An Interest Rate Decision Method for Risk-averse Portfolio Optimization
using Loan
Kiyoharu Tagawa
School of Science and Engineering, Kindai University, 3-4-1 Kowakae, Higashi-Osaka, 577-8502, Japan
Keywords:
Risk Analysis and Management, Mathematical Model of Loan, Portfolio Optimization.
Abstract:
Portfolio optimization using loan is formulated as a chance constrained problem in which the borrowing money
from loan can be invested in risk assets. The chance constrained problem is proven to a convex optimization
problem. The low interest rate of loan benefits borrowers. On the other hand, the high interest rate of loan
doesn’t benefits lenders because such a loan is not often used. For deciding a proper interest rate of loan that
benefits both borrowers and lenders, a new method is proposed. Experimental results show that the loan is
used completely to improve the efficient frontier if the interest rate is decided by the proposed method.
1 INTRODUCTION
Portfolio optimization is the process of determining
the best proportion of investment in different assets
according to some objective. The objective typically
maximizes factors such as expected return, and mini-
mizes costs like financial risk. Portfolio optimization
is one of the most challenging problems in the field
of finance. Therefore, a large number of works about
portfolio optimization have been reported (Mokhtar
et al., 2014; Mansini et al., 2014). In these works,
portfolio optimization has been discussed both in a
deterministic and in a stochastic domain, either in a
single period or in a multi-period framework.
In our previous work (Tagawa, 2019), portfolio
optimization using bank deposit and loan has been
formulated as a chance constrained problem in which
a non-risk asset called bank deposit is included in a
portfolio and the borrowing money from loan can be
invested in risk assets. It has been also proven that
the chance constrained problem is a multimodal opti-
mization problem having multiple optimal solutions.
Therefore, for solving the optimization problem, an
optimization method based on Differential Evolution
(DE) (Price et al., 2005) has been proposed.
The effect of the loan on portfolio optimization
has been also studied independently (Tagawa, 2020).
Portfolio optimization using only loan has been for-
mulated as a chance constrained problem. It has been
also proven that the chance constrained problem is
a convex optimization problem. Therefore, for solv-
ing the convex optimization problem, an interior point
method (Horst and Pardalos, 1995) has been used.
Experimental results show that the efficient frontier
is improved if the loan is used. Consequently, the low
interest rate of loan benefits borrowers. On the other
hand, the high interest rate of loan does not benefits
lenders because such a loan is not often used.
In this paper, portfolio optimization using loan is
studied more intensively. Specifically, a proper in-
terest rate of loan that benefits both borrowers and
lenders is considered. Then, a new method to decide
a proper interest rate of loan from an acceptable risk
is proposed. Experimental results show that the loan
is used completely to improve the efficient frontier if
the interest rate is decided by the proposed method.
The remainder of this paper is organized as fol-
lows. Section 2 explains conventional models for
portfolio optimization. Section 3 formulates a new
portfolio optimization problem using loan. Section 4
proposes a new method to decide a proper interest rate
of loan. Section 5 shows the results of numerical ex-
periments and discusses about them. Finally, Section
6 concludes this paper and mentions future work.
2 RELATED WORK
2.1 Definition of Portfolio
Let x
i
, i = 1, ··· , n be the proportion of i-asset
normalized by owned capital invested in n assets. A
portfolio is defined as x
x
x = (x
1
, ··· , x
n
)
n
. Since
Tagawa, K.
An Interest Rate Decision Method for Risk-averse Portfolio Optimization using Loan.
DOI: 10.5220/0009208400150024
In Proceedings of the 5th International Conference on Complexity, Future Information Systems and Risk (COMPLEXIS 2020), pages 15-24
ISBN: 978-989-758-427-5
Copyright
c
2020 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
15
we consider a long-only portfolio in a single period,
the portfolio x
x
x
n
is constrained as
x
1
+ x
2
+ ···+ x
n
= 1 (1)
where 0 x
i
, i = 1, ··· , n.
The unit investment in the i-asset provides return
ξ
i
over a single period operation. Each of asset
returns ξ
i
, i = 1, ··· , n is modeled by a random
variable following a normal distribution as
ξ
i
Normal(µ
i
, σ
2
i
). (2)
Incidentally, it is known that Normal distribution
can represent a fairly accurate model of asset returns
for portfolio optimization (Ruppert, 2011).
Let ρ
i j
be the correlation coefficient between ξ
i
and ξ
j
, i ̸= j. As well as the mean µ
i
and the standard
deviation σ
i
in (2), ρ
i j
is estimated statistically from
historical data (Rubio et al., 2012). In recent years,
an Artificial Intelligence (AI) method based on deep
learning is also reported to predict the future returns
of assets from market data (Obeidat et al., 2018).
The vector ξ
ξ
ξ = (ξ
1
, ··· , ξ
n
) of random returns in
(2) follows a multivariable normal distribution as
ξ
ξ
ξ Normal(µ
µ
µ, C
C
C) (3)
where the mean is given as µ
µ
µ = (µ
1
, ··· , µ
n
)
n
.
