Journal Pre-proof

of
Abstract

This paper deals with a portfolio optimization problem with sell orders, in
which the investor will immediately sell the risky assets once their prices reach

pro
the preset sell thresholds. Departing from the traditional portfolio models,
the investor needs to determine not only an investment proportion but also
a sell threshold for each risky asset. We formulate the portfolio problem as a
sequential decision problem and propose an automatic trading system with three
stages to determine the investment strategy at the beginning of each period. In
Stage 1, we formulate the return of an investment strategy as an LR-power fuzzy
re-
number based on the historical data and propose a fuzzy mean-semi-variance
portfolio optimization model with sell orders. In Stage 2, we design a multi-
objective genetic algorithm to solve the proposed model. In Stage 3, we select
an optimal investment strategy among the efficient solutions based on the fuzzy
lP
Value-at-Risk ratio. Moreover, we conduct two case studies in the real stock
market to illustrate the effectiveness and practicability of the proposed model
and algorithm. The comparison results show that the proposed trading system
has a better out-of-sample performance than the other ones.
Keywords: Fuzzy portfolio optimization, Automatic trading system, Sell
rna

order, Genetic algorithm

1. Introduction

Portfolio selection seeks optimal capital allocation among specific risky as-
sets, which is an important financial problem attracting the attention of many
Jou

academic researchers and practitioners. The pioneering mean-variance (MV)

portfolio selection model initially proposed by Markowitz (1952) laid the foun-
dation for the modern portfolio theory. After that, many scholars studied the
portfolio selection problem based on probability theory (Masmoudi & Fouad,
2018; Kalayci et al., 2019). As an alternative tool to probability theory, the
fuzzy set theory proposed by Zadeh (1965) can effectively handle the fuzzy and

2
 Journal Pre-proof

vague information with linguistic descriptions such as high return, low risk, etc.

of
On the basis of the fuzzy set theory, many scholars regarded the future returns
of risky assets as fuzzy numbers and proposed various fuzzy portfolio selection
models (Huang, 2009; Zhang et al., 2018).

pro
Although many research works on the portfolio selection problem in fuzzy
environment, no scholar has yet focused on the sell order behavior during the
investment process. The existing fuzzy portfolio models assume that the investor
reallocates his/her wealth at the beginning of the investment period and holds
the portfolio until the end of the period. However, in the highly developed
electronic financial market, the investor may sell the risky assets immediately

re-
to obtain a certain high return or avoid subsequent losses. Assume that the
investor adjusts the asset positions and further sets the sell orders for the risky
assets at the beginning of the investment period. During the investment process,
the sell orders will be executed once the prices of risky assets reach their sell
thresholds.
lP
We give a small example to justify why the sell order is important. Assume
that an investor purchases a risky asset with price p and plans to invest for
a period of time. We show the price path of the asset during the investment
period in Figure 1. Denote the lowest, highest, and closing prices of the asset
rna

during the investment period by l, h, and c, respectively. For the case without
the sell order, the investor holds the asset until the end of the period and gets
an investment return c/p. While, for the case with the sell order, the investor
presets a sell threshold, denoted by s, and sells the asset once its price reaches
s. If and only if s ∈ [l, h], the sell order is executed, and the investment return
is s/p. Furthermore, if s ∈ (c, h], the investor gets a higher return than the case
Jou

without the sell order. Since the price path of the asset is uncertain, how to
determine an appropriate sell threshold is a challenge. This paper is devoted to
dealing with the fuzzy portfolio optimization problem with sell orders to meet
this challenge. It can be foreseen that considering sell orders on the portfolio
selection model can help the investor captures more investment opportunities
during the investment process.

3
 Journal Pre-proof

of
h

Sell
Price

pro
p
c

Begin End
Time

Figure 1: The price path of the risky asset during an investment period.

re-
In this paper, we propose an automatic trading system to deal with the fuzzy
portfolio optimization problem with sell orders, in which the investor needs to
determine not only the portfolio but also the sell thresholds for the risky assets
lP
at the beginning of each period. The investor will sell a risky asset once its
price reaches its sell threshold. Firstly, we consider the return of an invest-
ment strategy as an LR-power fuzzy number and propose the WFF method to
determine its parameters. Then, we use the credibility expected value and semi-
variance of the investment strategy to measure its return and risk, respectively,
rna

and propose a fuzzy mean-semi-variance portfolio optimization model with sell

orders. Next, we design a multi-objective genetic algorithm (MOGA) to solve
the proposed model. Finally, we select an optimal investment strategy from the
efficient solutions based on the fuzzy Value-at-Risk ratio. We summarize the
main contributions as follows:
Jou

1) An automatic trading system for the fuzzy portfolio optimization problem

with sell orders is proposed.
2) An estimation method is proposed to determine the parameters of the LR-
power fuzzy return of a given investment strategy.
3) A multi-objective genetic algorithm is designed to solve the proposed model.
4) An optimal investment strategy is selected from the efficient solutions based

4
 Journal Pre-proof

on the fuzzy Value-at-Risk ratio.

of
We organize the remainder of this paper as follows. Section 2 reviews the
related works. Section 3 introduces some preliminaries. Section 4 establishes
an automatic trading system for the portfolio optimization problem with sell

pro
orders. Section 5 conducts two real case studies to illustrate the effectiveness of
the proposed trading system. Section 7 concludes this paper.

2. Literature review

Markowitz (1952) proposed the mean-variance (MV) portfolio selection mod-

re-
el which has laid the foundation for the modern portfolio theory. The MV mod-
el assumes that the returns of risky assets are random variables and measure
the investment return and risk in terms of the expected value and variance of
the portfolio, respectively. Inspired by the idea of the MV model, many ex-
tension works have been proposed, such as Merton (1972), Cai et al. (2000),
lP
Ray & Jenamani (2016), Bi et al. (2018).
The above studies characterize the returns of risky assets as random vari-
ables based on probability theory. However, there exist some non-probability
factors in the portfolio selection problem, such as political factors, expert opin-
rna

ions, and company performances. As an alternative tool to probability theory,

the possibility theory initially proposed by Zadeh (1965) is widely applied to s-
tudy the possibility portfolio selection problem. Carlsson et al. (2000) proposed
a fuzzy portfolio selection model with the highest utility score. Tsaur (2013)
proposed a fuzzy portfolio selection model with different investor risk attitudes.
Liu & Zhang (2015) proposed a multi-period possibilistic mean-semi-variance
Jou

portfolio selection model with minimum transaction lots. Lu & He (2020) pro-
posed an improved quantum-behaved particle swarm optimization algorithm to
solve the fuzzy portfolio selection models. Gong et al. (2021) proposed a regret
theory-based fuzzy multi-objective portfolio selection model considering higher
moments and DEA cross-efficiency. Tsaur et al. (2021) proposed a possibilistic

