This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 1

Mean-Semi-Entropy Models of Fuzzy Portfolio

Selection
Jiandong Zhou, Xiang Li∗ , Witold Pedrycz, Fellow, IEEE,

Abstract—In this paper, a concept of fuzzy semi-entropy is almost all investors would prefer a portfolio with a larger
proposed to quantify the downside uncertainty. Several prop- third order moment if the first and second moments are the
erties of fuzzy semi-entropy are identified and interpreted. By same. Then skewness was defined as the third central moment
quantifying the downside risk with the use of semi-entropy, two
mean-semi-entropy portfolio selection models are formulated and to measure the asymmetry degree of return distributions, and
a fuzzy simulation-based genetic algorithm is designed to solve mean-variance-skewness models were studied (Chunhachinda
the models to optimality. We carry out comparative analyses et al. [5]).
among the fuzzy mean-entropy models and the fuzzy mean-semi- All aforementioned portfolio selection models were con-
entropy models and demonstrate that the mean-semi-entropy structed within the framework of probability theory, and
models can significantly improve the dispersion of investment.
Several illustrative examples using stock dataset from the real- characterized security returns as random variables. However,
world financial market (China Shanghai Stock Exchange) also Lacagnina and Pecorella [12] pointed out that many non-
show the effectiveness of the models. probabilistic factors exist commonly in real financial markets
Index Terms—Fuzzy semi-entropy, Portfolio selection, Mean- so that the returns of risky securities contain other aspects of
semi-entropy model. uncertainty such as vagueness and ambiguity. Wang and Zhu
[35] suggested that the variations of stock markets can be more
efficiently addressed if fuzzy set based approach, quantitative
I. I NTRODUCTION and qualitative analysis, the expert’s knowledge and the man-
ager’s subjective opinions were better integrated into a port-
I N portfolio selection, we study how to allocate investor’s
wealth among a basket of securities to achieve a trade-
off between return and risk. Since Markowitz [24] proposed
folio selection model. Subsequently many stochastic portfolio
selection models are extended to handle uncertainty described
to measure investment return as the expected value and by fuzzy sets [8], [9], [15], [16], [18], [30], [34], [36].
measure investment risk as the variance from the expected For example, Huang [8] proposed the fuzzy mean-variance
value, numerous mean-variance models have been developed portfolio selection models subject to experts’ estimates, Li et
[3], [28], [32]. Noting that using variance as risk measure al. [15], [18] defined the skewness for fuzzy variables and
may sacrifice higher return, semivariance was introduced by proposed a fuzzy mean-variance-skewness model, and Vercher
Markowitz [25], which determines the undesirable deviation and Bermudez [34] proposed a fuzzy mean-downside risk-
of portfolio returns from the expected value. There have been skewness model for efficient portfolio selection.
studies on the properties and computation methods using semi- The study carried out by Philippatos and Wilson [29] was
variance. It has been shown that semi-variance is more reliable the pioneering work to associate entropy with a measure of
and convincing as a downside risk measure [4], [11], [23], risk in portfolio optimization. Simonelli [33] showed that
[26]. The variance and semi-variance based portfolio selection entropy regarded as a risk measure in portfolio is more
models depend only on the first and second order moments of convincing than variance to better wealth allocation strategy.
return distributions, which are not sufficient to explain higher Huang [7] employed fuzzy entropy as an alternative risk
moments associated with non-normal distributions. Arditti [1] measure and formulated two mean-entropy models. However,
and Arditti and Levy [2] asserted that higher moments cannot entropy expresses the uncertainty of both low and high extreme
be ignored unless the security returns are normally distributed. returns, which evidently will sacrifice the higher returns. The
Samuelson [31] showed that higher moments become relevant uncertainty that investors dislike is that the portfolio return
for investors to make decisions in portfolio selection and takes values lower than the expected return. Inspired by this
point of view, in this study, semi-entropy for fuzzy variables
J.D. Zhou is with the School of Economics and Management, Bei- is defined and fuzzy portfolio selection problems using semi-
jing University of Chemical Technology, Beijing 100029, China (e-mail: entropy as risk measure are discussed.
[email protected]).
X. Li is with the School of Economics and Management, Beijing The paper is structured as follows. In Section II, we reviews
University of Chemical Technology, Beijing 100029, China (e-mail: lixi- some preliminary concepts of fuzzy variables and credibility
[email protected]). measures. In Section III, we define semi-entropy for fuzzy
W. Pedrycz is with the Department of Electrical and Computer Engineering,
University of Alberta, Edmonton, T6R 2V4, AB, Canada; Department of variables and discuss its several properties. In Section IV,
Electrical and Computer Engineering, Faculty of Engineering, King Abdulaziz two fuzzy mean-semi-entropy portfolio selection models are
University, Jeddah, 21589, Saudi Arabia; Systems Research Institute, Polish proposed while in Section V, a fuzzy simulation-based genetic
Academy of Sciences, Warsaw, Poland (e-mail: [email protected]).
Manuscript received June 05, 2015; revised August 03, 2015; accepted algorithm is presented. We use several illustrative examples to
January 02, 2016 show the effectiveness of the proposed models and solution

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 2

method (Section VI). Finally, some conclusions and future The variance provides a measure of dispersion of the fuzzy
study directions are presented in Section VII. variable around its expected value.
Fuzzy entropy is employed to characterize the uncertainty
II. P RELIMINARIES
resulting from linguistic vagueness of the possible values of
In order to measure the chance of occurrence of a fuzzy fuzzy variables. Within the framework of credibility theory, Li
event, Zadeh defined possibility measure [37] and necessity and Liu [13] proposed Shannon-like entropy for both discrete
measure [38]. Although both of them have been shown to sat- variables and continuous variables. If ξ is a discrete fuzzy
isfy the properties of normality, nonnegativity and monotonic- variable taking values in {x1 , x2 , · · ·}, its entropy is defined
ity, they are not dual, i.e., the sum of possibilities/necessities as
for a fuzzy event and its complement is not equal to one. X∞

