Applied Mathematics and Computation 174 (2006) 1–12
www.elsevier.com/locate/amc
Mean–variance portfolio optimal
q
problem under concave transaction cost
a,*
Hong-Gang Xue , Cheng-Xian Xu b, Zong-Xian Feng c
a
School of Economics and Finance, Xi’an Jiaotong University, Xi’an 710049, China
b
Faculty of Science, Xi’an Jiaotong University, Xi’an 710049, China
c
JinHe Center for Economic Research, Xi’an Jiaotong University, Xi’an 710049, China
Abstract
In this paper, the classical mean–variance portfolio model is modified for calculating
a globally optimal portfolio under concave transaction costs. A non-decreasing concave
function is employed to approximate origin transaction cost function. The resulting
model is a D-C (difference of two convex functions) programming and a branch and
bound algorithm is designed to solve the problem. A series of numerical experiments
on the model is presented. The history data of nine stocks in Shan Xi province is used
in experiments, and efficient frontiers generated from the resulting model with different
limitations on investments are presented to show the effect of the model and the effi-
ciency of the algorithm solving the model.
2005 Elsevier Inc. All rights reserved.
Keywords: Mean–variance; Concave transaction cost; Globally optimal portfolio; Branch and
bound algorithm; Efficient frontier
q
This work is supported by National Natural Key Product Foundations of China 10231060.
*
Corresponding author. Address: Faculty of Science, XiÕan Jiaotong University, XiÕan 710049,
China.
E-mail address: [email protected] (H.-G. Xue).
0096-3003/$ - see front matter 2005 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2005.05.005
2 H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12
1. Introduction
The mean variance model for portfolio selection was first proposed by
Markowitz in 1952, and has been used by many companies and investors in
financial market. Assume that there exist n risky assets with expected return
rate ri of ith asset, i = 1, 2, . . . ,n. Then the model can be expressed as [8]
l T
min x Vx ð1 lÞRT x
2 ðM-VÞ
s.t. x 2 D \ S;
where x = (x1, x2, . . . , xn)T denotes the investment weight vector, R ¼
T
ðr1 ; r2 ; . . . ; rn Þ the expected return rate vector and V = (rij)n·n the variance–
covariance matrix of the return distributions of these n assets. Parameter
l 2 (0, 1] reflects the investorÕs attitude towards risk (which is measured by
the variance of the portfolio). The set D is a polytope that is generally formu-
lated by a series of linear P inequalities such as Ax 6 b. One typical case is the
capital budget constraint ni¼1 xi ¼ 1. The set S = {li 6 xi 6 ui, i = 1, 2, . . . , n}
is a rectangle and gives the low bound l = (l1, l2, . . . , ln)T and the upper bound
u = (u1, u2, . . . , un)T on the vector x.
In general, the variance–covariance matrix V is symmetric and positive def-
inite, and problem (M-V) is a convex quadratic programming. Effective meth-
ods are available for the solution of problem (M-V) [9]. The transaction cost
does not considered in problem (M-V). However, recent researches show that
the transaction cost is considerable in some investment actions. Konno and
Wijayake [5] studied a portfolio optimization problem under concave transac-
tion cost. The absolute deviation (AD) of the portfolioÕs return distribution is
used to measure the portfolioÕ risk, and the resulting problem is a Mean–AD
(MAD) model under concave transaction cost. A branch-and-bound algorithm
is proposed to solve the MAD model, and linear approximations are used to
result in linear sub-programming [5]. Krokhmal, Palmquist and Uryasev [7]
also took the transaction cost into account in their portfolio optimization
problem in which the CVaR is used to measure the portfolioÕs risk and trans-
action cost is expressed as a linear function.
In this paper, we consider the M–V model (M-V), but the transaction will
included into the model. Since the concave transaction cost is employed, the
resulting model is a D-C (difference of two convex functions) programming,
and it is difficult to find itÕs global solution. Pn However, the transaction cost func-
tion C(x) is separable, i.e., CðxÞ ¼ i¼1 C i ðxi Þ, a branch and bound algorithm
that is similar to [1,5,10,13,14] is proposed to solve the problem by using its
structure. Konno [5,6] proposed a branch and bound algorithm based on linear
underestimation to the concave transaction cost function to solve large scale
portfolio optimal problems with MAD model and concave transaction cost.
