ESG, Risk, and (Tail) Dependence: A B B A

ESG, Risk, and (tail) dependence
Karoline Baxa , Özge Sahinb , Claudia Czadob , Sandra Paterlinia,∗

a Department of Economics and Management, University of Trento, Trento, Italy
b Department of Mathematics, Technical University of Munich, Munich, Germany
Abstract
While environmental, social, and governance (ESG) trading activity has been a distinctive feature of financial mar-
kets, the debate if ESG scores can also convey information regarding a company’s riskiness remains open. Regu-
latory authorities, such as the European Banking Authority (EBA), have acknowledged that ESG factors can con-
tribute to risk. Therefore, it is important to model such risks and quantify what part of a company’s riskiness can
be attributed to the ESG ratings. This paper aims to question whether ESG scores can be used to provide infor-
mation on (tail) riskiness. By analyzing the (tail) dependence structure of companies with a range of ESG scores,
using high-dimensional vine copula modelling, we are able to show that risk can also depend on and be directly
associated with a specific ESG rating class. Empirical findings on real-world data show positive not negligible de-
pendencies between clusters determined by ESG scores, especially during the 2008 crisis.
Keywords: ESG scores, Risk, Dependence, Tail dependence, Vine Copula models
JEL classification: G32, C51, C58
1. Introduction
After the 2007-2009 financial crisis, many models which used to capture the dependence between a large num-
ber of financial assets were revealed as being inadequate during crisis. Moreover, research has shown that the
dependence structure of global financial markets has grown in importance in areas of optimal asset allocation,
multivariate asset pricing, and portfolio tail risk measures (Xu and Li, 2009). Consequently, the enormous losses
and the increased volatility in the global financial market elicited calls for an even more diligent risk management.
Over the past decade, the interest in socially responsible investments has grown exponentially (Auer and
Schuhmacher, 2016). The availability of non-financial data, including corporate social responsibility (CSR) or
environmental, social, and governance (ESG) data has skyrocketed, and has gained great interest from investors
∗ Corresponding author: Email address: [email protected]
Preprint submitted to SSRN May 15, 2021
Electronic copy available at: https://ssrn.com/abstract=3846739

for various reasons. These ESG scores indicate a level of ESG performances and are based on several criteria, given
by a rating institution (Bhattacharya and Sharma, 2019). The rating institutions use quantitative and qualitative
methods to assign an ESG score to a company (Berg and Lange, 2020). In brief, companies are awarded positive
scores for ESG responsible behavior, but are awarded negative scores for ESG irresponsible behavior.
The increase in demand in socially responsible investments stems from investors and asset managers pres-
sured by stakeholders to push companies to behave responsible and improve their ESG strategy (Henriksson et al.,
2019). The biennial 2018 Global Sustainable Investment Review states that over $30 trillion had been invested with
explicit ESG goals (Global Sustainable Investment Alliance (GSIA), 2018). According to a 2018 global survey, more
than 50% of the international asset owners are currently considering or already implementing ESG scores in their
investment strategy (Consolandi et al., 2020). Furthermore, Eccles and Klimenko (2019) state that this interest in
ESG assets is driven by the growing evidence of the positive impact of ESG materiality on financial performance.
Also, the European Banking Authority (EBA) stressed the importance of including ESG information into the reg-
ulatory and supervisory framework of EU credit institutions (European Banking Authority, 2018). Nevertheless,
using ESG factors has made the investment process noticeable more complex (Berg and Lange, 2020).
While more than 2000 empirical studies have been done analysing ESG factors and financial performance,
little is known about the dependence structure and associated risks (Friede, 2019; Shafer and Szado, 2018). This
is especially important as ESG scores are often linked to investment risk. The EBA define ESG risks as “the risks
of any negative financial impact to the institution stemming, from the current or prospective impacts of ESG
factors on its counterparties" (p. 28) (EBA, 2020). While regulators finally acknowledge the role played by ESG
factors in determining part of a companies risk, so far no guidelines are provided in how to best capture and
quantify such risk. If this is accurate, then ESG scores should possibly deliver some information on a company’s
ESG risk and its riskiness as a whole. This also means that companies that have the same ESG rating should
share similar risk characteristics and properties. This is especially important in times of increased volatility as
understanding the (tail) dependence structure among similarly rated assets is essential in order to set in place
effective risk management and diversification strategies (Ane and Kharoubi, 2003; Frahm et al., 2005; Malevergne
et al., 2005; Berg and Lange, 2020).
This research aims to contribute to the understanding of the dependence structure of assets belonging to the
same ESG class and provide some new indicators to capture such (tail) dependence. To model dependence we rely
on a high-dimensional vine copula modelling (Bedford and Cooke, 2001, 2002; Aas et al., 2009; Joe, 2014; Czado,
2019). This allows us to show to our knowledge for the first time that risk can also depend on ESG rating class.
Furthermore, by quantifying the overall dependence and tail dependence among assets that belong to the same

ESG class, we show that these dependencies are not negligible, especially in times of crisis.
The paper is structured as follows; Section 2 summarizes the literature and creates an understanding of ESG
scores and the occurrence of (tail) risk and dependence structures. while Section 3 introduces dependence mod-
elling and vine copulas. Section 4 then proposes the R-vine Copula ESG Risk model. The S&P 500 data is described
in Section 5. Then Section 6 reports the empirical results. Lastly, Section 7 provides an overall summary and out-
look for future research.
2. ESG Scores, Dependence, and Risk
Overall, ESG scores try to capture the amount of positive ESG disclosure and performance of a company (Bhat-
tacharya and Sharma, 2019). In general, companies are awarded grades measuring their ESG responsible behavior:
the higher the grade, the better the ESG performance. Acting as complementary non-financial information, ESG
scores can have the potential to increase the accuracy in performance forecasts and risk assessments (Achim and
Borlea, 2015).
In 2018, the European Commission published its Action Plan on Financing Sustainable Growth which provided
the European Union (EU) with a roadmap on sustainable finance and for future work across financial systems (Eu-
ropean Banking Authority, 2018). Furthermore, the European Banking Authority (EBA) stressed the importance of
including ESG information into the regulatory and supervisory framework of EU credit institutions (European
Banking Authority, 2018). It has also been shown that ESG factors may impact financial performance by substan-
tiating themselves in “financial or non-financial prudential risks, such as credit, market, operational, liquidity and
funding risks" (p.27) (EBA, 2020). According to the EBA, ESG risks are defined to materialize when ESG factors
have a negative impact on the financial performance or solvency (EBA, 2020). Furthermore, it is argued that the
materiality of ESG risks depends on the risks posed by ESG factors over different time frames (EBA, 2020). If this
is accurate, the ESG ratings should contain information about the company’s risk. Even though ESG has mostly
been defined in terms of risk by the regulator, and there is an ongoing debate on the effects of using ESG scores
on the financial performance research, there is no consensus on the (tail) dependence structure of assets within
and between each ESG rating class. Understanding the (tail) dependence structure of several assets, however, is
necessary in order to access inherent risks in the financial market.
Generally, market returns are assumed to follow a multivariate normal distribution; however, research has
shown that this is found to be not accurate in reality, and left tails are often heavier than right tails (Cont, 2001;
Jondeau and Rockinger, 2003). In the last years, dependencies between all financial asset returns have increased
due to globalization effects (Frahm et al., 2005). Therefore, modelling the dependence among assets and under-

standing the appearance of joint (tail) risk is especially important for asset pricing and risk management. Tail risk
or tail dependence is characterized as the probability of an extremely large negative (positive) return of an asset
given that the other asset yields an extremely large negative (positive) return,and is commonly quantified by the
so-called tail-dependence coefficient (Embrechts et al., 2001; Frahm et al., 2005; Xu and Li, 2009). Tail risk arises
when the likelihood of an extreme event that is more than three standard deviations away from the mean is more
likely to occur than shown by a normal distribution (Kelly and Jiang, 2014). Generally it is accepted that tail depen-
dence can be used as a proxy of systemic risk and tail risk has been linked to negative consequences for corporate
investment and risk-taking (Gormley and Matsa, 2011; Gormley et al., 2013; Shirvani and Volchenkov, 2019).
In the last years, companies with high ESG scores, so called ESG leaders, have experienced an intense influx
in investments from asset managers and institutional investors, possibly leading to an increased valuation (Berg
and Lange, 2020). This could lead to potential valuation risk which should be considered before any investment
decision (Dunn et al., 2018; Lioui, 2018; Pollard et al., 2018; Berg and Lange, 2020). Some researchers argue that
responsible ESG practices might mitigate the market’s perception of a company’s tail risk and, therefore, reduce
ex-ante expectations of a left-tail event (Shafer and Szado, 2018). In brief, they state that considering responsible
business practices when creating an equity portfolio can act as insurance against left-tail risk and, with that, pro-
tect company value. This is also in line with Wamba et al. (2020) who add that especially positive environmental
performance can act as an insurance for companies, reducing the probability of a negative event occurring and
with that reducing the company‘s systematic risk. Ashwin Kumar et al. (2016) add that positive ESG practices can
make a company less vulnerable to reputation, political and regulatory risk and thus leading to lower volatility
of cash flows and profitability’ (p. 292). Furthermore, positive ESG performance may generate more loyalty from
customers and employees plus, through that, protect companies from unforeseen harmful events, resulting in re-
duced tail risk (Shafer and Szado, 2018). Besides, better ESG performance allows companies to experience adverse
events less often and loose less value if they do occur (Minor, 2011).
Others recognize that improving ESG performance can also help with risk exposure (Giese et al., 2019). Maiti
(2020) finds that overall portfolios developed using the overall ESG as well as the individual E (Environment),
So (Social) and G (Governance) factors show generally better investment performance, implying policy modifi-
cations, while Breedt et al. (2019) show that ESG-tilted portfolios do not necessarily have higher risk-adjusted
returns. Other scholars have observed a mitigating effect of ESG performance on stock price crash risk if a com-
pany has less effective governance (Kim et al., 2014). Hoepner et al. (2016) found supporting evidence that ESG
engagements can be associated with subsequent reductions in downside risk. However, their findings are limited
as they are based on one single institutional investor. Additionally, Sherwood and Pollard (2017) finds that inte-