In order to derive the covariance matrix C
C
C in (3), a
matrix D
D
D is defined by using σ
i
in (2) as
D
D
D =
σ
1
0 ··· 0
0 σ
2
··· 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 ··· σ
n
. (4)
From the correlation coefficient ρ
i j
between ξ
i
and
ξ
j
, a coefficient matrix R
R
R is also defined as
R
R
R =
1 ρ
12
··· ρ
1n
ρ
21
1 ··· ρ
2n
.
.
.
.
.
.
.
.
.
.
.
.
ρ
n1
ρ
n2
··· 1
. (5)
From D
D
D in (4) and R
R
R in (5), C
C
C is obtained as
C
C
C = D
D
DR
R
RD
D
D. (6)
The return of a portfolio x
x
x
n
is defined as
r(x
x
x, ξ
ξ
ξ) =
n
i=1
ξ
i
x
i
= ξ
ξ
ξx
x
x
T
. (7)
According to the reproductive property of normal
distribution (Ash, 2008), the return in (7) also follows
a normal distribution as
r(x
x
x, ξ
ξ
ξ) Normal(µ
r
(x
x
x), σ
2
(x
x
x)) (8)
where the mean and the variance are given as
µ
r
(x
x
x) =
n
i=1
µ
i
x
i
= µ
µ
µx
x
x
T
(9)
σ
2
(x
x
x) = x
x
xC
C
C x
x
x
T
. (10)
2.2 Portfolio Optimization
By using the portfolio stated above, we explain basic
models used to formulate portfolio optimization.
2.2.1 Markowitz’s Model
In Markowitz’s model (Markowitz, 1952), the risk of
a portfolio x
x
x
n
is evaluated by the variance σ
2
(x
x
x)
shown in (10). Then, the risk is minimized keeping
an expected return µ
r
(x
x
x) larger than γ as
min σ
2
(x
x
x) = x
x
xC
C
C x
x
x
T
sub. to µ
r
(x
x
x) = µ
µ
µx
x
x
T
γ,
x
1
+ x
2
+ ···+ x
n
= 1,
0 x
i
, i = 1, ··· , n.
(11)
2.2.2 Roy’s Model
In Roy’s model (Roy, 1952), the risk of portfolio is
evaluated by the probability that the return r(x
x
x, ξ
ξ
ξ) in
(7) falls bellow a desired value γ . In order to
minimize the risk α (0, 1), portfolio optimization is
formulated as a chance constrained problem:
min α
sub. to Pr(r(x
x
x, ξ
ξ
ξ) γ) α,
x
1
+ x
2
+ ···+ x
n
= 1,
0 x
i
, i = 1, ··· , n.
(12)
where Pr(A ) is the probability that event A occurs.
2.2.3 Kataoka’s Model
Contrary to Roy’s model in (12), Kataoka’s model
(Kataoka, 1963) maximizes the desired value γ
of the return r(x
x
x, ξ
ξ
ξ) for an acceptable risk α (0, 0.5)
given by a probability. Then, portfolio optimization is
also formulated as a chance constrained problem:
max γ
sub. to Pr(r(x
x
x, ξ
ξ
ξ) γ) α,
x
1
+ x
2
+ ···+ x
n
= 1,
0 x
i
, i = 1, ··· , n.
(13)
2.3 Extended Models
There is a trade-off relationship between return and
risk. Therefore, by using a risk aversion indicator λ
[0, 1], Efficient Frontier model (Chang et al., 2000)
modifies Markowitz’s model defined in (11) as
min λσ
2
(x
x
x) (1 λ)µ
r
(x
x
x)
sub. to x
1
+ x
2
+ ···+ x
n
= 1,
0
x
i
,
i
=
1
,
···
,
n
.
(14)
COMPLEXIS 2020 - 5th International Conference on Complexity, Future Information Systems and Risk
16
By changing the value of λ [0, 1] in (14), we can
obtain the efficient frontier. The efficient frontier is a
continuous curve illustrating the trade-off between the
expected return (mean) and the risk (variance).
Genetic Algorithm (GA), Tabu Search (TS), and
Simulated Annealing (SA) have been applied to a
portfolio optimization problem based on the efficient
frontier model (Chang et al., 2000). Artificial Bee
Colony (ABC) algorithm has been also proposed for
solving a portfolio optimization problem based on the
efficient frontier model (Strumberger et al., 2018).
Portfolio optimization can be also formulated as a
multi-objective optimization problem as
min σ
2
(x
x
x) = x
x
xC
C
C x
x
x
T
max µ
r
(x
x
x) = µ
µ
µx
x
x
T
sub. to x
1
+ x
2
+ ···+ x
n
= 1,
0 x
i
, i = 1, ··· , n.
(15)
In order to obtain the efficient frontier for multi-
objective portfolio optimization problems, several
Multi-Objective Evolutionary Algorithms (MOEAs)
have been used successfully (Anagnostopoulos and
Mamanis, 2010; Ponsich et al., 2013).
