Portfolio Selection with Risk Aversion

Index by Optimizing over Pareto Set

Tran Ngoc Thang(B) and Nguyen Duc Vuong

School of Applied Mathematics and Informatics,

Hanoi University of Science and Technology, Hanoi, Vietnam
[email protected]

Abstract. In decision making theory, portfolio selection problems have

an important role, which suggest the best choices among many investing
alternatives. Within our article, we take into consideration the portfo-
lio selection problem with risk aversion index as an optimization prob-
lem over the efficient set of a convex biobjective programming problem
based on Markowitz mean-variance model. By using the outcome space
approach, we alter the considered problem to a convex programming
problem, which is solved efficiently by some computational tools. The
proposed algorithm is applied to optimize security portfolios and some
experiments are reported.

Keywords: Portfolio selection · Markowitz model · Risk aversion

index · Bi-objective programming · Pareto set

1 Introduction

In economy and finance, portfolio optimization helps investors research, under-

stand and make right decisions about their investments and strategies. Let’s
consider a set of potential portfolios, the process of selecting the best portfolio
according to some constraints is portfolio optimization.
Harry Markowitz was the first one to give a clear definition of modern portfo-
lio theory (see [1]) in which an investor wants to maximize a portfolio’s expected
return on any given amount of risk. Our research approaches the portfolio opti-
mization problem by the Mean-Variance model as a special case of bio-objective
convex programming (BOP ). To find an optimal solution in the Pareto set, we
consider an auxiliary factor. In other words, we solve the optimization problem
over the Pareto set. This problem belongs to NP-hard class even in the case we
have linear objective functions and a polyhedral feasible set. Therefore, we alter
the main problem to a convex programming problem.
Portfolio selection has attracted special attention of researchers. A lot of
results have been released in the fields of economics, finance, agriculture and
so on (see [2,4,5] and references therein). Some popular methods can be listed
as follows: Weighted Sum of Deviations (see [7]), Chebyshev Goal Programming

c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021
D.-T. Tran et al. (Eds.): ICISN 2021, LNNS 243, pp. 225–232, 2021.
https://doi.org/10.1007/978-981-16-2094-2_28
226 T. N. Thang and N. D. Vuong

(see [7]) or Lagrangian Multiplier (see [8]). Most of previous methods apply scar-
larization to transfrom Markowitz model to single-objective problem. However,
whether scarlarization is applied to non-linear functions or not is still a contro-
versial question. We approach problem (PS ) as an optimization problem over
the efficient set (PX ). Based on approaching by the outcome space, we propose
to transform problem (PX ) to a single-objective convex programming problem.
In Sect. 2, the theoretical preliminaries are presented to analyze the portfolio
selection problem with risk aversion index and propose the equivalent convex
programming problem. Section 3 introduces the algorithm to solve the problem
including some computational experiments as illustrations. A few conclusions
are given within the final section.

2 Theoretical Preliminaries

2.1 Portfolio Selection by Markowitz Model

Consider the portfolio determination problem as a bi-objective programming

problem based on Markowitz mean-variance model. Recall x = (x1 , x2 , ..., xn )T ,
in which xj represents the relative amount invested in asset number j. Since,
vector
n x is called a portfolio. Note that xj is positive for all j = 1, n and
j=1 j = 1. Return Lj of asset number j, j = 1, n is a random variable with
x
T T
expected value lj = E(Lj ). Notice L = (L1 , L2 , ..., L n ) , l = (l1 , l2 , ..., ln ) .
T n
Then, the whole portfolio has a return given as L x = j=1 Lj xj and expected
n
return given as E(x) = E(LT x) = j=1 lj xj . Calling co-variance matrix of ran-
dom vector
 R is Q = (σij )n×n . Then, variance of return is V(x) = V ar(LT x) =
n n
j=1 ij xj xi where σj is variance and ρij is correlation coefficient between
σ 2
i=1
Lj and Li , i, j = 1, 2, ..., n. Easy to see, matrix Q is symmetric and positive semi-
definite.
Thus, the profit is calculated based on the expected return or the average
value of the returns, while the risk is displayed by the variance of the returns
on the whole portfolio. Investors perform optimally on two objectives that are
maximally profitable and minimally risky functions with conditions bound on
the value of the portfolio. Then, the problem of optimizing the portfolio is given
as the following:
n

max E(x) = lj xj
j=1
n  n
(P S)

