi ≤ VaRi (p)} by

 
FXS | Xi ≤VaRi (p) (s) = Pr XS ≤ s | Xi ≤ VaRi (p) .

The conditional Value-at-Risk of the system at level q ∈ (0, 1) is

CoVaRS|i (q | p) := FS←| Xi ≤VaRi (p) (q).

The CoVaRS|i (q | p) admits an equivalent copula representation. Let Ui = Fi (Xi ) and

US = FS (XS ). The associated copula CiS , such that (Ui , US ) ∼ CiS , is uniquely specified,

4
  
for (u, v) ∈ (0, 1)2 , CiS (u, v) = FXi ,XS FX←i (u), FS← (v) . For each p ∈ (0, 1), define

CiS (p, v)
CS|i≤ (v | p) := Pr(US ≤ v | Ui ≤ p) = , v ∈ [0, 1].
p

Then

 
CoVaRS|i (q | p) = (FS ◦ CS|i≤ (· | p))← (q) = VaRS vS|i (q | p; CiS ) , (1)

and the copula-adjusted probability level is

n o
←
vS|i (q | p; CiS ) := CS|i≤ (q | p) = inf v ∈ [0, 1] : CS|i≤ (v | p) ≥ q . (2)

This representation makes explicit how the dependence structure of (Xi , XS ), encoded

by their joint copula CiS , will affect CoVaR via the adjusted probability level. In other

words, the copula reflects how the system’s return quantiles shift under the stress event. As

emphasized in Adrian & Brunnermeier (2016), Girardi & Ergün (2013), systemic impact

is better reflected by the shift in CoVaR than by the CoVaR level itself. For this end, we

introduce the risk measure ∆ CoVaR. Unlike the existing literature, we use the unconditional

VaR as the benchmark for normal market conditions. As shown in Section 4, this choice

gives ∆ CoVaR a clear interpretation in terms of the joint tail behavior of the risk variables

and its effect on systemic risk.

Definition 2.2 (∆CoVaR).

 
CoVaRS|i (q | p) − VaRS (q) VaRS vS|i (q | p; CiS )
∆ CoVaRS|i (q | p) := = − 1.
| VaRS (q)| | VaRS (q)|

Below are some basic properties of the copula-adjusted probability level, or CoVaR in the

uniform (0,1) scale. For the remainder, we omit the subscript when there is no ambiguity.

If (Xi , XS ) has copula C, then (−Xi , −XS ) has its reflected or survival copula C ∗ (u, v) =

u + v − 1 + C(1 − u, 1 − v); (−Xi , XS ) has 1-reflected copula, C 1∗ (u, v) = v − C (1 − u, v) =

u − C ∗ (u, 1 − v); (Xi , −XS ) has its 2-reflected copula, C 2∗ (u, v) = u − C (u, 1 − v).

5
Proposition 2.3. (a) Comonotonicity: v(q | p; C + ) = pq.

(b) Independence: v(q | p; C ⊥ ) = q.

(c) Countermonotonicity: v(q | p; C − ) = 1 − (1 − q)p.

(d) Coherence: if C1 ⪯c C2 (i.e. C1 (u, v) ≤ C2 (u, v) for all (u, v)) then for all p, q,

v(q | p; C2 ) ≤ v(q | p; C1 ).

(e) Reflection: v(q | p; C) = 1 − v(1 − q | p; C 2∗ ), v(q | p; C ∗ ) = 1 − v(1 − q | p; C 1∗ ).

(f) Range: pq ≤ v(q | p; C) ≤ q if C ⊥ ≺c C ≺c C + and q ≤ v(q | p; C) ≤ 1 − (1 − q)p if

C − ≺c C ≺c C ⊥ .

3 Tail regimes of the copula-adjusted level in CoVaR

This section studies the limiting behavior of the copula-adjusted probability level v(q | p; C),

which is central to the copula representation of CoVaR in Eq. (1). We show that, as

p ↓ 0, the limit of v(q | p; C) is determined by the limit laws of the conditional cumulative

distribution function (cdf) CS|i≤ (· | p) which in turn induces a classification of tail regimes.

We then link these regimes to the joint tail behavior of the variables through the tail

expansion of the copula, which yields a criterion to identify the tail regime and refined

characterization of v(q | p; C). Section 3.1 and Section 3.2 treat, respectively, the fixed q

case and the extreme quantile case q = p.

3.1 Tail regimes of v(q | p; C) for fixed q

Theorem 3.1. Suppose there exists a function A : (0, 1) → [0, 1] such that, for v ∈ (0, 1),

lim CS|i≤ (v | p) = A(v). (3)

p↓0

6
Assume that the convergence is locally uniform on (0, 1), i.e., for every 0 < a < b < 1,

sup CS|i≤ (v | p) − A(v) −−→ 0.

v∈[a,b] p↓0

Then the following hold:

(i) A is non-decreasing and continuous on (0, 1).

(ii) A(v) extends to a proper distribution function supported on [0, 1] with possible jumps

at 0 and 1 of sizes p0 := A (0+ ), p1 := 1 − A (1− ), 0 ≤ p0 , p1 ≤ 1, p0 + p1 ≤ 1.

With v(q | p; C) defined in Eq. (2), for every q ∈ (0, 1),



0, 0 < q < p0









lim v(q | p; C) = A ←
(q) ∈ (0, 1), p0 < q < 1 − p1 , provided A← is continuous at q .
p↓0 






1, 1 − p1 < q < 1



Remark 3.2. The boundary cdf limu→0 Pr (US ≤ v | Ui = u) has been studied in Hua & Joe

(2014); it assumes that the copula has a continuous first-order derivative with respect to

u. The boundary of CS|i≤ (v | u) has not, to our knowledge, been studied before. It does

not require that the copula have continuous derivatives. When the copula is continuously

differentiable in u, the two boundary cdfs coincide in the limit as u → 0, following from

l’Hopital’s rule. ◁

Theorem 3.1 characterizes the possible tail regimes of v(q | p; C) based on the structure of

the limiting conditional distribution A.

(i) Tail attraction

If p0 = 1 and p1 = 0, then the extended distribution of A(v) is a pure atom at 0.

Consequently,

v(q | p; C) → 0, q ∈ (0, 1).

7
 We refer to this regime as tail attraction, as the extremely small values of Ui pull US

towards the lower endpoint 0.

(ii) Tail repulsion

If p0 = 0 and p1 = 1, then the extended distribution of A(v) is a pure atom at 1.

Consequently,

v(q | p; C) → 1, q ∈ (0, 1).

We refer to this regime as tail repulsion, as conditioning on extremely small values of

Ui pushes US away from 0 and toward the opposite endpoint 1.

(iii) Tail balance

If p0 = p1 = 0, then the extended distribution of A(v) has no boundary atoms, so all

mass lies entirely in the interior. In this case,

v(q | p; C) → A← (q) ∈ (0, 1).

We call this regime tail balance, as extreme values of Ui do not force US to the

boundary but balance it in the interior.