Cardinality constraints restrict a portfolio to have
a specified number of assets. Specifically, several
assets to be invested in have to be selected from a
list of many assets. Thus, portfolio optimization in-
cluding cardinality constraints is usually formulated
as a mixed integer problem (Konno and Yamamoto,
2005). Furthermore, the number of assets has been
minimized by a multi-objective portfolio optimization
problem (Anagnostopoulos and Mamanis, 2010).
Portfolio optimization is often formulated based
on multiple periods. In order to evaluate the total
return over multiple periods, a risk function called
Mean Absolute Deviation (MAD) has been proposed
(Konno and Yamazaki, 1995). Conditional Value-at-
Risk (CVaR) has been also used to formulate portfolio
optimization considering the return expected through
multiple periods (Angelelli et al., 2008).
In the multi-period framework, transaction costs
have to be paid for any assets. Therefore, a portfolio
optimization problem considering the costs for selling
and buying assets to change the structure of portfolio
between periods has been formulated and solved by
using an extended GA (Aranha and Iba, 2007).
Currently, portfolio optimization is extended in
various ways. For example, Goal Programming (GP)
models have been proposed to compose a portfolio
of international mutual funds (Tamiz et al., 2013). A
deep learning network has been used to predict the
composite index of stock market (Pang et al., 2018).
The latest technology of AI has been also introduced
into portfolio optimization (Obeidat et al., 2018).
3 PROBLEM FORMULATION
Portfolio Optimization Problem using Loan (POPL)
is an extended version of Kataoka’s Model in (13).
The loan can be introduced into any models shown in
(11) to (13). Actually, Markowitz’s Model in (11) is
the most popular one. However, by using Kataoka’s
Model, we can decide the interest rate of loan from an
acceptable risk α (0, 0.5) given in advance.
3.1 Portfolio Including Loan
The borrowing money from loan is invested in risk
assets. Let x
0
be the proportion of loan used for
a portfolio x
x
x
n
. Let M , M > 0 be the upper
limit of the loan, which is specified by a multiple of
owned capital. If the loan is not used, the proportion
of loan is x
0
= 0. On the other hand, if the loan is used
up to the limit, the proportion of loan is x
0
= M.
Therefore, the constraints of POPL are
x
0
+ x
1
+ x
2
+ ···+ x
n
= 1,
M x
0
0, 0 x
i
, i = 1, ··· , n.
(16)
From the first constraint in (16), the proportion of
loan x
0
used for a portfolio x
x
x
n
is
x
0
= 1 1lx
x
x
T
(17)
where 1l
n
is a vector defined as 1l = (1, ··· , 1).
Let L be the interest rate of loan. The interest
rate L , L 0 is a constant value. Considering the
proportion of loan x
0
0 and L 0, the return r(x
x
x, ξ
ξ
ξ)
of a portfolio x
x
x
n
defined in (7) is revised as
g(x
x
x, ξ
ξ
ξ) = r(x
x
x, ξ
ξ
ξ) + L x
0
= ξ
ξ
ξx
x
x
T
+ L (1 1lx
x
x
T
)
= (ξ
ξ
ξ L 1l)x
x
x
T
+ L .
(18)
According to the reproductive property of normal
distribution (Ash, 2008), the return of POPL in (18)
also follows a normal distribution as
g(x
x
x, ξ
ξ
ξ) Normal(µ
g
(x
x
x), σ
2
(x
x
x)) (19)
where the mean is given as
µ
g
(x
x
x) = (µ
µ
µ L 1l)x
x
x
T
+ L . (20)
The variance σ
2
(x
x
x) in (19) is given by (10).
3.2 Portfolio Optimization using Loan
As stated above, POPL is formulated as an extended
version of Kataoka’s Model in (13). An acceptable
risk α (0, 0.5) is given in advance. Therefore, from
An Interest Rate Decision Method for Risk-averse Portfolio Optimization using Loan
17
Figure 1: Feasible region of POPL.
(16) and (18), POPL is also formulated as a chance
constrained problem:
max γ
sub. to Pr(g(x
x
x, ξ
ξ
ξ) γ) α,
x
1
+ x
2
+ ···+ x
n
M + 1,
x
1
+
x
2
+
···
+
x
n
1
,
0 x
i
, i = 1, ··· , n
(21)
where the proportion of loan x
0
[M, 0] does not
appear in (21) because it has been eliminated from
the constraints in (16) by using the equation in (17).
The chance constrained problem is usually hard
to solve directly (Pr
´
ekopa, 1995). However, we can
transform the above POPL in (21) into an equivalence
problem. Since the return g(x
x
x, ξ
ξ
ξ) of POPL follows
the normal distribution in (19), we can standardize the
chance constraint of POPL in (21) as
Pr
g(x
x
x, ξ
ξ
ξ) µ
g
(x
x
x)
σ(x
x
x)
γ µ
g
(x
x
x)
σ(x
x
x)
α. (22)
Furthermore, the probability in (22) is written as
Φ
γ µ
g
(x
x
x)
σ(x
x
x)
α (23)
where Φ : [0, 1] is the Cumulative Distribution
Function (CDF) of the standard normal distribution.
