Portfolio Optimization For Sustainable Investments

Portfolio optimization for sustainable
investments
Armin Varmaz
School of International Business Bremen

[email protected]
Corresponding Author
Christian Fieberg
University of Bremen
[email protected]
Thorsten Poddig
University of Bremen
[email protected]
October 4, 2022
Declarations of interest: none
Electronic copy available at: https://ssrn.com/abstract=3859616

Portfolio optimization for sustainable investments
Abstract
In mean-variance portfolio optimization, factor models can accelerate computation, re-

duce input requirements, facilitate understanding and allow easy adjustment to changing
conditions more effectively than full covariance matrix estimation. In this paper, we de-
velop a factor model-based portfolio optimization approach that takes into account aspects
of the environment, social responsibility and corporate governance (ESG). Investments in
assets related to ESG have recently grown, attracting interest from both academic research
and investment fund practice. Various literature strands in this area address the theoretical
and empirical relation among return, risk and ESG. Our portfolio optimization approach is
flexible enough to take these literature strands into account and does not require large-scale
covariance matrix estimation. An extension of our approach even allows investors to empir-
ically discriminate among the literature strands. A case study demonstrates the application
of our portfolio optimization approach.
Keywords: portfolio optimization, sustainable investment, investor preferences
JEL classification: G11, G12, G23, M14, Q5

I Introduction
Financial investment decisions that must take environmental, social, and corporate gover-
nance (ESG) issues into consideration constitute a fast-growing area in the investment and
banking industry. This growth is driven by the desire of investors to assess the ESG aspects
of firm conduct using nonfinancial data and institutional investors’ direct engagement in
ESG issues (Gillan and Starks, 2000, 2007; Grewal et al., 2016). Motivated by the increasing
popularity of sustainable investments, various approaches have been proposed to incorporate
ESG measures into the investment decision-making process (e.g., Bilbao-Terol et al., 2012;
Ballestero et al., 2012; Bilbao-Terol et al., 2013; Gasser et al., 2017; Pedersen et al., 2021;
Pástor et al., 2021; Steuer and Utz, 2022). A common approach is to extend the traditional
mean–variance approach (e.g., Hirschberger et al., 2013; Gasser et al., 2017; Pedersen et al.,
2021). We will refer to this approach as the extended mean–variance approach.
Investors face various challenges when applying the extended mean–variance approach
from the literature (e.g., Mynbayeva et al., 2022). In this paper, we will focus on three chal-
lenges. The first challenge is that the extended mean–variance approach requires investors
to provide preference parameters for return, risk and ESG. However, for an empirical and
practical application of the extended mean–variance approach, knowledge of investors’ pref-
erences and their parameters is key to implementing utility maximizing portfolios. From a
practical perspective, the choice of preference parameters is unclear. A strand of the litera-
ture estimates the preference parameters from market data (e.g., Jackwerth, 2000; Bollerslev
et al., 2011). Nevertheless, the estimated risk aversions describe an average investor and
hence cannot be easily adapted to a single investor. Another strand of the literature that
we follow links the preference parameters to levels of risk and returns (e.g., Das et al., 2010;
Alexander and Baptista, 2011; Bodnar et al., 2018a). For example, Das et al. (2010) impose
a risk constraint in portfolio optimization and determine the implied risk aversion coefficient.
It seems that the orientation toward risk can be more easily specified by using the level of

risk than by determining the risk aversion coefficients, which are abstract measures for most
investors.
The second challenge is that the extended mean–variance approach incorporates investors’
attitudes toward specific firm ESG characteristics. For empirical and practical portfolio
implementation, investors must assume a relation among ESG, return and risk. A first
strand of the literature argues that there is a direct relation between ESG characteristics and
stock returns (e.g., Friedman and Heinle, 2016; Barnett et al., 2020; Bolton and Kacperczyk,
2020). For example, investors could favor firms with high environmental standards. A second
strand of literature argues that ESG is related to a (systematic) risk factor (see Heinkel et al.,
2001; Luo and Balvers, 2017; Pedersen et al., 2021; Pástor et al., 2021; Hoepner et al., 2021).
For example, there could be an environmental risk factor (Bolton and Kacperczyk, 2020),
and firms with high (low) exposure to such environmental risk factor will offer high (low)
expected returns. The main difference between the two strands of the literature is their
assumption of the relation between ESG and firm risk. The first strand of the literature
assumes that the ESG characteristic (e.g., CO2 emissions) is related only to the expected
return but not to the risk of firms. The second strand of the literature assumes that exposure
to an ESG risk factor is related to firms’ expected return and risk. If the latter literature
is correct, investors will want to explicitly control this source of risk, which is not necessary
if the first strand of literature is correct. The challenge for investors is how to empirically
decide which literature strand is correct and whether their sustainable portfolios need to be
managed by ESG characteristics or by ESG factor loadings. The literature does not provide
an easy solution for how to decide whether the first or the second strand of the literature,
or possibly a combination of the strands, should be followed by investors when they prefer
to consider ESG aspects in their portfolio formation.
Finally, the third challenge is that due to the necessary covariance matrix estimation,
the extended mean–variance approaches discussed in the literature cannot be easily adapted
to unbalanced panels or an investment universe with a large number of assets because the

estimated covariance matrix is unreliable (e.g., Shanken, 1992; Bajeux-Besnainou et al.,
2012). When the number of stocks is larger than the number of historical return observations
per stock, the sample covariance matrix becomes singular (e.g., Ledoit and Wolf, 2003;
Bajeux-Besnainou et al., 2012). This challenge is also faced by mean–variance investors, but
extended mean–variance investors must still overcome the hurdle of estimating the covariance
matrix. The literature develops a large variety of approaches for covariance regularization
to improve the quality of large-scale covariance estimation (e.g., Laloux et al., 1999, 2000;
Ledoit and Wolf, 2003, 2004b,a; Tola et al., 2008; Bajeux-Besnainou et al., 2012; Bodnar
et al., 2018b; Plachel, 2019; Bian et al., 2020; Mynbayeva et al., 2022; Nguyen et al., 2022).
However, the estimation of large-scale covariance matrices remains a common challenge for
adopting traditional or extended mean–variance approaches.
In this paper, we propose a simple new portfolio optimization approach to incorporate
ESG into portfolio formation addressing the three challenges discussed above. Our contribu-
tion to the literature is fourfold. Our first contribution is that our approach links return, risk
and ESG preferences to the investor-desired levels of return, risk and ESG, enabling easy
adoption of our portfolio optimization approach in practice. Our second contribution is that
our approach allows investors to follow the abovementioned individual strands of literature,
or a combination thereof, regarding the relation among ESG, risk and return. The third con-
tribution is that after mild modifications of our portfolio optimization approach, we develop
a formal test that can help investors disentangle the strands of literature. This formal test
allows investors to determine whether a firm’s factor loading to an ESG risk factor or ESG
characteristic or both ESG factor loading and ESG characteristic explains stock return vari-
ations. It also allows investors to specify before portfolio optimization whether ESG factor
loading or ESG characteristic, or both or neither, are important for the formation of their
specific portfolios. Our fourth contribution is that based on the assumptions made in Laloux
et al. (2000), Ledoit and Wolf (2003) and Plachel (2019) regarding covariance shrinkage. We
show that the challenging estimation of a very large covariance matrix can be circumvented

and thus allows portfolio formation for mean–variance and extended mean–variance investors
to be applied, even to unbalanced panels and in situations with larger numbers of assets than
time-series observations.
Our paper develops a general methodology to form portfolios relying on the assumption of
an asset return model. In our derivations and applications in the main text, we use a single-
index model (e.g., the CAPM), which gives us the advantages of concreteness and simplicity.
However, the optimized portfolio approach is suitable for more general environments, as
shown in the appendix for multi-index models (e.g., APT). Additionally, for the sake of
concreteness, in the main part of the paper, the optimized portfolio approach considers a
single ESG firm characteristic. However, as shown in the appendix, the optimized portfolio
approach can address numerous firm characteristics. Examples of characteristics and risk
factors other than ESG that are discussed in the literature include size, value and momentum
(e.g., Fama and French, 1993, 1997, 2007). Even with a large number of assets, risk factors
and characteristics, the optimized portfolio approach is very simple to implement.
The paper proceeds as follows: Section II discusses the relation among risk, return and
ESG and, based on this discussion, develops our portfolio optimization approach. Section
III describes a formal test as a modification of the optimized portfolio approach. Section
IV demonstrates the empirical application of portfolio optimization and the test based on
stocks from the S&P 500 index. The empirical section presents a case study with the aim
of demonstrating the application of the portfolio approach. Section V summarizes and
concludes.
II Portfolio selection and asset return model
1 Extended mean–variance optimization program
Among others, Hirschberger et al. (2013), Utz et al. (2014), Gasser et al. (2017) and Pedersen
et al. (2021) proposed the objective function (1), which assumes a sustainable investor whose

objective function with regard to portfolio formation incorporates return, risk and an ESG
characteristic.
1
(1) max αµp − λσp2 + θp
w 2
µp is the portfolio return with µp = w| µ, where w is an N × 1 vector of portfolio weights of

P
N risky assets with i wi = 1; and µ is a N × 1 vector of the expected excess returns µi over
the risk-free rate on N risky assets. The superscript | indicates the transpose operation. σp2 is
the portfolio return variance with σp2 = w| V w, where V is an N × N positive semidefinite
covariance matrix of asset returns across N firms. θp is the portfolio ESG rating with
θp = w| θ, where θ is an N × 1 vector of the ESG characteristic of each single asset.
Consistent with the literature, the additivity of the ESG characteristic is assumed (e.g.,
Drut, 2010; Gasser et al., 2017; Pedersen et al., 2021). α, λ and (α, λ, ∈ R) are scalars
that represent investor preferences for return, risk and ESG, respectively. In line with the
literature, the inputs µ, V and θ of the optimization are assumed to be given, and the
investors must specify their preferences (α, λ, ). The solution of optimization (1) for
optimal weights is given in Equation (2),
α −1 h
(2) w= V µ + V −1 θ + V −1 1
λ λ λ
where h is the Lagrangian multiplier with
−1 −1 −1

h = λ 1| V −1 1 − α1| V −1 µ 1| V −1 1 − 1| V −1 θ 1| V −1 1
2 Linking preferences to levels
This section reformulates the optimization program (1). The reformulation is motivated by
the challenges of specifying investor preferences for return, risk and ESG. Many investors are
unable to consistently specify their preferences and have difficulty quantifying the values for

α, λ, and . Without these specific quantities, the optimization program (1) will not lead to a
portfolio that maximizes investors’ utility. Instead of specifying preference values, investors
find it easier to formulate desired portfolio properties in terms of expected return, risk and
ESG. Even the determination of the preferences in experiments requires (in the first step)
that the investors state their desired levels of return, risk and ESG (Harrison and Rutström,
2008).
The idea of linking the risk aversion coefficient to the risk level is not new (Das et al.,
2010; Alexander and Baptista, 2011). The literature suggests that the attitude toward risk
can be more easily specified by using the risk level than by determining an abstract risk aver-
sion coefficient. We extend the literature by linking ESG preferences to the level of an ESG
characteristic. Specifically, we consider the levels for return and ESG as constraints of the
reformulated optimization. The constraints must satisfy the condition that the portfolio re-
turn and portfolio ESG reflect the respective preferences. Since portfolio return and portfolio
ESG are constraints, the objective function in (3) includes only the portfolio variance:
1 |
(3) min w Vw
w 2
(4) s.t. w| 1 = 1
(5) w| µ = µ∗p
(6) w| θ = θp∗
The levels of portfolio return µ∗p and portfolio ESG rating θp∗ are investor specific and chosen
by each investor individually. Intuitively, the portfolio obtained by (3)–(6) is a minimum
variance portfolio with a desired level of return µ∗p and ESG characteristic θp∗ , i.e., it is an
efficient portfolio. If the levels µ∗p and θp∗ are consistently set to the investor’s preferences α
and , the optimization program (3)–(6) is equivalent to maximization program (1). We show
in appendix A the equivalence of the desired levels of return, risk and ESG of a portfolio with
the corresponding investor’s preferences. By equivalence, we refer to our reformulation of

the approach in Equation (1), as shown by the optimization program in Equations (3)–(6).1
The reformulation of the portfolio optimization (1) allows for the preferences for return,
risk and ESG to be replaced by the desired levels of return and ESG in the determination of
a minimum variance portfolio. The reformulation enables easier adoption and allows for a
simplification of the portfolio optimization program, which in turn helps investors incorporate
ESG aspects into their portfolios. The equivalence of the optimization in (1) and (3)–(6) is
also observed by Gasser et al. (2017) and Utz et al. (2014). For example, Gasser et al. (2017)
analyze three extreme cases in which the investor has a preference for return, for ESG or for
risk.
3 Asset return models and portfolio optimization for sustainable
investors
3.1 The relation among risk, return and ESG: A qualitative description. This
section describes two strands of the literature regarding the relation among risk, return and
ESG. The aim is to introduce two views on ESG from the literature and to discuss their
consequences for portfolio formation. The first strand of the literature argues that there is
a direct relation between ESG characteristics and stock returns (e.g., Friedman and Heinle,
2016; Bolton and Kacperczyk, 2020; Li et al., 2019). For example, investors could favor
firms with high environmental standards. A second strand of the literature argues that
ESG is related to a (systematic) risk factor (see Heinkel et al., 2001; Luo and Balvers, 2017;
Pedersen et al., 2021; Pástor et al., 2021; Hoepner et al., 2021). For example, there could
be an environmental risk factor (Bolton and Kacperczyk, 2020), and firms with high (low)
loadings to the environmental risk factor offer high (low) expect returns.
The key difference between these two strands of literature is easiest to explain with an
1
The intuition behind the equivalence is that after the optimal portfolio weights from (1) are found, we
can calculate the values for the portfolio return and ESG for the optimal portfolio. If we set the values of
portfolio return and ESG as µ∗p and θp∗ in the portfolio optimization (3)–(6) and solve for optimal weights,
we will obtain the same result.

