Quality & Quantity (2006) 40:457–468 © Springer 2006

DOI 10.1007/s11135-005-1054-0

Portfolio Selection under Maximum Minimum

Criterion*

YUANYAO DING1,2
1
Faculty of Business, Ningbo University, Ningbo 315211, Zhejiang, P.R.China; 2 School
of Statistics, Renmin University of China, Beijing 100872, P.R.China;
E-mail: [email protected].

Abstract. In this paper, we studied the problem of risky portfolio selection under uncertainty.
Different from risk-return analytical methodology, we formulated a model under maximum
minimal criterion of uncertain decision-making theory. If the investor had no any distribu-
tion information of the returns and (s)he knew the variation scopes of the returns by his/her
knowledge of the market information or experts’ evaluations of the alternative risky assets,
then we showed that the optimal portfolio strategy of the model under maximal minimal
criterion could be obtained by solving linear programming. If the returns were known to be
normal distributed, the investor’s optimal portfolio strategy could be obtained by solving a
nonlinear programming. The paper also provided an algorithm to solve this programming.
At last, the paper compared this model with Markowitz’s mean-varience (M-V) model and
Young’s minmax model, and pointed out the distinctions and similarities between our model
and the other two.

Key words: Decision-making, maximum minimum criterion, portfolio selection, uncer-

tainty, maximum loss risk

1. Introduction
There are widespread fields in practice involving optimal portfolio selec-
tion. Prominent examples include (1) asset allocation for pension plans and
insurance companies; (2) social resources allocation and management of
welfare programs; (3) security selection for stock and bond portfolio man-
agers; and (4) risk management for large public corporations. In these and
other situations, the decision maker has to consider uncertainty of out-
comes and balance between bearing risk and seeking return according to
his preference. Portfolio selection theory deals with how to form a satisfy-
ing portfolio among numerous assets.
It has been suggested that modern portfolio theory was introduced
by Markowitz (1952). In the theory of Markowitz, the return of any

∗
Supported in part by Program for NCET, in part by the Key Project of Chinese Min-
istry of Education 104053.
458 YUANYAO DING

portfolio is represented by its expected value or its mean and the risk
is measured with the variance of its return; investors are assumed to
like return and dislike risk; and they will choose a portfolio to max-
imize the expected profit given a certain fixed investment risk level or
choose a portfolio to minimize the investment risk given a certain fixed
return level. This is the famous mean-variance (M-V) model. The M-V
model of Markowitz is a single period static portfolio planning model.
In the past five decades, the M-V model has been extended and devel-
oped in numerous directions which mainly focused in four respects: (1)
To simplify the computational complexity of the M-V model, for exam-
ple, Sharpe (1963) simplified the M-V model by a single index model
to reduce the model parameters, Konno and Yamazaki (1991) presented
the mean-absolute deviation model which can be solved by linear pro-
gramming although the absolute deviation is analogical to standard devi-
ation in meaning, and Wang and Xia (2002) proposed a model with
order of expected returns in the M-V framework. (2) Because people
have different understanding about risk, different risk measures have be
defined and applied to mean-risk analysis of portfolio selection. Arzac and
Bawa (1977) used the loss probability to quantify risk, Fishburn (1977),
Harlow (1991) and Lucas et al. (1998) used downside moment of return to
quantify risk, and Lemus (1999) used quantile-based risk measures. (3) To
extend the M-V model from a single period to multi-periods, for example,
Li and Ng (2000) studied the multi-period M-V model. (4) Several authors,
as Yaoshimoto (1996) and Konno and Wijayanayake (2000), considered
transaction cost in portfolio selection models. In the aforementioned lit-
erature, portfolio selection models are all constructed or discussed in the
framework of return-risk analytical methodology as the M-V model.
However, any optimal choice of assets is based on the uncertainty of
their prices or returns. Since that different investors may behave different,
different investments have different characteristics and the information of
market may be incomplete or asymmetry, there does not exist a portfolio
or portfolio strategy adaptive to all investors. And it is well known that the
return-risk analytical methodology have their limitations in practice.
In the present paper, the maximum minimum criterion in uncertain deci-
sion-making theory is used to portfolio selection and a new model is con-
structed and compared with some other models. Throughout the paper,
short sell is not allowed. This paper is organized as follows. Section 2
presents a new model for portfolio selection under uncertainty by using
maximum minimum criterion; Section 3 gives some applicable remarks on
this new model. Section 4 provides the results for solving the model. Sec-
tion 5 compares the model under maximal minimal criterion with the
M-V model of Markovitz and Young’s minmax portfolio selection model.
Section 6 summarizes the conclusions.
PORTFOLIO SELECTION UNDER MAXIMUM MINIMUM CRITERION 459

