Rebalancing An Investment Portfolio in The Presence of Transaction Costs
Rebalancing An Investment Portfolio in The Presence of Transaction Costs
Rebalancing An Investment Portfolio in The Presence of Transaction Costs
Transaction Costs
John E. Mitchell1
Department of Mathematical Sciences,
Rensselaer Polytechnic Institute,
Troy, NY 12180.
[email protected]
http://www.rpi.edu/˜mitchj
Stephen Braun2
Warren & Selbert, Inc.
Santa Barbara, CA 93101
[email protected]
Abstract
The inclusion of transaction costs is an essential element of any realistic
portfolio optimization. In this paper, we consider an extension of the standard
portfolio problem in which transaction costs are incurred to rebalance an in-
vestment portfolio. The Markowitz framework of mean-variance efficiency is
used with costs modelled as a percentage of the value transacted. Each secu-
rity in the portfolio is represented by a pair of continuous decision variables
corresponding to the amounts bought and sold. In order to properly repre-
sent the variance of the resulting portfolio, it is necessary to rescale by the
funds available after paying the transaction costs. We show that the resulting
fractional quadratic programming problem can be solved as a quadratic pro-
gramming problem of size comparable to the model without transaction costs.
Computational results for two empirical datasets are presented.
1
Research partially supported by NSF Grant number CCR–9901822.
2
Research supported in part by the National Science Foundation’s VIGRE Program (Grant DMS-
9983646).
1 Introduction
Constructing a portfolio of investments is one of the most significant financial deci-
sions facing individuals and institutions. A decision-making process must be devel-
oped which identifies the appropriate weight each investment should have within the
portfolio. The portfolio must strike what the investor believes to be an acceptable
balance between risk and reward. In addition, the costs incurred when setting up a
new portfolio or rebalancing an existing portfolio must be included in any realistic
analysis. In this paper, we show that proportional transaction costs can be incor-
porated in a manner which makes the resulting optimization problem a quadratic
program.
Essentially the standard portfolio optimization problem is to identify the optimal
allocation of limited resources among a limited set of investments. Optimality is
measured using a tradeoff between perceived risk and expected return. Expected
future returns are based on historical data. Risk is measured by the variance of those
historical returns.
When more then one investment is involved, the covariance among individual in-
vestments becomes important. In fact, any deviation from perfect positive correlation
allows a beneficial diversified portfolio to be constructed. Efficient portfolios are allo-
cations that achieve the highest possible return for a given level of risk. Alternatively,
efficient portfolios can be said to minimize the risk for a given level of return. These
ideas earned their inventor a Nobel Prize and have gained such wide acceptance that
countless references could be cited; however, the original source is Markowitz [20],
discussed in detail in his classic text [21]. A recent summary of the motivation for
this model appears in Markowitz [22].
One standard formulation of the portfolio problem minimizes a quadratic risk
measurement with a set of linear constraints specifying the minimum expected port-
folio return, E0 , and enforcing full investment of funds. The decision variables xi
are the proportional weights of the ith security in the portfolio. Here n securities are
under consideration. Additionally, µ is the column vector of expected returns and
Q is the positive semidefinite covariance matrix. Short selling is not allowed, so the
proportions xi are restricted to be nonnegative. This formulation is:
1 T
min 2
x Qx
T
s.t. µ x ≥ E0
eT x =1 (1)
x ≥ 0,
1
In this paper, we describe a method for finding an optimal portfolio when propor-
tional transaction costs have to be paid. These costs vary linearly with the amount of
a security bought or sold. Our method requires the solution of a quadratic program
of similar size to the Markowitz model, and then the solution of a linear programming
problem. Our method allows different costs for different securities, and different costs
for buying and selling. Our model captures the feature that transaction costs are paid
when a security is bought or sold and the transaction cost reduces the amount of that
particular security that is available. In particular, both the risk and the return in our
model are measured using the portfolio arising after paying the transaction costs.
