A multi-period fuzzy portfolio optimization model with minimum transaction lots
A multi-period fuzzy portfolio optimization model with minimum transaction lots
Accepted Manuscript
PII: S0377-2217(14)00900-X
DOI: 10.1016/j.ejor.2014.10.061
Reference: EOR 12614
Please cite this article as: Yong-Jun Liu, Wei-Guo Zhang, A multi-period fuzzy portfolio optimiza-
tion model with minimum transaction lots , European Journal of Operational Research (2014), doi:
10.1016/j.ejor.2014.10.061
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Highlights
T
IP
• A genetic algorithm is designed for solution.
CR
US
AN
M
ED
PT
CE
AC
1
ACCEPTED MANUSCRIPT
T
IP
CR
Abstract
US
with minimum transaction lots. Based on possibility theory, we formulate a mean-
semivariance portfolio selection model with the objectives of maximizing the ter-
AN
minal wealth and minimizing the cumulative risk over the whole investment hori-
zon. In the proposed model, we take the return, risk, transaction costs, diversifi-
cation degree, cardinality constraint and minimum transaction lots into considera-
M
tion. To reflect investor’s aspiration levels for the two objectives, a fuzzy decision
technique is employed to transform the proposed model into a single objective
ED
markets to demonstrate the idea of our model and the effectiveness of the designed
algorithm.
CE
∗
Corresponding author. Tel.: +86208711421. E-mail address:[email protected]
(Y.-J. Liu); [email protected]; [email protected]. (W.G. Zhang).
1. Introduction
T
basket of securities. To realize this idea, the proper portfolio model must be pre-
IP
sented. Most of existing portfolio selection models have been proposed on the
assumption of a perfect fractionability of the investments, which are difficult to
CR
implement. In real world, each security has its minimum transaction lot. So
it is necessary to consider rounds. To reflect this realistic characteristic of se-
US
curity, some researchers have proposed a series of mixed-integer programming
models. Speranza (1996) proposed a mixed-integer programming model on the
AN
basis of the mean absolute deviation model in Konno and Yamazaki (1991) by
taking minimum transaction lots and maximum number of securities into consid-
eration, and designed a simple two-phase heuristic algorithm to solve the proposed
M
the problem. In Kellerer et al. (2000), a portfolio selection model with fixed
costs and minimum transaction lots was proposed and two linear programming
PT
based heuristic algorithms were designed for solution. Konno and Wijayanayake
(2001) discussed portfolio optimization problem with concave transaction costs
CE
model. Konno and Wijayanayake (2002) discussed the market illiquidity effects
and investigated a portfolio optimization problem with D. C. transaction costs and
minimal transaction unit constraints. Recently, Lin and Liu (2008) proposed three
possible models for portfolio selection problems with minimum transaction lots,
and devised genetic algorithms to solve them. Baixauli-Soler et al. (2011) em-
3
ACCEPTED MANUSCRIPT
ployed VaR as risk measure to investigate an asset allocation problem under real
constraints, such as minimum transaction lots and non-linear cost structure.
Notice that all the models mentioned above are single period portfolio selec-
tion models. However, in real world, investors tend to invest long-term invest-
ment. The investors should adjust their wealth from time to time. So it is nature
T
to investigate multi-period portfolio selection problems. Numerous researchers
IP
have studied multi-period portfolio optimization problems, see for instance, El-
CR
ton and Gruber (1974), Fama (1970) and Hakansson (1971). Li and Ng (2000)
made breakthrough result for dynamic portfolio. In their model, they employed
the idea of embedding the problem in a tractable auxiliary problem to investigate
US
the mean-variance formulation in multi-period portfolio selection and obtained the
corresponding mean-variance efficient frontier. After that, Zhu et al. (2004) used
AN
the same approach to study the continuous-time dynamic multi-period problem by
taking risk control into account. Rocha and Kuhn (2012) formulated a dynamic
M
tion problem with multiple rival risk and return scenarios. Fonseca and Rustem
(2012) applied linear decision rules to a robust multiperiod international portfolio.
PT
concerned with multi-period sequential decision problems for financial asset al-
location and presented a multi-period portfolio optimization model with control
AC
policies. Briec and Kerstens (2009) proposed a general approach for multi-horizon
mean-variance portfolio analysis. Fu et al. (2010) considered continuous-time
mean-variance portfolio selection with borrowing constraint, i.e, under different
interest rates for borrowing and lending, rendering the market incomplete. Fu et
al. (2014) considered the optimal asset allocation problem in a continuous-time
4
ACCEPTED MANUSCRIPT
T
random variables with probability distributions. The basic assumption of them
IP
is that the future situation of risky assets can be correctly reflected by its histor-
CR
ical data. However, since financial markets are complex and ever-changing, this
kind of assumption is hard to ensure in real world. As is well known, portfolio
decision-making is often affected by many non-probabilistic factors including so-
US
cial, economic, political, people’s cognitive and psychological factors, etc. For
example, REITs and the financial crisis mentioned by Basse et al. (2009) are also
AN
affected by aforementioned human factors, which have caused some instability
in markets affecting the correlations between the returns of different asset classes
M
and that this observation is a reminder of the fact that the correlation matrices of
returns on risky assets regularly used in financial optimizations are not necessarily
ED
stable over time. Thus, the influence of experts’ experiences and knowledge, and
investors’ subjective opinions on portfolio selection cannot be neglected (such as
PT
Tanaka and Guo (1999), Fang et al. (2006), Vercher et al. (2007), Liu et al. (2012)
and so on). Due to the influence of above-mentioned non-probabilistic factors,
CE
the future situation of risky assets are usually characterized by fuzzy uncertainty
such as vagueness and ambiguity. Namely, decision makers are usually provided
AC
5
ACCEPTED MANUSCRIPT
tors into consideration, fuzzy approaches are more suitable than probabilistic ap-
proaches in characterizing the uncertainty in real financial markets as pointed by
Liu and Zhang (2013). With the widely use of fuzzy set theory in Zadeh (1965),
more and more researchers have realized that they could use the fuzzy set theory
to handle the vagueness and ambiguity, see for example, Tanaka and Guo (1999),
T
Carlsson et al. (2002), Fang et al. (2006), Vercher et al. (2007), Zhang et al.