2. Portfolio Selection Model under Maximum Minimum Criterion

Assume that n risky assets Si (i = 1, 2, . . . , n) with random return rate
Ri (i = 1, 2, . . . , n), respectively, are available for the investor during the
planning horizon.
Let xi be the fraction of the initial wealth allocated to asset i
(i = 1, 2, . . . , n). If there does not exist a riskless asset for investment, then
the vector of the weights X = (x1 , . . . , xn )T ∈ R n should satisfy to

eT X = 1, X ≥ 0. (1)

Here, and subsequently, e = (1, 1, ..., 1)T is the vector of all 1s of the appro-
priate dimension. The Equation (1) implies that the decision maker invests
his entire fund to the risky assets. The inequality of (1) implies that it is
not allowed to sell short for risky assets. X = (x1 , . . . , xn )T is called an invest
strategy or feasible portfolio if it satisfies
(1).
Let R = (R1 , R2 , . . . , Rn )T , r(R|X) = ni=1 xi Ri = XT R. For an investor
with a utility function of deterministic wealth increment W: u = u(W ), if
he takes an invest strategy X at the beginning of the time period, he
would get an actual utilty value u(r(R|X)w0 ) at the end of the period,
here w0 is the total initial wealth to invest. By expected utility theory, the
rational investor would choose X to maximize his expected utility U (X) =
E[u(r(R|X)w0 )] while R is random or uncertain with a known probability
distribution. Markowitz’s M-V model was derived under the case that the
investor’s utility is a quadratic function or R is normal distributed.
However, because it is difficult to specify the utility function accurately,
it is almost impossible to calculate expected value of the utility function
even though the distribution of random vector R is given. Further more, it
is often very difficult in the real world to describe exactly the distribution
of R owing to the short of historical materials or difficult to get sample
materials of some risky assets (e.g., R & D items), and the incomplete and
asymmetry of market information. In deed, the only thing people can do
is to give an estimation or assumption of the distribution or give a credi-
ble variation scope of R depending on experts’ experience and knowledge
or subjective belief of those available assets to be invested; especially the
assets are such as new technology or new production.
Assume that before making decision, by means of quality forecasting
method such as Delphi method including experts experience and scenario
analysis method or some other forecasting techniques, the investor could
have knowledge that there exists a credible scope D ⊂ R n with a certain
credible level not less than α (subjective or objective probability) in which
the uncertain return vector R would realize in the future. And no other
information could be available for making decision. In this situation, as an
investor of risk or uncertainty aversion, he may choose an optimal or a
460 YUANYAO DING

satisfactory portfolio X to satisfy the following model:

max min[u(r(R|X)w0 )]
X R
(2)
R∈D
s.t.
eT X = 1, X ≥ 0.
In fact, because the utility function is monotonically increasing, model (2)
can be expressed as the following equivalent form:
max min[r(R|X)]
X R
(2’)
R∈D
s.t.
eT X = 1, X ≥ 0.
This is a portfolio selection model under maximal minimal criterion,
here r(R|X) = XT R.
• This model does not depend on the form of utility function, but the
restricted condition R ∈ D may be different with different investors and
depends on the market information and the investor’s knowledge.
• In a complete market, if R ∈ D is only the revealed information by
the market, the portfolio derived by this model could be used as the
benchmark portfolio for any investors.
• If P (R ∈ D) ≥ α and R ∗ is the solution of this model, then −r(R ∗ |X)
implies the maximum loss value of portfolio X in a certain confidence
level (refer to Studer (1999) for the maximum loss as a risk measure, it
is similar to VaR) and thus optimal portfolio X∗ has the minimal risk
of maximal loss. Thus this model is to minimize the maximum loss risk
value.