example. Suppose that Firm A has a low ESG characteristic. Firm A is a supplier to a
high ESG firm and consequently has cash flows that are highly correlated with the ESG risk
factor. Due to the correlation, the firm has a high factor loading to the ESG risk factor
with a positive risk premium. Firm B also has a high ESG characteristic and low factor
loading on the ESG risk factor. Firm B is engages in sustainability, but its cash flow does
not depend on the ESG risk factor. Except for ESG risk factor loadings and characteristics,
both firms are exposed to the same risk factors with the same factor loading values. If the
first strand of the literature is correct, then Firm B will have higher returns relative to Firm
A. If instead, the second strand of literature is the correct description of empirical returns,
then Firm A will have higher returns relative to Firm B. However, in this case, Firm A will
contribute more to portfolio risk, and consequently, investors prefer to control for this source
of the risk explicitly in their portfolio optimization, which is not necessary if the first strand
of the literature is correct. Each case has a considerable influence on portfolio formation.
Furthermore, each case will result in completely different portfolio allocations depending on
which strand of the literature an investor follows.
The first strand of the literature assumes that there is a direct relation between ESG
characteristics and asset returns (e.g., Friedman and Heinle, 2016; Barnett et al., 2020; Li
et al., 2019). According to this view, there is an ESG reward that is significantly different
from zero, and there is no ESG risk premium. The covariance matrix and hence the asset
risk is not related to the ESG characteristic. Consequently, a firm with a high value of
ESG offers different (higher or lower) returns and the same risk relative to an otherwise
identical firm with a low value of ESG. In this vein, Friede et al. (2015) conduct a meta-
analysis of 60 meta-studies including more than 2,000 single studies. They document that
90% of the empirical studies found a nonnegative relation between the ESG characteristic
and financial performance.2 Based on the list of “100 Best CSR companies in the world”, Li
2
In the subsequent text, we will assume a positive reward for the ESG characteristic. However, the ESG
reward can also be negative. For example, firms with high CO2 emissions may be avoided by investors (e.g.,
Bolton and Kacperczyk, 2020).
10

et al. (2019) construct a value-weighted portfolio that offers statistically significant annual
abnormal returns of 1.81% and 1.26% by controlling for Carhart four factors and the Fama-
French five-factors model, respectively. There are several approaches in the literature to
explain why firms with high values of ESG characteristics earn higher returns. Bénabou and
Tirole (2010) argue that the ESG activity of a firm can affect its value in three ways: (1)
firms can “do well by doing good” (e.g., reducing workplace injury); (2) firms can maximize
shareholder value when they exercise ESG behavior on behalf of stakeholders (e.g., customers
paying more for high-priced fair-trade products); and (3) firms can destroy value when
they engage in projects for the benefit of managers (e.g., managers donate to their favorite
charities to benefit privately). Edmans (2011) observe that high ESG firms may enjoy
satisfied employees, fewer environmental risks, good governance or loyal customers. Such
firms generate higher than expected earnings due to their greater efficiency. Due to the higher
earnings, these firms offer higher returns. Investors following this strand of the literature
want to include assets with ESG characteristics with a positive relation to return in their
portfolio. However, they do not need to control for ESG risk since the ESG characteristic
is assumed to have no relation to asset comovements. In portfolio optimization, we can
expect to overweight firms with high ESG characteristics because such firms contribute only
to higher portfolio returns without increasing portfolio risk.
The second strand of the literature follows the idea of traditional asset pricing and as-
sumes an ESG risk factor that governs the expected return and the covariances among assets
(see Heinkel et al., 2001; Luo and Balvers, 2017; Pedersen et al., 2021; Pástor et al., 2021;
Hoepner et al., 2021). Accordingly, there is an ESG risk premium, which is significantly
different from zero, but no ESG characteristic reward. There are different explanations for
the existence of an ESG risk factor in the literature. Bolton and Kacperczyk (2020) sug-
gest that an ESG risk factor reflects investors’ ESG preferences that affect their portfolio
formation. Investor preferences for specific ESG characteristics can lead to excess demand
for some stocks. Among others, Heinkel et al. (2001), Luo and Balvers (2017) and Hoep-
11

ner et al. (2021) argue that a permanent shift in ESG preferences will permanently change
the efficiency frontier and can lead to an ESG risk factor premium. Maiti (2021) suggests
forming risk factors based on sorts on size and ESG dimensions and finds improvement in
model performance for EURO Stoxx constituents when an ESG risk factor is added to Fama
and French (1993) three– and four–factor models. Pástor et al. (2021) and Pedersen et al.
(2021) show that a sufficiently large number of investors with preferences for ESG will lead
to risk factor premiums. Thus, as long as empirical studies do not incorporate an appro-
priate ESG risk factor, they will observe positive abnormal returns for ESG stocks while in
reality, these stocks earn a risk premium (Sharfman and Fernando, 2008; Oikonomou et al.,
2012; Albuquerque et al., 2019; Flammer, 2015). In portfolio optimization, firms with high
loadings on an ESG risk factor are not necessarily overweighted because such firms increase
portfolio return and risk simultaneously. Here, it depends on the return-risk trade-off of
whether these firms enter the portfolio. The consequences for portfolio allocation are quite
different from the first strand of the literature. In section III, we describe a formal test that
helps investors empirically decide which strand of the literature they should follow in their
portfolio formation because this heavily influences portfolio allocation.
3.2 The relation among risk, return and ESG: The formal description. This
section describes our asset return model. The introduction of the asset return model helps
to capture the discussion from the literature about the relation among return, risk and ESG.
Let R denote an N × F matrix of F observations of a system of N random variables,
representing F returns on a universe of N stocks. Our assumptions follow the literature
dealing with the estimation procedure for large covariance matrices in the portfolio context
(e.g., Ledoit and Wolf, 2003; Bodnar et al., 2018b; Plachel, 2019).
Assumption 1. The number of stocks N and the number of observations F are fixed and
finite.
Assumption 2. Asset returns have finite first and second moments.
12

Assumption 3. The return rit on an asset i at time t is described by the factor model (7).
K−1
X
(7) rit = Et−1 (rit ) + βik f˜kt + ei f˜et + uit
k=1
where βik is the loading of asset i on factor k, f˜kt ∼ N (0, σf2k ) is the factor’s k innovation
at time t, ei is the asset i’s loading on the ESG risk factor, and f˜et ∼ N (0, σf2e ) is the factor
innovation of the ESG risk factor at time t. Finally, uit ∼ N (0, σu2 ) is the residual asset
return. According to (7), there are K risk factors, where the K-th risk factor is the ESG
risk factor. The expected return is
K−1
X
(8) Et−1 (rit ) = rf t + βik µk + ei µe + θi c
k=1
where µk (µe ) is the premium on the risk factor k (ESG risk factor), θi is the ESG charac-
teristic of the asset i, and c ∈ R is the ESG reward. The covariance matrix for the asset
return is
(9) V = BV f B | + RV
where V f is the K ×K diagonal covariance matrix of orthogonal factor returns, B is a N ×K

matrix of asset factor loadings to K risk factors and RV is the N × N covariance matrix
of residual asset returns. The residual returns are identically and independently distributed
(i.i.d. residual returns) and hence Cov(ui , uj ) = 0, ∀i 6= j and Cov(ui , uj ) = σu2 , ∀i = j.3
Assumption 4. The risk factors have positive variances, that is, V ar(fk ) > 0, ∀k.
In the last subsection, we described two strands of the literature about the relation
among risk, return and ESG. Applying the equations (7)–(9), we can show the differences
3
There are some remarks on the assumption 3. For our portfolio approach, the i.i.d. residual returns are
an important feature of the asset return model (7). We only need that the risk factors explain a significant
part of the asset variances and covariances to approximate the i.i.d. assumption. The i.i.d. assumption is a
standard assumption in the literature (e.g., Pedersen et al., 2021; Daniel et al., 2020; Ledoit and Wolf, 2003).
13

between the views in the literature in a formal way. The first strand of the literature argues
that the ESG characteristic is important to investors and that there is no ESG risk factor.
Consequently, we remove the ESG risk factor from the factor model (7). Then, the expected
return for firms is
K−1
X
Et−1 (rit ) = rf t + βik µk + θi c
k=1
and the covariance σij between assets i and j is
K−1
X
σij = βik βjk σr2k + σuij
k
where σr2k describes the variance of the k-th risk factor return, σuij = 0, ∀i 6= j and σuij =
σu2 , ∀i = j. Thus, the ESG characteristic affects the expected return but not the risk.
Note that since the ESG risk factor is not included, it affects neither the returns nor the
covariances.
The second strand of the literature assumes an ESG risk factor. Accordingly, the expected
return of an asset is
K−1
X
Et−1 (rit ) = rf t + βik µk + ei µe
k=1
and the covariance σij between assets i and j is
K−1
X
σij = βik βjk σr2k + ei ej σr2e + σuij
k
where σr2e describes the variance of the ESG risk factor return. Thus, in this literature, the
asset loading to the ESG risk factor drives its expected return and risk.
3.3 Implications for portfolio formation. The asset return model (7)–(9) helps de-
velop a general portfolio optimization program without the need to estimate expected returns
or the covariance matrix. For the sake of simplicity and concreteness, in this subsection, we
rely on a return model that is described by the market risk factor (CAPM), an ESG risk
14

factor and an ESG characteristic reward.4 We start with the general setting where we allow
for an ESG reward and an ESG risk factor premium simultaneously (which might be the case
in reality), although the literature usually assumes the one or the other case. Accordingly,
the expected return is
(10) Et−1 (rit ) = µi = rf t + βi µrm + ei µe + θi c
and the covariance between the assets i and j is
(11) σij = βi βj σr2m + ei ej σr2e + σuij
For the sake of simplicity, it is assumed that the investors know the true βi and ei to avoid
addressing estimation errors in the presentation of the optimized portfolio framework. The
variance of investors’ portfolio, which is the objective function of the optimization program
(3)–(6), can be written according to the equation (11) as
σp2 = w| V w
= w| βσr2m β | + eσr2e e| + RV w

= w| βσr2m β | w + w| eσr2e e| w + w| RV w
= βp σr2m βp + ep σr2e ep + σu2 w| Iw
where β is the N × 1 vector of asset factor loadings to the market risk factor, e is N × 1
vector of asset factor loadings to the ESG risk factor and I is an N × N identity matrix.
If the portfolio factor loadings βp = N
P PN
i=1 wi βi and ep = i=1 wi ei are constrained in the
portfolio optimization to be at the investor-desired levels βp∗ and e∗p , then the terms βp σr2m βp
and ep σr2e ep from the portfolio variance calculation become constants in the optimization.
Adding a constant to the objective function will not change the optimal solution for the asset
4
The description of general cases with K risk factors and M characteristics is left to our appendix D.
15

weights. Consequently, we can remove the terms βp∗ σr2m βp∗ and e∗p σr2e e∗p from the objective
function. Additionally, the scaling of the sum of squared asset weights with the constant
residual variance (σu2 w| Iw) does not change the optimal solution for the asset weights since
σu2 is a constant and identical for all assets due to the i.i.d. assumption in the asset return
model.5 Therefore, the objective function of our optimization approach in (3)–(6) simplifies
to w| w.
Similarly, the investor-specific levels βp∗ , e∗p and θp∗ determine the portfolio return since
µ∗p =w| µ = w| (rf t 1 + βµrm + eµr + θc)
=rf t w| 1 + w| βµrm + w| eµre + w| θc
=rf t + βp µrm + ep µre + θp c
Instead of specifying the level of expected return µ∗p (see constraint (5)), we can also control
for the portfolio factor loadings to the market risk factor (i.e., βp∗ ) and to the ESG risk factor
(i.e., e∗p ) as we already control for the portfolio ESG θp∗ in the optimization program (3)–(6)
(see constraint (6)).
Thus, by assuming the asset return model (7)–(9), the portfolio optimization program
(3)–(6) results in
1 |
(12) min w w
w 2
(13) s.t. w| 1 = 1
(14) w| β = βp∗
(15) w| θ = θp∗
(16) w| e = e∗p
where the constraint (14) constrains the loading to the market risk factor βp∗ , and the con-
5
We provide sketches of the proofs for general cases with K risk factors in appendix B.
16

straint (16) constraints the portfolio loading to the ESG risk factor.
For the sake of clarity, we gather the variables on the left-hand side of the constraints
(13)–(16) into a N × 4 matrix X with
X = [1, β, θ, e]
and the variables on the right-hand side of the constraints are gathered into a 4 × 1 vector
b with
b = [1, βp∗ , θp∗ , e∗p ]|
yielding the refined optimization program in Equations (17)–(18).
1 |
(17) min w w
w 2
(18) s.t. X |w = b
The optimal portfolio weights are obtained by minimizing the Lagrange function L(w, κ) ≡
1 |
2
w w − κ| (X | w − b), where κ| is the Lagrange multiplier. The solution for the optimal
portfolio weights is given in equation (19).
(19) w| = b| (X | X)−1 X |
The optimal solution of Equation (19) differs across investors because they can choose differ-
ent values for the elements of b. Notably, vector b can account for any other investor-specific
portfolio. Thus, if vector b is
|
b = 1 0.66 0.5 1.8
the investor aims to determine a portfolio that is fully invested and exhibits a beta to
the market of 0.66, an ESG factor loading of 1.8 and an ESG portfolio characteristic of
17