From (23), we can derive the equivalence problem
of the chance constrained problem in (21) as
max γ(x
x
x) = µ
g
(x
x
x) + Φ
1
(α)σ(x
x
x)
sub. to x
1
+ x
2
+ ···+ x
n
M + 1,
x
1
+ x
2
+ ···+ x
n
1,
0 x
i
, i = 1, ··· , n.
(24)
The equivalence problem in (24) is a deterministic
one. Thus, we don’t need to evaluate the probability
that appears in (21). The deterministic optimization
problem in (24) is also called POPL in this paper.
Figure 1 illustrates the feasible region of POPL for
the case of n = 2. The feasible region is denoted by
the gray area between two hyper-planes. If a portfolio
x
x
x
n
doesn’t use the loan (x
0
= 0), it exists on the
lower plane: 1lx
x
x
T
= 1. On the other hand, if a port-
folio x
x
x
n
uses the loan up to the limit (x
0
= M),
it exists on the upper plane: 1lx
x
x
T
= M + 1.
3.3 Solution of Problem
We consider the solution of POPL in (24).
Lemma 1. The standard deviation σ(x
x
x) defined by
(10) is convex (Tagawa, 2019).
Proof. Since the covariance matrix C
C
C in (6) is positive
semi-definite, it can be decomposed as
σ(x
x
x) =
x
x
xC
C
C x
x
x
T
=
x
x
x A
A
AA
A
A
T
x
x
x
T
=
y
y
yy
y
y
T
(25)
where C
C
C = A
A
AA
A
A
T
and y
y
y = x
x
x A
A
A
n
.
From (25), σ(x
x
x) is a norm. The norm meets the
triangle inequality for any θ [0, 1] as
σ(θx
x
x + (1 θ)
ˆ
x
x
x) σ(θ x
x
x) + σ((1 θ)
ˆ
x
x
x). (26)
The right side of (26) can be transformed as
σ(θx
x
x) =
θy
y
y(θ y
y
y)
T
= θ
y
y
yy
y
y
T
= θσ(x
x
x). (27)
From (26) and (27), we have
σ(θx
x
x + (1 θ)
ˆ
x
x
x) θ σ(x
x
x) + (1 θ)σ(
ˆ
x
x
x). (28)
From (28), σ(x
x
x) in (10) is a convex function.
Theorem 1. The objective function γ(x
x
x) of POPL in
(24) is concave. In other words, γ(x
x
x) is convex.
Proof. From (20) and γ(x
x
x) in (24), we have
θγ(x
x
x) + (1 θ)γ(
ˆ
x
x
x) γ(θ x
x
x + (1 θ)
ˆ
x
x
x)
= Φ
1
(α)×
(θσ(x
x
x) + (1 θ)σ(
ˆ
x
x
x) σ(θ x
x
x + (1 θ)
ˆ
x
x
x)).
(29)
From Lemma 1 and Φ
1
(α) < 0 for α (0, 0.5),
the right side of (29) is negative. Hence, we have
γ(θx
x
x + (1 θ)
ˆ
x
x
x) θ γ(x
x
x) + (1 θ)γ(
ˆ
x
x
x). (30)
From (30), γ(x
x
x) in (24) is a concave function.
Since all constraints of POPL in (24) are linear,
the feasible region of POPL is convex. Furthermore,
from Theorem 1, POPL is a convex optimization
problem. If x
x
x
n
is a local optimal solution of a
convex optimization problem, the solution x
x
x
n
is
guaranteed to be a global optimal one of the convex
optimization problem (McCormick, 1983).
The gradient of γ(x
x
x) in (24) can be derived as
∇γ(x
x
x) = (µ
µ
µ L 1l)+ Φ
1
(α)
x
x
xC
C
C
x
x
xC
C
C x
x
x
T
. (31)
The global optimal solution x
x
x
n
of POPL in
(24) satisfies either of the following two conditions.
∇γ(x
x
x
) = 0
0
0 holds.
Some constraints in (24) are active with x
x
x
n
.
COMPLEXIS 2020 - 5th International Conference on Complexity, Future Information Systems and Risk
18
4 INTEREST RATE OF LOAN
4.1 Proper Interest Rate
We think about a proper interest rate of loan L
that benefits borrowers to get much return.
Let x
x
x
n
be a portfolio of POPL in which the
loan is not used as x
0
= 0. Therefore, 1lx
x
x
T
= 1 holds.
For x
x
x
n
, the objective function in (24) is
γ(x
x
x) = µ
g
(x
x
x) + Φ
1
(α)σ(x
x
x)
= (µ
µ
µ L 1l)x
x
x
T
+ L + Φ
1
(α)σ(x
x
x)
= µ
µ
µx
x
x
T
+ L (1 1lx
x
x
T
) + Φ
1
(α)σ(x
x
x)
= µ
µ
µx
x
x
T
+ Φ
1
(α)σ(x
x
x)
= µ
r
(x
x
x) + Φ
1
(α)σ(x
x
x) = γ
0
(x
x
x).
