Quantitative Finance

ISSN: 1469-7688 (Print) 1469-7696 (Online) Journal homepage: https://www.tandfonline.com/loi/rquf20

Comparative issues between linear and non-

linear risk measures for non-convex portfolio
optimization: evidence from the S&P 500

Panos Xidonas & George Mavrotas

To cite this article: Panos Xidonas & George Mavrotas (2014) Comparative issues between
linear and non-linear risk measures for non-convex portfolio optimization: evidence from the
S&P 500, Quantitative Finance, 14:7, 1229-1242, DOI: 10.1080/14697688.2013.868027

To link to this article: https://doi.org/10.1080/14697688.2013.868027

Published online: 25 Feb 2014.

Submit your article to this journal

Article views: 254

View related articles

View Crossmark data

Citing articles: 2 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=rquf20
Quantitative Finance, 2014
Vol. 14, No. 7, 1229–1242, http://dx.doi.org/10.1080/14697688.2013.868027

Comparative issues between linear and non-linear

risk measures for non-convex portfolio
optimization: evidence from the S&P 500
PANOS XIDONAS*† and GEORGE MAVROTAS‡
†Decision Support Systems Laboratory, School of Electrical & Computer Engineering, National Technical University of
Athens, Athens, Greece
‡Laboratory of Industrial & Energy Economics, School of Chemical Engineering, National Technical University of Athens,
Athens, Greece

(Received 13 March 2013; accepted 14 November 2013)

This article focuses on inferring critical comparative conclusions as far as the application of
both linear and non-linear risk measures in non-convex portfolio optimization problems. We seek to
co-assess a set of sophisticated real-world non-convex investment policy limitations, such as
cardinality constraints, buy-in thresholds, transaction costs, particular normative rules, etc. within
the frame of four popular portfolio selection cases: (a) the mean-variance model, (b) the mean-semi
variance model, (c) the mean-MAD (mean-absolute deviation) model and (d) the mean-semi MAD
model. In such circumstances, the portfolio selection process reflects to a mixed-integer bi-objective
(or in general multiobjective) mathematical programme. We precisely develop all corresponding
modelling procedures and then solve the underlying problem by use of a novel generalized
algorithm, which was exclusively introduced to cope with the above-mentioned singularities. The
validity of the attempt is verified through empirical testing on the S&P 500 universe of securities.
The technical conclusions obtained not only confirm certain findings of the particular limited
existing theory but also shed light on computational issues and running times. Moreover, the results
derived are characterized as encouraging enough, since a sufficient number of efficient or Pareto
optimal portfolios produced by the models appear to possess superior out-of-sample returns with
respect to the benchmark.

Keywords: Portfolio optimization; Linear and non-linear risk measures; Non-convex policy
constraints; S&P 500
JEL Classifications: C, C6, C61

1. Problem setting extremely high, even when very fast computer systems and
solvers are exploited.
The mean-variance formulation (Markowitz 1952) is target The last issue was successfully confronted by some very
for severe criticism, mainly because of the difficulty of its effective techniques developed by Markowitz and Perold
implementation. As an example of the corresponding (1981) and Perold (1984). To avoid manipulations with very
predicament, consider an institution that follows a number dense covariance matrices, these approaches focused on
of 250 stocks will need 31,125 correlation coefficients for methods for diagonalizing the covariance matrix structure,
implementing the underlying framework. But it seems quite which enable one to go directly from historical data to a
unlikely that analysts will be able to directly estimate such diagonalized covariance matrix structure, just by adding
huge and complex correlation structures, within the feasible only a few extra variables and constraints to the formula-
organizational limits of a traditional asset management tion. The downside of the above methods is that one can
house. Additionally, the computational difficulty associated generally figure that there is a loss of information. Thus,
with solving a large-scale quadratic programming problem there is likely to be a difference between the frontier
with very dense covariance matrix on a real-time basis is obtained and the one desired. Konno and Suzuki (1992)
proposed yet another diagonalization strategy. It is interest-
ing because it does not involve a loss of information. Being
*Corresponding author. Email: [email protected]
© 2014 Taylor & Francis
1230 P. Xidonas and G. Mavrotas