The portfolio rebalancing problem has similarities to the index tracking prob-
lem [1, 8, 12]. See Zenios [29] for a discussion of portfolio optimization models. See
Bertsimas et al. [5] for a case study in portfolio optimization. The optimal solution
to the portfolio optimization problem is sensitive to the data Q and µ, so estimating
this data accurately is the subject of current research; see Chopra and Ziemba [10] or
Bengtsson and Holst [3] for example. Stochastic programming approaches to portfolio
optimization have been investigated in [4, 11, 14, 15, 16, 24, 26, 27] and elsewhere;
such approaches work with sets of scenarios.
This paper is organized as follows. We turn to the portfolio rebalancing problem
in §2. First, we motivate the cost model and provide examples of costs that fit this
model before introducing an essential change of variables. Alternative approaches
to this problem are also discussed in §2. Our model is a quadratic programming
problem, and the nature of the solutions to that problem is discussed in §3. If the
return requirement is low, the nature of the solutions can change; this is the subject
of §4. Computational results for two empirical datasets are presented in §5. Finally,
we offer concluding remarks in §6.
1. Capital gains taxes are a security-specific selling cost that can be a major con-
sideration for the rebalancing of a portfolio.
2
high selling cost and low buying cost could be used to express optimism that a
security may outperform its forecast.
Let ui and vi represent the amount bought and sold (respectively) of security i.
The amount invested in each of the securities will be
x = x̄ + u − v. (2)
We assume proportional transaction costs. Let cBi and cSi denote the transaction
cost of buying and selling one unit of security i, respectively. We assume 0 ≤ cB < e,
0 ≤ cS < e and cS + cB > 0. We let x0 denote the total amount spent on transaction
costs, so
x0 = cTB u + cTS v. (3)
The total amount invested in the securities, after paying transaction costs, will be
1 − x0 . We obtain the constraint
eT x = 1 − cTB u − cTS v.
This equation can be used in place of (1) to give a model for minimizing the variance
of the resulting portfolio subject to meeting an expected return of E0 > 0 in the
presence of proportional transaction costs. The resulting model is
1 T
min 2
x Qx (5)
T
s.t. µ x ≥ E0 (6)
x−u+v = x̄ (7)
T T
(cB + e) u + (cS − e) v = 0 (8)
u, v, x ≥ 0. (9)
A user might also require restrictions such as limiting the proportion of assets
that can be invested in a group of securities. We can express this as a homogeneous
constraint on x. For example, if security 1 must constitute no more than 10% of the
resulting portfolio, we can impose the constraint
n
X
9x1 − xi ≤ 0.
i=2
3
When there are no transaction costs to be paid, one dollar is always available for
P
investment, i.e. ( ni=1 xi = 1). This assumption is implicit in the standard risk mea-
sure. However, for nonzero transaction costs that implicit assumption is no longer
valid. One dollar is not available for investment, costs will be paid to rebalance. The
appropriate objective is therefore
1 T
2
x Qx
f(x) := . (10)
(1 − x0 )2
Here x0 is again the amount paid in transaction costs. Therefore (1 − x0) is the
actual amount available for investment, so we are choosing to scale the standard risk
measurement by the square of the dollar amount actually invested.
This gives the fractional quadratic programming problem (F QP ) which we will
solve to find the optimal portfolio for a given expected return.
xT Qx
min 2(1−cT T 2
B u−cS v)
T
s.t. µ x ≥ E0
(F QP ) x − u + v = x̄
(cB + e)T u + (cS − e)T v = 0
Ax ≤ 0
u, v, x ≥ 0.
Analytically, notice that with zero transaction costs then x0 ≡ 0 and we recover
the standard risk measurement. So our choice does pass the first test required of any
theoretical extension; recover the previous result. This choice also makes dimensional
sense given the quadratic numerator.
Our choice of this fractional objective function also makes intuitive sense. For
nonzero transaction costs, there are conflicting effects at work within the portfolio.
Thinking of fixed x̄ then as the transaction cost percentage is increased, you expect
smaller absolute amounts of principal will be available for investment. But in order
to get the same payoff (µT x) on a smaller amount of principal you will need to reach
for higher returns. This should correlate to taking on higher levels of risk. Our
fractional choice effectively boosts the risk measurement for these transaction cost
depleted portfolios.