min V(x) = σij xj xi
i=1 j=1

s.t. x1 + x2 + ... + xn = 1, xj ≥ 0, j = 1, n.
Here, the risk function can be written in quadratic form V(x) = xT Qx. Because
Q is positive semi-definite so V(x) is a convex function. Also easy to see, return
 Portfolio Selection with Risk Aversion Index 227

function E(x) is linear. So problem (PS ) is a special case of convex biobjective

programming which can be given as the following

Min{f (x)|x ∈ X }, (BOP )

where X is a nonempty compact convex set and the objective function f =

(f1 , f2 )T is convex on X . Recall that a vector function f (x) is called convex
on X if its component functions fi (x), i = 1, 2, are convex on X . The problem
of selecting a portfolio is known as finding the efficient solution set of problem
(BOP ).

2.2 Portfolio Selection Problem with Risk Aversion Index

As mentioned in the previous subsection, it is stated by Markowitz portfolio

theory that an speculator should select a portfolio from the efficient solution
set of problem (BOP ), depending on his risk aversion. The investor can select
a portfolio by maximizing an utility function, such as U (E(x), V(x)) = E(x) −
A ∗ V(x), where A is the risk aversion coefficient, which is a non-negative real
number. We can replace A by 1/λ − 1 with a real number λ ∈ [0, 1]. Then, we
consider the function Ū (E(x), V(x)) = −λ ∗ U (E(x), V(x)) = λ ∗ (−E(x)) + (1 −
λ) ∗ V(x). We can see that function Ū (E(x), V(x)) is a weighted composition of
return and risk. Selecting a portfolio is solved by minimizing this function over
the Pareto set of (BOP ). In general, the formulation of the main problem is
given as

min{Φ(x)|x ∈ XE }, (PX )

where Φ(x) = λ, f (x) and XE is the efficient solution set for problem (BOP ),
i.e. Recall that XE contains every x0 ∈ X which satisfies that there is no x ∈ X
where f (x0 ) ≥ f (x), f (x0 ) = f (x).
As usual, the notation y 1 ≥ y 2 , where y 1 , y 2 ∈ Rm , is used to indicate yi1 ≥ yi2
for all i = 1, . . . , m.
It is widely acclaim that, in general, the set XE is nonconvex and given
implicitly as the form of a standard mathematical programming problem, even
in the case m = 2 and we have linear objective functions f1 , f2 as well as the
polyhedral feasible set X . Hence, problem (PX ) is a global optimization problem
and belongs to NP-hard problem class. In this article, we propose to transform
the main problem to a convex programming problem.

2.3 The Equivalent Convex Programming Problem

Let Y = {y ∈ Rm |y = f (x) for some x ∈ X }. Normally, the set Y represents the

outcome set for problem (BOP ).
Let Q ⊂ Rm be a nonempty set. MinQ is denoted as the efficient set of Q.
Note that MinQ = {q 0 ∈ Q| q ∈ Q : q 0 ≥ q, q 0 = q}. It is clear that a point
q 0 ∈ MinQ if Q ∩ (q 0 − Rm m m
+ ) = {q }, where R+ = {y ∈ R |yi ≥ 0, i = 1, . . . , m}.
0
228 T. N. Thang and N. D. Vuong

Since the functions fi , i = 1, . . . , m, are continuous and X ⊂ Rn is a

nonempty compact set, the outcome set Y is also compact set in Rm . There-
fore, the efficient set MinY is nonempty [3]. Let YE be denoted as the efficient
outcome set for problem (BOP ), where YE = {f (x)|x ∈ XE }. We can infer by
definition that YE = MinY. The relationship between the efficient solution set
XE and the efficient set MinY is described as follows.

Proposition 1.
i) For any y 0 ∈ MinY, if x0 ∈ X satisfies f (x0 ) ≤ y 0 , then x0 ∈ XE .
ii) For any x0 ∈ XE , if y 0 = f (x0 ), then y 0 ∈ MinY.

Let Z = Y + Rm + = {z ∈ R
m
| z ≥ y for some y ∈ Y}. Apparently, Z is not
empty. Also, it is a full-dimension closed set but it is nonconvex in general (see
Fig. 1) The following interesting property of Z will be used in the sequel.

Proposition 2. MinZ = MinY.

The problem (PX ) can be reformulated by outcome-space approach as

min ϕ(y) s.t. y ∈ YE . (PY )

Combining YE = MinY and Proposition 2, problem (PY ) can be rewritten as

min ϕ(y) s.t. y ∈ MinZ. (PZ )

Thus, we solve problem (PZ ) instead of problem (PY ).