3. Some Applicable Remarks of the Model

In practice of risky portfolio selection, usually the decision maker would
make in advance the evaluation, forecast or estimation of the risky assets
and provide a lower level or variation scope of return rates of the assets.
In pervasive situation, the D in the restricted condition of model (2)
may be given after forecasting or project evaluating in one of the follow-
ing two forms:
(i) D = D1 = {R|ai ≤ Ri ≤ bi , i = 1, 2, . . . , n}, here ai and bi (i = 1, 2, . . . , n)
are given constants;
(ii) D = D2 = {R|ci ≤ Ri , i = 1, 2, . . . , n}, here ci (i = 1, 2, . . . , n) are given
constant.
In some special situations, R is assumed as multivariate normal distribu-
tion. Denote E(Ri ) = µi (i = 1, 2, . . . , n), µ = (µ1 , µ2 , . . . , µn )T , and cov(R) =
PORTFOLIO SELECTION UNDER MAXIMUM MINIMUM CRITERION 461

 = (σij2 )n×n ,  > 0. Then µ and  could be estimated or forecast in some

way.
Because (R − µ)T  −1 (R − µ) ∼ χ 2 (n), it has P ((R − µ)T  −1 (R − µ) ≤
χα (n)) = α or P ((R − µ)T  −1 (R − µ) ≥ χ1−α
2 2
(n)) = α, where χα2 (n) is the
α quartile of χ distribution with free degree of n. Now the D in the
2

restricted condition of model (2) may be given by one of the following two
forms:
 
(iii) D = D3 = R|(R − µ)T  −1 (R − µ) ≤ χα2 (n) , 
(iv) D = D4 = R|(R − µ)T  −1 (R − µ) ≥ χ1−α2
(n) .
For each given X, which satisfies to (1), let R ∗ = (R1∗ , R2∗ , . . . , Rn∗ )T be a
solution of the following linear programming:

min XT R, s.t. R ∈ D. (3)

R

Thereby R ∗ might be dependent on X. Then it is obvious that the solution

of model (2 ), X ∗ = (x1∗ , x2∗ , . . . , xn∗ )T , can be obtained by solving the follow-
ing programming:
max XT R ∗
X (4)
s.t. eT X = 1, X ≥ 0.
Up to now, to solve model (2) has been transformed to solve (3) and (4)
iteratively.

4. Optimal Portfolio under Maximal Minimal Criterion

In order to solve model (2’) for getting the optimal portfolio under max-
imal minimal criterion, let’s consider model (3) first. By denoting A =
(a1, a2 , . . . , an )T , B = (b1 , b2 , . . . , bn )T , C = (c1 , c2 , . . . , cn )T , we have the
following results:

LEMMA 1.
(i) If D = D1 = {R|A ≤ R ≤ B}, then the solution of model (3) R ∗ exists
and
1
[(b + ai ) − sign(xi )(bi − ai )] for xi = 0,
Ri = 2 i
∗
arbitrary for xi = 0.
Here sign() is the sign function.
(ii) If D = D2 = {R|C ≤ R} and X ≥ 0, then the solution of model (3) exists
and:

∗ ci for xi > 0,
Ri =
arbitrary for xi = 0.
462 YUANYAO DING

(iii) If D = D3 = {R|(R − µ)T  −1 (R − µ) ≤ χα2 (n)}

√ and X = 0, then the solu-
χ 2 (n)
tion of model (3) R ∗ exists and R ∗ = µ − √ Tα X.
X X
(iv) If D = D4 = {R|(R − µ)T  −1 (R − µ) ≥ χ1−α
2
√(n)} and X = 0, then the
χ 2 (n)
solution of model (3) exists and R ∗ = µ + √ 1−α X.
XT X

Proof. It is obvious that (i) and (ii) hold.

For (iii), let f1 (R) = X T R, g1 (R) = (R − µ)T  −1 (R − µ) − χα2 (n), then
while D = D3 model (3) can be expressed as the normal programming prob-
lem:
min f1 (R), s.t. g1 (R) ≤ 0. (5)
R

The corresponding Kuhn–Tucher conditions are:

∇f1 (R ∗ ) = −λ∇g1 (R ∗ ),
λg1 (R ∗ ) = 0,
(6)
g1 (R ∗ ) ≤ 0,
λ ≥ 0.

Since ∇f1 (R ∗ ) = X,
√ ∇g
∗ −1 ∗
1 (R ) = −2 (R − µ) and X = 0, it is easy to
√
χ (n)
2
XT X
verify that R ∗ = µ − √ Tα X and λ = √ satisfy (6). Because f1 (R) is
X X 2 χα (n)
2

linear and g1 (R)

√ 2is convex so that problem (5) is a convex programming,
∗ χ (n)
thus R = µ − √ Tα X is the global solution of programming (5).
X X
For (iv), since the process is similar to prove (iii), it is omitted.