Without the denominator in the objective function, it is possible to obtain optimal
solutions which involve both buying and selling a particular security, so both ui > 0
and vi > 0 for some security i. Of course, in practice this is not a desirable strategy.
Nonetheless, it does reduce the measure of risk 12 xT Qx. The incidence of solutions of
this form is particularly noticeable for low values of E0 . Normalizing by the amount
invested in securities prevents this undesirable outcome, as we shall see.
The fractional objective f(x) can be made quadratic using the technique of re-
placing the denominator by the square of the reciprocal of a variable. This is a
straightforward extension of the technique of Charnes and Cooper [9] for fractional
programs where the objective is a ratio of linear functions and the constraints are
4
linear. Let
1
t := (11)
1− cTB u − cTS v
and then define
û := tu, v̂ := tv, x̂ := tx. (12)
Note that since u and v are constrained to be nonnegative, we must have t ≥ 1.
Note that we now have t − cTB û − cTS v̂ = 1. The constraints (6)–(8) can be multiplied
through by t. Thus, the fractional quadratic program (F QP ) is equivalent to the
quadratic programming problem (QP ):
1 T
minx̂,û,v̂,t 2
x̂ Qx̂
T
s.t. µ x̂ − E0 t ≥ 0
(QP ) x̂ − û + v̂ − x̄t = 0
(cB + e)T û + (cS − e)T v̂ = 0
− cTB û − cTS v̂ + t = 1
Ax̂ ≤ 0
x̂, û, v̂, t ≥ 0.
Once we find a solution (x̂∗, û∗, v̂ ∗, t∗) to (QP ), we can obtain a solution (x∗ , u∗, v ∗)
to the original problem (F QP ) by rescaling x̂, û and v̂, so x∗ = t1∗ x̂∗, u∗ = t1∗ û∗ , and
v ∗ = t1∗ v̂ ∗.
The efficient frontier is found by optimizing (QP ) for different values of E0. In
Figure 1, we graph the efficient frontier for a nine-security problem, with three dif-
ferent choices for the transaction costs, namely zero costs, 3% costs for each buy
and sell decision, and 5% costs for each buy and sell decision. The initial portfolio
is equally weighted in the nine securities. The optimal objective function is plotted
against the value of E0 . Optimizing this quadratic program creates the situation
where the c = 0% frontier extends furthest into the risk/return plane. Other trans-
action cost efficient frontiers, abbreviated TCEF, are pulled back from that limit as
seen in Figure 1. It is apparent that transaction costs reduce the range of investment
choice.
The model (QP ) is the one that we propose to solve in order to find an optimal
level of risk for a given level of return in the presence of proportional transaction
costs. In the next two sections we examine properties of the solutions to (QP ). First,
we conclude this section by considering other methods for portfolio rebalancing that
have been discussed in the literature.
Adcock and Meade [1] add a linear term for the costs to the original Markowitz
quadratic risk term and minimize this quantity. This requires finding an appropriate
balance between the transaction costs and the risk. The model assumes a fixed rate
of transaction costs across securities. The risk is measured in terms of the adjusted
portfolio before transaction costs are paid.
Konno and Wijayanayasake [18] consider a cost structure that is considerably
more involved than ours, with the result that the model is far harder to solve.
5
The Transaction Cost Efficient Frontier
0.14
c=0%
c =3
%
0.12 c=5%
0.1
2
Risk per Dollar Invested
0.08
0.06
0.04
0.02
0
0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Expected Return
Figure 1: The initial portfolio is located by a circle. Notice that as the level of
transaction costs c increases, the curves shift left. Increased transaction costs reduce
investment choice.
6
Yoshimoto [28] considers a similar transaction cost model to ours and proposes
a nonlinear programming algorithm to solve the problem. Computational results
indicate that ignoring transaction costs can result in inefficient portfolios.
An alternative model is to reduce the vector of expected returns µ by the trans-
action costs. This gives a model of the form
1 T
min 2
x Qx
T
s.t. µ x̄ + (µ − cB )T u − (µ + cS )T v ≥ E0
x − u + v = x̄
T
e x = 1
Ax ≤ 0
u, v, x ≥ 0.