based upon Cholesky factorization, it requires that a covari- integer variables are involved. Practical portfolio optimiza-
ance matrix be initially in hand and that it be invertible. tion problems like those stated above can be solved by the
While this method, when applicable, will produce a true conventional integer linear programming software within a
non-dominated frontier, the difficulty is that most large practical amount of time. Nevertheless, the whole context is
covariance matrices are not invertible (Steuer et al. 2011). becoming even more complicated, when multiple investment
An additional factor that seriously augments the complex- objectives are to be simultaneously optimized (Lee and
ity of the portfolio optimization process has to do with the Chesser 1980, Roman et al. 2007, Steuer et al. 2007,
need of co-assessing a set of sophisticated real-world Xidonas and Mavrotas 2012).
investment constraints (Branke et al. 2009). Such The current paper aspires to effectively deal with all the
constraints may be: (a) the maximum number of securities above issues, since few propositions, if any, are concerned
to be included in the portfolio; (b) certain buy-in thresholds, with the introduction of a generalized mixed-integer multi-
which demand at least some minimum amount of the stock objective portfolio optimization framework for modelling
to be bought, if this stock is included in the portfolio; (c) and analytically solving real-world situations, in the pres-
the transaction costs volume; and (d) particular normative ence of: (a) multiple investment objectives, (b) cardinality
rules, such as the 5-10-40 constraint, which defines upper constraints, (c) transaction costs and (d) compliance norms.
limits in the weight of each individual security or sum of Our primary target is to infer critical comparative conclu-
weights of the same issuer’s securities in the portfolio. sions as far as the application of both linear and non-linear
Indeed, an optimal solution of a large-scale quadratic risk measures in non-convex portfolio optimization
programming problem usually contains many non-zero problems. On this basis, we seek to co-assess all the above-
elements. However, this is extremely inconvenient since mentioned sophisticated investment policy limitations,
significant amount of transaction costs have to be paid in within the frame of four popular portfolio selection cases:
order many different stocks to be bought in a small amount. (a) the mean-variance model, (b) the mean-semi variance
Also, it may be impossible to purchase very small amounts model, (c) the mean-MAD (mean-absolute deviation) model
of stock, due to trading restrictions in the minimum transac- and (d) the mean-semi MAD model.
tion units. Moreover, a very large number of assets in a The validity of the attempt is verified through empirical
portfolio, as a result of the optimization process, may force testing on the S&P 500 universe of securities. The technical
the fund manager to eliminate stocks with smaller weight, conclusions obtained not only confirm certain findings of
in order to get around the difficulty. But these cut-off the particular limited existing theory but also shed light on
manipulations may distort the substance of the portfolio computational issues and running times. Moreover, the
optimization process to the extent that the resulting portfo- results derived are characterized as encouraging enough,
lio standard deviation is considerably larger than the since a sufficient number of efficient or Pareto optimal port-
obtained through the exact model. folios produced by the models, appear to possess superior
All the above problems can be properly formulated by out-of-sample returns with respect to the benchmark.
introducing integer variables (Canakgoz and Beasley 2009). The paper proceeds as follows: In section 2 we present
However, solving portfolio optimization problems containing the proposed methodological framework for dealing with
integer variables has been considered to be an intractably dif- bi-objective or in general multiobjective portfolio selection
ficult problem. The main reason is that the standard model problems. In section 3, we meticulously analyse the whole
uses variance (or standard deviation) as the measure of risk. model building procedure. Finally, the empirical results and
Unfortunately, there exists virtually no efficient algorithm for concluding remarks are given in sections 4 and 5.
analytically solving non-linear integer programming
problems. For this reason, many researchers propose the line-
arization of the portfolio selection problem. Following 2. Proposed framework
Sharpe (1971), Konno and Yamazaki (1991), Ogryczak
(2000), Mansini et al. (2003) and Angelelli et al. (2007) The proposed methodological approach is based on the
represent milestone attempts towards this direction, i.e. theoretical framework of multiobjective mathematical
exploiting absolute deviation as an alternative measure of risk programming (MMP). This section highlights the intercon-
and formulating the portfolio selection problem as a linear nection between this approach and the major issue of intro-
one. ducing multiple investment objectives in the real-world
As a risk measure, absolute deviation is consistent with portfolio management practice.
the von Neumann’s principle of rational decision-making, The general MMP model is defined as follows:
namely the principle of maximization of expected utility
(Konno and Yamamoto 2005), while proportional to standard max½ f1 ðxÞ; f2 ðxÞ; . . .; fp ðxÞjx 2 S (1)
deviation when security returns are multivariate normally
distributed (Konno and Yamazaki 1991). Also, absolute devi- where x is the vector of decision variables, fi ðxÞ;
ation is essentially represented as a linear system of equa- i ¼ 1; 2; . . .; p are the objective functions and S is the feasi-
tions when the random variables are distributed over a ble region. As already noted, the concept of optimality in
discrete set of points, so that the portfolio optimization prob- MMP is replaced with that of efficiency or Pareto optimality.
lem reduces to a linear programming problem. This means According to Hwang and Masud (1979), the methods for
that the mean-absolute deviation model possess critical solving MMP problems can be classified into three catego-
advantages over the (non-linear) mean-variance model when ries, according to the phase in which the decision-maker
 Comparative issues between linear and non-linear risk measures 1231

(DM), i.e. the investor in the case of portfolio selection, is p − 1 objective functions that will be used as constraints.
involved in the decision-making process, by expressing his The calculation of the range of the objective functions over
preferences: the a priori methods, the interactive methods the efficient set is not a trivial task. While the best value is
and the generation or a posteriori methods. easily attainable as the optimal of the individual optimiza-
Generation methods are the less popular due to the tion, the worst value over the efficient set (nadir value) is
underlying computational burden and the lack of widely not. The most common approach is to calculate these
available software. However, they have some significant ranges from the pay-off matrix (the table with the results
advantages. The solution process within a generation from the individual optimization of the p objective
method is divided in two phases: The first phase concerns functions). On this basis, the nadir value is usually
the generation of the efficient solutions, while the second approximated as the minimum of the corresponding column.
phase concerns the involvement of the DM when the aggre- However, even in this case, someone must be sure that the
gate information is on the table. Consequently, the DM is obtained solutions from the individual optimization of the
involved only in the second phase, having at hand all the objective functions are indeed efficient solutions. In
possible alternatives, i.e. the whole Pareto optimal set of presence of alternative optima, there is no guarantee that a
portfolios. Hence, generation methods are very favourable solution obtained by commercial software is an efficient
in cases where the DM is hardly available and the interac- solution. In order to overcome this ambiguity, we propose
tion with him is difficult, an extremely crucial issue within the use of lexicographic optimization for every objective
the typical process of investment management. Besides, the function in order to construct the pay-off matrix,
fact that none of the potential solutions has been left uncov- exclusively with efficient solutions. It is well known that
ered reinforces the DM’s confidence on the final decision. lexicographic optimization provides only efficient solutions
One of the most widely used generation method is the (Miettinen 1998). A simple remedy in order to bypass the
ε-constraint method. In the ε-constraint method, one of the difficulty of estimating the nadir values of the objective
objective functions is optimized using the other objective functions is to define reservation values for the objective
functions as constraints, incorporating them in the constraint functions. A reservation value acts like a lower (or upper
part of the model as shown below (Chankong and Haimes for minimization objective functions) bound. Values worse
1983). Without loss of generality, the first objective than the reservation value are outside the feasible region
function is optimized and the rest p − 1 objective functions and are not allowed.
are treated as constraints:

maxf f1 ðxÞj f1 ðxÞ  ei ; i ¼ 2; 3; . . .; p ^ x 2 Sg (2) 2.2. Avoidance of weakly Pareto optimal solutions
The second point of attention is that the optimal solution of
By parametrical variation in the right-hand side of the problem (2) is guaranteed to be an efficient solution only if
constrained objective functions, the efficient solutions of the all p − 1 objective function constraints are binding. If some
problem are obtained. of the p − 1 constraints that correspond to the objective func-
Several versions of the ε-constraint method have been tions are non-binding, then problem (2) may have alternative
appeared in the literature trying to improve its performance optima that may produce better values for the p − 1 objective
or adapt it to a specific type of problems (Hamacher et al. functions (as the relevant constraints are not binding). In this
2007). However, despite its advantages over the weighting case, the optimal solution of (2) may not be a non-domi-
method, the ε-constraint method has three points that need nated solution of the multiobjective problem. In order to
special attention in its implementation: (a) the calculation of overcome this defect, we propose the transformation of the
the range of the objective functions over the efficient set; objective function constraints to equalities by explicitly
(b) the guarantee of efficiency of the obtained solution; and incorporating the appropriate slack (for minimization objec-
(c) the increased solution time for problems with several tives) or surplus (for maximization objectives) variables. At
(more than two) objective functions. the same time, these slack or surplus variables are used as a
Mavrotas (2009) addresses the above issues are by second term (with lower priority) in the objective function,
introducing a novel version of the ε-constraint method, the forcing the programme to produce only efficient solutions.
so-called augmented ε-constraint method (AUGMECON). The new problem becomes:
The main innovations introduced within the AUGMECON (
method are: (a) the use of lexicographic optimization; (b) Xp