IP
(2007), Zhang et al. (2009), Zhang et al. (2010), Barak et al. (2013), and Liu and
CR
Zhang (2013). Though great progress has been made in fuzzy portfolio selection,
most of existing models are single period models. Recently, Sadjadi et al. (2011)
discussed a fuzzy multi-period portfolio optimization problem with different rates
US
for borrowing and lending. Zhang et al. (2012) proposed a possibilistic mean-
semivariance-entropy model for multi-period fuzzy portfolio selection. Liu et al.
AN
(2012) presented four multi-period fuzzy portfolio optimization models by using
multiple criteria. To our knowledge, the researches about multi-period fuzzy port-
M
folio selection problem with real constrains are few. The purpose of this paper is
to investigate multi-period portfolio selection problem in fuzzy environment with
ED
the return rate of a portfolio. The investment risk is quantified by the lower pos-
sibilistic semivariance of the return rate of portfolio and the diversification degree
AC
is measured by the proportion entropy in Kapur (1990). Since the proposed model
is a bi-objective programming problem, a S-shape membership function is used to
express investor’s satisfaction degree for each objective and then transform it into
a corresponding single objective mixed-integer nonlinear programming problem.
After that, we design a novel genetic algorithm for solution.
6
ACCEPTED MANUSCRIPT
T
S-shape fuzzy membership functions to express investor’s aspiration levels for
IP
the two objectives. By using the nonlinear S-shape membership functions, we
CR
transform the proposed model into a single programming problem. In Section 4,
we design a genetic algorithm to solve the proposed model. In Section 5, we use
an empirical application in Chinese stock markets to demonstrate the idea of our
US
model and the effectiveness of the designed algorithm for solution. In Section 6,
we conclude this paper by some remarks.
AN
2. Preliminaries
M
In this section, let us first review some basic conceptions about fuzzy variables,
which we need in the following sections.
ED
Let A be a fuzzy number, i.e. such fuzzy subset A of the real line R with a
membership function µA : R −→ [0, 1], that (see Dubois and Prade (1980)):
PT
R; λ ∈ [0, 1]);
III) µA is upper semicontinuous;
AC
IV) supp(A) is bounded, where supp(A) = cl{x ∈ R|µA (x) > γ} and cl is the
γ-level set of A as [A]γ = [a(γ), a(γ)] (∀γ ∈ [0, 1]) closure operator.
Denote the family of fuzzy numbers as F. For any A ∈ F, we denote the
γ-level set of A as [A]γ = [a(γ), a(γ)] (γ ∈ [0, 1]). Fuzzy numbers can also be
considered as possibility distributions. If A ∈ F is a fuzzy number and x ∈ R
7
ACCEPTED MANUSCRIPT
a real number then µA (x) can be interpreted as the degree of possibility of the
statement ‘x is A’.
Let A ∈ F be a fuzzy number with [A]γ = [a(γ), a(γ)] (γ ∈ [0, 1]). Carls-
son and Fullér (2001) defined the possibilistic mean value of fuzzy number A as
follows
T
Z 1
IP
E(A) = γ(a(γ) + a(γ))dγ. (1)
0
CR
In Saeidifar and Pasha (2009), if we set the weighted function as f (γ) = 2γ,
then we have the lower and upper possibilistic variances of A with the following
forms
−
V ar (A) = 2
Z US1
γ(E(A) − a(γ))2 dγ, (2)
AN
0
Z 1
+
V ar (A) = 2 γ[E(A) − a(γ)]2 dγ. (3)
0
M
bers. Carlsson and Fullér (2001) defined the possibilistic mean value about the
linear combination of A and B as follows
PT
8
ACCEPTED MANUSCRIPT
T
Var− ( λi Ai ) = λ2i Var− (Ai ) + 2 λi λj Cov− (Ai , Aj ), (7)
i=1 i<j=1
IP
i=1
Xn
+
P
n P
n
Var ( λi Ai ) = λ2i Var+ (Ai ) + 2 λi λj Cov+ (Ai , Aj ). (8)
CR
i=1 i=1 i<j=1
According to the definitions above, we can obtain the following formulas for a
trapezoidal fuzzy number A = (a, b, α, β) with tolerance interval [a, b], left width
1 − a−x
US
α > 0 and right width β > 0 if its membership function takes the following form
, if a − α ≤ x ≤ a,
AN
α
1, if x ∈ [a, b],
µA (x) =
1 − x−b , if b ≤ x ≤ b + β,
β
M
0, otherwise.