H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12 3
Numerical tests on 200 assets and 60 simulation sceneries showed that the algo-
rithm can solve large scale problems effectively.
Subdivision processes play an important role in proving the convergence of
a branch and bound method. The concept of ‘‘normal rectangular subdivision’’
introduced in [3,4,11] will be used in branch and bound algorithms for the solu-
tion of the problem with set S. The class of normal subdivision includes well-
know exhaustive bisection, x-subdivision, adaptive bisection. Xue and Xu
[13,14] proposed a branch and bound algorithm based on the largest distance
bisection, and it is proved that the largest distance bisection is also a normal
rectangular subdivision. Solve these problems from 20 to 160 variables effec-
tively. Some evidence showed that the x-subdivision is better than the exhaus-
tive bisection and the adaptive bisection [10], while numerical tests in [13,14]
showed that the largest distance bisection is more efficient than the x-subdivi-
sion. Hence, we will use the largest distance bisection in the proposed branch-
and-bound algorithm for the solution of the resulting D-C programming
problems.
The rest of the paper is organized as follows. In Section 2, we will construct
the mean–variance model under concave transaction cost and describe the new
branch and bound algorithm for solving it. In Section 3, we will disrupt the de-
tail process of the largest distance bisection method. In Section 4, we will com-
pare the case that the transaction cost exist with the other case that the
transaction cost exist, and we will analysis the effective of the investment upper
bound and low bound for efficient frontier by numerical tests. Conclusions are
given in Section 5.
2. The M–V portfolio problem under concave transaction cost
In this section, we will give the mean–variance portfolio problem under con-
cave transaction costs.
The transaction cost associated with a portfolio x = (x1, x2, . . . , xn)T is usu-
ally defined as the sum of individual transaction cost on each asset
Xn
CðxÞ ¼ C i ðxi Þ;
i¼1
where Ci(xi) is the individual cost on the ith asset. Ci(xi) generally is a non-
decreasing concave function up to certain point ^xi . It is assumed that Ci(xi),
i = 1, 2, . . . ,n are smooth enough. Then the first-order derivative of the trans-
action cost function has the property C 0i ðxi Þ P 0 and the second-order deriva-
tive of Ci(x) satisfies C 00i ðxi Þ < 0 for all i = 1, 2, ,n. Such properties of the cost
function coincide with the rule of margin cost decreasing, and can be described
by the curve in Fig. 1.
4 H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12
Fig. 1. concave transaction cost.
However, the unit transaction cost increases beyond the point ^xj , due to the
‘‘ill-liquidity’’ effects. Konno and Wijayanayake explained this economic phe-
nomena in [5], and made an assumption that the amount of investment on the
jth asset is below the critical point ^xj . The value of ^xj can be obtained by finding
the inflexion point of original transaction cost function. Under such an
assumption, the transaction cost is a well specified concave function and can
be calculated by the transaction cost table of the agent.
Under the concave transaction cost, the portfolioÕs net return is RTx C(x),
and the portfolio optimal problem can be expressed as follows:
min f ðxÞ
ðQ0 Þ
s.t. x 2 D \ S;
where the objective function
f ðxÞ ¼ pðxÞ þ uðxÞ; ð1Þ
l
pðxÞ ¼ xT Vx
2
is the convex part which denotes the portfolioÕs risk, and
Xn
uðxÞ ¼ ui ðxi Þ; ui ðxi Þ ¼ ð1 lÞðC i ðxÞ ri xi Þ; i ¼ 1; 2; . . . ; n ð2Þ
i¼1
is the portfolioÕs net return. Problem (Q0) is a D-C programming.