The duality is intuitive and important in the theory, which H[ξ] = S(ν(xi )), (5)
ensures the consistency with the law of excluded middle i=1

and the law of contradiction. In addition, human thinking is where S(t) = −t ln t−(1 − t) ln(1 − t) (See Fig. 1). If ξ is a
always dominated by the duality. In order to construct a dual continuous fuzzy variable, its entropy is defined as
measure, Liu and Liu [19] defined credibility measure as the
Z +∞
average of possibility and necessity measures. This measure
was redefined by Li and Liu [14] as a set function satisfying H[ξ] = S(ν(x))dx. (6)
−∞
the normality, monotonicity, duality and partial maximality
axioms. Nowadays, credibility measure has been applied to The continuous and differentiable function S(t) has the
fuzzy variables, such as expected value (Liu and Liu [19]), following properties: it is nonnegative, unimodal and concave;
variance (Liu [21]), skewness (Li et al. [15]), entropy (Li and it takes the minimum value 0 at t = 0 and t = 1; it takes the
Liu [13]), cross-entropy (Qin et al. [30]) and others. maximum value ln 2 at t = 0.5.
Let Θ be a nonempty set, and let Γ be a σ-algebra over
Θ. Each element A ∈ Γ is called an event. It is necessary S(t)
to assign each event A a real number Cr{A} ∈ [0, 1], i.e.,
the credibility measure, which indicates the chance that A
will occur. In order to ensure that the set function satisfies ln2
certain mathematical properties, Li and Liu [14] proposed the
following four axioms:
Axiom 1. Cr{Θ} = 1.
Axiom 2. Cr{A1 } ≤ Cr{A2 } whenever A1 ⊆ A2 .
c
Axiom 3. Cr{A} S + Cr{A } = 1 for any event A.
t
Axiom 4. Cr{ i Ai } = supi Cr{Ai } for any collection of
events {Ai } with supi Cr{Ai } < 0.5. 0 0.5
Suppose that ξ is a fuzzy variable with membership function
µ. Then for any set B of <, the credibility of ξ ∈ B was Fig. 1. Function S(t) = −t ln t − (1 − t) ln(1 − t)
defined by Liu and Liu [19] as
1  The membership functions and entropies for equipossible
Cr{ξ ∈ B} = sup µ(x) + 1 − sup µ(x) . (1) fuzzy variable, triangular fuzzy variable and trapezoidal fuzzy
2 x∈B x∈B c
variable are summarized in Table I. For more details on
Note that the coefficient 1/2 ensures that the duality property credibility measure and fuzzy entropy, the reader may refer
holds. The function ν(x) = Cr{ξ = x} with x ∈ < is to [17].
called the credibility function (Li [17]). Conversely, if ξ is
a fuzzy variable, then its membership function is derived TABLE I
from the credibility function by taking the expression µ(x) = M EMBERSHIP FUNCTIONS AND THEIR ENTROPIES
min{2ν(x), 1} for all x ∈ <.
Fuzzy variable Membership function Entropy
The expected value [19] of fuzzy variable ξ is 1, if a ≤ x ≤ b
Z +∞ Z 0 (a, b) (b − a) ln 2
0, otherwise.
E[ξ] = Cr{ξ ≥ x}dx − Cr{ξ ≤ x}dx (2) x−a
0 −∞ , if a ≤ x ≤ b
b−a
provided that at least one of the two integrals is finite, which (a, b, c) c−x (c − a)/2
, if b < x ≤ c
is a type of Choquet integral. c−b
0, otherwise.
The variance [21] of fuzzy variable ξ is x−a
, if a ≤ x ≤ b
b−a
V [ξ] = E (ξ − E[ξ])2
 
(3) 1, if b < x ≤ c (b − a + d − c)/2+
(a, b, c, d)
which is actually the expected value of the nonnegative fuzzy d−x (c − b) ln 2
, if c < x ≤ d
variable (ξ − E[ξ])2 as d−c
Z ∞ 0, otherwise.

V [ξ] = Cr{(ξ − E[ξ])2 ≥ x}dx. (4)

0

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 3

III. S EMI - ENTROPY Let ξ be a continuous fuzzy variable with some finite
In this section, we present the definition of fuzzy semi- expected value e. Since S(0) = 0, the semi-entropy can be
entropy and prove its important properties. rewritten as Z e


Definition 3.1: Suppose that ξ is a discrete fuzzy variable Sh [ξ] = S ν(x) dx. (12)
−∞
taking values in {x1 , x2 , · · ·} with some finite expected value
e. Then the semi-entropy is defined as Theorem 3.4: Let ξ be a continuous fuzzy variable with
∞ some finite expected value e. Then we have Sh [ξ] > 0.
Proof: Without any loss of generality, we assume that e ≥ 0. It
X
S ν(xi )− ,


Sh [ξ] = (7)
i=1
follows from the nonnegativity of function S(t) that Sh [ξ] ≥ 0.
Especially, if Sh [ξ] = 0, we have ν(x) = 0 for all x ≤ e,
where S(t) = −t ln t − (1 − t) ln(1 − t) and which implies that Cr{ξ ≤ x} = 0 for all x ≤ 0, Cr{ξ ≥
ν(xi ), if xi ≤ e x} = 1 for all 0 ≤ x ≤ e and
(
ν(xi )− = (8) Z +∞ Z +∞
0, otherwise. E[ξ] = Cr{ξ ≥ x}dx = e + Cr{ξ ≥ x}dx > e.
0 e
Let ξ be a discrete fuzzy variable taking values in
{x1 , x2 , · · ·} with some finite expected value e. Since S(0) = The contradiction indicates that Sh [ξ] > 0.
0, the semi-entropy can be rewritten as Theorem 3.5: Let ξ be a continuous fuzzy variable with
X
some finite expected value e. Then for any real numbers a, b
Sh [ξ] = S ν(xi ) . (9) with a ≥ 0, we have
xi ≤e
Sh [aξ + b] = aSh [ξ]. (13)
Theorem 3.1: Suppose that ξ is a discrete fuzzy variable
Proof: If a = 0, then (13) is trivial. If a > 0, we have E[aξ +
taking values in {x1 , x2 , · · ·} with some finite expected value
b] = ae + b and Cr{aξ + b = x} = ν((x − b)/a) for any
e. Then we have Sh [ξ] ≥ 0.
x ∈ <. Then it follows from the definition of semi-entropy
Proof: It follows immediately from the nonnegativity of
that
function S(t). Z ae+b
Theorem 3.2: Suppose that ξ is a simple fuzzy variable Sh [aξ + b] = S ν((x − b)/a) dx


taking values in {x1 , x2 , · · · , xN } with some finite expected
Z−∞e
value e. Then we have Sh [ξ] ≤ n ln 2, where n is the overall

= aS ν(x) dx
number of xi satisfying xi ≤ e. The equality holds if and only −∞
if ν(xi ) = 0.5 for all xi ≤ e.
= aSh [ξ].
Proof: Since function S(t) reaches its maximum value equal
to ln 2 at t = 0.5, we have The proof is complete.
X
Theorem 3.6: Let ξ be a continuous fuzzy variable with
Sh [ξ] = S ν(xi ) ≤ n ln 2, some finite expected value e. Then we have Sh [ξ] ≤ H[ξ].
xi ≤e Proof: It follows from the nonnegativity of function S(t) that
and the equality holds if and only if ν(xi ) = 0.5 for all xi ≤ e. Z e


Z +∞


The proof has been completed. Sh [ξ] = S ν(x) dx ≤ S ν(x) dx = H[ξ]. (14)
−∞ −∞
Theorem 3.3: Suppose that ξ is a discrete fuzzy variable
taking values in {x1 , x2 , · · ·} with some finite expected value For a symmetrical continuous fuzzy variable, since ν(e +
e. Then we have Sh [ξ] ≤ H[ξ]. x) = ν(e − x) for all x ≥ 0, the semi-entropy takes half of
Proof: According to (8), we have ν(xi )− = ν(xi ) if xi ≤ the value of entropy.