(iv) Mixed tail regimes

We refer to the case as mixed tail regimes when the extended limiting distribution

has at least one boundary atom and is not a single point mass. It places mass p0 at 0,

mass p1 at 1 and the remaining 1 − p0 − p1 on the interior. The behavior of v(q | p; C)

exhibits mixed regimes depending on how the total mass is distributed between the

two boundary atoms and the interior component.

A central contribution of this paper is to link the behavior of the limiting conditional

distribution A(v) with the joint tail expansion of the copula so that we can characterize

the asymptotic regimes of the quantiles based on tail properties of the copula, i.e., the tail

order and the tail dependence function. In Proposition 3.3, we showed that when the copula

8
has strong lower tail dependence with κ = 1, the conditional limit A(v) will have a strictly

positive mass at 0 with size equal to the limit of its tail dependence function.

Proposition 3.3. Assume the copula C admits a lower tail expansion with a tail order

κ = 1, as u ↓ 0,

C(uw1 , uw2 )
= b(w1 , w2 ) + R(u; w1 , w2 ), (4)
u

in which R (u; w1 , w2 ) → 0 uniformly whenever uw1 → 0 and uw2 → 0. The tail dependence

function b(w1 , w2 ) is monotone and bounded with b∞ := limr→∞ b(1, r) = supr>0 b(1, r) ∈

(0, 1]. Then the function A defined in Eq. (3) satisfies

A(0+ ) = b∞ ∈ (0, 1].

The results in the following lemma aim to relate the behavior of v(q | p; C) to the tail

expansion of the copula and its 2-reflected copula. The proof follows from the definition of

C 2∗ .

Lemma 3.4. Let C be a bivariate copula. Define its 2-reflection

C 2∗ (u, v) := u − C(u, 1 − v), (u, v) ∈ [0, 1]2 ,

which is the copula associated with the random vector (Xi , −XS ). The conditional distribution

function

2∗
CS|i≤ (v | p) := Pr(1 − US ≤ v | Ui ≤ p) = 1 − CS|i≤ (1 − v | p).

Assume that for every v ∈ (0, 1), A(v) = limp↓0 CS|i≤ (v | p) exists. Then A2∗ (v) :=
2∗
limp↓0 CS|i≤ (v | p) also exist and satisfy

A2∗ (v) = 1 − A(1 − v), v ∈ (0, 1).

Theorem 3.5. Assume the copula C admits a lower tail expansion with tail order κ ≥ 1,

C(uw1 , uw2 )
= a(w1 , w2 ) ℓ(u) + R(u; w1 , w2 ), u ↓ 0, (5)
uκ

9
where ℓ is slowly varying at 0, R (u; w1 , w2 ) → 0 uniformly whenever uw1 → 0 and uw2 → 0.

When κ = 1, the expansion reduces to Eq. (4) and the tail order function a coincides with

the tail dependence function b.

Assume in addition, the following boundary conditions on the tail order function,

(B1) As r ↓ 0, a(1, r) = a1 r + o(r) for some constant a1 > 0.

(B2) If κ = 1, let H(r) = a(1, r), which is continuous and strictly increasing on (0, ∞),

with limr↓0 H(r) = 0, limr→∞ H(r) = b∞ ∈ (0, 1]. If κ > 1, then a(1, r) = O(rρ ) for

some 0 < ρ ≤ κ − 1 as r → ∞.

Then the following hold.

(i) The limiting conditional cdf A has a strictly positive mass at 0, i.e. A(0+ ) > 0, if and

only if C has strong lower tail dependence, i.e., κ = κ(C) = 1. Moreover, when κ = 1,

A(0+ ) = b∞ and for every q ∈ (0, b∞ ),

v(q | p; C) ∼ H −1 (q) p ↓ 0.

(ii) The limiting conditional cdf A has a strictly positive mass at 1, i.e. 1 − A(1− ) > 0,

if and only if the 2-reflected copula C 2∗ has strong lower tail dependence, i.e., κ2∗ =

κ(C 2∗ ) = 1. In that case, 1 − A(1− ) = b2∗ 2∗

∞ , and for every q ∈ (1 − b∞ , 1),

v(q | p; C) ∼ 1 − (H 2∗ )−1 (1 − q) p ↑ 1.

Remark 3.6. The conclusion of the theorem relies on two technical conditions, the uniform

expansion of the copula and growth rate of the tail order function. If the conditions fail, the

tail order κ alone no longer determines the boundary behavior of the limiting conditional

cdf. It is possible to have κ > 1 while A(0+ ) > 0 and the conditional quantiles converge

at non-polynomial rates that are not captured by the expansion used in the theorem. The

Gumbel copula and the Gaussian copula are two examples with tail order κ > 1 and A in

10
Eq. (3) degenerate at 0. For these two families, the tail expansion in Eq. (5) is not uniform;

the remainder in the expansion inflates away from the diagonal. ◁

Copula κ A(v) v(q | p) v(p)

Clayton(θ > 0) 1 1 p(q −θ − 1)−1/θ p2

Gumbel∗ (δ > 1) 1 1 H −1 (q; δ)p p2

1 −1
IPS∗ (θ > 0) 1+ 1 − e−FΓ (v;θ)
FΓ (− ln(1 − q); θ) (Γ(θ + 1))−1/θ p 2−1/θ (θ > 1)
θ

{Γ(θ + 1)}−1 pθ (0 < θ < 1)

Frank (θ > 0) 2 (1 − e−θv )(1 − e−θ )−1 − 1θ log(1 − q(1 − e−θ )) θ−1 (1 − e−θ )p

n 1
o 1
Gumbel2∗ (δ > 1) 1+δ 0 1 − exp −(δq)1/δ (− log(p))1− δ (δp)1/δ (− log(p))1− δ

Clayton2∗ (θ > 0) θ+2 0 1 − ((1 − q)−θ − 1)−1/θ p 1 − θ−1/θ p1−1/θ (θ > 1)

p1−θ (0 < θ < 1)

Table 1: Tail order κ, limiting conditional cdf A(v), copula-adjusted level v(q | p; C) and

v(p; C) for common parametric copula families. A superscript ∗ denotes the reflected copula,

1∗ the 1-reflected, and 2∗ the 2-reflected copula. The IPS copula is an Archimedean copula

constructed from the integrated positive stable Laplace transform; see Section 4.11 of Joe

(2014) for details.

Theorem 3.5 clarifies the link between the behavior of A, v(q | p; C) and the tail expansion of

the copula. When the copula C has tail order 1, the tail mass is asymptotically concentrated

in the lower left corner. The limiting conditional cdf A has positive mass at 0, and the

conditional quantiles converge toward 0 at a linear rate in p. Conversely, if the 2-reflection

of C has tail order 1, then the tail mass is asymptotically concentrated in the upper left

corner. In that case, A assigns positive mass to 1, and the conditional quantiles converge

toward 1 at a linear rate in p.

