Economic Modelling 62 (2017) 194–206

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.elsevier.com/locate/econmod

Can energy commodity futures add to the value of carbon assets? MARK
a,b,⁎ c d
Xiaoqian Wen , Elie Bouri , David Roubaud
a
Institute of Chinese Financial Studies, Southwestern University of Finance and Economics, China
b
Collaborative Innovation Center of Financial Security, Southwestern University of Finance and, Economics, China
c
USEK Business School, Holy Spirit University of Kaslik (USEK), Lebanon
d
Montpellier Business School, Montpellier, France

A R T I C L E I N F O A BS T RAC T

JEL classification: This paper examines whether energy commodity futures are an attractive asset class for helping investors
C51 manage carbon risk. We use futures prices for EU allowances (EUAs) and four energy-related commodities
G11 (crude oil, coal, natural gas, and electricity) in Phase II and about half of Phase III of the European Union
Q43 Emissions Trading Scheme. Both static and generalized autoregressive score dynamic copulas are used to model
Keywords: the dependence between the EUAs and the four energy commodity futures prices, with an emphasis on the
Carbon asset performance of two different portfolio strategies (diversified portfolios and hedged portfolios) and the resulting
Energy commodity futures effect on risk mitigation in the carbon market. Our empirical results show that despite the superiority of the
Portfolio management
hedged portfolios in increasing the risk-adjusted returns of carbon assets, the dynamic diversified portfolios are
Copulas
much preferred for reducing variance and the downside risks of carbon assets. Of the four energy commodity
futures examined, coal (electricity) is found to be the most (least) attractive in terms of mitigating the carbon
risk. These results are confirmed in both sub-sample and out-of-sample analyses.

1. Introduction period experienced over-allocation, leading to a fall in the price of

The European Union Emissions Trading Scheme (EU ETS) is the
largest greenhouse gas (GHG) emissions trading system in operation
and is the world's foremost tool for managing the reduction of carbon
dioxide (CO2), the main GHG (Bing et al., 2015). Launched on January
1, 2005, the EU ETS is a cap-and-trade system in which mandated
companies are allocated permits to emit CO2. These carbon permits,
called EU allowances (EUAs), are traded on financial markets. EUAs,
which have emerged as a factor of production, give their holders the
right to emit one ton of CO2 (or its equivalent). Small energy users (i.e.,
cleaner companies) whose emissions are below their allowable limits
(caps) can sell their unused EUAs to large energy users (i.e., polluting
companies) whose CO2 emissions exceed their caps. As of December
2015, the EU ETS covers more than 12,000 heavy energy-using
installations in power and industry plants across the 28 EU countries
and Iceland, Liechtenstein, Norway, and Croatia. To date, the EU ETS
has had three trading periods, known as operational phases: Phase I
(January 1, 2005 to December 31, 2007), Phase II (January 1, 2008 to
December 31, 2012), and Phase III (January 1, 2013 to 31 December
2020). Each phase has its own specific compliance requirements.1
Phase I was used to test price formation in the carbon market; this pilot
allowances. Phase II overlapped with the first commitment period of
the Kyoto Protocol, in which the involved countries had committed to
meet their emissions reduction targets. Despite the inclusion of the
aviation industry into the scheme on January 1, 2012, Phase II
featured a surplus of unused allowances that weighed on the carbon
price. Furthermore, the economic slowdown coupled with the global
financial crisis led to a cut in both emissions and demand. Phase III
featured a major reform with the introduction of an EU-wide cap on
emissions (reduced by 1.74% each year) and the progressive replace-
ment of the free allocation of allowances by auctioning.
Now in operation for over 10 years, the EU ETS has gained great
popularity among traditional and alternative asset managers and has
become an important investment vehicle for managing asset risks
(Subramaniam et al., 2015). With more worldwide climate regulations
in force to encourage sustainable development, carbon has become a
valuable asset for heavy energy-using plants and industries. Managing
the carbon asset price is essential for a large number of those plants
and industries. In addition to hedging the upside risk of the carbon
price, which potentially shifts more costs to companies demanding
carbon allowances, hedging the downside risk of the carbon price is
becoming more urgent. As noted in a recent news article by Thomson

⁎
Corresponding author at: Institute of Chinese Financial Studies, Southwestern University of Finance and Economics, China.
E-mail addresses: wxqkou@sina.com (X. Wen), eliebouri@usek.edu.lb (E. Bouri), d.roubaud@Montpellier-BS.com (D. Roubaud).
1
Detailed explanations of the EU ETS phases are available at http://ec.europa.eu/clima/policies/ets/index_en.htm.

http://dx.doi.org/10.1016/j.econmod.2016.12.022
Received 6 August 2016; Received in revised form 12 November 2016; Accepted 21 December 2016
Available online 06 January 2017
0264-9993/ © 2016 Elsevier B.V. All rights reserved.
X. Wen et al.
Reuters,2 carbon prices have fallen dramatically in recent years, partly
because of a worsening global economic outlook and declining levels of
GHG emissions, which has led to a slowdown in the demand for carbon
credits. Furthermore, the oversupply of carbon allowances by the UN
climate panel has led to a surplus of emission permits, resulting in a
large imbalance between the demand for and supply of carbon credits
(Balcilar et al., 2016).
The fact that EUAs are a factor of production suggests that changes
in EUA prices are closely related to the dynamics of other energy
commodity markets. For example, Reboredo (2014) points out that the
state of macroeconomic indicators and financial markets affects carbon
and energy prices, leading to an interaction between them; Zhang and
Sun (2016) show that sharp changes in carbon prices in recent years
have been closely correlated with energy prices. However, given that
decreases (increases) in energy prices can increase (reduce) companies’
energy use, which may increase (mitigate) demands for carbon
allowances, a low positive or negative relationship between the energy
price and the carbon asset price is also expected. Accordingly, the
diversification potential of energy futures for carbon assets is worthy of
attention. Therefore, we study the dependence between energy com-
modity futures and carbon prices with emphasis on portfolio diversi-
fication and hedging potential for the carbon asset.
Most of the studies that have addressed the market linkages
between carbon and energy markets have focused on the time-varying
correlations on average and on the dynamic volatility spillovers.
Although the work of Marimoutou and Soury (2015) is an important
step in modeling the dependence between CO2 emission spot prices
and a set of commodity prices (Brent crude oil, natural gas, coal and S
& P energy index) for extreme cases using copulas, it overlooks the
effects on portfolio diversification and hedging strategies and lacks in-
depth sub-sample and out-of-sample analyses. Reboredo (2013a) is
another interesting work examining EUA and crude oil market
dependence using copulas; however, this work is restricted to the
dependence between the price of carbon and the price of one type of
fossil energy, crude oil. As for the risk management of the carbon asset,
prior studies mostly focus on the effectiveness of carbon derivative
markets (see, among others, Daskalakis et al., 2009; Narayan et al.,
2015; Balcilar et al., 2016; Philip and Shi, 2016; Xu et al., 2016).
This paper contributes to the literature in the following three
dimensions. First, unlike previous studies, which generally focused
on the effectiveness of carbon derivative markets in managing the
carbon risk, this paper seeks more diverse hedging tools for the carbon
asset. Specifically, it emphasizes hedging the risks of (extreme)
decreases in the carbon price. This timely and urgent issue has been
rarely explored thus far, and this paper provides a detailed analysis of
the extent to which various energy commodity futures can offer
protection for the carbon asset. Second, copula functions, which are
capable of capturing both the average and tail dependence between
markets and of establishing a more effective multivariate distribution
of asset returns, are used to model the relationship between carbon and
other energy commodity futures markets. Specifically, in addition to
the static copulas, the generalized autoregressive score (GAS) specifica-
tion of Creal et al. (2013) is applied to characterize the dynamics of
copula parameters; such modeling, which is still relatively novel in the
field of energy economics, is more sensitive to correlation shocks,
which enables us to better capture the dependency structure. On the
basis of the copula information, various portfolios are constructed;
then, our portfolio management has the advantage of flexibly and fully
characterizing the market dependence structure. Third, all of the
analyses are conducted in different market phases/environments
(sub-samples) and out-of-sample as well, thus making the empirical
2
“Carbon offsets near record low, worst performing commodity,” Reuters (August 5,
2011).http://www.reuters.com/article/2011/08/05/us-carbon-low-
idUSTRE77442920110805.
195
Economic Modelling 62 (2017) 194–206
evaluations more comprehensive and robust.
We provide evidence of symmetric tail dependence between the
returns of carbon assets and energy commodity futures and find that
despite the superiority of the hedged portfolios in increasing the risk-
adjusted returns of carbon assets, the dynamic diversified portfolios are
much preferred for reducing both variance and the downside risks of
carbon assets. Of the four energy commodity futures examined in this
paper, coal futures (electricity futures) are found to be the most (least)
attractive in terms of mitigating the carbon (extreme) risk; these results
are confirmed in both sub-sample and out-of-sample analyses.
The remainder of the paper is structured as follows. Section 2
reviews the related work. Sections 3 and 4 present the empirical model
and the data, respectively. Section 5 presents the results. Section 6
concludes the paper.
2. Literature review
Since the launch of the EU ETS in 2005, the body of research
literature on carbon markets has grown rapidly. The first strand of
research focuses on the efficiency of the European carbon market
(Daskalakis and Markellos, 2008; Krishnamurti and Hoque, 2011;
Philip and Shi, 2016), its price discovery (Rittler, 2012; Benz and
Hengelbrock, 2009; Schultz and Swieringa, 2014; Narayan et al.,
2015), and its price dynamics (Benz and Trück, 2009; Daskalakis
et al., 2009; Conrad et al., 2012; Zhu et al., 2014).
A second strand of research has emphasized the factors affecting
European carbon prices. Several studies have provided evidence that
carbon allowance prices are affected by weather conditions (Mansanet-
Bataller et al., 2007; Alberola et al., 2008; Bredin and Muckley, 2011;
Liu and Chen, 2013). Various economic activity indicators and financial
factors have also been found to be empirically linked with carbon
prices. Chevallier (2009a) examines the macroeconomic determinants
of EUA prices and reports a weak association between carbon prices
and stock and bond variables. Bredin and Muckley (2011) find
significant relationships between carbon and stock prices and the level
of industrial production. Furthermore, Chevallier (2011) argues that
exogenous recessionary shocks on the economy have an adverse effect
on carbon prices. Reboredo (2014) finds that oil prices, which are
closely linked to economic activity indicators and financial variables,
may transmit financial uncertainties to carbon prices. Other studies
have provided evidence on the linkages between European carbon
prices and electricity stock returns (Oberndorfer, 2009; Veith et al.,
2009; Tian et al., 2016), arguing that emission prices introduce
additional costs to power generators, bringing about more volatility
in their cash flows.
A third strand of research has concentrated on the linkages between
European carbon prices and electricity prices (Sijm et al., 2006; Lu
et al., 2012; Aatola et al., 2013; de Menezes et al., 2016) based on the
rationale that the electricity sector is affected by carbon prices because
of its large contribution to total EU CO2 emissions. In addition,
substitutions between fuel oil, coal, and natural gas affect the price of
carbon credits because of the difference in CO2 emissions across these
fossil energies. Motivated by this strand of research, a fourth strand
argues that energy prices are important drivers of carbon prices
(Convery et al., 2007; Mansanet-Bataller et al., 2007; Hammoudeh
et al., 2014; Hammoudeh et al., 2015). Power plants can reduce their
CO2 emissions and thereby the cost of producing one unit of electricity
by switching from high-carbon-density fuels such as oil and coal to
lower-carbon-density fuels such as natural gas. This leads the carbon
and energy markets to interconnect. Accordingly, Bunn and Fezzi
(2007) indicate that carbon prices are highly responsive to gas prices.
Fezzi and Bunn (2009) show that electricity prices are driven by carbon
and natural gas prices. Chevallier (2009b) focuses on fuel-switching
behavior and shows that the relative prices of coal and gas are drivers
of carbon prices.
Numerous studies have examined the linkages between carbon and
X. Wen et al. Economic Modelling 62 (2017) 194–206