grating ESG emerging market equities into institutional portfolios can create higher returns and lower downside
risks compared to non-ESG equity investments, while Shafer and Szado (2018) add that better ESG performance
significantly reduces the perceived one-month ahead tail risk.
Looking at a country’s creditworthiness, Capelle-Blancard et al. (2019) show that countries with above-average
ESG scores are linked to reduced default risk and smaller sovereign bond yield spreads. Additionally, Breuer et al.
(2018) show that the cost of equity is reduced when a company invests in CSR and vice versa, given that it is
located in a country with high investor protection. Moreover, Li et al. (2017) find that companies in regions of
high social trust tend to have lower tail risk. A possible reason why environmental practices have the power to
control tail risk could be that a company’s good social records will be more valuable in the long run due to a lower
frequency of litigation (Goldreyer and Diltz, 1999). Additionally, higher environmental standards are valuable to
shareholders due to companies avoiding litigation costs, reputation losses, and environmental hazards (Chan and
Walter, 2014). Furthermore, ESG can drive an asymmetric return pattern in which social responsible investment
(SRI) funds (using positive rather than negative screenings) outperform conventional funds in times of crisis but
underperform in calm periods (Nofsinger and Varma, 2014). Moreover, Bae et al. (2019) find that CSR reduces the
costs of high leverage and decreases losses in market share when firms are highly leveraged. However, it is possible
that ESG risk is non-diversifiable. Here Dorfleitner et al. (2016) provide evidence that ESG ratings are subject to a
non-diversifiable risk component as it depends on the overall market.
Despite these extensive efforts, the effect of ESG scores on dependence and (tail) risks has not yet been clearly
understood. While investors have been aware of the need to use more sophisticated models to assess the depen-
dence behavior of assets since the financial crisis, such models, allowing investors to quantify dependence and
(tail) risk among ESG-based classes of assets, have yet to be introduced. In this paper, we focus on dependence
modelling via vine copulas (Czado, 2019), which are described in the next sections.
3. Dependence Modelling and Vine Copulas
Rarely a distribution follows the strict spherical assumptions with the same type of dependence structure
across the components of the distribution as implied by correlation (Dorey and Joubert, 2005). While Pearson
correlation has been used as a measure of pairwise dependence, modern risk management requires a thorough
stochastic understanding beyond linear correlation. The copula approach accounts for complex dependency pat-
terns, such as non-symmetry and dependence in the extremes.
A copula C is a cumulative distribution function (cdf) with uniform marginals on the unit interval. Sklar’s
theorem (1959) states that if F is a continuous d -dimensional distribution function for X = (X 1 , . . . , X d )> with a

univariate cdf F p (x p ) of a continuous random variable X p for p = 1, . . . , d with its realizations x p , the joint distri-
bution function F can be written as
F (x 1 , . . . , x d ) = C (F 1 (x 1 ), . . . , F d (x d )). (1)
The corresponding density is

d
Y
f (x 1 , . . . , x d ) = c(F 1 (x 1 ), . . . , F d (x d )) · f p (x p ), (2)
p=1
¢>
where c is the d-dimensional copula density of the random vector F = F 1 (X 1 ), . . . , F d (X D ) ∈ [0, 1]d and f p (x p ) is
¡
the associated univariate marginal density of F p (x p ) for p = 1, . . . , d .
Different copula types can accommodate flexible dependence patterns in the bivariate case (d = 2), as shown
in Table 1.1 Nevertheless, the existing parametric families of multivariate copulas are not as flexible as the bivari-
ate copula families to represent complex dependence patterns. For instance, the multivariate Gaussian copula
does not accommodate any tail dependence and has been criticized immensely after the 2008 financial crisis (Li,
2000; Salmon, 2012; Puccetti and Scherer, 2018; Czado, 2019). The multivariate Student’s t copula apprehends
tail dependence but does not capture any asymmetry in the tails. Furthermore, these multivariate exchangeable
Archimedean copulas become inflexible as the dimension increases since they model the dependence between
a large number of pairs of variables using not more than two parameters. If no dependence is found, and the
random variables are independent, the independence copula best models their behavior.
Itau1 Copulas BB Copulas

Properties t F N C J G I BB1 BB7 BB8
Positive Dependence X X X X X X - X X X
Negative Dependence X X X - - - - - - -
Tail Asymmetry - - - X X X - X X X
Lower Tail Dependence X - - X - - - X X -
Upper Tail Dependence X - - - X X - X X X
Table 1: Parametric copula families and their properties without rotations and reflections. Notation of copula families:t = Student’s t, F = Frank,
N = Gaussian, C = Clayton, J = Joe, G = Gumbel, I = Independence, BB1 = Clayton- Gumbel, BB7 = Joe-Clayton, BB8= Extended Joe
1 Copula families for which the parameter estimation by Kendall’s τ inversion is available without rotations
To accommodate a great variety of dependence structures in higher dimensions and overcome the issues of
multivariate elliptical, Archimedean, and even other copulas; vine copulas (so-called pair copula constructions)
were made operational for data analysis by Aas et al. (2009). They are a class of copulas which are based on con-
ditioning ideas first proposed by Joe (1996) and further more developed by Bedford and Cooke (2001, 2002). The
d ·(d −1)
approach allows to construct any d -dimensional copula and its density by 2 bivariate copulas and their den-
1 To extend the range of dependence, counterclockwise rotations and reflections of the copula density are included. This allows the Gumbel,
Clayton, Joe, BB1, BB7 and BB8 to also accommodate negative dependence (τ < 0). For more details, we refer to Chapter 3 of Czado (2019).

sities. Furthermore, the expression can be represented by an undirected graphical structure involving a set of
linked trees, i.e., regular vine (R-vine) (Bedford and Cooke, 2001).
A d -dimensional R-vine consists of d − 1 nested trees; each tree Tm has a node set Vm and an edge set E m
for m = 1, . . . , d − 1. The nodes in the tree level one and m + 1 are given by V1 = {1, . . . , d } and Vm+1 = E m for
m = 1, . . . , d −2, respectively. Thus, the edges in the tree level m are the nodes in the tree level m +1. The proximity
condition requires that if an edge connects two nodes in Tm+1 , their associated edges in Tm must have a shared
node in Tm .
To express a vine copula’s density, each edge in a R-vine is linked to a bivariate copula. Therefore, it is com-
putationally tractable and uniquely determined (Kurowicka and Cooke, 2006). Let c e a ,e b ;D e be a bivariate copula
density associated with an edge e in a regular vine between node a and b, with D e being the conditioning set of
e = {a, b} and F p denoting the marginal distribution function for p = 1, . . . , d . Then a d -dimensional vine copula
density c under the simplifying assumption can be constructed as follows:
dY
−1 Y ¡ ¡ ¢
c(F 1 (x 1 ), . . . , F d (x d ) = c e a ,e b ;D e F e a |D e (x e a |x D e ), F e b |D e x e b |x D e ) , (3)
m=1 e∈E m
where x D e = (x z )z∈D e is a subvector of x = (x 1 , . . . , x d )> ∈ Rd , F e a |D e is the conditional distribution function of the
random variable X e a |X D e = x D e , which can be calculated recursively (Joe, 1996). The set D e is called the condition-
ing set, and the indices e a , e b form the conditioned set. With the simplifying assumption, a bivariate copula density
does not depend on the specific value of its conditioning variables but only on the conditioning values through its
arguments. For more details, we refer to Stoeber et al. (2013).
Figure 1 shows a 6-dimensional R-vine with its edge labels. It has five tree levels, and the node set in the first
tree level is V1 = {1, 2, 3, 4, 5, 6}. The edge set in the second tree is E 2 = {{1, 4; 2}, {1, 5; 2}, {2, 3; 1}, {1, 6; 3}}. In the higher
tree levels, it holds V2 = E 1 , V3 = E 2 , V4 = E 3 , V5 = E 4 and V6 = E 5 . The tree levels are selected sequentially by
a top-down approach, i.e., the selection starts with the first tree T1 and continues tree by tree up to the last tree
Td −1 . The chosen trees satisfy the proximity condition. For instance, in T2 the nodes {1, 3} and {3, 6} are joined
since they share a common node {3} as shown in Figure 1. After the association of each edge of the R-vine with a
bivariate copula density, we can write its vine copula density as in Equation (4), where the edge e = {1, 4; 2} in E 2 is

associated with a bivariate copula density c 1,4;2 , for example.
¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢
c F 1 (x 1 ), . . . , F 6 (x 6 ) =c 13 F 1 (x 1 ), F 3 (x 3 ) · c 12 F 1 (x 1 ), F 2 (x 2 ) · c 36 F 3 (x 3 ), F 6 (x 6 ) · c 24 F 2 (x 2 ), F 4 (x 4 ) · c 25 F 2 (x 2 ), F 5 (x 5 )
¡ ¢ ¡ ¢ ¡ ¢
· c 14;2 F 1|2 (x 1 |x 2 ), F 4|2 (x 4 |x 2 ) · c 15;2 F 1|2 (x 1 |x 2 ), F 5|2 (x 5 |x 2 ) · c 23;1 F 2|1 (x 2 |x 1 ), F 3|1 (x 3 |x 1 )
¡ ¢ ¡ ¢
· c 16;3 F 1|3 (x 1 |x 3 ), F 6|3 (x 6 |x 3 ) · c 26;1,3 F 2|1,3 (x 2 |x 1 , x 3 ), F 6|1,3 (x 6 |x 1 , x 3 )
(4)
¡ ¢ ¡ ¢
· c 35;1,2 F 3|1,2 (x 3 |x 1 , x 2 ), F 5|1,2 (x 5 |x 1 , x 2 ) · c 34;1,2 F 3|1,2 (x 3 |x 1 , x 2 ), F 4|1,2 (x 4 |x 1 , x 2 )
¡ ¢ ¡ ¢
· c 46;1,2,3 F 4|1,2,3 (x 4 |x 1 , x 2 , x 3 ), F 6|1,2,3 (x 6 |x 1 , x 2 , x 3 ) · c 45;1,2,3 F 4|1,2,3 (x 4 |x 1 , x 2 , x 3 ), F 5|1,2,3 (x 5 |x 1 , x 2 , x 3 )
¡ ¢
· c 56;1,2,3,4 F 5|1,2,3,4 (x 5 |x 1 , x 2 , x 3 , x 4 ), F 6|1,2,3,4 (x 6 |x 1 , x 2 , x 3 , x 4 ) .
T1 T2 T3 T4 T5
1,3 1,6;3
3 1 1,3 3,6 1,6;3 2,6;1,3 4,6;1,2,3
4,6
2,6
3,6 1,2
2,3
;1,
;1,
;1
2, 3
3
5,6;1,2,3,4
6 2 1,2 2,3;1 3,4;1,2
,3
2
3,5
;2
1, 5
2
;1,
2, 4
2, 5
;1,
1,4
;1,
3,4
;2
4,5
2
4 5 2,4 2,5 1,4;2 1,5;2 3,5;1,2 4,5;1,2,3
Figure 1: Example of a 6-dimensional regular vine with edge labels.
To specify a vine copula model, a vine tree structure and the associated bivariate copula families with cor-
responding parameters are needed. For the selection of vine tree structures, we follow the sequential top down
approach proposed by Dißmann (2013) and Dißmann et al. (2013). It starts with the tree level one and finds
the maximum spanning tree, where each edge has a predefined weight, e.g., the absolute value of the empirical
Kendall’s τ between the nodes forming the edge. Then, from the set of bivariate copula families in Table 1, we
select the optimal pair copula families using the Akaike Information Criterion (AIC) (Akaike, 1998), which has
been shown to have good copula family selection properties and high accuracy (Brechmann, 2010). Since each
pair copula family can be specified sequentially and independently at the same tree level, an alternative selection
criteria like the BIC would not be needed to induce sparsity at this step. In the next step, we compute the pseudo
copula data for the second tree level T2 , using the estimated pair copulas in the first tree level T1 . These input data
is similarly used to select the second tree level T2 of the vine and corresponding pair copula families with their
parameters. The approach continues sequentially up to the last tree level. Following this method, we estimate
the vine tree structure together with its pair copula families and parameter values. As we choose a pair copula
family corresponding to each edge separately, the parameters estimation step is at most a two-dimensional opti-
mization, which is computationally efficient. Moreover, the performance of this method is satisfactory compared