(32)
Theorem 2. Let x
x
x
n
be a solution of POPL in
which the loan is not used as x
0
= 0. The solution can
be improved by borrowing money from the loan if the
interest rate of loan L meets the condition:
γ
0
(x
x
x) > L (33)
where γ
0
(x
x
x) = µ
r
(x
x
x) + Φ
1
(α)σ(x
x
x).
Proof. Let’s consider a new portfolio
ˆ
x
x
x = κx
x
x, κ > 1.
The new portfolio
ˆ
x
x
x
n
borrows money as
ˆx
0
= 1 1l
ˆ
x
x
x
T
= 1 κ 1lx
x
x
T
= 1 κ < 0 (34)
where ˆx
0
is the proportion of loan for
ˆ
x
x
x
n
.
The objective function value of
ˆ
x
x
x
n
is
γ(
ˆ
x
x
x) = (µ
µ
µ L 1l)
ˆ
x
x
x
T
+ L + Φ
1
(α)σ(
ˆ
x
x
x)
= κ(µ
µ
µ L 1l)x
x
x
T
+ L + κΦ
1
(α)σ(x
x
x)
= κγ
0
(x
x
x) + L (1 κ 1lx
x
x
T
).
(35)
From (35) and 1lx
x
x
T
= 1, the difference between
the returns of
ˆ
x
x
x
n
and x
x
x
n
is
γ(
ˆ
x
x
x) γ
0
(x
x
x)
= (κ 1)γ
0
(x
x
x) + L (1 κ 1lx
x
x
T
)
= (κ 1)(γ
0
(x
x
x) L).
(36)
If the condition in (33) is satisfied, we have
γ
(
ˆ
x
x
x) > γ
0
(x
x
x). (37)
Therefore,
ˆ
x
x
x
n
is better than x
x
x
n
.
Theorem 3. Let x
x
x
n
be a solution of POPL that
uses the loan. The portfolio x
x
x
n
borrows money
from the loan up to the limit such as x
0
= M if
γ(x
x
x) > L. (38)
Figure 2: Portfolio
ˆ
x
x
x
2
is better than x
x
x
2
.
Proof. Let’s consider a new portfolio
ˆ
x
x
x = κx
x
x, κ > 1.
The new portfolio
ˆ
x
x
x
n
borrows much money than
the current one x
x
x
n
as
ˆx
0
= 1 1l
ˆ
x
x
x
T
= 1 κ 1lx
x
x
T
< 1 1lx
x
x
T
= x
0
0
(39)
where ˆx
0
is the proportion of loan for the new
portfolio
ˆ
x
x
x
n
, while x
0
is the proportion of
loan for the current portfolio x
x
x
n
.
The objective function value of x
x
x
n
is
γ(x
x
x) = µ
g
(x
x
x) + Φ
1
(α)σ(x
x
x)
= (µ
µ
µ L 1l)x
x
x
T
+ L + Φ
1
(α)σ(x
x
x).
(40)
The objective function value of
ˆ
x
x
x
n
is
γ(
ˆ
x
x
x) = µ
g
(
ˆ
x
x
x) + Φ
1
(α)σ(
ˆ
x
x
x)
= κ(µ
µ
µ L 1l)x
x
x
T
+ L + κΦ
1
(α)σ(x
x
x).
(41)
From (40) and (41), the gap between them is
γ(
ˆ
x
x
x) γ(x
x
x) = (κ 1)(γ(x
x
x) L). (42)
From (42) and κ > 1, if the condition in (38) is
satisfied,
ˆ
x
x
x
n
is better than x
x
x
n
as
γ(
ˆ
x
x
x) > γ(x
x
x). (43)
Consequently, every portfolio x
x
x
n
of POPL
that satisfies the condition in (38) can be improved
proportionally to the amount of debt.
Please notice that if the condition shown in (33)
is satisfied by an interest rate L and a portfolio
x
x
x
n
(x
0
= 0), the condition in (38) is also satisfied
by the interest rate L and a new portfolio
ˆ
x
x
x
n
( ˆx
0
< 0) generated as
ˆ
x
x
x = κ x
x
x, κ > 0. That is because
the relation γ(
ˆ
x
x
x) > γ
0
(x
x
x) > L holds. Besides, from
Theorem 3, the portfolio
ˆ
x
x
x
n
exists on the upper
plane such as 1l
ˆ
x
x
x
T
= M + 1. Figure 2 illustrates the
above x
x
x
n
and
ˆ
x
x
x
n
for the case of n = 2.
An Interest Rate Decision Method for Risk-averse Portfolio Optimization using Loan
19
4.2 Interest Rate Decision Method
The low interest rate of loan benefits borrowers. On
the other hand, the high interest rate doesn’t benefit
lenders because such a loan is not often used. Hence,
we propose a method to decide an interest rate of loan
that benefits both borrowers and lenders.
We formulate a sub-problem of POPL in the case
that the loan is not used. From (32) and 1lx
x
x
T
= 1, the
sub-problem of POPL can be formulated as
max γ
0
(x
x
x) = µ
r
(x
x
x) + Φ
1
(α)σ(x
x
x)
sub. to x
1
+ x
2
+ ···+ x
n
= 1,
0 x
i
, i = 1, ··· , n.