The method implicitly assumes that transaction costs are paid at the end of the
period, impacting both the risk and the return. If the transaction costs must be paid
at the beginning of the period then care must be taken in the sale of assets to pay
the transaction costs, in order to ensure that the resulting portfolio has securities in
the same proportion. Further, the return calculation assumes a return on the amount
paid in transaction costs, so this constraint needs to be modified.
If the only transaction cost is a fixed cost per transaction then transaction costs
can be controlled by placing an upper bound on the number of transactions. This
gives rise to a quadratically constrained integer programming problem. This approach
has been investigated widely; see, for example, Perold [25], Bienstock [6] or Lee and
Mitchell [19]. The presence of the integrality restriction makes the formulation far
harder to solve than the one presented in this paper. Placing an explicit limit on the
number of transactions is only an indirect way to limit transaction costs in the case
of proportional transaction costs and may well result in suboptimal solutions. The
constraint is rather crude. For example, it would require modification in the case that
transaction costs differ from one security to another.
There has been some interest in portfolios that can be modified continuously, start-
ing with Merton [23], with subsequent research including that by Albeverio et al [2],
Davis and Norman [13], and Kamin [17]. These methods can only solve problems
with a very limited number of securities.
7
If the return constraint is not active at the optimal solution, then it is possible
that an optimal solution will not be complementary. We discuss this situation in §4,
where we also show that a complementary solution can always be found efficiently
even in this situation.
The Karush-Kuhn-Tucker (KKT) optimality conditions for the quadratic program
(QP ) require that (x̂, û, v̂, t) be primal feasible and that the following equations can
be satisfied:
AT π + Qx̂ − wµ + sr − sx = 0 (13)
−sr + y(cB + e) − zcB − su = 0 (14)
sr + y(cS − e) − zcS − sv = 0 (15)
−wE0 − x̄T sr + z = 0 (16)
w(µT x̂ − E0 t) = 0 (17)
x̂T sx = 0 (18)
ûT su = 0 (19)
v̂ T sv = 0. (20)
Here, w is a nonnegative scalar, y and z are free scalars, π is a nonnegative m-vector,
sr is a free n-vector, and sx , su , and sv are nonnegative n-vectors. We have exploited
the fact that t ≥ 1 in any feasible solution in deriving (16).
Assume that ûi > 0 and v̂i > 0 for some security i. We argue that we must then
have w = 0 for the KKT conditions to hold. If ûi > 0 and v̂i > 0, it follows from
(19) and (20) that sui = svi = 0. Adding together the ith components of (14) and (15)
gives
(y − z)(cB i + cS i ) = 0.
Since cS + cB > 0, we immediately obtain y = z. Now adding together the whole of
(14) and (15) gives su + sv = 0, so su = sv = 0 and sr = ye. Substituting into (16)
for sr and using the facts that y = z and eT x̄ = 1 gives wE0 = 0. Thus, we have
proved the following theorem.
Theorem 1 If the optimal solution to (QP ) has a strictly positive Karush-Kuhn-
Tucker multiplier for the return constraint µT x̂ − E0 t ≥ 0 then the optimal solution
is complementary.
This theorem does not make any assumptions about the rank of Q. In particular,
the result still holds even if there is a risk-free security, that is, a security for which
there is no variability in the return. Any vector in the nullspace of Q can be regarded
as a risk-free security. By a change of variables, this can be regarded as a single
security, and the corresponding row and column of Q for this security are zero.
8
constraint may be satisfied strictly in the solution to (QP ). Alternatively, it may
be that the presence of the homogeneous constraints Ax ≤ 0 results in an optimal
solution that satisfies the return constraint strictly. If the return constraint is not
active at the optimal solution, then an optimal solution may not be complementary.
Nonetheless, there is an optimal solution that is complementary, and that solution
can be found efficiently. We first give an example of such a situation in §4.1, we then
show that alternative optimal solutions exist, including one that is complementary
in §4.2, and then we show how the complementary solution can be found efficiently
in §4.3.