energy commodity prices using multivariate GARCH models. For and fEUA,(REUA,t) and fenergy(Renergy,t) are, respectively, the
example, Chevallier (2012) tests the ability of multivariate GARCH marginal densities of REUA,t and Renergy,t. Therefore, the log-like-
models such as BEKK, constant conditional correlation (CCC), and lihood function is
DCC to capture the dynamics of the correlations between the oil, gas,
log g = log c + log fEUA + log fenergy . (3)
and CO2 variables. Koch (2014) uses a multivariate GARCH model and
shows that the conditional correlation between the carbon and energy We denote the parameters in c, fEUA, and fenergy as θc, θEUA, and
markets depends on market uncertainty conditions. Using similar θenergy, respectively, and rewrite Eq. (3) as
models, Zhang and Sun (2016) find some significant volatility spil-
L (θ ) = Lc (θc ) + LEUA (θEUA) + Lenergy (θenergy ), (4)
lovers and correlations between carbon prices and three energy prices
(coal, natural gas, and Brent crude oil). Balcilar et al. (2016) use a where Lkis the log-likelihood function of the copula (k=c), EUA
Markov regime-switching DCC GARCH model and focus on the (k=EUA), and energy commodity futures (k=energy) densities.
volatility linkages across energy and carbon prices and on hedging Based on the two-stage estimation procedure proposed by Joe
strategies for carbon risk. (1997) known as inference for the margins (IFM), the parameters of
The multivariate GARCH models are advantageous in capturing the each univariate model are first obtained via maximum likelihood; then,
spillovers/interactions between markets; however, they overlook the the marginal CDFs are applied to the standardized residuals
tail dependence and have the restricted assumption of multi-normality uˆ t = FEUA (zEUA, t ; θˆEUA), and vˆt = Fenergy (zenergy, t ; θˆenergy ) for t=1 to T is
for asset returns. To overcome these limitations, copula functions have computed. The parameters of the copulas are obtained by solving
become attractive tools for modeling market dependence in recent T
years; these functions are known to be advantageous in building θˆc = arg max
θc
∑ ln c (uˆt , vˆt ; θc).
effective multivariate distributions of asset returns and in capturing t =1 (5)
average dependence and a rich pattern of tail dependence. Using
copulas, Reboredo (2013a) examines the dependence between the EUA
price and the crude oil price; Marimoutou and Soury (2015) model the
tail dependence between CO2 emission spot prices and a set of 3.2. Marginal distribution models
commodity prices (Brent crude oil, natural gas, coal, and the S & P
energy index). These two studies provide a solid foundation for our It is well known that financial asset returns (including commod-
paper. However, as noted in the introduction, by using copulas, we ities) have some stylized features such as fat tails, serial correlation,
further contribute to the debate on dependences across energy futures leverage effects, and conditional heteroscedasticity. To capture these
(crude oil, coal, natural gas and electricity), and carbon futures prices; stylized features of EUA and other energy commodity futures returns,
moreover, we emphasize the hedging and portfolio diversification autoregressive moving average (ARMA) models with generalized auto-
potential of energy commodity futures for carbon assets with different regressive conditional heteroscedasticity (GARCH) errors for returns
sample compositions. (Rt) are used.
The conditional mean follows an ARMA (p, q) process given by
p q
3. Methodology
Ri, t = μi + ∑ φi,j Ri,t −j − ∑ θi,j εi,t −j + εi,t ,
j =1 j =1 (6)
3.1. Copula functions
where p, q are non-negative integers that can be selected by the
When developing portfolio management strategies, the average Schwarz information criterion (SIC), which is known to lead to a
movements and the joint extreme movements across variables are parsimonious specification, and φi,j, θi,j are, respectively, the AR and
important (Reboredo, 2013b). Given the flexibility and effectiveness of MA parameters (i=EUA, oil, coal, natural gas and electricity). εi,t
copula functions in characterizing these movement patterns, we use follows a skewed-t distribution:
copulas to model the joint distribution of carbon asset returns and εi, t = σi, t zi, t , (7)
other energy commodity futures returns and obtain valuable informa-
tion on the average dependence and tail dependence. zi, t ~skewed − t (zi |ηi , ϕi ). (8)
By Sklar's theorem, a two-dimensional joint distribution function G The density function of the skewed-t distribution of Hansen (1994)
with continuous marginals FX and FY has a unique copula representa- is
tion such that G(x, y)=C(FX (x), FY(y)). A joint distribution function
⎧ bc (1 +
⎪
1 bz + a 2 −η +1/2 a
can then be decomposed into marginal distributions and the depen- ( )) , z<−
skewed − t (z|η, ϕ) = ⎨
η−2 1−ϕ b
dence structure described by a copula. In this paper, REUA,t and .
⎪ bc (1 + 1 bz + a 2 −η +1/2 a
Renergy,t (energy=oil, coal, natural gas, electricity) are EUA and ⎩ (
η−2 1+ϕ
)) , z ≥ −b
(9)
energy commodity futures returns, respectively, at time t.
The values of a, b, and c are defined as
Their conditional cumulative distribution functions (CDFs) are den-
oted by FEUA(REUA,t) and Fenergy(Renergy,t), respectively. η−2 Γ (η + 1/2)
a ≡ 4ϕc , b ≡ 1 + 3ϕ 2 − a 2 , c ≡ ,
u=FEUA(REUA,t) and v=Fenergy(Renergy,t) are distributed as con- η−1 π (η − 2) Γ (η /2) (10)
tinuous uniform variables on (0, 1), and the copula function C (u, v) is
defined by the joint CDF of the returns: where η is the kurtosis parameter and ϕ is the asymmetry parameter.
These are restricted to 2 < η < ∞ and –1 < ϕ < 1. Like the Student's t-
G (REUA, t , Renergy, t ) = C (FEUA (REUA, t ), Fenergy (Renergy, t )). (1) distribution, it is well defined only for η > 2; the skewness exists for η >
By differentiating all of the conditional CDFs, the conditional joint 3 and the kurtosis exists only if η > 4. In the case of ϕ > 0 (ϕ < 0), the
density is then given by mode of the density is to the left (right) of 0 and the variable is skewed
to the right (left).
∂G (REUA, t , Renergy, t ) In Eq. (7), σi,t is the square root of the conditional variance. Next,
g (REUA, t , Renergy, t ) = = c (FEUA (REUA, t ), Fenergy (Renergy, t ))
∂REUA, t ∂Renergy, t we consider two of the most popular GARCH models for modeling the
× fEUA (REUA, t ) × fenergy (Renergy, t ), conditional variance of asset returns: a standard GARCH and an
(2)
asymmetric GARCH. Following Brooks (2002), a GARCH family model
where c (u, v ) = ∂ 2C (u, v )/∂u∂v is a conditional copula density function with 1 lag order can sufficiently capture the volatility clustering in asset

196
X. Wen et al. Economic Modelling 62 (2017) 194–206

returns. Few financial studies have used higher-order models; there- by Patton (2006), the GAS specification is more sensitive to depen-
fore, the order of all of the GARCH processes are specified as 1 (see, dence shocks, and it thus better captures the variability in the market
among others, Guo et al., 2016; Wen et al., 2012). dependence (Avdulaj and Barunik, 2015).3 Consider a copula with
The standard GARCH (1, 1) model for the returns on asset i is dynamic parameters C(δt(γ)). To keep the parameters within a
particular range, the GAS specification applies a strictly increasing
σi2, t = ωi + αi εi2, t −1 + βi σi2, t −1, (11) transformation (e.g., log, logistic, arc tan) to the copula parameter and
where αi represents the ARCH term that measures the effect of past models the evolution of the transformed parameter, which is denoted
innovations on current variance and βi represents the GARCH term by ft:
that measures the effect of past variance on current variance. The ft = h (δt ) ⇔ δt = h−1 ( ft ), (15)
degree of persistence of the variance shock is measured by the sum of
the ARCH and GARCH parameters (αi+βi). where ft +1 = c + βGAS ft + αGAS It−1/2 st , (16)
Considering the asymmetry in the conditional variance process, the
GJR-GARCH (1, 1) model for returns on asset i is given by ∂
st = log(ut , vt ; δt ),
∂δ (17)
σi2, t = ωi + αi εi2, t −1 + βi σi2, t −1 + γi εi2, t −1 Ii, t −1. (12)
It = Et −1 [st st′] = I (δt ). (18)
Compared to Eq. (11), the additional variable Ii,t−1 in Eq. (12) is a
Then, the future value of the copula parameter is a function of a
dummy variable that measures the asymmetric response of the
constant, the current value, and the score of the copula-likelihood, It−1/
conditional variance to negative shocks. The dummy variable takes a 2
st. As the Gumbel copula and its rotation require the parameter to be
value of 1 in response to negative shocks and 0 in response to positive
greater than one, we follow Patton (2013) and set the function of δt to
shocks. If the asymmetric coefficient of the conditional variance γi is
be δt =1+exp(ft); for the Gaussian and Student-t copulas, we set the
significantly positive, a negative shock leads to a higher future
correlation to be δt=(1–exp(−ft))/(1+exp(−ft)) to ensure that the
conditional variance than a positive shock of the same magnitude.
correlation lies in the range (−1, 1).