The constraints in (Q0) is the same as in problem (M-V). In general,
1
n
6 ui 6 ^xi . If short sells are not permitted, then li P 0, else li 6 0. If we set
1
xPi ¼ n for all i = 1, 2, . . . , n, then the1 investment weight vector x 2 S, and
n 1 T
i¼1 x i ¼ 1, so x 2 D. That is, x ¼ n
; . . . ; n
is a feasible point, and hence,
the feasible set is nonempty.
Theorem 2.1. Suppose the concave transaction cost functions Ci(xi)
i = 1, 2, . . . , n are continues, then the global solution of the problem (Q0) exists.
H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12 5
Proof. It is clear that the feasible set D \ S is nonempty, closed and bounded.
While the assumption implies that the objective function is continues on the
feasible region. Therefore, the conclusion comes from the fact that the mini-
mum exists for a continuous function on a closed and bounded nonempty
set. h
3. The branch and bound algorithm
In this section, we will present the proposed branch and bound method in
which the largest distance rectangular subdivision process will be used to divide
problems into sub-problems.
3.1. The description of the algorithm
Let S0 = {li 6 xi 6 ui, i = 1, 2, . . . ,n} be the initial rectangle. We replace the
individual concave functions ui(xi) in u(x) by underestimated linear functions
w0i ðxi Þ over S0 (see Fig. 2)
w0i ðxi Þ ¼ di xi þ gi ; i ¼ 1; 2; . . . ; n; ð3Þ
where
ui ðui Þ ui ðli Þ
di ¼ ; gi ¼ ui ðli Þ di li ; i ¼ 1; 2; . . . ; n. ð4Þ
ui li
Let
g0 ðxÞ ¼ pðxÞ þ w0 ðxÞ;
then g0(x)
Pnis the convex envelope of the function f(x) over the set S0, where
w0 ðxÞ ¼ i¼1 w0i ðxi Þ. We solve the quadratic underestimated approximation
to (Q0),
Fig. 2. The underestimated linear function.
6 H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12
minfg0 ðxÞ ¼ pðxÞ þ w0 ðxÞjx 2 D; l 6 x 6 ug. ðQ0 Þ
(Q0 ) is a convex quadratic programming problem. The active set method in
[2,12] can be used to effectively solve ðQ0 Þ. Let x0 be an optimal solution
ðQ0 Þ, then we obtain an lower bound g0 ðx0 Þ and a upper bound f ðx0 Þ of the
optimal value f(x*) of the problem (Q0) according to the following theorem.
Theorem 3.1. Let x0 be an optimal solution of problem ðQ0 Þ and x* be a global
optimal solution of problem (Q0). Then the following relations hold
g0 ðx0 Þ 6 f 6 f ðx0 Þ; ð5Þ
where f* = f(x*).
Proof. It follows from g0(x) 6 f(x) "x 2 [l, u], that:
g0 ðx0 Þ ¼ minfg0 ðxÞjx 2 D; l 6 x 6 ug
6 minff ðxÞjx 2 D; l 6 x 6 ug ¼ f 6 f ðx0 Þ.