5
 Journal Pre-proof

mean-variance portfolio selection model considering the guaranteed return rate

of
for excess investment.
Considering the possibility measure is not self-dual, Liu & Liu (2002) pro-
posed the credibility measure to characterize the occurrence chance of fuzzy

pro
events, which is widely used to model the uncertain returns of risky assets.
Huang (2008) proposed a mean-semi-variance portfolio selection model under
the credibility framework. Li et al. (2010) proposed a fuzzy mean-variance-
skewness portfolio selection model considering the asymmetry of fuzzy return-
s. Vercher & Bermúdez (2015) proposed a credibility mean-absolute deviation
framework for the multi-objective portfolio selection problem with cardinality

re-
constraint. Guo et al. (2016) assumed that risky assets have different invest-
ment horizons and proposed a multi-period credibility mean-variance portfolio
selection model with different investment horizons. Jalota et al. (2017b) pro-
posed an automatic process to generate the parameters of the LR-power fuzzy
returns for risky assets and proposed four credibilistic multi-criteria portfolio
lP
selection models. Mohebbi & Najafi (2018) proposed a multi-period credibility
portfolio selection model based on scenario tree. Wang et al. (2018) proposed a
bi-objective portfolio selection model with objectives of fuzzy Share ratio and
fuzzy Value-at-Risk ratio. Garcı́a et al. (2019) proposed a credibility multi-
rna

objective portfolio selection model considering environment, social and gover-

nance scores. Zhang (2019) proposed a random credibility portfolio selection
model with different convex transaction costs. Guo et al. (2020) proposed a
fuzzy hidden Markov-switching portfolio selection model considering the capital
gain tax and transaction costs. Garcı́a et al. (2020) proposed two credibility
mean-semivariance/CVaR-liquidity portfolio selection models with some realis-
Jou

tic constraints. Mehlawat et al. (2020) proposed two fuzzy multi-objective port-
folio selection models by taking variance and conditional value at risk as risk
measures, respectively, and designed a DEA method with multiple inputs and
outputs to evaluate the performances of the portfolios. Li et al. (2021) proposed
a random credibility multi-objective portfolio selection model for large-scale se-
curities data and design an intelligent hybrid algorithm to solve it. Gupta et al.

6
 Journal Pre-proof

(2021) use the coherent fuzzy numbers to model the investor’s perception of

of
the stock market and proposed two multi-period fuzzy portfolio optimization
models.
Due to the convenience of linear membership function, many fuzzy portfo-

pro
lio selection models regard the returns of risky assets as triangular or trapezoid
fuzzy numbers, such as Qin (2015), Mashayekhi & Omrani (2016), Liu & Zhang
(2018), Zhou et al. (2018), Garcı́a et al. (2020). However, the linear mem-
bership function may not be quite fit for the historical data and other infor-
mation. The LR-power fuzzy number, a more generic fuzzy number whose
shape is more variable, is proved to be a good choice to model the fuzzy

re-
returns (Vercher & Bermúdez, 2013; Jalota et al., 2017a; Garcı́a et al., 2019;
Yang et al., 2021). Moreover, most of the traditional fuzzy portfolio selection
models assume that the fuzzy returns of the risky assets are independent. The
applications of the above models do not yield a distributive investment, which
contradicts with the traditional portfolio theory (Inuiguchi & Tanino, 2000).
lP
Based on the above considerations, Vercher & Bermúdez (2015) designed an
estimation method to approximate the parameters of the fuzzy return of a
given portfolio directly instead of each risky asset independently. Inspired by
Vercher & Bermúdez (2015), we propose an estimation method, named weight-
rna

ed fuzzy frequency (WFF), to determine the parameters of the LR-power fuzzy

return for a given investment strategy. There are two main differences between
the proposed estimation method and that of Vercher & Bermúdez (2015). One
is that this study considers the transaction costs and sell orders in the modeling
process. The other is that Vercher & Bermúdez (2015) determines the parame-
ters based on the sample percentiles of the portfolio’s historical returns, whereas
Jou

this study determines the parameters based on the fuzzy frequencies and fuzzy
percentiles of the returns of the portfolio under the historical market states.
Although there are abundant achievements in the fuzzy portfolio selection
models, we have not found one that considering sell orders. Noting that the sell
order is an important investment behavior in the financial market, we introduce
it into the fuzzy portfolio optimization and propose an automatic trading system

7
 Journal Pre-proof

with sell orders to enrich the theoretical implications for the fuzzy portfolio

of
theory. In the trading system, we design a multi-objective genetic algorithm to
solve the proposed model and apply the fuzzy VaR ratio to select an optimal
investment strategy from the efficient solutions. With the inputs of trading

pro
price data, the trading system automatically generates the investment strategy
involving the investment portfolio and sell orders for the risky assets, which
can effectively provide decision support for investors in the face of the portfolio
management problem with sell orders.

3. Preliminaries

re-
In this section, we introduce some preliminaries used in the following discus-
sion.

3.1. Credibility theory

Definition 1. (Liu & Liu, 2002) Let ξ be a fuzzy number with membership
lP
function µξ . For a set A ⊆ R, define the credibility measure of the statement
“ξ ∈ A” by the average of the possibility and necessity measures, i.e.,
( )
1
Cr{ξ ∈ A} = sup µξ (x) + 1 − sup µξ (x) (1)
2 x∈A x∈Ac

Obviously, the credibility measure is self-dual, i.e., for any set A ⊆ R, Cr{ξ ∈
rna

A} + Cr{ξ ∈ Ac } = 1.

Definition 2. (Liu & Liu, 2002) Let ξ be a fuzzy number. Define its credibility
expected value by
∫ +∞ ∫ 0
E[ξ] = Cr{ξ ≥ x}dx − Cr{ξ ≤ x}dx (2)
0 −∞
Jou

provided that at least one of the above two integrals is finite.

Definition 3. (Huang, 2008) Let ξ be a fuzzy number with finite expected

value. Define its credibility semi-variance by
[( )2 ]
−
V − [ξ] = E (ξ − E[ξ]) (3)

where (·)− = max{−(·), 0}.

8
 Journal Pre-proof

Definition 4. (Wang et al., 2018) Let L be a fuzzy number representing the

of
loss of an investment. Define the Value-at-Risk (VaR) of the investment under
confidence level θ ∈ (0, 1) by

VaRθ [L] = sup {x : Cr{L ≤ x} ≤ θ} (4)

pro
It can be seen that VaRθ [L] indicates that the greatest loss of the investment
under confidence level θ.

Definition 5. (Wang et al., 2018) Let r be a fuzzy number representing the

future return rate of an investment. Denote the loss of the investment by −r.
Define the fuzzy VaR ratio of the investment under confidence level θ by
re- V R[R] =
E[r] − rf
VaRθ [−r]
(5)

where rf is the risk-free rate.