3.3. Static and GAS dynamic copula approach

3.4. Portfolio strategies design

A very attractive advantage for copulas is tail dependence, which

measures the probability of variables going up and down together in
extreme market cases. The lower and upper tail dependence can be
written in terms of copulas as follows:
C (v , v )
λL (v ) = lim P [X ≤ Fc−1 (v )|Y ≤ Fs−1 (v )] = lim ,
v →0 v →0 v (13)
1 − 2v + C ( v , v )
λU (v ) = lim P [X ≥ Fc−1 (v )|Y ≥ Fs−1 (v )] = lim
v →1 v →1 1−v
where λL (v ), λU (v ) ∈ [0, 1]. In the case of λL > 0 (λU > 0), lower
(upper) tail dependence is exhibited, indicating a non-zero probability
of observing extremely small (large) values for both series.
As noted above, the copulas provide information on the average
dependence in addition to the tail dependence. The average depen-
dence given by linear correlation ρ, Kendall's τ, or Spearman's rho can
be obtained from the dependence parameter of the copulas.
Implications for portfolio establishments can be obtained from the
copula dependence information. For instance, if the average depen-
dence between two asset returns is negative or weakly positive, it
means that one asset can provide diversification benefits for the other
asset on average, whereas if the average dependence is high and
positive, then, with no constraints of short sales, the risk of holding a
long position of one asset can be hedged by holding some short position
of the other asset. Similarly, if the tail dependence is large (very close to
zero), then weak (strong) diversification benefits are expected under
extreme market circumstances, and in the case of high tail dependence,
hedging the extreme risk by taking short positions can also be
considered.
The specification of the copula function is crucial in determining
whether one asset can hedge the risks of the other asset. We consider
four kinds of copula—Gaussian, Student-t, Gumbel, and Rotated
Gumbel—to capture different patterns of tail dependence between
markets, including tail independence and symmetric and asymmetric
tail dependence (the details of these copula functions are provided in
Designing portfolio strategies is a vital element of risk management.
Given their popularity among private investors, fund and index
providers, pension funds, endowments, quantitative funds, exchange-
traded products, and long-term investors (e.g., Anderson et al., 2012;
Benartzi and Thaler, 2001; Bessler and Wolff, 2015; Lien et al., 2002;
Reboredo, 2013a,b), five portfolio strategies are considered: naïve
diversification, diversified risk-parity, optimal portfolio, OLS hedging,
, and dynamic hedging. The estimates of the conditional mean and
(14)
conditional variance for each asset return (Section 3.2) and the
dependence structure between different assets (Section 3.3) are needed
to form these strategies.
The naïve diversification strategy is advantageous because it is
simple to implement and can eliminate the transaction costs induced
by adjustments of the dynamic asset weights. It allocates wealth equally
to different assets, thus giving equal weight to energy commodity
futures and carbon assets for each paired asset. The diversified risk-
parity strategy allocates wealth to each asset in the portfolio based on
risk, with low volatility assets earning a higher premium per unit of
volatility than high volatility assets. The weight for asset i (i=EUA,
energy-related futures) is given by
1/σˆi2, t
wi, t = 2
.
∑i =1 (1/ σˆi2, t ) (19)
The optimal portfolio strategy aims at minimizing the risk of
portfolios without reducing the expected return. Unlike the diversified
risk-parity strategy, it not only takes into account the risk of asset
returns but also the correlations between assets. According to Kroner
and Ng (1998), the optimal weight of carbon assets in the portfolio at
time t is given by
2
σˆenergy ,t − σ
ˆEUA − energy, t
wtEUA = 2 2
.
σˆEUA, t − 2σ
ˆEUA − energy, t + σˆenergy ,t (20)
2 2
In both Eqs. (19) and (20), σ̂EUA , t and σ̂energy, t are computed via Eq.

Appendix A). This consideration is important because the joint move-
ments of asset price returns in upside and downside markets could be
different; correspondingly, the portfolio strategies should be different.
Given that the dependence between markets can be time-varying,
we adopt the GAS model of Creal et al. (2013) to further describe the
parameters in these four copulas. Compared with the method proposed
(11) or Eq. (12). In Eq. (20), σ̂EUA − energy, t is the conditional covariance of
energy commodity futures and carbon assets, which is obtained by the
conditional volatility of asset returns in Eqs. (11) or (12) and the
3
More details regarding the comparison of these two approaches can be seen in
Section 3.1 in Creal et al. (2013).

197
X. Wen et al.
dependence information from the best copula model fit. According to
Kroner and Ng (1998), in Eq. (20), when wtEUA > 1, wtEUA=1, and
when wtEUA < 0, wtEUA=0. For energy commodity futures, the
optimal weight is 1-wtEUA. Portfolios based on the naïve diversifica-
tion strategy, diversified risk-parity strategy, and optimal portfolio
strategy are all diversified portfolios; they are denoted as Portfolio 1,
Portfolio 2, and Portfolio 3, respectively.
In addition to holding long positions on both energy commodity
futures and carbon assets in a portfolio like the above diversified
portfolios, we can also consider a hedge portfolio consisting of a long
position on carbon assets and a short position on energy commodity
futures. More specifically, a long position of 1 dollar on carbon assets is
hedged by a short position of β dollars on energy commodity futures.
Given that the static and simple hedging strategy based on an OLS
regression is found to be superior to the dynamic hedging strategy in
some studies (e.g., Lien et al., 2012; Moosa, 2003), we also investigate
whether such a conclusion holds in this study. Based on an OLS
regression, the static β is the slope coefficient for regression of the
energy commodity returns on the carbon asset returns. The time-
varying βt is defined as
σˆenergy − EUA, t
βt = . period. We use the futures prices for EUAs because the futures market
2
σˆenergy ,t
Portfolios based on the static and time-varying β are hedge
portfolios denoted as Portfolio 4 and Portfolio 5, respectively.
Portfolio 1 and Portfolio 4 are static portfolio strategies; Portfolio 2,
Portfolio 3, and Portfolio 5 are dynamic portfolio strategies.
3.5. Portfolio performance measures
To evaluate the effectiveness of Portfolios 1–5 for hedging the risk
of the carbon asset price, we use four measures: risk-adjusted return
increase, variance reduction, value-at-risk (VaR) reduction, and ex-
pected shortfall (ES) reduction.
We denote the portfolio consisting of the carbon asset only as the
benchmark. The risk-adjusted return increase (RR Inc) is given by
RRportfolio i − RRbenchmark
RR Inc .portfolio i = ,
RRbenchmark
where RRportfolio i (RRbenchmark) is the risk-adjusted return of
portfolio i [i=1,2,3,4,5] (benchmark), which is measured as the mean
portfolio return as a ratio of its standard deviation. A positive value
indicates a risk-adjusted return increase compared with the bench-
mark.
The variance reduction (Var Red) is written as
Varbenchmark − Varportfolio i
Var Red .portfolio i = ,
Varbenchmark
where Var is the variance of the portfolio (benchmark) return. A
positive value indicates a variance reduction in comparison with the
benchmark.
We also use measures of VaR (value-at-risk) reduction and ES
(expected shortfall) reduction for extreme downside risk reduction (the
details of calculating the VaR and ES are provided in Appendix B). A
positive value indicates extreme downside risk reduction in comparison
with the benchmark.
The VaR (ES) reduction is calculated as follows:
1 T VaRbenchmark, t − VaRportfolio i, t
VaR Red .portfolio i = ∑t =1
T VaRbenchmark , t
1 T ESbenchmark, t − ESportfolio i, t
ES Red .portfolio i = ∑t =1
T ESbenchmark , t
Economic Modelling 62 (2017) 194–206
800 50
45
700
Energy commodity futures price
40
600
35
500 30
EUA price
400 25
300 20
15
200
10
100 5
0 0
03/08 03/09 03/10 03/11 03/12 03/13 03/14 03/15 03/16
Crude oil Coal Natural gas Electricity EUA
Fig. 1. Energy commodity futures and EUA futures prices.
4. Data
We use daily prices of EUA futures for the carbon asset obtained
from the Thomson Reuters DataStream database. As Phase I (from
2005–2007) was the test period of the EU ETS, our sample period
begins from Phase II, covering the March 3, 2008 to January 8, 2016
(21) aids the price discovery process by transferring information to the spot
market, and thus it is advantageous for investors to use the futures
market when designing risk and trading strategies because the price
signal is of better quality (Reboredo, 2013a; Rittler, 2012; Ortas and
Alvarez, 2016).
In addition to the EUA futures, other energy commodity futures for
the same period are (i) the NYMEX (New York Mercantile Exchange)
WTI crude oil futures (oil); (ii) the ARA (Argus/McCloskey) coal
futures (coal); (iii) the ICE (International Commodities Exchange)
UK natural gas futures prices (natural gas); and (iv) the EEX
(European Energy Exchange) electricity futures prices (electricity).
The price data for crude oil futures are from the EIA, and those of
coal, natural gas, and electricity futures are from the Thomson Reuters
DataStream. As the EUA futures and EEX electricity futures are priced
in euros and the ICE natural gas futures are priced in pounds sterling,
we convert them to US dollars for consistency with the WTI crude oil
futures and ARA coal futures prices. Exchange rate data are from the
(22) Federal Reserve Bank of St. Louis.
The price development of EUA futures and other energy commodity
futures during the March 3, 2008 to January 8, 2016 period is depicted
in Fig. 1. It is worth noting that for each variable under study,
continuous time series are constructed using the one-year daily
settlement prices in the year before the futures contracts on that
variable expired.4 Fig. 1 clearly shows the negative effect of the 2008
global financial crisis (GFC). Because of the economic slowdown and
lower energy demands, the carbon emissions market plummeted along
(23) with other energy commodity markets from the second half of 2008.
From the beginning of 2009, all markets gradually recovered, with the
price slightly increasing in the following two years. A similar downward
trend in the market is again observed from mid-2011, especially in the
EUA carbon market. Based on Balcilar et al. (2016), this seems to
coincide with a prolonged crisis in the Eurozone that led to a
widespread economic slowdown, driving energy demand down.
Recently, because of the worsening global economic outlook and
continuing weak energy demand, the prices of the carbon asset and
other energy commodity futures have remained low.
Table 1 provides the descriptive statistics of the asset returns, with
returns calculated as 100 times the difference in the log of the futures
, prices. Compared with other energy commodity futures, the EUA
(24)
carbon market experienced the greatest price return volatility. All of
.
(25)
4
For a detailed description of how we computed the continuous price series for each of
the variables under study, refer to Zhang and Sun (2016).
198
X. Wen et al. Economic Modelling 62 (2017) 194–206

Table 1
Descriptive statistics for returns (%).