(44)
In the same way with Theorem 1, we can prove
that the sub-problem of POPL in (44) is also a convex
optimization problem. Furthermore, the gradient of
γ
0
(x
x
x) in (44) can be derived as
∇γ
0
(x
x
x) = µ
µ
µ + Φ
1
(α)
x
x
xC
C
C
x
x
xC
C
C x
x
x
T
. (45)
From Theorem 2, the procedure of the proposed
interest rate decision method is stated as follows:
Step 1: Give an acceptable risk α (0, 0, 5).
Step 2:
By solving the sub-problem of POPL in (44)
with the above risk α, get a solution x
x
x
n
.
Step 3: Choose a proper value for the interest rate of
loan L in the range from 0 to γ
0
(x
x
x
).
If γ
0
(x
x
x
) 0 holds in Step 2, we should give up the
investment in assets. That is because POPL doesn’t
have any solutions that generate profits. Otherwise,
we need to increase the acceptable risk α (0, 0.5) in
Step 1. Then, we look for another L again.
From Theorem 3, if the interest rate of loan L
is decided by the proposed method, POPL in (24) can
be written by a rather simple form as
max γ(x
x
x) = µ
g
(x
x
x) + Φ
1
(α)σ(x
x
x)
sub. to x
1
+ x
2
+ ···+ x
n
= M + 1,
0 x
i
, i = 1, ··· , n.
(46)
Please notice that the portfolio
ˆ
x
x
x = κx
x
x
generated
by the optimal solution x
x
x
n
of the sub-problem
of POPL in (44) is not guaranteed to be an optimal
solution of POPL in (46). Therefore, we have to solve
POPL in (46) seriously under the interest rate of loan
L decided by using the proposed method.
5 NUMERICAL EXPERIMENT
For solving convex optimization problems shown in
(24), (44), and (46), an interior point method provided
Table 1: Mean and variance of asset return by port0.
ξ
i
ξ
1
ξ
2
ξ
3
ξ
4
µ
i
0.05 0.06 0.07 0.08
σ
2
i
0.10
2
0.20
2
0.15
2
0.25
2
Table 2: Correlation coefficient by port0.
ρ
i j
ξ
1
ξ
2
ξ
3
ξ
4
ξ
1
1.0 0.7 0.1 0.4
ξ
2
0.7 1.0 0.5 0.2
ξ
3
0.1 0.5 0.1 0.3
ξ
4
0.4 0.2 0.3 1.0
Table 3: Index of data set and number of assets.
Data set Index n
port1 Hang Seng 31
port2 DAX 85
by MATLAB (L
´
opez, 2014) is employed. In order to
enhance the performance of the interior point method,
the gradients of objective functions, namely ∇γ(x
x
x) in
(31) and ∇γ
0
(x
x
x) in (45), are used explicitly. As stated
above, the optimality of the solutions x
x
x
n
obtained
the interior point method have been also verified.
5.1 Problem Instances
Instances of POPL are defined by using three data sets
of assets, which are named port0, port1, and port2.
The data set called port0 is given by Table 1 and
Table 2. The port0 consists of n = 4 assets. Table 1
shows the mean µ
i
and variance σ
2
i
of asset returns
ξ
i
, n = 1, ··· , n. Table 2 shows the correlation
coefficient ρ
i j
between asset returns ξ
i
and ξ
j
.
The data sets called port1 and port2 are provided
by OR-Library (Beasley, 1990). The data set contains
means, variances, and a coefficient matrix of n asset
returns. Table 3 shows the capital market indices of
those data sets and the numbers of their assets.
5.2 Fixed Interest Rate
From each of the data sets, POPL is formulated as
shown in (24). By changing the value of the risk
α (0, 0.5), POPL in (24) is solved repeatedly. A
constant value is used for the interest rate of loan
L regardless of the value of α (0, 0.5).
Figure 3 shows the efficient frontier evaluated for
POPL of port0, where the upper limit of loan is given
as M = 2. Three different interest rates, L = 0.03,
0.04, and 0.05, are compared in Figure 3. “None” is
the efficient frontier when the loan is not used.
Figure 4 shows the proportion of loan x
0
0 for
each portfolio shown in Figure 3. Since “None”
COMPLEXIS 2020 - 5th International Conference on Complexity, Future Information Systems and Risk
20
Figure 3: Efficient frontier for port0 with M = 2.
Figure 4: Proportion of loan x
0
for port0 with M = 2.
Figure 5: Efficient frontier for port0 with M = 3.
Figure 6: Proportion of loan x
0
for port0 with M = 3.
doesn’t use the loan, it keeps x
0
= 0 in Figure 4.
Figure 5 shows the efficient frontier evaluated for
port0 with M = 3. Figure 6 shows the proportion of
loan x
0
[M, 0] for each portfolio in Figure 5.
From Figure 3 and Figure 5, we can confirm that
the efficient frontier is further improved by borrowing
Figure 7: Efficient frontier for port1 with M = 2.
Figure 8: Proportion of loan x
0
for port1 with M = 2.
Figure 9: Efficient frontier for port1 with M = 3.