0.5. The example highlights the property of the proposed portfolio optimization program to
include the combination of the strands of the discussed literature. According to our example,
the assumed investor believes that both ESG characteristics and ESG factor loading are
important to explain the variation of the empirical returns. The two described strands of
the literature assume that either ESG characteristics or ESG risk factors are important. The
portfolio formation for these investor types is described in the next subsection.
Assumptions 1.–4. in section 3.2 and the resulting asset return model (7)–(9) thus sim-
plify the optimization problem considerably. Our assumptions follow the literature that
addresses the estimation and shrinkage of large covariance matrices where the number of
assets N exceeds the number of observations F . In contrast to the literature on the reg-
ularization of covariance matrices, our approach does not improve the estimation of the
covariance matrix, but it shows that for the purpose of portfolio formation, the inclusion of
the covariance matrix in the objective function is not necessary as long as the portfolio risk
is constrained by the portfolio factor loadings.
Additionally, the reformulated optimization program (12)–(16) allows easier adaptation
because the reformulated optimization relies only on the desired levels of risk factor loadings
and ESG characteristics of the portfolio. For investors, the specification of these quantities is
often easier than the determination of the corresponding preferences. Our portfolio approach
allows investors to find an optimal portfolio for unbalanced panels and a large number of
assets because the estimation of the expected returns µ and the covariance matrix V is
circumvented. All we need is that the factor loadings can be estimated in such a way that
the residual returns are approximately i.i.d. Consequently, as long as the factor loadings can
be estimated, neither a balanced data panel with the same observation number nor a larger
number of observations than assets, i.e., N < F , is necessary.
3.4 ESG characteristic and ESG risk factor portfolios. The optimization program
(12)–(16) follows from our general asset return model (7)–(9). However, the two strands
18

of the literature assume investors that must care about either an ESG characteristic (e.g.,
CO2 emissions) or an ESG risk factor (e.g., carbon risk factor as proposed by Bolton and
Kacperczyk (2020)). For each investor type, a slightly different optimization program is
needed.
If the investors must care about an ESG characteristic but not an ESG risk factor, then
they solve
1 |
(20) min w w
w 2
(21) s.t. w| 1 = 1
(22) w| β = βp∗
(23) w| θ = θp∗
The solution for optimal weights w|θ from optimization program (20)–(23) is
(24) w|θ = b|θ (X |θ X θ )−1 X |θ
where bθ = [1, βp∗ , θp∗ ] and X θ = [1, β, θ].

If the investors must care about an ESG risk factor but not about an ESG characteristic,
then they solve
1 |
(25) min w w
w 2
(26) s.t. w| 1 = 1
(27) w| β = βp∗
(28) w| e = e∗p
19

The solution for optimal weights w|e from optimization program (25)–(28) is
(29) w|e = b|e (X |e X e )−1 X |e
where be = [1, βp∗ , e∗p ] and X e = [1, β, e].
3.5 Comparison with shrinking techniques. The observation that the expected re-
turns µ and the covariance matrix V can be estimated by risk factor models goes back at
least to Sharpe (1963). Similarly, Jacobs et al. (2005), Maillet et al. (2015), Carroll et al.
(2017) and Ledoit and Wolf (2022) demonstrate the use of factor models to estimate the
expected returns and the covariance matrix for portfolio optimization. This literature pro-
poses a two-step approach for portfolio formation with a risk factor model. In the first step,
the expected returns and the covariance matrix are estimated by applying Equation (7). In
the second step, the expected returns and the estimated covariance matrix are plugged into
the optimization programs (1) or (3)–(6), which are then solved for optimal weights. This
literature usually does not exploit a simplification of the optimization program which, as
shown above, is possible when we assume the asset return model (7)–(9).
Another strand of literature exploits the factor structure of asset returns to regularize
the covariance matrix directly when the number of assets is larger than the number of period
observations (e.g., Laloux et al., 2000; Ledoit and Wolf, 2003; Tola et al., 2008; Kolm et al.,
2014; Bodnar et al., 2018b; Plachel, 2019; Mynbayeva et al., 2022; Nguyen et al., 2022). This
strand of the literature is interested in how to reliably estimate the covariance matrix and
not in portfolio optimization. By assuming a factor model structure, we can considerably
simplify portfolio optimization for investors. The estimation of the full covariance matrix,
which is regarded in the literature as difficult (e.g., Bajeux-Besnainou et al., 2012; Maillet
et al., 2015; Daniel et al., 2020), becomes obsolete with our approach. In appendix C,
we present results from a simulation study that compares our portfolio optimization (12)–
(16) with portfolio optimizations with the same constraints but a slightly different objective
20

function including a covariance matrix, i.e., σp2 = w| V w. Specifically, the covariance matrix
is estimated with one of the following approaches: 1. simple sample estimation, 2. covariance
regularization by Ledoit and Wolf (2003), 3. covariance regularization by Ledoit and Wolf
(2004a), and 4. covariance regularization by Ledoit and Wolf (2004b). We then compare
the differences between true asset weights and estimated asset weights, where the asset
weights are calculated either by our portfolio optimization approach without a covariance
matrix or by the portfolio optimization approaches with inclusion of the covariance matrix.
The results confirm that covariance regularizations improve portfolio optimization, as the
Euclidean norm of the differences between true and estimated asset weights is smaller relative
to sample covariance estimates. However, our portfolio optimization approach performs at
least as well as the covariance regularization techniques proposed in the literature and saves
one crucial estimation step.
III The disentangling test between the competing ESG
explanations
In the last section, the asset return model (Equations (7)–(9)) can incorporate views from
the literature that assume that either ESG characteristics or ESG factor loading is impor-
tant for the cross-section of asset returns. Our portfolio optimization program (12)–(16)
even allows the investor to follow a mixed model, where ESG characteristics and ESG factor
loadings are important. From an investor’s perspective, it is empirically unclear what the
actual relation among risk, return and ESG looks like. Making matters even worse, Berchicci
and King (2020) show that the relation among return, risk and ESG is sensitive to the pe-
riod, market, data provider and methods applied in the literature. Whether the first or the
second strand of the literature, or a combination thereof, describes the expected returns in
the cross-section of asset returns thus requires an empirical clarification. However, this dis-
tinction is crucial for the specification of the optimization problem because it determines the
21

final portfolio allocation. In this section, we show how our proposed portfolio optimization
(12)–(16) can be applied to empirically decide whether the first (ESG characteristic) strand
of the literature, the second (ESG risk factor), both strands or neither are important for
asset return variation. Additionally, knowledge of the correct model can help in building
more appropriate models for return and risk management for practical portfolio manage-
ment and research. For example, suppose there is an ESG risk factor, but investors specify
the portfolio ESG characteristic and not the portfolio factor loading to the ESG risk factor.
Then, they will probably not attain the desired expected return µ∗p because they do not
consider portfolio ESG factor loading to be an important part of the expected return. The
investors assume their portfolio return is µp = βp µrm + θp c, but the actual portfolio return is
µp = βp µrm +ep µresg . Consequently, the expected portfolio returns will vary depending on the
P
portfolio ESG factor loading ( wi ei ). It is important to note that this fluctuation of the ex-
pected portfolio return is due to the unconsidered ESG risk factor. Additionally, the realized
portfolio return will fluctuate around the expected value (i.e., it has a standard deviation).
Similarly, investors do not control for a potential source of systematic risk because portfolio
variance will depend on the variance of the ESG risk factor. Since the investors’ portfolio is
exposed to ESG risk, the true portfolio variance (σp2 = βp2 σr2m + e2p σr2e ) will always be greater
than the expected portfolio variance (σp2 = βp2 σr2m ) because according to our assumption 4,
the risk factor variance is always positive.
To derive the disentangling test, for the sake of simplicity and concreteness, we rely on the
optimization program (12)–(16) and maintain the assumption of a single ESG characteristic
and two risk factors (market portfolio and an ESG risk factor). However, the presented
disentangling test is not limited to this setup. Appendix E shows the implementation of
the test for a large number of risk factors and characteristics. The solution for the optimal
weights of the optimization program (12)–(16) is given in equation (19) and repeated here
w| = b| (X | X)−1 X |
22

The solution will vary across investors depending on the choices of the elements of vector
b. For the purpose of empirical application of the disentangling test, we must consider
two additional aspects. First, we consider the time dimension of portfolio returns. Second,
we form several portfolios with varying values of factor loadings to risk factors and ESG
characteristics. Our optimization approach is sufficiently flexible to easily incorporate both
aspects. We define W t as an N × P matrix of the optimal weights of P portfolios at time
t, t ∈ {1, 2, . . . , F } and r t as the vector of realized asset returns at time t. The realized,
out-of-sample returns on the portfolios at time t are
−1
(30) λt = W |t−1 r t = D X |t−1 X t−1 X |t−1 r t
where X t is the matrix X at time t, and λt is a consistent vector of the realized returns
of each of the considered portfolios in t. D is a P × 4 matrix, the rows of which are the
right-hand sides of the optimization constraints (b| ). Equation (30) can thus determine the
returns for P different portfolios in a single step at time t. For convenience, we define a
P × F matrix Λ. Each row of Λ represents a time series of realized returns on one of P
portfolios at time t, t ∈ {1, 2, . . . , F }.
Our disentangling test builds on the idea of tracking portfolios. In our optimization
program (12)–(16), we can easily form tracking portfolios by setting one value in the vector
b to one and all other values to zero. If we set the first value of b to 1, then the tracking
portfolio has zero loading to the market risk factor and to the ESG risk factor, and its
ESG characteristic is also zero. Such a portfolio earns a risk-free rate at time t for the
investors (see equation (8)). If we set only the second (third, fourth) element of b to 1 and
the others to zero, then the portfolio tracks the market risk premium (ESG characteristic
reward, ESG risk factor premium) and is a zero-net-investment portfolio. While zero–net–
investment portfolios are rare in practical portfolio management, they are useful for our
disentangling test. Under the null hypothesis of a true risk factor model, the traditional
23

approach is to test whether the regression constant from a regression of portfolio returns
on risk factors is significantly different from zero (e.g., Li et al., 2019). By forming zero–
net–investment tracking portfolios, we can always directly test this hypothesis because the
returns on such portfolios are always without the risk-free return component of asset returns.
This is also true when asset returns are not excess returns. Consequently, the realized returns
on zero–net–investment tracking portfolios can be used directly in traditional approaches to
test whether the regression constant is significantly different from zero.6 By means of zero–
net–investment, we obtain tracking portfolio returns that are easily tested in the regression
framework.
Consequently, the choice for the matrix D as a 4 × 4 identity matrix in our disentangling
test allows the calculation of the tracking portfolio returns µ for 4 portfolios at time t7 , i.e.,
−1
(31) λt = X |t−1 X t−1 X |t−1 r t ∀t
where the elements of the 4 × 1 vector λt represent the returns of tracking portfolios for the
risk-free rate, the market risk factor realizations, the realization of the ESG characteristic
reward and the realizations of the ESG risk factor premium, respectively.
Specifically, the literature proposes applying spanning tests (e.g., Black et al., 1972;
Fama and French, 1993; Li et al., 2019; Zhu et al., 2019; Li and Sun, 2022) to test whether
(arbitrarily formed) portfolios are priced by a model. Assume that each time series of the
6
On the other hand, a full–investment tracking portfolio earns the risk-free rate and the risk premiums.
Then, in the traditional approach, we must test whether the regression constant is significantly different
from rf , i.e., the test would be b0 − rf = 0, where b0 is the regression constant.
7
Our appendix E elaborates in more detail why the identity matrix for the purpose of the disentangling
test is an appropriate choice.
24

four portfolios is regressed on the two risk factor returns rm and re
 
rp1,t = b0,p1 + b1,p1 rm,t + b2,p1 resg,t + εp1,t , . . . , ∀t
 
rp2,t = b0,p2 + b1,p2 rm,t + b2,p2 resg,t + εp2,t , . . . , ∀t
Λ=
 

(32)
r
 p3,t = b0,p3 + b1,p3 rm,t + b2,p3 resg,t + εp3,t , . . . , ∀t

 
rp4,t = b0,p4 + b1,p4 rm,t + b2,p4 resg,t + εp4,t , . . . , ∀t
= BF F | + E
where each row of Λ represents a time series of realized returns on one of the four portfolios
at time t, t ∈ {1, 2, . . . , F }; F is a F × 3 matrix of ones (constant) and two columns of
returns for the market portfolio and the ESG risk factor, respectively. B F is a 4 × 3 matrix
of the regression coefficients, i.e., the constant and two risk factor loadings b1 and b2 from the
spanning test (32) for the four portfolios. Again, the four specific portfolios are considered
for simplicity and concreteness. The appendix shows a general case with many risk factors
and/or characteristics. E is a 4 × F matrix of the random residual returns.
The important consequence from the choice of the values in D as the identity matrix is
that we know the expected values for the factor loadings from the spanning tests (32). If
the two risk factor model is valid, the expected values (E) of the coefficients in B F are
 