Subsequently let’s turn to discuss the solution of model (2’).

4.1. d = d1
In this case, the portfolio selection model under maximal minimal criterion
is:
max min [XT R]
X R
s.t. A≤R≤B (7)
eT X = 1, X ≥ 0.
From (i) of Lemma 1, the optimal portfolio solution of model (7) can be
obtained by solving the following linear programming:
max XT A
X (8)
s.t. eT X = 1, X ≥ 0.
While ai (i = 1, 2, . . . , n) are not equal, model (8) will have a unique solu-
tion.
PORTFOLIO SELECTION UNDER MAXIMUM MINIMUM CRITERION 463

4.2. d = d2
In this case, the portfolio selection model under maximal minimal criterion
is:
max min [XT R]
X R
s.t. C ≤R (9)
eT X = 1, X ≥ 0.
From (ii) of Lemma 1, the optimal portfolio solution of model (9) can be
obtained by solving the following linear programming:
max XT C
X (10)
s.t. eT X = 1, X ≥ 0.
While ci (i = 1, 2, . . . , n) are not equal, model (10) will have a unique solu-
tion.

4.3. d = d3
In this case, the portfolio selection model under maximal minimal criterion
can be expressed as:

max min [XT R]

X R
s.t. (R − µ)T  −1 (R − µ) ≤ χα2 (n) (11)
eT X = 1, X ≥ 0.
From (iii) of Lemma 1, the optimal portfolio solution of model (11) can
be obtained by solving the following nonlinear programming:
√
max XT µ − χα XT X
X (12)
s.t. XT e = 1, X ≥ 0.

Here χα = χα2 (n). In order to solve (12), let’s consider the following
programming first:
√
max XT µ − χα XT X
X (13)
s.t. XT e = 1.

Proposition 1. Let a = eT  −1 e, b = µT  −1 e, c = µT  −1 µ, then while

the optimal portfolio strategy X ∗ of programming (13) exists and
χα2 ≥ c,
   
∗ 1 −1 1
X =  µ+ aχα − (ac − b ) −b e .
2 2 (14)
aχα2 − (ac − b2 ) a
464 YUANYAO DING


Proof. It is easy to verify that there exist λ∗ = a1 aχα2 − (ac − b2 ) − b

and X∗ = √ 2 1  −1 µ + a1 aχα2 − (ac − b2 ) −b e which satisfy to
aχα −(ac−b )
2

the Kuhn–Tucker Conditions of (13):

√χα X − µ + λe = 0,
X T X
λ(1 − XT e) = 0,
λ ≥ 0.

Thus X∗ is the solution of (13).

By considering the relationships between the solution of model (12) and
that of model (13), we get the following results:

T
Proposition 2. Let x ∗ = x1∗ , x2∗ , . . . , xn∗ be the optimal solution of
model (13), then
(a) if x ∗ ≥ 0, then x ∗ is also the optimal solution of model (12);
(b) if there exist some negative components in vector x ∗ , denoted by
xi1 ∗ , xi2 ∗ , . . . , xik ∗, then the optimal solution of model (12) must be in
k
the set of C = Cit , where Cit is the set consisted of all the feasible
t=1
solutions with xit = 0 of model (13).

Proof. The proof of (a) is obvious.

For (b), If we suppose x̃ = (x̃1 , x̃2 , . . . , x̃n )T ≥ 0 is the optimal solution of
k
model (12) and x̃ ∈
/ C, that is x̃ ∈ C̄it, then the components of x̃ satisfy
t=1
to x̃it > 0 (t = 1, 2, . . . , k) and the left components√ of x̃ are all nonnegative.
Let x̂(λ) = λx̃ + (1 − λ)x ∗ , g(λ) = x̂ T µ − χα x̂ T  x̂, λ ∈ [0, 1], it is easy to
see that x̂ is the feasible solution of model (13).
√ −1  −1