γ)β] (∀γ ∈ [0, 1]). By Eq. (1), the possibilistic mean value of A can be denoted as
a+b β−α
E(A) = + . (9)
PT
2 6
Using Eqs. (2-3), the upper and lower possibilistic variances of A can be, respec-
CE
tively, represented as
b−a α+β 2 β2
V ar+ (A) = ( + ) + , (10)
2 6 18
AC
b−a α+β 2 α2
V ar− (A) = ( + ) + . (11)
2 6 18
9
ACCEPTED MANUSCRIPT
T
For the better understanding of our paper, we first introduce all the notations that
IP
will be used in the following sections. For i = 1, 2, . . . , n and t = 1, 2, . . . , T , we
CR
let
xt,i
xt,n+1
US
the number of transaction lots invested in security i at period t;
rt,i the return rate of security i at period t, where rt,i = (at,i , bt,i , αt,i , βt,i );
ct,i the transaction cost of the minimal transaction unit on security i at period t;
PT
10
ACCEPTED MANUSCRIPT
Most of existing portfolio selection models have been proposed on the as-
sumption that the investments are perfect fractionability, in such a way that the
investment proportion of portfolio for each security could be represented by real
numbers. However, in real financial market, the transaction of securities are usu-
T
ally associated with minimum lots or rounds. For different kinds of securities, the
IP
minimum lots may be different, for instance 100, 500 or 1000 shares. Let us take
CR
A-share in Chinese security markets for example. Its minimal unit is 100 shares.
The worth for each transaction lot is equal to pt,i = Ni yt,i , where yt,i is the market
US
price of security i at period t and Ni is the number of minimum units of security
i required as minimum quantity. xt,i pt,i denotes as the part of the total available
wealth that the investor decides to invest on security i at period t.
AN
In this section, we employ the crisp possibilistic mean value of the net return
on the portfolio at each period to measure the return of portfolio. The risk on
ED
the return rate of portfolio at each period is quantified by the lower possibilistic
variance and the diversification degree is measured by the proportion entropy in
PT
Kapur (1990).
For the transaction costs, we assume in the sequel that the transaction costs
CE
at period t is a V shape function of the difference between the tth period port-
folio xt = (xt,1 , xt,2 , . . . , xt,n , xt,n+1 ) and the t − 1th period portfolio xt−1 =
(xt−1,1 , xt−1,2 , . . . , xt−1,n , xt−1,n+1 ). Thus, the transaction costs of security i at
AC
Similar to Yoshimoto (1996), the absolute term on the right hand side of Eq. (12)
11
ACCEPTED MANUSCRIPT
−
(xt,i − xt−1,i )pt,i }; d+
t,i and dt,i satisfy the following relationships
T
−
(xt,i − xt−1,i )pt,i = d+
t,i − dt,i ,
IP
− −
d+ +
t,i · dt,i = 0, dt,i , dt,i ≥ 0, i = 1, 2, . . . , n, t = 1, 2, . . . , T.
CR
Notice that d− +
t,i and dt,i cannot be zero at the same time. Hence, we can find
− −
out that if the difference d+ +
t,i − dt,i is positive (negative), dt,i (dt,i ) becomes the
difference.
US
Since the transaction costs is not needed for the risk-free asset, by Eq. (13),
the total transaction costs of the portfolio at period t can be represented by
AN
X n
−
Ct = ct,i (d+
t,i + dt,i ). (14)
i=1
M
The return on the portfolio xt = (xt,1 , xt,2 , . . . , xt,n , xt,n+1 ) at period t can be
expressed by
ED
n
X
RP,t = xt,i pt,i rt,i + rf (t)xt,n+1 , t = 1, 2, . . . , T, (15)
i=1
P
n
PT
Xn n
X
Wt+1 = xt,i pt,i (1 + rt,i ) + (Wt − xt,i pt,i − Ct )(1 + rf (t)). (16)
i=1 i=1
Solving Eq. (16) recursively, the terminal wealth obtained at the end of period T
AC
can be represented by
T
Y X n
T X T
Y
WT +1 =W1 (1 + rf (t)) + xt,i pt,i (rt,i − rf (t)) (1 + rf (i))
t=1 t=1 i=1 i=t+1
T
X T
Y
− Ct (1 + rf (t)). (17)
t=1 t=1
12
ACCEPTED MANUSCRIPT
T
−
− ct,i (d+
t,i + dt,i ) (1 + rf (t)). (18)
IP
t=1 i=1 t=1
From the assumption above, the return rate of security i at period t, rt,i =
CR
(at,i , bt,i , αt,i , βt,i ), is trapezoidal fuzzy variable for all i = 1, 2, . . . , n and t =
1, 2, . . . , T . By the the sup-min extension principle in Zadeh (1965), we can con-
US
clude that the return of the portfolio at period t (i.e., RP,t ), the return rate of
portfolio at the tth period rp,t and the terminal wealth WT +1 are all trapezoidal
AN
fuzzy variables. By Eqs. (9) and (18), we have
T X
X n YT
at,i + bt,i βt,i − αt,i
E(WT +1 ) = xt,i pt,i ( + − rf (t)) (1 + rf (i))
2 6
M
Let wt = (wt,1 , wt,2 , . . . , wt,n , wt,n+1 ) be the investment proportion of the port-
folio at period t. By Eq. (9), the return rate of the portfolio at period t can be
PT
expressed by
n
X
CE
P
n+1
AC
xt,i pt,i
where wt,i = 1 and wt,i = P
n
−
(i = 1, 2, . . . , n).