3.2. Weighted fuzzy frequency method

lP
Definition 6. A fuzzy number ξ is said to be an LR-power fuzzy number if its
membership function is of the following form
 ( )δ

 a−x

 1 − , if α < x < a

 α

rna




 1, if a ≤ x ≤ b
µξ (x) = ( )κ (6)

 x−b

 1− , if b < x < b + β

 β





0, otherwise

where α, β > 0 are the left and right spreads, respectively; δ, κ > 0 are the left
and right shape parameters. Denote it by ξ = (a, b, α, β)δ,κ .
Jou

We show the membership functions of LR-power fuzzy number ξ = (a, b, α, β)δ,κ

with different values δ and κ in Figure 2.
Next, we formulate an uncertain quantity as an LR-power fuzzy number and
design an estimation method to determine its parameters based on its sample
data.

9
 Journal Pre-proof

of
1

(x) >1 >1

=1 =1

pro
<1 <1

0
a- a b b+
x

Figure 2: The membership functions of ξ = (a, b, α, β)δ,κ with different values

δ and κ. re-
Firstly, we introduce the definitions of fuzzy frequency and fuzzy quantile
for a given sequence data.
lP
Definition 7. Let x = (x1 , . . . , xN ) be a sequence data of uncertain quantity
ξ. For x ∈ R, define the fuzzy frequency of x in the sequence x by
N
∑ { }
|x − xj |
Fx (x) = hj max 1 − ,0 (7)
η
rna

j=1

where η > 0 is the width parameter and hj ∈ [0, 1] is the important degree of
the sample data xj , j = 1, . . . , N .

We can be seen that the fuzzy frequency Fx (x) indicates the weighted
“times” that x appears in the sequence data x.

Definition 8. Let x = (x1 , . . . , xN ) be a sequence data of uncertain quantity

Jou

ξ. For p ∈ [0, 1), define the fuzzy p-quantile of the sequence x by

{ ∫z }
−∞ x
F (x)dx
Qx (p) = sup z ∈ R : ∫ +∞ ≤p (8)
−∞ x
F (x)dx

Then, we employ [Qx (0.05), Qx (0.95)], [Qx (0.25), Qx (0.75)], and [Qx (0.45),
Qx (0.55)] to fit the support set [a − α, b + β], 0.5-level set [a − α · 0.51/δ , b +

10
 Journal Pre-proof

β · 0.51/κ ], and core set [a, b] of the LR-power fuzzy number ξ = (a, b, α, β)δ,κ ,

of
respectively, i.e., a = Qx (0.45), b = Qx (0.55), α = a−Qx (0.05), β = Qx (0.95)−
ln 0.5 ln 0.5
b, δ = ln((a−Qx (0.25))/α) , and κ = ln((Qx (0.75)−b)/β) .

3.3. Genetic algorithm

pro
Genetic algorithm (GA) is an evolutionary algorithm based on the simulation
of the biological evolution principle, which is widely used in solving the portfolio
selection problems, such as Bermúdez et al. (2012), Vercher & Bermúdez (2015)
and Guo et al. (2016). In GA, a solution is encoded by an individual in the
search space, and a group of individuals in the search space form a population.

re-
In the iteration process, each individual will be evaluated and endowed with
a fitness value. Then, the crossover and mutation operators are executed to
produce the offspring, in which the individuals with higher fitness have great
chance of being selected as the parents of the operators. We briefly introduce
the main process of the basic genetic algorithm:
lP
Step 1 Initialize the population in the search space randomly;
Step 2 Calculate the fitness of the individuals in the population;
Step 3 Execute the crossover and mutation operators to generate offspring;
Step 4 Execute the selection operator to save some individuals to the next
rna

generation;
Step 5 Repeat Steps 2 to 4 until termination condition is reached, and report
the best individual in the final population as the optimal solution for
the optimization problem.

4. Automatic trading system

Jou

In this section, we design an automatic trading system for the fuzzy portfolio
problem with sell orders.

4.1. Problem description and notations

We consider a portfolio optimization problem with a basket of risky assets.
Assume that the investor adjusts his/her investment strategy at the beginning

11
 Journal Pre-proof

of each period. We formulate the portfolio optimization problem as a sequential

of
decision problem, in which the investor needs to construct a portfolio and preset
sell orders for the risky assets at the beginning of each period. The sell order
of a risky asset is represented by a sell threshold different from its current

pro
price. For each risky asset, the investor will sell it once its price reaches its
sell threshold during each period. We consider two types of sell orders: one
is the limit order, which preset a sell threshold higher than the current price;
the other is the stop order, which preset a sell threshold lower than the current
price. In fact, we only care about whether the sell orders are executed, which
can be identified according to the highest and lowest prices of the risky assets

re-
during the investment period. Specifically, the limit sell order for a risky asset
is executed if its highest price is higher than or equal to the sell threshold, while
the sell stop order for a risky asset is executed if its lowest price is lower than
or equal to the sell threshold.
We introduce some notations used in the following discussion.
lP
• n: the number of the risky assets;
• w: the length of estimation window;
• T : the total periods of the stock data, which is w plus the periods of invest-
ment;
rna

c
• rt,i : the closing return of asset i during period t, which is its closing price
divided by its closing price of the previous period, i = 1. . . . , n, t = 1, . . . , T ;
h
• rt,i : the highest return of asset i during period t, which is its highest price
divided by its closing price of the previous period, i = 1. . . . , n, t = 1, . . . , T ;
l
• rt,i : the lowest return of asset i during period t, which is its lowest price
divided by its closing price of the previous period, i = 1. . . . , n, t = 1, . . . , T ;
Jou

• xt,i : the investment proportion of asset i during period t, i = 1, . . . , n;

• st,i : the sell threshold of asset i during period t, i = 1, . . . , n;
• xt : the portfolio of period t, which is represented by xt = (xt,1 , . . . , xt,n );
• st : the sell threshold vector of period t, which is represented by st = (st,1 , . . . , st,n );
• (xt ; st ): the investment strategy of period t;

12
 Journal Pre-proof

• Rt (xt ; st ): the LR-power fuzzy return of the investment strategy (xt ; st ).

of
Assume that the investor joins the market at the beginning of period w + 1
and plans to invest for T − w periods. We collect the historical data from
period 1 to period w, including the closing, highest and lowest returns of the n

pro
risky assets. For period t = w + 1, . . . , T , the investor determines the portfolio
xt = (xt,i , . . . , xt,n ) and the sell threshold vector st = (st,1 , . . . , st,n ) (st,i ̸= 1)
at the beginning of this period and updates the data of period t at the end of
this period. Assume that the investment is self-financed and the short selling
∑n
is not allowed, i.e., i=1 xt,i = 1 and xt,i ≥ 0. We regard the return of the
investment strategy Rt (xt ; st ) as an LR-power fuzzy number, and adopt the
re-
WFF method introduced in subsection 3.2 to determine its parameters based
on the historical data of the recent previous w periods. There are two reasons
for selecting the LR-power fuzzy number. The first one is that the LR-power
fuzzy number is more generic than the commonly used trapezoid one, and its
lP
membership function possesses a more flexible shape that can better fit the
distribution of historical data. The other is that it has good structure and
properties, and its related mathematical characteristics, such as expected value,
lower semi-variance, and fuzzy VaR, can be easily calculated.
rna