EUA Oil Coal Natural gas Electricity

Mean −0.070 −0.057 −0.059 −0.037 −0.074

SD 3.368 2.530 1.690 2.855 2.812
Skewness −0.917*** 0.080*** −0.854*** 2.361*** 0.115**
Kurtosis 17.391*** 5.108*** 8.235*** 28.453*** 4.329***
JB 25001.041*** 2134.759*** 5782.770*** 68004.924*** 1536.249***
Q(15) 74.579*** 57.827*** 54.499*** 31.572*** 50.384***
Q2(15) 207.501*** 1695.824*** 1732.554*** 18.188 1439.200***
ARCH(5) 0.048 20.716*** 17.748*** 0.001 18.821***
Pearson Correlation 1 0.236 0.207 0.181 0.297

Notes. Daily observations are 1962 in total for the period of Mar 3, 2008 to Jan 8, 2016. The Jarque-Bera (J-B) statistic tests for the null hypothesis of normality in the sample returns
distribution. The Q (15) (Q2(15)) is the Ljung-Box Q test of serial correlation up to 15 lags in the (squared) returns. ARCH (5) is the Lagrange Multiplier (LM) test for the autoregressive
conditional heteroscedasticity (ARCH) up to 5 lags. Pearson correlation is the correlation coefficient between EUA and other energy commodity futures (including Oil, Coal, Natural Gas
and Electricity) returns. ***, **, * indicate statistical significance at the 1%, 5% and 10% level respectively.

the asset return distributions are skewed, leptokurtic, and non-normal. Table 2
The Ljung–Box statistics suggest the presence of serial correlation in all Empirical copula for EUA and energy futures returns.
returns, and the ARCH effect is very significant in the squared returns
EUA – Oil 37 31 22 19 21 20 20 13 7 7
of crude oil, coal, and electricity futures. Finally, the relatively low 41 19 26 18 17 17 18 14 16 10
values of the Pearson linear correlations indicate the diversification 20 26 24 21 16 24 18 17 17 13
potential of portfolios consisting of EUA futures and other energy 21 24 25 21 21 14 15 19 22 14
commodity futures. 24 24 15 24 19 18 17 22 19 14
17 16 17 24 15 24 24 19 21 19
We also conduct a preliminary examination of the dependence 13 20 15 25 24 13 24 22 22 18
structure between the EUA carbon market and other energy commod- 9 13 21 16 21 20 18 21 28 29
ity markets. With the empirical data, and following existing studies 11 6 18 13 19 23 19 25 24 38
(e.g., Reboredo, 2011, 2013a, 2013b; Wen et al., 2012), we rank each 4 17 13 15 23 23 23 24 20 35
EUA – Coal 46 23 31 19 9 20 13 17 14 5
pair of EUA and other energy asset returns in ascending order and
31 25 20 22 13 27 17 9 17 15
divide each series evenly into 10 bins. The observations with the lowest 23 16 25 20 15 24 20 19 16 18
values are included in bin 1 and those with the highest values are 17 27 15 20 16 28 17 20 18 18
included in bin 10. As our focus is on the values of one series associated 13 20 22 22 24 28 16 20 16 15
with the values of another series, we count the number of observations 18 22 18 20 9 34 19 20 22 14
10 18 15 17 12 31 31 20 21 21
in each cell (i, j). The results of the empirical copula table are shown in 14 17 16 18 17 25 18 22 20 29
Table 2. Given that more observations are found in cells (1, 1) and (10, 13 11 18 21 15 23 23 23 27 22
10), tail dependence is expected between the carbon market and other 12 17 16 17 8 14 22 26 25 40
energy markets; as the differences in the number of observations in EUA – Natural gas 43 34 24 18 10 17 14 12 10 15
34 24 25 15 16 14 17 15 18 18
these cells are not significant, symmetric tail dependence is indicated.
21 30 27 16 24 13 14 19 16 16
26 18 22 25 18 19 19 18 14 17
5. Empirical results 15 23 17 26 24 23 19 18 17 14
13 12 15 20 30 28 26 22 18 12
12 14 17 26 22 27 21 21 17 19
5.1. Estimates of marginal distribution models
11 19 18 14 16 25 17 26 26 24
15 12 17 18 19 15 24 15 30 31
Based on the selection criteria of SIC (Beine and Laurent, 2003), 7 10 14 18 17 15 25 30 30 31
the best-performing marginal distribution models are AR(2)-GJR(1,1) EUA – Electricity 43 33 24 15 20 20 18 11 7 6
43 27 26 17 16 13 18 11 17 8
for the EUA futures returns, AR(0)-GJR(1,1) for the WTI crude oil
23 26 23 35 15 15 21 13 15 10
futures returns, AR(1)-GARCH(1,1) for the coal futures returns, and 17 28 28 18 20 21 16 19 16 13
AR(0)-GARCH(1,1) for both the natural gas and electricity futures 23 19 24 21 26 19 16 17 17 14
returns. The parameter estimates for these marginal models are 18 16 15 18 25 23 21 21 22 17
reported in Table 3. 11 15 16 23 21 22 20 28 21 19
6 14 14 16 21 23 23 18 36 25
In the conditional mean equations, the terms of AR(2) and AR(1)
10 7 13 14 18 20 21 23 24 46
are significant for the EUA and coal, respectively, implying that the 3 11 13 19 14 20 22 35 21 39
lagged returns have a significant effect on current returns. For crude
oil, natural gas, and electricity, the current returns are not related to Notes. Daily observations for each series are 1962 in total. EUA futures returns are
the returns in previous periods. In the conditional volatility equations, ranked along the horizontal axis in ascending order (from top to bottom), whereas energy
commodity futures returns (including Oil, Coal, Natural Gas and Electricity futures
the coefficients of lagged variance and lagged squared residuals are
returns) are ranked along the vertical axis in ascending order (from left to right). Each
highly significant in all cases, suggesting that the volatility at time t
box includes the number of observations that belong to the respective quantiles of EUA
depends on both the volatility and the relevant information at time t−1. and energy commodity series.
A leverage effect exists in the EUA and oil prices, but such an
asymmetric effect is not found for the coal, natural gas, or electricity market, a variety of geopolitical and economic events are relevant,
prices, which is consistent with the findings of Efimova and Serletis suggesting the presence of an asymmetric response to the negative
(2014); in addition, we find evidence of an asymmetric effect in the price shocks in these two markets. In contrast, according to
carbon price but not in the coal price. Generally speaking, financial Hammoudeh et al. (2015), coal and natural gas are cheaper and more
turbulence, policy shifts, geopolitical tensions, and extreme events may abundant than crude oil, and electricity is regulated with no major
induce an asymmetric response to shocks. For the EUA market, shifts surprises in price, suggesting that the asymmetric responses to
between different market phases are obvious, whereas for the crude oil

199
X. Wen et al. Economic Modelling 62 (2017) 194–206

Table 3 Table 4
Parameter estimates of the marginal distribution models. Parameter estimates of the static copulas.

EUA Oil Coal Natural gas Electricity EUA- Oil EUA-Coal EUA-Natural gas EUA-Electricity

Conditional mean equation Static Gaussian Copula

µi −0.012 −0.035 −0.040 −0.013 −0.029 ρ 0.261*** 0.216*** 0.246*** 0.340***
φi,1 (0.051) (0.038) (0.029) (0.043) (0.043) LL (0.022) (0.022) (0.022) (0.021)
φi,2 −0.004 0.141*** AIC 69.471 46.751 61.101 120.541
(0.023) (0.022) −136.940 −91.500 −120.201 −239.080
−0.057**
(0.022) Static Student-t Copula
ρ 0.262*** 0.219*** 0.255*** 0.342***
Conditional variance equation v (0.022) (0.020) (0.022) (0.020)
ωi 0.080*** 0.015* 0.029*** 0.023** 0.013 LL 30.198 11.592*** 11.989*** 25.431*
αi (0.027) (0.008) (0.009) (0.009) (0.008) AIC (93.998) (3.322) (3.508) (15.239)
βi 0.048*** 0.018** 0.095*** 0.036*** 0.049*** 70.565 52.735 67.820 122.105
γi (0.015) (0.008) (0.016) (0.007) (0.009) −137.125 −101.464 −131.633 −240.204
ϕi 0.921*** 0.948*** 0.895*** 0.962*** 0.950***
ηi (0.011) (0.008) (0.014) (0.006) (0.008) Static Gumbel Copula
0.048** 0.065*** −0.041 0.088*** −0.061* δ 1.175*** 1.145*** 1.169*** 1.252***
(0.019) (0.015) (0.032) (0.031) (0.033) LL (0.020) (0.018) (0.019) (0.021)
−0.056* −0.070** 5.096*** 3.864*** 8.789*** AIC 59.443 45.345 53.921 106.334
(0.032) (0.033) (0.617) (0.341) (1.222) −116.885 −88.688 −105.839 −210.666
5.943*** 9.149***
(0.770) (1.354) Static Rotated Gumbel Copula
δ 1.169*** 1.143*** 1.173*** 1.243***
Diagnosis tests LL (0.019) (0.018) (0.019) (0.021)
Q(15) 0.301 0.896 0.783 0.264 0.559 AIC 57.338 43.876 58.577 102.168
Q2(15) 0.936 0.012 0.160 0.936 0.480 −112.674 −85.751 −115.152 −202.333
ARCH(2) 0.994 0.478 0.882 0.997 0.986
ARCH(5) 0.999 0.903 0.975 0.999 0.999 Notes. This table provides parameter estimates of the static copulas with standard errors
K-S test 0.404 0.610 0.725 0.572 0.404 in parentheses. Log-likelihood (LL) and Akaike information criterion (AIC) values
A-D test 0.681 0.620 0.576 0.675 0.594 adjusted for small-sample bias are provided for the copula models. ***, **, * indicate
statistical significance at the 1%, 5% and 10% level respectively.
Notes. This table provides parameter estimates of marginal distribution models with
standard errors in parentheses, and p-values for the diagnosis tests of (squared)
EUA–electricity. This implies that the tail dependence for EUA–coal is
standardized residuals. Parameters of marginal distribution models are defined in Eqs.
(6)–(9) and Eqs. (11) and (12). For K-S (Kolmogorov-Smirnov) and A-D (Anderson- the lowest, while that for EUA–electricity is the highest, again
Darling) tests, p-values less than 0.05 indicate a rejection of the null hypothesis that the indicating that coal (electricity) futures may provide the highest
particular model is well specified at the 5% level. ***, **, * indicate statistical significance (lowest) diversification benefits for the carbon asset in extreme market
at the 1%, 5% and 10% level respectively. cases. Table 5 presents the estimates of the GAS dynamic copulas.
Given the high significance of the parameters in the GAS specification,
negative price shocks are not statistically significant in these three the static copulas may not be adequate for describing the dependence
energy markets. structure. The value of βGAS is very close to but less than 1 in all cases,
The Ljung–Box statistics and ARCH effect tests suggest the absence indicating strong persistence in the market dependence.5
of serial correlation and the ARCH effect in (squared) returns in almost Because different copulas have different average and tail depen-
all cases. In addition, the p-values for the K–S (Kolmogorov–Smirnov) dence characteristics and because the specifications of the copula
and A–D (Anderson–Darling) tests show that all of the models pass the functions are important for portfolio management, we need to choose
tests at the 5% level, indicating that the probability integral transforms the optimal copula for characterizing the dependence structure be-
are uniform (0, 1). Thus, the marginal distribution models can tween the EUA futures market and other energy commodity futures
adequately describe the distributions of individual asset returns, and markets. Following previous studies (see, among others, Guo et al.,
the copula functions can correctly capture the co-movements between 2016; Reboredo, 2013b; Wen et al., 2012), we select the optimal copula
the EUA and other energy commodity markets. for each pair of assets based on the log-likelihood and AIC values. As
shown in Tables 4 and 5, the static Student-t copula is the optimal
5.2. Estimates of copulas static copula for all pairs of EUAs and other energy commodity futures,
whereas the GAS dynamic Student-t copula is the best among the
Table 4 presents the estimates of the static copulas. For the dynamic copulas. Thus, a symmetric dependence structure between the
Gaussian and Student-t copulas, the dependence parameter ρ is carbon and other energy commodity asset returns is confirmed.
positive and significant in all paired assets, as the ρ value is positive Combining the static and time-varying copulas, we can see that for
but relatively low; therefore, energy commodity futures for the carbon all pairs of EUAs and other energy commodity futures, the log-
asset are expected to have diversification benefits. Among all pairs of likelihood and AIC values show the GAS dynamic Student-t copula to
EUA and energy commodity futures, the correlation coefficient for be the best fit among the copula functions examined in this paper.
EUA–coal is the lowest, while that for EUA–electricity is the highest, Thus, the dependence between the EUA and other energy futures
indicating that coal futures may provide the highest diversification markets is time-varying. Furthermore, the EUA futures market is
benefits for EUA while electricity futures offer the lowest. The degrees linked with the other energy-related commodity futures markets to
of freedom for the Student-t copula are not low, indicating the the same degree regardless of whether markets are booming or
existence of tail dependence in all paired assets. As for the Gumbel crashing. Thus, the models that assume tail independence between
copula and its rotation, both of which characterize asymmetric tail the markets would underestimate the extreme risks.
dependence, the parameters are significant and indicate positive
dependence; the similar values of the parameters imply that the upper
and lower tail dependence are similar as well. The lowest value of the 5
It should be noted that the β parameter in GAS models plays the same role as (α+β)
parameter δ is found for EUA–coal, while the highest value is found for in a GARCH (1, 1) model; see Example 1 in Creal et al. (2013).