i=1 Wt − ct,i (d+
t,i +dt,i )
i=1
From Eq. (11), the lower possibilistic semivariance of the return rate on the
portfolio at period t can be calculated by
n
X n
− bt,i − at,i βt,i + αt,i 2 1 X
V ar (rp,t ) = [ wt,i ( + )] + ( wt,i αt,i )2 . (21)
i=1
2 6 18 i=1
13
ACCEPTED MANUSCRIPT
Thus, the cumulative lower possibilistic semivariance over the whole holding pe-
riod can be computed by
XT Xn n
bt,i − at,i βt,i + αt,i 2 1 X
V (x) = {[ wt,i ( + )] + ( wt,i αt,i )2 }. (22)
t=1 i=1
2 6 18 i=1
T
Based on the analysis above, we quantify the return of security by the possi-
IP
bilistic mean value and the risk by the lower possibilistic variance. We assume
that the investor wants to maximize the terminal wealth and minimize the cumu-
CR
lative risk on the return rates of portfolios over T investment periods. At the same
time, he requires the return rate and the diversification degree of portfolio at each
US
period must achieve or exceed their corresponding preset minimum levels and the
risk of the return rate on portfolio must be smaller than the given maximum risk
AN
tolerance level at each period. Moreover, he restricts to hold at most K assets in
portfolio at each period. Then, the multi-period portfolio selection problem can
M
ED
PT
CE
AC
14
ACCEPTED MANUSCRIPT
T
PT Pn Pn
min V (x) = {[ w (
bt,i −at,i
+
βt,i +αt,i 2
)] + 1
( wt,i αt,i )2 }
t,i 2 6 18
IP
t=1 i=1 i=1
Pn Pn
s.t. x p + ct,i (d+ −
t,i + dt,i ) ≤ E(Wt ), (23)
t,i t,i
CR
i=1 i=1
Pn
a +b β −α
wt,i ( t,i 2 t,i + t,i 6 t,i ) + wt,n+1 rf (t) ≥ r(t), (24)
i=1
Pn P n
b −a β +α
(P1 ) [ wt,i ( t,i 2 t,i + t,i 6 t,i )]2 + 18 1
( wt,i αt,i )2 ≤ δt , (25)
−
i=1
P
n+1
i=1
wt,i ln(wt,i ) ≥ et , US i=1
(26)
AN
P
n+1
xt,i pt,i
wt,i = 1 (wt,i = , i = 1, 2, . . . , n), (27)
Wt −
Pn
ct,i (d+ −
i=1 t,i +dt,i )
i=1
P
n+1
sign(wt,i ) ≤ K, (28)
M
i=1
0 ≤ lt,i ≤ xt,i ≤ ut,i , xt,i ∈ Z; (29)
d+ −
t,i − dt,i = (xt,i − xt−1,i )pt,i ,
ED
d+ , d− ≥ 0, d+ · d− = 0, i = 1, 2, . . . , n, t = 1, 2, . . . , T,
t,i t,i t,i t,i
where Eq. (23) indicates that the sum of both the total invested amount at the
PT
, i.e., Wt ; Eq. (24) suggests that the return rate of the portfolio at period t must be
no less than the given minimum expected return rate level r(t); Eq. (25) requires
AC
that the risk of the return rate of the portfolio at period t must be smaller than
or equal to the preset maximum risk tolerance level δt ; Eq. (26) means that the
diversification degree of the portfolio at period t must achieve or exceed the given
diversification degree et ; Eq. (27) requires that the sum of weights at period t
must be unity; Eq. (28) represents that the maximum number of assets in the
15
ACCEPTED MANUSCRIPT
T
jectives. To remove the effects of their incommensurability, similar to Hwang and
IP
Yoon (1981), we need to perform normalization operation on each objective as
CR
follows
E(WT +1 ) − WM
dW = , (30)
WP IS − WN IS
dV =
US
V (x) − VM
VP IS − VN IS
,
spectively; EM and VM represent the middle aspiration levels for terminal wealth
and the cumulative risk over the whole holding periods, respectively. After nor-
ED
malization, the value of each objective function is kept in the range [0, 1]. Thus,
we can compare them because their values are independent of the scale of the
PT
16
ACCEPTED MANUSCRIPT
T
µE (x) = , (33)
1 + exp(−θW dW )
IP
1
µV (x) = , (34)
1 + exp(−θV dV )
CR
where θW and θV denote as investor’ s satisfaction degrees for terminal wealth
and the cumulative risk over the whole investment horizon, respectively. Fig. 1
US
shows the membership function about the goal for terminal wealth. Fig. 2 shows
the membership function about the goal for the cumulative risk over the whole
AN
investment horizon.
Remark: Notice that θW and θV determine the shapes of membership functions
µE (x) and µV (x), respectively, where θW > 0 and θV > 0. The larger parameters
M
E ( x)
1
PT
0.5
0
CE
EM E (WT 1 )
V ( x)
0.5
0
VM V ( x)
17
ACCEPTED MANUSCRIPT
Based on the maximization principle in Bellman and Zadeh (1970), the fol-
lowing relationship can be obtained
T
IP
Then, the bi-objective portfolio selection problem (P1 ) can be formulated as the
following single objective mixed-integer nonlinear programming problem (P2 ):
CR
max η
s.t. µE (x) ≥ η,
(P2 )
USµV (x) ≥ η,
x ∈ D,
AN
By Eqs. (33-34), the problem (P2 ) can be rewritten as the following form:
max η
M
s.t. η + exp(−θW dW )η ≤ 1,
(P3 )
η + exp(−θV dV )η ≤ 1,
ED
x ∈ D.
4. Solution algorithm
PT
For the mix-integer nonlinear programming problem (P3 ), the traditional op-
timization approaches cannot be used to solve the multi-period portfolio selec-
CE
tion problem with minimum transaction lots and cardinality constraints. For this
reason, we propose a novel genetic algorithm to solve the multi-period portfolio
AC
18
ACCEPTED MANUSCRIPT
for solving different types of optimization problems since they make no restric-
tive assumptions about the solution space. Up to now, many researchers have
used the GAs to solve complex mixed-integer nonlinear programming problems.