4.1 An example
Consider, for example, a problem with two securities, with initial portfolio x̄ =
(0.5, 0.5), with buying and selling transaction costs of 1%, with desired expected
return for the portfolio of E0 = 29, and with expected returns for the securities of
µ = (25, 35). Assume the covariance matrix is equal to
" #
6 −2
Q= .
−2 4
Assume there are no side constraints of the form Ax ≤ 0. We also impose the upper
bounds on u and v of e − x̄ and x̄, respectively. The problem is then solved using
CPLEX called from AMPL. The optimal solution value is reported as 0.7140204261,
achieved by " # " #
0.42077 0.50000
u= , v=
0.48194 0.42096
If we require that a security be either bought or sold, and not both, then the optimal
solution value is reported as 0.7142653036, achieved by
" # " #
0 0.07206
u= , v= .
0.07063 0
If this problem is solved to higher precision, it has multiple optimal solutions with
value 0.714286, including noncomplementary solutions and a complementary solution
with u2 = 0.07061 and v1 = 0.07204. The complementary solution can be recovered
from a noncomplementary solution, as we argue later.
9
to û, v̂, and t, respectively, that maintain feasibility and which leave x̂ unchanged.
The new point will be
Theorem 2 Let ∆û and ∆v̂ satisfy ∆û − ∆v̂ = x̄ and (cB + cS )T ∆û = (cS − e)T x̄.
Let ∆t = −1. Let α be a scalar chosen so that û0 ≥ 0 and v̂ 0 ≥ 0. The point defined
in equations (21)–(24) is optimal for (QP ).
Proof: We show first that the point is feasible in (QP ), and then we show that
0 ∗
x̂ = x̂ .
We have
(cB + e)T û0 + (cS − e)T v̂ 0 = α((cB + e)T ∆û + (cS − e)T ∆v̂)
from feasibility of w∗
= α((cB + e)T ∆û + (cS − e)T (∆û − x̄))
from an assumption of the theorem
= α((cB + cS )T ∆û − (cS − e)T x̄)
= 0 from an assumption of the theorem
and
If Q has full rank then any alternative optimal solution must satisfy û−v̂+tx̄ = x̂∗.
We can then show that any optimal solution to (QP ) must take the form given in
Theorem 2.
10
Theorem 3 If Q has full rank then any optimal solution to (QP ) must take the form
given in (21)–(24) and the directions must satisfy the hypotheses of Theorem 2, after
rescaling so that ∆t = −1.
Proof: Since Q has full rank, we must have ∆û − ∆v̂ + tx̄ = 0.
If ∆t = 0 then ∆û = ∆v̂, but this is impossible if ∆û and ∆v̂ also satisfy
(cB + e)T ∆û + (cS − e)T ∆v̂ = 0, since cS + cB > 0.
If ∆t 6= 0, we can assume ∆t = −1, rescaling if necessary. It follows that
∆û − ∆v̂ = x̄. We also have that (cB + e)T ∆û + (cS − e)T ∆v̂ = 0, so (cB + cS )T ∆û =
(cS − e)T x̄.
For j 6= i :
(
x̄j if v̂j∗ = 0
∆ûj = (25)
0 otherwise
(
−x̄j if v̂j∗ > 0
∆v̂j = (26)
0 otherwise
For i :
1 X
∆ûi = ((cS − e)T x̄ − (cBj + cSj )∆ûj ) (27)
cBi + cSi j6=i
∆v̂i = ∆ûi − x̄i (28)
For t :
∆t = −1. (29)
Lemma 1 If either û∗j or v̂j∗ is zero then the given direction ensures that û0j v̂j0 = 0
for any choice of α, where û0 and v̂ 0 are defined in (21) and (22), respectively.
Proof: If v̂j∗ > 0 then both û∗j and ∆ûj are zero, so û0j = 0. If v̂j∗ = 0 then ∆v̂j = 0
so v̂j0 = 0. The result follows.
Lemma 2 For the given direction, both ∆ûi and ∆v̂i are negative.
Proof: We have ∆ûj ≥ 0 for j 6= i. It follows that ∆ûi < 0 and therefore that
∆v̂i < 0.
11
Lemma 3 The given direction satisfies the conditions of Theorem 2.