4.2. Estimation for the LR-power fuzzy return

Given an investment strategy (xt ; st ), we generate its LR-power fuzzy return

by the following two steps. Firstly, we evaluate its performance under the
historical market states of the recent previous w periods and generate a sequence
return data for the investment strategy. Then, we employ the WFF method
introduced in 3.2 to determine the parameters of Rt (xt ; st ) = (at , bt , αt , βt )δt ,κt .
Jou

(t′ )
We introduce a binary variable, i.e., χt,i , to indicate whether the sell order
of asset i is executed during period t under the market state of period t′ , t′ =
t − w, . . . , t − 1, i.e.,

(t′ )
χt,i = I(st,i > 1)I(st,i ≤ rth′ ,i ) + I(st,i < 1)I(st,i ≥ rtl′ ,i ) (9)

13
 Journal Pre-proof

where I(·) is the indicator function, taking value 1 if (·) is true and 0 otherwise.

of
If the sell order of asset i is executed, its return is the sell threshold; otherwise,
its return is the closing return. Thus, the return of asset i under the market
state of period t′ is

pro
(t′ ) (t′ ) (t′ )
rt,i = χt,i st,i + (1 − χt,i )rtc′ ,i (10)

Denote the transaction cost rate of asset i by ci and the initial investment
proportion of asset i at the beginning of period t by x0t,i (before adjusting).
Adopting the V-type transaction cost function, we formulate the transaction
cost of trading asset i under the market state of period t′ as
( )
(t)
re- (t)
ct,i = ci |xt,i − x0t,i | + xt,i st,i χt,i

The net return of the investment strategy (xt ; st ) under the market state of
period t′ is
(11)

n (
∑ )
(t′ ) (t′ ) (t′ )
rt = xt,i rt,i − ct,i (12)
lP
i=1

(t−w) (t−1)
The sequence return data for the investment strategy is rt = (rt , . . . , rt ).
We consider that the recent sample is more important to the future state, and
thus assume
0.6(t′ − t + w)
rna

′
htt = 0.4 + , t′ = t − w, . . . , t − 1.
t−1
According to Definition 7, for x ∈ R, the fuzzy frequency of x in rt is
t−1
∑ ′
′ x − rtt
Frt (x) = htt max{1 − , 0}.
η
t′ =t−w

According to Definition 8, for p ∈ [0, 1), the fuzzy p-quantile of rt by

{ ∫z }
Jou

F (x)dx
−∞ rt
Qrt (p) = sup z ∈ R : ∫ +∞ ≤p
−∞ rt
F (x)dx
According to the discussion in Subsection 3.2, we can obtain the parameters
of fuzzy return Rt (xt ; st ) = (at , bt , αt , βt )δt ,κt as follows: at = Qrt (0.45), bt =
ln 0.5
Qrt (0.55), αt = at − Qrt (0.05), βt = Qrt (0.95) − bt , δt = ln((at −Qrt (0.25))/αt )
,
ln 0.5
and κt = ln((Qrt (0.75)−bt )/βt )
.

14
 Journal Pre-proof

4.3. Mean-semi-variance model

of
Under the return-risk framework, we measure the investment return and risk
during period t by the expected value and semi-variance of Rt (xt ; st ), respec-
tively. The expected value of Rt (xt ; st ) is
∫ ∞ ∫

pro
0
E [Rt (xt ; st )] = Cr{Rt (xt ; st ) ≥ x}dx − Cr{Rt (xt ; st ) ≤ x}dx
0 −∞
(13)
at + bt κt βt δt αt
= + −
2 2(κt + 1) 2(δt + 1)
The semi-variance of Rt (xt ; st ) is
∫ E[Rt (xt ;st )]
V − [Rt (xt ; st )] = 2 (E [Rt (xt ; st )] − x) Cr{Rt (xt ; st ) ≤ x}dx
−∞

+
re-
(E [Rt (xt ; st )] − at + αt )
2
−
2
αt2
δt + 2
δ +1
−
αt (E [Rt (xt ; st )] − at )
δt + 1
(at − E [Rt (xt ; st )]) t max {at − E [Rt (xt ; st )] , 0}
αtδt (δt + 1)(δt + 2)
κ +1
(E [Rt (xt ; st )] − bt ) t max {E [Rt (xt ; st )] − bt , 0}
+
βtκt (κt + 1)(κt + 2)
lP
(14)
If the sell thresholds are set to be very high or low, the corresponding sell
orders are useless because they are almost impossible to be executed. Thus, we
set the lower and upper bounds of the selling threshold st,i as
rna

st,i = min rtl′ ,i and st,i = max rth′ ,i

t−w≤t′ ≤t−1 t−w≤t′ ≤t−1

respectively, which makes the sell order of asset i to be executed under at least
one historical market state.
To sum up, we formulate the fuzzy mean-semi-variance model for the port-
folio optimization problem with sell orders as


 max E [Rt (xt ; st )]
Jou








 min V − [Rt (xt ; st )]



 ∑ n
s.t. xt,i = 1 (15)



 i=1



 xt,i ≥ 0 i = 1, . . . , n





 st,i ≤ st,i ≤ st,i , st,i ̸= 1, i = 1, . . . , n

15
 Journal Pre-proof

A feasible investment strategy (x∗t ; s∗t ) is said to be a Pareto optimal solu-

of
tion or efficient solution of model (15) if there does not exist another feasi-
ble investment strategy (xt ; st ) such that E [Rt (xt ; st )] ≥ E [Rt (x∗t ; s∗t )] and
V − [Rt (xt ; st )] ≤ V − [Rt (x∗t ; s∗t )] with at least one strict inequality. For an effi-

pro
cient solution (x∗t ; s∗t ), the risk-return pair (V − [(x∗t ; s∗t )] , E [(x∗t ; s∗t )]) is said to
be an efficient point in the criterion space, and all the efficient points form the
efficient frontier.

4.4. Multi-objective genetic algorithm

The traditional GA is used to solve unconstrained single objective program-

re-
ming problems. Considering model (15) is a bi-objective model with some con-
straints, we find two challenges in applying the GA to solve the proposed model.
One is how to handle the constraints, and the other is how to evaluate the indi-
viduals based on the multiple objectives. In the MOGA, we design an encoding
method to handle the constraints of model (15) and propose a fitness function
lP
to evaluate a series of individuals based on their dominance relationship and
isolation degree.