Li and Gen (1996) and Yokota et al. (1996) proposed genetic algorithms based on
penalty function to solve mixed-integer nonlinear programming problems. Deep
T
et al. (2009) designed a real code genetic algorithm for solving integer and mixed
IP
integer constrained optimization problems. Recently, some researchers use GAs
CR
for optimizing the portfolio selection problems. Lin and Liu (2008) designed
a genetic algorithm with penalty function to solve portfolio selection problems
with minimum transaction lots. Soleimani et al. (2009) used genetic algorithm to
US
solve Markowitz-based portfolio selection with minimum transaction lots, cardi-
nality constraints and regarding sector capitalization. Baixauli-Soler et al. (2011)
AN
proposed a multi-objective genetic algorithm for solving mean-VaR portfolio se-
lection problem with minimum transaction lots and non-linear transaction costs.
M
In this section, we design a novel genetic algorithm on the basis of the solution
algorithm in Deep et al. (2009) to solve the problem (P3 ). We first introduce its
ED
19
ACCEPTED MANUSCRIPT
T
the integer part of X. To make the individual satisfy the cardinality constraint,
IP
we keep the largest K variables in each group. After truncation operation and
CR
cardinality repair, we normalize the remainder components in each group to keep
them satisfy the budget constraint. Furthermore, we calculate the violent value of
the individual X, denoted it as V iol(X).
US
Evaluation function: The evaluation function of this algorithm is formulated as
follows
AN
f (Xi ), if Xi is feasible,
eval(Xi ) = P
m (36)
fworst + max{gj (Xi ), 0} if otherwise,
M
j=1
where fworst is the objective function value of the worst feasible solution at cur-
ED
rent population. From Eq. (36), we can find out that, if Xi is an infeasible solu-
tion, then its evaluation value depends on both the amount of constraint violation
value and the population of solutions at hand. However, if Xi is a feasible so-
PT
lution, then its evaluation value is equal to the objective function value f (Xi ).
Pm
max{gj (Xi ), 0} is the sum of the violent values of Xi both inequality con-
CE
j=1
straints and equality constraints (where gj (Xi ) denotes as the left side term of the
jth inequality constraint minus its right side term). For the equality constraints
AC
in the problem (P3 ), we can transform them into corresponding inequality con-
straints by using a tolerance. If there are no feasible solutions in the population,
then fworst is set to zero. Notice that Eq. (36) is an evaluation function without
penalty factor for constraint handling.
For any given two solutions, we compare them by the following rules in Deb
20
ACCEPTED MANUSCRIPT
(2000):
1) A feasible solution is always preferred over an infeasible one;
2) Between two feasible solutions, the one with better objective function is pre-
ferred;
3) Between two infeasible solutions, the one with smaller constraint violation is
T
preferred.
IP
Crossover operation: The crossover operation is performed by Laplace crossover
CR
in Deep and Thakur (2007a). The Laplace crossover is summarized as follows.
For two randomly selected parents C1 and C2 , we generate two offsprings as fol-
lows
0
US
C1 = C1 + λ|C1 − C2 |,
AN
0
C2 = C2 + λ|C1 − C2 |,
t,i t,i
λt,i = (37)
φ + ϕlog(ν ), if c ≥ r,
t,i t,i
PT
crossover operation, we can obtain four individuals, i.e., two parents and two off-
springs. Then, we perform constraint-handling operations on the two offsprings.
AC
And then, we check the feasibility of the four individuals and select the two better
individuals as the individuals for next generation.
Mutation operation: The mutation operation is based on power mutation in Deep
and Thakur (2007b). In this process, a random number r is generated by the power
distribution r = (r1 )p , where r1 is a random number in (0,1) and p is the mutation
21
ACCEPTED MANUSCRIPT
x−xl
where y = xu −x
; xl and xu are the lower and upper bounds on the value of the de-
T
cision variable; y is a random number in (0,1). Similar to crossover operation, we
IP
perform the constraint-handling operation on the two obtained offsprings. After
checking the feasibility of the two offsprings and their parents, we select the two
CR
better individuals for next generation.
US
Selection operation: We use the common spinning the roulette wheel technique
to perform the selection operation. To do that, we can keep the better chromo-
AN
somes have more chance to produce offspring. Notice that, for the problem (P3 ),
the chromosome with a higher objective value is better, i.e. the chromosome with
M
a higher fitness value is better. Before performing the spinning the roulette wheel,
we first introduce a rank-based evaluation function denoted by eval(Ck ) to calcu-
ED
22
ACCEPTED MANUSCRIPT
After repeating this process pop− size times, we can obtain pop− size parents for
next generation.
Stopping criterion: If the stopping criterion (maximum number of iterations
Gmax ) is satisfied, or a given number of generations was reached without improve-
ment of the fitness function values, then we terminate the process of iteration.
T
The procedure of the designed algorithm can be summarized as follows:
IP
Step 1: Set parameters: population size pop− size, crossover probability Pc , mu-
CR
tation probability Pm and maximum iteration Gmax ;
Step 2: Randomly generate pop− size solutions in the bound constraints of vari-
ables;
US
Step 3: Calculate the fitness value of each individual;
Step 4: Update the chromosomes by crossover and mutation operations
AN
Step 5: Select the chromosomes by spinning the roulette wheel;
Step 6: Repeat the second to fifth steps for a given number of cycles;
M
Step 7: Terminate the repetition cycle when a given maximum number of itera-
tions is reached. Report the best chromosome as the optimal solution.