11
Most common copula families fall into a single regime for v(q | p; C). Table 1 summarizes the

tail order κ, the limiting conditional cdf A(v), and the behavior of v(q | p; C) for common

copula families. The first two rows, corresponding to the Clayton and reflected Gumbel

copula families, both exhibit strong lower tail dependence: A is degenerate at 0, there is

tail attraction for all q ∈ (0, 1), with v(q | p; C) → 0 at a linear rate. The middle two rows,

corresponding to the reflected IPS and Frank copula families. In both cases κ > 1, A(v)

is strictly increasing, and v(q | p; C) → A−1 (q) for all q ∈ (0, 1), reaching a tail balance.

The last two rows correspond to 2-reflected copula families, where the reflection operation

converts families with only positive dependence to families with only negative dependence.

In these cases, A(v) is degenerate at 1, and there is tail repulsion: v(q | p; C) → 1 for all

q ∈ (0, 1).

The reflected IPS copula family has cdf

 
C 2∗ (u, v; θ) = u + v − FΓ (xθ + y θ )1/θ ; θ , x = FΓ−1 (u; θ), y = FΓ−1 (v; θ), θ > 0,

where FΓ (·; θ) and FΓ−1 (·; θ) are the cdf and quantile functions for the Gamma(θ, 1) distri-

bution. It has negative dependence for 0 < θ < 1, independence for θ = 1 and positive

dependence for θ > 1. The tail order is κ(θ) = 1 + θ−1 > 1. This family is useful for κ > 1

in (B2) of Theorem 3.5. It can be shown directly that its tail order function in Eq. (5) is

a(w1 , w2 ) = (w1 + w2 )κ − w1κ − w2κ with ℓ(u) = [Γ(θ + 1)]1/θ /κ. Then a(1, r) ∼ κrκ−1 as

r → ∞, so that the condition in (B2) is satisfied with ρ = κ − 1.

Mixed tail regimes arise when the tail mass is split between the lower and upper corners.

This is less common among standard parametric copula families. The Student-t copula has

tail order 1 in all four corners, which results in concentration of tail mass in both the lower

and upper corners, and thus exhibits a mixed tail regime.

Example 3.7. Let Cρ,ν be the bivariate Student tν copula with correlation ρ ∈ (−1, 1) and

12
degrees of freedom ν > 0. Its lower tail dependence function is

n h w2 −1/ν io n h w1 io
b (w1 , w2 ; ρ, ν) = w1 Tν+1 K ρ − ( ) + w2 Tν+1 K ρ − ( )−1/ν ,
w1 w2
q
where Tν+1 is the univariate t CDF with ν+1 degrees of freedom and K = (ν + 1)/ (1 − ρ2 ) .

Let H(r) = b(1, r; ρ, ν) and H 2∗ (r) = b(1, r; −ρ, ν).

For v ∈ (0, 1), A(v) = supr>0 H(r) = Tν+1 (ρK) = q ∗ , and has a jump of size p1 = 1 − q ∗ at

1. It thus exhibits a mixed tail regime depending on q. For q < q ∗ , it has tail attraction v(q |

p; C) ∼ H −1 (q)p → 0. For q > q ∗ , it has tail repulsion, v(q | p; C) ∼ 1−(H 2∗ )−1 (1−q)p → 1.

For q = q ∗ , v(q ∗ | p; C) → 21 . ◁

3.2 Tail regimes of v(q | p; C) for extreme-level q

In systemic risk measurement, we are particularly interested in extreme quantile levels,

typically with the same tail probability used for both the institution and the system, i.e.,

q = p. Under this setting, Definition 2.1 simplifies to:

CoVaRS|i (p) = VaRS (v(p; CiS )),

←
and the adjusted level v(p; CiS ) = CS|i≤ (p | p). A key difference is that, under q = p,

the right-hand-side target probability becomes smaller order than p. Driving v(p) to the

lower endpoint, as a result, does not require the same strength of tail concentration as in

Section 3.1. Theorem 3.8 derives the limit of v(p) based on the limiting conditional cdf,

showing that v(p) → 0 holds broadly except in the degenerate case where mass concentrates

in the upper left corner, in which case higher-order terms determine the limit.

Theorem 3.8. With A(v) defined in Eq. (3),

(i) If A(0+ ) = p0 > 0, as p ↓ 0,

v(p; C) → 0.

13
 (ii) If A(0+ ) = p0 = 0 and A(1− ) = 1−p1 ∈ (0, 1], and v− = inf{v ∈ [0, 1] : A(v) > 0} = 0,

then as p ↓ 0,

v(p; C) → 0.

(iii) If p0 = 0, p1 = 1, then A(v) = 0 for all v ∈ (0, 1), it would depend on the higher-order

terms. Assume there exists δ > 0, a non-decreasing function B : (0, 1) → (0, ∞), and

ℓ is slowly varying with ℓ0 := limp↓0 ℓ(p) ∈ (0, ∞) such that uniformly for v ∈ (0, 1)

C(p, v) = B(v)p1+δ ℓ(p) + o(p1+δ ).

(a) 0 < δ < 1: v(p; C) → 0.

(b) δ = 1: v(p; C) → B ← (ℓ−1

0 ) if B
←
is continuous at ℓ−1
0 .

(c) δ > 1: v(p; C) → 1.

The next theorem characterizes the behavior of v(p; C) using the lower tail expansion of

the copula. It provides a criterion in terms of the tail order and yields explicit convergence

rates for v(p; C). The case of κ ∈ [1, 2] has been studied in Li & Joe (2026).

Theorem 3.9. Assume the copula C admits the lower-tail expansion as described in

Theorem 3.5. Then

v(p; C) → 0 ⇐⇒ 1 ≤ κ < 2 + ρ.

Moreover as p ↓ 0,


O(p3−κ ),


 1 ≤ κ ≤ 2,
v(p; C) =
 κ−2
O(p1− ρ ),

2 < κ < 2 + ρ.



The last column of Table 1 shows v(p; C) for common copula families. Clayton and reflected

Gumbel copulas have κ = 1, so v(p) = O(p2 ). The reflected IPS copula has tail order

κ = 1 + 1/θ and ρ = κ − 1 for (B2) of Theorem 3.5; then v(p) = O(p2−1/θ ) when θ > 1

(1 < κ ≤ 2) and v(p) = O(p1/(κ−1) ) = O(pθ ) when 0 < θ < 1 (2 < κ < 2+(κ−1)). The Frank

14
copula has tail orthant independence with κ = 2 for all −∞ < θ < ∞, and hence v(p) = O(p).

The 2-reflected Gumbel and Clayton copulas exhibit negative quadrant dependence with

κ > 2. For the former, it has 2 < κ < 2+ρ for all δ > 1, and v(p) = O(p1/δ (− log p)1−1/δ ) ↓ 0.

For the latter, 2 < κ < 2 + ρ when 0 < θ < 1, so v(p) = O(p1−θ ) when 0 < θ < 1. However,

when θ > 1, we instead have v(p) ∼ 1 − θ−1/θ p(θ−1)/θ ↑ 1.