b
 0,p1 = r f 0 0 
 
 0 b1,p2 = 1 0 
(33) E (B F ) = 
 

b = 0 0 0 
 0,p3 
 
0 0 b2,p4 = 1
The disentanglement between the two strands of the literature regarding the relation among
return, risk and ESG is mainly determined by the values of b2,p4 and b0,p3 . Accordingly, we
expect to observe no significant coefficients on the risk factor (b1,p1 , b2,p1 , first row in B F ),
25

with the first portfolio and a regression constant being close to the risk-free rate.8 Regarding
the second portfolio, we expect a factor loading of 1 to the market portfolio returns (b1,p2 = 1)
and zero for the remaining coefficients (b0,p2 , b2,p2 ). If the first strand of the literature (ESG
characteristics) is correct, then we expect to see a significant regression constant for portfolio
three (b0,p3 ). Since the portfolio is formed to have no loadings on the market or ESG risk
factors but a positive ESG characteristic, a significant regression coefficient likely indicates
that there is a relation between return and ESG characteristic. A positive (negative) constant
b0,p3 suggests that investors are asking (bidding on) ESG stocks in excess. For the second
strand of the literature to be correct, the regression constant b0,p3 for the third portfolio
must be zero. Similarly, if the second strand of the literature (ESG factor loading) is a valid
description of asset return variations, we expect for the fourth portfolio a factor loading
of 1 to the ESG risk factor (b2,p4 = 1) and zero for the remaining coefficients (b0,p4 , b1,p4 ).
Consequently, the constant (b0,p3 ) and the risk factor factor loadings (b1,p3 , b2,p3 ) will be zero
since the portfolio is formed to have factor loadings of zero for the ESG and market risk
factors. In empirical applications, it is possible that both ESG characteristics and ESG
factor loading add to the explanation of asset returns. If such a mixed model is correct, then
we expect to see a significant regression constant b0,p3 for the third portfolio and a significant
regression coefficient for the ESG risk factor with a value of one (b2,p4 = 1) for the fourth
portfolio.
The use of our optimized portfolios as test assets has a key advantage compared to the
common use of the Black et al. (1972) test since we can provide the expected values of
the regression slope coefficients. The traditional Black et al. (1972) test focuses only on
an evaluation of the regression constant b0 . Our approach allows us to interpret both the
regression constant b0 and the slope coefficients (b1 and b2 ) for the congruent risk factors.
We refer to appendix F for a detailed description of the interpretation of the results from
our disentangling test. The appendix provides an intuition with a two–assets portfolio and
8
If we were to use excess returns, then the regression constant would be zero.
26

with a simulation study. In the simulation study, we assume two risk factors and no ESG
characteristic reward. The simulation study does not offer additional results but provides
insight into how to interpret the results from equations (32) and (33).
IV Empirical application
1 Data sample
This section aims to demonstrate the empirical application of our portfolio optimization pro-
gram in a case study. The case study also shows how to apply and interpret the results from
our disentangling test. Importantly, the aim of the case study is to demonstrate the appli-
cation and not to decide whether the cross-section of asset returns is described by an ESG
characteristic or by an ESG risk factor. The case study assumes extended mean–variance
investors who are seeking to maximize their utility. These investors will first apply the dis-
entangling test to analyze whether there is a relation of asset returns to ESG characteristics
and/or to ESG factor loading. After the type of relation is determined, these investors form
their portfolios accordingly.
For the case studies, we rely on the database “Refinitiv ASSET4” (ASSET4) to acquire an
independent external ESG measure (e.g., Gasser et al., 2017; Hoepner et al., 2021). ASSET4
offers more than 250 key indicators of environmental, social and governance performance.
The key indicators from these areas are aggregated into an overall ESG measure (TESGS).
We follow the literature and use the aggregated ESG measure as the ESG characteristic of
an asset (e.g., Gasser et al., 2017; Pedersen et al., 2021). The ESG characteristic has values
ranging between 0 and 100, indicating the lowest and highest scores, respectively. Due to
the demonstrative character of our case study to show the application of our portfolio opti-
mization and the disentangling test, we assume that the ESG characteristics from ASSET4
correctly indicate the ESG level of a firm. However, Berg et al. (2022) show that the cor-
relation between ESG characteristics from different data providers is low; hence, the results
27

from empirical studies, including our case studies, are not comparable if they use different
data providers.
We conduct our empirical study with the constituents of the S&P 500 index. The data
of firms from the S&P 500 index are obtained from “Refinitiv Datastream”. To obtain the
ESG characteristics, we match firms from the S&P 500 index that have data in the ASSET4
database. We acquire the monthly constituent list, the monthly total return index and
the aggregated ESG characteristics via the Refinitiv Datastream from December 1989 to
December 2019. The monthly asset excess returns are calculated based on the total return
indices, which include dividend payments and the risk-free rate. The time series length in
our final data sample is limited due to the availability of ESG data. The ESG characteristics
are available from January 2001 to December 2019.
The constituents of the S&P 500 index are regularly updated because the index includes
shares of the 500 largest US firms by market valuation. The number of firms in the index
in a given month is always 500. As the constituents of the index change over time, our data
sample becomes unbalanced. However, our optimized portfolio approach does not require
balanced data samples because at time t, we do not need the covariance matrix, which is
otherwise estimated over the last t − H observations, whereas H typically indicates the last
5 years. It requires to include the factor loadings to risk factors and the ESG characteristics.
The ESG characteristic is known at time t, and the factor loadings at time t can be estimated
for each asset even if the time series of the observations do not have the same length. The
solution for our portfolio optimization program requires the data on ESG characteristics and
on factor loadings to be available at time t to calculate the portfolio return at time t + 1.
Consequently, the number of assets in the portfolio may vary over time due to the data
availability of the ESG characteristic. Our final sample consists of 686 individual assets with
an ESG rating that were members of the S&P 500 index during the 2001–2019 period. The
data sample includes assets without a full time series of observations. Table 1 classifies the
stocks included in the data sample by ESG score levels and provides the descriptive statistics
28

of the returns. According to Table 1, numerous assets are assigned ESG scores between 21
and 80. The return-risk reward (µ/σ) increases with the ESG characteristic because the
portfolio risk decreases as the values of the ESG characteristic increase.
Table 1: Descriptive statistics of monthly returns of assets at different ESG score levels in %. µ is the simple average
return; σ is its standard deviation; P(25) and P(75) report the 25th and 75th percentiles, respectively; and N is the time series
average number of assets in each portfolio. The statistics are calculated for equally weighted portfolios, monthly sorted on ESG
characteristics into portfolios with ESG ranges as indicated in the first column; i.e., the portfolio for the ESG range of 0–20
can change every month since the firm’s ESG characteristic can become less than or greater than 20. Then, we calculate the
average return in month t for the portfolio and repeat the exercise until the last period. The procedure results in a time series
of monthly portfolio returns.
ESG Median µ σ P(25) P(75) µ/σ N

0-20 1.53 1.14 6.69 -2.09 5.12 0.170 5
21-40 1.23 1.02 5.11 -1.51 3.55 0.199 131
41-60 1.21 0.97 4.81 -1.30 3.67 0.202 238
61-80 1.29 1.08 4.71 -1.00 3.55 0.229 258
81-100 1.36 1.01 4.19 -0.95 3.46 0.241 54
We assume a two-risk factor world. The risk factors are a market portfolio and an ESG
risk factor. Regarding the market portfolio, we define its returns as the value-weighted
returns of all assets available at the time of portfolio formation. Regarding the ESG risk
factor, Maiti (2021) proposes forming a zero-net-investment portfolio in line with the factor
construction of Fama and French (1993). In line with Maiti (2021), in each month, the
ESG risk factor-mimicking portfolio goes long into a high-ESG portfolio and short into a
low-ESG portfolio. The high-ESG portfolio consists of assets with an ESG characteristic
greater than the 75th percentile in the month of portfolio formation, and its return is a
simple average of the asset returns. Similarly, the low-ESG portfolio consists of assets with
an ESG characteristic less than the 25th percentile in the month of portfolio formation, and
its monthly return is a simple average of the asset returns.
Table 2 reports the descriptive statistics of our two risk factors, ESG characteristics
θ, factor loadings to the market portfolio β and ESG risk factor e. The mean µ and the
standard deviation σ are time series values in % for the risk factors (rm and rESG ) and
panel data values for the asset- and time-specific θ, β and e. The factor loadings β and
e are calculated in each month in simple time series regressions based on the previous five
29

Table 2: Descriptive statistics of the risk factor returns, ESG characteristics, and factor loadings to the risk factors. The
return values are in %. rm is the return on the market portfolio, rESG is the return on the ESG risk factor, θ is the firm’s ESG
characteristic, β is the firm’s factor loading to the market return, and e is the firm’s factor loading to the ESG risk factor. µ is
the average, while σ is its standard deviation. The correlations are calculated for the time series (rm , rESG ) and in panel (θ,
β, e) variables.
rm rESG θ β e
µ 0.74 0.15 58.94 1.08 -0.49
σ 4.06 1.72 17.75 0.51 1.30
Correlation of the risk factor time series
rm 1.00
rESG -0.35 1.00
Correlation of factor loadings and ESG characteristic
θ β e
θ 1.00
β -0.03 1.00
e 0.16 -0.26 1.00
years of observations. The monthly average return of 0.74% and its standard deviation
of 4.06% are considerably higher for the market portfolio relative to the ESG risk factor
(0.15% and 1.72%, respectively). The average firm in our sample has an ESG characteristic
greater than 50, a factor loading to the market portfolio of approximately 1 and a negative
factor loading to the ESG risk factor of approximately −0.5. The time series correlation
between the variables is calculated for the two risk factors and is found to be negative. The
panel correlations are calculated for the asset- and time-specific variables θ, β and e and
are found to be low, particularly the correlation between the ESG characteristic and ESG
factor loading. The correlation values suggest that the ESG characteristic does not simply
translate into a specific value of ESG factor loading. For example, assets with high ESG
values do not automatically have high ESG factor loadings.
2 Case study
This subsection presents the application of the disentangling test and of the portfolio op-
timization program. In the first step, we will test the relation of asset returns to ESG
30

characteristics and to ESG factor loading by applying our disentangling test, which is pre-
sented in section III. In the second step, depending on the results from the disentangling
test, we will form the portfolios either according to the first strand of the literature (ESG
characteristic) or the second strand of the literature (ESG factor loading) or a combination
of both strands of the literature or neither of them.
Table 3 reports the results from the first step, which is the disentangling test shown in
equation (32). In each month, we determine the weights of the out-of-sample excess returns
on four portfolios in the subsequent month (as specified by Equation (30)).
Table 3: Results of the test of the four portfolios with varying b from Equation (30) based on empirical data. The table
columns b0 , b1 , b2 and R2 show the estimated OLS coefficients from the regression (32), where the portfolio excess returns are
regressed on the risk factors. The values of the regression coefficients marked in bold are statistically significantly different from
zero at the 5% level and are calculated with HAC-robust standard errors. The standard errors are Newey-West corrected with
a lag number of 4. R2 is the adjusted R2 .
b0 b1 b2 R2
rp1 0.00 -0.38 -2.26 0.10
rp2 -0.01 0.78 0.09 0.72
rp3 0.00 0.01 0.03 0.12
rp4 0.00 0.01 0.54 0.52
Table 3 shows the coefficients from a time series regression of optimized portfolio excess
returns on the two risk factors. The reported coefficients in the table are OLS coefficients,
but the standard errors are corrected for heteroscedasticity and autocorrelation (HAC cor-
rected). Due to our disentangling optimization approach, we know the expected values of
the regression coefficients. If the first strand of the literature is in line with the data, then we
expect to see a significant regression constant (b0 ) for portfolio 3 and insignificant regression
coefficients b2 for portfolios 3 and 4. If the second strand of the literature correctly describes
the asset return variation, then the risk factor model must be true, and accordingly, we
expect to observe significant coefficients close to one for the market portfolio (b1 ) and the
ESG risk factor (b2 ) in portfolios 2 and 4, respectively, and insignificant regression constants
(b0 ). A mix of the results would indicate that a mixed model of ESG characteristics and
ESG factor loading could be true. If no regression coefficients for portfolios 3 and 4 are
31