ED
5. Empirical study
PT
To illustrate the idea of our model and the designed algorithm, let us con-
sider an empirical application in Chinese stock markets. Assume that there are
CE
10 stocks and a risk-free asset in a financial market for trading. The exchange
codes of the 10 stocks are 600005, 600068, 000858, 600009, 000031, 000056,
AC
23
ACCEPTED MANUSCRIPT
asset to simulate the future transaction. Here, we collect the weekly closing pric-
ing of these stocks in six years from Jan. 2004 to Jan. 2010 as the sample data.
We set each two years as an investment period to handle these historical data. We
assume that the return rates of stocks are trapezoidal fuzzy variables. The corre-
sponding trapezoidal possibility distributions of these stocks at each period which
T
are shown in the following Table 1. Assume that the market price of each stock
IP
can be reflected by its historical information. Table 2 shows the market prices of
CR
these stocks at the date of transaction for each period.
Assume that the transaction cost rate is 0.003, i.e., ct,i = 0.003 (t = 1, 2, 3; i =
1, 2, . . . , 10). The expected return rate levels of the portfolios on the three invest-
ment periods are set as 0.0245, 0.0267 and 0.0333, respectively. The return rate
US
of risk-free asset at each period is set as 0.02. The risk tolerance levels of the
portfolios on the three investment periods are set as 0.015, 0.018 and 0.024, re-
spectively. The maximum number of assets in the portfolio at each period is set to
AN
6, i.e., K = 6. The diversification degree of the portfolio at each period is set to
1.2. Here, we set EM = 1.12 × 105 and VM = 0.025.
Table 1. The possibility distributions of the returns of assets at each period.
M
t Stock 1 2 3 4 5 6 7 8 9 10
a1,i -0.0066 -0.0641 -0.0106 0.0187 -0.0009 0.0248 -0.0649 -0.0166 -0.0623 -0.0071
ED
b1,i 0.0335 0.0409 0.0252 0.0423 0.0425 0.0399 0.0411 0.0329 0.0220 0.0439
t=1
α1,i 0.1195 0.0711 0.1723 0.1167 0.2119 0.2055 0.1142 0.2152 0.0592 0.2947
β1,i 0.1206 0.2179 0.2082 0.1389 0.4283 0.3212 0.3052 0.4316 0.1589 0.3429
a2,i -0.0407 -0.1059 -0.0741 -0.0024 0.0247 -0.1032 0.0135 -0.0193 0.0193 0.0198
PT
b2,i 0.0166 0.0450 -0.0296 0.0233 0.0406 0.0089 0.0172 0.0196 0.0665 0.0483
t=2
α2,i 0.1539 0.0275 0.0241 0.1291 0.2819 0.0913 0.1922 0.2105 0.1952 0.2104
β2,i 0.3342 0.2805 0.5995 0.3055 0.3914 0.3710 0.3221 0.3550 0.2179 0.2385
CE
a3,i 0.0258 -0.1311 0.0371 0.0323 -0.0805 -0.0042 -0.0063 0.0174 0.0204 -0.0462
b3,i 0.0362 0.0317 0.0726 0.0496 0.0562 0.0630 0.0446 0.0289 0.0711 0.0501
t=3
α3,i 0.2561 0.1861 0.2945 0.3458 0.2808 0.2425 0.1280 0.1011 0.2585 0.1620
AC
β3,i 0.2800 0.3261 0.3685 0.4145 0.6293 0.3142 0.3200 0.1703 0.3405 0.4777
t Stock 1 Stock 2 Stock 3 Stock 4 Stock 5 Stock 6 Stock 7 Stock 8 Stock 9 Stock 10
t = 1 6.41 1.93 3.49 5.93 22.31 14.94 49.09 30.20 26.33 26.49
t = 2 9.92 1.23 5.75 8.25 20.72 6.96 86.62 12.61 36.08 21.56
t = 3 29.11 6.06 24.51 14.12 18.57 12.34 178.96 25.51 145.91 82.00
24
ACCEPTED MANUSCRIPT
T
4. Table 3 shows the transaction lots of the 10 stocks at each period. Table 4
gives the invested amounts of the 10 stocks and the risk-free asset at each period.
IP
The satisfaction degree η, terminal wealth, net profit and total transaction cost are
summarized in Table 5.
CR
Table 3. The transaction lots of 10 stocks at each period.
(θE , θV ) t Stock 1 Stock 2 Stock 3 Stock 4 Stock 5 Stock 6 Stock 7 Stock 8 Stock 9 Stock 10
(20, 80)
t=1
t=2
t=3
0
0
0
0
2
0
1
0
5
18
17
0
US
11
19
9
1
24
0
0
23
0
3
3
0
0
0
0
0
4
0
0
5
0
9
3
0
0
0
0
0
AN
t=1
(50, 50) t=2 0 0 12 19 0 3 0 0 6 0
t=3 0 2 9 1 0 4 0 0 4 0
t=1 0 0 0 13 21 0 0 5 0 0
M
Table 4. The invested amounts of 10 stocks and risk-free asset (RFA) at each period.