e,
and ν(xi
)− = 0 if xi > e, which implies that S ν(xi )− ≤ Example 3.1: The semi-entropy values for equipossible
S ν(xi ) for all i and then Sh [ξ] ≤ H[ξ]. The proof is fuzzy variable, triangular fuzzy variable and trapezoidal fuzzy
complete. variable are shown as follows. The detailed calculations are
Theorem 3.3 states that semi-entropy assumes smaller val- given in the Appendix.
ues than entropy. The reason is that semi-entropy expresses Let ξ = (a, b) be an equipossible

fuzzy variable. Then its
only the downside uncertainty, while entropy captures the semi-entropy is (b − a) ln 2 /2.
overall uncertainty. Let ξ = (a, b, c) be a triangular fuzzy variable with expected
Definition 3.2: Suppose that ξ is a continuous fuzzy variable value e = (a + 2b + c)/4. Define ζ(x) = x2 ln x − (1 −
with some finite expected value e. Then its semi-entropy is x)2 ln(1 − x). Then it has been proved that the semi-entropy
defined as is
Z +∞ 
Sh [ξ] = S ν(x)− dx,


(10)  (b − a)ρ − (b − a)ζ(ρ), if e ≤ b

−∞ Sh [ξ] = b−a (15)
where S(t) = −t ln t − (1 − t) ln(1 − t) and

 + (c − b)ζ(τ ), otherwise
2
ν(x), if x ≤ e
(
where ρ = (2b+c−3a)/8(b−a) and τ = (3c−2b−a)/8(c−b).
ν(x)− = (11) Especially, if ξ is a symmetric triangular fuzzy variable with
0, otherwise. b − a = c − b, then its semi-entropy is Sh [ξ] = (b − a)/2.

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 4

Let ξ = (a, b, c, d) be a trapezoidal fuzzy variable with Suppose that the ξ has a joint credibility function ν.
expected value e = (a + b + c + d)/4. Then it has been proved Randomly generate vectors yi with credibilities νi for all
that the semi-entropy Sh [ξ] = i = 1, 2, · · · , N . According to Li (2013) [17], the procedures

(b − a)ρ − ζ(ρ), if e ≤ b to estimate the expected value are summarized in the form of
Algorithm 1.




 (b−a)+ (a+c+d−3b) ln 2

+ , if b < e ≤ c (16) Algorithm. 1: Fuzzy simulation for expected value

 2 4 Step 1 Set e = 0.
 b−a + (c−b) ln 2 + ζ(τ ), Randomly generate vectors y1 , y2 , · · · , yN and compute the

 Step 2
otherwise

credibilities ν1 , ν2 , · · · , νN with
2 νk = ν(yk ).
where ρ = (b + c + d − 3a)/8(b − a) and τ = (3d−a−b− Step 3 Set two numbers
c)/8(d−c). a = f (x, y1 ) ∧ f (x, y2 ) ∧ · · · ∧ f (x, yN ),
b = f (x, y1 ) ∨ f (x, y2 ) ∨ · · · ∨ f (x, yN ).
Step 4 Randomly generate a real number r from [a, b].
IV. F UZZY M EAN -S EMI - ENTROPY M ODELS Step 5 If r ≥ 0, set e → e + Cr{f (x, ξ) ≥ r}.
Suppose that we consider m securities. Let xi denote the Otherwise, set e → e − Cr{f (x, ξ) ≤ r}.
Step 6 Repeat N times from the fourth to sixth steps.
investment proportion in security i, and let ξi stand for the Step 7 Return E = a ∨ 0 + b ∧ 0 + e · (b − a)/N .
fuzzy return of the ith security for i = 1, 2, · · · , m. We
use expected value to measure the investment return and According to Definition 3.1, the semi-entropy can be sim-
semi-entropy to measure the investment risk. Conservative ulated through the numerical integration of function S(ν(x)).
investors generally set lower risk tolerances, and then pursue Then the procedures to approximate the semi-entropy can be
the maximum return among portfolios satisfying this risk summarized as Algorithm 2.
constraint,
 Algorithm. 2: Fuzzy simulation for semi-entropy value

 max E[x1 ξ1 + x2 ξ2 + · · · + xm ξm ] Step 1 Set h = 0, k = 0 and M = 0.
s.t. Sh [x1 ξ1 + x2 ξ2 + · · · + xm ξm ] ≤ γ

Step 2 Randomly generate vectors y1 , y2 , · · · , yN and compute the
(17)
 x1 + x2 + · · · + xm = 1 credibilities ν1 , ν2 , · · · , νN with
νk = ν(yk ).

xi ≥ 0, i = 1, 2, · · · , m,

Step 3 Set two numbers
where γ denotes the maximum risk level that the investor a = f (x, y1 ) ∧ f (x, y2 ) ∧ · · · ∧ f (x, yN ),
b = f (x, y1 ) ∨ f (x, y2 ) ∨ · · · ∨ f (x, yN ).
can tolerate. Conversely, aggressive investors generally set Step 4 Estimate the expected value e = f (x, ξ) by Algorithm 1.
higher return levels, and then pursue the minimum risk among Step 5 If f (x, yk ) ≤ e, set M = M + 1.
portfolios satisfying this return constraint, Step 6 Compute sk = S(νk ). If f (x, yk ) ≥ e,
 sk = −νk ln νk − (1 − νk ) ln(1 − νk ),
 min Sh [x1 ξ1 + x2 ξ2 + · · · + xm ξm ] otherwise, sk = 0.
Step 7 Set h → h + sk .

s.t. E[x1 ξ1 + x2 ξ2 + · · · + xm ξm ] ≥ α

(18) Step 8 If k < N , set k = k + 1 and go to step 5.

 x1 + x2 + · · · + xm = 1 Step 9 Return h(b − a)/M as the semi-entropy.
xi ≥ 0, i = 1, 2, · · · , m,