This gives the conclusion. h
Theorem 3.1 indicates that if
uðx0 Þ w0 ðx0 Þ 6 ð6Þ
is satisfied with a given tolerance , then x0 is an approximate solution of (Q0)
with error less than . If (6) does not hold, we will use an NRS process to divide
the problem (Q0) into two subproblems:
minff ðxÞjx 2 D; x 2 S 1 g ðQ1 Þ
and
minff ðxÞjx 2 D; x 2 S 2 g; ðQ2 Þ
where the sub-rectangles S1 and S2 are generated from S0
S 1 ¼ fx j ls 6 xs 6 hs ; lj 6 xj 6 uj ; j ¼ 1; 2; . . . ; n; j 6¼ sg; ð7Þ
S 2 ¼ fxjhs 6 xs 6 us ; lj 6 xj 6 uj ; j ¼ 1; 2; . . . ; n; j 6¼ sg: ð8Þ
Using a similar way, we can get two underestimated quadratic programming
programs to the branched subproblems (Q1) and (Q2) by replacing the function
u(x) with new underestimated linear functions w1(x) and w2(x), where
X 0
w1 ðxÞ ¼ wi ðxi Þ þ w1s ðxs Þ ð9Þ
i6¼s
H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12 7
and
X
w2 ðxÞ ¼ w0i ðxi Þ þ w2s ðxs Þ. ð10Þ
i6¼s
The following is the underestimated quadratic programming to problem
(Q1):
minfg1 ðxÞ ¼ pðxÞ þ w1 ðxÞjx 2 D; x 2 S 1 g: ðQ1 Þ
If problem (Q1 ) is infeasible, then problem (Q1) is also infeasible, and we will
delete problem (Q1). Otherwise, let x1 be an optimal solution of (Q1 ), then
we obtain an low bound g1 ðx1 Þ and a upper bound f ðx1 Þ for the optimal value
f(x1) of problem (Q1), where x1 is an optimal solution of (Q1). If g1 ðx1 Þ > f ðx0 Þ,
then f ðx1 Þ P g1 ðx1 Þ > f ðx0 Þ (according to Theorem 3.1), and problem (Q1) will
be deleted from further consideration. Otherwise, if f ðx1 Þ g1 ðx1 Þ < , then
problem (Q1) is solved with x1 being an approximate solution, if f ðx1 Þ <
f ðx0 Þ, x1 will replace x0 as an approximation to the optimal solution. If
f ðx1 Þ g1 ðx1 Þ P , the set S1 will be further divided into two subsets to gener-
ate two new subproblems. Repeat this process until no subproblems exists.
Now we give the detailed description of the proposed branch-and-bound
algorithm.
Algorithm 1 (The branch-and-bound algorithm).
Step 0. Let l0 = l, u0 = u, and give tolerance > 0. Solve ðQ0 Þ to obtain an
optimal solution x0 , and set x0 ¼ x0 ; a0 ¼ f ðx0 Þ; b0 ¼ bðQ0 Þ ¼
g0 ðx0 Þ, Q = {Q0} and k = 0.
Step 1. Delete all Qi from Q with b(Qi) > ak . Let Q be the set of remaining
subproblems. If Q = /, terminate and xk is an global optimal solu-
tion of (Q0).
Step 2. Select a problem (Qk) from Q
minff ðxÞjx 2 V ; x 2 S k g ðQk Þ
such that
bk ¼ bðQk Þ ¼ minfbðQj Þ; Qj 2 Qg
and divide Sk into Sk,1 and Sk,2 using a NRS process, generate two
subproblems Qk,1 and Qk,2, Set Q = Qn(Qk).
Step 3. For i = 1, 2, solve quadratic underestimate problem Qk;i to obtain opti-
mal solutions xk;i . If Qk;i is infeasible, then Q = Q, else Q = Q [ {Qk,i,
and set bk;i ¼ gk;i ðxk;i Þ; ak;i ¼ f ðxk;i Þ.
Step 4. Set ak+1 = min{ak, ak,1, ak,2}, xk+1 = argmin{ak+1}, k = k + 1 and then
go to Step 1.
The convergence of the algorithm is given by the following theorem. Since the
proof process is similar to one of Theorem 2.2 in [14] and hence is omitted here.
8 H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12
Theorem 3.2. The sequence xk generated by the algorithm above converges to an
optimal solution of (Q0) as k ! 1.
3.2. Normal rectangle subdivisions
Various NRS processes for rectangle subdivision are available [10,13].
Numerical tests in [13] show that the the largest distance bisection (LDB) strat-
egy is efficient for solving D-C programming [10,13,14], and hence the LDB
strategy will be employed in Algorithm 1.
Suppose that a rectangle S k ¼ fxjlki 6 xi 6 uki ; i ¼ 1; 2; ; ng is selected for
further division in Algorithm 1. For simplicity, we will denote the two sub-rect-
angles obtained from a bisection method as S+1, S+2, that is
S þ1 ¼ fxjlks 6 xks 6 hs ; lkj 6 xkj 6 ukj ; j ¼ 1; 2; . . . ; n; j 6¼ sg; ð11Þ
S þ2 ¼ fxjhs 6 xks 6 uks ; lkj 6 xkj 6 ukj ; j ¼ 1; 2; . . . ; n; j 6¼ sg. ð12Þ
Different choices in the index s and the point hks generate different rectangle
subdivisions. The largest distance bisection choose the index s and the point
hks in the following way.