4 Tail limits of ∆CoVaR and systemic risk implications

In quantifying systemic risk impact, the primary concern is its effect on extreme system

losses. Therefore, ∆ CoVaRS|i (q | p) is most informative when evaluated at small q. When

q = p, Definition 2.2 reduces to

VaRS (v(p; C)) − VaRS (p)

∆CoVaRS|i (p) = ,
| VaRS (p)|

which measures the percentage change in the system’s VaR under distress conditions

relative to normal conditions. Its behavior will depend on the the limiting behavior of v(p),

determined by the tail dependence structure of the copula, and VaR, determined by the

marginal distribution of XS , as detailed in the following theorem.

Theorem 4.1. Let (Xi , XS ) be the returns of institution i and the system, with joint cdf

FXi ,S (x, s) = Pr(Xi ≤ x, XS ≤ s) = CiS (FXi (x), FXS (s)). Assume that

(A1) FS lies in the minimum-domain of attraction of a univariate generalized extreme value

(GEV) distribution with lower–tail index ξ = ξS ≥ 0. Hence, as p ↓ 0



−p−ξ L∗ (1/p),


 ξ > 0,
VaRS (p) ∼

−H −1 (− log p),

ξ = 0.



where L∗ is slowly varying at ∞ and H −1 has extended regular variation at ∞ with

index γ ∈ (0, ∞].

15
(A2) The copula CiS admits a lower expansion of Eq. (5), with tail order κ = κiS ∈ [1, 2 + ρ)

so that v(p; C) behaves as described in Theorem 3.9.

Then, as p ↓ 0, the following hold:

(i) Heavy marginal tail (ξ > 0)



−(2−κ)ξ
−p → −∞, 1 ≤ κ < 2,








∆ CoVaRS|i (p) ∼ −ξ
1 − a0 ∈ (−∞, 1), κ = 2, with a(1, a0 ) = 1.






 (κ−2)ξ
1 − p ρ → 1, 2 < κ < 2 + ρ.



(ii) Light marginal tail (ξ = 0)



1 − (3 − κ)γ < 0, 1 ≤ κ < 2,









∆ CoVaRS|i (p) → 0, κ = 2,







1 − (1 − κ−2 )γ ∈ (0, 1], 2 < κ < 2 + ρ.


ρ

This result provides a clear interpretation of ∆ CoVaR based on the joint tail behavior of

the distribution, through the copula tail and the marginal tail.

When the copula exhibits strong to intermediate tail dependence (i.e., κ ∈ [1, 2)),

∆ CoVaRS|i (p) is eventually negative. This reflects a risk amplification effect: under the

distress of institution i, the system is pushed deeper into the tail. The amplification factor

is determined by the marginal tail distribution. If FS lies in the Fréchet domain (ξ > 0),

the amplification scales as −p−(2−κ)ξ and diverges as p ↓ 0. If FS lies in the Gumbel domain

(ξ = 0), the amplification converges to the finite limit 1 − (3 − κ)γ , with larger γ (faster

growth of H −1 ) producing a stronger effect.

When 2 < κ < 2 + ρ, which can occur with negative dependence, ∆ CoVaRS|i (p) is

asymptotically positive. This reflects a risk attenuation effect: under the distress of

institution i, the system’s tail losses become less extreme. The attenuation factor is between

16
0 and 1, and tends to 1 when ξ > 0. Note, when v(p) → 1, the factor can exceed 1 or even

diverges to ∞.

In the middle case with tail-orthant independence (κ = 2), if ξ = 0 then ∆ CoVaRS|i (p) → 0,

indicating that the system’s tail risk is asymptotically unchanged. If ξ > 0, the sign depends

on the value of a0 = limp→0 r(p; C). ∆ CoVaRS|i (p) is asymptotically negative when a0 < 1

and positive when a0 > 1.

5 Estimation of CoVaR and its asymptotic properties

Estimation of CoVaR, based on the copula representation in Eq. (1), involves estimating the

marginal Value-at-Risk (VaR) and the copula-adjusted level v(q | p; C). In this work, we

propose a minimum-distance estimation approach for v(q | p; C), which represents CoVaR

on the uniform scale, and can then be combined with the marginal distribution to estimate

CoVaR on the original scale.

5.1 Empirical estimators of tail dependence

Let {(Xi , Yi ), i ∈ {1, 2, . . . , }} be an infinite sequence of independent random vectors with

common bivariate distribution function F and univariate continuous margins FX , FY . Denote

by Xi:n be the i-th smallest among the first n random variables {X1 , . . . , Xn }. Define the

transformed variables Ui = FX (Xi ), and Vi = FY (Yi ) and their empirical distribution

function is
n
1X
Cn (u, v) = 1 {Ui ≤ u, Vi ≤ v} , u, v ∈ (0, 1).
n i=1

In practice, Ui and Vi are unobserved, we replace them with pseudo-observations from ranks,

n n
Ui,n = n−1 Vi,n = n−1
X X
1 (Xj ≤ Xi ) , 1 (Yj ≤ Yi ) .
j=1 j=1

17
Using these pseudo-observations, the empirical copula function is
n
1X
Cb n (u, v) = 1 {Ui,n ≤ u, Vi,n ≤ v} , u, v ∈ (0, 1).
n i=1

Proposition 5.1. Let CXY be the bivariate copula of (Ui , Vi ) = (FX (Xi ), FY (Yi )). The

conditional cdf is CY |X≤ (v | p) = CXY (p, v)/p, for v ∈ (0, 1). Let pn = kn /n, and define the

sequence of empirical estimators,

n
Cbn (pn , v) 1 X  kn 
kn ,n (v) :=
Ab = 1 Ui,n ≤ , Vi,n ≤ v , v ∈ (0, 1).
pn kn i=1 n

Suppose there exists a function A : (0, 1) → (0, 1) such that as p ↓ 0,

sup CY |X≤ (v | p) − A(v) −→0.

v∈(0,1)
√
Then as n → ∞, with kn → ∞, kn /n → 0 , kn / n −→ ∞,

P
sup Abkn ,n (v) − A(v) −
→ 0.
v∈(0,1)

Proposition 5.2. Suppose the copula CXY admits a lower tail expansion in Eq. (5) with

tail order κ = 1 and tail dependence function b(w1 , w2 ; C).

With pn = kn /n, for fixed (w1 , w2 ) ∈ S1 := {(w1 , w2 ) : w1 > 0, w2 > 0, w1 + w2 = 2}, let
n
Cbn (pn w1 , pn w2 ) n X kn kn
 
b
b
kn ,n (w1 , w2 ) := = 1 Ui,n ≤ w1 , Vi,n ≤ w2 .
pn kn i=1 n n

Then as n → ∞, kn /n → 0, kn → ∞,

P
sup b
b
kn ,n (w1 , w2 ) − b(w1 , w2 ; C) −→ 0.
(w1 ,w2 )∈S1

5.2 Minimum distance estimation of CoVaR

The proposed estimation procedure is based on the classical principle of minimum distance

estimation (Parr & Schucany 1982), where one fits a parametric model by aligning it as

18
closely as possible to an empirical counterpart. A similiar idea has been used in Einmahl

et al. (2012) and Nolde et al. (2025) for extreme value estimation.