to the full log likelihood method, which could be computationally intractable in high dimensions (Haff, 2012).
More details about R-vines are given by Kurowicka and Cooke (2006); Kurowicka, D. and Joe (2011); Joe (2014); and
Czado (2019).
4. The R-Vine Copula ESG Risk Model
S,q
In the following, we introduce the ESG class dependence indicator D j (τ) with τ as our dependence mea-
S,q
sure and the ESG class lower tail dependence indicator D j (λ) with λ as the lower tail dependence measure to
capture the overall and left-tail ESG class dependence for each asset j within each industrial sector S and period
q. Inspiration for these indicators arises from Brechmann and Czado (2013), who propose to model dependence
across geographical equity sectors. As ESG score values are weighted differently within industrial sectors, we fit
a separate R-vine copula ESG risk model for any set of assets belonging to a specific industrial sector S in period
q. Within each industrial sector S and period q, depending on the ESG scores, assets can be attributed to four
different ESG class groups (k ∈ A, B,C , D) following the thresholds given by Refinitiv (2019). Assets are given a
D grade for an ESG score lower than 25, C for an ESG score between 25 and lower than 50, B for an ESG score
between 50 and lower than 75 and finally A for an ESG score between 75 and 100. The lower the ESG score, the
S,q
poorer the relative ESG performance. We compute the mean ESG score (E SG j ) for each asset j in sector S over
a specific time period q. We then use these values to group each asset j within each sector S in period q into ESG
S,q S,q
class groups (R j ) and compute the ESG class indices per ESG class k (I t ,k , i.e. the ESG class index A in sector S
S,q S,q
in period q on trading day t is I t ,A ) by linearly combining asset returns (Y t , j ,k defined as the return of an asset j in
S,q
sector S in period q on trading day t belonging to ESG class k) in the same mean ESG class (R j ) using the market
S,q
capitalization weights. Later, in the R-vine model, we connect the assets to their belonging ESG class index (I t ,k )
S,q
and industrial sector index (I t ) at time t. As shown in Figure 2 and 3 we use only the variable names such as I S,q ,
S,q S,q S,q S,q S,q
I A , A1 etc. to denote the nodes. This association data is given by I t , I t ,A , Y t ,1,A .
As can be seen, we fix the first two tree levels of the R-vine. The first tree level specification allows us to cap-
ture the behaviour of an asset in a period compared to its ESG peers within the same sectors as well as the the
dependence between ESG class index and industrial sector index. In the second tree level we capture the de-
pendence of an asset with its sector after the effect of its ESG class index is removed. Thus, fixing the first two
trees is a necessary step in our R-vine copula ESG risk model. The R-vine model is structured as follows. For the
first tree level T1 shown in Figure 2, the central node is always chosen to be the sector index {I S,q } while the ad-
S,q S,q S,q S,q
jacent nodes are the ESG class indices: Index A {I A }, Index B {I B }, Index C {IC }, Index D {I D }. Then assets
S,q S,q S,q S,q S,q S,q S,q S,q
{A 1 , ..., A S,q }, {B 1 , ..., B S,q }, {C 1 , ...,C S,q }, and {D 1 , ..., D S,q } within the sector S in period q are linked to their
nA nB nC nD

S,q
belonging ESG class index, where n k denotes the number of assets in sector S which are in ESG class k in period
q. The number of assets can differ overall across sectors S and also across ESG class in different time periods q
within sectors S as assets can change ESG class over time. Thus, the assets are in contiguous groups based on
S,q S,q S,q
the mean ESG score in time period q. In the second tree level T2 , the nodes {I A , I S,q }, {I B , I S,q }, {IC , I S,q }, and
S,q
{I D , I S,q }, which are the edges in T1 , are sequentially ordered. The new order is motivated by the ESG Reuters
thresholds which divides assets in contiguous groups based on the mean ESG scores. Additionally, the other edges
S,q S,q S,q S,q
of tree T1 are connected to an edge with a common ESG class index. For example, {A 1 , I A }, ..., {A S,q ,IA }
nn A
S,q S,q S,q S,q S,q S,q S,q S,q S,q S,q S,q
is connected to {I A , I S,q }, {B 1 , I B }, ..., {B S,q , I B } to {I B , I S,q }, {C 1 , IC }, ..., {C S,q , IC } to {IC , I S,q
}, and
nB nC
S,q S,q S,q S,q S,q
{D 1 , I D }, ..., {D S,q , I D } to {I D , I S,q }.
n D
S,q
B2
S,q
B S,q S,q
nB B S ,q B1
B2
q
n S ,q , S S,
S,q
B I ,q q I
B S, , B
,I B
B1
S,q
S,q
Index B (I B )
S,q S,q
A S,q q
C S,q
nA A S ,q S, nC
I B , I S,q
n S ,q , S q ,I C
S,q
A I ,q S , ,q
A C nS C
S,q S,q S,q S,q S,q S,q

A 2 ,I A I A , I S,q IC , I S,q C2 , IC
S,q S,q S,q S,q
A2 Index A (I A ) Sector Index (I S,q ) Index C (IC ) C2
q
S,
q I
S, , A C S ,q
A1 1
, I S ,q
I D , I S,q
C
S,q
S,q S,q
A1 C1
S,q
Index D (I D )
q
S, D S ,q
I
q , D
S , ,q 1
, I S ,q
S,q
D nS D D
D 2 ,I D
S,q
S,q S,q
D S,q D1
nD
S,q
D2
Figure 2: The first tree level T1 of the regular vine used in the vine copula risk model for sector S in time period q.
We then let the additional trees be chosen by the Dißmann Algorithm (Dißmann, 2013) as explained in Sec-
tion 3. Here, the most important dependencies are measured by absolute weights |w| and modelled first in the
maximum spanning tree. We choose the empirical Kendall’s τ between the copula data associated with an edge
as the weight measure w for the edge as it gives a bivariate monotone association and its values ranges from
[−1, 1] (Joe, 2014). Kendall’s τ is a ranked based dependence measure robust to outliers that fits the aim of our
analysis and can be defined in terms of copulas for two continuous random variables (X 1 , X 2 ) with copula C as
τ=4 C (u 1 , u 2 )dC (u 1 , u 2 ) − 1. An alternative overall weight measure could be Spearman’s ρ. Furthermore, we

R
[0,1]2
estimate a lower tail dependence coefficient λ. When Student’s t and Clayton (and BB1 and BB7) copulas are cho-
sen as best fit among the bivariate copula families given in Table 1, we can use this information to also estimate the
10

S,q S,q
C 2 ,IC
S,q S,q S,q S,q S,q S,q

A , IA C , IC S,q S,q D , ID
C 2 ,I
S,q S,q S,q
C 1 , IC
S,q
nA nC nD
A S ,q C S ,q S,
q
q
S,
S,q ; I
n S ,q n S ,q q ;I D
A ,I S , C ,I S ,
q ;I C S,
q q S, ,I
; I S ,q ; I S ,q S,
q
S,q
C
q ,I ,q
A C S,
C1 D nS D
S,q S,q S,q S,q S,q S,q S,q S,q S,q S,q S,q
A 2 , I S,q ; I A I A , I B ; I S,q I B , IC ; I S,q IC , I D ; I S D 2 , I S,q ; I D
A 2 ,I A I A ,I S,q I B ,I S,q IC ,I S,q I D ,I S,q D2 , ID
q
q S, D S,
S, B S ,q
q ;I B 1 q
q ;I A S, 1
,I S , ,I S ,
S, q ,I q q
S,
q ,I S, ,q ; I S ,q ; I S ,q
B nS B B D
S,q
A1
, I S,q ; I B
S,q
B2
A1 , I A B S,q , IB B1 , IB D1 , ID
nB
S,q S,q
B 2 ,I B
Figure 3: The second tree level T2 of the regular vine used in the vine copula risk model for sector S in time period q.
lower tail dependence coefficient λ which ranges from 0 to 1 and is defined for a bivariate distribution with copula
C (t , t ) 2
C as λl ower = limx→0+ P (X 2 ≤ F 2−1 (t ) | X 1 ≤ F 1−1 (t )) = limx→0+ . Other methods could include estimating a
t
specific distribution or a family of distributions; or working with a non-parametric model. We refer to Frahm et al.
(2005) for further reading on estimation methods and Czado et al. (2013) for additional weight measures. Fixing
the first two trees and using these dependence measures allow us to capture the (tail) dependence of each asset
with its belonging ESG class in order to understand joint dependence behaviour.
4.1. Dependence Indicators
First, we fit the R-vine tree structure, described in the previous section, with the specified first and second tree
3
level to each sector S and time period q. Using the fitted copula parameters and families, we can estimate the
Kendall’s τ and the implied lower tail dependence coefficient λ of an asset j with its ESG class k and that of an ESG
S,q S,q
class index I t ,k with its sector index I t .
S,q
We now introduce a new ESG class dependence indicator D j (τ), which we can compute for each asset j with
its ESG class k in period q within each sector S. It accounts for the conditional dependencies of the asset j to
S,q S,q
other assets o j , ..., o j given by the fitted vine copula model, where j l is the number of assets which occur in
1 l −1
2 The lower tail dependence coefficient is also non-zero when the fitted bivariate copula class is Student’s t, Clayton, 180° Joe, 180° Gumbel,
BB1, BB7, or 180° BB8 in our model. For other bivariate copula families, we have zero lower tail dependence coefficient.
3 We fit two different R-vine copula model classes. In the first, we allow only for the bivariate itau copulas, for which the estimation by
Kendall’s τ inversion is available (Student’s t, Frank, Gaussian, Clayton, Joe, Gumbel, and Independence copula). Using these, we also allow
for asymmetry in the upper and lower tails, as shown in Table 1. These copula families have a one-to-one relationship between their copula
parameter and Kendall’s τ. In the second model, we include additional copulas with two parameters. Including more bivariate copula families
such as BB1 (a combination between both extreme cases of Clayton copula and Gumbel copula), BB7 (combination of Joe copula and Clayton
copula), and the extreme-value copula BB8 (extended Joe) does not necessarily improve our model fits. To choose our optimal model we use
the Bayesian Information Criteria (BIC) as it tends to select a parsimonious model that reasonably approximates the density (Schwarz, 1978).
The selection of BB copula families is very rare. The information criteria are given for one sector in the Appendix C, all other sectors are
available from the authors upon request. Applying the Vuong Test by (Vuong, 1989) with the parsimonious BIC correction and considering
BIC values of the model fits indicate that itau bivariate copula family set is preferred in our data, except for the Real Estate sector in period
q 1 = 2006−2010 and q 2 = 2011−2015. Therefore, we allowed for BB copula fits in these two cases. Since the number of parameters differ in the
bivariate copula families that we considered, we used the parsimonious BIC correction in the Vuong test to favor more parsimonious models.
11