significant, neither strand of the literature with respect to ESG is a correct description of
asset return variation.
There are three notable results of our test. First, the regression constant of portfolio 2
is significant, which suggests that the risk factor model cannot fully price portfolio returns.
However, the constant is economically less relevant. Second, the returns on portfolios 1
(zero–beta portfolio) and 3 (ESG characteristic portfolio) are not significantly exposed to the
market risk factor. However, both portfolios are somewhat exposed to the ESG risk factor.
Interestingly, the factor loading of portfolio 1 to the ESG factor is high and insignificant,
while the factor loading of portfolio 3 to the ESG factor is significant but economically
rather low. Third, the returns on portfolios 2 and 4 are exposed to congruent risk factors
and significantly different from zero. However, our disentangling test suggests that the factor
loadings are expected to be 1. Since the (untabulated) HAC standard errors are 0.08 and
0.14 for the coefficient to the market factor for portfolio 2 (b1 ) and the coefficient to the
ESG risk factor for portfolio 4, respectively, only the former is statistically insignificantly
different from 1. The coefficient b2 for the ESG risk factor is significantly less than 1, which
is not in line with our expectation. According to the disentangling test, the coefficient must
be insignificantly different from 1. There are several explanations for the result. First,
an explanation for the lower than 1 value might be that the optimization relies on the
estimated factor loadings e, which are subject to the errors-in-variable problem, i.e., they
are not perfectly correlated with the true factor loadings. Second, the reason for the result
might be our formation of the ESG risk factor. We form the ESG risk factor in line with the
literature, but it may not fully capture an ESG risk premium. Third, our model includes
only two risk factors, the market and ESG risk factors. There could be additional important
risk factors or characteristics that we did not include in our illustrative case study.
The results of the disentangling test suggest that the ESG characteristics do not add to
the explanation of stock returns in the cross-section. However, the two risk factor models
cannot fully explain the time series of the estimated risk factor premiums. Nevertheless,
32

notably, the empirical investigation presented here is only of a demonstrative nature. In
a real-world application, additional tests would be necessary to determine the additional
drivers of the asset return variation. For our illustrative example, we will refrain from
conducting additional tests and assume that the ESG characteristics do not add to the
explanation of the return variation in our sample and that the risk factor model performs
sufficiently well for our purpose. Accordingly, extended mean–variance investors will follow
the second strand of the literature and consider the ESG factor loading only in the portfolio
formation.
Table 4 reports descriptive statistics for three portfolio returns. The portfolios are formed
according to the optimization program (25)–(28). The three portfolios are for illustrative
purposes only. In portfolios one to three, the loadings to the ESG risk factor are arbitrarily
set to −1.25, −0.5 and 0.25. Then, in each month, we determine the weights of three
portfolios according to equation (29) and calculate the out-of-sample return in the next
month. We calculate the weighted average return according to equation (30) and repeat the
calculations for all months t = 1, 2, . . . , F − 1. We obtain for each portfolio a time series
of portfolio returns. These time series of portfolio returns are used to calculate descriptive
statistics in Table 4.
Table 4: Descriptive statistics of the monthly returns of portfolios for investors following second strand of the literature at
different ESG risk factor factor loadings in %. “P” is the number of the portfolio. “Budget” is the sum of the portfolio weights,
β is the factor loading to the market factor and e is the factor loading to the ESG risk factor. “Budget”, β and e define the
investor-specific portfolio. µ, σ and “SR” are the mean, standard deviation and Sharpe ratio of the out-of-sample portfolio
returns, respectively.
A. Investors’ setting (pre-formation) B. Portfolio statistics (post-formation)

P Budget β e µ σ SR
1 1 0.66 -1.25 0.73 3.47 0.21
2 1 0.66 -0.5 0.77 3.25 0.24
3 1 0.66 0.25 0.82 3.08 0.27
Part A of Table 4 reports the setting of the portfolio optimization. Part B of Table
4 shows the out-of-sample descriptive statistics of the portfolio returns. We present the
monthly average out-of-sample return (column µ), the standard deviation of the portfolio
33

return (column σ) and the Sharpe ratio (column SR), which is calculated as µ/σ. In Table
4, all of the portfolios are set to be fully invested (column Budget) and have an arbitrarily
chosen beta of 0.66 (column β). We assume risk-averse investors who are only willing to
accept portfolio risks less than the market risk. The portfolios differ in how the factor
loadings to the ESG risk factor (column e) are set. While these values are chosen arbitrarily,
the ascending order of the factor loadings to the ESG risk factor can confirm by example the
relation between the ESG factor loading and portfolio return from the disentangling test.
We observe that the portfolio return (column µ) increases with increasing values of the
ESG factor loading from portfolios 1 to 3. Based on the results from the disentangling test,
the increasing portfolio return is indeed not surprising. Similarly, the portfolio risk decreases
from portfolios 1 to 3. Because the market factor loading is held constant and the portfolio
variances increase with the squared value of the factor loading (i.e., V ar(rp ) = βp2 σr2m +e2p σr2e ),
the result is plausible and is in line with the observation from the disentangling test that the
returns are exposed, however not perfectly, to the ESG risk factor.
Our case study and the presented results, which are purely demonstrative, reflect the
ongoing discussion in the academic literature. Our results indicate that the inclusion of ESG
aspects in portfolio optimization may be beneficial for the out-of-sample return and Sharpe
ratio. However, our results from the disentangling test are inconclusive. The results do not
support the first strand of the literature, where ESG characteristics are important for the
explanation of the asset returns. The results are only weakly in favor of the second strand
of the literature, which assumes the existence of an ESG risk factor. More research than
an illustrative case study is needed to determine whether the ESG characteristic, the factor
loading to an ESG risk factor, a mixture of the two or neither explains the asset returns.
This is all the more important as Berchicci and King (2020) show that the relation among
return, risk and ESG is sensitive to the period, market, data provider and methods applied
in the literature.
34

V Conclusion
This paper makes important contributions to the literature. First, motivated by the increas-
ing popularity of ESG aspects in the investment fund- and banking industries, we introduce
and demonstrate an optimized portfolio approach that has technical and practical advantages
over the incorporation of ESG into the traditional mean–variance approach. In particular,
we link the preference coefficients to return, risk and ESG to levels of return, risk and ESG.
The latter are considered easier to understand by investors relative to preference coefficients.
Additionally, based on the assumption of i.i.d. residual returns, we show how portfolio opti-
mization is conducted without the estimation of the covariance matrix. Second, based on the
proposed portfolio formation, we introduce a simple test that allows investors to distinguish
between competing explanations for the relation among ESG, risk and return.
The two-risk factor model and a single (ESG) characteristic are used for concreteness
and simplicity. Our optimized portfolio approach is suitable for more general environments,
as shown in the appendix for arbitrage pricing theory (APT). In fact, the optimized port-
folio approach can address a very large number of (arbitrarily chosen) firm characteristics.
Even with a large number of assets, risk factors and characteristics, the optimized portfolio
approach is very simple to implement. Additionally, our approach can easily be adopted
to create factor tracking portfolios (by setting the respective factor loading constraint to
one), factor neutral portfolios (by setting the respective factor loading constraint to zero),
style tracking portfolios (by setting the respective characteristic constraint to one), style
neutral portfolios (by setting the respective characteristic constraint to zero) or mixtures
of these approaches. Another appealing feature of our optimization approach is that an
analytical solution can be derived. However, numerical solutions will have to be adopted in
cases in which additional weight constraints or a subset selection must be considered. Our
test can also help with the more general debate (e.g. Fama and French, 1993; Daniel and
Titman, 1997; Daniel et al., 2020; Davis et al., 2000) regarding whether the factor loadings
35

to risk factors (such as SMB or HML) or congruent characteristics (such as firm size and the
book-to-market-equity ratio) explain returns.
The models and corresponding portfolio optimizations presented in our paper consider a
broad range of empirical and theoretical literature. However, there are still limitations to our
approach. There is empirical evidence that tail-risk measures are related to ESG character-
istics since companies with high ESG characteristics are less vulnerable to company-specific
negative events (e.g., Diemont et al., 2016; Krüger, 2015). Since portfolio optimization
assumes an investor who accepts variance as the appropriate risk measure, our approach
cannot account for alternative risk measures (e.g., expected shortfall, value at risk).
References
Albuquerque, R., Koskinen, Y., and Zhang, C. (2019). Corporate social responsibility and
firm risk: Theory and empirical evidence. Management Science, 65(10):4451–4469.
Alexander, G. J. and Baptista, A. M. (2011). Portfolio selection with mental accounts and
delegation. Journal of Banking & Finance, 35(10):2637–2656.
Bajeux-Besnainou, I., Bandara, W., and Bura, E. (2012). A krylov subspace approach to
large portfolio optimization. Journal of Economic Dynamics and Control, 36(11):1688–
1699.
Ballestero, E., Bravo, M., Pérez-Gladish, B., Arenas-Parra, M., and Plà-Santamaria, D.
(2012). Socially responsible investment: A multicriteria approach to portfolio selection
combining ethical and financial objectives. European Journal of Operational Research,
216(2):487–494.
Barnett, M., Brock, W., and Hansen, L. (2020). Pricing uncertainty induced by climate
change. Review of Financial Studies, 33(3):1024–1066.
36

Bénabou, R. and Tirole, J. (2010). Individual and corporate social responsibility. Economica,
77(305):1–19.
Berchicci, L. and King, A. (2020). Evidence on social and financial performance: mapping the
empirical garden of forking paths. Academy of Management Proceedings, 2020(1):17546.
Berg, F., Kölbel, J. F., and Rigobon, R. (2022). Aggregate confusion: The divergence of esg
rating. Review of Finance, pages 1–30.
Bian, Z., Liao, Y., O’Neill, M., Shi, J., and Zhang, X. (2020). Large-scale minimum vari-
ance portfolio allocation using double regularization. Journal of Economic Dynamics and
Control, 116:103939.
Bilbao-Terol, A., Arenas-Parra, M., and Cañal-Fernández, V. (2012). Selection of socially

responsible portfolios using goal programming and fuzzy technology. Information Sciences,
189:110–125.
Bilbao-Terol, A., Arenas-Parra, M., Cañal-Fernández, V., and Bilbao-Terol, C. (2013). Se-
lection of socially responsible portfolios using hedonic prices. Journal of Business Ethics,
115(3):515–529.
Black, F., Jensen, M., and Scholes, M. (1972). The capital asset pricing model: Some
empirical tests. In Jensen, M., editor, Studies in the theory of capital markets. Praeger
Publishers Inc.
Bodnar, T., Okhrin, Y., Vitlinskyy, V., and Zabolotskyy, T. (2018a). Determination and
estimation of risk aversion coefficients. Computational management science, 15(2):297–
317.
Bodnar, T., Parolya, N., and Schmid, W. (2018b). Estimation of the global minimum vari-
ance portfolio in high dimensions. European Journal of Operational Research, 266(1):371–
390.
37

Bollerslev, T., Gibson, M., and Zhou, H. (2011). Dynamic estimation of volatility risk
premia and investor risk aversion from option-implied and realized volatilities. Journal of
econometrics, 160(1):235–245.
Bolton, P. and Kacperczyk, M. (2020). Do investors care about carbon risk? Technical
report, National Bureau of Economic Research, Cambridge, MA.
Carroll, R., Conlon, T., Cotter, J., and Salvador, E. (2017). Asset allocation with correlation:
A composite trade-off. European Journal of Operational Research, 262(3):1164–1180.
Daniel, K., Mota, L., Rottke, S., and Santos, T. (2020). The cross-section of risk and returns.
The Review of Financial Studies, 33(5):1927–1979.
Daniel, K. and Titman, S. (1997). Evidence on the characteristics of cross sectional variation
in stock returns. The Journal of Finance, 52(1):1–33.
Das, S., Markowitz, H., Scheid, J., and Statman, M. (2010). Portfolio optimization with
mental accounts. Journal of financial and quantitative analysis, 45(2):311–334.
Davis, J. L., Fama, E. F., and French, K. R. (2000). Characteristics, covariances, and average
returns: 1929 to 1997. The Journal of Finance, 55(1):389–406.
Diemont, D., Moore, K., and Soppe, A. (2016). The downside of being responsible: Corporate
social responsibility and tail risk. Journal of Business Ethics, 137(2):213–229.
Drut, B. (2010). Sovereign bonds and socially responsible investment. Journal of Business
Ethics, 92(1):131–145.
Edmans, A. (2011). Does the stock market fully value intangibles? Employee satisfaction
and equity prices. Journal of Financial Economics, 101(3):621–640.
Fama, E. F. and French, K. R. (1993). Common risk factors in the returns on stocks and
bonds. Journal of Financial Economics, 33(1):3–56.
38

Fama, E. F. and French, K. R. (1997). Industry costs of equity. Journal of Financial
Economics, 43(2):153–193.
Fama, E. F. and French, K. R. (2007). Disagreement, tastes, and asset prices. Journal of
Financial Economics, 83(3):667–689.
Flammer, C. (2015). Does corporate social responsibility lead to superior financial perfor-
mance? A regression discontinuity approach. Management Science, 61(11):2549–2568.
Friede, G., Busch, T., and Bassen, A. (2015). Esg and financial performance: aggregated
evidence from more than 2000 empirical studies. Journal of Sustainable Finance & In-
vestment, 5(4):210–233.
Friedman, H. L. and Heinle, M. S. (2016). Taste, information, and asset prices: Implications
for the valuation of CSR. Review of Accounting Studies, 21(3):740–767.
Gasser, S. M., Rammerstorfer, M., and Weinmayer, K. (2017). Markowitz revisited: Social
portfolio engineering. European Journal of Operational Research, 258(3):1181–1190.
Gillan, S. L. and Starks, L. T. (2000). Corporate governance proposals and shareholder

activism: The role of institutional investors. Journal of Financial Economics, 57(2):275–
305.
Gillan, S. L. and Starks, L. T. (2007). The evolution of shareholder activism in the United
States. Journal of Applied Corporate Finance, 19(1):55–73.
Grewal, J., Serafeim, G., School, A. Y. H. B., and 2016, U. (2016). Shareholder activism on
sustainability issues. Technical report, Harvard Business School, Boston, MA.
Harrison, G. W. and Rutström, E. E. (2008). Risk aversion in the laboratory. In Risk

aversion in experiments. Emerald Group Publishing Limited.
Heinkel, R., Kraus, A., and Zechner, J. (2001). The effect of green investment on corporate
behavior. Journal of Financial and Quantitative Analysis, 36(4):431–449.
39