4.4.1. Encoding and decoding

Define the search space by
rna

SP = {p = (p1 , . . . , p2n ) : 0 ≤ pj ≤ 1, j = 1, . . . , 2n} (16)

A population in a genetic algorithm comprises finite individuals or chromosomes,

each of which is encoded by a real-value vector p in the search space. For
p ∈ SP , the actual investment proportion and sell threshold of asset i are
decoded by

Jou

n
∑


 xt,i = pi / pi
i=1 , i = 1, . . . , n (17)



st,i = st,i + pn+i (st,i − st,i )
It can be seen that the solution represented by any individual in the search
space satisfies all the constraints of model (15) except for st,i ̸= 1. If st,i = 1, we
randomly generate a sufficient small positive number ϵ, and let pn+i ← pn+i + ϵ.

16
 Journal Pre-proof

4.4.2. Fitness

of
We design a fitness function to evaluate a series of individuals. The higher
the fitness of an individual, the greater chances of its reproduction during the
iteration process.

pro
If an individual is not inferior to the other for both two objectives and
superior to the latter for at least one objective, we say that the former dominates
the latter. An individual which is not dominated by anyone is called a non-
dominated individual. We assign the individuals into different dominance classes
based on their dominance relationship (see Figure 3). The higher the dominance
class of an individual, the higher its fitness.

re-
Expected value

lP
Class 1
Class 2
Class 3
Class 4
Class 5
Class 6

Semi-variance
rna

Figure 3: The dominance classes.

Assume there are N individuals needed to be evaluated, denoted by p1 , . . . , pN .

We employ the following procedure to determine the dominance classes of the
individuals.
Jou

Step 1 Initialize the dominance classes for all individuals as class 1, i.e., Cj = 1,
j = 1, . . . , N .
Step 2 Find all the non-dominated individuals from the current individual set,
and denote their corresponding index set by D.
Step 3 Let Cj ← Cj + 1 for j ∈
/ D.

17
 Journal Pre-proof

Step 4 Remove the above non-dominated individuals. If the current individual

of
set is empty, terminate the procedure; otherwise, return to Step 2.

Considering the necessity for the above procedure, we present an algorithmic

scheme to find the non-dominated individuals from a number of individuals.

pro
Assume there are N ′ individuals p1 , . . . , pN ′ . Denote the semi-variance and
expected value of pj by Vj and Ej , respectively, j = 1, . . . , N ′ . Without loss of
generality, assume V1 ≤ V2 ≤ · · · ≤ VN ′ . We employ the following procedure to
find all the non-dominated solutions from the N ′ individuals.

Step 1 Let S be a N × 3 matrix

 
re- 
S=.

1
 ..

N
V1
..
.
VN ′
E1
..
.
EN ′




(18)

Step 2 Initialize j = 1 and d = N ′ .

lP
Step 3 Let I1 = sign(Sj+1,2 − Sj,2 ) and I2 = sign(Sj+1,3 − Sj,3 ).
Step 4 If I1 < I2 , delete row j from S, and let j ← max{j −1, 1} and d ← d−1;
if I1 > I2 , delete row j + 1 from S, and let d ← d − 1; if I1 = I2 , let
j ← j + 1.
Step 5 If j < d, return to Step 3; otherwise, terminate the procedure, and
rna

report S.

The first column of S indicates the indexes for the non-dominated solutions,
and the second and third columns of S indicate their risk-return pairs.
Furthermore, we evaluate the individuals in the same dominance class based
on their isolation degree. To encourage diversity, we endow higher fitness for
Jou

an individual with higher isolation degree. Denote the number of the dominant
class by K. Assume there are Nk individuals in dominant class k, and denote
them by pk,1 , . . . , pk,Nk . Denote the semi-variance and expected value of pk,j
by Vk,j and Ek,j , respectively, j = 1, . . . , Nk , k = 1, . . . , K. Without loss
of generality, assume V1 ≤ V2 ≤ · · · ≤ VN ′ . Define the isolation degree of

18
 Journal Pre-proof

individual pk,j by

of
 ( )
 1 Ek,2 − Ek,1 Vk,2 − Vk,1

 1 + + , j=1

 2 Ek,Nk − Ek,1 Vk,Nk − Vk,1

 ( )
 1 Ek,j+1 − Ek,j−1 Vk,j+1 − Vk,j−1
Dk,j = + , j = 2, . . . , Nk − 1

 2 Ek,Nk − Ek,1 Vk,Nk − Vk,1

pro

 ( )

 1 Ek,Nk − Ek,Nk −1 Vk,Nk − Vk,Nk −1

1 + + , j = Nk
2 Ek,Nk − Ek,1 Vk,Nk − Vk,1
(19)

To sum up, define the fitness of individual pk,j by

Dk,j
F (pk,j ) = K − k + (20)
M

re-
where M is a sufficient large positive number. It can be seen that in a series
of individuals, the higher the dominance class of an individual, the higher its
fitness; while in the same dominance class, the higher the isolation degree of an
individual, the higher its fitness.
lP
4.4.3. Crossover
Crossover operation produces a pair of offspring based on two parents, which
are selected from the population using the tournament selection method. That
is, we randomly select five individuals from the current population and take
rna

the best two of them as the parents. Denote the better patent by p1 =
(p1,1 , . . . , p1,2n ) and the other by p2 = (p2,1 , . . . , p2,2n ). Then, we structure
two offspring as follows:

p′1 = p2 + rand1 · (p1 − p2 ) (21)

p′2 = p1 + rand2 · (b − p1 ) (22)

Jou

where rand1 , rand2 are randomly generated on interval [0, 1] and b = (b1 , . . . , b2n )
with bi = I(p1,i ≥ p2,i ), i = 1, . . . , 2n.

4.4.4. Mutation
Mutation operation produces one offspring based on a parent, which is se-
lected from the population using the tournament selection method. That is, we

19
 Journal Pre-proof

randomly select five individuals from the current population and take the best

of
one as the parent. Denote the parent by p = (p1 , . . . , p2n ). Then, we structure
one offspring as follows:

 p′i = pi + rand1 · (I(rand2 ≥ 0.5) − pi ), i = l

pro
(23)

p′i = pi , i ̸= l

where rand1 , rand2 are randomly generated on interval [0, 1] and l is randomly
generated on {1, . . . , 2n}.

4.4.5. Main procedure

re-
We summarize the main procedure of the proposed MOGA as follows:

Step 1 Set parameters: population size Np and maximum generation gmax .

Step 2 Let g = 1, and randomly generate an initial population involving Np
individuals in the search space.
Step 3 Calculate the fitness of the individuals of the current population. If
lP
g > gmax , terminate the procedure and report the dominant class 1 of
the population as the target efficient frontier; otherwise, go to the next
step.
Step 4 Repeat the crossover operation Np /2 times to obtain Np new individu-
rna

als, and denote them by pc1 , . . . , pcNp .