Following this general principle, we define a criterion function Qn (θ) as the integrated

distance between an empirical tail functional and its model-based counterpart, and estimate

the parameter θ by minimizing Qn (θ). The choice of the tail functional is guided by the

theoretical results established in Section 3, which provides a basis to determine the regimes

of v(q | p; C) based on the tail order in the lower and upper left corner, κ and κ2∗ .

With the assumptions in Theorem 3.5, we consider the following cases.

(i) Tail attraction

If κ = 1 and κ2∗ > 1, the tail mass is concentrated in the lower left corner. We

assume a parametric model for the tail dependence function in the lower left corner

b(w1 , w2 ; θ) with H(r) = b(1, r; θ) → 1 as r → ∞, and construct the criterion function

Z
Qkn ,n (θ) = b
b
kn ,n (w1 , w2 ) − b (w1 , w2 ; θ) dw1 dw2 , (6)
S1

where bbkn ,n (w1 , w2 ) is the empirical tail dependence function defined in Proposition 5.2.

By Theorem 3.5, the conditional quantiles can be estimated as, for q ∈ (0, 1),

v̂n (q | pn ) = H −1 (q; θ̂n ) · pn , (7)

which converges to 0 as pn → 0.

(ii) Tail repulsion

If κ > 1 and κ2∗ = 1, the tail mass is concentrated in the upper left corner, and

we will assume a parametric model for b(w1 , w2 ; C 2∗ ). The criterion function will

be constructed based on bbkn ,n (w1 , w2 ) of the 2-reflected data. By Theorem 3.5, the

conditional quantiles can be estimated by

v̂n (q | pn ) = 1 − (H 2∗ )−1 (1 − q; θ̂n ) · pn ,

19
 which converges to 1 as pn → 0.

(iii) Tail balance

If κ > 1 and κ2∗ > 1, then by Theorem 3.5, A(v) does not have any boundary atoms.

We therefore assume a proper parametric distribution A(v; θ) and define the criterion

function as
Z 1
Qkn ,n (θ) = Abkn ,n (v) − A(v; θ) dv, (8)
0

where Abkn ,n is the empirical estimator for the conditional cdf defined in Proposition 5.1.

The conditional quantile at level q is then estimated by inverting the fitted model:

v̂n (q | pn ) = A← (q; θ̂n ). (9)

The three cases above fall into a single tail regime, which covers the behavior of most

parametric copula families. As discussed in Section 3, however, mixed tail regimes may

occur; for example, a tail order of 1 in both corners indicates a mixed regime. In this case,

one could model using the first-order expansion of Student-t copula, as in Example 3.7, and

construct a criterion function that targets both corners,

Z
Qkn ,n (ρ, ν) = kn ,n (w1 , w2 )−b(w1 , w2 ; ρ, ν)
b
b + bb2∗
kn ,n (w1 , w2 )−b(w1 , w2 ; −ρ, ν) dw1 dw2 .
S1

The next two theorems establish the consistency of v̂n (q | pn ), the estimator of CoVaR on

the uniform scale. To recover CoVaR on the original scale, we can plug it into a marginal

quantile estimator, such as

 
\| pn ) = Fb −1 v̂n (q | pn ) ,
CoVaR(q S,n

whose consistency follows from the continuous-mapping theorem, once the marginal quantile

estimator is uniformly consistent on a neighbourhood of v(q | p; C).

Theorem 5.3. Assume the model A(·; θ) for the boundary cdf is correctly specified.

20
(A1) The parameter set Θ ⊂ Rp is non-empty, closed, and bounded (hence compact).

(A2) There exists a true parameter value θ0 ∈ Θ such that

C(p, v)
lim = A(v; θ0 ), uniformly in v ∈ (0, 1).
p↓0 p

(A3) For every θ ∈ Θ, the function v 7→ A(v; θ) is a proper distribution function on [0, 1],

continuous and strictly increasing on (0, 1). Moreover, the mapping (v, θ) 7→ A(v; θ)

is jointly continuous on [0, 1] × Θ.

(A4) For every ε > 0,

Z 1
inf A(v; θ) − A(v; θ0 ) dv > 0.
∥θ−θ0 ∥≥ε 0

Let θ̂n be the minimizer of the criterion function in Eq. (8), and the plug-in estimator of

the extreme conditional quantile in Eq. (9).

√
Under Assumptions (A1) to (A4), as n → ∞, kn → ∞, pn = kn /n → 0, and kn / n → ∞,

P
θ̂n −
→ θ0 ,

and for every fixed q ∈ (0, 1),

P
v̂n (q | pn ) − v(q | pn ; C) −
→ 0.

Theorem 5.4 treats the regime where v(q | p; C) converges to zero. In this case, we establish

consistency by showing the ratio converges to one in probability. Theorem 3.1 in Nolde

et al. (2025) presents a similar result for the adjustment factor in the case q = p.

Theorem 5.4. Assume the model for tail dependence function b(·; θ) is correctly specified

and assumed κ = 1.

(A1) The parameter space Θ ⊂ Rp is a compact subset.

21
(A2) There exists a true parameter θ0 ∈ Θ ⊂ Rp such that, for every

(w1 , w2 ) ∈ S1 := {(w1 , w2 ) : w1 > 0, w2 > 0, w1 + w2 = 2},

C(uw1 , uw2 )
= b(w1 , w2 ; θ0 ) + R(u; w1 , w2 ), sup |R(u; w1 , w2 )| −−→ 0.
u (w1 ,w2 )∈S1 u↓0

(A3) The map (w1 , w2 , θ) 7→ b(w1 , w2 ; θ) is continuous on S1 × Θ.

(A4) For every θ ∈ Θ the function

H(r; θ) := b(1, r; θ), r > 0,

is a proper, absolutely continuous distribution function on (0, ∞).

(A5) For every ε > 0,

Z
inf |b (w1 , w2 ; θ) − b (w1 , w2 ; θ0 )| dw1 dw2 > 0.
{θ:∥θ−θ0 ∥≥ε} S1

Let θ̂n minimize the sample objective Qkn ,n (θ) defined in Eq. (6), and let v̂n (q | pn ) be the

conditional quantile estimator in Eq. (7), Then, as n → ∞, kn → ∞, kn /n → 0,

P
θ̂n −
→ θ0 ;

For every fixed q ∈ (0, 1), as pn → 0,

v̂n (q | pn ) P
−
→ 1.
v(q | pn )

The finite-sample performance of the proposed estimation approach is assessed through

an extensive simulation study. The parametric extreme value models, along with detailed

estimation procedures and results, are reported in Section B of the supplementary material.