where α denotes the predetermined minimum return level. Genetic algorithm (GA) introduced by Holland [10] is a
population-based optimization which resemblances the natural
V. F UZZY S IMULATION -BASED G ENETIC A LGORITHM evolution process known as a survival of the fittest and has
In this section, we elaborate on a fuzzy simulation-based been applied to efficiently solve the complex non-convex
genetic algorithm (GA) to deal with complex non-convex optimization problems existing in many practical areas [6].
mean-semi-entropy portfolio selection models. The specificity of fuzzy simulation-based GA mainly lies
Generally speaking, it is impossible to calculate the ex- in initialization and feasibility checking procedures. Taking
pected value and semi-entropy analytically except when the se- model (17) as an example, we briefly summarize the fuzzy
curity returns obey some special distributions such as equipos- simulation-based GA as follows:
sible membership functions, triangular membership functions, Step 1 A solution x = (x1 , x2 , · · · , xm ) is represented by the
trapezoidal membership functions and others. As an appli- chromosome C = (c1 , c2 , · · · , cm ) via
cation of Monte-Carlo method, fuzzy simulation technique ci
was proposed by Liu and Liu [19], which will be used to xi = , ∀i = 1, 2, · · · , m
c1 + c2 + · · · + cm
approximate the expected value and semi-entropy numerically.
The detailed introduction to fuzzy simulation techniques can where the values of genes c1 , c2 , · · · , cm are restricted to the
be found in [17], [20]. In this paper, we employ fuzzy interval [0, 1], which ensures that x1 + x2 + · · · + xm = 1 and
simulation to approximate the expected value and fuzzy semi- xi ≥ 0 with i = 1, 2, · · · , m always hold.
entropy of the portfolio return, which can be expressed as Randomly generate a chromosome C = (c1 , c2 , · · · , cm )
from the hypercube [0, 1]m . By using fuzzy simulation, we
U : x → E[f (x, ξ)], U : x → Sh [f (x, ξ)] determine the semi-entropy Sh [x1 ξ1 + x2 ξ2 + · · · + xm ξm ]
where x = (x1 , x2 , · · · , xm ) is a solutionPvector, ξ = and check its feasibility in the following way:
m
(ξ1 , ξ2 , · · · , ξm ) is a fuzzy vector, f (x, ξ) = i=1 xi ξi is a if Sh [x1 ξ1 + x2 ξ2 + · · · + xm ξm ] > γ, return 0;
2m-dimensional real valued function. otherwise, return 1;

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 5

where 1 means a feasible solution and 0 denotes an infeasible TABLE II

one. F UZZY SECURITY RETURNS
If C has been checked to be feasible, it is taken to be an No. Fuzzy return Entropy Semi-entropy
initial chromosome. Otherwise, we repeat the above process 1 (−0.2, 2.1, 2.5) 1.35 0.82
until a feasible chromosome has been produced. Finally, repeat 2 (−0.1, 1.9, 3.0) 1.55 0.85
3 (−0.4, 3.0, 4.0) 2.20 1.29
the above process pop size times and obtain the collection of 4 (−0.1, 2.0, 2.5) 1.30 0.78
initial chromosomes C1 , C2 , · · · , Cpop size . 5 (−0.6, 3.0, 4.0) 2.30 1.35
Step 2 Calculate the objective values for all chromosomes 6 (−0.2, 2.5, 3.0) 1.60 0.97
7 (−0.2, 3.0, 3.5) 1.85 1.14
by using fuzzy simulation, and then rank the chromosomes 8 (−0.4, 2.5, 4.0) 2.20 1.21
according to the objective values. The higher the expected 9 (−0.3, 2.8, 3.2) 1.75 1.09
value is, the better the chromosome is. Give a parameter θ ∈ 10 (−0.3, 2.0, 2.5) 1.40 0.84
(0, 1) and compute the values of the rank-based evaluation
function as follows
probability Pc = 0.8 and mutation probability Pm = 0.2 as
Eval(Ci ) = θ(1 − θ)i−1 , i = 1, 2, · · · , pop size.
often encountered in the literature [15], [30].
Note that i = 1 refers to the best chromosome while i = Assume that the investor is aggressive with return level
pop size pertains to the worst one. α ∈ [1.55, 2.40], which means that 1.55 is the min-
Step 3 The selection of chromosomes is done by spinning the imum expected return and 2.40 is the maximum ex-
roulette wheel method such that the better chromosomes will pected return among these ten securities. Taking α set
have more chance to produce offsprings. The selection process as 1.55, 1.65, 1.75, 1.85, 1.90, 1.95, 2.05, 2.15, 2.25, 2.35, 2.40
includes main procedures as: respectively, running the fuzzy simulation-based GA yields the
1) Compute the reproduction probability as best portfolios shown in Table III, from which we can see that
k
the expected return increases with the risk measured by semi-
entropy getting higher.
X
P0 = 0, Pi = Eval(Ci ), k = 1, 2, · · · , pop size;
i=1
Within Markowitz’s mean-risk framework that an optimal
portfolio can be obtained by maximizing the mean and mini-
2) Randomly generate a real number p from (0, pop size]; mizing the risk, entropy and variance have been recognized as
3) Select the chromosome Ci under the condition that Pi−1 < two practically useful risk measures for portfolio optimization
p ≤ Pk , k = 1, 2, · · · , pop size; [7], [15], [8], [40]. We illustrate the effectiveness of using
4) Obtain pop size chromosomes by repeating steps 2 and 3 semi-entropy as a risk measure for portfolio selection by
pop size times. comparing it with entropy and variance at different expected
Step 4 Update the chromosomes by crossover and mutation return level α (See Fig. 3). We can see that semi-entropy has
operations [15], [30], and use fuzzy simulation to calculate the positive correlation relationship with entropy and variance, that
semi-entropy and expected value for checking the feasibility is, semi-entropy can serve as an alternative risk measure of
and ranking the chromosomes. portfolio optimization.
Step 5 Repeat the second to fourth steps Generation times. In particular, taking α = 1.90 in the proposed mean-semi-
Step 6 Return the best found chromosome. entropy model, the best portfolio distributively invests securi-
ties 2, 4, 7, 9 with proportions around 15%, and securities 1,
VI. I LLUSTRATIVE E XAMPLES 3, 5, 6, 8, 10 with proportions less than 10%. In Huang [7],
In this section, we present two numerical examples to with the same return level 1.90 for mean-entropy model, the
illustrate the effectiveness of the proposed mean-semi-entropy optimal portfolio centrally invests to securities 2, 4 and 7 with
models. Assume that there are ten securities whose returns proportions 17.17%, 53.57% and 29.26%, respectively. By the
are modeled by triangular fuzzy variables ξi = (ai , bi , ci ) proposed fuzzy simulation-based GA, taking the minimum
with i = 1, 2, · · · , 10. In order to make comparisons with the expected return level α = 1.90 for mean-variance model yields
mean-entropy model (Huang [7]), we take the values ai , bi the optimal portfolio which allocates wealth into securities 4
and ci as in the Huang’s work, refer to Table II. The values and 7 with proportions 42.29% and 57.71%, respectively.
of entropy and semi-entropy associated with each triangular The results of comparative analysis imply that the mean-
fuzzy variable are also reported in the same table. semi-entropy model leads to more distributed investments
Example 6.1: We simulate the expected return curves as than the mean-entropy model and mean-variance-model, which
functions of the number of sample points (N ) in fuzzy coincides to the expression commonly used in investments:
simulation as well as Generation, pop size, probability of “never put all your eggs in one basket”. In addition, since
crossover (Pc ) and probability of mutation (Pm ) in GA, the fuzzy returns of securities given in Table II are all
which are shown by Fig. 2. Note that the expected return asymmetric, using semi-entropy as downside risk measure has
converges as N ≥ 2800, Generation ≥ 24, pop size ≥ 28, the advantage to characterize the investor’s real perception of
0.7 ≤ Pc ≤ 0.95 and 0.05 ≤ Pm ≤ 0.35. Therefore, we investment risk and the requirement of avoiding sacrificing
in the ensuing experiments have decided to use the following potential expected return.
values of the parameters: N = 2800, Generation = 24 and Example 6.2: Assume that the investor is conservative
pop size = 28 in what follows. In addition, we set crossover with risk level γ ∈ [0.78, 1.35], where 0.78 is the mini-