Step 5 Repeat the mutation operation Np times to obtain Np new individuals,
and denote them by pm m
1 , . . . , pNp .

Step 6 Calculate the fitness of the all individuals p1 , . . . , pNp , pc1 , . . . , pcNp ,
pm m
1 , . . . , pNp .

Step 7 Preserve the best Np individuals as the next population. Let g ← g + 1,

Jou

and return to Step 3.

4.5. Fuzzy Value-at-Risk ratio

The last stage of portfolio decision-making is to choose one from the efficient
solutions for investing. We select the investment strategy with the greatest fuzzy
VaR ratio from all the efficient solutions obtained by the MOGA algorithm.

20
 Journal Pre-proof

Define the loss of the investment by L = −(Rt (xt ; st ) − 1). The VaR of the

of
investment under confidence level θ (θ > 0.5) is

V aRθ [L] = sup{x : Cr{L ≤ x} ≤ θ}

= sup{x : Cr{−(Rt (xt ; st ) − 1) ≤ x} ≤ θ} (24)

pro
1
= 1 − bt − βt (1 − 2γ) κt

The fuzzy VaR ratio of the investment is

E[Rt (xt ; st )] − rf
V R[Rt (xt ; st )] = (25)
V aRθ [L]

4.6. Main steps of the automatic trading system

re-
We apply the automatic trading system to solve the (T − w)-period portfolio
optimization problem with sell orders. Denote the initial investment portfolio
of period t by x0t = (x0t,1 , . . . , x0t,n ), t = w + 1, . . . , T . Naturally, we have x0t,i ≥ 0
∑n ∑n
and i=1 x0t,i ≤ 1 (the proportion (1 − i=1 x0t,i ) of current wealth is held in
lP
cash). Assume that x0w+1 is given by the investor according to the distribution
of his/her initial wealth. Then, for each period t, starting from t = w + 1, we
receive the return data of period t at the end of this period, and update the
initial investment proportion of asset i for next period by
c
(1 − χt,i )xt,i rt,i
rna

x0t+1,i = , t = w + 1, . . . , T (26)
rt

where χt,i is a binary variable indicating whether the sell order of asset i is
executed during period t and rt is the net return of the investment strategy
′
(xt ; st ) during period t. The definition of χt,i and rt are similar to that of χtt,i
′
and rtt (see Eqs. (9)-(12)).
To sum up, we conclude the main steps of the trading system for the portfolio
Jou

optimization problem as follows:

Step 1 Initialize t = w + 1 and give the initial portfolio vector x0w+1 .

Step 2 Take the recent previous w period data as input of the WFF method and
formulate the mean-semi-variance portfolio optimization model with sell
orders.

21
 Journal Pre-proof

Step 3 Using the MOGA solve the model to obtain its approximate efficient

of
frontier with a series of efficient solutions.
Step 4 Calculate the fuzzy VaR ratio of all the above efficient solutions and
choose the largest one as the investment strategy for period t.

pro
Step 5 Let t ← t+1, and update the initial investment proportion x0t,i according
to Eq. (26).
Step 6 If t > T , terminate the procedure; otherwise, return to Step 2.

5. Real case studies

re-
In this section, we conduct two real case studies to illustrate the effective-
ness and practicability of the proposed portfolio optimization strategy. In case
study 1, we exemplify the performance of the proposed MOGA algorithm and
automatic trading system in detail. In case study 2, we briefly illustrate the
effectiveness of the proposed automatic trading system involving a number of
lP
risky assets.

5.1. Case study 1

In this case study, we consider a portfolio problem with the stocks selected
from the New York Stock Exchange (NYSE). We randomly select twenty-seven
rna

stocks from the Dow Jones Industrial Average (DJIA) index and divide them
into three datasets, and list their codes in Table 1. We collect their daily
closing, highest and lowest returns with 471 observation days from January
2018 to December 2019. Set the length of the estimation window as w = 350.
Assume that the investor joins the market at the beginning of day t = 351
and he/she plans to invest for 121 days. Assume that the initial portfolio of
Jou

the first investment day is even investment, i.e., x0351 = (1/9, . . . , 1/9), and the
transaction cost rate of asset i is ci = 0.0015, i = 1, . . . , 9.

5.1.1. Performance of the MOGA

To illustrate the effectiveness of the MOGA, we show its realizations in
solving model (15) on the three datasets at the first investment day t = 351.

22
 Journal Pre-proof

Table 1: The codes of the twenty-seven stocks in case study 1.

of
DS 1 DS 2 DS 3
No. Code No. Code No. Code

pro
1 MCD.N 1 IBM.N 1 TRV.N
2 MMM.N 2 DIS.N 2 VZ.N
3 MRK.N 3 JNJ.N 3 HD.N
4 PG.N 4 AXP.N 4 UNH.N
5 WBA.O 5 BA.N 5 MSFT.O
6 WMT.N 6 CAT.N 6 CSCO.O
7
8
9
AAPL.O
INTC.O
NKE.N
re- 7
8
9
CVX.N
JPM.N
KO.N
7
8
9
GS.N
V.N
AMGN.O

Set the population size and maximum generation of the MOGA as Np = 50

lP
and gmax = 3000, respectively. We take the first 350 historical daily data as
input of the WFF method and show the risk-return pairs of the initial and final
population in the mean-semi-variance space in Figure 4. From Figure 4, we
can see that the MOGA can obtain a series of diversified efficient points in the
rna

final population on the three datasets. Moreover, we compare its performance

with that of the other two intelligent algorithms, the genetic algorithm (GA)
proposed by Bermúdez et al. (2012) and the particle swarm optimization (PSO)
proposed by Wang & Watada (2013). We set the parameters of the above two
algorithms according to the original literature. The population size, elitism rate,
mutation rate, and maximum generation of the GA is 2700, 0.2, 0.1, and 300,
Jou

respectively, and the population size, inertia weight, self-learning rate, social
learning rate, and maximum generation of the PSO is 250, 0.2, 2, 2, 3000,
respectively. We show the comparative results of the three algorithms in Figure
5 and their running time in Table 2. From Figure 5 and Table 2, we find that
the MOGA gets the best performance with the least running time among the

23
 Journal Pre-proof

three algorithms, which indicates that the MOGA is an efficient algorithm in

of
solving the proposed model.
10-3 10-3
2 2
Initial population Initial population
Final population 1.5 Final population
1
1

pro
0 0.5
Expected value

Expected value
0
-1
-0.5

-2 -1

-1.5
-3
-2

-4 -2.5
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Semi-variance -5 Semi-variance -5
10 10

(a) DS 1 (b) DS 2

1.5

0.5
re- 10-3
Initial population
Final population
Expected value

-0.5

-1
lP
-1.5

-2
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Semi-variance -5
10

(c) DS 3

Figure 4: The risk-return pairs of the initial and final populations in case study
rna

1 (t = 351).