22
6 Empirical systemic risk analysis

This section applies the proposed framework to empirical data from the U.S. financial

market. The goal is to assess systemic risk contributions and exposures using CoVaR and

∆CoVaR, with an emphasis on how joint tail behavior and tail dependence with the system

shape these measures and reveal differences in the systemic roles of assets and institutions.

6.1 CoVaR and ∆CoVaR in the Copula-AR-GARCH model

Let Rt = (Rti , RtS ) denote the bivariate return vector at time t, consisting of returns

for institution i and the system S, for t = 1, 2, . . . , T . The return process is adapted to

the filtration {Ft }Tt=0 , generated by the observed history: Ft := σ (Rs : s ≤ t) . Under the

Copula-AR-GARCH model (Jondeau & Rockinger 2006):

Rtj = µtj + σtj Ztj , Ztj ∼ FZj , j ∈ {i, S},

2
where: µtj = E(Rtj | Ft−1 ), σtj = Var(Rtj | Ft−1 ), and Ztj are standardized innovations

with marginal distribution FZj . The AR and GARCH dynamics are given by:
pj
2 2 2
X
µtj = µj + ϕjk Rt−k,j , σtj = β0j + β1j Zt−1,j + β2j σt−1,j .
k=1

The joint conditional distribution of returns (Rti , RtS ) is determined by the copula CiS

linking the innovations:

r − µti  r − µ 
 
ti tS tS
FRt (rti , rtS | Ft−1 ) = CiS FZi , FZS .
σti σtS

Let the distress event of institution i, given the past information, be

n o
Di,t (p) := Rti ≤ VaRi|t (p)

and the enlarged information set be the smallest σ-field that includes both the past market

information and the event that institution i is in distress:

Hi,t (p) := Ft−1 ∨ σ (Di,t (p)) = σ (Ft−1 ∪ {Di,t (p)}) .

23
Then, the system CoVaR conditional on the information set Hi,t (p) is

 
CoVaRS|i,t (p) = µtS + σtS FZ−1
S
rS|i (p) · p (10)

with the adjustment factor rS|i (p) = CS|i≤← (p | p)/p.

The time-varying ∆CoVaR is given by

CoVaRS|i,t (p) − VaRS|t (p) σtS FZ−1

S
(rS|i (p) · p) − σtS FZ−1
S
(p)
∆ CoVaRS|i,t (p) = = −1 .
| VaRS|t (p)| |µtS + σtS FZS (p)|

For financial returns, the conditional mean µtS is typically negligible relative to the tail

quantiles, so we obtain,
 
FZ−1
S
rS|i (p) · p − FZ−1
S
(p)
∆ CoVaRS|i (p) ≈ , (11)
|FZ−1
S
(p)|

and Theorem 4.1 would apply.

The explicit formulas clarify the main drivers of the two risk measures and how they

should be interpreted. CoVaRS|i,t (p) in Eq. (10) measures the absolute level of system loss

conditional on institution i being in distress. As a high-frequency risk metric, it is largely

driven by changes in market volatility. ∆ CoVaRS|i (p) in Eq. (11), on the other hand, is

determined more by the dependence and market regime, captured by the adjustment factor

rS|i (p) and the tail index of FZS . A smaller adjustment factor and a larger tail index both

contribute to a more negative ∆ CoVaRS|i , indicating a greater contribution of firm-level

distress to systemic risk.

6.2 Systemic risk impact of the financial sector

A large literature, e.g., Bernal et al. (2014), documents the buildup of systemic risk in the

financial sector before the 2008 crisis, and the sector remains central to macroprudential

regulation. However, structural changes in the economy and the rapid expansion of the

technology sector raise the question of whether the systemic role of financial institutions has

24
 0.09 −0.8
Mean adjustment factor

Mean Delta CoVaR

0.08
Sector Sector
−1.0
broker−dealer broker−dealer
depository depository
insurance insurance
0.07 real−estate real−estate

−1.2

0.06

−1.4

2005 2010 2015 2020 2005 2010 2015 2020

Year Year

Figure 1: Mean estimates of rS|i (p) (left) and ∆ CoVaRS|i (p)(right) for large U.S. financial

institutions with p = 0.05, segmented by depository, broker-dealers, insurance companies,

and real estate firms. The estimates for rS|i (p) and ∆ CoVaRS|i (p) are calculated based on

the daily log return data from June 2000 to June 2025 using five-year rolling window.

broker−dealer depository
0.0600 0.065

0.0575

0.060

0.0550

0.055
0.0525
adjustment factor

0.0500
0.050
insurance real−estate
0.12 0.13

0.11
0.10

0.09
0.08

0.07

0.06

0.05
2005 2010 2015 2020 2005 2010 2015 2020
year

Figure 2: Path of the estimated adjustment factor rS|i (p) at p = 0.05 for 35 institutions that

are consistently among the top 100 firms by market capitalization throughout 2000-2025.

25
shifted over time. We address this question by estimating ∆ CoVaRS|i for large U.S. financial

institutions, using the S&P 500 index as a proxy for the system. For each rolling five-year

window over 2000–2025, we identify the 100 largest firms by market capitalization and

model their daily returns using marginal AR-GARCH specifications with skew- t innovations.

Based on the resulting filtered innovations, we then estimate the adjustment factor rS|i (p)

following the procedure proposed in Section 5.

Candidate models are derived from the tail expansions of parametric copulas listed in

Table 1, with further details provided in Section B.1 of the Supplementary Material. As

discussed in Section 5.2, the model choices are determined by the tail regime. In practice,

we tell the regime based on the pair of empirical tail dependence coefficients in the lower-left

and upper-left corners, (λ̂, λ̂2∗ ). A positive λ̂ with λ̂2∗ close to zero indicates tail attraction,

in which case b (w1 , w2 ) is modeled using the lower tail dependence function of the reflected

Gumbel or Student-t copula. A positive λ̂2∗ with λ̂ close to zero indicates tail repulsion,

and b∗2 (w1 , w2 ) is modeled using the lower tail dependence function of the Clayton copula.

If both coefficients are close to zero, the regime is classified as tail balance, and A(v) is

modeled using the boundary cdf of the reflected IPS copula. Following the simulation

experiments in the Supplementary Material, kn = 100 is used in the analysis.

Fig. 1 reports the mean estimates of rS|i (p) and ∆ CoVaRS|i (p) at p = 0.05, grouped

by Standard Industrial Classification (SIC) into depository institutions, broker-dealers,

insurance firms, and real estate firms. Fig. 2 as complementary to the aggregate view,

shows the time path of rS|i (p) for institutions that remained in the top 100 by market

capitalization throughout 2000–2025.