Proposition 3. If f is convex, the set Z = f (X) + R2+ is convex in R2 .

As Z is a nonempty convex subset in R2 by Proposition 3, the fact that MinZ

is homeomorphic to a line segment [a, b] ⊂ R is proven in [4]. By geometry, it is
easily seen that the optimal solution set of the following problem

min{y2 | y ∈ Z, y1 = y1I } (PS )

contains only one solution y S and similarly, the problem

min{y1 | y ∈ Z, y2 = y2I } (PE )

has a unique optimal solution y E . Since Z is convex, problems (PS ) and (PE )
are convex programming problems. If y S = y E then XE = {y S } and y S is the
only optimal solution of problem (PX ), else if y S = y E then MinZ is a curved
line on the boundary of Z with starting point y S and the end point y E such that

y1E > y1S and y2S > y2E . (1)

Note that we also get the efficient solutions xS , xE ∈ XE such that y S = f (xS )
and y E = f (xE ) while solving problems (PS ) and (PE ). For the convenience,
xS , xE is called to be the efficient solutions respect to y S , y E , respectively. In the
 Portfolio Selection with Risk Aversion Index 229

second case, we consider problem (PZ ) where ϕ(y) = λ, y. Direct computation
reveals the line through y S and y E equation is c, y = α, in which
⎧  
⎨c = E 1 S , S 1 E ,
y1 −y1 y2 −y2
E E (2)
⎩α = Ey1 S + Sy2 E .
y −y 1y −y
1 2 2

From (1), it is easily seen that the vector c is strictly positive. Now, let Z̃ = {y ∈
Z | c, y ≤ α} and Γ = ∂ Z̃ \(y S , y E ), where (y S , y E ) = {y = ty S +(1−t)y E | 0 <
t < 1} and ∂ Z̃ is the boundary of the set Z̃.
It is clear that Z̃ is a compact convex set because Z is convex. By the
definition and geometry, we can see that Γ contains every extreme points of Z̃,
moreover, MinZ = Γ . Consider the following convex problem

minλ, y s.t. y ∈ Z̃. (CP 0 )

that has the explicit reformulation as follows

min λ, y
s.t. f (x) − y ≤ 0 (CP 1 )
x ∈ X , c, y ≤ α,

where the vector c ∈ R2 and the real number α is determined by (2). Because of
the convexity of the objective function f and the feasible set X , problem (CP 1 )
is a convex programming problem with n + 2 real variables.

Fig. 1. The convex set Z̃

Proposition 4. If (x∗ , y ∗ ) ∈ arg min (CP 1 ), then x∗ ∈ arg min (PX ).

x x

Proof. It is well known that a convex programming problem with the linear
objective function has an optimal solution which belongs to the extreme point
230 T. N. Thang and N. D. Vuong

set of the feasible solution set [6]. Therefore, problem (CP 0 ) has an optimal
solution y ∗ ∈ Γ . This fact and YE = MinY implies that y ∗ ∈ MinZ. Since
MinZ ⊂ Z̃, it implies that y ∗ ∈ arg min (PZ ).
x
Since MinZ = YE = MinY, by definition, we have λ, y ∗  ≤ λ, y for all
y ∈ YE and y ∗ ∈ MinY. Then
λ, y ∗  ≤ λ, f (x), ∀x ∈ XE . (3)
Because (x∗ , y ∗ ) ∈ arg min (CP 1 ), f (x∗ ) ≤ y ∗ . By Proposition 1, x∗ ∈ XE .
x
Furthermore, f (x∗ ) ∈ Y and y ∗ ∈ MinY. The definition of efficient points infers
that y ∗ = f (x∗ ). Combining this fact and (3), we get Φ(x∗ ) ≤ Φ(x) ∀ x ∈ XE
which means x∗ ∈ arg min (PX ). 
x

3 Procedures and Computing Experiments

The procedure for solving problem (PX ) is established by Proposition 4.
Procedure 1.
Step 1. Solve problem (PS ) and problem (PE ) to find the efficient points y S , y E
and the efficient solutions xS , xE respect to y S , y E , respectively.
Step 2. If y S = y E Then Go to Step 3. Else The procedure is terminated (y S
∈ arg min (PX )).
x
Step 3. Solve problem (CP 1 ) to find an optimal solution (x∗ , y ∗ ). The procedure
is terminated (x∗ ∈ arg min (PX )).
x
Below are the example to illustrate Procedure 1.
Example 1. Consider problem (PX ), where XE is the efficient solution set to
problem (BOP ) with f1 (x) = x21 + 1, f2 (x) = (x2 − 3)2 + 1,

X = x ∈ R2 | (x1 − 1)2 + (x2 − 2)2 ≤ 1, 2x1 − x2 ≤ 1 ,

and Φ(x) = λ1 f1 (x) + λ2 f2 (x). The computational results are shown in Table 1.