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 6

2.5 2.5

2.3 2.3
2.1
2.1
1.9
1.9 1.7
1.7 1.5

Expected return
Expected return

1.5 1.3
1.1
1.3
0.9
1.1 0.7
0.9 N×102 0.5
Generation Pc
0.7 Pop-size 0.3 Pm

0.5 4 0
8 12 16 20 24 28 32 36 40 44 48 0.01 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.95

Fig. 2. The expected return curves as functions of parameters N, Generation, pop size, Pc and Pm

TABLE III
O PTIMAL WEALTH ALLOCATION (%) FOR DIFFERENT RETURN LEVEL α

Return level 1.55 1.65 1.75 1.85 1.90 1.95 2.05 2.15 2.25 2.35 2.40
Semi-entropy 0.41 0.48 0.54 0.66 0.69 0.72 0.79 0.95 1.10 1.17 1.29
Security 1 10.39 11.12 7.45 9.29 8.41 7.08 10.75 7.85 6.35 0.00 0.00
Security 2 11.85 1.10 10.73 5.52 17.20 2.89 0.00 2.95 0.00 0.00 0.00
Security 3 4.58 7.25 14.07 18.55 4.27 16.98 27.33 41.49 55.81 82.66 1.00
Security 4 14.63 14.92 12.76 11.41 16.53 5.97 0.00 0.00 0.00 0.00 0.00
Security 5 12.24 8.22 7.06 9.29 1.79 8.97 9.80 6.57 0.00 0.00 0.00
Security 6 10.68 11.80 10.38 6.05 9.30 9.04 13.13 13.72 8.25 0.00 0.00
Security 7 3.23 7.70 11.01 19.11 15.03 14.51 13.75 17.36 21.87 17.34 0.00
Security 8 6.48 15.05 7.90 12.97 4.85 13.31 11.23 0.00 0.00 0.00 0.00
Security 9 3.93 13.58 8.88 7.10 14.71 14.76 14.00 11.56 14.07 0.00 0.00
Security 10 21.98 9.27 9.75 0.72 7.92 6.49 0.00 0.00 0.00 0.00 0.00

TABLE IV
O PTIMAL WEALTH ALLOCATION (%) FOR DIFFERENT RISK LEVEL γ

Risk level 0.78 0.85 0.92 0.99 1.06 1.13 1.20 1.27 1.35
Expected return 1.41 1.58 1.73 1.82 1.93 2.07 2.15 2.26 2.35
Security 1 4.61 3.12 9.60 7.89 5.46 8.15 0.00 0.00 0.00
Security 2 5.95 5.16 13.58 9.40 5.40 3.39 8.19 0.00 0.00
Security 3 13.66 16.37 12.40 14.10 21.84 17.70 20.11 19.96 0.00
Security 4 7.86 2.57 9.35 9.37 7.33 3.38 6.14 0.00 0.00
Security 5 17.57 28.56 11.70 14.77 18.28 23.94 38.86 57.73 1.00
Security 6 8.95 8.32 6.81 4.77 4.05 8.09 0.00 0.00 0.00
Security 7 11.31 10.5 9.69 5.10 5.03 0.00 8.12 0.00 0.00
Security 8 14.41 10.64 11.74 16.08 17.13 18.82 13.13 22.31 0.00
Security 9 12.26 9.21 5.39 11.08 8.80 15.98 0.00 0.00 0.00
Security 10 3.44 5.56 9.74 6.43 6.68 0.00 5.45 0.00 0.00

mum semi-entropy value and 1.35 is the maximum semi- closing price in the next month, pi is the closing price at
entropy value of these ten securities. Consider γ to be present, and di is the estimated dividends during the month.
0.78, 0.85, 0.92, 0.99, 1.06, 1.13, 1.20, 1.27 and 1.35. By per- By using the simple estimation approach proposed by Zhang
forming the fuzzy simulation-based GA, the best portfolios are et al. [39], the return rates of the 12 securities are regarded
shown in Table IV. Since security 5 has the highest expected to be asymmetric triangular and trapezoidal fuzzy variables,
return, the investment proportion on it gradually increases, and i.e., ξi = (ai , bi , ci ), i = 1, 2, · · · , 8 and ξi = (ai , bi , ci , di ),
reaches 100% at the highest risk level γ = 1.35. i = 9, · · · , 12, respectively, given in Table V.

Example 6.3: To further illustrate the optimization idea of Assume that the investor is conservative with the maximum
using semi-entropy for portfolio selection problem, we assume bearable risk level γ = 0.27. After a run of the aforementioned
that the investor has 12 securities for portfolio investment in fuzzy simulation-based GA solution method with parameters
different industries from the China Shanghai Stock Exchange. set as N = 2800, Generation = 24, pop size = 28, Pc =
The investment duration is one month. The return for the ith 0.8 and Pm = 0.2, we solve the model (17) and obtain the
security, denoted by ξi , is defined as ξi = (p∗i + di − pi )/pi , following optimal portfolio shown in Table VI.
i = 1, 2, · · · , 12, respectively, where p∗i is the estimated The investment wealth is mainly allocated into five stocks

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 7

TABLE V
F UZZY RETURNS OF THE 12 RISKY SECURITIES FROM C HINA S HANGHAI S TOCK E XCHANGE

No. Security code Fuzzy return Expected value Semi-entropy

1 601985 (−0.4103, 0.1535, 0.4398) 0.0841 0.2340
2 600221 (−0.1987, 0.2907, 0.7015) 0.2711 0.2311
3 601857 (−0.5321, 0.3752, 0.4188) 0.1593 0.3060
4 600036 (−0.1812, 0.1993, 0.4392) 0.1641 0.1659
5 600795 (−0.2109, 0.2092, 0.6923) 0.2250 0.2410
6 601766 (−0.1325, 0.0992, 0.4129) 0.1197 0.1500
7 601899 (−0.2753, 0.2709, 0.7921) 0.2647 0.2688
8 601106 (−0.4312, 0.2517, 0.4412) 0.1283 0.2566
9 601988 (−0.2512, 0.0919, 0.3821, 0.6312) 0.2135 0.2082
10 600016 (−0.3319, −0.1309, 0.1401, 0.8107) 0.1220 0.1736
11 600093 (−0.1538, 0.1032, 0.3019, 0.5013) 0.1882 0.1541
12 601362 (−0.2188, 0.0799, 0.1981, 0.4632) 0.1306 0.1646

2.4 Assume that the investor is aggressive with the prede-

Entropy
Semi-entropy
termined minimum return level α = 0.2400. After a run
2.1 Variance of the aforementioned fuzzy simulation-based GA solution
method with parameters set: N = 3000, Generation = 25,
1.8
pop size = 28, Pc = 0.8 and Pm = 0.2, we solve the
1.5 model (18) and obtain the optimal portfolio, which mainly
allocates into eleven stocks of all the 300 stocks, while invests
1.2
almost zero (less than 0.001) to the remaining 289 stocks. The
0.9
wealth proportions for the eleven stocks are shown in Fig. 4.
The expected return of the optimal portfolio is 0.2517 and the
0.6 corresponding semi-entropy is 0.2473.