the guarantee for producing only Pareto optimal solutions max f1 ðxÞ þ d  si j fi ðxÞ  si ¼ ei ; i ¼ 2; 3; . . .;
i¼2
(thus avoiding weakly Pareto optimal solutions); (c) the )
algorithmic acceleration of the computational process, using þ
p ^ x 2 S; si 2 R ð3Þ
the early exit from the nested loops; and (d) the implemen-
tation of the algorithm. A brief and intuitive description of
the above novelties is provided below. where δ is a small number, usually between 10−3 and 10−6.
The above formulation (3) of the ε-constraint method
produces only efficient solutions and avoids the generation
2.1. Construction of the pay-off matrix
of weakly efficient solutions. In order to avoid any scaling
In order to properly apply the ε-constraint method, we must problems, it is also recommended to replace the si in the
have the range of every objective function, at least for the second term of the objective function by si/rgi, where rgi is
1232 P. Xidonas and G. Mavrotas

the range of the i-th objective function (as calculated from Multiobjective optimization results to the acquisition of
the payoff matrix). Thus, the objective function of the the efficient surface of Pareto optimal portfolios, from
ε-constraint method becomes: which the final selection has to be made. The effective
representation of this information, in such problem types,
( )
Xp
si i.e. selecting portfolios under multiple investment
max f1 ðxÞ þ d  (4) objectives, is a matter of crucial importance. An efficient
i¼2
rg i
way to delineate the solutions space is by means of precise
and illustrative graphs (see the figures introduced through
The proposed version of the ε-constraint method that paragraphs 4.2–4.5).
corresponds to model (3) with the objective function (4)
reflects to the so-called augmented ε-constraint method or
AUGMECON method.
3. Model building

2.3. Early exit from the loops In this section, we develop the detailed equations of the
models used. For clarity, decision variables are denoted by
It must be noted that a similar, to above mentioned, attempt
lowercase letters, while parameters are expressed by capi-
which uses slack and surplus variables in order to transform
tals. In the first subsection, we describe the common part of
the ε-constraints into equality constraints is proposed by
the models that incorporates the objective function of return
Ehrgott and Ryan (2002) in their elastic constraint method.
and the assumed policy constraints. Particular emphasis is
The main innovation in the elastic constraint method is the
given to the modelling of the 5-10-40 policy requirement
way it handles infeasibilities in the ε-constraint problem
which has become common practice in modern portfolio
(surplus variables that are incorporated with penalty coeffi-
management. In the subsequent sections, we describe the
cients in the objective function). In AUGMECON, the case
required equations for modelling the four risk measures,
of infeasibilities is treated differently, using the early exit
namely variance, semi-variance, mean-absolute deviation
from the nested loops. This is an innovative addition to the
and semi-mean absolute deviation. We complete the section
algorithm that is activated whenever problem (3) becomes
with the summary characteristics of the four models.
infeasible for some combination of ei. The early exit from
the loops works as follows: The bounding strategy for each
one of the objective function starts from the more relaxed
formulations (lower bound for a maximization objective 3.1. Common part across the four models
function or upper bound for a minimization) and move to
strictest (individual optima). In this way, when we arrive at 3.1.1. The objective function of return maximization.
an infeasible solution there is no need to perform the Assume an existing portfolio Po in the beginning of the
remaining runs of the loop (as the problem will become analysis, where the weight of the i-th security in it is
even stricter and thus remains infeasible) and we force an denoted as Wi0 . Then, for the maximization of the expected
exit from the loop. In problems with many objective portfolio’s capital return we have:
functions (as the portfolio selection problem may have) the
early exit from the loops significantly improves the solution X
N
time. max Z1 ¼  i  wi  Bi  wþ  Si  w Þ
ðR (5)
i i
i¼1

2.4. Implementation where R  i is the average capital return of the i-th security
AUGMECON’s implementation is also characterized by a over the assumed historical period, wi is the weight of the
couple of attractive attributes. From the pay-off matrix we i-th security, wþ 
i and wi are the weights of the i-th security
obtain the range of each one of the p − 1 objective that should be bought and sold according to the results of
functions that are going to be used as constraints. Then we the analysis (with regard to the existing weight Wi0 ), Bi and
divide the range of the i-th objective function to qi equal Si are the buying and selling commissions and N is the
intervals using qi − 1 intermediate equidistant grid points. number of securities in the investment universe.
Thus, we have in total qi + 1 grid points that are used to
vary parametrically the right-hand side of the i-th objective
function. The total number of runs becomes
(q2 + 1) × (q3 + 1) × … × (qp + 1). A desirable 3.1.2. The constraints. The constraints of the problem
characteristic of the AUGMECON method (and generally are the required balance equations as well as the policy
the ε-constraint philosophy) is that we have total control of constraints. The specific constraints that implement the
the density of the efficient set representation, by properly 5-10-40 rule are described separately in a subsequent
assigning the values to the qi. The higher the number of sub-section.
grid points the denser is the representation of the efficient The securities’ weight quantities are controlled through
set but with the cost of higher computation times. A the following balance equation:
trade-off between the density of the efficient set and the
computation time is always advisable. wi ¼ Wi0 þ wþ 
i  wi (6)
 Comparative issues between linear and non-linear risk measures 1233