Because g (λ) = −χα T
x̂  x̂ (x̃ − x ∗ )T  (x̃ − x ∗ ) − x̂ T  x̂ ×
 2   2   
(x̃ − x ∗ )T  x̂ , and (x̃ − x ∗ )T  x̂ < (x̃ − x ∗ )T  (x̃ − x ∗ ) x̂ T  x̂ for
 > 0, we have g  (λ) < 0 and therefore g(λ) is a strict convex function. Fur-
ther, because x ∗ is the optimal solution of model (13), we have g(0) > g(λ)
for any λ ∈ (0, 1], and thus g(λ)is a strict decreasing function of λ ∈ [0, 1].
It follows g(λ) > g(1) for any λ ∈ [0, 1).
Because x̂(λ) is continuous about λ, and x̂(1) = x̃ ,there exists a λ close
to 1 so that the components of x̂(λ) satisfy to x̂it > 0 (t = 1, 2, . . . , k) with
the other components nonnegative. Therefore x̂(λ) is also a feasible solu-
tion of model (12) and
 √
g(λ) = x̂(λ)T µ − χα x̂(λ)T  x̂(λ) > x̃ T µ − χα x̃ T  x̃ = g(1).
PORTFOLIO SELECTION UNDER MAXIMUM MINIMUM CRITERION 465

This contradicts to that x̃ = (x̃1 , x̃2 , . . . , x̃n )T ≥ 0 is the optimal solution of

model (13).
Hence the assumption x̃ ∈ / C is not correct and part (b) is proved true until
now.

According to Propositions 1 and 2, the optimal portfolio solution of
model (12) can be obtained by the following steps of algorithm:
Step 1: To solve model (13) and get the solution x ∗ by using formula
(14).
Step 2: If x ∗ ≥ 0, then x ∗ is the solution of model (12) and stop the pro-
∗ ∗ ∗
cess, else if x ∗ has negative components xi1 , xi2 , . . . , xik , then go
to next step.
Step 3: In model (13), after adding a constraint xij = 0 to it (that means
to cancel out the asset Sij from the n alternative assets), we can
get a new solution of model (13) denoted by X−ij and the corre-
sponding objective value denoted by Oij (j = 1, 2, . . . , k ). If Oit =
max{Oij , j = 1, 2, . . . , k}, then let x ∗ = X−it and go to Step 2.
j

4.4. d = d4
In this case, the portfolio selection model under maximal minimal criterion
can be expressed as:

max min
R
[XT R]
X
s.t. (R − µ)T  −1 (R − µ) ≥ χ1−α
2
(n) (15)
T
e X = 1, X ≥ 0.

From (iv) of Lemma 1, the optimal portfolio solution of model (15) can be
obtained by the following nonlinear programming:
√
max XT µ + χ1−α XT X
X (16)
s.t. XT e = 1, X ≥ 0.

By substituting constant χα with constant −χ1−α in model (12), we can

get model (16). Therefore, the procedure to solve model (16) is the same as
that stated in Section 4.3.

5. Comparison with other Models

5.1. comparison with m-v model
Markowitz (1952) presented a quantitative approach of portfolio analysis.
His prescription still remains the dominant technique in use today. Marko-
witz formulated the portfolio choice problem by the so-called M-V model,
466 YUANYAO DING

namely to maximize expected return for a given level of variance and to

minimize variance for a given level of expected return. If we use the same
notations above mentioned, the M-V model can be expressed as follows:

max XT µ
X
s.t. X T X ≤ σ02 (17)
XT e = 1, X ≥ 0.

Or equivalent as

min XT X
X
s.t. X T µ ≥ µ0 (18)
XT e = 1, X ≥ 0.

Or in the equivalent form with the same efficient frontier:

max XT µ − 21 λXT X
X (19)
s.t. XT e = 1, X ≥ 0,

where parameter σ02 represents the maximum variance level which can be
accepted by an investor, parameter µ0 represents the minimal return level
required by an investor, and parameter λ represents the constant absolute
risk aversion degree of an investor.
Arguments have been made that the M-V model is appropriate only
if the investor’s utility is quadratic or the joint distribution of return is
normal.
From the above, we know that our new model is independent of the
investor’s utility and especially suitable for the situation that sample or
historical data of returns is difficult to fetch. If the joint distribution of
returns was known as normal and both mean and variance-covariance of
the returns could be estimated, the optimal portfolio given by our model
might be also M-V efficient. The latter could be seen from comparing
model formula (12) with model formula (19), and its proof is omitted.