(θE , θV ) t Stock 1 Stock 2 Stock 3 Stock 4 Stock 5 Stock 6 Stock 7 Stock 8 Stock 9 Stock 10 RFA
t = 1 0.00 386.00 1745.00 6523.00 53544.00 0.00 0.00 12080.00 0.00 0.00 25499.17
PT
(20, 80) t = 2 0.00 0.00 10350.00 7425.00 0.00 2088.00 0.00 0.00 32472.00 0.00 51485.48
t = 3 0.00 606.00 41667.00 1412.00 0.00 3702.00 0.00 0.00 43773.00 0.00 15693.98
t = 1 0.00 0.00 0.00 11267.00 51313.00 0.00 0.00 15100.00 0.00 0.00 12396.98
CE
(50, 50) t = 2 0.00 0.00 6900.00 15675.00 0.00 2088.00 0.00 0.00 21648.00 0.00 65674.29
t = 3 0.00 4902.00 22059.00 1412.00 0.00 4936.00 0.00 0.00 58364.00 0.00 26250.02
t = 1 0.00 0.00 0.00 7709.00 46851.00 0.00 0.00 13165.00 0.00 0.00 25548.31
AC
(80, 20) t = 2 0.00 0.00 4025.00 9900.00 0.00 0.00 0.00 0.00 43296.00 6468.00 50621.64
t = 3 0.00 0.00 4025.00 0.00 0.00 0.00 0.00 0.00 43773.00 6468.00 31966.25
25
ACCEPTED MANUSCRIPT
Table 5. Satisfaction degree η, terminal wealth, net profit and total transaction cost.
(θE , θV ) Terminal wealth Net profit Total transaction cost Satisfaction degree η
T
From Tables 3, 4 and 5 above, we find that different investment preferences
IP
lead to different investment strategies. If preference parameter vector (θE , θV ) is
CR
set as (20,80), it means that the investor requires the less vagueness on µV (X)
than that on µE (X). Namely, the investor is conservative. In this case, the in-
US
vestor should select a risk-free asset and four stocks (i.e., Stocks 1, 2, 4 and 5)
for constructing portfolio at period 1. The buying lots of the four stocks are 2,
1, 7 and 37, respectively. The investment amounts on the four stocks and the
AN
risk-free asset at period 1 are 1282.00, 193.00, 4151.00, 82547.00 and 19680.76
CNY, respectively. Then, at the beginning of period 2, the investor will read-
M
just his portfolio. After adjustment, the investor will select Stocks 2, 3, 4, 6, 7,
9 and the risk-free asset for constructing portfolio. The transaction lots of the
ED
696.00, 17324.00, 10824.00 and 58940.51 CNY, respectively. Then, at the be-
ginning period 3, the investor should adjust his portfolio again. After adjustment,
CE
Stocks 3, 7, 9 and the risk-free asset are selected for constructing portfolio. The
transaction lots on the selected three stocks are 7, 2 and 3, respectively. The
AC
corresponding investment amounts on the three stocks and the risk-free asset are
17157.00, 35792.00, 43773.00 and 10201.54 CNY, respectively. In this case, the
obtained terminal wealth is 112713.18 CNY, the net profit is 12713.18 CNY, the
total transaction cost is 622.81 CNY and the investor’s satisfaction degree about
the investment strategy is 0.8231. However, if the investor shows a neutral attitude
26
ACCEPTED MANUSCRIPT
toward both terminal wealth and cumulative risk, we set the preference parameter
vector (θE , θV ) as (50,50). The transaction lots of the 10 stocks of the three invest-
ment periods are shown in lines 5-7 in Table 4 and the corresponding investment
amounts of the eleven assets in the three investment periods are given in lines 5-7
in Table 5. In this case, the terminal wealth is 112431.37 CNY, the net profit is
T
12431.37 CNY, the total transaction cost is 638.99 CNY and the investor’s sat-
IP
isfaction degree about the investment strategy is 0.9108. If the investor requires
CR
the less vagueness on µE (X) than that on µV (X). Here, the preference parame-
ter vector (θE , θV ) is set as (80,20). The transaction lots of the 10 stocks in the
three investment periods are shown in lines 8-10 in Table 4 and the correspond-
US
ing investment amounts of the eleven assets on the three investment periods are
given in lines 8-10 in Table 5. The obtained terminal wealth is 112123.27 CNY,
AN
the net profit is 12123.27 CNY, the total transaction cost is 666.24 CNY and the
satisfaction degree about the two decision objectives is 0.7432. If the investor is
M
not satisfied with any of the obtained investment strategy, he could reset the value
of θE and θV .
ED
6. Conclusion
PT
semivariance model for multi-period fuzzy portfolio selection problem with mini-
mum transaction lots. In the proposed model, we consider six criteria, i.e., return,
AC
27
ACCEPTED MANUSCRIPT
each objective and then the proposed model is transformed into a single objective
mix-integer nonlinear programming problem. A genetic algorithm is designed
for solving the proposed mixed-integer nonlinear programming problem. A nu-
merical example is given to illustrate the idea of our model and demonstrate the
effectiveness of the designed algorithm. The computation results show that the
T
proposed model can express investors’ intention by varying the parameter values
IP
of the satisfaction degree.
CR
Acknowledgements
US
We thank Editor Lorenzo Peccati and the anonymous reviewers for their pen-
etrating remarks and suggestions, which led to an improved version of this pa-
AN
per. This research was supported by the National Natural Science Foundation of
China (No.70825005), GDUPS (2010) and the Fundamental Research Funds for
the Central Universities (2014ZP0005).
M
References
ED
Barak, S., Abessi, M., & Modarres, M. (2013). Fuzzy turnover rate chance
constraints portfolio model. European Journal of Operational Research, 228
AC
(1), 141-147.
Basse, T., Friedrich, M. and Vazquez Bea, E. (2009). REITs and the financial
crisis: Empirical evidence from the US. International Journal of Business
and Management, 4(11), 1-8.
28
ACCEPTED MANUSCRIPT
T
Calafiore, G. C., 2008. Multi-period portfolio optimization with linear con-
IP
trol policies. Automatica, 44 (10), 2463-2473.
CR
Carlsson, C., & Fullér, R. (2001). On possibilistic mean value and variance
of fuzzy numbers. Fuzzy Sets and Systems, 122 (2), 315-326.