which means that the weight of the i-th security in the incorporating appropriate additional binary variables in
resultant portfolio is the initial, plus the weight of the spe- order to express the discontinuities of the feasible domain.
cific security that has been bought, minus the weight that Given that this specific rule is incorporated in the mathe-
has been sold. When starting from scratch Wi0 = 0 for all i. matical programming model and is not applied a posteriori,
The capital completeness and no short sales allowance it is guaranteed that all the generated Pareto optimal portfo-
are introduced with the following equation: lios comply with this rule.
One very interesting attempt to incorporate the 5-10-40
X
N rule, within a mathematical programming framework, is the
wi ¼ 1 and wi  0 (7) one of Branke et al. (2009). However, an improved
i¼1 modelling formulation is proposed in the present paper.
More precisely, the auxiliary decision variables that are
The model is also equipped with binary variables, in order to used for the issuers which surpass the 5% weight do not
control the existence (bi = 1) or non-existence (bi = 0) of the express its total weight, but only the additional weight over
i-th security in the portfolio. The use of binary variables 5%. In this way, a number of auxiliary constraints are
allows the following cardinality constraint equation, provid- avoided and in the cases that were tested the solution
ing direct control of the number of securities in the portfolio: obtained faster than applying the Branke et al. (2009)
model. In the specific examples with a universe of 477
X
n
SL  bi 6 SU (8) securities and a historical horizon of 104 weeks (=2 years),
i¼1 our 5-10-40 formulation was on average 4.96% more
efficient in computational time than the model of Branke
where SL and SU are the minimum and maximum number et al. (2009).
of securities allowed to participate in the portfolio. More specifically, the 5-10-40 rule is modelled as
Two more constraint equations are introduced for the follows:
exact calibration of the i-th security’s lower and upper X
weight in the portfolio: wi 6 0:05  ð1 þ oj Þ; j ¼ 1; 2; . . .; S (12)
i2Pj
wi  WL  bi  0; i ¼ 1; 2; . . .; N (9)
X
wi  0:05 6 aj 6 0:05  oj ; j ¼ 1; 2; . . .; S (13)
wi  WU  bi  0; i ¼ 1; 2; . . .; N (10) i2Pj

where WL and WU are the minimum and maximum allowable

security weights in the portfolio. It is noteworthy that without X
S X
S
the binary variables we cannot model equation (9) 0:05  oj þ aj  0:4 (14)
(minimum buy-in threshold). j¼1 j¼1

Moreover, for the proper tuning of the portfolio’s beta

coefficient (market risk), we formulate the following two where Pj is a portfolio that contains the securities of the
inequalities: j-th issuer, oj are binary variables which indicate if the j-th
issuer is an “overachiever” i.e. if the j-th issuer’s weight is
X
N less than 5% then oj = 0, if it is more than 5% then oj = 1,
BetaL  Betai  wi  BetaU (11) aj are the auxiliary continuous variables associated with the
i¼1 j-th issuer (receiving zero value if the issuers’ weight is less
than or equal to 5% or the value of the additional weight
where Betai is the beta coefficient of the i-th security and over 5% if the issuers’ weight is greater than 5%) and S is
BetaL and BetaU are the lower and upper allowable the number of issuers. It must be stressed that if there is
portfolio beta bounds correspondingly. only one security per issuer, then S = N.

3.1.3. Formulation of the 5-10-40 rule. The proposed 3.2. Modelling risk
modelling approach also incorporates the popular 5-10-40
directive, a rule that is based on the established legislative In this subsection, the four versions of the second objective
framework for mutual funds. According to this specific rule, function that quantifies risk along with the accompanying
securities of the same issuer are allowed to weight a constraints are described.
maximum of 5% of the mutual fund’s net asset value.
However, in the special case that the total weight of all the
issuer’s securities with amounts between 5 and 10% is less 3.2.1. Risk as variance (VAR model). In our approach
than 40%, they are allowed to amount a maximum of 10%. we do not use the covariance matrix but the original data in
The complexity accrues from the fact that the 5-10-40 order to calculate the variance implementing a formulation
constraint is non-convex, thus ruining any initial linearity similar to Konno and Yamazaki (1991). In the present case
assumptions. However, the difficulty is managed by with 500 securities and 2 years (=104 weeks) historical
1234 P. Xidonas and G. Mavrotas

weekly data, instead of a 500 × 500 matrix we use a order to avoid the quadratic terms in the objective function
104 × 500 which means substantial computational economy. implied by the incorporation of variance, they introduce the
The minimization of portfolio’s variance can be written as: linear version of risk quantified by the portfolio’s MAD:
 
!2 T X 
1X T XN 1X 
N
 i Þ
min Z2 ¼ iÞ
wi  ðRit  R min Z2 ¼  wi  ðRit  R (20)
T t¼1 i¼1
(15) T t¼1  i¼1 

where T is the number of historical periods and Rit is the As in the previous case of variance, T additional positive
return of the i-th security during the t-th historical period. continuous variables (yt) are used for the representation of
On this basis, T additional positive continuous variables (yt) each period’s absolute deviation from the mean. However,
are used for the representation of each period’s deviation in the MAD formulation instead of T equations we need the
from the mean, resulting in the following T constraints that following 2 × T inequalities as constraints in order to
are added to the model: properly capture the effect of absolute deviation:

X X
N
N
 i Þ  yt ¼ 0;  i Þ þ yt > 0;
wi  ðRit  R t ¼ 1; 2; :::; T (21)
wi  ðRit  R t ¼ 1; 2; . . .; T (16)
i¼1
i¼1

Consequently, the objective function that minimizes

variance is: X
N
 i Þ  yt 6 0;
wi  ðRit  R t ¼ 1; 2; :::; T (22)
i¼1
1 X
T
min Z2 ¼ y2t (17)
T t¼1 Then, the objective function that minimizes MAD is
transformed to:
It is observed that the second objective function of the
VAR model is quadratic, so that a Mixed Integer Quadratic 1X T

Programming (MIQP) solver is required for its solution. min Z2 ¼ yt (23)

T t¼1

3.2.2. Risk as semi-variance (SVAR model). From the

early steps of his modern portfolio theory, Markowitz 3.2.4. Risk as semi-mean absolute deviation (SMAD
recognizes the importance of semi-variance as an alternative model). Following the concepts of semi-variance (only
risk measure (Markowitz 1952). Semi-variance refers to the downside risk is to be minimized), Speranza (1993)
“downside risk”. It implies that we are interested in proposes the use of the mean-absolute semi-deviation
minimizing the underachievement part (those parts that are (SMAD). The modelling of the specific measure follows the
under the mean) of the total variance and we are not inter- modelling of MAD from Konno and Yamazaki (1991)
esting in the overachievement (which is actually desirable if described in the previous subsection, with the following
we talk about returns). The modelling of semi-variance is -difference: we use only the relation (21) from the two
performed in the same way as the modelling of variance inequalities that model the absolute value, implying that we
with the following modifications: the decision variable yt are only interested in downside deviations from the mean.
that expresses the deviation from the mean splits into the Therefore, the model for SMAD is the same as that of
downside deviation y þ MAD but in the place of inequalities (21) and (22) we use
t and the upside deviation yt .
only the following:
X
N
 i Þ  yþ þ y ¼ 0;
wi  ðRit  R t ¼ 1; 2; :::; T (18) X
N
t t  i Þ þ yt > 0;
wi  ðRit  R t ¼ 1; 2; . . .; T (24)
i¼1
i¼1