the conditioning set, when j is in the conditioned set with the fitted vine in sector S and period q. We use the
associated Kendall’s τ and its estimate τ̂ as a dependence measure. Our ESG class dependence indicator of asset
j in sector S and period q is defined by:
|τ̂ j ,I S,q |
S,q k
D j (τ) = , ∀S,q, j . (5)
|τ̂ j ,I S,q | + |τ̂ j ,I S,q |I S,q | + |τ̂ j ,o S,q |I S,q ,I S,q | + . . . + |τ̂ j ,o S,q |I S,q ,I S,q ,o S,q ,...,o S,q |
k S k j1 k S l k j1 j l −1
To exclude unrealistic assumptions, we assume that asset j is not independent of all other assets and include
all Kendall’s τ only in absolute value. Similarly, we can also define the following ESG class lower tail dependence
S,q
indicator D j (λ) for each asset j with its ESG class k within each sector S and period q. Here λ is denotes the
lower tail dependence coefficient, with estimate λ̂. Thus, we have
|λ̂ j ,I S,q |
S,q k
D j (λ) = , ∀S,q, j . (6)
|λ̂ j ,I S,q | + |λ̂ j ,I S,q |I S,q | + |λ̂ j ,o S,q |I S,q ,I S,q | + . . . + |λ̂ j ,o S,q |I S,q ,I S,q ,o S,q ,...,o S,q |
k S k j1 k S l k j1 j l −1
The values for both indicators are in the interval [0, 1]. Therefore, we obtain an indication on how strong the
(tail) dependence of the asset j in sector S of period q to its ESG class k, given its (conditional) dependence to its
S,q S,q
sector I S,q and to other assets o j , ..., o j . Fitting the vine copula models allows to capture these complex de-
1 l −1
pendency structures, especially, as they allow for asymmetric and tail dependence. To get a more comprehensive
understanding, instead of only looking at the dependence of an asset j with its ESG class index alone, the (tail)
dependence indicators are standardized as ratios as given in Equation (5) and Equation (6).
S,q
The ESG class dependence indicator D j (τ) as shown in Equation (5) gives an indication of how strong the
absolute dependence between an asset and its belonging ESG class in relation to all other assets in the same ESG
S,q
class is. The ESG class lower tail dependence indicator D j (λ) defined in Equation (6) quantifies the strength of
the dependence within the lower-left-quadrant tail of an asset’s return and its associated ESG class index in rela-
tion to all other assets within the fitted R-vine copula model. The design of the indicators allows to capture the
share of dependencies an asset has with other assets within the same ESG class and sector. This helps to under-
stand if there are common behaviors - especially for tail dependence. Common dependencies could indicate that a
time series (e.g., returns) exhibit comovement and, therefore, could share some risk properties. As many investors
tilt their portfolio towards comparable large ESG score, using an inclusion or exclusion approach, understanding
the dependence between similarly ESG rated assets is necessary in order to be able to capture comovements and
promote diversification.
12

5. The Data
We consider daily prices and yearly environmental, social, and governance (ESG) data of 334 US companies
j , constituents of S&P 500 index, for which the yearly ESG scores (E SG Sy, j ) from Refinitiv 4 , which have values
between 0 and 100, are available in the period from 3 January 2006 to 31 December 2018 (i.e. a total of t = 3271
trading days). This time period is chosen in order to have the largest possible time period for the largest number
of assets to get a comprehensive analysis of assets belonging to the S&P 500. The overall time frame is split into
three different time periods q: q 1 = 2006 − 2010, q 2 = 2011 − 2015, and q 3 = 2016 − 2018. The first two intervals are
made of 5 years of data each, where the first interval includes the 2008 financial crisis. The last interval consists
of 3 years of data and their ESG scores, according to Refinitiv are to be considered not definitive as they were
not yet five years old when downloading the data (in November 2020). This means that the ESG scores can be
modified from the provider post-publication. Furthermore, the S&P 500 market capitalization weights (M Sj ) from
S,q S,q
1 January 2015 are used in combination with the ESG classes (R j ) created by the mean ESG scores (E SG j ) of
each asset j and the Refinitiv ESG thresholds k, explained in the prior section, to compute ESG threshold indices
S,q S,q S,q S,q S,q
(I t ,k ), which correspond to the nodes Index A {I A }, Index B {I B }, Index C {IC }, and Index D {I D } for trading
day t in period q in Figure 2.5 These indices are computed for each of the 10 sectors S and 3 time periods q
S,q
individually. Additionally, for each industrial sector S the S&P 500 sector index (I t ) per period q is also included.
As the constituents of the ESG scores are weighted differently within each sector and it has been shown by Ashwin
Kumar et al. (2016) that ESG scores affect various sectors to a different degree, we estimate a R-vine model for
each industrial sector S and time period q separately (Eikon, 2020). In particular, we have 50 assets for Consumer
Cyclicals, 35 for Healthcare, 23 for Utilities, 51 for Technology, 19 for Basic Material, 21 for Real Estate, 49 for
Financials, 41 for Industrial, 16 for Energy and 31 for Consumer Non-Cyclicals. These assets are grouped given
S,q S,q
their mean ESG score (E SG j ) into four different ESG class groups (R j ) per time period q as shown in Table B.7
in Appendix B. Figure 4 shows the distribution of ESG scores for the different sectors during the different times
periods, from the oldest at the bottom to the most recent ones at the top. It is clear that overall ESG scores have
improved in time, possibly also due to the increasing pressure faced by companies to disclose information as well
as the implementation of new policies according to ESG criteria.
4 The ESG scores considered are the ones available from Refinitiv, the financial and risk business unit of Thomas Reuters. Refinitiv offers a
comprehensive ESG database and covers 70 percent of global market capitalization, across more than 450 different ESG metrics (Eikon, 2020).
These percentile-ranked ESG scores are designed to measure the company’s relative ESG performance, commitment, and effectiveness across
ten main topics, including emissions, environmental product innovation, and human rights, all of which are based publicly-reported data
(Eikon, 2020).
5 All notation are reported in the Appendix A.5.
13

Consumer Cyclicals Consumer Non−Cycicals Basic Materials
2018 2018 2018
2017 2017 2017
2016 2016 2016
2015 2015 2015
2014 2014 2014
2013 2013 2013
Year
Year
Year
2012 2012 2012
2011 2011 2011
2010 2010 2010
2009 2009 2009
2008 2008 2008
2007 2007 2007
2006 2006 2006
25 50 75 25 50 75 25 50 75
ESG Score ESG Score ESG Score
Technology Utilities Financials

2018 2018 2018
2017 2017 2017
2016 2016 2016
2015 2015 2015
2014 2014 2014
2013 2013 2013
Year
Year
Year
2012 2012 2012
2011 2011 2011
2010 2010 2010
2009 2009 2009
2008 2008 2008
2007 2007 2007
2006 2006 2006
25 50 75 25 50 75 25 50 75
Energy Industrials Real Estate

2018 2018 2018
2017 2017 2017
2016 2016 2016
2015 2015 2015
2014 2014 2014
2013 2013 2013
Year
Year
Year
2012 2012 2012
2011 2011 2011
2010 2010 2010
2009 2009 2009
2008 2008 2008
2007 2007 2007
2006 2006 2006
25 50 75 25 50 75 0 25 50 75
Healthcare
2018
2017
2016
2015
2014
2013
Year
2012
2011
2010
2009
2008
2007
2006
25 50 75
ESG Score
Figure 4: Development of ESG scores per sector S over the years
To get a preliminary understanding if ESG scores can provide additional information on the riskiness of a com-
pany, the empirical 95% Value at Risk (VaR), computed as the empirical quantile of the asset return distribution
based on daily data for each asset j in the period q grouped according to their mean ESG scores across all sectors
S is presented in Figure 5. Notice VaR values are smaller and exhibit larger variability during the time period 2006-
2010, due to the 2008 financial crisis, while estimates becomes larger and, therefore, less extreme in the other time
periods. It is also quite evident, especially for the time periods 2006-2010 and 2011-2016 that ESG scores seem to
be capable to provide information on the tail risk of an asset, as better rated companies tend to have also less neg-
ative VaR values. Remember also that ESG scores for 2016-2018 are yet not definitive. Differences tend to diminish
14

in the last two periods for classes A, B, and C as also companies tend to improve in their ESG scores, as shown in
Figure 4, and markets recover from the 2008 financial crisis. Still, ESG class D, which typically contain companies
which have not yet fully disclosed ESG information and are also lower rated than other companies exhibit worst
VaR levels also for the time interval 2016-2018.
95% Value at Risk
0.000
−0.025
Value at Risk 95%
−0.050
−0.075
−0.100
2006−2010 2011−2015 2016−2018

Interval
ESG Class A B C D
Figure 5: Empirical 95% Value at Risk classified by ESG class and time period
Summing up, empirical data suggest that ESG scores can provide information on tail riskiness of companies
and allow to group assets which might share similar tail-risk characteristics, at least in the first two time periods.
The difference diminishes in time but this could be due to the fact that companies tend to overall improve their
ratings reducing the number of assets in ESG class D and differences across ESG classes, except the ESG class D
which tend to become smaller, especially in period of lower market turmoil. Plots for 99% VaR and 95% and 99%
Expected Shortfall show similar behaviour.
6. Empirical Results
In this section, we report the main results utilizing the proposed R-vine copula ESG risk model. After esti-
mating the marginal distribution according to the IFM – see for details Appendix A.5, we check the normalized
15

contour and pair plots of the pseudo-copula data to identify deviations from a Gaussian dependence structure.
Then, two R-vine models for each sector S and time period q are fitted. Overall we fit 2 · 10 · 3 = 60 (2 different
R-vine model classes, 10 sectors S, and 3 time periods q) individual R-vine models following the structure shown
in Figures 2 and 3. The first R-vine model is fitted allowing only for the itau copula families and their rotations and
reflections, while in the second R-vine model we allow for all parametric copula families and their rotations and
reflections (all copula families are presented in Table 1). When possible (as described in Section 4.1) the simpler
model including only the itau copula families is used in the analysis, if not the R-vine model using all parametric
copula families is chosen. A decision table for each sector S and time period q can be found in Table C.10, Ap-
pendix C.2. We find that only for the first two time periods q 1 and q 2 of the Real Estate sector, the model including
all parametric copulas fits better.