Hirschberger, M., Steuer, R. E., Utz, S., Wimmer, M., and Qi, Y. (2013). Computing the
nondominated surface in tri-criterion portfolio selection. Operations Research, 61(1):169–
183.
Hoepner, A. G. F., Oikonomou, I., Sautner, Z., Starks, L. T., and Zhou, X. (2021). ESG
shareholder engagement and downside risk. Technical report, SSRN.
Jackwerth, J. C. (2000). Recovering risk aversion from option prices and realized returns.
The Review of Financial Studies, 13(2):433–451.
Jacobs, B. I., Levy, K. N., and Markowitz, H. M. (2005). Portfolio optimization with factors,
scenarios, and realistic short positions. Operations Research, 53(4):586–599.
Kolm, P. N., Tütüncü, R., and Fabozzi, F. J. (2014). 60 years of portfolio optimization: Prac-
tical challenges and current trends. European Journal of Operational Research, 234(2):356–
371.
Krüger, P. (2015). Corporate goodness and shareholder wealth. Journal of Financial Eco-
nomics, 115(2):304–329.
Laloux, L., Cizeau, P., Bouchaud, J.-P., and Potters, M. (1999). Noise dressing of financial
correlation matrices. Physical review letters, 83(7):1467.
Laloux, L., Cizeau, P., Potters, M., and Bouchaud, J.-P. (2000). Random matrix the-
ory and financial correlations. International Journal of Theoretical and Applied Finance,
3(03):391–397.
Ledoit, O. and Wolf, M. (2003). Improved estimation of the covariance matrix of stock
returns with an application to portfolio selection. Journal of empirical finance, 10(5):603–
621.
Ledoit, O. and Wolf, M. (2004a). Honey, i shrunk the sample covariance matrix. The Journal
of Portfolio Management, 30(4):110–119.
40

Ledoit, O. and Wolf, M. (2004b). A well-conditioned estimator for large-dimensional covari-
ance matrices. Journal of multivariate analysis, 88(2):365–411.
Ledoit, O. and Wolf, M. (2022). The power of (non-) linear shrinking: A review and guide
to covariance matrix estimation. Journal of Financial Econometrics, 20(1):187–218.
Li, F. W. and Sun, C. (2022). Information acquisition and expected returns: Evidence from
edgar search traffic. Journal of Economic Dynamics and Control, page 104384.
Li, Z., Minor, D. B., Wang, J., and Yu, C. (2019). A learning curve of the market: Chasing al-
pha of socially responsible firms. Journal of Economic Dynamics and Control, 109:103772.
Luo, H. A. and Balvers, R. J. (2017). Social screens and systematic investor boycott risk.
Journal of Financial and Quantitative Analysis, 52(1):365–399.
Maillet, B., Tokpavi, S., and Vaucher, B. (2015). Global minimum variance portfolio opti-
misation under some model risk: A robust regression-based approach. European Journal
of Operational Research, 244(1):289–299.
Maiti, M. (2021). Is esg the succeeding risk factor? Journal of Sustainable Finance &
Investment, 11(3):199–213.
Mynbayeva, E., Lamb, J. D., and Zhao, Y. (2022). Why estimation alone causes markowitz
portfolio selection to fail and what we might do about it. European Journal of Operational
Research, 301(2):694–707.
Nguyen, V. A., Kuhn, D., and Mohajerin Esfahani, P. (2022). Distributionally robust in-
verse covariance estimation: The wasserstein shrinkage estimator. Operations Research,
70(1):490–515.
Oikonomou, I., Brooks, C., and Pavelin, S. (2012). The impact of corporate social per-
formance on financial risk and utility: A longitudinal analysis. Financial Management,
41(2):483–515.
41

Pástor, L., Stambaugh, R. F., and Taylor, L. A. (2021). Sustainable investing in equilibrium.
Journal of Financial Economics, 142(2):550–571.
Pedersen, L. H., Fitzgibbons, S., and Pomorski, L. (2021). Responsible investing: The
esg-efficient frontier. Journal of Financial Economics, 142(2):572–597.
Plachel, L. (2019). A unified model for regularized and robust portfolio optimization. Journal
of Economic Dynamics and Control, 109:103779.
Shanken, J. (1992). On the estimation of beta-pricing models. Review of Financial Studies,

5(1):1–55.
Sharfman, M. P. and Fernando, C. S. (2008). Environmental risk management and the cost
of capital. Strategic Management Journal, 29(6):569–592.
Sharpe, W. F. (1963). A simplified model for portfolio analysis. Management science,

9(2):277–293.
Steuer, R. E. and Utz, S. (2022). Non-contour efficient fronts for identifying most preferred
portfolios in sustainability investing. European Journal of Operational Research.
Tola, V., Lillo, F., Gallegati, M., and Mantegna, R. N. (2008). Cluster analysis for portfolio
optimization. Journal of Economic Dynamics and Control, 32(1):235–258.
Utz, S., Wimmer, M., Hirschberger, M., and Steuer, R. E. (2014). Tri-criterion inverse
portfolio optimization with application to socially responsible mutual funds. European
Journal of Operational Research, 234(2):491–498.
Zhu, Z., Duan, X., Sun, L., and Tu, J. (2019). Momentum and reversal: The role of short
selling. Journal of Economic Dynamics and Control, 104:95–110.
42

Appendices
A Sketches of proofs for lining preferences to levels
1 A motivating example
The equivalence of preference formulation and level formulation for the portfolio optimization
is easiest to show starting with the traditional Markowitz portfolio optimization as presented
in optimization program (34).
1
(34) max w| µ − λw| V w
w 2
s.t. w| 1 = 1
where µ is a N × 1 vector of expected returns, V is a N × N covariance matrix, and w is a

N × 1 vector of asset weights in the portfolio. 1 is a N × 1 vector of ones, and λ is a scalar.
−1
The solution for the optimal weights is w∗ = λ1 V −1 µ − 1| V −1 µ 1| V −1 1 V −1 1 . The

portfolio return is
1 | −1 −1 | −1
µp = µ| w∗ = µ V µ − 1| V −1 µ 1| V −1 1

(35) µ V 1
λ
After the expected returns µ and the covariance matrix V are estimated, the investors’
choice of the risk aversion parameter λ determines the expected return on the portfolio µp .
We can conversely ask investors for the desired portfolio return µ∗p and then solve the
equation (35) with respect to the risk aversion parameter, i.e.,
µ| V −1 µ
λ= −1
µ∗p + 1| V −1 µ 1| V −1 1 µ| V −1 1

The key insight is that the risk aversion parameter determines the expected portfolio return
43

because µ and V are estimated, known values. However, the desired return level µ∗p reveals
the risk preference and the risk aversion parameter of investors. Consequently, the Markowitz
risk aversion parameter is related to the choice of the desired return of the investor and vice
versa.
2 ESG, return and risk preferences
Analogous to the traditional Markowitz optimization, there is a relation between the desired
levels of return, risk and ESG of a portfolio and the respective preference values. To illustrate
this point, based on the solution (2), we can write the expected excess return of an investor’s
portfolio µ∗p according to Equation (36).
α | −1 h
(36) µ∗p = µ| w = µ V µ + µ| V −1 θ + µ| V −1 1
λ λ λ
Similarly, the portfolio ESG characteristic is
α | −1 h
(37) θp∗ = θ | w = θ V µ + θ | V −1 θ + θ | V −1 1
λ λ λ
and the portfolio variance is
1
(38) σp2∗ = w| V w = 2
(αµ + θ + h1)| V −1 (αµ + θ + h1)
λ
As before, the Lagrange multiplier is
−1 −1 −1

h = λ 1| V −1 1 − α1| V −1 µ 1| V −1 1 − 1| V −1 θ 1| V −1 1
and depends only on the choices of α, λ and .

The key insight is that if an investor specifies the desired levels of return, risk and ESG for
a portfolio, i.e., the left-hand side of Equations (36)–(38), then the corresponding preferences
44

are also determined. It follows that the investor’s consistent indication of both, either her
preferences or her desired level for return, risk and ESG, will lead to the same portfolio.
Equations (36)–(38) can be used to solve for three unknown parameters based on the levels
of return, risk and ESG or based on the corresponding preferences. The preferences (α, λ, )
and portfolio properties (µ∗p , σp2∗ , θp∗ ) must be consistent because otherwise, the portfolio
would not be optimal.
For our portfolio optimization program, the investor only needs to specify µ∗p and θp∗
because we minimize the portfolio variance. An optimal portfolio for investors must have
minimum variance for a given level of return and ESG. The minimum variance is provided
by our objective function.
B Sketches of the proofs for removing the covariances
and expected returns from the optimization prob-
lem (12)–(15)
We assume that the F × N matrix of F asset return observations R can be decomposed into
systematic and idiosyncratic parts, e.g., according to factor models or by means of principal
component analysis. Equation (39) shows the decomposition
(39) R = F B| + E
where F is the F × K matrix of returns for the K factor and B is the N × K matrix of
factor loadings. E is the F × N matrix of i.i.d. residual returns. The covariance matrix is
(40) V = BV f B | + RV
45

where V f is the K × K diagonal covariance matrix of factor returns and RV is the N × N
diagonal covariance matrix of i.i.d. residual returns. The diagonal elements of RV have the
same value due to the i.i.d. assumption, i.e., var(εi ) = σε2 . The variance of the portfolio is
σp2 = w| (BV f B | + RV ) w
= w| BV f B | w + w| RV w
= b| V f b + σε w| Iw
where I is the N × N identity matrix and b is a K × 1 vector, whose elements represent

factor loading values of the portfolio. For our optimization program, the values in b are
prespecified and constant; hence, the optimal portfolio weights for the minimum variance
portfolio are obtained by minimizing the sum of squared weights, i.e.,
(41) min w| w
s.t. w| 1 = 1
w | B = b|
C Comparison between our approach and covariance
shrinking approaches
Our portfolio optimization approach does not require the estimation of the covariance matrix.
Several approaches have been proposed in the literature to regularize the covariance matrix.
This section presents the results from a simulation study comparing our approach with the
popular approaches proposed by Ledoit and Wolf (2003), Ledoit and Wolf (2004a) and Ledoit
and Wolf (2004b). We conduct the simulation study following the next steps:
1. Simulate a N × 1 vector of asset factor loading β with βi ∼ N (1, 0.5), i = 1, . . . , N and

N ∈ {50, 100, 200, 500, 1000}.
46

2. Simulate a N × 1 vector of asset characteristics θ with θi ∼ N (0, 1), i = 1, . . . , N and
N ∈ {50, 100, 200, 500, 1000}
3. Simulate a F × 1 vector of factor returns f ∼ N ( 0.08
12
, √0.2
12
) with F = 120
4. Simulate asset return matrix R with rit = βi ft + θi c + εit , where c = 0.04/12 and
V ar(ε) ∼ N (0.025, s) with s ∈ {0.00, 0.01, 0.02} and Cov(εi , εj ) = 0. If V ar(εi ) < 0,
then we set V ar(εi ) = 0.025.
5. Estimate the factor exposure β̂i based on the sample observations from steps 2–4.
6. Calculate the covariance V meth , where meth indicates the estimation method. We apply
the following: 1. the simple sample estimation (V sample ),9 2. covariance regularization
by Ledoit and Wolf (2003) (V L1 ), 3. covariance regularization by Ledoit and Wolf
(2004a) (V L2 ), 4. covariance regularization by Ledoit and Wolf (2004b) (V L3 ), 5. true
p
covariance matrix applying true values of β and σf = 0.2/ (12).
7. Find the optimal asset weights applying each covariance matrix from step 5 in the new
optimization program (12)–(16) and for the other optimization approach with the same
constraints but where the covariance matrix is included in the objective function (i.e.,
σp2 = w| V w). The factor loading of the portfolio is set to 1 (βp∗ = 1), the portfolio
characteristic is set to zero (θp∗ = 0) and the sum of portfolio weights must be 1.
8. Calculate the Euclidean norm for the difference between true asset weighs (calculated
with true covariances and true betas) and calculated weights from step 7.
9. Repeat steps 1–8 1000 times
Our simulation reflects important style facts from practical portfolio management. The
number of assets can be much larger relative to the number of empirical observations. In
step 4, we allow the residual variances to differ between assets and hence to (mildly) violate
our assumption 3. In step 5, the factor loadings are estimated, while in the main text, we
assume the investor knows the true values to keep the descriptions simple and tractable.
9
The sample covariance matrix serves only for comparison purposes in our simulation because it is not
positive semidefinite. We use the historical covariance matrix to illustrate the usefulness of alternative
methods.
47