200
X. Wen et al. Economic Modelling 62 (2017) 194–206

0.7
Table 5
Parameter estimates of the GAS dynamic copulas. 0.6
0.5
EUA-Oil EUA-Coal EUA-Natural gas EUA-Electricity 0.4
0.3
GAS dynamic Gaussian Copula 0.2
c 0.551*** 0.427*** 0.540*** 0.740*** 0.1
αGAS (0.093) (0.136) (0.073) (0.104) 0
βGAS 0.020** 0.014*** 0.017* 0.022***
-0.103/08 03/09 03/10 03/11 03/12 03/13 03/14 03/15 03/16
LL (0.008) (0.005) (0.010) (0.008)
AIC 0.984*** 0.994*** 0.975*** 0.986*** EUA-Oil EUA-Coal EUA-Gas EUA-Electricity
(0.013) (0.005) (0.027) (0.011)
Fig. 2. Dynamic correlations between energy commodity futures and EUA futures
81.987 64.458 69.450 139.424
−157.962 −122.904 −132.888 −272.836 prices.

GAS dynamic Student-t Copula

0.06
c 0.009*** 0.001 0.006*** 0.011***
0.05
αGAS (0.001) (0.001) (0.001) (0.001)
βGAS 0.042*** 0.021*** 0.027*** 0.045*** 0.04
v (0.009) (0.000) (0.008) (0.012) 0.03
LL 0.983*** 0.999*** 0.988*** 0.985*** 0.02
AIC (0.003) (0.000) (0.000) (0.000) 0.01
25.063*** 48.309*** 11.628*** 26.525*** 0
(0.003) (0.034) (0.029) (0.046) 03/08 03/09 03/10 03/11 03/12 03/13 03/14 03/15 03/16
83.584 64.709 76.321 141.640
−159.148 −121.398 −144.622 −275.260 EUA-Oil EUA-Coal EUA-Gas EUA-Electricity

Fig. 3. Dynamic tail dependence between energy commodity futures and EUA futures
GAS dynamic Gumbel Copula
prices.
c −0.020*** −0.005** −0.041*** −0.017***
αGAS (0.000) (0.002) (0.004) (0.000)
βGAS 0.085*** 0.064*** 0.071*** 0.074*** 2008 to the beginning of 2012, a turbulent period that encompasses
LL (0.000) (0.007) (0.023) (0.018) both the GFC and the Eurozone debt crisis, both of which led to large
AIC 0.989*** 0.998*** 0.977*** 0.988***
shocks to most financial assets and thus to intensified correlations.
(0.002) (0.000) (0.001) (0.002)
71.668 59.408 59.118 122.940 Conversely, lower correlation coefficients are observed from the second
−137.324 −112.804 −112.224 −239.868 half of 2012 to the end of 2014, a period during which global economic
conditions gradually recovered, leading to lower risk synergy between
GAS dynamic Rotated Gumbel Copula carbon and energy commodity futures. Starting in 2016, with the
c −0.028*** −0.007 −0.096 −0.019
worsening global economic outlook, commodity prices (including
αGAS (0.004) (0.009) (0.653) (0.016)
βGAS 0.105*** 0.067*** 0.167 0.101 energy commodity prices) slumped continually, and the carbon price
LL (0.015) (0.025) (0.507) (0.077) faced great pressure and experienced further falls, again leading to
AIC 0.986*** 0.997*** 0.945*** 0.988*** intensified correlations. From this viewpoint, the dependence between
(0.000) (0.004) (0.206) (0.011)
the EUA and energy commodity futures returns is higher in Phase II of
71.801 58.996 68.819 124.295
−137.590 −111.980 −131.626 −242.578 the EU ETS and lower in Phase III.
The tail dependence is presented in Fig. 3. The lowest tail
Notes. This table provides parameter estimates of the time-varying copulas with standard dependence is found in EUA–coal, while the highest tail dependence
errors in parentheses. Log-likelihood (LL) and Akaike Information Criterion (AIC) values is exhibited in EUA–natural gas. These findings can be attributed to the
adjusted for small-sample bias are provided for the copula models. ***, **, * indicate fact that under extreme circumstances, power plants are expected to
statistical significance at the 1%, 5% and 10% level respectively.
have greater motivation or pressure to contribute to cleaner production
or to offset production costs; thus, they switch to the much cheaper,
The dynamic correlations between the EUA and energy commodity
more abundant, and cleaner energy source, natural gas. This strength-
futures returns from the best-fit copula, the GAS dynamic Student-t
ens the co-movements of the natural gas price and the carbon price. In
copula, are displayed in Fig. 2. It can be seen that the correlations
addition, because of the GFC and the Eurozone debt crisis experience
between the EUA and other energy commodity futures returns are very
in Phase II of the EU ETS, tail dependence is found to be generally
volatile: they can be at a positive low level and at a relatively high level,
higher in Phase II than in Phase III.
suggesting that both diversification and hedging strategies should be
considered for managing the carbon asset risk. The correlations of
5.3. Portfolio performance
EUA–electricity are the highest, while those of EUA–coal are the
lowest. For the EUA–electricity pair, the strong relationship can be
The dynamic average dependence and tail dependence between the
attributed to the fact that the electricity sector is the main contributor
EUA futures market and other energy commodity futures markets
to carbon emissions; further reasons may include strong regulations, a
presented above are crucial for investors to hedge their exposure to
lack of energy substitution, and the inelastic demand in the electricity
carbon price movements and downside risks. Using the various
sector in the short run (Hammoudeh et al., 2015). The relatively weak
portfolio strategies described in Section 3.4, different portfolios for
relationship of EUA–coal can be attributed to the fact that coal is much
managing the carbon risk are established. In addition, portfolio
less clean than crude oil and natural gas, the cheapness and abundance
performance is considered to determine the extent to which the
of natural gas and environmental regulatory constraints have caused
different portfolios can reduce the carbon risk.
coal to lose its competitiveness in power plants, leading to the weak
Table 6 presents the risk–return evaluations of portfolios consisting
relationship between coal and the EUA. The highest (lowest) correla-
of EUAs and other energy commodity futures.6 Compared with the
tions of EUA–electricity (EUA–coal) indicate that coal (electricity)
diversified portfolios, the hedged portfolios, including the static
futures have the highest potential for diversifying (hedging) the carbon
asset.
For all paired assets, high correlation coefficients are observed from 6
To save space, detailed information about portfolio weights is not provided; it is
available from the authors upon request.