Proof: For every component, we have ∆û − ∆v̂ = x̄. Further, from the definition
of ∆ûi, we have (cB + cS )T ∆û = (cS − e)T x̄.
It follows from these lemmas that the direction defined by (25)–(29) maintains
feasibility and optimality, so the point given in (21)–(24) is optimal for any α for which
it is nonnegative. Further, the direction does not introduce noncomplementarity in
any component. The maximum positive possible step length α can be found using a
minimum ratio test. If the first component to be driven to zero is either û0i or v̂i0 then
we have removed the noncomplementarity in the ith component, without introducing
noncomplementarity in any other variable. It may be that some other component v̂j0
is driven to zero first. In this case, we calculate a new direction using (25)–(29) and
repeat. If x̄j > 0 then, in the new direction, we have ∆ûj > 0 and ∆v̂j = 0 which
makes ∆ûi and ∆v̂i more negative. Further, the number of negative components ∆v̂j
is decreased by at least one. Thus, we immediately have the following theorem.
Recall that x0 is the total amount spent on transaction costs. It follows from (11)
that t = 1/(1 − x0 ) so x0 = 1 − (1/t), so reducing t will reduce the amount spent in
transaction costs. If an optimal solution w∗ to (QP ) is noncomplementary, then the
value t can be reduced using the direction given in (25)–(29). Therefore, minimizing
t while leaving x̂ unchanged will result in a complementary optimal solution. This
suggests solving the following linear programming problem.
min ∆t
s.t. −cTB ∆û − cTS ∆v̂ + ∆t = 0
(LP t ) ∆û − ∆v̂ + x̄∆t = 0
∆û ≥ −û∗
∆v̂ ≥ −v̂ ∗
Note that a feasible solution to this linear program will also satisfy (cB + e)T ∆û +
(cS − e)T ∆v̂ = 0, since eT x̄ = 1.
The procedure given above for finding a complementary optimal solution to (QP )
is not affected by the presence of a risk-free security. Of course, the optimal solution
is to place all assets in the risk-free security giving an objective function value of zero,
if this is feasible. If the risk-free security is a combination of the original securities,
then it may not be possible to achieve a risk-free portfolio, even with very low values
for E0 . Theorem 3 assumes there are no risk-free securities, so alternative optimal
solutions may exist that do not give the same value for x̂. In this situation, a direction
12
∆x̂ in the nullspace of Q may lead to alternative optimal solutions. The minimum
transaction cost solution can be found by solving the following generalization of (LP t ):
min ∆t
s.t. Q∆x̂ = 0
∆x̂ − ∆û + ∆v̂ − x̄∆t = 0
(LP rf ) T T
(cB + e) ∆û + (cS − e) ∆v̂ = 0
−cTB ∆û − cTS ∆v̂ + ∆t = 0
µT ∆x̂ − E0 ∆t ≥ E0 t∗ − µT x̂∗
A∆x̂ ≤ −Ax̂∗
∆û ≥ −û∗
∆v̂ ≥ −v̂ ∗
5 Computational Results
We discuss two portfolios in this section, a nine-security one due to Markowitz [21],
and a portfolio consisting of the thirty stocks in the Dow Jones Industrial Average.
Solving the problem (QP ) for different values of E0 will give a transaction cost effi-
cient frontier (TCEF). The relationship of the TCEF to the efficient frontier with no
transaction costs is explored in this section, as well as other properties of the TCEF.
We investigated the Markowitz 9-security portfolio with all transaction costs equal
to 5% and the initial portfolio equally divided among the securities. For the return
and risk data, see Braun [7]. Figure 2 shows which securities are involved at each
point along the TCEF. The vertical scale is the minimum required expected return
(ie. the horizontal coordinate) of the efficient frontier plots.
Figure 2 contains a great deal of useful information which can be accessed by visual
inspection. For example, looking down a column, changes occurring in the portfolio
as the required expected return is decreased can easily be seen. Also, horizontal
comparisons between panels allow for comparisons to be made between the no-cost
and c=5% frontiers. Note that the highest returning portfolio on the no-cost frontier
has no counterpart on the c=5% TCEF. Likewise, the least risky-least returning
point on the c=5% TCEF has no counterpart on the no-cost frontier. This is a
manifestation of the leftward shift seen in Figure 1. With no transaction costs the
highest returning portfolio is always full investment in the single highest yielding
individually efficient security. This is security #5 for the Markowitz dataset. However
with the introduction of nonzero transaction costs, that is no longer the case.