S,q
The data fitted in each R-vine model includes the daily log-return data (Y t , j ,k ) for each sector S, the daily sector
S,q S,q
index (I t ), as well as the daily ESG class indices (I k ) in period q and ESG class k. In terms of complexity,
this means that depending on the sector S, we are dealing with a 20-dimensional data to 55-dimensional data
S,q
set. For example, the Energy sector S includes j = 16 assets and the mean ESG scores of these assets (E SG j ,
where S belongs to the Energy sector) belong to all four threshold ESG scores (k ∈ A, B,C , D), in the time period
S,q S,q S,q S,q
q 1 = 2006−2010, therefore, we have to compute four ESG class indices (I t ,A 1 , I t ,B 1 , I t ,C 1 , I t ,D 1 , where S is the Energy
sector) for each trading day t in period q 1 using the market capitalization weights (M Sj ) for the Energy sector S.
Furthermore, we also include the Energy sector index (I tS , where S belongs to the Energy sector). Then together
with the logarithmic returns of these assets (Y tS, j , where S belongs to the Energy sector) we create a 21-dimensional
dataset for this period q 1 to fit both R-vine models. Figures 6 and 7 show the tree level T1 and T2 of the itau based
R-vine model of the Energy sector. The same node notation as in Figures 2 and 3 are used.
For all sectors, we find that for the first tree T1 we have mostly Student’s t copulas fitted. This indicates the
presence of both, upper and lower tail dependence. Rarely the Gaussian and Frank copula families are chosen
which do not model tail dependence. The independence copula is also chosen infrequently in the first tree level
indicating the existence of dependence. In the second tree T2 the number of independence copulas, indicating a
decrease in model complexity, and Frank, Gaussian, and Clayton copulas increase, indicating less pronounced tail
dependence. Looking at the example of the Energy sector in Figures 6 and 7, we illustrate these relationships. The
exact number of itau copula families fitted for each sector S can be found in the Table 2. The number parametric
copula families fitted are available from the authors upon request.
16

C hevr on S,q H al l i bur t on S,q
Occi d ent al .pt l S,q C onocophi l l i ps S,q
t (0
0)
t (0.6
.80
t (0
)
)
.69 .74
) t (0
3)
M ar at hon.oi l S,q t (0.62) t (0.6 Hess S,q
S,q
IB
B aker.Hug es S,q
Schl umber g er S,q
N (−0.05)
t (0.66)
)
.81
t (0
Devon.Ener g y S,q
t (0.62)
t (1) N (−0.05) N (−0.05)
t (0.65)
S,q S,q
Exxon Mobil IA Energy Index (I S,q ) IC
t (0
.56 Apache S,q
)
N (−0.04)
t (0.62)
V al er o.Ener g y S,q
W i l l i ams S,q
9) S,q t (0.8
t (0.6 ID 1)
P i oneer S,q Eog .r es S,q
t (0.64)
N at i onal .Oi l .W el l S,q
Figure 6: The first tree level T1 of the fitted regular vine using only the itau copula families for the sector S = Energy in period q 1 = 2006 − 2010.
S,q S,q S,q S,q
I S,q denotes the Energy sector Index, while I A ,I B ,IC ,I D indicate the ESG class index for the Energy sector. A letter at an edge with a
number inside the parenthesis refers to its bivariate copula family with its estimated Kendall’s τ. As we allow for rotations and reflections
Kendall’s τ can also be negative. (See Table 1 for copula family abbreviations. )
S,q
C onocophi l l i ps S,q , I B
S,q
H al l i bur t on S,q , I B
S,q S,q
M ar at hon.oi l ,S,q I B Occi d ent al .pt l S,q , I B
S,q S,q
C hevr on S,q , I B Hess S,q , I B
I(0)
I(0)
2)
t (0
.0
−0
.02
C(
)
I(0 )
) .03 S,q
P i oneer,S,q , I D
J(0
)
I(0
t (0.67) t (0.69) t (0.66) F(0.04)

S,q S,q S,q S,q S,q
I A ,I S,q I B ,S,q IC ,I S,q I D ,I S,q Eog .r es S,q I D
G(−
0.03
0) t (0 )
1 ) .01
0. .02 )
J(− J(0
S,q S,q
Apache S,q ,IC N at i onalOi l .W el l ,S,q , I D
)
03
I(0)
.
S,q 1 S,q
−0
E xxon S,q , I A W i l l i ams S,q ,IC

C(
C(
I(0)
−0
.02
S,q
)
Devon.Ener g y S,q ,IC
S,q S,q
V al er o.Ener g y S,q ,IC B aker.Hug es S,q ,IC
S,q
Schl umber g er S,q ,IC
Figure 7: The second tree level T2 of the fitted regular vine using only the itau copula families for the sector S = Energy in period q 1 = 2006 −
S,q S,q S,q S,q
2010. I S,q denotes the Energy Sector index, while I A ,I B ,IC ,I D indicate the ESG class index for the Energy sector. A letter at an edge with
a number inside the parenthesis refers to its bivariate copula family with its estimated Kendall’s τ. As we allow for rotations and reflections
Kendall’s τ can also be negative. (See Table 1 for copula family abbreviations.)
17

Sector BM.a C. Cyc.b N.Cyc.c Ene.d Fin.e H.Care f Ind.g R. Est. h Tech.i Uti.j
2006-2010 T1 T1 T1 T1 T1 T1 T1 T1 T1 T1
Student’s t 18 44 30 16 50 33 44 20 50 23
Clayton 0 0 1 0 0 2 0 1 0 0
Frank 1 6 1 0 1 1 0 0 4 2
Gaussian 3 0 2 4 2 2 0 0 0 0
Gumbel 0 4 1 0 0 0 1 0 1 0
Independence 0 0 0 0 0 0 0 3 0 0
Joe 0 0 0 0 0 0 0 0 0 0
2011-2015 T1 T1 T1 T1 T1 T1 T1 T1 T1 T1
Student’s t 19 43 31 15 50 34 41 20 50 23
Clayton 1 0 0 0 0 2 0 0 0 2
Frank 0 7 0 1 0 0 2 0 3 0
Gaussian 0 0 4 1 2 1 2 0 2 0
Gumbel 0 4 0 0 0 0 0 0 0 0
Independence 2 0 0 3 0 0 0 1 0 0
Joe 0 0 0 0 0 1 0 3 0 0
2016-2018 T1 T1 T1 T1 T1 T1 T1 T1 T1 T1
Student’s t 18 41 29 15 50 33 41 18 50 23
Clayton 2 0 0 1 0 3 0 1 0 3
Frank 0 10 2 1 0 0 3 0 1 0
Gaussian 1 1 2 1 1 1 0 2 1 0
Gumbel 0 2 1 1 1 0 0 0 2 0
Independence 1 0 0 0 0 0 0 2 0 0
Joe 0 0 0 0 0 0 0 0 0 0
Table 2: Bivariate itau copula families and independence copula fitted for Tree 1 (T1 ). To simplify, no difference in the counts are made
based on rotations and reflections. These are available on request by the authors. (a Basic Materials, b Consumer Cyclicals, c Consumer
Non-Cyclicals, d Energy, e Financials, f Healthcare, g Industrials, h Real Estate, I Technology, J Utility)
6.1. Dependence Estimation

S,q S,q
We then compute both dependence indicators, D j (τ) and D j (λ) of asset j in sector S in period q belonging
to ESG class k using the optimal R-vine model.
S,q 1 X S,q
D̄ k (τ) = S,q
D j 0 (τ), ∀S,q,k , (7)
n k j 0 ∈[1,nS ]
S,q
j 0 :R 0 =k
j
18

S,q 1 X S,q
D̄ k (λ) = S,q
D j 0 (λ), ∀S,q,k , (8)
n k j 0 ∈[1,nS ]
S,q
j 0 :R 0 =k
j
S,q S,q
where n k corresponds to the number of assets in sector S, which are in ESG class k in period q, i.e., n k =
S,q
1, ∀S,q,k and R j 0 is the ESG class group based on the mean ESG score of an asset j 0 in sector S in period
P
j 0 ∈[1,n S ]
S,q
j 0 :R =k
j0
q.
S,q S,q
The mean values for the overall and lower tail dependence indicator D̄ k (τ) and D̄ k (λ) are reported in Table
3 and 4. Here the hyphen indicates that no dependence indicator can be computed due to missing values or we
chose not to compute the dependence indicator for groups with only one asset as they dominate the dependence
indicator. Boxplots for both dependence indicators are given in Figure D.8 and Figure D.9 in the Appendix D.
Time Interval 2006-2010 2011-2015 2016-2018
Sector | ESG class A B C D A B C D A B C D
Basic Materials - 0.30 0.29 - 0.25 0.35 0.27 - 0.23 0.28 0.25 -
Consumer Cyclicals 0.25 0.22 0.18 0.18 0.16 0.21 0.15 0.15 0.16 0.15 0.17 0.21
Consumer Non-Cyclicals 0.22 0.23 0.23 - 0.22 0.26 0.19 - 0.16 0.20 0.16 -
Energy - 0.29 0.25 0.30 0.24 0.31 0.30 - 0.23 0.28 0.33 -
Financials - 0.25 0.24 - 0.17 0.23 0.21 - 0.19 0.21 0.18 -
Healthcare 0.25 0.23 0.22 0.24 0.26 0.24 0.28 - 0.22 0.22 0.22 -
Industrials - 0.22 0.25 0.22 0.19 0.25 0.25 0.21 0.17 0.23 0.19 -
Real Estate - - 0.38 0.24 0.32 0.27 0.26 - 0.23 0.26 0.22 -
Technology 0.19 0.17 0.23 0.19 0.17 0.16 0.22 - 0.18 0.15 0.15 -
Utility - 0.32 0.30 - - 0.32 0.24 - 0.33 0.35 0.26 -
S,q
Table 3: D̄ k (τ) for each Sector and ESG Class using the itau copula families
S,q S,q
The ESG class dependence indicators, D j (τ) and D̄ k (τ), suggest a dependence of an assets with its ESG
class given its belonging sector. Looking at Table 3, we find that the majority of assets has a dependence between
0.2 and 0.4, and is thereby not negligible. Finding dependence means that assets that belong to the same ESG class
exhibit dependence among each other. We find that the largest ESG class dependence in q 1 = 2006 − 2010 for ESG
class A is found for Consumer Cyclicals as well as Healthcare assets, also we notice that many sector S in period
q 1 do not have enough assets in the ESG class A to compute a mean value. The largest ESG class dependence for
ESG class B assets is found in the Utility sector. While the largest dependence for ESG class C is found in the Real
Estate sector. The Energy sector shows to have the highest mean dependence in the ESG class D. When looking
19