We are interested in minimizing only the downside

Then, the objective function that minimizes the
deviation so that only y
t participates in the objective
mean-absolute semi-deviation is transformed to:
function:

1X T
1X T
min Z2 ¼ yt (25)
min Z2 ¼ ðy Þ2 (19) T t¼1
T t¼1 t

3.3. Summary of models’ characteristics

3.2.3. Risk as mean-absolute deviation (MAD model).
In 1991, Konno and Yamazaki propose the use of Conclusively, for a model with N securities, S issuers and T
mean-absolute deviation (MAD) as alternative measure of historical data points, the characteristics of the four models
risk in their seminal work (Konno and Yamazaki 1991). In are shown in table 1.
 Comparative issues between linear and non-linear risk measures 1235

Table 1. Model characteristics.

Model Type Objective functions Continuous variables Binary variables Constraints

VAR MIQP Return-variance 3N + S + T N+S 3N + 3S + T + 6

SVAR MIQP Return-semi variance 3N + S + 2T N+S 3N + 3S + T + 6
MAD MIP Return-MAD 3N + S + T N+S 3N + 3S + 2T + 6
SMAD MIP Return-semi MAD 3N + S + T N+S 3N + 3S + T + 6

Notes: N: Number of securities, S: Number of issuers, T: Number of historical data.

This means that for a problem with 477 securities, the actual performance (i.e. its ex-post return) for the next
same number of issuers and 104 data points (=2 years of quarter. The portfolio’s actual performance (ex-post
weekly data), the number of continuous variables is 2012 performance) is calculated by multiplying the actual returns
(2116 for the SVAR model), the number of binary variables of the securities for the specific quarter with the weights of
is 954 and the number of constraints is 2972 (3076 for the the securities as we obtain them for the specific Pareto
MAD model). Although the obtained models are large in portfolio from the bi-objective optimization. In this way, we
size, as it will be presented in the next section, the efficient can compare the efficient frontiers using their actual
frontiers can be generated in a few minutes. performance. The performance of the S&P 500 index is
also calculated for the next quarter in order to be compared
with the Pareto optimal portfolios’ performances. Our
4. Empirical evidence from the S&P 500 purpose is not to compare individual portfolios, but to
compare efficient fronts as obtained by the four models.
For each period, we run the four models described in
4.1. Design of the computational experiment
section 3 (VAR, SVAR, MAD and SMAD). These models
The proposed methodology has been applied to one of the differ only in the expression of risk while all the other
most popular stock market indices, the S&P 500. The characteristics are the same. From the four bi-objective
empirical testing process has two basic characteristics: (a) a optimizations we obtain four efficient fronts, one for each
rolling historical optimization horizon of two-yrs constant risk model. We repeat this process for the four periods
length, starting from 01-01-08, and (b) an out-of-sample shown in figure 1.
evaluation period of one year, separated in four quarters. Regarding the required model’s parameters described in
For each one of them, we perform the out-of-sample section 3, we use some indicative, close to reality values, as
analysis. The two years are the calculation period and the most of them are subjective. Specifically, the portfolio’s
three months is the testing period (i.e. ‘out-of-sample’ as it beta coefficient is constrained between 0.7 and 1.3
was not taken under consideration during the calculations). (equation (11)), while the buying and selling commissions
Consequently, the validation phase covers in total a were set to 0.1 and 0.15% correspondingly (equation (5)).
three-yrs period (see figure 1). After preliminary analysis it Regarding the cardinality constraints and due to the 5-10-40
was found that 23 from the 500 securities do not have norm, which forces the minimum number of securities that
adequate data and therefore we used 477 of them. participate in the portfolio to be 16, only an upper bound
The granulation of the historical data is weekly, which was determined to 30 (equation (8)). Moreover, the
means that we use weekly closing prices to calculate the minimum buy-in threshold was set to 3% (equation (9)).
return of the securities as well as of the S&P 500 index. The GAMS language (version 23.5) is used for modelling
This means that the two-year rolling historical horizon has and the CPLEX 12.2 solver is used for the optimization.
104 data points. The calculated efficient frontier is repre- The GAMS models are properly developed to implement
sented by 51 points that correspond to 51 Pareto optimal the augmented ε-constraint method (section 2) that produces
(efficient) portfolios. For each one of them we calculate its the efficient frontier. The Mixed Integer Programming

Figure 1. The empirical testing process.