Table 2: The running time of the three algorithms in case study 1 (t = 351).

DS 1 DS 2 DS 3
MOGA GA PSO MOGA GA PSO MOGA GA PSO
208s 285s 312s 213s 284s 318s 211s 286s 307s
Jou

5.1.2. Performance of the automatic trading system

To illustrate the effectiveness and practicability of our proposed trading sys-
tem, we compare its out-of-sample performance with that of the DJIA and the
following five portfolio selection models:

24
 Journal Pre-proof

10-3 10-3
2 2

of
1.5
1
1

0 0.5
Expected value

Expected value
0
-1
-0.5

-2 MOGA -1

pro
GA
PSO
-1.5
-3 MOGA
-2 GA
PSO
-4 -2.5
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Semi-variance -5 Semi-variance -5
10 10

(a) DS 1 (b) DS 2

10-3
1.5

0.5

re-
Expected value

-0.5

-1
MOGA
GA
-1.5 PSO

-2

-2.5
0 0.5 1 1.5 2 2.5 3 3.5
Semi-variance -5
10

(c) DS 3
lP
Figure 5: The efficient frontiers obtained by the three algorithms in case study
1 (t = 351).

(1) The 1/N model, which allocates a fraction 1/N of the available wealth to
rna

each of the selected stocks at the beginning of each day.

(2) The mean-variance model proposed by Markowitz (1952), in which the min-
imum expected return of each day is set as 1.0005.
(3) The fuzzy mean-variance model proposed by Carlsson et al. (2000), in which
the risk aversion parameter is set as 2.46.
(4) The credibility mean-absolution semi-deviation model proposed by Vercher & Bermúdez
Jou

(2015), in which the minimum skewness is set as 0, and the lower and upper
bounds on the investment proportions are set as 0.01 and 0.4, respectively.
(5) Our proposed automatic trading system without sell orders.

For each day t, starting from t = 351, we adopt the most recent previous
350 daily data as inputs of the model to determine the investment strategy.

25
 Journal Pre-proof

Next, the actual returns of the investment strategies are computed based on the

of
data of day t. Repeating the above process until the end of the investment, we
obtain 121 out-of-sample daily returns for each investment strategy. We show
the cumulative daily returns of the strategies in Figure 6 and list some statistics

pro
of their out-of-sample daily returns in Table 3, including the maximum (MAX),
minimum (MIN), average (AVE), standard deviation (SD), Sharp ratio (SR),
total transaction costs (TTC), and final wealth (FW).
From Figure 6, we find that the cumulative returns of our proposed strategy
with sell orders are higher than that of other strategies most of the time on the
three datasets. From Table 3, we conclude three findings as follows. Firstly,

re-
the final wealth of our proposed strategy with sell orders is the highest among
the strategies, and its total transaction costs are much higher than the other
strategies. It can be seen that our proposed trading system with sell orders is
an effective high-frequency trading strategy, and it obtains a higher return than
the other models despite enormous transaction costs. Secondly, the maximum
lP
return of the proposed strategy with sell orders is lower than that of at least one
other strategy on datasets 2 and 3, and the minimum return of the proposed
strategy is the highest on the three datasets. It can be seen that due to selling
the risky asset in advance, introducing sell orders into the investment process
rna

may miss some high returns while achieves a significant effect in avoiding huge
losses. Thirdly, the average return and Sharp ratio of the proposed strategy is
higher than that of other strategies while its standard deviation is lower than
that of other strategies on all datasets, which indicates that the proposed s-
trategy with sell orders obtains a higher return and lower risk than the other
strategies. In addition, focusing on the two most direct performance indicators,
Jou

the Sharp ratio (SR) and final wealth (FW), we find that our proposed trading
system without sell orders performs second only to that with it, i.e., its per-
formance is not as good as our trading system with sell orders but better than
that of the DJIA and other strategies. To sum up, we can conclude that the
proposed trading system is practical to meet the portfolio management problem
and considering the sell orders can improve its performance effectively.

26
 Journal Pre-proof

1.3 1.14

of
DJIA 1.12 DJIA
1.25 1/N 1/N
Mar. 1.1 Mar.
Car. Car.
1.2 Ver. 1.08 Ver.
Cumulative return

Cumulative return
Our (without) Our (without)
Our (with) 1.06 Our (with)
1.15
1.04
1.1
1.02

pro
1.05
0.98
1
0.96

0.95 0.94
350 375 400 425 450 475 350 375 400 425 450 475
Period Period

(a) DS 1 (b) DS 2

1.2

DJIA
1/N
1.15 Mar.
Car.
Ver.
Cumulative return

re-
1.1

1.05

1
Our (without)
Our (with)

0.95
350 375 400 425 450 475
Period

(c) DS 3
lP
Figure 6: The out-of-sample performances of the investment strategies in case
study 1.

5.2. Case study 2

rna

To illustrate the effectiveness of the automatic trading system for large-scale

cases, we conduct another case study with a number of risky assets in the Hong
Kong Stock Exchange (HKSE). We select forty stocks from the Hang Seng Index
(HSI), whose codes are listed in Table 4. Collect their daily historical data with
473 observation days from January 2018 to December 2019. We also set the
Jou

length of the estimation window as w = 350. Assume that the investor joins
the market at the beginning of day t = 351, and he/she plans to invest for
123 days. Assume that the initial portfolio of the first investment day is even
investment, i.e., x0351 = (1/40, . . . , 1/40), and the transaction cost rate of asset
i is ci = 0.0015, i = 1, . . . , 40.
Similar to case study 1, we compare the out-of-sample performance of the

27
 Journal Pre-proof

Table 3: The statistics of the out-of-sample returns in case study 1.

of
MAX MIN AEV SD SR TTC FW
DJIA 1.0148 0.9695 1.0006 0.0078 0.0735 − 1.0682

pro
1/N 1.0194 0.9690 1.0012 0.0082 0.1428 0.0014 1.1466
Mar. 1.0159 0.9753 1.0005 0.0076 0.0668 0.0044 1.0599
Car. 1.0172 0.9648 1.0008 0.0092 0.0877 0.0178 1.0974
DS 1
Ver. 1.0172 0.9711 1.0012 0.0080 0.1449 0.0047 1.1452
Our 1.0156 0.9739 1.0012 0.0075 0.1572 0.0231 1.1485
(without)

Our 1.0228 0.9948 1.0021 0.0065 0.3186 0.2475 1.2827

(with)