Figure 10: Proportion of loan x
0
for port1 with M = 3.
much more money. Furthermore, from Figure 4 and
Figure 6, we can see that the loan is always used up
to the limit (x
0
= M) regardless of its value M.
Figure 7 shows the efficient frontier evaluated for
port1 with M = 2. Figure 8 shows the proportion of
loan x
0
[M, 0] for each portfolio in Figure 7.
An Interest Rate Decision Method for Risk-averse Portfolio Optimization using Loan
21
Figure 11: Efficient frontier for port2 with M = 2.
Figure 12: Proportion of loan x
0
for port2 with M = 2.
Figure 13: Efficient frontier for port2 with M = 3.
Figure 14: Proportion of loan x
0
for port2 with M = 3.
Figure 9 shows the efficient frontier evaluated for
port1 with M = 3. Figure 10 shows the proportion of
loan x
0
[M, 0] for each portfolio in Figure 9.
Figure 11 shows the efficient frontier evaluated for
port2 with M = 2. Figure 12 shows the proportion of
loan x
0
[M, 0] for each portfolio in Figure 11.
Table 4: Interest rate of loan by proposed method.
α 0.05 0.10 0.15 0.20 0.25
port0 0.005 0.010 0.015 0.020 0.020
port1 0.001 0.001 0.005
port2 0.001 0.002 0.004 0.006 0.008
α 0.30 0.35 0.40 0.45
port0 0.025 0.030 0.035 0.040
port1 0.005 0.010 0.015 0.020
port2 0.010 0.010 0.012 0.015
Figure 15: Efficient frontier for port0.
Figure 16: Proportion of loan x
0
for port0.
Figure 13 shows the efficient frontier evaluated for
port2 with M = 3. Figure 14 shows the proportion of
loan x
0
[M, 0] for each portfolio in Figure 13.
From Figure 12 and Figure 14, the loan is not used
at all when the interest rate is high (L = 0.03).
From Figure 3 to Figure 14, we can confirm that
the use of loan works well for improving the efficient
frontier. Besides, the loan is always used up to the
limit as x
0
= M. The lower interest rate provides
higher return and benefits borrowers. On the other
hand, the high interest rate of loan doesn’t benefit
lenders because such a loan is not often used. The
high interest rate of loan doesn’t benefit borrowers,
either. That is because the efficient frontier can’t be
improved for borrowers without using the loan.
5.3 Variable Interest Rate
From each of the data sets, POPL is formulated as
shown in (46). According to the proposed method,
COMPLEXIS 2020 - 5th International Conference on Complexity, Future Information Systems and Risk
22
Figure 17: Efficient frontier for port1.
Figure 18: Proportion of loan x
0
for port1.
Figure 19: Efficient frontier for port2.
Figure 20: Proportion of loan x
0
for port2.
a proper interest rate of loan L is decided for
each of the risk α (0, 0.5) as shown in Table 4.
The loan can’t be used for port1 when the risk is too
low (α 0.1). That is because the optimal solution
x
x
x
n
of the sub-problem of POPL in (44) doesn’t
meet γ
0
(x
x
x
) 0. For each pair of α and L shown in
Table 4, POPL in (46) is solved repeatedly.
Figure 15 shows the efficient frontier evaluated for
port0, where the proper interest rate L in Table 4
is used for each risk α (0, 0.5). Two different upper
limits, M = 2 and M = 3, are compared in Figure 15.
Figure 16 shows the proportion of loan for each
portfolio shown in Figure 15. From Figure 16, we can
confirm that every portfolio except “None” borrows
money from the loan up to the limit as x
0
= M.
Figure 17 shows the efficient frontier evaluated for
port1 with the interest rate of loan L in Table 4.
The loan is not used for the low risks, α = 0.05 and
α = 0.1, in Figure 17. Figure 18 shows the proportion
of loan for each portfolio shown in Figure 17.
Figure 19 shows the efficient frontier evaluated for
port2 with L in Table 4. Figure 20 shows the
proportion of loan for each portfolio in Figure 19.
From Figure 15 to Figure 20, we can confirm that
the loan with the proper interest rate L is always
used for improving the efficient frontier regardless of
the acceptable risk α (0, 0.5). Furthermore, we can
see that the return γ(x
x
x) of the optimal solution x
x
x
n
for POPL depends not only on the risk α but also on
the upper limit of loan M. Specifically, we can get
more return by borrowing more money.
6 CONCLUSION
Portfolio optimization using loan has been formulated
as POPL and solved in this paper. The emphasis of
our work is on the proposal of an interest rate decision
method for POPL. From an acceptable risk α, the pro-
posed method can derive a proper interest rate of loan
L that benefits both borrowers and lenders. Thereby,
POPL in (24) can be also written by a rather simple
form in (46). Finally, from the result of the numeri-
cal experiment, we have confirmed that the efficient
frontier is improved by using the loan completely.
As mentioned above, the proposed method in this
paper benefits both borrowers and lenders. Therefore,
we can expect that the proposed method contributes
to economic revitalization through active investment
using loan. On the other hand, we have to choose
the acceptable risk α (0, 0.5) carefully to use the
proposed method safely and effectively.