1236 P. Xidonas and G. Mavrotas

(MIP) module of CPLEX 12.2 is used for the MAD and the last portfolios (with id = …49, 50, 51) correspond to
SMAD models and the MIQP module is used for the VAR those closer to minimum risk. The horizontal red line is the
and the SVAR models. The characteristics of the PC that performance of the S&P 500 index in the first quarter of
was used for the empirical testing are: Intel Core i5 at 2010. It is evident that almost all the Pareto optimal
2.5 GHz with 64 bit OS and 4 GB RAM. The portfolios outperform the index. In order to have a better
computational process of the empirical testing phase for comparative view of the four models, we calculate the
each period is described in the next subsections, while in ‘cumulative distribution’ of the Pareto optimal portfolios’
subsection 4.6 the summary and discussion of the results is performance across the efficient frontier. In figure 3, we see
presented. the curves of these cumulative distributions. As it was
mentioned, MAD and SMAD model provide almost
identical results and they are depicted with the same curve
(MAD-SMAD).
4.2. Run #1: optimization period 1/1/08–31/12/09, test The x-axis is the bin scale from 1 to 99% and the y-axis
period 1/1/10–31/3/10 is the portfolio’s return. Every point (x, y) in the curves
We implement the four models of section 3 using a two-yrs indicate that the x% of the Pareto optimal portfolios have
historical optimization horizon, from 01-01-08 to 31-12-09, return lower than y. For example, in figure 3 we can
in order to obtain the Pareto optimal portfolios and the effi- observe that 10% of the Pareto optimal portfolios (x-axis)
cient frontiers. The solution time for the 4 runs varied from obtained from the MAD (and the SMAD) model have
94″ (SMAD model) to 3712″ (VAR model) with an Intel return less than 11% (y-axis), which means that 90% of
Core i5 M460 2.53 GHz processor. It was observed that the them have return more than 11%. The upper is located one
results from the MAD and the SMAD model are almost curve in the graph, the better is the model in comparison to
identical regarding the Pareto optimal portfolios (slight the others. From figure 3 we can see that the four models
differences only in the second decimal of some weights). are close enough. For low returns the SVAR model outper-
At the end of the first quarter of 2010, we use the forms, while for higher returns the MAD models are better.
securities’ actual returns from 01-01-10 to 31-03-10, for The VAR model is always in between.
calculating the actual returns of the obtained Pareto optimal
portfolios. We also calculate the S&P 500 return for the
same quarter. The actual returns of the 51 Pareto optimal
4.3. Run #2: optimization period 1/4/08–31/03/10, test
portfolios from the four models are shown in figure 2.
period 1/4/10‒30/6/10
The direction for generating the Pareto front is from
return maximization towards risk minimization. Therefore, Subsequently, we implement the four models of section 3
the first portfolios (with id = 1, 2, 3…) correspond to those shifting the historical optimization horizon three months
closer to maximum return based on the historical data and ahead, from 01-04-08 to 31-03-10, in order to obtain the

(a) VAR model (b) SVAR model

25 25

20 20
Actual return %

Actual return %

15 15

10 10

5 5

0
0

(c) MAD model (d) SMAD model

25 25

20 20
Actual return %
Actual return %

15 15

10 10

5 5

0 0

Portfolio Portfolio

Figure 2. Ex-post performance of the Pareto optimal portfolios for the 1st period (Q1, 2010).
 Comparative issues between linear and non-linear risk measures 1237

Pareto optimal portfolios and the efficient frontiers. It was VAR and the SVAR model. In addition, these positive
also observed, as in the first quarter, that the results from returns correspond to portfolios that emerge from the Pareto
the MAD and the SMAD model are almost identical front which is in the risk minimization area (i.e. portfolios
regarding the Pareto optimal portfolios. with id >30).
At the end of the second quarter of 2010, we use the As in the previous case, we calculate the cumulative
securities’ actual returns from 01-04-10 to 30-06-10, for distributions across the Pareto front and we depict them In
calculating the actual returns of the obtained Pareto optimal figure 5. As it was mentioned, MAD and SMAD model
portfolios. We also calculate the S&P 500 return for the provide almost identical results and they are depicted with
same quarter. The actual returns of the 51 Pareto optimal the same curve (MAD-SMAD).
portfolios from the four models are shown in figure 4. It is easily observed that now the MAD-SMAD model
The horizontal red line is the performance of the S&P outperforms the VAR and the SVAR models for almost
500 index in the second quarter of 2010. Again, all the entire Pareto fronts. The top 90% of the Pareto
the Pareto optimal portfolios outperform the index which is optimal portfolios from the MAD-SMAD model outper-
‒7.7%. However, the picture is quite different than the first form the top 90% of those of the VAR and the SVAR
quarter. More portfolios have negative returns. The MAD model. The dominance becomes even more apparent in
and SMAD model attain more positive returns than the the top 30%.

25

20
Return (%)

15

10

0
10%

15%

20%

25%

30%

35%

40%

45%

50%

55%

60%

65%

70%

75%

80%

85%

90%

95%

99%
1%

5%

Percentage of Pareto portfolios with return less than vertical axis' value
VAR SVAR MAD-SMAD

Figure 3. Cumulative distribution of ex-post returns for the Pareto fronts 1st period (Q1, 2010).

(a) VAR model 3

(b) SVAR model
3

2 2

1 1

0 0
Actual return %
Actual return %

-1 -1

-2 -2

-3 -3

-4 -4

-5 -5

-6 -6

-7 -7

-8 -8

(c) MAD model (d) SMAD model

7 7
6 6
5 5
4 4
3 3
Actual return %
Actual return %

2 2
1 1
0 0
-1 -1
-2 -2
-3 -3
-4 -4
-5 -5
-6 -6
-7 -7
-8 -8

Portfolio Portfolio

Figure 4. Ex-post performance of the Pareto optimal portfolios for the 2nd period (Q2, 2010).
1238 P. Xidonas and G. Mavrotas

Figure 5. Cumulative distribution of ex-post returns for the Pareto fronts 2nd period (Q2, 2010).

4.4. Run #3: optimization period 1/7/08–30/06/10, test index which is 6.7%. However, the picture is quite different
period 1/7/10–30/9/10 than the second quarter. The SVAR and the VAR models
perform better than those of MAD and SMAD. This is also
We shift the historical optimization horizon three months
confirmed by the cumulative distribution curves depicted in
ahead, from 01-07-08 to 30-06-10, in order to obtain the
figure 7.
Pareto optimal portfolios and the efficient frontiers. The
It is easily observed that now the SVAR model
results from the MAD and the SMAD model are again
dominates, the VAR model is second and the MAD-SMAD
almost identical regarding the Pareto optimal portfolios,
model provides significantly lower returns in the Pareto
confirming the observations of the first two runs.
front.
At the end of the third quarter of 2010, we use the
securities’ actual returns from 01-07-10 to 30-09-10, for
calculating the actual returns of the obtained Pareto optimal
portfolios. The actual returns of the 51 Pareto optimal
4.5. Run #4: optimization period 1/10/08–30/09/10, test
portfolios from the four models are shown in figure 6.
period 1/10/10–31/12/10
As before, the horizontal red line is the performance of
the S&P 500 index (for the third quarter of 2010 this time). In the last run, we shift the historical optimization horizon
Again, all the Pareto optimal portfolios outperform the three months ahead, from 01-10-08 to 30-09-10, in order to

(a) VAR model (b) SVAR model

17 17
16 16
15 15
14 14
13 13
12 12
Actual return %

Actual return %

11 11
10 10
9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0

(c) MAD model (d) SMAD model

13 13
12 12
11 11
10 10
Actual return %

Actual return %

9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0

Portfolio Portfolio

Figure 6. Ex-post performance of the Pareto optimal portfolios for the 3rd period (Q3, 2010).
 Comparative issues between linear and non-linear risk measures 1239