Overall, it suggests that the financial sector remains the major contributor to systemic risk

with the adjustment factors close to the lower bound p = 0.05 (compare the comonotonicity

bound in Proposition 2.3 when q = p) and large negative values of ∆CoVaRS|i . Over time,

26
 ETF Financial Tech ETF Financial Tech

AIG
−1.0 −1.0
SCHW

ORCL PNC INTC

BK WFC
NVDA AMD BAC QCOM META
GS
MSFT C MET
TQQQ AVGO MS AMZN
AMD ARKW
QQQ XLY GOOGL JPM
QCOM
−1.2 −1.2
SSO XLI ARKQ AAPL
AVGO INTC AAPL
BAC ARKQ UPRO VUG
NVDA
WFC BK C AMZN
ORCL
MS ARKW MSFT
META AIG
QQQ GOOGL PNC MET
SSO XLY GS JPM
−1.4 VUG XLI SCHW −1.4
UPRO TQQQ

0.05 0.06 0.07 0.08 0.09 0.10 0.05 0.06 0.07 0.08 0.09 0.10

Figure 3: CoVaRS|i (p) versus the adjustment factor rS|i (p) at p = 0.05 for major U.S.

financial institutions, large technology firms, and equity ETFs, based on daily returns for

two sample periods: 2012–2018 (left panel) and 2018–2025 (right panel).

as shown in Fig. 2, both banks and insurance firms exhibit a mild U-shaped pattern: their

systemic importance rises from the early 2000s, peaks around the financial crisis and its

aftermath, and then declines gradually in recent years. Real-estate firms, however, show a

different pattern. Their systemic impact increases sharply in the years leading up to the

crisis, declines markedly afterward, and then remains relatively stable in recent years.

6.3 Cross-sectional contributions to systemic risk

This section reveals cross-sectional differences in systemic risk contributions among large-

capitalization U.S. firms across major sectors. Fig. 3 and Fig. 4 plot the estimated

∆CoVaRS|i (p) against the adjustment factor rS|i (p) at p = 0.05 for firms spanning fi-

nancials, technology, consumer, energy, health care, real estate, and utilities over two sample

periods, 2012–2018 and 2018–2025.

As in Fig. 3, financial institutions and technology firms exhibit the strongest tail dependence

with the system, and this translates to strong tail attraction between firm-level distress and

system losses. It signals that these firms are likely systemically central: stress originating in

them tends to transmit rapidly and intensify system-wide tail losses, making them potential

originators and amplifiers of systemic risk.

27
 Consumer Energy Health care Real estate Utilities Consumer Energy Health care Real estate Utilities

0.0 0.0

EXC
DUK GIS KMB
D SO ABBV MRK PFE
AEP O
CL MO
−0.5 XEL −0.5
DLR PSA D
PSA NEE
AEP
NEE
VTR OXY
VTR DUK JNJ O DLR
GIS MRK
SRE WMT UNH LLY XEL
EOG AMT WMT XOM SO EQIX AMT
−1.0 −1.0 PM CVX PG
LLY KO SLB
PFE PEP PM PG SRE EXC EOG COP
COP KMB PEP SLB
CVX KO
JNJ OXY EQIX MO
CL
XOM ABBV
UNH
−1.5 −1.5

0.2 0.4 0.6 0.2 0.4 0.6

Figure 4: ∆ CoVaRS|i (p) versus the adjustment factor rS|i (p) at p = 0.05 for firms across

consumer, energy, health care, real estate, and utilities sectors, based on daily returns for

two sample periods: 2012–2018 (left panel) and 2018–2025 (right panel).

Firms from the other sectors, as shown in Fig. 4, also exhibit negative ∆CoVaRS|i , indicating

a positive contribution to systemic risk. But their tail behaviors are more heterogeneous,

combining both tail attraction and tail balance, and their adjustment factors and ∆CoVaRS|i

span a wider range.

In Fig. 3, we notice that the entities with the most negative ∆CoVaRS|i are large equity

exchange-traded funds (ETFs) and mutual funds. As discussed in financial literature

(Ramaswamy 2011), these investment vehicles typically do not originate distress, but they

act as important channels that amplify systemic losses.

Comparing the two periods, ∆CoVaRS|i exhibits an overall upward shift. This change is

because the innovation distribution has a less heavier tail in 2018-2025 compared to 2012-

2018. The behavior of rS|i (p) reveals changes in the dependence structure. In particular we

observe a clear reordering of systemic importance. In 2012–2018, systemic risk contributions

are dominated by financial institutions, while most technology firms, aside from a few

exceptionally large firms such as Microsoft and Google, have relatively modest impacts.

This pattern changes markedly in 2018–2025, with systemic importance shifting toward

the technology sector and away from financial firms. As shown in Fig. 4, this trend is

particularly evident for companies such as NVIDIA, AMD, and Broadcom, which have

28
 Commodities Fixed income Fx currency Commodities Crypto Fixed income Fx currency

1.0 1.0

VGLT VGIT IEF FXY

BND
AGG SHY TLT
SCHP UUP
0.5 0.5
GLD FXF FXY
TIP
SHY VGIT VGLT
UDN
UUP UNG IEF TLT
UNG SCHP
0.0 SLV FXE 0.0 FXF TIP AGG
DBA PPLT BND
FXE
FXB UDN
USO
DBA
GLD
−0.5 −0.5
PPLT XMR−USD
DBC
ADA−USD
USO SLV DBC
FXB SOL−USD
−1.0 −1.0 BTC−USD LTC−USD

−1.5 −1.5

0 2 4 6 8 0 2 4 6 8

Figure 5: ∆ CoVaRi|S (p) versus the adjustment factor ri|S (p) at p = 0.05 for fixed income

securities, cryptocurrency, commodities, computed based the daily returns from 2012–2018

(left panel) and 2018–2025 (right panel).

expanded rapidly with surging demand for AI infrastructure.

This shift is consistent with the findings in Section 6.2, which indicate a declining systemic

impact of the financial sector. At the same time, it points to an emerging source of systemic

risk in the technology sector, particularly among firms closely tied to the AI boom.

6.4 Systemic risk exposures and hedging assets

The ∆CoVaR in Eq. (11), when conditioning on system-wide distress, measures an asset’s

exposure to systemic risk. A value near zero indicates that the asset’s lower-tail distribution

is largely unaffected by systemic stress while a large positive ∆CoVaRi|S implies that the

asset tends to gain value when the system is under distress.

This information is useful for portfolio risk management as CoVaRi|S can help investors and

risk managers identify assets that improve portfolio resilience during episodes of system-wide

stress. Fig. 5 and Fig. 6 plot the estimated ∆CoVaRi|S (p) against ri|S (p) at p = 0.05, for a

wide range of assets including fixed income securities, commodities, and cryptocurrencies

based on daily returns over 2012–2018 and 2018–2025.