0.3 0.25

0
1.55 1.65 1.75 1.85 1.9 1.95 2.05 2.15 2.25 2.35 2.4
Return level() 0.2

Fig. 3. Relationship between entropy, semi-entropy, variance and return level

α
0.15
Proportion(%)

TABLE VI
O PTIMAL ALLOCATION AMONG THESE 12 SECURITIES (%) WITH γ = 0.27
0.1

No. 1 2 3 4 5 6
Allocation(%) 0.00 43.65 0.00 5.72 13.53 17.02
No. 7 8 9 10 11 12 0.05
Allocation(%) 0.00 0.00 11.05 0.00 9.03 0.00

0 24
600221, 600795, 601766, 601988 and 600093 with total Code
69

93

52

52

39

77

52

80

14

65

05
02

04

03

04

00

01

00

00

02

01

60
60

60

60

60

60

60

60

60

60

60

proportion of 94.28%. The expected return of the optimal

portfolio is 0.2392 and the corresponding semi-entropy is Fig. 4. Wealth allocation of the 11 invested stocks from 300 stocks (China
0.2536. This example carried out illustrates that semi-entropy Shanghai Stock Exchange)
as downside risk measure matches reality more exactly when
the security returns are demonstrated by asymmetric fuzzy
variables by using real financial market data. In light of results VII. C ONCLUSIONS
in Table VI, the obtained optimal allocation strategy shows the In this paper, we defined the concept of fuzzy semi-
robustness of the proposed portfolio optimization approach. entropy to measure the downside uncertainty, and proposed
Example 6.4: To illustrate the effectiveness of the proposed two mean-semi-entropy portfolio selection models by taking
model when there are large number of risky securities, we semi-entropy as risk measure. Compared with the fuzzy mean-
give a case study in which 300 stocks from the China entropy model [7], the advantage of the proposed mean-
Shanghai Stock Exchange are randomly chosen for further semi-entropy models is that it can lead to more distributive
portfolio optimization. Similarly, by analyzing the historical investment, which is shown by numerical experiments. Ad-
data and other necessary information of the 300 securities, the ditionally, two examples using stock dataset from the real-
return rates of them are regarded as triangular fuzzy variables world financial market (China Shanghai Stock Exchange)
ξi = (ai , bi , ci ) with ci − bi = bi − ai , i = 1, 2, · · · , 300. Part show the effectiveness of the models for real portfolio selection
of the data of ξi are given in Table VII. problem. The methodology presented here is quite general

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 8

TABLE VII
T RIANGULAR FUZZY RETURNS OF 300 STOCKS FROM C HINA S HANGHAI S TOCK E XCHANGE

No. Security code Fuzzy return Expected value Semi-entropy

1 600402 (−0.2917, 0.2505, 0.7206) 0.2325 0.2586
2 600631 (−0.4138, 0.1364, 0.5073) 0.0916 0.2441
3 600438 (−0.2303, 0.2193, 0.3615) 0.1425 0.1719
4 600120 (−0.2085, 0.1737, 0.4943) 0.1583 0.1804
5 600196 (−0.2905, 0.2348, 0.5887) 0.1920 0.2330
··· ··· ··· ··· ···
299 600190 (−0.4257, 0.1930, 0.4132) 0.0934 0.2407
300 600045 (−0.1951, 0.1993, 0.5734) 0.1942 0.1937

and can be extended to the portfolio selection problems in where ρ = (2b+c−3a)/8(b−a) and τ = (3c−2b−a)/8(c−b).
stochastic and other types of uncertain environments. Especially, if ξ is a symmetric triangular fuzzy variable with
In future studies, we plan to conduct empirical studies to b − a = c − b, then its semi-entropy is Sh [ξ] = (b − a)/2.
test the effectiveness of mean-semi-entropy models in real-life Proof: For triangular fuzzy variable ξ = (a, b, c), we have
financial market. In addition, we will extend the mean-semi- 
x−a
entropy model to consider transaction cost and skewness for

 , if x ∈ [a, e] and a < e ≤ b < c


 2(b − a)
explaining higher moments with non-normal distributions, as 
x−a


, if x ∈ [a, b] and a < b ≤ e < c

in Liu et al. [22], Zhang et al. [40] and Li et al. [15]. We could 

also construct robust portfolios (Nguyen and Lo [27]) based
−
ν(x) = 2(b − a)
(c − x)

on preferences defined over the available securities, instead 
, if x ∈ [b, e] and a < b ≤ e < c


2(c − b)

of their exact expected returns. Finally, it will be valuable to 



apply the mean-semi-entropy models and their extensions to 

0, otherwise.
study the multi-period portfolio selection problems, portfolio
adjusting problems, optimal project selection and adjusting The computing of semi-entropy breaks down into two cases.
problems, capital budgeting problems, subcontractor selection If e ≤ b, the semi-entropy is
Z e 
problems and so on. (6b − 5a − c)2 6b − 5a − c

x−a
S dx = ln +
a 2(b − a) 64(b − a) 8(b − a)
ACKNOWLEDGMENT
2b + c − 3a (2b + c − 3a)2 2b + c − 3a
This work was supported by the National Natural Sci- − ln .
8 64(b − a) 8(b − a)
ence Foundation of China (Nos. 71371027, 91224001), Bei-
jing Nova Program (No. Z14111000180000), and the State Otherwise, if e > b, the semi-entropy is
Z b   Z e  
Key Laboratory of Rail Traffic Control and Safety (No. x−a c−x b−a
RCS2014K011). S dx + S dx = +
a 2(b−a) b 2(c−b) 2
(3c−2b−a)2 3c−2b−a (5c−6b+a)2 5c−6b+a
A PPENDIX ln − ln .
64(c−b) 8(c−b) 64(c−b) 8(c−b)
This section presents the calculation processes for the semi-
entropy values of equipossible fuzzy variable, triangular fuzzy Denote ρ = (2b+c−3a)/8(b−a) and τ = (3c−2b−a)/8(c−b)
variable and trapezoidal fuzzy variable. for simplification. Define ζ(x) = x2 ln x − (1 − x)2 ln(1 − x).
Theorem 7.1: Let ξ = (a, b) be an equipossible fuzzy Then the semi-entropy becomes

variable with expected
value e = (a + b)/2. Then its semi-  (b − a)ρ − (b − a)ζ(ρ), if e ≤ b

entropy is (b − a) ln 2 /2. Sh [ξ] =


Proof: For any x ∈ [a, b], we have S ν(x) = −0.5 ln 0.5 − b−a

 + (c − b)ζ(τ ), otherwise
(1−0.5) ln(1−0.5) = ln 2. Then it follows from the definition 2
of semi-entropy that Particularly, if ξ is a symmetric triangular fuzzy variable with
Z e b − a = c − b, which means e = b, then it can be easily
calculated that the semi-entropy is Sh [ξ] = (b − a)/2. The