5.2. comparison with young’s minimax model

Young (1998) introduced a minimax portfolio selection criterion under
which the optimal portfolio minimizes the maximum loss (or negative gain)
over all the past historical periods, subject to a restriction on the mini-
mum acceptable average return across observed time periods. The optimal
PORTFOLIO SELECTION UNDER MAXIMUM MINIMUM CRITERION 467

portfolio is the solution of the following linear program:

max Mp
Mp ,X


n
s.t. xi Rit − Mp ≥ 0
i=1

n (20)
xi R̄i ≥ G
i=1

n
xi ≤ 1, xi ≥ 0, i = 1, 2, . . . , n,
i=1

where Mp is the minimal return on portfolio, Rit the return on one dollar
invested in security i in time period t, R̄i the average return on security i,
and G is the minimum required average return rate on the portfolio.
As Young said, the minimax portfolio is optimal with respect to the
data set {Rit }. If one lacks both historical data on past returns and a pre-
dictive probability model for future returns (then the data can be simu-
lated), the minimax method will not be applicable.
From our model mentioned above, we know that our model could be
applicable even if we have no historical data or a probability model to
predict future returns. If investor has historical data on past returns, then
(s)he might define the variation scope of the future returns on the his-
torical data. In this case, by formulae (7) and (9), the optimal portfolio
under our model could be given by solving the following equivalent linear
programming:
max Mp
Mp ,X


n
s.t. xi Rit − Mp ≥ 0 (21)
i=1

n
xi = 1, xi ≥ 0, i = 1, 2, . . . , n.
i=1

Comparing model (21) with model (20), we can find that the minimal
return of the portfolio defined by model (20) will be not less than that of
the portfolio defined by model (21), even though the average return of the
former might be lager than that of the latter.

6. Conclusions
In this paper, we studied the problem of risky portfolio selection under
uncertainty. Different from risk-return analytical methodology, we formu-
lated a model under maximum minimum criterion of uncertain decision-
making theory. If the investor had no any distribution information of the
returns and he knew the variation scopes of the returns by his knowledge
468 YUANYAO DING

of the market information or experts’ evaluation of the alternative risky

assets, then we proved that the optimal portfolio strategy of the model
under maximum minimum criterion could be obtained by solving linear
programming. If the returns were known to be normal distributed, the
investor’s optimal portfolio strategy could be obtained by solving a qua-
dratic programming. The paper also provided an algorithm to solve this
quadratic programming, which is based on the fact that this quadratic
programming has an analytical solution while short sell of risky assets is
allowed. At last, the paper compared this model with Markowitz’s M-V
model and Young’s minmax model, and pointed out the distinctions and
similarities between our model and the other two.

References
Arzac, E. & Bawa, V. (1977). Portfolio choice and equilibrium in capital markets with safety-
first investers. Journal of Financial Economics 4: 277-288.
Fishburn, P. C. (1977). Mean-risk analysis with risk associated with below-target returns.
American Economic Review 67: 116–126.
Harlow, W. V. (1991). Asset allocation in a downside risk framework. Financial Analysis
Journal 47: 28–40.
Konno, H. & Yamazaki, H. (1991). Mean absolute deviation portfolio optimization model
and its application to Tokyo stock market. Management Science 37: 519–531.
Konno, H. & Wijayanayake, A. (2000). Portfolio optimization problem under concave trans-
action costs and minimal transaction unit constraints. Mathematical Programming Series
B 89: 205–216.
Lemus R. G. (1999). Portfolio optimization with quantile-based risk measures, Ph.D. Thesis,
Massachusetts Institute of Technology.
Li, D. & Ng, W. (2000). Optimal dynamic portfolio selection: multi-period mean-variance
formulation. Mathematical Finance 10: 387–406.
Lucas, A. & Klaassen, P. et al. (1998). Extreme returns, downside risk and optimal asset
allocation. Journal of Portfolio Management 25: 71–79.
Markowitz, H. (1952). Portfolio selection. Journal of Finance 7: 77–91.
Sharpe, W. F. (1963). A simplified model for portfolio analysis. Management Science 9: 277–
293.
Studer, G. (1999). Risk measurement with maximum loss. Mathematical Methods of Opera-
tions Research 50: 121–124.
Wang, S. & Xia, Y. (2002). Portfolio Selection and Asset Pricing. New York: Springer.
Young, M. R. (1998). A minimax portfolio selection rule with linear progamming solution.
Management Science 44: 673–683.
Yaoshimoto, A. (1996). The mean-variance approach to portfolio optimization subject to
transition costs. Journal of the Operations Research Society of Japan 39: 99–117.