For structured hedging instruments shown in Fig. 6, we observe strong tail repulsion and

large positive ∆CoVaRi|S . Some of these instruments are explicitly designed to move against

29
 Hedge Instrument Hedge Instrument

2.0 2.0

DOG
DOG
PSQ
TZA PSQ
1.5 1.5
RWM VIXY
HDGE RWM
SJB VIXY BTAL TZA
VIXM
1.0 1.0 VIXM

BTAL

HDGE
0.5 0.5
SJB

0.0 0.0

0 5 10 15 20 0 5 10 15 20

Figure 6: ∆ CoVaRi|S (p) versus the adjustment factor ri|S (p) at p = 0.05 for hedging

instruments based on daily returns from 2012–2018 and 2018–2025.

the system, such as inverse equity ETFs and VIX-linked products. Others provide protection

through exposure to specific stress channels; for example, SJB, which holds short positions

in high-yield corporate bonds, gains from the deterioration of credit markets during systemic

crises.

Fig. 5 identifies several natural hedging assets, including long-duration U.S. Treasury bond

ETFs (e.g., TLT, VGLT, and IEF) and currency ETFs (e.g., FXY, FXF). These assets

exhibit weak tail repulsion, or negative tail balance, indicating that they tend to appreciate

or remain stable under systemic distress.

The right panel of Fig. 5 reports the estimates for cryptocurrencies. Although these assets

are often viewed as isolated from the traditional financial assets (Corbet et al. 2018), our

results show that they still display weak tail attraction, or positive tail balance. In addition,

their marginal tail indices are noticeably larger than those of other asset classes, pointing

to greater exposure to idiosyncratic extreme shocks.

The sign reversal in the bond-stock correlation has been widely documented in the economics

literature, see e.g., Campbell et al. (2025), and is often linked to shifts in the macroeconomic

regime. Comparing the two panels of Fig. 5 suggest a similar change in tail dependence.

The ri|S (p) of fixed-income bonds and safe-haven currencies decreases and ∆CoVaRi|S

becomes less positive, and in some cases, turns negative. This indicates that their hedging

30
effectiveness against systemic risk declines over time, with these assets increasingly comoving

with system losses instead of offsetting them.

7 Discussion

In this work, we develop a theoretical and inferential framework for conditional Value-at-Risk

(CoVaR) based on copula and extreme value theory. It clarifies how the tail dependence

structure determines the limiting behavior of CoVaR and ∆CoVaR, thereby providing a

principled basis for interpreting these measures in systemic risk analysis.

The paper also contributes an empirical analysis based on the proposed framework, highlight-

ing several important findings. Our results show that, although the financial sector remains

the major contributor to systemic risk, its impact has weakened in recent years, coinciding

with the rising systemic importance of major technology firms. From the perspective of

macroprudential regulation, this shift suggests an emerging source of systemic risk from the

technology sector. The 2007–2009 financial crisis illustrates how the repricing of housing

assets became systemic when compounded with the structure of the financial system. Our

findings point to a related but distinct concern: even though the shocks originate outside

the traditional financial sector, they may generate system-wide effects as major technology

firms have become deeply embedded in the system’s downside dependence structure. An

important direction for future research is therefore to develop systemic risk monitoring and

stress-testing frameworks that account for the potential repricing of AI-linked assets and

their impacts for system-wide stability.

Data availability statement

The data used in this study were obtained through Wharton Research Data Services

(WRDS) and Yahoo Finance. The replication package, including the code and instructions

31
for accessing the data, is available in a GitHub repository.

Disclosure statement

The authors report there are no competing interests to declare.

References
Adrian, T. & Brunnermeier, M. K. (2016), ‘CoVaR’, American Economic Review

106(7), 1705–1741.

Ascorbebeitia, J., Ferreira, E. & Orbe, S. (2022), ‘The effect of dependence on European

market risk. a nonparametric time varying approach’, Journal of Business & Economic

Statistics 40(2), 913–923.

Bernal, O., Gnabo, J.-Y. & Guilmin, G. (2014), ‘Assessing the contribution of banks,

insurance and other financial services to systemic risk’, Journal of Banking & Finance

47, 270–287.

Bernard, C. & Czado, C. (2015), ‘Conditional quantiles and tail dependence’, Journal of

Multivariate Analysis 138, 104–126.

Borio, C. & Zhu, H. (2012), ‘Capital regulation, risk-taking and monetary policy: a missing

link in the transmission mechanism?’, Journal of Financial Stability 8(4), 236–251.

Campbell, J. Y., Pflueger, C. & Viceira, L. M. (2025), Bond-stock comovements, Technical

report, National Bureau of Economic Research.

Corbet, S., Meegan, A., Larkin, C., Lucey, B. & Yarovaya, L. (2018), ‘Exploring the dynamic

relationships between cryptocurrencies and other financial assets’, Economics Letters

165, 28–34.

32
Einmahl, J. H. J., Krajina, A. & Segers, J. (2012), ‘An M-estimator for tail dependence in

arbitrary dimensions’, The Annals of Statistics 40(3), 1764–1793.

Fan, Y., Härdle, W. K., Wang, W. & Zhu, L. (2018), ‘Single-index-based CoVaR with very

high-dimensional covariates’, Journal of Business & Economic Statistics 36(2), 212–226.

Girardi, G. & Ergün, A. T. (2013), ‘Systemic risk measurement: multivariate GARCH

estimation of CoVaR’, Journal of Banking & Finance 37(8), 3169–3180.

Hua, L. & Joe, H. (2011), ‘Tail order and intermediate tail dependence of multivariate

copulas’, Journal of Multivariate Analysis 102(10), 1454–1471.

Hua, L. & Joe, H. (2014), ‘Strength of tail dependence based on conditional tail expectation’,

Journal of Multivariate Analysis 123, 143–159.

International Monetary Fund (2011), Global financial stability report: Grappling with crisis

legacies, Technical report, International Monetary Fund, Washington, DC.

Jaworski, P. (2017), ‘On conditional value at risk for tail-dependent copulas’, Dependence

Modeling 5(1), 1–19.

Joe, H. (2014), Dependence Modeling with Copulas, CRC Press, Boca Raton, FL.

Jondeau, E. & Rockinger, M. (2006), ‘The copula-GARCH model of conditional dependencies:

An international stock market application’, Journal of International Money and Finance

25(5), 827–853.

Li, X. & Joe, H. (2023), ‘Estimation of multivariate tail quantities’, Computational Statistics

& Data Analysis 185, 107761.

Li, X. & Joe, H. (2026), ‘Properties of CoVaR based on tail expansions of copulas’, Journal

of Multivariate Analysis 211, 105510.

33
Nolde, N. & Zhou, C. (2021), ‘Extreme value analysis for financial risk management’, Annual

Review of Statistics and Its Application 8(1), 217–240.

Nolde, N., Zhou, C. & Zhou, M. (2025), ‘Tail risk in the tail: Estimating high quantiles

when a related variable is extreme’, Journal of the American Statistical Association .

Parr, W. C. & Schucany, W. R. (1982), ‘Minimum distance estimation and components

of goodness-of-fit statistics’, Journal of the Royal Statistical Society Series B: Statistical

Methodology 44(2), 178–189.

Ramaswamy, S. (2011), Market structures and systemic risks of exchange-traded funds,

Technical report, Bank for International Settlements.

34