Sh [ξ] = ln 2dx = (b − a) ln 2 /2.
a proof is complete.
The proof is complete. Theorem 7.3: Let ξ = (a, b, c, d) be a trapezoidal fuzzy
Theorem 7.2: Let ξ = (a, b, c) be a triangular fuzzy variable with expected value e = (a + b + c + d)/4. Then its
variable with expected value e = (a + 2b + c)/4. Define semi-entropy is
ζ(x) = x2 ln x − (1 − x)2 ln(1 − x). Then its semi-entropy 
 (b − a)ρ − ζ(ρ), if e ≤ b
is 


  2(b−a) + (a+c+d−3b) ln 2

 (b − a)ρ − (b − a)ζ(ρ), if e ≤ b
 Sh [ξ] =
4
, if b < e ≤ c
Sh [ξ] =


b−a  b−a + (c−b) ln 2 + ζ(τ ),


+ (c − b)ζ(τ ), otherwise otherwise.

 
2 2

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 9

where ρ = (b + c + d − 3a)/8(b − a) and τ = (3d−a−b− [3] M.J. Best and R.R. Grauer, “The efficient set mathematics when mean-
c)/8(d−c). variance problems are subject to general linear constraints,” Journal of
Economics and Business, vol. 42, no. 2, pp. 105-120, 1990.
Proof: For trapezoidal fuzzy variable ξ = (a, b, c, d), we have [4] F. Choobineh and D. Branting, “A simple approximation for semivari-
 x−a ance,” European Journal of Operational Research, vol. 27, no. 3, pp.
 , if x ∈ [a, e] and a < e ≤ b < c < d 364-370, 1986.
2(b − a)

[5] P. Chunhachinda, K. Dandapani, S. Hamid and A.J. Prakash, “Portfolio



selection and skewness: Evidence from international stock markets,”

 x−a


 , if x ∈ [a, b] and a < b ≤ e ≤ c < d Journal of Banking and Finance, vol. 21, no. 2, pp. 143-167, 1997.
2(b − a)




 [6] M. Gen and R. Cheng. Genetic algorithms and engineering optimization.
 New York, Wiley, 2000.
0.5, if x ∈ [b, e] and a < b < e ≤ c < d




 [7] X. Huang, “Mean-entropy models for fuzzy portfolio selection,” IEEE

 Transactions on Fuzzy Systems, vol. 16, no. 4, pp. 1096-1101, 2008.
ν(x)− = x−a [8] X. Huang, “Mean-variance models for portfolio selection subject to
, if x ∈ [a, b] and a < b < c ≤ e < d experts’ estimations,” Expert Systems with Applications, vol. 39, no. 5,


 2(b − a) pp. 5887-5893, 2012.




 0.5, if x ∈ [b, c] and a < b < c ≤ e < d [9] D. Huang, S. Zhu, F.J. Fabozzi and M. Fukushima, “Portfolio selection


 under distributional uncertainty: A relative robust CVaR approach,” Eu-
 ropean Journal of Operational Research, vol. 203, no. 1, pp. 185-194,
d−x


 2010.

 , if x ∈ [c, e] and a < b < c ≤ e < d


 2(d − c) [10] J. Holland, Adaptation in Natural and Artificial Systems. Ann Arbor,

 MI: University of Michigan Press,1975.
0,

otherwise. [11] P.D. Kaplan and R.H. Alldredge, “Semivariance in risk-based index
construction: Quantidex global indexes,” Journal of Investing, vol. 6, no.
We consider here three cases. If e ≤ b, its semi-entropy is 2, pp. 82-87, 1997.
Z e  [12] V. Lacagnina and A. Pecorella, “A stochastic soft constraints fuzzy
(7b−5a−c−d)2 7b−5a−c−d

x−a model for a portfolio selection problem,” Fuzzy Sets and Systems, vol.
S dx = ln + 157, no. 10, pp. 1317-1327, 2006.
a 2(b − a) 64(b − a) 8(b − a)
[13] P. Li and B. Liu, “Entropy of credibility distributions for fuzzy vari-
ables,” IEEE Transactions on Fuzzy Systems, vol. 16, no. 1, pp. 123-129,
b+c+d−3a (b+c+d−3a)2 b+c+d−3a
− ln . 2008.
8 64(b − a) 8(b − a) [14] X. Li and B. Liu, “A sufficient and necessary condition for credibility
measures,” International Journal of Uncertainty, Fuzziness & Knowledge-
If b < e ≤ c, its semi-entropy is Based Systems, vol. 14, no. 5, pp. 527-535, 2006.
Z b   Z e [15] X. Li, Z. Qin and S. Kar, “Mean-variance-skewness model for portfolio
x−a b−a selection with fuzzy returns,” European Journal of Operational Research,
S dx + S(0.5)dx = + vol. 202, no. 1, pp. 239-247, 2010.
a 2(b − a) b 2
[16] X. Li, B. Shou and Z. Qin, “An expected regret minimization portfolio
(a + c + d − 3b) ln 2 selection model,” European Journal of Operational Research, vol. 218,
. no. 2, pp. 484-492, 2012.
4 [17] X. Li, Credibilistic Programming. Springer-Verlag, Heidelberg, 2013.
Otherwise, if e > c, its semi-entropy is [18] X. Li, S.N. Guo and L.A. Yu, “Skewness of fuzzy numbers and its
applications in portfolio selection,” IEEE Transactions on Fuzzy Systems,
Z b   Z c Z e   vol. 23, no. 6, pp. 2135-2143, 2015.
x−a d−x
S dx+ S(0.5)dx+ S dx [19] B. Liu and Y.-K. Liu, “Expected value of fuzzy variable and fuzzy
a 2(b − a) b c 2(d − c) expected value models,” IEEE Transactions on Fuzzy Systems, vol. 10,
no. 4, pp. 445-450, 2002.
b − a (3d − a − b − c)2 3d − a − b − c [20] B. Liu, Theory and Practice of Uncertain Programming. Heidelberg,
= + ln + Germany: Physica-Verlag, 2008.
2 64(d − c) 8(d − c)
[21] B. Liu, Uncertainty Theory: An Introduction to its Axiomatic Founda-
tions. Berlin, Germany: Springer-Verlag, 2004.
(a + b + 5d − 7c)2 a + b + 5d − 7c [22] S. Liu, S.Y. Wang and W. Qiu, “Mean-variance-skewness model for port-
(c − b) ln 2 − ln .
64(d − c) 8(d − c) folio selection with transaction costs,” International Journal of Systems
Science, vol. 34, no. 4, 255-262, 2003.
Denote ρ = (b+c+d−3a)/8(b−a) and τ = (3d−a−b−c)/8(d− [23] J.C. Mao, “Models of capital budgeting, EV vs ES,” Journal of financial
c) for simplification. Define ζ(x) = x2 ln x−(1−x)2 ln(1−x). and quantitative analysis, vol. 4, no. 5, pp. 657-675, 1970.
[24] H. Markowitz, “Portfolio selection,” Journal of Finance, vol. 7, no. 1,
Then the semi-entropy is pp. 77-91, 1952.
 [25] H. Markowitz, Portfolio Selection: Efficient diversification of invest-