DS 2
1/N
Mar.
Car.
Ver.
re-
1.0177 0.9681 1.0004 0.0082 0.0528 0.0011 1.0496
1.0313 0.9768 1.0006 0.0071 0.0843 0.0031 1.0723
1.0282 0.9797 1.0003 0.0074 0.0393 0.0157 1.0326
1.0195 0.9723 1.0005 0.0074 0.0665 0.0053 1.0581
Our 1.0347 0.9781 1.0007 0.0073 0.0946 0.0036 1.0836
(without)
lP
Our 1.0269 0.9968 1.0009 0.0059 0.1552 0.2262 1.1150
(with)

1/N 1.0162 0.9701 1.0008 0.0078 0.1011 0.0010 1.0958

Mar. 1.0149 0.9752 1.0007 0.0071 0.0915 0.0033 1.0786
Car. 1.0250 0.9603 1.0010 0.0088 0.1120 0.0093 1.1207
DS 3
Ver. 1.0169 0.9719 1.0005 0.0074 0.0699 0.0062 1.0609
rna

Our 1.0167 0.9746 1.0010 0.0073 0.1402 0.0152 1.1281

(without)

Our 1.0212 0.9970 1.0012 0.0051 0.2297 0.2175 1.1503

(with)

proposed strategy with that of the HSI and other five models. We show the
cumulative daily returns of the five strategies in Figure 5 and list the statistics
of their out-of-sample daily returns in Table 5. As shown in Figure 7, the
Jou

cumulative returns of our proposed strategy with sell order is much higher than
that of the other strategies most of the time. It can be seen that the other
strategies perform poorly due to the downturn of the market, while our proposed
strategy with sell orders still obtains a good performance. Moreover, we can
see from Table 5 that the maximum return, minimum return, average return,

28
 Journal Pre-proof

Table 4: The codes of the forty stocks in case study 2.

of
No. 1 2 3 4 5 6 7 8
Code 0002.HK 0003.HK 0669.HK 0005.HK 0006.HK 0027.HK 0883.HK 0386.HK
No. 9 10 11 12 13 14 15 16

pro
Code 1093.HK 0762.HK 2382.HK 1928.HK 0017.HK 1109.HK 0700.HK 2018.HK
No. 17 18 19 20 21 22 23 24
Code 1044.HK 0688.HK 3328.HK 2318.HK 2628.HK 1398.HK 2388.HK 0823.HK
No. 25 26 27 28 29 30 31 32
Code 2313.HK 1038.HK 1299.HK 1177.HK 0857.HK 0011.HK 0101.HK 0012.HK
No. 33 34 35 36 37 38 39 40
Code 0066.HK 2319.HK 2020.HK 0175.HK 0939.HK 3988.HK 0941.HK 2007.HK

re-
Sharp ratio, and final wealth of the proposed strategy with sell orders are higher
than the other strategies. In addition, the Sharp ratio and final wealth of our
proposed trading system without sell orders are the second only to that with it,
i.e., its performance is not as good as our trading system with sell orders but
lP
better than that of the HSI and other strategies. We conclude that the proposed
automatic trading system has a good performance for the large-scale case.
1.4
DJIA
1/N
1.3 Mar.
rna

Car.
Ver.
Our (without)
1.2
Cumulative return

Our (with)

1.1

0.9
Jou

0.8
350 375 400 425 450 475
Period

Figure 7: The out-of-sample performances of the investment strategies in case

study 2.

29
 Journal Pre-proof

Table 5: The statistics of the out-of-sample returns in case study 2.

of
MAX MIN AEV SD SR TTC FW
HSI 1.0390 0.9715 0.9999 0.0102 -0.0049 − 0.9876

pro
1/N 1.0361 0.9718 1.0004 0.0104 0.0392 0.0016 1.0444
Mar. 1.0347 0.9781 0.9994 0.0086 -0.0733 0.0133 0.9209
Car. 1.0270 0.9552 1.0003 0.0122 0.0272 0.0246 1.0323
Ver. 1.0377 0.9795 0.9991 0.0089 -0.0970 0.0345 0.8945
Our 1.0377 0.9806 1.0008 0.0088 0.0868 0.0568 1.0930
(without)

Our 1.0501 0.9954 1.0026 0.0096 0.2758 0.2851 1.3761

(with)

6. Discussion
re-
In this paper, we initially considered the sell orders in the fuzzy portfolio op-
timization problem and proposed an automatic trading system with sell orders,
in which the investment strategy, consisting of the portfolio and sell thresholds
lP
of the risky assets, is generated automatically at the beginning of each period.
The proposed trading system provides an intelligent procedure of dynamically
reallocating asset positions and finding the best sell moments, which enriches
the theoretical implications for the modern portfolio theory. Furthermore, we
rna

shed light on the practical implications of our study as the following two aspects.
Firstly, we provided a decision support system with fully automatic operation
for investors in the financial management problem. With the inputs of trading
price data, which can be obtained easily from the financial market, the trading
system automatically generates the investment strategy using the optimality
theory and soft computing technology. In the experiments using actual stock
Jou

data in the NYSE and HKSE, the proposed trading system gets better out-of-
sample performance than some existing ones. The above discussion indicates
that the proposed trading system is easy to implement and is practical in the
real financial management problem, which can effectively guide investors to meet
the portfolio problem with sell orders.

30
 Journal Pre-proof

Secondly, we revealed the importance of sell orders and proposed an in-

of
telligent framework to determine the appropriate sell thresholds for the risky
assets. The traditional portfolio selection model assumes that investors hold
the portfolio until the end of the investment period. Under this assumption,

pro
the investment return is only related to the relative price of the assets in the
period but has nothing to do with the price paths of the assets. We realized
that introducing the sell order into the portfolio optimization can help investors
capture more investment opportunities in the investment process. Appropriate
sell orders can effectively find the sell moments for the risky assets in the rapid-
ly changing financial market. We embedded sell orders into the fuzzy portfolio

re-
optimization decision-making process and provided a realizable way to pursue
high return and low risk.

7. Conclusions
lP
This paper studied a sequential portfolio optimization problem with sell
orders in fuzzy environment, in which the investor needs to determine an in-
vestment proportion and a sell threshold for each risky asset. We proposed an
automatic trading system to determine the investment strategy at the begin-
ning of each period. Firstly, taking the historical data as inputs, we designed
rna

the WFF method to formulate the fuzzy return of an investment strategy as an

LR-power fuzzy number. Then, we proposed a fuzzy mean-semi-variance port-
folio optimization problem with sell orders. Next, we designed a multi-objective
genetic algorithm to solve the proposed model and select one from the efficien-
t solutions based on the fuzzy Value-at-Risk ratio. Finally, we discussed two
real case studies to illustrate the effectiveness of the proposed trading system.
Jou

The results showed that the proposed automatic trading system has a better
out-of-sample performance than some other ones without sell orders.
There are some interesting extensions of our study worth to be studied.
Firstly, it is a meaningful topic to further investigate the implementation and
application of the proposed trading system in some emerging markets such as fu-

31