The largest distance bisection
Algorithm 2 (Largest distance bisection (LDB)).
Step 1. Calculate the slopes of lines of the underestimation functions wki ðxi Þ
ui ðuki Þ ui ðlki Þ
di ¼ ; i ¼ 1; 2; . . . ; n:
uki lki
Express the distance between ui(xi) and wki ðxi Þ for xi 2 ½lki ; uki
d i ðxi Þ ¼ ui ðxi Þ wki ðxi Þ ¼ ui ðxi Þ di xi gi ; i ¼ 1; 2; . . . ; n:
Step 2. Maximize the distance function di(xi) to get the solution hki . Let d i ðhki Þ
be the maximum.
Step 3. Determine the index s from
d s ðhks Þ ¼ maxfd i ðhki Þ; i ¼ 1; 2; . . . ; ng: ð13Þ
Step 4. Determine the bisection point hks . One method is to set hks ¼ xks , and
another is to set hks as the point calculated at Step 2.
It is proved in [13] that the LDB method is a NRS process that ensures the
convergence of Algorithm 1.
H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12 9
4. Numerical tests
Numerical experiments on the proposed algorithm have been conducted
on portfolio optimal problem (Q0) with nine stocks from Shan Xi province
in China. Weekly prices for these stocks are used in numerical calculations.
The program was coded using MatLab and tested on Pentium Pro
1794 MHZ with 256 Mbyte memory. The parameter value e = 105 is used in
the algorithm to terminate the iteration for tests on both the subdivisions. Qua-
dratic functions are used to fit transaction cost functions. In problem (Q0), we
set polytop D = {xjx1 + x2 + + x9 = 1}.
Fig. 3 presents the efficient frontiers generated from the the M–V portfolio
models without transaction cost (the model (M–V)) and with transaction cost
(the model (Q0). The low bound and upper bound on investment are li = 0 (i.e.,
short sell is forbidden), and ubi = 0.35, i = 1, 2, . . . , n. It can be observed that the
efficient frontier without transaction is higher than the efficient frontier with
transaction cost, and the difference between the returns with and without trans-
action cost under the same risk level increases as the risk increases.
Fig. 4 presents the efficient frontiers from the model (Q0) under three differ-
ent upper bounds on investment. The short sell is still not permitted, i.e., li = 0,
–3
x 10 Mean – Variance portfolio
2.6
without transaction cost
with transaction cost
2.5
2.4
return rate
2.3
2.2
2.1
1.9
0.0108 0.011 0.0112 0.0114 0.0116 0.0118 0.012
risk(variance)
Fig. 3. Efficient frontiers with and without transaction cost.
10 H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12
–3
x 10 Mean – Variance portfolio
2.6
ub=0.325
ub=0.3
ub=0.275
2.5
2.4
return rate
2.3
2.2
2.1
1.9
0.011 0.0112 0.0114 0.0116 0.0118 0.012 0.0122 0.0124 0.0126 0.0128
risk(variance)
Fig. 4. Efficient frontiers under difference investment upper bound.
i = 1, 2, . . . ,n. It can be observed from the figure that higher in values of the
upper bound, higher the efficient frontier. It also be found from the calculation
results that with higher values in upper bound, the diversify of the resulting
portfolio increases. Thus, the transaction cost increases, and the net return
of the portfolio decrease under the same level of risks. It can also be observed
from the figure that the differences between efficient frontiers increase as the le-
vel of risks increases. Larger the upper bound, higher the increment on the net
return. This implies that the net returns can be increased by increasing the
upper bound on the limitation of the investment without increasing the level
of risks.