A general observation is that the portfolios along the TCEF are not simply related
to the portfolios along the no transaction cost efficient frontier. Sometimes, entirely
new securities are involved. Sometimes, buy and sell decisions are reversed. So
Figure 2 highlights that the introduction of costs changes the portfolio rebalancing
problem dramatically and that the optimal solutions are also quite different.
We applied our solution strategy to the problem of rebalancing portfolios com-
posed of the 30 stocks which currently make up the Dow Jones Industrials Average.
All securities were involved initially, with proportions varying from 1% to 5%. The
13
High
High
<−−−− Expected Return −−−−>
Low
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Securities involved: Securities involved:
c=0% Frontier TCEF
Figure 2: These panels show which securities are involved along the no-transaction
cost efficient frontier (left) and the c=5% TCEF (right). If no symbol is present, then
a security was liquidated. The symbols 4 and 5 represent buy and sell decisions
that did not result in liquidation of the security. The symbol 2 represents a security
in which the position is unchanged.
buying and selling costs varied from security to security, from 0% to 5%. For the val-
ues of the parameters x̄, cB , cS as well as for the return and risk data, see Braun [7].
In Figures 3 and 4, we present visualizations of the optimal portfolios for the no
transaction cost efficient frontier and the TCEF.
6 Conclusions
The results of this paper will allow the incorporation of transaction costs into portfolio
optimization problems, in a manner that leads to intuitive and sensible allocations.
The model calculates the risk of the resulting portfolio, weighted by the amount
invested after paying transaction costs. The model can be formulated as a quadratic
programming problem of size comparable to the model with no transaction costs, so
14
Securities involved: c=0% Frontier
High
<−−−− Expected Return −−−−>
Low
0 5 10 15 20 25 30
Security Number
Figure 3: Involved Securities for the Dow Jones Portfolio: This figure shows which
securities are involved along the c=0% frontier.
References
[1] C. J. Adcock and N. Meade. A simple algorithm to incorporate transactions costs
in quadratic optimization. European Journal of Operational Research, 79(1):85–
94, 1994.
[2] S. Albeverio, L. J. Lao, and X. L. Zhao. On-line portfolio selection strategy with
prediction in the presence of transaction costs. Mathematical Methods of OR,
54:133–161, 2001.
15
Securities involved: TCEF
High
<−−−− Expected Return −−−−>
Low
0 5 10 15 20 25 30
Security Number
Figure 4: Involved Securities for the Dow Jones Portfolio: This figure shows which
securities are involved along the TCEF. Here, overlay comparisons with Figure 3 are
possible.
16
[7] S. E. Braun. Solving a quadratic programming problem subject to orthogonality
constraints. PhD thesis, Mathematical Sciences, Rensselaer Polytechnic Insti-
tute, Troy, NY, December 2001.
[8] T.-J. Chang, N. Meade, J. E. Beasley, and Y. M. Sharaiha. Heuristics for cardi-
nality constrained portfolio optimisation. Computers and Operations Research,
27:1271–1302, 2000.
[10] V. K. Chopra and W. T. Ziemba. The effect of errors in means, variances, and
covariances on optimal portfolio choice. In W. T. Ziemba and J. M. Mulvey, edi-
tors, Worldwide asset and liability modeling, pages 53–61. Cambridge University
Press, 1998.
[11] A. Consiglio and S. Zenios. Designing portfolios of financial products via inte-
grated simulation and optimization models. Operations Research, 47(2):195–208,
1999.
[14] S.-E. Fleten, K. Høyland, and S. W. Wallace. The performance of stochastic dy-
namic and fixed mix portfolio models. European Journal of Operational Research,
140(1):37–49, 2002.
17
[20] H. Markowitz. Portfolio selection. Journal of Finance, 7:77–91, 1952.
18