201
X. Wen et al. Economic Modelling 62 (2017) 194–206

Portfolio 4 and dynamic Portfolio 5, most significantly increase the Table 6

risk-adjusted returns (RR) of the benchmark portfolio consisting of the Risk-return evaluations for portfolios of EUA and energy commodity futures.
EUA carbon asset only, whereas the diversified portfolios, including the
RR Inc. Var Red. VaR Red. ES Red.
diversified risk–parity weighted Portfolio 2 and optimally weighted
Portfolio 3, perform best in reducing the variance of the benchmark Benchmark – – – –
portfolio. Within the five portfolio strategies for the EUA and energy
EUA-Oil Portfolio 1 −0.286 0.521 0.250 0.367
futures, Portfolio 4 of EUA–oil and EUA–electricity, and Portfolio 5 of
Portfolio 2 −0.238 0.670 0.328 0.431
EUA–coal and EUA–natural gas perform best in increasing the risk- Portfolio 3 −0.333 0.671 0.353 0.456
adjusted returns, whereas Portfolio 3 of EUA–oil and EUA–natural Portfolio 4 0.238 0.056 −0.136 0.021
gas, and Portfolio 2 of EUA–coal and EUA–electricity, provide the Portfolio 5 −0.143 0.055 0.010 0.137
highest reductions in variance. Furthermore, among the several choices EUA-Coal Portfolio 1 −0.524 0.635 0.370 0.465
Portfolio 2 −1.000 0.801 0.520 0.598
of energy commodity futures, the EUA carbon asset with coal futures
Portfolio 3 −0.952 0.801 0.526 0.600
improves the risk–return pattern of the benchmark portfolio most; Portfolio 4 0.333 0.043 −0.102 0.049
despite the fact that electricity futures are not the worst choice for Portfolio 5 0.619 0.055 −0.004 0.127
improving carbon asset returns, they are the last option for investors to EUA-Natural gas Portfolio 1 −0.048 0.494 0.234 0.355
Portfolio 2 −0.810 0.651 0.328 0.431
mitigate carbon risk.
Portfolio 3 −0.905 0.654 0.325 0.429
The risk evaluation of downside risk gains for a 99% confidence Portfolio 4 0.095 0.033 −0.102 0.047
level using the best fitting copula, the GAS dynamic Student-t copula, is Portfolio 5 0.143 0.050 −0.003 0.124
reported in Table 6. The dynamic diversified portfolio is again found to EUA-Electricity Portfolio 1 −0.381 0.452 0.185 0.307
be the superior strategy for carbon risk management.7 The strategy of Portfolio 2 −0.524 0.605 0.281 0.387
Portfolio 3 −0.333 0.535 0.246 0.358
Portfolio 3 (2) provides the largest reduction of VaR and ES of EUA
Portfolio 4 0.381 0.088 −0.200 −0.032
carbon assets when portfolios consist of EUA and oil/coal (natural gas/ Portfolio 5 0.286 0.088 0.046 0.166
electricity). Furthermore, adding coal (electricity) futures to the EUA
carbon asset provides the largest (smallest) reduction in the extremely Notes. This table shows the risk-return evaluations for portfolios of EUA and other
negative risks of the benchmark portfolio.8 energy futures. Portfolio 1–5 is the portfolio with the naïve diversification strategy,
diversified risk-parity strategy, optimal portfolio strategy, OLS regression hedging
Taken together, the dynamic diversified portfolios are much pre-
strategy and dynamic hedging strategy, respectively. RR is the risk-adjusted returns of
ferred to the hedged portfolios in mitigating the risks of EUAs. portfolios. Var is the variance of portfolio returns. RR Inc represents the RR increase is
Furthermore, among the energy commodity assets, coal futures are the difference between the RR of portfolio i (i=1, 2, 3, 4, 5) and the Benchmark portfolio
the best choice for diversifying the (extreme) risks and improving the consisting of the carbon asset only (positive values indicate RR increase). Var Red. is the
returns of the EU carbon market. The superiority of coal futures in Var decrease between the Var of portfolio i (i=1, 2, 3, 4, 5) and Benchmark (positive
improving the risk–return pattern of the EUA futures is due to their values indicate Var reduction). VaR (ES) is the average of dynamic VaR (ES) of portfolio.
VaR (ES) Red. is the VaR (ES) reduction, i.e., the average difference between the
lowest average dependence and tail dependence. Similarly, the highest
dynamic VaR (ES) of portfolio i (i=1, 2, 3, 4, 5) and Benchmark (positive values indicate
average dependence of EUA–electricity makes electricity futures the VaR /ES reduction).
last option for reducing the variance of carbon assets. Although the tail
dependence of EUA–electricity is not the highest among all paired Table 7 reports the risk–return evaluations of portfolios consisting
assets, electricity futures remain the worst choice for mitigating the of EUA and energy commodity futures in these two phases, with the
extreme risk of the carbon asset. This may be attributed to the confidence level for VaR and ES still being 99%.9 As Phase II
differences in the calculations of tail dependence and the extreme risk experienced the 2008 GFC and the 2009 Greek debt crisis, the EUA
measures of VaR and ES, for example, the lower tail dependence is carbon asset had greater VaR and ES and lower risk-adjusted returns
obtained by making the confidence level infinitely close to 1, while the in Phase III when the world was experiencing economic slowdown. The
downside risk measure of VaR and ES are obtained by setting the volatility in Phase III is larger than that in Phase II, which is
confidence level to 99%. The various portfolio strategies could be attributable to the fact that EUAs experienced extremely high volatility
another reason as well. at the beginning of 2013 because of the switch from Phase II to Phase
III.
In Phase II, the diversified portfolio strategies (Portfolios 1, 2, and
5.4. Robustness checks
3) provide the highest increase in the risk-adjusted returns of carbon
assets. In terms of variance reduction, VaR, and ES mitigation, the
5.4.1. Sub-sample analysis
best-performing strategies are still the dynamic diversified portfolios
Given that the EU ETS is composed of three phases—Phase I
even though they vary from the diversified risk–parity weighted and
(2005–2007), a test period, Phase II (2008–2012), the first commit-
optimally weighted portfolios. Among the several choices of hedged
ment period (associated with stricter caps as a number of industrialized
assets for EUAs, natural gas futures are found to be optimal for
nations committed themselves to binding emissions targets during this
improving the risk-adjusted returns of the benchmark, and coal futures
period), and Phase III (2013–2020), the second commitment period,
(electricity futures) are the best (worst) choice for reducing the
which has gone through a number of changes including an overall EU
(extreme) market risks. In Phase III, instead of increasing the risk-
cap, tighter limits on the use of offsets, and a move from allowances to
adjusted returns, the diversified portfolio strategies cause the bench-
auctioning—we attempt to further analyze whether our results hold in
mark returns to be worse; in contrast, the hedged portfolios can
the different phases. Our sample period covers almost all of Phase II
multiply the benchmark returns, with Portfolio 4 of the static hedged
and more than three years of Phase III.
portfolio including EUA and electricity futures performing best. The
dynamic diversified portfolios provide the largest reduction of (ex-
7
We tested the accuracy of the VaR for each portfolio using the likelihood ratio test for treme) risks, and coal futures (electricity futures) are the best (worst)
“correct conditional coverage” proposed by Christoffersen (1998). The results show that alternative for mitigating the risk of carbon assets. It is interesting to
the model can accurately calculate the VaR of portfolios in almost all cases. To save
space, these figures are not provided. More details are available from the authors upon
request.
8 9
The same conclusion can also be obtained based on estimates of VaR for 99.9%, In the sub-sample analysis, to keep the model consistent and for convenient
99.5% and 95% confidence level. To save space, these figures are not provided, but they comparison with the full sample analysis, we used the same marginal distribution
are available from the authors upon request. models and copula as for the full sample.

202
X. Wen et al. Economic Modelling 62 (2017) 194–206

Table 7 5.4.2. Out-of-sample analysis

Risk-return evaluations for portfolios of EUA and energy commodity futures in the sub- Some researchers argue that commodity portfolio benefits may be
samples.
exaggerated in in-sample analyses and are not observable in out-of-
RR Inc. Var Red. VaR Red. ES Red. sample analyses (e.g., Bessler and Wolff, 2015; Daskalaki and
Skiadopoulos, 2011). For this reason, a more realistic assessment of
Benchmark – – – – the out-of-sample analysis is conducted to further evaluate the
sensitivity of our main results.
Panel A: Phase II subsample (Mar 4, 2008–Dec 31, 2012)
EUA-Oil Portfolio 1 0.306 0.373 0.230 0.235 Following Patton (2013), we estimate the marginal distribution
Portfolio 2 0.250 0.461 0.282 0.279 models and copulas using an in-sample period (from March 4, 2008 to
Portfolio 3 0.313 0.463 0.286 0.286 May 24, 2013, including 1308 observations) and evaluate it on the
Portfolio 4 −0.056 0.118 0.077 0.090 remaining 654 observations (the out-of-sample period including 1962
Portfolio 5 0.028 0.122 0.081 0.097
– 1308 = 654 observations). We use the “fixed window” estimation
EUA-Coal Portfolio 1 −0.028 0.539 0.328 0.343
Portfolio 2 0.167 0.657 0.420 0.420 whereby the models (including the marginal distribution models and
Portfolio 3 0.194 0.659 0.425 0.385 the optimal GAS dynamic Student-t copula) are estimated just once
Portfolio 4 0.115 0.092 0.045 0.073 with the data from [1, 1308],10 and the out-of-sample estimates are
Portfolio 5 0.306 0.116 0.061 0.083
then obtained using the in-sample estimates. Patton (2013) points out
EUA-Natural Portfolio 1 0.389 0.305 0.157 0.171
Gas Portfolio 2 0.121 0.445 0.251 0.249 that such a method is useful when the model is computationally
Portfolio 3 0.105 0.452 0.253 0.252 intensive to estimate.11 After obtaining the forecasts of the correlations
Portfolio 4 −0.056 0.057 0.027 0.052 between the EUAs and other energy commodity futures, their marginal
Portfolio 5 0.028 0.091 0.048 0.071 conditional means, and their variances following Siburg et al. (2015),
EUA- Portfolio 1 0.300 0.256 0.150 0.159
we forecast the VaR and ES as follows: (1) for K=5000, simulate K
Electricity Portfolio 2 0.203 0.348 0.199 0.201 (k)
Portfolio 3 0.188 0.357 0.201 0.206 observations uT+1, v(k)T+1 (k=1, …, K) from the Student-t copula; (2)
Portfolio 4 −0.029 0.171 0.109 0.125 convert u T+1 and v(k)T+1 to z(k)T+1,1 and z(k)T+1,2, respectively,
(k)

Portfolio 5 −0.006 0.178 0.123 0.135 using the quantile function of the skewed-t distribution; (3) transform
z(k)T+1,j into the simulated return R(k)T+1,j=μT+1,j+σT+1,jz(k)T+1,j,
Panel B: Phase III subsample (Jan 2, 2013–Jan 6, 2016)
EUA-Oil Portfolio 1 −14.590 0.659 0.251 0.447
where μT+1,j and σT+1,j (j=1, 2) are the forecasted one-step-ahead
Portfolio 2 −19.209 0.869 0.447 0.595 conditional mean and volatility values, respectively; (4) compute the
Portfolio 3 −18.148 0.869 0.440 0.591 simulated portfolio returns as R(k)T+1,P=wT+1,1R(k)T+1,1+wT
Portfolio 4 2.788 0.007 −0.168 0.140 (k)
+1,2R T+1,2, where wT+1,j (j=1, 2) is determined by the forecasted
Portfolio 5 1.431 0.003 −0.162 0.144
variance and correlations and becomes a constant for the strategies of
EUA-Coal Portfolio 1 −10.929 0.726 0.385 0.543
Portfolio 2 −38.508 0.937 0.664 0.763 Portfolios 1 and 4; (5) compute the VaR as the value of the qth
Portfolio 3 −39.070 0.937 0.665 0.764 percentile in the distribution of the simulated portfolio returns for day
Portfolio 4 1.965 0.002 −0.180 0.137 T+1 and the ES as the mean value for situations in which the simulated
Portfolio 5 2.040 0.007 −0.182 0.134
portfolio returns exceed the VaR; and (6) update the information set
EUA-Natural Portfolio 1 −10.930 0.671 0.301 0.480
gas Portfolio 2 −26.983 0.852 0.462 0.602
and repeat the above steps for day T+2 and so forth.
Portfolio 3 −28.263 0.853 0.462 0.600 Table 8 presents the risk–return evaluations for the out-of-sample
Portfolio 4 2.894 0.010 −0.158 0.147 portfolios of the EUAs and energy commodity futures, with the
Portfolio 5 2.301 0.019 −0.145 0.153 confidence level for VaR and ES still being 99%. As in the main results,
EUA- Portfolio 1 −16.791 0.636 0.218 0.414
the hedged portfolios are superior to the diversified portfolios in
Electricity Portfolio 2 −21.143 0.847 0.401 0.550
Portfolio 3 −19.141 0.850 0.397 0.549 increasing the risk-adjusted returns of the EUA carbon asset, whereas
Portfolio 4 4.863 0.017 −0.141 0.152 the dynamic diversified portfolios are better in terms of reducing the
Portfolio 5 3.482 0.007 −0.129 0.160 variance of the benchmark. For EUAs and oil (electricity), Portfolio 3 is
the best strategy for mitigating variance, and for EUAs and coal
Notes. This table shows the risk-return evaluations for portfolios of EUA and other
(natural gas), Portfolio 2 is optimal. To decrease the downside risks
energy futures in the sub-samples. The observations in Phase II and Phase III are 1208
and 754, respectively. Portfolio 1–5 is the portfolio with the naïve diversification of EUAs, dynamic diversified strategies are again confirmed to be
strategy, diversified risk-parity strategy, optimal portfolio strategy, OLS regression optimal, with the optimally weighted portfolio strategy (Portfolio 2)
hedging strategy and dynamic hedging strategy, respectively. RR is the risk-adjusted performing best in all cases. Moreover, electricity futures are the best
returns of portfolios. Var is the variance of portfolio returns. RR Inc represents the RR asset choice for improving the risk-adjusted returns of the benchmark,
increase is the difference between the RR of portfolio i (i=1, 2, 3, 4, 5) and the which may be because the relatively high correlations between the EUA
Benchmark portfolio consisting of the carbon asset only (positive values indicate RR
and electricity futures returns make the hedged portfolio superior in
increase). Var Red. is the Var decrease between the Var of portfolio i (i=1, 2, 3, 4, 5) and
Benchmark (positive values indicate Var reduction). VaR (ES) is the average of dynamic increasing the returns of the benchmark. However, coal futures
VaR (ES) of portfolio. VaR (ES) Red. is the VaR (ES) reduction, i.e., the average (electricity futures) are best (worst) for reducing (extreme) market
difference between the dynamic VaR (ES) of portfolio i (i=1, 2, 3, 4, 5) and Benchmark risks. Thus, the diversification benefits still exist in the out-of-sample
(positive values indicate VaR /ES reduction). model.