Table 1. Results of Example 1 by computation

λ x∗ y∗ Φ(x∗ )
(0.0, 1.0) (0.9612, 2.9991) (1.9412, 1.0000) 1.0000
(0.2, 0.8) (0.4460, 2.8325) (1.1989, 1.0281) 1.0622
(0.5, 0.5) (0.2929, 2.7071) (1.0858, 1.0858) 1.0858
(0.8, 0.2) (0.1675, 2.5540) (1.0208, 1.1989) 1.0622
(1.0, 0.0) (0.0011, 2.0435) (1.0000, 1.9326) 1.0000
(−0.2, 0.8) (0.9654, 2.9992) (1.9412, 1.0000) 0.4118
(0.8, −0.2) (0.0011, 2.0435) (1.0000, 1.9326) 0.4146
 Portfolio Selection with Risk Aversion Index 231

Example 2. (see [8], pp. 7) There are five stocks with the list of ticker symbols as
DGX, AAPL, BAC, IBM, MSFT. Chronicled stock cost and profit installment
from 2002 to 2007 are used to calculate the anticipated return and covariance
matrix of five stocks (Table 2).

Table 2. Expected Returns and Covariance Matrix of chosen stocks

Stock Expected Profit Covariance Matrix

DGX AAPL BAC IBM MSFT
DGX 1.006% 0.0056 0.0014 0.0004 0.0024 −0.0002
AAPL 4.085% 0.0014 0.0127 0.0013 0.0024 0.0010
BAC 1.236% 0.0004 0.0013 0.0016 0.0010 0.0006
IBM 0.400% 0.0024 0.0024 0.0010 0.0065 0.0030
MSFT 0.513% −0.0002 0.0010 0.0006 0.0030 0.0039

We use Matlab to resolve the problem, then the speculator can get an insight
into the Expected Return and Risk for any risk aversion index A in Fig. 2. From
the figure, it can be revealed that when A varies from 20 to 50 or greater, there
is not a significant change in Expected Return. The same trend is also applied
to Risk when A ranges between 5 and 50 or higher.

Fig. 2. Expected Return and Risk vs. Risk Aversion Index

232 T. N. Thang and N. D. Vuong

4 Conclusion
Within this research, a method has been developed to solve the issue of portfolio
selection with risk aversion index by proposing the equivalent single-objective
optimization problem over Pareto Set. Practical experiments reveal that this
method required little computational effort in comparison to other methods
because only some convex programming need solving. Therefore, it can be widely
applied in most fields where exist portfolio optimization.

Acknowledgment. This research is funded by Hanoi University of Science and Tech-

nology (HUST) under grant number T2018-TT-007.

References
1. Markowitz, H.: Portfolio selection. J. Fin. 7(1), 77–91 (1952)
2. Kim, N.T.B., Thang, T.N.: Optimization over the efficient set of a bicriteria convex
programming problem. Pac. J. Optim. 9, 103–115 (2013)
3. Luc, D.T.: Theory of Vector Optimization. Springer, Berlin (1989)
4. Phu, H.X.: On efficient sets in R2 . Vietnam J. Math. 33, 463–468 (2005)
5. Thoai, N.V.: Reverse convex programming approach in the space of extreme criteria
for optimization over efficient sets. J. Optim. Theory Appl. 147(2), 263–277 (2010).
https://doi.org/10.1007/s10957-010-9721-2
6. Tuy, H.: Convex Analysis and Global Optimization. Kluwer (1998)
7. Roberts, M.C., Dizier, A.S., Vaughan, J.: Multiobjective Optimization: Portfolio
Optimization Base on Goal Programming Methods (2012)
8. Duan, Y.C.: A multi-objective approach to portfolio optimization. Rose Hulman
Undergraduate Math. J. 8(1), Article 12 (2007)