18

16

14

12

Return (%)
10

0
10%

15%

20%

25%

30%

35%

40%

45%

50%

55%

60%

65%

70%

75%

80%

85%

90%

95%

99%
1%

5%

Percentage of Pareto portfolios with return less than vertical axis' value
VAR SVAR MAD - SMAD

Figure 7. Cumulative distribution of ex-post returns for the Pareto fronts 3rd period (Q3, 2010).

obtain the Pareto optimal portfolios and the efficient This is evident from the higher values in the average
frontiers. The results from the MAD and the SMAD model stocks/portfolio. In general, the number of stocks/portfolio
are again almost identical regarding the Pareto optimal port- varies from 16 to 23. Notice that the spontaneous upper
folios, confirming the observations of the first three runs. bound for the selected stocks is implied by the 3%
At the end of the third quarter of 2010, we use the minimum buy-in threshold and it is 34 stocks, while the
securities’ actual returns from 01-10-10 to 31-12-10, for forced upper bound implied by equation (8) is 30 stocks.
calculating the actual returns of the obtained Pareto optimal Therefore, we see that the number of stocks in the portfolio
portfolios. The actual returns of the 51 Pareto optimal does not reach its upper limit in none of the tested models
portfolios from the four models are shown in figure 8. across the four periods. The lower bound is implied by the
As before, the horizontal red line is the performance of 5-10-40 rule and it is 16 stocks. In addition, we observe
the S&P 500 index (for the third quarter of 2010 this time). that the MAD and SMAD model involve systematically
Now the proportion of portfolios that are better than the more stocks in their Pareto front in comparison to the VAR
S&P 500 index is lower for all the models. The correspond- and SVAR models. By the term ‘involved stocks’ we mean
ing cumulative distribution curves are depicted in figure 9. those stocks that appear in at least one of the 51 Pareto
The picture now is mixed with the MAD-SMAD model optimal portfolios. The VAR model has the lower number
to dominate in lower returns and the SVAR model in the of involved stocks.
higher returns. We can also observe that about 80% of A significant issue is the solution time. The VAR model
the portfolios from the MAD-SMAD model are better than is by far the more demanding in computation time as it was
the S&P 500 index, while this share drops to 65% for the also found in previous works (e.g. Konno and Yamazaki
SVAR model. 1991). Moreover, these solution times were obtained with
increased gap threshold (0.5%) in the MIQP solver of
CPLEX. For the other models the gap threshold was set to
0% assuring that the obtained solution is the global
4.6. Results and discussion optimum in MIP and MIQP problems (the gap threshold
The first critical observation is that the MAD and the indicates the difference between the best found solution and
SMAD model provide practically the same Pareto optimal the best possible solution in Integer Programming). The
portfolios. This is in accordance with the theory found in SVAR model is faster than the VAR model in all cases as
the literature regarding the MAD and the SMAD model well as the SMAD model is faster than the MAD. It must
(Speranza 1993). In addition, the absolute equality of the be noticed that although the SVAR model is MIQP, it has
generated Pareto optimal portfolios in such complex prob- competitive solution times with the two MIP models (MAD
lems is a noteworthy confirmation of the theory. Moreover, & SMAD) that are supposed to be easier to solve.
it has to be stressed that on average SMAD is 35% faster Regarding the ex-post evaluation of the models with the
than the MAD calculation. The summary of the results is out-of-sample data, the next three columns of table 2 offer
shown in table 2. The first three columns refer to the opti- fruitful information. This information in combination with
mization phase i.e. the generation of the efficient frontier the relevant charts can lead to the following conclusions.
and the related Pareto optimal portfolios. First, the portfolios of the generated efficient fronts
A second observation regards the number of securities significantly outperform the S&P 500 index with only a
(stocks) in the Pareto optimal portfolios. The MAD and the few exceptions (e.g. the last period). This is an indication
SMAD model systematically add more stocks in the Pareto that using this kind of tools (multiobjective optimization
optimal portfolios than the VAR and the SVAR models. and generation the efficient front with mathematical
1240 P. Xidonas and G. Mavrotas

(a) VAR model (b) SVAR model

14 14
13 13
12 12
11 11
10 10
Actual return %

Actual return %
9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0

(c) MAD model (d) SMAD model

14 14
13 13
12 12
11 11

Actual return %
10
Actual return %

10
9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0

Portfolio Portfolio

Figure 8. Ex-post performance of the Pareto optimal portfolios for the 4th period (Q4, 2010).

16

14

12

10
Return (%)

0
10%

20%

30%

40%

50%

75%

85%

95%
15%

25%

35%

45%

55%

60%

65%

70%

80%

90%

99%
1%

5%

Percentage of Pareto portfolios with return less than vertical axis' value
VAR SVAR MAD- SMAD

Figure 9. Cumulative distribution of ex-post returns for the Pareto fronts 4th period (Q4, 2010).

programming) may keep investors on the safe side. The last front area” that constantly produces the best ex-post
two columns present the average first-quarter return over all performances. From model to model these areas differ.
the Pareto optimal portfolios and the corresponding standard For example, in period 4 for the VAR and the SVAR
deviation. It can be seen that there is no clear winner, as model, the good area (with higher returns) is the maxi-
from period to period things change. In general, when the mum return area (with portfolios 1–12), while for the
market moves upwards, the SVAR model seems to be the MAD and SMAD model the “good area” is the middle
winner. However, when the market moves downwards (e.g. area (portfolios 20–35). Despite this difference, we can
period 2), the MAD and the SMAD models are the best. see that in all the models and across all the periods the
Regarding the dispersion of returns across the Pareto front middle area is constantly on the safe side without big
as quantified by their standard deviation (last column of decreases in the portfolios’ returns. This can be used as
table 2), the SVAR model is the one with the less variation an argument in favour of multiobjective approaches
in all periods except the last one. implying that a safe policy must be pursued away from
Observing the charts with the portfolios’ performance the individual optima (the extremes of the efficient
one may conclude that there is no a particular “efficient frontier).
 Comparative issues between linear and non-linear risk measures 1241

Table 2. Summary results from running the bi-objective model for the four periods.