 (b − a)ρ − ζ(ρ), if e ≤ b ments. New York: Wiley, 1959.


 b−a (a+c+d−3b) ln 2
 [26] H. Markowitz, P. Todd, G. Xu and Y. Yamane, “Computation of mean-
+ , if b < e ≤ c semivariance efficient sets by the critical line algorithm,” Annals of
Sh [ξ] =
 2 4 Operations Research, vol. 45, no. 1, pp. 307-317, 1993.
[27] T.D. Nguyen and A. Lo, “Robust ranking and portfolio optimization,”

 b−a + (c−b) ln 2 + ζ(τ ),


otherwise. European Journal of Operations Research, vol. 221, no. 2, 407-416, 2012.

2 [28] A.F. Perold, “Large-scale portfolio optimization,” Management Science,
vol. 30, no. 10, pp. 1143-1160, 1984.
The proof is complete.
[29] G.C. Philippatos and C.J. Wilson, “Entropy, market risk, and the
selection of efficient portfolios,” Applied Economics, vol. 4, no. 3, pp.
R EFERENCES 209-220, 1972.
[30] Z. Qin, X. Li and X. Ji, “Portfolio selection based on fuzzy cross-
[1] F.D. Arditti, “Risk and the required return on equity,” Journal of Finance, entropy,” Journal of Computational and Applied mathematics, vol. 228,
vol. 22, no. 1, pp. 19-36, 1967. no. 1, pp. 139-149, 2009.
[2] F.D. Arditti and H. Levy, “Portfolio efficiency analysis in three moments: [31] P. Samuelson, “The fundamental approximation theorem of portfolio
The multiperiod case,” Journal of Finance, vol. 30, no. 3, pp. 797-809, analysis in terms of means, variances, and higher moments,” Review of
1975. Economic Studies, vol. 37, pp. 537-542, 1970.

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2016.2543753, IEEE
Transactions on Fuzzy Systems

IEEE TRANSACTIONS ON FUZZY SYSTEMS 10

[32] W.F. Sharpe, “Capital asset prices: A theory of market equilibrium under Witold Pedrycz received the M.Sc., Ph.D. and
conditions of risk,” Journal of Finance, vol. 19, no. 3, pp. 425-442, 1964. D.Sci. degrees from the Silesian University of Tech-
[33] M.R. Simonelli, “Indeterminacy in portfolio selection,” European Jour- nology, Gliwice, Poland. He is currently a Professor
nal of Operational Research, vol. 163, no. 1, pp. 170-176, 2005. and Canada Research Chair (CRC computational
[34] E. Vercher and J.D. Bermudez, “A possibilistic mean-downside risk- intelligence) with the Department of Electrical and
skewness model for efficient portfolio selection,” IEEE Transactions on Computer Engineering, University of Alberta, Ed-
Fuzzy Systems, vol. 21, no. 3, pp. 585-595, 2013. monton, AB, Canada. In 2009, he was elected as a
[35] S. Wang and S. Zhu, “On fuzzy portfolio selection problems,” Fuzzy foreign member of the Polish Academy of Sciences,
Optimization and Decision Making, vol. 1, no. 4, pp. 361-377, 2002. Warsaw, Poland. He is the author of 14 research
[36] B. Wang, S. Wang and J. Watada, “Fuzzy-portfolio-selection models monographs covering various aspects of computa-
with value-at-risk,” IEEE Transactions on Fuzzy Systems, vol. 19, no. 4, tional intelligence and software engineering. He is
pp. 758-769, 2011. also with the Department of Electrical and Computer Engineering Faculty of
[37] L.A. Zadeh, “Fuzzy sets as a basis for a theory of possibility,” Fuzzy Engineering, King Abdulaziz University, Jeddah, Kingdom of Saudi Arabia.
Sets and Systems, vol. 1, no. 1, pp. 3-28, 1978. His main research interests include computational intelligence, fuzzy modeling
[38] L.A. Zadeh, “A theory of approximate reasoning,” Machine Intelligence, and granular computing, knowledge discovery and data mining, fuzzy control,
vol. 9, no. 1, pp. 149-194, 1979. pattern recognition, knowledge-based neural networks, relational computing,
[39] W.G. Zhang, X.L. Zhang and W.J. Xu, “A risk tolerance model for and software engineering. He has published numerous papers in this area.
portfolio adjusting problem with transaction costs based on possibilistic Dr. Pedrycz was elected as a Fellowof the Royal Society of Canada in 2012.
moments”, Insurance: Mathematics and Economics, vol. 46, no. 3, pp. He has been amember of numerous program committees of IEEE conferences
493-499, 2010. in the area of fuzzy sets and neurocomputing. He is intensively involved in
[40] W.G. Zhang, Y.J. Liu and W.J. Xu, “A possibilistic mean-semivariance- editorial activities. He is the Editor-in-Chief of Information Sciences, and
entropy model for multi-period portfolio selection with transaction costs,” serves as an Associate Editor of the IEEE TRANSACTIONS ON FUZZY
European Journal of Operational Research, vol. 222, no. 2, pp. 341-349, SYSTEMS. In 2007, he received a prestigious Norbert Wiener Award from
2012. the IEEE Systems, Man, and Cybernetics Council. He received the IEEE
Canada Computer Engineering Medal in 2008. In 2009, he received a Cajastur
Prize for soft computing from the European Centre for Soft Computing for
“pioneering and multifaceted contributions to granular computing.”

Jiandong Zhou received the B.S. degree from

School of Management and Economics, Beijing
Institute of Technology (BIT), Beijing, China, in
2014. He is now working towards the M.S. degree
under the supervision of Prof. Xiang Li at School of
Economics and Management, Beijing University of
Chemical Technology.
His current research interests include the areas of
computational social science, text mining, operations
research, and smart transportation.

Xiang Li received the B.S. degree from Jilin

University, Changchun, China, in 2004 and the
Ph.D. degree in operations research and cybernet-
ics from Tsinghua University, Beijing, China, in
2008.
He is currently a Professor with Beijing Uni-
versity of Chemical Technology. He is the author
or coauthor of more than 60 articles on interna-
tional journals, including Transportation Research
Part B; Transportation Research Part C, Informa-
tion Sciences; European Journal of Operational Re-
search; IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION
SYSTEMS; IEEE TRANSACTIONS ON FUZZY SYSTEMS, and IEEE
TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS-PART A,
which have been cited more than 500 times on Web of Science and 1200
times on Google Scholar. His research interests include uncertainty theory,
fuzzy programming, and its applications in sustainable railway transport. He
serves as an Associate Editor of Information Sciences and an Editorial Board
Member of International Journal of General Systems.

1063-6706 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.