at the second time interval q 2 after the 2008 financial crisis one can see that for ESG score class A, the Real Estate
assets, for ESG score class B, the Basic Material assets and for ESG class C, the show the Energy assets show the
highest ESG score class dependence. Now in ESG class D only two sectors have enough assets to compute the
mean dependence.
In the last time interval q 3 one can see that the overall dependence among similarly rated asset is decreased
compared to the earlier time periods . Furthermore, the Utility assets are found to have the highest ESG class
dependence for ESG score A and B. Here, the Energy assets show the highest dependence in ESG scores class C. In
the ESG class D, only one sector is left. This is the result of the improvement of the ESG scores within each sector
as shown in Figure 4.
Overall, we find overall dependence among assets which are similarly rated and across sector is present. Dur-
ing q 3 = 2016 − 2018, when ESG scores are not yet definitive, the dependence is found to be smaller. Such results
might change in the case that ESG scores are updated to take into consideration further information or disclosure
from the companies.
Time Interval 2006-2010 2011-2015 2016-2018
Basic Materials - 0.82 0.68 - 0.86 0.95 0.93 - 0.79 0.75 0.74 -
Consumer Cyclicals 0.84 0.63 0.51 0.48 0.72 0.66 0.55 0.60 0.39 0.33 0.36 0.72
Consumer Non-Cyclicals 0.66 0.54 0.60 - 0.80 0.79 0.75 - 0.51 0.70 0.53 -
Energy - 0.78 0.73 0.90 0.65 0.79 0.78 - 0.77 0.56 0.81 -
Financials - 0.69 0.64 - 0.64 0.76 0.67 - 0.60 0.61 0.54 -
Healthcare 0.86 0.64 0.61 0.65 0.80 0.68 0.81 - 0.51 0.54 0.48 -
Industrials - 0.69 0.69 0.69 0.61 0.76 0.81 0.83 0.49 0.64 0.56 -
Real Estate - - 0.85 0.80 0.94 0.78 0.74 - 0.40 0.69 0.67 -
Technology 0.75 0.55 0.61 0.60 0.58 0.59 0.63 - 0.58 0.46 0.57 -
Utility - 0.81 0.84 - - 0.89 0.85 - 0.78 0.83 0.72 -
S,q
Table 4: D̄ k (λ) for each Sector and ESG Class using the itau copula families
S,q S,q
Furthermore, the ESG class tail dependence indicator D j (λ) and its mean value D̄ k (λ) are computed as
stated in Equation (6) and Equation (8). Due to the design of the indicator and the number of copula families with
zero tail dependence coefficients, the magnitude of the indicator is close to 1. Still, it is possible to compare the
lower tail dependence across sectors S and ESG classes k.
As expected due to the design of the indicator, we find that the tail dependence is higher than the overall
dependence. Furthermore, for some sectors the tail dependence is explicitly lower than for other sectors. In the
20

first time interval for ESG class A and B, the dependence is lowest for the Consumer Non-Cyclicals. In the ESG class
C, again this is true for the Consumer Non-Cyclicals as well as the Consumer Cyclicals, Financials, Healthcare, and
Technology assets. Lastly, the lowest dependence is found in the Consumer Cyclicals assets for ESG class D.
In the second interval q 2 , the lowest dependence is found for the Technology sector in ESG class A and ESG
class B, and the the Consumer Cyclicals in ESG class C and D. Again not many sectors with enough assets are
available in ESG class D.
In the final interval, q 3 , we find the lowest tail dependence for Consumer Cyclicals assets for each ESG class.
Overall, the tendency seems to be a decrease in tail dependence in the last time interval. This is in line with the
findings of the overall dependence in Table 3. However, we find that some sectors including Utilities and the Basic
Materials maintain their relative high score throughout all three intervals. As the ESG scores in the last interval are
not yet definite, they might be adapted in the next years by the provider. Nevertheless, these ESG scores can still
be used by investors for their decision making and dependence analysis.
Overall, by introducing an indicator to capture overall dependence among assets with the similar ESG scores
and across sector, we are capable to quantify dependence . We see that such dependence is not negligible, with
values often between 0.2 and 0.4, which tend to increase during crisis. Also, by quantifying explicitly tail depen-
dence, we are able to provide a measure of it, and, therefore, show that tail dependence tends to be higher during
crisis. Still, as the overall ESG dependence vary between 0.2 and 0.4, the idiosyncratic component for each stock
as well as some other effects could still play a relevant role.
7. Conclusion
Starting from an empirical analysis on real-world financial data, we notice that that ESG rating classes can
provide information on (tail) dependence and tail risk. In fact, as Figure 5 shows, especially in the time periods
2006-2010 and 2011-2015, assets with better ESG rating seem to exhibit less tail risk. However, as companies tend
to improve their rating throughout time, these differences tend to diminish, especially in the interval 2016-2018,
in which, however, Refinitiv ESG scores cannot yet be considered definitive. Based on such observations, we pro-
pose the R-vine copula ESG risk model to capture dependence and tail dependence among assets belonging to
the same industrial sector based on their ESG class and among classes. Vine copula models are a flexible tool to
capture financial times series characteristics including non-symmetry and dependence in the extremes, as it can
rely on a pair copula construction. Furthermore, especially, the R-vine model is able to flexible model the depen-
dence structures using the ESG scores and the given ESG thresholds as it also allows for different and not only
recursive conditioning orders making the model less restrictive. Moreover, as for each level an arbitrary bivariate
21

copula can be specified, every complex dependence structure can be captured effectively and optimally. After
estimating the R-vine model, we can not only compute all the conditional dependencies among assets as well as
specify their interactions as modelled by different copulas families, but we can also introduce two new ESG class
dependence indicators that capture both overall dependence and also tail dependence. Furthermore, we are ca-
pable to quantify the dependence component that relies on belonging to a specific ESG class and industrial sector.
Such components are not negligible as they vary between 0.2 and 0.4, with some even larger values during 2008
financial crisis. We expand these findings further by quantifying also tail dependence among assets belonging to
the same ESG class and industrial sector. We notice that such tail dependence is also stronger during crisis.
The understanding and estimation of such dependence is of utmost importance for setting up adequate risk
management and mitigation tools as well as building portfolios, ideally also ESG diversified and resilient to crises.
Current popular ESG inclusion approaches that focus on picking only assets in the highest ESG rating classes
could have indeed possibly benefit in the past from better VaR values but such behavior is not clear for the most
recent interval, where ESG classes are overlapping. In fact, picking assets with the highest ESG scores does not
lead to better VaR values necessarily and could instead results in applying too much pressure on a specific set of
assets, without a clear benefit. The constant trend in improving ESG scores might be a factor behind the lack of
VaR differentiation between the classes A, B, and C in the last time interval, joint to the fact that such ESG scores
are yet not definitive. Still, we notice that ESG class D assets tend to exhibit poorer VaR values than other ESG
classes, suggesting that ESG disclosure might also have some indirect and positive effect on the company risk
management. High on the agenda, the current model could be used to develop new ESG investment strategies
and ESG based risk mitigation and management modelling tools.
22

Appendix A. Notation and Computation
Appendix A.1. Data
We introduce the mathematical indices, data sets and their notations used in the paper.
Type of Data Notation
Sector S = 1, . . . , 10
Trading day t = 1, ..., 3271
Year y = 2006, . . . , 2018
Period q = 1, 2, 3
Total number of assets in sector S n S in Section Appendix B, Table B.7
Asset j in sector S j = 1, . . . , n S
ESG class k ∈ {A, B,C , D}
ESG score of asset j in sector S in year y E SG Sy, j
Log return of asset j in sector S in period q on trading day t S,q

Y t , j ,k
belonging to ESG class k
S,q
S&P 500 log return for sector S on trading day t in period q It
Market capitalization weight of asset j in sector S (by 1.01.2015) M Sj

S,q
Mean ESG score of asset j in sector S and period q E SG j
S,q
ESG class of asset j in sector S and period q Rj
S,q
ESG class weight of asset j in sector S in period q αj
S,q
Values of ESG class k in sector S and period q at trading day t I t ,k
Table A.5: Mathematical Indices, datasets, and their notation used in the paper.
As mentioned in Section 5, we worked with the mean ESG score and the corresponding ESG class of an asset j
in sector S for period q. Accordingly, we calculated the assets’ ESG class weights, which were then used to calculate
their corresponding ESG class values, as defined in Table A.5. Their computations are given as follows.
S,q
Appendix A.2. Mean ESG score of asset j in sector S and period q (E SG j ) :
S,q 1 X
E SG j = E SG Sy, j for ∀ j ,S,q , (A.1)
|P q | y∈P q
where P 1 = [2006, 2010], P 2 = [2011, 2015], P 3 = [2016, 2018], and |P q | denotes the number of years in P q .
23

S,q
Appendix A.3. ESG class of asset j in sector S and period q (R j ) :
We have ∀ j ,S,q :

 S,q
A, if E SG j ∈ [75, 100],








 S,q
 B,
 if E SG j ∈ [50, 75),
S,q
Rj = (A.2)
 S,q
C, if E SG j ∈ [25, 50),









 D, otherwise.

S,q
ESG class weight of asset j in sector S in period q (α j ):
S,q
M Sj
αj = for ∀ j ,S,q . (A.3)
M Sj0
P
j 0 ∈[1,n S]
S,q S,q
j 0 :R 0 =R j
j
S,q
Appendix A.4. Values of ESG class k in sector S and period q at trading day t (I t ,k ):
S,q S,q S,q

α j 0 · Y t , j 0 ,k
X
I t ,k = for ∀S,q,k and t ∈ T q , (A.4)
j 0 ∈[1,n S ]
S,q
j 0 :R =k
j0
where T1 = [1, 1260], T2 = [1261, 2517], T3 [2518, 3271].
Appendix A.5. Two-Step Inference for Margins
As financial data are strongly dependent on past values and not uniformly distributed on [0, 1]d , which is the
necessary input for a copula, a two-step inference for margins (IFM) approach is followed. This approach as been
investigated by Joe (2005). We follow a parametric marginal model and estimate the margins first, we then use
the estimated marginal distributions to transform the data on the copula scale by defining the pseudo-copula
data. This allows us to remove the marginal time dependence by utilizing standard univariate time series models
and then proceed with standardized residuals obtained from these models. We fit a generalized autoregressive
conditional heteroskedasticity (GARCH) model with Student t innovations to our data, allowing for time-varying
volatility and volatility clustering.
24

Parameters Notation
The set of trading days in period q p q with p 0 = ;, p 1 = {1, . . . , 1260}, p 2 = {1261, . . . , 2517}, p 3 = {2518, . . . , 3271}
S&P 500 log returns for sector S in period q I S,q =(I 1+|p
S S
, . . . , I |p )> ∈ R|p q |
q−1 | q−1 |+|p q |
Matrix of log returns Y tS, j in sector S and period q for T ∈ t q Y S,q = [Y1S , . . . , YnSS ] ∈ R|p q |×nS , where Y jS0 = (Y1+|p
S
0 , . . . , Y|pS 0 )> , j 0 ∈ [1, n S ]
q−1 |, j q−1 |+|p q |, j
S,q S,q S,q S,q S,q S,q S,q