note that the portfolios perform much better in Phase III than in the
previous Phase II. Specifically, the variance reduction of coal futures
for the carbon asset can be as high as 0.937 in Phase III; this figure is
even higher than the hedging effectiveness of carbon futures against its
spot risk provided by Balcilar et al. (2016).
Overall, consistent with the results in the full sample, the dynamic
diversified portfolios provide the largest reduction of (extreme) risks,
and coal futures (electricity futures) are found to be superior (inferior)
in mitigating the carbon (extreme) risk.
10
In the out-of-sample analysis, keeping the model consistent and for the convenience
of comparison with the full sample analysis, we used the same marginal distribution
models and copula as considered for the full sample.
11
In addition to the out-of-sample estimation used in this paper, a “recursive” or
“expanding” window in which the forecast for observation t is based on data in the
interval [1, t−1] and a “rolling” window using data only in the interval [t-T, t−1] (T is
assumed to be the observations of in-sample period) are also used for out-of-sample
estimation. However, the former easily increases the computational burden and the latter
has to “throw away” observations from the start of the in-sample period (Patton, 2013).

203
X. Wen et al. Economic Modelling 62 (2017) 194–206

Table 8 periods during which the linkages between asset returns vary drama-
Risk-return evaluations for out-of-sample portfolios of EUA and energy commodity tically.
futures.
One of the main findings to emerge from our empirical analyses is
RR Inc. Var Red. VaR Red. ES Red. that investors in the European carbon market can use the weak
dependence between carbon and some energy-related futures to build
Benchmark – – – – more efficient investment portfolios to prevent (extreme) investment
risks. In particular, our findings suggest that among several energy-
EUA-Oil Portfolio 1 −1.593 0.548 0.258 0.377
Portfolio 2 −2.207 0.714 0.388 0.488 related assets, coal futures are the best choice for diversifying the risks
Portfolio 3 −2.206 0.715 0.388 0.488 of EUAs, while electricity futures are the least attractive; dynamic
Portfolio 4 0.281 0.013 0.003 0.147 diversified portfolios are much preferred to hedged portfolios for
Portfolio 5 0.313 −0.020 0.025 0.158 mitigating the risks of the European carbon market; these results are
EUA-Coal Portfolio 1 −1.000 0.708 0.423 0.507
confirmed in both the sub-sample and out-of-sample analyses. It is also
Portfolio 2 −3.188 0.868 0.598 0.666
Portfolio 3 −3.406 0.865 0.603 0.668 noteworthy that the portfolios perform much better in Phase III than in
Portfolio 4 0.063 0.000 −0.022 0.128 Phase II.
Portfolio 5 0.469 −0.042 0.003 0.142 The implications of the above findings for investors are summarized
EUA-Natural gas Portfolio 1 −1.156 0.594 0.324 0.430
as follows. First, our findings suggest that among several energy-
Portfolio 2 −2.281 0.692 0.402 0.501
Portfolio 3 −2.500 0.688 0.410 0.507 related assets, coal futures have the lowest average/tail dependence
Portfolio 4 0.219 0.012 0.000 0.145 with the carbon asset; they also show that coal futures can reduce the
Portfolio 5 0.250 −0.003 0.020 0.154 carbon (extreme) risk to the largest extent but cannot necessarily
EUA-Electricity Portfolio 1 −1.781 0.500 0.221 0.340 provide the largest increase in the carbon return; this means that the
Portfolio 2 −2.227 0.673 0.339 0.438
low average dependence and tail dependence between the two assets
Portfolio 3 −2.125 0.682 0.344 0.445
Portfolio 4 0.500 0.033 0.035 0.167 are likely to be the robust indicators of risk diversification, but they do
Portfolio 5 0.594 −0.015 0.073 0.189 not imply that one asset can necessarily improve the return of the other
asset. Second, although the dependence between carbon and energy
Notes. This table shows the risk-return evaluations for out-of-sample portfolios of EUA futures can be relatively high, given their low and positive dependence
and other energy futures. The observations in out-of-sample are 654. Portfolio 1–5 is the
in most cases, diversified portfolios are much preferred to hedged
portfolio with the naïve diversification strategy, diversified risk-parity strategy, optimal
portfolio strategy, OLS regression hedging strategy and dynamic hedging strategy, portfolios for mitigating the carbon risk. For risk-averse investors, this
respectively. RR is the risk-adjusted returns of portfolios. Var is the variance of portfolio finding suggests that the diversified portfolio strategy is better than the
returns. RR Inc represents the RR increase is the difference between the RR of portfolio i hedged portfolio strategy. Third, the portfolio performance is much
(i=1, 2, 3, 4, 5) and the Benchmark portfolio consisting of the carbon asset only (positive better in Phase III than in Phase II, which implies that the financial
values indicate RR increase). Var Red. is the Var decrease between the Var of portfolio i stresses during Phase II can weaken the benefits of the diversification/
(i=1, 2, 3, 4, 5) and the Benchmark (positive values indicate Var reduction). VaR (ES) is
hedging of energy-related futures for the carbon asset, while with the
the average of dynamic VaR (ES) of portfolio. VaR (ES) Red. is the VaR (ES) reduction,
i.e., the average difference between the dynamic VaR (ES) of portfolio i (i=1, 2, 3, 4, 5) major reform (an EU-wide cap on emissions and a progressive
and Benchmark (positive values indicate VaR /ES reduction). replacement of free allocation of allowances by auctioning) in Phase
III, these benefits can be strengthened. In particular, in Phase III of EU
6. Conclusions ETS, the high variance reduction of coal futures for the carbon assets
indicates that coal futures are expected to be an even more attractive
As is suggested by Balcilar et al. (2016), to effectively achieve the hedging tool for the carbon asset within the new market mechanism.
targets of emission reductions, continuing emission trading alone is not This finding is also helpful for policy makers in regulating the risks of
enough; the development of risk management strategies for carbon risk carbon markets.
is also necessary. Modeling the dependence between EUA futures and Further research should consider more diverse copula functions to
crude oil futures, coal futures, natural gas futures, and electricity model market dependence; it should also consider the use of higher
futures is important for investors’ and policy-makers’ understanding of frequency data for carbon and energy commodity assets to identify
the co-movements of EUAs and each of these energy futures and of the potential implications for high-frequency traders. Furthermore, trans-
implications for portfolio diversification and hedging strategies. Given action costs and dynamic portfolio weights should be taken into
that most previous research has focused on overall co-movements, thus account in portfolio management.
overlooking the relationships during extreme market movements, and
has focused on the effectiveness of carbon derivative markets, we use Acknowledgments
copula functions to consider both the average dependence and the tail
dependence between EUAs and various energy commodity futures The authors would like to thank the editor and the anonymous
markets to provide more comprehensive information about hedging reviewer for their helpful comments. The financial support offered by
the carbon risk to economic actors. This exploration aids in the the National Natural Science Foundation of China (NSFC) (No.
construction of more resilient portfolios that are able to resist stress 71601157) is gratefully acknowledged.

Appendix A. Copula functions

Appendix A.1. Gaussian Copula

The Gaussian copula, which is regarded as the benchmark of copula functions, characterizes the tail independence. It is given by
CG (ut , vt ; ρ) = Φ (Φ−1 (ut ), Φ−1 (vt )), (A.1)
−1 −1
where Φ is the bivariate standard normal CDF with linear correlation ρ (−1 < ρ < 1); Φ (ut) and Φ (vt) are standard normal quantile functions; in
the Gaussian copula, λU = λL = 0 .

204
X. Wen et al. Economic Modelling 62 (2017) 194–206

Appendix A.2. Student-t Copula

The Student-t copula is another common choice for the dependence structure:
CS (ut , vt ; ρ , v ) = T (t −1 (ut ), t −1 (vt )), (A.2)
−1 −1
where T is the bivariate Student-t CDF with a degree-of-freedom parameter v and linear correlation ρ (−1 < ρ < 1), and t (ut) and t (vt) are the
quantile functions of the univariate Student-t distribution, with v as the degree-of-freedom parameter. The Student-t copula also features symmetric
tail dependence as λU = λL = 2tv (− v 1 − ρ / 1 + ρ ) > 0 ; tv(·) is the CDF of the Student-t distribution with degree of freedom v, and ρ is the linear
correlation coefficient.

Appendix A.3. Gumbel Copula and its rotation

The Gumbel copula, an extreme value copula, has a higher probability concentrated in the upper tail, and its rotation puts the focus on the lower
tail dependence. The Gumbel copula and its rotation are presented by

CG (ut , vt ; δ ) = exp(−((− log ut )δ + (− log vt )δ )1/ δ ), (A.3)

CRG (ut , vt ; δ ) = ut + vt − 1 + CG (1 − ut , 1 − vt ; δ ), (A.4)

where δ ∈ (1, ∞); from it, Kendall's τ, the non-linear measure of market dependence, is (δ−1)/δ, the upper tail dependence is λU=2–2 1/δ
, and the
lower tail dependence is λL=0; for its rotation, the upper tail dependence is λU=0 and the lower tail dependence is λL=2–21/δ.

Appendix B. Value-at-risk and expected shortfall

VaR (value-at-risk) for an asset or portfolio return is defined as follows at the confidence level of (1-p) at time t:
Pr(Rt < VaRt |ψt−1) = p , (B.1)
where ψt-1 is the information set at time t. According to Eq. (B.1), given a time horizon, we have (1-p) confidence that the loss will be no larger than
VaR. However, as this measure only gives a probability of extreme loss, ES (expected shortfall) is also measured; it is given by
ES = E [Rt |Rt < VaRt ( p )]. (B.2)
ES is a VaR-related risk measure that reveals the loss size when the loss is larger than VaR.
We use a Monte Carlo simulation to calculate the VaR and ES of the portfolio returns. We simulate 5000 values of the uniform variates of the
given copula model at each time (time t=1 to T) and invert the marginal cumulated distribution function to obtain standardized residuals for each
asset. These simulated standardized residuals are then used to compute portfolio returns with the portfolio weights calculated in Section 3.4.
Finally, we measure VaR as the qth quantile of the empirical distribution of these simulated portfolio returns and ES as the expected loss exceeding
VaR.
The VaR of the carbon asset is calculated as
−1
VaRt = μ t − Fskew − t (ϕ, η) ( p ) σt , (B.3)
−1
where Fskew − t (ϕ, η) ( p ) is the (1-p) quantile of the skewed-t distribution with the kurtosis parameter η and the asymmetry parameter ϕ. Based on the
VaR, the ES of the carbon asset is the expected loss that exceeds its VaR.