Table 5 presents the averages of the Euclidean norm for each covariance estimation
method (Panel A to Panel E) for each sample size and for varying residual variances. Panel
E presents the averages for our portfolio optimization approach that does not require an
estimation of the covariances. The results in Table 5 suggest that the covariance regula-
tion approaches proposed by Ledoit and Wolf (2003), Ledoit and Wolf (2004a), Ledoit and
Wolf (2004b) and our approach are capable of delivering asset weights close to true values.
Additionally, the assumption of identical residual variances is not crucial for covariance reg-
ularization and for our approach. When we allow the residual variances to vary across assets
(columns V ar(ε) = 0.025 + N (0, 0.01) and V ar(ε) = 0.025 + N (0, 0.02)), the Euclidean
norms do not deteriorate considerably. As expected, the estimation of the historical covari-
ance matrix for the cases with N > F results in a matrix that is not positive semidefinite,
and the optimization program is not well defined. The asset weights based on the historical
covariance matrix are estimated only for the purpose of comparison across the methods.
The results confirm that covariance regularizations improve portfolio optimization, as the
Euclidean norm of the differences between true and estimated asset weights is smaller rela-
tive to sample covariance estimates. However, our portfolio optimization approach performs
at least as well as the covariance regularization techniques proposed in the literature and
saves one crucial estimation step.
48

Table 5: Euclidean norm of the difference between true asset weights and calculated portfolio weights based on varying
approaches
N V ar(ε) = 0.025 V ar(ε) = 0.025 + N (0, 0.01) V ar(ε) = 0.025 + N (0, 0.02)
sample
Panel A: V
50 0.296 0.376 0.355
100 0.379 0.427 0.414
200 21.055 342.844 30.257
500 488.871 381.340 556.267
1000 624.808 624.593 770.940
Panel A: V L1
50 0.302 0.411 0.431
100 0.293 0.427 0.428
200 0.318 0.439 0.377
500 0.326 0.408 0.438
1000 0.322 0.398 0.387
Panel C: V L2
50 0.297 0.422 0.429
100 0.324 0.488 0.453
200 0.505 0.520 0.409
500 0.609 0.418 0.406
1000 0.545 0.395 0.344
Panel D: V L3
50 0.306 0.289 0.342
100 0.302 0.312 0.345
200 0.313 0.323 0.278
500 0.313 0.272 0.341
1000 0.311 0.299 0.299
Panel E: New approach
50 0.282 0.253 0.311
100 0.275 0.293 0.332
200 0.303 0.316 0.271
500 0.312 0.270 0.340
1000 0.310 0.298 0.299
D Portfolio optimization with L risk factors and M
characteristics
Let asset returns be generated by the asset return model under assumptions 1, 2 and 4 and
the refined assumption 3a (equations (42)–(44)).
49

Assumption 3a. The return rit on asset i at time t is described by the factor model (42).
K
X
(42) rit = Et−1 (rit ) + βik f˜kt + uit
k=1
where βik is the loading of asset i on factor k and f˜kt ∼ N (0, σf2k ) is the factor’s k innovation
at time t. uit ∼ N (0, σu2 ) is the residual asset return. The expected return is
K
X M
X
(43) Et−1 (rit ) = rf t + ( βik µk + θi cm )
k=1 m=1
| {z }
excess return
where rf t is the return on the risk-free asset at time t, µk is the premium on risk factor k, θi
is the characteristic of asset i and cm ∈ R is the reward for characteristic m. The covariance
matrix for the asset return is
(44) V = BV f B | + RV
where V f is the K × K diagonal covariance matrix of orthogonal factor returns, B is the

N × K matrix of asset factor loadings to K risk factors, and RV is the N × N covariance
matrix of residual asset returns. The residual returns are identically and independently
distributed (i.i.d. residual returns); hence, Cov(ui , uj ) = 0, ∀i 6= j and Cov(ui , uj ) = σu2 , ∀i =
j.
We again reformulate the portfolio optimization program (3)–(6). The variance of the
portfolio is
σp2 = w| (BV f B | + RV ) w
= w| BV f B | w + w| RV w
= β | V f β + σε w| Iw
50

where I is the N × N identity matrix and β is a K × 1 vector, whose elements represent
factor loading values of the portfolio.
Similarly, the investor-specific levels b and θ determine the portfolio return since
µ∗p = w| µ = w| (rf 1 + Bµk + Θc) = rf w| 1 + w| Bµk + w| Θc = rf + β | µk + θ | c
where µk is the K ×1 vector of risk factor premiums, B is a N ×K matrix of factor loadings,

c is the M × 1 vector of characteristic rewards, Θ is the N × M matrix of characteristics
and θ is the M × 1 vector, whose elements represent characteristic values of the portfolio.
By the same arguments as given in appendix A (linking preferences to levels) and ap-
pendix B (adding (scaling) with (by) a constant does not change the optimal weights), the
portfolio optimization program (3)–(6) results in
1 |
(45) min w w
w 2
(46) s.t. w| 1 = 1
(47) w| B = β|
(48) w| Θ = θ|
where the constraint (47) constrains the loadings to risk factors and the constraint (48)
constrains the portfolio characteristics.
For the sake of clarity, we gather the variables on the left-hand side of the constraints
(13)–(16) into a N × (1 + K + L) matrix X with
X = [1, B, Θ]
and the variables on the right-hand side of the constraints are gathered into a (1 + K + L) × 1
vector g with
g = [1, β | , θ | ]|
51

yielding the refined optimization program in Equations (49)–(50).
1 |
(49) min w w
w 2
(50) s.t. X |w = g
The solution for the optimal portfolio weights is given in equation (51).
(51) w| = g | (X | X)−1 X |
The optimal solution of Equation (51) differs across investors because they can choose dif-
ferent values for the elements of g.
E Disentangling test with L risk factors and M char-
acteristics
For the purpose of the disentangling test, we must calculate the portfolio returns at time
t, t = 1, 2, . . . , F . Based on the solution for optimal portfolio weights in equation (51), the
return rpt on portfolio p at time t is
−1
(52) rpt = w|t−1 r t = g | X |t−1 X t−1 X |t−1 r t
By the choices of the elements in vector g, the investors can specify the desired portfolio
features and earn different returns.
We can calculate the returns on P different portfolios in one step when we observe that the
portfolios for one investor differ only with respect to the values in g. Let G = [g 1 , g 2 , . . . , g P ]
be a (1 + K + M ) × P matrix of right-hand-side values of portfolio constraints. The number
of portfolios P can be greater than, less than or equal to (1 + K + M ). For the purpose of
the disentangling test, we are interested in the special case P = 1 + K + M because we test
52

whether the K risk factors and the M characteristics are priced while one portfolio tracks
the risk–free return. In appendix section D equation (51), we show how the optimization for
a single portfolio works for the general case of K risk factors and M characteristics. We can
calculate the realized return for all (1 + K + M ) portfolios at time t as
−1
(53) λt = W |t−1 r t = G| X |t−1 X t−1 X |t−1 r t ∀t
where λt is a P = (1 + K + M ) × 1 vector of portfolio returns at time t, and W t is for the

purpose of the disentangling test a N × (1 + K + M ) matrix of asset weights. Note that
the dimensions of λt and W t are P × 1 and N × P , respectively, if we simply calculate the
return on P portfolios.
To test whether the K risk factors and M characteristic portfolios are priced, among a
large number of potential approaches, presumably the easiest approach is to create tracking
portfolios and to test whether the created tracking portfolio loads on the corresponding risk
factor. In our portfolio optimization program, the tracking portfolios are easily obtained by
setting the respective element of the constraints (g) to the value 1 and all other elements to
zero. For example, if we set the first value of g to 1, then the portfolio is a zero-factor-loading
and zero-characteristic portfolio, which earns the risk–free rate at time t for the investors.
If we set the second (K + 2) element of g to 1, then the portfolio tracks the risk premium
(characteristic reward) of the first risk factor (first characteristic).
By setting the constraint for the factor loading or the characteristic to 1 in vector g, and
all other elements of g being zero, the resulting tracking portfolio is a zero-net-investment.
While zero–net–investment portfolios are rare in practical portfolio management, they are
useful for our disentangling test. Under the null hypothesis of the true risk factor model, the
traditional approach is to test whether the regression constant from a regression of portfolio
returns on risk factors is significantly different from zero. As we form zero–net–investment
tracking portfolios, we can always directly test this hypothesis because the returns on such
53

portfolios are always without the risk-free return component of asset returns. To understand
this point, consider a two–assets zero-investment portfolio that tracks the first risk factor.
The tracking portfolio is formed to be exposed only to the first risk factor. The sum of asset
weights is zero; hence, the portfolio weights are w1 + w2 = 0 ⇔ w2 = −w1 . According to the
risk factor model (42), the realized return on this portfolio is

rp = w1 rf t + β11 µ1 + β11 f˜1t − w1 rf t + β21 µ1 + β21 f˜1t

= (w1 − w1 ) rf t + w1 β11 µ1 + β11 f˜1t − β21 µ1 − β21 f˜1t
= µ1 (w1 β11 − w1 β21 ) + f˜1t (w1 β11 − w1 β21 )
= µ1 βp1 + f˜1t βp1 = µ1 + f˜1t
The realized returns on such a portfolio can directly be used in traditional approaches to
test whether the regression constant is significantly different from zero. On the other hand,
if we use the full–investment portfolio, i.e., w1 + w2 = 1, the return on such a portfolio is

rp = w1 rf t + β11 µ1 + β11 f˜1t + w2 rf t + β21 µ1 + β21 f˜1t

= (w1 + w2 ) rf t + w1 β11 µ1 + β11 f˜1t + w2 β21 µ1 + β21 f˜1t
= rf t + µ1 (w1 β11 + w2 β21 ) + f˜1t (w1 β11 + w2 β21 )
= rf t + µ1 βp1 + f˜1t βp1 = rf t + µ1 + f˜1t
In this example, the full-investment tracking portfolio earns the risk-free rate and the risk
premiums. Then, in the traditional approach, we need to test whether the regression constant
is significantly different from rf , i.e., the test would be b0 −rf = 0 with b0 being the regression
constant. By means of zero–net–investment, we obtain portfolio returns that are easily tested
in the regression framework.
Consequently, the choice for the matrix G as the (1 + K + M ) × (1 + K + M ) identity
matrix in our disentangling test allows the calculation of the tracking portfolio returns µ for
54

(1 + K + M ) portfolios at time t, i.e.,
−1
(54) λt = X |t−1 X t−1 X |t−1 r t ∀t
where the elements of the (1+K +M )×1 vector λt represent the returns of tracking portfolios
for the risk–free rate, the k-th factor return realizations and the m-th characteristic reward
realizations, respectively. For example, the APT factor premium µk is the return on a
portfolio that mimics the k-th risk factor such that µk = N
P
i=1 wik ri , where wik is the weight
of asset i in portfolio k + 1. As the calculation of realized returns λt is done in each time

period t, we can define a (1 + K + M ) × F matrix Λ = [λ1 , λ2 , . . . , λF ].
The solution presented in Equations (53) and (54) offers a good interpretation within asset
pricing theory for the first (1 + K) portfolios. The first portfolio (represented by the first row
in W |t ) is a full-investment portfolio ( N
P
i wi = 1; recall that the first column of X consists
of ones), the portfolio loadings βP k and portfolio characteristics θP m of which are both zero.
Consequently, the portfolio has no loading to either a risk factor or a characteristic reward.
Thus, it has the same expected return as the risk-free rate. If one would use excess return
r, this portfolio would earn an expected excess return of zero. The portfolios represented
by the second to K + 1 rows in W |t require zero net investment. Furthermore, the k + 1
row in W |t represents a portfolio with a portfolio loading of one to the k-th risk factor.
Additionally, the k + 1 portfolio is not exposed to any other risk factor or characteristic
reward. Thus, the k + 1 portfolio has a loading of zero on the remaining K − 1 factors and a
portfolio characteristic of zero on all M characteristic rewards. Such a portfolio, the loading
βP k to the k-th risk factor of which is one and the portfolio loadings to K − 1 remaining risk
factors and the portfolio characteristics of which are zero, simply mimics the ex-post return
realizations of the k-th risk factor. The portfolio of the k-th risk factor is created to comply
with our null hypothesis of a true asset pricing model. If the asset pricing model is true,
the candidate asset pricing model prices without a pricing error each of the portfolios of the
55

k-th risk factor.
The remaining M characteristic portfolios are formed to have the portfolio characteristic
θmP for the m-th characteristic of one and the factor loadings to the risk factors and the
remaining firm characteristics of zero. Such a portfolio does not replicate the return real-
izations of risk factors. If a characteristic reward exists, the characteristic portfolio could
mimic its past return realizations, which is, however, unobservable since we can only observe
the asset characteristic but not the characteristic reward.
After forming the tracking portfolio for each time t according to equation (54), the next
step in our disentangling test is the regression of the portfolio returns in Λ on the risk factor
realization as in equation (42), i.e.,
(55) Λ = BF F | + E
where each row of Λ represents a time series of realized returns on one of the (1 + K + M )
portfolios at time t, t ∈ {1, 2, . . . , F }; F is a F × (K + 1) matrix of ones (constant) and K
columns of returns of risk factors. B F is a (1 + K + M ) × (K + 1) matrix of regression
coefficients, i.e., the constant and the risk factor factor loadings β from the spanning test
(55). E is a (1 + K + M ) × F matrix of the random residual returns.
The important consequence from the choice of the values in G as identity from Equation
(54) is that we know the expected values for the factor loadings from the spanning tests (32).
If the risk factor model is valid, the expected values (E) of the coefficients in B F from the
regression of portfolio returns are
 