Data matrix for sector S in period q X = [Y ,I , I A , I B , IC , I D ] ∈ R|p q |×P
Columns of the data matrix X S,q d = 1, . . . , P

S,q S,q S,q
Column d of the data matrix X S,q Xd =(X 1,d , . . . , X |p )> ∈ R|p q |
q |,d
S,q S,q
Conditional variance vector of X t ,d on trading day t (σd ,t )2
S,q
Estimated degree of freedom for X t ,d ν̂Sd
S,q S,q
Estimated distribution function of the innovation distribution for time series X t ,d F̂ d ( · ; ν̂Sd )
S,q S,q
Estimated u-data for an observation X t ,d in sector S and period q û t ,d where t ∈ T q
Table A.6: Notation of the GARCH and Copula model.
As an input of a R-vine model in sector S and period q, we have a data matrix X S,q defined in Table A.6. Overall,
we have |S| × |q| = 10 × 3 = 30 vine copula risk models.
In sector S and period q, we fit a GARCH(1,1) models with appropriate error distribution for a marginal time
S,q
series, X d , and estimate the parameters of the following model:
S,q S,q S,q S,q S,q

εd ,t = σt ,d · z t (σd ,t )2 = γ0 + γ1 · (εt −1,d )2 + β1 · (σt −1,d )2 (A.5)
where (z t )t >1 is a sequence of normal random independent and identically distributed random variables sat-
isfying the standard assumptions E [z t ] = 0 and v ar [z t ] = 1 and follows a Student’s t distribution. Then using the
cumulative distribution function of the standardized Student’s t distribution, we determine the pseudo-copula
data as probability integral transformation (PIT), i.e.
S,q
 
S,q
X t ,d
û t ,d := F̂ d  ; ν̂Sd  . (A.6)
σ̂St,d
Following this two-step approach allows us to convert data to the copula scale, which can be used for estima-
tion of the copula parameter of the chosen bivariate copula family.
25

Appendix B. Number of Assets for each ESG class per Sector and interval
Time Interval 2006-2010 2011-2015 2016-2018
Basic Materials 1 7 11 0 3 11 5 0 5 10 4 0
Consumer Cyclicals 2 15 23 10 4 25 15 6 9 27 12 2
Consumer Non-Cyclicals 6 11 13 1 9 17 4 1 9 18 4 0
Energy 1 6 6 3 3 9 3 1 4 9 3 0
Financials 1 17 30 1 3 24 22 0 6 13 30 0
Healthcare 2 13 16 3 6 18 9 1 13 16 5 0
Industrials 1 15 20 5 3 22 14 2 7 24 10 0
Real Estate 1 1 14 4 2 7 10 1 6 10 4 0
Technology 6 16 21 8 7 25 18 1 10 32 9 0
Utilities 0 13 10 0 0 15 8 0 2 18 3 0
Table B.7: Number of Assets for each ESG class per sector S and time period q
Appendix C. Model Fit
Comparison of the itau and parametric R-vine models.
Appendix C.1. Energy Sector Model Fit
In the following the BIC are compared from three different vine models for the Energy sector and time period.
In every model the preselect feature of rvinecopulib is not activated. The output of all additional sectors and time
intervals are available from the authors upon request. See Table 1 for copula family abbreviations.
loglik BIC
Itau 2006-2010 32393.89 -63074.45
Par 2006-2010 32363.61 -62899.67
Itau 2011-2015 30135.86 -58637.47
Par 2011-2015 29643.34 -57495.43
Itau 2016-2018 11526.09 -21727.11
Par. 2016-2018 11005.10 -20671.87
Table C.8: Model Fit - Energy
26

Interval Statistic Schwarz Statistic Schwarz p-value
2006-2010 2.012 5.806 6.383e-09
2011-2015 4.470 5.182 2.194e-07
2016-2018 6.210 6.289 3.190e-10
Table C.9: Vuong Test - Energy
Appendix C.2. Decision Table Model Fit
Sector BM.a C. Cyc.b N.Cyc.c Ene.d Fin.e H.Care f Ind.g R. Est. h Tech.i Uti.j
2006-2010 Itau Itau* Itau* Itau* Itau* Itau* Itau* Par* Itau* Itau*
2011-2015 Itau* Itau* Itau* Itau* Itau* Itau* Itau* Par* Itau* Itau*
2016-2018 Itau Itau* Itau* Itau * Itau* Itau* Itau* Itau Itau* Itau
Table C.10: Model Decision based on parsimonious Schwarz correction and BIC (Schwarz, 1978). The copula families corresponding to each
itau and parametric model are given in Table 1. (a Basic Materials, b Consumer Cyclicals, c Consumer Non-Cyclicals, d Energy, e Financials, f
Healthcare, g Industrials, h Real Estate, I Technology, J Utility)

* Statistically Significant at α = 5%
Appendix D. Boxplots of Dependence Indicator
Similar to Tables 3 and 4, ESG classes which contain only one asset are not plotted below.
S,q
Appendix D.0.1. ESG class dependence indicator D j (τ)
ESG class dependence indicator

2006−2010
Utilities
Technology
Real Estate
Industry
Healthcare
Financials
Energy
Consumer Non−Cyclicals
Consumer Cyclicals
Basic Materials
2011−2015
Utilities
Technology
Different Sectors
Real Estate
Industry
Healthcare
Financials
Energy
Consumer Cyclicals
Basic Materials
2016−2018
Utilities
Technology
Real Estate
Industry
Healthcare
Financials
Energy
Consumer Cyclicals
Basic Materials
0.00 0.25 0.50 0.75 1.00
Dependence Measure
ESG Class A B C D
Figure D.8: ESG class dependence indicator for each time interval
27

S,q
Appendix D.0.2. ESG class lower tail dependence indicator D j (λ)
ESG class lower tail dependence indicator

2006−2010
Utilities
Technology
Real Estate
Industry
Healthcare
Financials
Energy
Consumer Cyclicals
Basic Materials
2011−2015
Utilities
Technology
Different Sectors
Real Estate
Industry
Healthcare
Financials
Energy
Consumer Cyclicals
Basic Materials
2016−2018
Utilities
Technology
Real Estate
Industry
Healthcare
Financials
Energy
Consumer Cyclicals
Basic Materials
0.00 0.25 0.50 0.75 1.00
Dependence Measure
ESG Class A B C D
Figure D.9: ESG class lower tail dependence indicator for each time interval
28

References
Aas, K., Czado, C., Frigessi, A., Bakken, H., 2009. Pair-copula constructions of multiple dependence. Insurance: Mathematics and Economics
44, 182–198. URL: http://dx.doi.org/10.1016/j.insmatheco.2007.02.001, doi:10.1016/j.insmatheco.2007.02.001.
Achim, M.V., Borlea, S.N., 2015. Developing of ESG score to assess the non-financial performances in Romanian companies. Procedia Eco-
nomics and Finance 32, 1209–1224.
Akaike, H., 1998. Information theory and an extension of the maximum likelihood principle, in: Selected papers of hirotugu akaike. Springer,
pp. 199–213.
Ane, T., Kharoubi, C., 2003. Dependence structure and risk measure. Journal of Business 76, 411–438.
Ashwin Kumar, N.C., Smith, C., Badis, L., Wang, N., Ambrosy, P., Tavares, R., 2016. ESG factors and risk-adjusted performance: a new quan-
titative model. Journal of Sustainable Finance & Investment 6, 292–300. URL: https://doi.org/10.1080/20430795.2016.1234909,
doi:10.1080/20430795.2016.1234909.
Auer, B.R., Schuhmacher, F., 2016. Do socially (ir) responsible investments pay? New evidence from international ESG data. The Quarterly
Review of Economics and Finance 59, 51–62.
Bae, K.H., El Ghoul, S., Guedhami, O., Kwok, C.C.Y., Zheng, Y., 2019. Does corporate social responsibility reduce the costs of high leverage?
Evidence from capital structure and product market interactions. Journal of Banking & Finance 100, 135–150.
Bedford, T., Cooke, R.M., 2001. Probability density decomposition for conditionally dependent random variables modeled by vines. Annals of
Mathematics and Artificial intelligence 32, 245–268.
Bedford, T., Cooke, R.M., 2002. Vines : A New Graphical Model for Dependent Random Variables 30, 1031–1068.
Berg, E., Lange, K.W., 2020. Enhancing ESG-Risk Modelling-A study of the dependence structure of sustainable investing.
doi:http://hdl.handle.net/2445/169669.
Bhattacharya, S., Sharma, D., 2019. Do environment, social and governance performance impact credit ratings: a study from India. Interna-
tional Journal of Ethics and Systems 35, 466–484.
Brechmann, E., 2010. Truncated and simplified regular vines and their applications .
Brechmann, E.C., Czado, C., 2013. Risk management with high-dimensional vine copulas: An analysis of the Euro Stoxx 50. Statistics and Risk
Modeling 30, 307–342. doi:10.1524/strm.2013.2002.
Breedt, A., Ciliberti, S., Gualdi, S., Seager, P., 2019. Is ESG an Equity Factor or Just an Investment Guide? The Journal of Investing 28, 32–42.
Breuer, W., Müller, T., Rosenbach, D., Salzmann, A., 2018. Corporate social responsibility, investor protection, and cost of equity: A cross-
country comparison. Journal of Banking & Finance 96, 34–55.
Capelle-Blancard, G., Crifo, P., Diaye, M.A., Oueghlissi, R., Scholtens, B., 2019. Sovereign bond yield spreads and sustainability: An empirical
analysis of OECD countries. Journal of Banking & Finance 98, 156–169.
Chan, P.T., Walter, T., 2014. Investment performance of “environmentally-friendly” firms and their initial public offers and seasoned equity
offers. Journal of Banking & Finance 44, 177–188.
Consolandi, C., Eccles, R.G., Gabbi, G., 2020. How material is a material issue? Stock returns and the financial relevance and financial intensity
of ESG materiality. Journal of Sustainable Finance & Investment , 1–24.
Cont, R., 2001. Empirical properties of asset returns: stylized facts and statistical issues. Journal of Quantitative Finance 1, 223–236.
Czado, C., 2019. Analyzing Dependent Data with Vine Copulas. Springer , 242.
Czado, C., Jeske, S., Hofmann, M., 2013. Selection strategies for regular vine copulae. Journal de la Société Française de Statistique 154, 174–191.
Dißmann, J., 2013. Statistical Inference for Regular Vines and Application. Master’s thesis. Technical University of Munich.
29