For future work, we will extend POPL based on a
multi-period framework. Furthermore, we would like
to include cardinality constraints into POPL.
An Interest Rate Decision Method for Risk-averse Portfolio Optimization using Loan
23
REFERENCES
Anagnostopoulos, K. P. and Mamanis, G. (2010). A port-
folio optimization model with three objectives and
discrete variables. Computers & Operations Re-
search, 37:1285–1297.
Angelelli, E., Mansini, R., and Speranza, M. G. (2008). A
comparison of MAD and CVaR models with real fea-
tures. Journal of Banking & Finance, 32:1188–1197.
Aranha, C. and Iba, H. (2007). Modelling cost into a ge-
netic algorithm-based portfolio optimization system
by seeding and objective sharing. In Proc. IEEE
Congress on Evolutionary Computation, pages 196–
203, Singapore.
Ash, R. B. (2008). Basic Probability Theory. Dover, Down-
ers Grove.
Beasley, J. E. (1990). OR-Library: distributing test prob-
lems by electronic mail. Journal of the Operational
Research Society, 41(11):1069–1072.
Chang, T.-J., Meade, N., Beasley, J. E., and Sharaiha,
Y. M. (2000). Heuristics for cardinality constrained
portfolio optimization. Computers & Operations Re-
search, 25:1271–1302.
Horst, R. and Pardalos, P. M. (1995). Handbook of Global
Optimization. Kluwer Academic Publishers.
Kataoka, S. (1963). Stochastic programming model.
Econometrica, 31(1/2):181–196.
Konno, H. and Yamamoto, R. (2005). Integer program-
ming approaches in mean-risk models.
Computational
Management Science, 2(4):339–351.
Konno, H. and Yamazaki, H. (1995). Mean-absolute de-
viation portfolio optimization model and its applica-
tions to Tokyo stock market. Management Science,
37(5):519–531.
L
´
opez, C. P. (2014). MATLAB Optimization Techniques.
Springer.
Mansini, R., Ogryczak, W., and Speranza, M. G. (2014).
Twenty years of linear programming based on port-
folio optimization. European Journal of Operational
Research, 234:518–535.
Markowitz, H. (1952). Portfolio selection. The Journal of
Finance, 7(1):77–91.
McCormick, G. P. (1983). Nonlinear Programming. John
Wiley & Sons.
Mokhtar, M., Shuib, A., and Mohamad, D. (2014). Mathe-
matical programming models for portfolio optimiza-
tion problem: a review. International Journal of
Mathematical and Computational Sciences, 8(2):428–
435.
Obeidat, S., Shapiro, D., and Lemay, M. (2018). Adap-
tive portfolio assets allocation optimization with deep
learning. IARIA International Journal on Advances in
Intelligent Systems, 11(1&2):25–34.
Pang, X., Zhou, Y., Wang, P., Lin, W., and Chang, V.
(2018). Stock market prediction based on deep long
short term memory neural networks. In Proc. COM-
PLEXIS 2018, pages 102–108.
Ponsich, A., Jaimes, A. L., and Coello, C. A. C. (2013).
A survey on multiobjective evolutionary algorithms
for the solution of the portfolio optimization problem
and other finance and economics applications. IEEE
Trans. on Evolutionary Computation, 17(3):321–343.
Pr
´
ekopa, A. (1995). Stochastic Programming. Kluwer Aca-
demic Publishers.
Price, K., Storn, R. M., and Lampinen, J. A. (2005). Dif-
ferential Evolution: A Practical Approach to Global
Optimization. Springer.
Roy, A. D. (1952). Safty first and the holding of assets.
Econometrica, 20(3):431–449.
Rubio, F., Mestre, X., and Palomar, D. (2012). Performance
analysis and optimal selection of large minimum vari-
ance portfolios under estimation risk. IEEE Journal of
Selected Topics in Signal Processing, 6(4):337–350.
Ruppert, D. (2011). Statistics and Data Analysis for Finan-
cial Engineering. Springer.
Strumberger, I., Tuba, E., Bacanin, N., Beko, M., and Tuba,
M. (2018). Hybridized artificial bee colony algorithm
for constrained portfolio optimization problem. In
Proc. IEEE Congress on Evolutionary Computation,
pages 1–8, Rio de Janeiro, Brazil.
Tagawa, K. (2019). Group-based adaptive differential evo-
lution for chance constrained portfolio optimization
using bank deposit and bank loan. In Proc. IEEE
Congress on Evolutionary Computation
, pages 1557–
1563, Wellington, New Zealand.
Tagawa, K. (2020). Chance constrained portfolio opti-
mization using loan. In Proc. The 12th International
Conference on Information, Process, and Knowledge
Management, pages 1–6, Valencia, Spain.
Tamiz, M., Azmi, R. A., and Jones, D. F. (2013). On se-
lecting portfolio of international mutual funds using
goal programming with extended factors. European
Journal of Operation Research, 226:560–576.
COMPLEXIS 2020 - 5th International Conference on Complexity, Future Information Systems and Risk
24