rf 0 0 ... 0 
 
 0 β1,p2 = 1 0 ... 0 
(56) E (B F ) = 
 

0 0 β3,p3 = 1 ... 0 
 
 
0 0 0 . . . βK,p(1+K+M )
56

or, if we use excess asset returns
 
0 0 0 ... 0 
 
0 β1,p2 = 1 0 ... 0 
(57) E (B F ) = 
 

0 0 β3,p3 = 1 ... 0 
 
 
0 0 0 . . . βK,p(1+K+M )
F Interpretation of the disentangling test results
1 A motivating example
For more intuition on the results of our test, we analyze the return of the four optimized
portfolios in a simple two-assets case. The idea of our test is to use the optimization program
(12)–(16) to form portfolios, which are exposed only with specific values to the variables of
interest. The expected returns for each of the portfolios are easily derived in the two-
assets case. Of course, the two–assets case only serves illustration purposes since in the
two–assets case, we would face an optimization problem with four restrictions but only two
(free) weights. Such an optimization program is undetermined. As we only illustrate the
interpretation, we abstract from the computation problems. The return on the first portfolio
(formed with (31)) in the two-assets case is given in Equation (58).
(58) E(rp1 ) =w1 E(r1 ) + w2 E(r2 )
=w1 (rf + β1 µrm + e1 µresg + θ1 c) + w2 (rf + β2 µrm + e2 µresg + θ2 c)
= (w1 + w2 ) rf + (w1 β1 + w2 β2 ) µrm + (w1 e1 + w2 e2 ) µresg + (w1 θ1 + w2 θ2 ) c

| {z } | {z } | {z } | {z }
=1 βp∗ =0 e∗p =0 θp∗ =0
=rf
57

The first portfolio is a full-investment portfolio that is not exposed to any risk factors and
earns the risk–free return. If the null hypothesis of a true risk factor model is valid, in a re-
gression of realized portfolio returns on the two risk factors, i.e., rp1t = b0 +b1 rmt +b1 resg +uit ,
this portfolio must have no loading to any risk factor, and the regression constant must equal
the risk-free rate. Otherwise, there would be additional risk factors or firm characteristics
that are empirically important but not considered in our portfolio optimization program
(12)–(16). If we use excess returns, i.e., asset return minus the risk-free rate for each time t,
then under the null hypothesis of the true asset pricing model, the regression constant must
be close to zero and statistically insignificant.
The second portfolio is exposed only to the market portfolio. We form a zero–net–
investment portfolio with a market factor loading of one and no loading to the ESG risk
factor and no ESG characteristic. The return on this portfolio in the two-assets case is
E(rp2 ) =w1 E(r1 ) + w2 E(r2 )

| {z } | {z } | {z } | {z }
=0 βp∗ =1 e∗p =0 θp∗ =0
=µrm
By setting the desired value to βp∗ = 1 and wp∗ , e∗p , θp∗ = 0, the optimization program (12)–(16)
finds a zero-net investment portfolio of firms 1 and 2 with the desired market factor loading
of one, the desired ESG factor loading of zero and the desired ESG characteristic of zero. The
return on this portfolio tracks the realizations of market excess return.10 In a spanning test
regression of the return of this portfolio on the risk factors, i.e., rp2t = b0 +b1 rmt +b1 resg +uit ,
this portfolio must have a factor loading of one to the market portfolio and no factor loading
to other risk factors and an insignificant regression constant close to zero. If we would
10
This example shows that our approach can also be adopted for factor/index tracking purposes or to
create factor-neutral portfolios. Using our optimization approach for these tasks is even more interesting
since an analytical solution can be determined.
58

not impose the zero–net–investment constraint, then the expected portfolio return would
be E(rp2 ) = rf + µrm . Then, in a regression of portfolio returns on the risk factors, the
regression constant must be close to the risk-free rate.
The third portfolio helps test whether the ESG characteristic is related to return. It is a
zero–net investment portfolio with an ESG characteristic of one and no factor loading to the
market risk or to the ESG risk factor. The return on this portfolio in the two-assets case is
E(rp3 ) =w1 E(r1 ) + w2 E(r2 )

| {z } | {z } | {z } | {z }
=0 βp∗ =0 e∗p =0 θp∗ =1
=c
The portfolio tracks the reward of the ESG characteristic. For a spanning test, we run a
regression of the realized returns of this portfolio on the two risk factors. It follows that,
for the ESG characteristic explanation to be a correct description of empirical returns, the
portfolio return must not be exposed to any risk factor, and the regression constant must
be significant. The regression constant represents the return unexplained by the risk factor
model, which does not include the ESG characteristic return.
The fourth portfolio tests whether the ESG risk factor is relevant in modeling expected
returns. Consequently, we set the portfolio ESG factor loading to 1, and we control for
portfolio beta and portfolio ESG characteristics in such a way that they are zero. Then, the
expected portfolio return is
E(rp4 ) =w1 E(r1 ) + w2 E(r2 )

| {z } | {z } | {z } | {z }
=0 βp =0 ep =1 θp =0
=µresg
59

The portfolio tracks the return of the ESG risk factor. For the ESG risk factor explanation
to be true, a necessary condition is that, in a regression of the realized returns of this
portfolio on the market return and ESG risk factor, this portfolio must have a factor loading
statistically not distinguishable from one to the ESG risk factor and no factor loading to the
market risk or a significant constant. Otherwise, the ESG risk factor explanation from the
literature is not a full description of the empirical returns.
2 Simulation
For the purpose of an additional simplified demonstration of how the results of our test must
be interpreted, we conduct a simulation. More specifically, we simulate the null hypothesis
of a true risk factor model as provided in Equation (59).
(59) rit − rf t = bi1t rm,t + bi2t resg,t + it
Accordingly, the market portfolio and ESG risk factor are priced. We simulate the values
of an ESG firm characteristic, but there is no characteristic reward, i.e., c = 0. Therefore,
the ESG characteristic does not affect the cross-section of the returns. In empirical studies,
a firm characteristic is very often correlated with its congruent risk factor loading (e.g.,
Fama and French, 1993; Daniel and Titman, 1997; Daniel et al., 2020; Davis et al., 2000).
Therefore, we assume that the ESG characteristic of each firm is correlated in the time series
with its firm ESG factor loading bi2t with a correlation coefficient of 0.9. The correlation of
0.9 considerably exceeds the empirical values reported in the literature. There was no time
series correlation between the risk factors.
Table 6 reports the values used in the simulation for the factor model that generates
the simulated returns. All of the values of the simulations conform to the assumption of a
two–risk factor model world and are chosen arbitrarily. The market return premium is two
times greater than the ESG risk premium, while both risk factors have the same standard
60

Table 6: Values of the variables in the return simulation rit − rf t = bi1t rm,t + bi2t resg,t + it . Column µ (σ, N, F) reports
the selected mean (standard deviation, number of assets, and number of time series observations) of the variables. The values
of the risk factors rm and resg are drawn from a normal distribution with parameters µ and σ. The values of factor loadings
b1 and b2 and the characteristic θ are generated by an VAR(1) vector–autoregressive model, in which the diagonal elements of
the coefficient matrix are A = 0.999. The time series correlation between b2 and θ (an ESG characteristic) is 0.9.
µ σ N F
rm 0.006 0.03 1 600

resg 0.003 0.03 1 600
b1 0.835 0.74 2000 600
b2 0.162 0.62 2000 600
θ 45.69 17.09 2000 600
i 0 0.03 2000 600
deviation of returns. The values of the simulated risk factor returns and the residual return
i are drawn from a normal distribution with parameters µ and σ. The correlation between
the 2000 simulated asset returns is induced by the risk factor loadings.
Table 7: Results from one run of the simulation. The table reports columns b0 , b1 , b2 and R2 , which show the estimated OLS
coefficients from the regression (32), where the portfolio returns are regressed on the risk factors. The values of the regression
coefficients marked in bold are statistically significant at the 5% level and are calculated with HAC-robust standard errors. The
standard errors are Newey-West corrected with a lag number of 4. R2 is the adjusted R2 .
b0 b1 b2 R2
rp1 0.000 -0.002 0.001 0.001
rp2 0.000 1.001 0.000 0.998
rp3 0.000 -0.002 0.001 0.001
rp4 0.000 0.001 1.000 0.999
∗∗∗
indicates significance at the 1% level
Table 7 reports the results of our test based on optimized portfolios. Each row in the
table presents the results of the spanning test for one of the four portfolios, whose returns
are calculated by Equation (30). The first portfolio is expected to have no factor loading to
a risk factor and no abnormal returns. If we include the risk-free rate in our model, then this
portfolio will track the risk–free return. The test confirms this expectation because all of the
regression coefficients are zero. The second portfolio is exposed to the market portfolio but
not the ESG risk factor or ESG characteristic. Therefore, we should observe a significant
factor loading of 1 only to the market portfolio. We observe a value of zero of the constant of
61

the regression, suggesting that the assumed risk factor model can price the portfolio returns.
Consistent with the null hypothesis, the factor loading to the market portfolio is very close
to the ideal value of 1 in the second optimized portfolio. The remaining coefficients of the
second portfolio are, as expected, close to zero. The third portfolio is formed on the values
of the ESG characteristic. According to the simulation of a true risk factor model, there is
no factor loading to any risk factor despite the high correlation between ESG factor loading
and the ESG characteristic. The simulation results confirm the null hypotheses. Our test of
the fourth portfolio shows that it is exposed only to the ESG risk factor, which is consistent
with the null hypothesis.
The values of the coefficients b1 and b2 are notable. When constructing the optimized
portfolios, we precisely impose the estimated coefficient values reported in Table 7 for the
calculation of asset weights in varying elements of D as an identity matrix from Equation
(31). The returns on the third portfolio representing characteristic rewards are not exposed
to any simulated factor return despite the high correlation between ESG factor loading and
the characteristic because the optimization imposes the constraint that the portfolio loading
to the two risk factors is zero. The estimated coefficients from the time-series regressions
with the simulated returns suggest that the optimized portfolios help distinguish between
the ESG risk factor and ESG characteristic stories. Notably, the R2 is quite low for the
characteristic portfolios.
62

Portfolio Optimization For Sustainable Investments

Uploaded by

Copyright:

Available Formats

Portfolio Optimization For Sustainable Investments

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

Portfolio Optimization For Sustainable Investments

Uploaded by

Copyright:

Available Formats

Portfolio optimization for sustainable

School of International Business Bremen

Electronic copy available at: https://ssrn.com/abstract=3859616

In mean-variance portfolio optimization, factor models can accelerate computation, re-

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

II Portfolio selection and asset return model

1 Extended mean–variance optimization program

Electronic copy available at: https://ssrn.com/abstract=3859616

µp is the portfolio return with µp = w| µ, where w is an N × 1 vector of portfolio weights of

where h is the Lagrangian multiplier with

−1 −1 −1

2 Linking preferences to levels

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

3 Asset return models and portfolio optimization for sustainable

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Assumption 2. Asset returns have finite first and second moments.

Electronic copy available at: https://ssrn.com/abstract=3859616

where V f is the K ×K diagonal covariance matrix of orthogonal factor returns, B is a N ×K

Electronic copy available at: https://ssrn.com/abstract=3859616

and the covariance σij between assets i and j is

and the covariance σij between assets i and j is

Electronic copy available at: https://ssrn.com/abstract=3859616

(10) Et−1 (rit ) = µi = rf t + βi µrm + ei µe + θi c

and the covariance between the assets i and j is

(11) σij = βi βj σr2m + ei ej σr2e + σuij

= βp σr2m βp + ep σr2e ep + σu2 w| Iw

Electronic copy available at: https://ssrn.com/abstract=3859616

µ∗p =w| µ = w| (rf t 1 + βµrm + eµr + θc)

=rf t w| 1 + w| βµrm + w| eµre + w| θc

=rf t + βp µrm + ep µre + θp c

Electronic copy available at: https://ssrn.com/abstract=3859616

yielding the refined optimization program in Equations (17)–(18).

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

(24) w|θ = b|θ (X |θ X θ )−1 X |θ

where bθ = [1, βp∗ , θp∗ ] and X θ = [1, β, θ].

Electronic copy available at: https://ssrn.com/abstract=3859616

(29) w|e = b|e (X |e X e )−1 X |e

where be = [1, βp∗ , e∗p ] and X e = [1, β, e].

Electronic copy available at: https://ssrn.com/abstract=3859616

III The disentangling test between the competing ESG

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

ESG Median µ σ P(25) P(75) µ/σ N

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

Electronic copy available at: https://ssrn.com/abstract=3859616

and depends only on the choices of α, λ and .

(59) rit − rf t = bi1t rm,t + bi2t resg,t + it