References scheme. Energy Econ. 33 (2), 353–362.

Aatola, P., Ollikainen, M., Toppinen, A., 2013. Price determination in the EU ETS
market: theory and econometric analysis with market fundamentals. Energy Econ.
36, 380–395.
Alberola, E., Chevallier, J., Chèze, B., 2008. The EU emissions trading scheme: the effects
of industrial production and CO2 emissions on carbon prices. Int. Econ. 116 (4),
93–125.
Anderson, R.M., Bianchi, S.W., Goldberg, L.R., 2012. Will my risk parity strategy
outperform? Finance Anal. J. 68, 75–93.
Avdulaj, K., Barunik, J., 2015. Are benefits from oil-stock diversification gone? New
evidence from a dynamic copula and high frequency data?. Energy Econ. 51, 31–44.
Balcilar, M., Demirer, R., Hammoudeh, S., Nguyen, D.K., 2016. Risk spillovers across the
energy and carbon market and hedging strategies for carbon risk. Energy Econ. 54,
159–172.
Beine, M., Laurent, S., 2003. Central bank intervention and jumps in double long
memory models of daily exchange rates. J. Empir. Finance 10, 641–660.
Benartzi, S., Thaler, R.H., 2001. Naïve diversification strategies in defined contribution
saving plans. Am. Econ. Rev. 91, 79–98.
Benz, E., Trück, S., 2009. Modeling the price dynamics of CO2 emission allowances.
Energy Econ. 31, 4–15.
Benz, E., Hengelbrock, J., 2009. Liquidity and price discovery in the European CO2
futures market: An intraday analysis. Carbon Markets Workshop, LSE 5 May 2009.
Bessler, W., Wolff, D., 2015. Do commodities add value in multi-asset portfolios? An out-
of-sample analysis for different investment strategies. J. Bank. Finance 60, 1–20.
Bing, X., Bloemhof-Ruwaard, J., Chaabane, A., Van der Vorst, J., 2015. Global reverse
supply chain redesign for household plastic waste under the emission trading
scheme. J. Clean. Prod. 103, 28–39.
Bredin, D., Muckley, C., 2011. An emerging equilibrium in the EU emissions trading
Brooks, C., 2002. Introductory Econometrics for Finance. Cambridge University Press,
New York.
Bunn, D., Fezzi, C., 2007. Interaction of European carbon trading and energy prices,
Fondazione Eni Enrico Mattei Working Papers, Paper 123.
Chevallier, J., 2009a. Carbon Futures and Macroeconomic Risk Factors: A View From the
EU ETS. Energy Economics, 31, pp 614–625.
Chevallier, J., 2009b. Energy risk management with carbon assets. Int. J. Glob. Energy
Issues 32, 328–349.
Chevallier, J., 2011. Macroeconomics, finance, commodities: interactions with carbon
markets in a data-rich model. Econ. Model. 28 (1–2), 557–567.
Chevallier, J., 2012. Time-varying correlations in oil, gas and CO2 prices: an application
using BEKK, CCC and DCC-MGARCH models. Appl. Econ. 44, 4257–4274.
Christoffersen, P.F., 1998. Evaluating interval forecasts. Int. Econ. Rev. 39, 841–862.
Conrad, C., Rittler, D., Rotfuß, W., 2012. Modeling and explaining the dynamics of
European Union Allowance prices at high-frequency. Energy Econ. 34 (1), 316–326.
Convery, F.J., Redmond, L., 2007. Market and price developments in the European
Union Emissions Trading Scheme. Rev. Environ. Econ. Policy 1 (1), 88–111.
Creal, D., Koopman, S.J., Lucas, A., 2013. Generalized autoregressive score models with
applications. J. Appl. Econ. 28, 777–795.
Daskalaki, C., Skiadopoulos, G.S., 2011. Should investors include commodities in their
portfolios after all? New evidence. J. Bank. Finance 35, 2606–2626.
Daskalakis, G., Markellos, R.N., 2008. Are the European carbon markets efficient? Rev.
Futur. Mark. 17, 103–128.
Daskalakis, G., Psychoyios, D., Markellos, R.N., 2009. Modeling CO2 emission allowance
prices and derivative: evidence from the European trading scheme. J. Bank. Finance
33, 1230–1240.
Efimova, O., Serletis, A., 2014. Energy market volatility modeling using GARCH. Energy
Econ. 43, 264–273.
Fezzi, C., Bunn, D., 2009. Structural interactions of European carbon trading and energy

205
X. Wen et al.
prices. J. Energy Mark. 4, 53–69.
Guo, Y., Wen, X., Wu, Y., Guo, X., 2016. How is China's coke price related with the world
oil price? The role of extreme movements. Econ. Model. 58, 22–33.
Hammoudeh, S., Nguyen, D.K., Reboredo, J.C., Wen, X., 2014. Dependence of stock and
commodity futures markets in China: implications for portfolio investment. Emerg.
Mark. Rev. 21, 183–200.
Hammoudeh, S., Lahiani, A., Nguyen, D.K., Sousa, R.M., 2015. An empirical analysis of
energy cost pass-through to CO2 emissions prices. Energy Econ. 49, 149–156.
Hansen, B.E., 1994. Autoregressive conditional density estimation. Int. Econ. Rev. 35,
705–730.
Joe, H., 1997. Multivariate Models and Dependence Concepts. Monographs in Statistics
and Probability. Chapman and Hall, London, 73.
Koch, N., 2014. Dynamic linkages among carbon, energy and financial markets: a smooth
transition approach. Appl. Econ. 46, 715–729.
Krishnamurti, C., Hoque, A., 2011. Efficiency of European emissions markets: lessons
and implications. Energy Policy 39, 6575–6582.
Kroner, K.F., Ng, V.K., 1998. Modeling asymmetric movements of asset prices. Rev.
Finance Stud. 11, 844–871.
Lien, D., Tse, Y.K., Tsui, A.C., 2002. Evaluating the hedging performance of the constant-
correlation GARCH model. Appl. Finance Econ. 12, 791–798.
Liu, H.H., Chen, Y.C., 2013. A study on the volatility spillovers, long memory effects and
interactions between carbon and energy markets: the impacts of extreme weather.
Econ. Model. 35, 840–855.
Lu, X., Salovaara, J., McElroy, M., 2012. Implications of the recent reductions in natural
gas prices for emissions of CO2 from the US power sector. Environ. Sci. Technol. 46
(5), 3014–3021.
Mansanet-Bataller, M., Pardo, A., Valor, E., 2007. CO2 prices, energy and weather.
Energy J. 28, 73–92.
Marimoutou, V., Soury, M., 2015. Energy markets and CO2 emissions: analysis by
stochastic copula autoregressive model. Energy 88, 417–429.
de Menezes, L.M., Houllier, M.A., Tamvakis, M., 2016. Time-varying convergence in
European electricity spot markets and their association with carbon and fuel prices.
Energy Policy 88, 613–627.
Moosa, I.A., 2003. The sensitivity of the optimal hedging ratio to model specification.
Finance Lett. 1, 15–20.
Narayan, P.K., Sharma, S.S., 2015. Is carbon emissions trading profitable? Econ. Model.
47, 84–92.
Oberndorfer, U., 2009. EU emission allowances and the stock market: evidence from the
electricity industry. Ecol. Econ. 68, 1116–1126.
Ortas, E., Alvarez, I., 2016. The efficacy of the European Union Emissions Trading
Scheme: depicting the co-movement of carbon assets and energy commodities
206
Economic Modelling 62 (2017) 194–206
through wavelet decomposition. J. Clean. Prod. 116, 40–49.
Patton, A.J., 2006. Modelling asymmetric exchange rate dependence. Int. Econ. Rev. 47,
527–556.
Patton, A.J., 2013. Copula Methods for Forecasting Multivariate Time Series. Handbook
of Economic Forecasting Vol. 2. Elsevier North-Holland, Amsterdam, Chapter 16.
Philip, D., Shi, Y., 2016. Optimal hedging in carbon emission markets using Markov
regime switching models. J. Int. Finance Mark. Inst. Money 43, 1–15.
Reboredo, J.C., 2011. How do crude oil prices co-move? A copula approach. Energy
Econ. 33, 948–955.
Reboredo, J.C., 2013a. Modeling EU allowances and oil market interdependence.
Implications for portfolio management. Energy Econ. 36, 471–480.
Reboredo, J.C., 2013b. Is gold a safe haven or a hedge for the US dollar? Implications for
risk management. J. Bank. Finance 37, 2665–2676.
Reboredo, J.C., 2014. Volatility spillovers between the oil market and the European
Union carbon emission market. Econ. Model. 36, 229–234.
Rittler, D., 2012. Price discovery and volatility spillovers in the European Union
emissions trading scheme: a high-frequency analysis. J. Bank. Finance 36, 774–785.
Schultz, E., Swieringa, J., 2014. Catalysts for price discovery in the European Union
Emissions Trading System. J. Bank. Finance 42, 112–122.
Siburg, K.F., Stoimenov, P., Weiß, G.N.F., 2015. Forecasting portfolio value-at-risk with
nonparametric lower tail dependence estimates. J. Bank. Finance 54, 129–140.
Sijm, J., Neuhoff, K., Chen, Y., 2006. CO2 cost pass through and windfall profits in the
power sector. Clim. Policy 6 (1), 49–72.
Subramaniam, N., Wahyuni, D., Cooper, B.J., Leung, P., Wines, G., 2015. Integration of
carbon risks and opportunities in enterprise risk management systems: evidence
from Australian firms. J. Clean. Prod. 96, 407–417.
Tian, Y., Akimov, A., Roca, E., Wong, V., 2016. Does the carbon market help or hurt the
stock price of electricity companies? Further evidence from the European context. J.
Clean. Prod. 112, 1619–1626.
Veith, S., Werner, J.R., Zimmermann, J., 2009. Capital market response to emission
rights returns: evidence from the European power sector. Energy Econ. 31, 605–613.
Wen, X., Wei, Y., Huang, D., 2012. Measuring contagion between energy market and
stock market during financial crisis: a copula approach. Energy Econ. 34,
1435–1446.
Xu, L., Deng, S.J., Thomas, V.M., 2016. Carbon emission permit price volatility reduction
through financial options. Energy Econ. 53, 248–260.
Zhang, Y.J., Sun, Y.F., 2016. The dynamic volatility spillover between European carbon
trading market and fossil energy market. J. Clean. Prod. 112, 2654–2663.
Zhu, B., Ma, S., Chevallier, J., Wei, Y., 2014. Modelling the dynamics of European carbon
futures price: a Zipf analysis. Econ. Model. 38, 372–380.