Optimization phase Evaluation phase

Stocks per portfolio Total number of stocks Solution Proportion greater than Average Standard
Model (average) involved time (s) index (%) return deviation

Run #1 VAR 17.10 47 3712.37 100 15.57 2.40

SVAR 17.31 47 164.41 100 16.26 2.52
MAD 17.41 49 132.58 98 15.78 3.71
SMAD 17.41 49 93.63 98 15.78 3.71
Run #2 VAR 16.96 39 994.73 100 −1.33 1.81
SVAR 17.35 47 221.20 100 −2.19 1.37
MAD 17.47 50 148.85 100 −0.06 2.77
SMAD 17.49 50 78.07 100 −0.05 2.78
Run #3 VAR 17.24 42 3076.83 100 12.75 1.25
SVAR 17.20 46 97.44 98 13.32 1.80
MAD 17.59 50 142.45 88 9.44 1.84
SMAD 17.59 50 101.63 88 9.44 1.84
Run #4 VAR 16.86 43 3220.62 55 9.87 1.68
SVAR 17.24 46 131.59 61 10.69 2.11
MAD 17.59 55 206.43 80 10.54 1.61
SMAD 17.47 55 134.23 80 10.54 1.61

5. Conclusions Canakgoz, N. and Beasley, J., Mixed-integer programming

approaches for index tracking and enhanced indexation. Eur. J.
Oper. Res. 2009, 196, 384–399.
Our purpose in this article was to derive comparative
Chankong, V. and Haimes, Y., Multiobjective Decision Making:
conclusions as far as the application of both linear and Theory and Methodology, 1983 (Elsevier: New York).
non-linear risk measures in non-convex portfolio optimiza- Ehrgott, M. and Ryan, D., Constructing robust crew schedules
tion problems. On this basis, four popular portfolio selection with bi-criteria optimization. J. Multi-Criteria Decis. Anal.,
cases were examined: (a) the mean-variance model, (b) the 2002, 11, 139–150.
Hamacher, H.W., Pedersen, C.R. and Ruzika, S., Finding represen-
mean-semi variance model, (c) the mean-MAD (mean-abso-
tative systems for discrete bi-criterion optimization problems.
lute deviation) model and (d) the mean-semi MAD model. Oper. Res. Lett., 2007, 35, 336–344.
We also aimed at simultaneously co-assessing a set of Hwang, C.L. and Masud, A., Multiple Objective Decision Making.
sophisticated real-world non-convex investment policy Methods and Applications: A State of the Art Survey. Lecture
limitations, such as cardinality constraints, buy-in Notes in Economics and Mathematical Systems, Vol. 164, 1979
(Springer-Verlag: Berlin).
thresholds, transaction costs, particular normative rules, etc.
Konno, H. and Suzuki, K., A fast algorithm for solving large
In all of the above situations, the portfolio selection process scale mean-variance models by compact factorization of
reflects to a mixed-integer bi-objective (or in general multi- covariance matrices. J. Oper. Res. Soc. Jpn., 1992, 35(1),
objective) mathematical programme. We meticulously 93–104.
developed all corresponding modelling procedures and then Konno, H. and Yamamoto, R., Integer programming approaches
in mean-risk models. Comput. Manage. Sci., 2005, 2,
solved the underlying problem by the use of a novel gener-
339–351.
alized algorithm, exclusively designed to cope with the Konno, H. and Yamazaki, H., Mean-absolute deviation portfolio
above-mentioned characteristics. optimization model and its application to Tokyo Stock Market.
The validity of the attempt was verified through empirical Manage. Sci., 1991, 37(5), 519–531.
analysis on the S&P 500 securities. The technical Lee, S. and Chesser, D., Goal programming for portfolio
management. J. Portfolio Manage., 1980, 6(3), 22–26.
conclusions obtained not only confirmed certain findings of
Mansini, R., Ogryczak, W. and Speranza, M.G., LP solvable
the particular limited existing theory but also shed light on models for portfolio optimization: A classification and
computational issues and running times. Moreover, the computational comparison. IMA J. Manage. Math., 2003, 14,
results derived are characterized as encouraging enough, 187–220.
since a sufficient number of efficient or Pareto optimal port- Markowitz, H., Portfolio selection. J. Financ., 1952, 7(1), 77–91.
Markowitz, H. and Perold, A., Portfolio analysis with factors and
folios produced by the models, appear to possess superior
scenarios. J. Financ., 1981, 36(14), 871–877.
out-of-sample returns with respect to the benchmark. Mavrotas, G., Effective implementation of the ε-constraint method
in multiobjective mathematical programming problems. Appl.
Math. Comput., 2009, 213(2), 455–465.
References Miettinen, K.M., Nonlinear Multiobjective Optimization, 1998
(Kluwer Academic: Boston, MA).
Angelelli, E., Mansini, R. and Grazia Speranza, M., A comparison Ogryczak, W., Multiple criteria linear programming model for
of MAD and CVaR models with real features. J. Bank. Financ., portfolio selection. Ann. Oper. Res., 2000, 97, 143–162.
2007, 32(7), 1188–1197. Perold, A., Large scale portfolio optimization. Manage. Sci., 1984,
Branke, J., Scheckenbach, B., Stein, M., Deb, K. and Schmeck, 30(10), 1143–1160.
H., Portfolio optimization with an envelope-based multiobjective Roman, D., Darby-Dowman, K. and Mitra, G., Mean-risk models
evolutionary algorithm. Eur. J. Oper. Res., 2009, 199, 684–693. using two risk measures: A multi-objective approach. Quant.
Finance, 2007, 7(4), 443–458.
1242 P. Xidonas and G. Mavrotas

Sharpe, W., A linear programming approximation for the general Steuer, R., Qi, Y. and Hirshberger, M., Comparative issues in large
portfolio analysis problem. J. Financ. Quant. Anal., 1971, 6, scale mean-variance efficient frontier computation. Decis.
1263–1275. Support Syst., 2011, 51, 250–255.
Speranza, M.G., Linear programming models for portfolio optimi- Xidonas, P. and Mavrotas, G., Multiobjective portfolio optimiza-
zation. Finance, 1993, 14, 107–123. tion with non-convex policy constraints: Evidence from the
Steuer, R., Qi, Y. and Hirshberger, M., Suitable-portfolio investors, Eurostoxx 50. Eur. J. Financ., 2012, doi:10.1080/
non-dominated frontier sensitivity and the effect of multiple 1351847X.2012.733718.
objectives on standard portfolio selection. Ann. Oper. Res.,
2007, 152, 297–317.