US
Carlsson, C., Fullér, R., & Majlender, P. (2002). A possibilistic approach
to selecting portfolios with highest utility score. Fuzzy Sets and Systems,
AN
131(1), 13-21.
(2/4), 311-338.
Deep, K., Singh, K. P., Kansal, M. L., & Mohan C. (2009). A real coded ge-
netic algorithm for solving integer and mixed integer optimization problems.
Applied Mathematics and Computation, 212 (2), 505-518.
29
ACCEPTED MANUSCRIPT
Deep, K., & Thakur, M. (2007a). A new crossover operator for real coded
genetic algorithms. Applied Mathematics and Computation, 188 (1), 895-
912.
Deep, K., & Thakur, M. (2007b). A new mutation operator for real coded
T
genetic algorithms. Applied Mathematics and Computation, 193 (1), 211-
IP
230.
CR
Dubois, D., & Prade, H. (1980). Fuzzy sets and systems: Theory and appli-
cations. Academic Press, New York.
US
Elton, E. J., & Gruber, M. J. (1974). On the optimality of some multiperiod
portfolio selection criteria. Journal of Business, 47 (2), 231-243.
AN
Fama, E. F. (1970). Multiperiod consumption-investment decisions. The
American Economic Review, 60 (1), 163-174.
M
Fang, Y., Lai, K. K., & Wang, S. Y. (2006). Portfolio rebalancing model
with transaction costs based on fuzzy decision theory. European Journal of
ED
Fu, C. P., Lari-Lavassani, A., & Li, X. (2010). Dynamic mean-variance port-
AC
30
ACCEPTED MANUSCRIPT
Gülpınar, & N., Rustem, B. (2007). Worst-case robust decisions for multi-
period mean-variance portfolio optimization. European Journal of Opera-
tional Research, 183(3), 981-1000.
T
general theory of portfolio choice. Journal of Finance, 26 (4), 857-884.
IP
Holland, J. H. (1975). Adaptation in natural and artificial systems. Ann Ar-
CR
bor: University of Michigan Press.
Hwang, C. L., & Yoon, K. (1981). Multiple attribute decision making: Meth-
US
ods and applications, Springer-Verlag, Heidelberg.
Li, Y., & Gen, M. (1996). Nonlinear mixed integer programming problems
AN
using genetic algorithm and penalty function. Proceedings of 1996 IEEE Int.
Conf. on SMC, pp. 2677-2682.
M
Li, D., & Ng, W. L., 2000. Optimal dynamic portfolioselection: multiperiod
mean-variance formulation. Mathematical Finance, 10 (3), 387-406.
ED
Liu, Y. J., Zhang, W. G., & Xu, W. J. (2012). Fuzzy multi-period portfolio
CE
Lin, C. C., & Liu, Y. T. (2008). Genetic algorithms for portfolio selection
problems with minimum transaction lots. European Journal of Operational
Research, 185 (1), 393-404.
31
ACCEPTED MANUSCRIPT
T
under concave transaction costs and minimal transaction unit constraints.
IP
Mathematical Programming, 89 (2), 233-250.
CR
Konno, H., & Wijayanayake, A. (2002). Portfolio optimization under D.C.
transaction costs and minimal transaction unit constraints. Journal of Global
Optimization, 22 (1/4), 137-154.
US
Konno, H., Yamazaki, H. (1991). Mean-absolute deviation portfolio opti-
AN
mization model and its application to Tokyo Stock Market. Management
Science, 37, 519-531.
M
Mansini, R., & Speranza, M. G. (1999). Heuristics algorithms for the port-
ED
Sadjadi, S. J., Seyedhosseini, S. M., & Hassanlou, Kh. (2011). Fuzzy mul-
tiperiod portfolio selection with different rates for borrowing and lending.
Applied Soft Computing, 11 (4), 3821-3826.
Saeidifar, A., & Pasha, E. (2009). The possibilistic moments of fuzzy num-
32
ACCEPTED MANUSCRIPT
T
and regarding sector capitalization using genetic algorithm. Expert Systems
IP
with Applications, 36 (3), 5058-5063.
CR
Speranza, M. G. (1996). A heuristic algorithm for a portfolio optimization
model applied to the Milan stock market. Computers & Operations Re-
search, 23 (5), 433-441.
US
Tanaka, H., & Guo, P. (1999). Portfolio selection based on upper and lower
AN
exponential possibility distributions. European Journal of Operational Re-
search, 114 (1), 115-126.
M
Vercher, E., Bermúdez, J. D., & Segura, J. V. (2007). Fuzzy portfolio op-
timization under downside risk measures. Fuzzy Sets and Systems, 158 (7),
ED
769-782.
Yokota, T., Gen, M., & Li, Y. X. (1996). Genetic algorithm for non-linear
mixed integer programming problems and its applications. Computers & In-
AC
33
ACCEPTED MANUSCRIPT
T
Zhang, W. G., Wang, Y. L., Chen, Z. P., & Nie, Z. K. (2007). Possibilistic
IP
mean-variance models and efficient frontiers for portfolio selection problem.
CR
Information Sciences, 177 (13), 2787-2801.
Zhang, W. G., Zhang, X. L., & Xiao, W. L., 2009. Portfolio selection under
US
possibilistic mean-variance utility and a SMO algorithm. European Journal
of Operational Research, 197 (2), 693-700.
AN
Zhu, S. S., Li, D., & Wang, S. Y. (2004). Risk control over bankcruptcy in
dynamic portfolioselection: A generalized mean-variance formulation. IEEE
M
34