Dißmann, J., Brechmann, E.C., Czado, C., Kurowicka, D., 2013. Selecting and estimating regular vine copulae and application to financial
returns. Computational Statistics & Data Analysis 59, 52–69.
Dorey, M., Joubert, P., 2005. Modelling copulas: an overview. The Staple Inn Actuarial Society , 1–27.
Dorfleitner, G., Halbritter, G., Nguyen, M., 2016. The risk of social responsibility–is it systematic? Journal of Sustainable Finance & Investment
6, 1–14.
Dunn, J., Fitzgibbons, S., Pomorski, L., 2018. Assessing risk through environmental, social and governance exposures. Journal of Investment
Management 16, 4–17.
EBA, 2020. EBA Discussion paper on management and supervision of ESG risks for credit institutions and investment firms. Technical Report
October.
Eccles, R.G., Klimenko, S., 2019. The investor revolution. Harvard Business Review 97, 106–116.
Eikon, T.R., 2020. Thomson Reuters ESG Score.
Embrechts, P., Lindskog, F., McNeil, A., 2001. Modelling dependence with copulas. Rapport technique, Département de mathématiques,
Institut Fédéral de Technologie de Zurich, Zurich 14.
European Banking Authority, 2018. Sustainable Finance. URL: https://www.eba.europa.eu/financial-innovation-and-fintech/sustainable-finance
Frahm, G., Junker, M., Schmidt, R., 2005. Estimating the tail-dependence coefficient: properties and pitfalls. Insurance: Mathematics and
Economics 37, 80–100.
Friede, G., 2019. Why don’t we see more action? A metasynthesis of the investor impediments to integrate environmental, social, and gover-
nance factors. Business Strategy and the Environment 28, 1260–1282.
Giese, G., Lee, L.E., Melas, D., Nagy, Z., Nishikawa, L., 2019. Foundations of ESG investing: how ESG affects equity valuation, risk, and perfor-
mance. The Journal of Portfolio Management 45, 69–83.
Global Sustainable Investment Alliance (GSIA), 2018. 2018 Global Sustainable Investment Review. Technical Report. URL:
http://www.gsi-alliance.org/wp-content/uploads/2019/06/GSIR_Review2018F.pdf.
Goldreyer, E.F., Diltz, J.D., 1999. The performance of socially responsible mutual funds: Incorporating sociopolitical information in portfolio
selection. Managerial Finance .
Gormley, T.A., Matsa, D.A., 2011. Growing out of trouble? Corporate responses to liability risk. The Review of Financial Studies 24, 2781–2821.
Gormley, T.A., Matsa, D.A., Milbourn, T., 2013. CEO compensation and corporate risk: Evidence from a natural experiment. Journal of Ac-
counting and Economics 56, 79–101.
Haff, I.H., 2012. Comparison of estimators for pair-copula constructions. Journal of Multivariate Analysis 110, 91–105.
Henriksson, R., Livnat, J., Pfeifer, P., Stumpp, M., 2019. Integrating ESG in portfolio construction. Journal of Portfolio Management 45, 67–81.
Hoepner, A.G.F., Oikonomou, I., Sautner, Z., Starks, L.T., Zhou, X., 2016. {ESG} Shareholder Engagement and Downside Risk. {SSRN} Electronic
Journal doi:10.2139/ssrn.2874252.
Joe, H., 1996. Families of m-variate distributions with given margins and m (m-1)/2 bivariate dependence parameters. Lecture Notes-
Monograph Series , 120–141.
Joe, H., 2005. Asymptotic efficiency of the two-stage estimation method for copula-based models. Journal of Multivariate Analysis 94, 401–419.
doi:10.1016/j.jmva.2004.06.003.
Joe, H., 2014. Dependence modeling with copulas. CRC press.
Jondeau, E., Rockinger, M., 2003. Testing for differences in the tails of stock-market returns. Journal of Empirical Finance 10, 559–581.
Kelly, B., Jiang, H., 2014. Tail risk and asset prices. The Review of Financial Studies 27, 2841–2871.
Kim, Y., Li, H., Li, S., 2014. Corporate social responsibility and stock price crash risk. Journal of Banking & Finance 43, 1–13.
30

Kurowicka, D., Cooke, R.M., 2006. Uncertainty analysis with high dimensional dependence modelling. John Wiley & Sons.
Kurowicka, D. and Joe, H., 2011. Dependence modeling - handbook on vine copulae. World Scientific Publishing Co., Singapore.
Li, D.X., 2000. On default correlation: A copula function approach. The Journal of Fixed Income 9, 43–54.
Li, X., Wang, S.S., Wang, X., 2017. Trust and stock price crash risk: Evidence from China. Journal of Banking & Finance 76, 74–91.
Lioui, A., 2018. Is ESG Risk Priced? SSRN Electronic Journal doi:http://dx.doi.org/10.2139/ssrn.3285091.
Maiti, M., 2020. Is ESG the succeeding risk factor? Journal of Sustainable Finance & Investment , 1–15.
Malevergne, Y., Pisarenko, V., Sornette, D., 2005. Empirical distributions of stock returns: between the stretched exponential and the power
law? Quantitative Finance 5, 379–401.
Minor, D.B., 2011. Corporate citizenship as insurance: Theory and evidence. University of California, Berkley .
Nofsinger, J., Varma, A., 2014. Socially responsible funds and market crises. Journal of Banking & Finance 48, 180–193.
Pollard, J.L., Sherwood, M.W., Klobus, R.G., 2018. Establishing ESG as risk premia. Journal of Investment Management 16, 32–43.
Puccetti, G., Scherer, M., 2018. Copulas, credit portfolios, and the broken heart syndrome , 114–130doi:10.1515/demo-2018-0007.
Refinitiv, 2019. Environmental, Social and Governance (ESG) Scores from Refinitiv. ESG Scores Methodology , 1–18URL:
https://www.refinitiv.com/content/dam/marketing/en_us/documents/methodology/esg-scores-methodology.pdf.
Salmon, F., 2012. The formula that killed Wall Street. Significance 9, 16–20.
Schwarz, G., 1978. Estimating the dimension of a model. The Annals of Statistics 6, 461–464.
Shafer, M., Szado, E., 2018. Environmental, social, and governance practices and perceived tail risk. Accounting & Finance
doi:10.1111/acfi.12541.
Sherwood, M.W., Pollard, J.L., 2017. The risk-adjusted return potential of integrating ESG strategies into emerging market eq-
uities. Journal of Sustainable Finance & Investment 8, 26–44. URL: https://doi.org/10.1080/20430795.2017.1331118,
doi:10.1080/20430795.2017.1331118.
Shirvani, A., Volchenkov, D., 2019. A regulated market under sanctions: On tail dependence between oil, gold, and Tehran stock exchange
index doi:arXiv:1911.01826.
Sklar, M., 1959. Fonctions de repartition an dimensions et leurs marges. Publications de l’Institut de Statistique de l’Universite de Paris 8,
229–231.
Stoeber, J., Joe, H., Czado, C., 2013. Simplified pair copula constructions—limitations and extensions. Journal of Multivariate Analysis 119,
101–118.
Vuong, Q.H., 1989. Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica: Journal of the Econometric Society ,
307–333.
Wamba, L.D., Sahut, J.M., Braune, E., Teulon, F., Others, 2020. Does the optimization of a company’s environmental performance reduce
its systematic risk? New evidence from European listed companies. Corporate Social Responsibility and Environmental Management 27,
1677–1694.
Xu, Q., Li, X.M., 2009. Estimation of dynamic asymmetric tail dependences: an empirical study on Asian developed futures markets. Applied
Financial Economics 19, 273–290.
31

ESG, Risk, and (Tail) Dependence: A B B A

Uploaded by

Copyright:

Available Formats

ESG, Risk, and (Tail) Dependence: A B B A

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ESG, Risk, and (Tail) Dependence: A B B A

Uploaded by

Copyright:

Available Formats

ESG, Risk, and (tail) dependence

Karoline Baxa , Özge Sahinb , Claudia Czadob , Sandra Paterlinia,∗

JEL classification: G32, C51, C58

∗ Corresponding author: Email address: [email protected]

Preprint submitted to SSRN May 15, 2021

Electronic copy available at: https://ssrn.com/abstract=3846739

et al., 2005; Berg and Lange, 2020).

Electronic copy available at: https://ssrn.com/abstract=3846739

look for future research.

2. ESG Scores, Dependence, and Risk

necessary in order to access inherent risks in the financial market.

Electronic copy available at: https://ssrn.com/abstract=3846739

Electronic copy available at: https://ssrn.com/abstract=3846739

significantly reduces the perceived one-month ahead tail risk.

non-diversifiable risk component as it depends on the overall market.

3. Dependence Modelling and Vine Copulas

terns, such as non-symmetry and dependence in the extremes.

Electronic copy available at: https://ssrn.com/abstract=3846739

bution function F can be written as

The corresponding density is

the associated univariate marginal density of F p (x p ) for p = 1, . . . , d .

Itau1 Copulas BB Copulas

Electronic copy available at: https://ssrn.com/abstract=3846739

density c under the simplifying assumption can be constructed as follows:

where x D e = (x z )z∈D e is a subvector of x = (x 1 , . . . , x d )> ∈ Rd , F e a |D e is the conditional distribution function of the

arguments. For more details, we refer to Stoeber et al. (2013).

Electronic copy available at: https://ssrn.com/abstract=3846739

4 5 2,4 2,5 1,4;2 1,5;2 3,5;1,2 4,5;1,2,3

Figure 1: Example of a 6-dimensional regular vine with edge labels.

Electronic copy available at: https://ssrn.com/abstract=3846739

4. The R-Vine Copula ESG Risk Model

Electronic copy available at: https://ssrn.com/abstract=3846739

S,q S,q S,q S,q S,q S,q

τ=4 C (u 1 , u 2 )dC (u 1 , u 2 ) − 1. An alternative overall weight measure could be Spearman’s ρ. Furthermore, we

Electronic copy available at: https://ssrn.com/abstract=3846739

S,q S,q S,q S,q S,q S,q

4.1. Dependence Indicators

Electronic copy available at: https://ssrn.com/abstract=3846739

j in sector S and period q is defined by:

lower tail dependence coefficient, with estimate λ̂. Thus, we have

Electronic copy available at: https://ssrn.com/abstract=3846739

as the implementation of new policies according to ESG criteria.

Electronic copy available at: https://ssrn.com/abstract=3846739

Technology Utilities Financials

Energy Industrials Real Estate

Figure 4: Development of ESG scores per sector S over the years

Electronic copy available at: https://ssrn.com/abstract=3846739

VaR levels also for the time interval 2016-2018.

95% Value at Risk

2006−2010 2011−2015 2016−2018

Expected Shortfall show similar behaviour.

Electronic copy available at: https://ssrn.com/abstract=3846739

all parametric copulas fits better.

Electronic copy available at: https://ssrn.com/abstract=3846739

Occi d ent al .pt l S,q C onocophi l l i ps S,q

Schl umber g er S,q

P i oneer S,q Eog .r es S,q

N at i onal .Oi l .W el l S,q

t (0.67) t (0.69) t (0.66) F(0.04)