Fig. 5 presents the efficient frontiers with transaction cost under different
low bounds on investment. The upper bound is ui = 0.3 and the values on
low bound are li = 0.01, 0, 0.01, i = 1, 2, . . . ,n. It can be observed from the fig-
ure that the lower the low bound on the investment, the higher the position of
the efficient frontier. This is because when the low bound on investment de-
creases, the diversity of the resulting investment reduces, that is, the investment
will be on few assets, and hence the transaction cost will be reduce and the net
return increases. This will give the higher efficient frontier. Another observa-
H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12 11
–3
x 10 Mean–Variance portfolio
2.6
lb=0.0 1
lb=0
2.5 lb=0.01
2.4
return rate
2.3
2.2
2.1
1.9
0.011 0.0115 0.012 0.0125
risk(variance)
Fig. 5. Efficient frontiers under difference investment low bound.
tion from the figure is that under the same return requirement, the portfolio
with lower value of the low bound on investment has lower level of risks that
the portfolio with higher values of low bound. This implies that investors can
reduces their risks by reducing their low bound on investment without reducing
the requirement on net returns, and hence, short sells are a valuable investment
strategy in reducing risks on investment.
5. Conclusion
In this paper, the classical mean–variance portfolio model is modified to
take the concave transaction costs so that it can approximate the conditions
in real market well. A branch and bound algorithm is given for the solution
of the mean–variance portfolio optimal models. Numerical tests on the model
using the branch-and-bound algorithm with the data of 9 stocks in Shan Xi
province show that the efficient frontier with the transaction costs taken into
account will be lower than the efficient frontier without transaction cost. Tests
with different low bounds and upper bounds on investment have been made to
show the effects of the transaction cost.
12 H.-G. Xue et al. / Appl. Math. Comput. 174 (2006) 1–12
References
[1] J.E. Falk, R.M. Soland, An algorithm for separable nonconvex programming problems,
Manage. Sci. 15 (1996) 550–569.
[2] P.E. Gill, W. Murray, M.H. Wright, Practical Optimazition, Academic Press. Inc, London,
1982.
[3] R. Horst, H. Tuy, Global Optimization Deterministic Approaches, third ed., Springer, Berlin,
New York, 1997.
[4] R. Horst, P.M. Pardalos, N.V. Thoai, Introduction to Global Optimization, Kluwer
Academic Publishers, London, 1995.
[5] H. Konno, A. Wijayanayake, Portfolio optimazation problem under concave transaction costs
and minimal transaction unit constrains, Math. Program., Ser. B 89 (2001) 233–250.
[6] H. Konno, P.T. Thach, H. Tuy, Optimization on Low Rank Nonconvex Structures, Kluwer
Academic Publishers, Dordrecht, 1997.
[7] P. Krokhmal, J. Palmquist, S. Uryasev, Portfolio optimization with conditional Value-at-Risk
objective and constrains, Research Report 2001-9, ISE Department, University of Florida,
September 2001.
[8] H.M. Markowitz, Portfolio selection, J. Finan. 7 (1) (1952) 77–79.
[9] M.C. Steinbach, Markowitz revisited: single-period and multi-period mean–variance models,
Woking Paper, Konrad-Zentrum für Informationstechnik Berlin, Preprint SC-99-30, August,
1999.
[10] Q.P. Thai, Le Thi Hoai An, Pham Dinh Tao, Decomposition branch and bound method for
globally solving linear constrained indefinate quadratic programming problems, Operat. Res.
Lett. 17 (1995) 215–220.
[11] H. Tuy, Convex Analysis and Global Optimization, Kluwer Academic Publishers, Dordrecht,
1998.
[12] R.T. Rockafellar, Convex Analysis, Princeton University Press, Princeton, NJ, 1970.
[13] H. Xue, C. Xu, A branch and bound algorithm for Separable Concave Programming,
J. Computat. Math. 6 (2004) 895–904.
[14] H. Xue, C. Xu, A branch and bound algorithm for solving a class of D-C programming, Appl.
Math. Comput. 165 (2) (2005) 291–302.
