A Simple Algorithm for Optimal Portfolio Selection with Fixed Transaction Costs Author(s): Nitin R. Patel and Marti G. Subrahmanya Source: Management Science, Vol. 28, No. 3 (Mar., 1982), pp. 303-314 Published by: INFORMS Stable URL: http://www.jstor.org/stable/2630883 . Accessed: 06/10/2011 05:25 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range ofcontent in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected]. INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science. http://www.jstor.org
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
8/2/2019 A Simple Algorithim for Capturing Portfolio
A Simple Algorithm for Optimal Portfolio Selection with Fixed Transaction Costs
Author(s): Nitin R. Patel and Marti G. SubrahmanyaSource: Management Science, Vol. 28, No. 3 (Mar., 1982), pp. 303-314Published by: INFORMSStable URL: http://www.jstor.org/stable/2630883 .
Accessed: 06/10/2011 05:25
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .http://www.jstor.org/page/info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.
A SIMPLE ALGORITHM FOR OPTIMAL PORTFOLIOSELECTION WITH FIXED TRANSACTION COSTS*
NITIN R. PATELt AND MARTI G. SUBRAHMANYAMt
The general optimal portfolio selection problem witlhfixed transaction costs is a complex
mathematical programming problem. However, by placing reasonable restrictions on thevariance-covariance matrix of returns, it is possible to simplify the solutioniof the problem.
Specifically, if the structure of returns between securities is such that the pairwise correlation
coefficients are approximately the same, a fairly simple algorithm which requires little
computational effort can be employed. This method can also be extended to.the case where
changes in the information set necessitate a revision of an existing portfolio.
The pioneering work of Markowitz ([13] and [14]) in portfolio analysis has served as
the basis for the development of much of modern financial theory. Two directions of
enquiry have been pursued in the literature:one, the development of modern portfoliotheory to optimize selection of portfolios under different scenarios and the other, the
derivation of a theory of pricing of capital assets under uncertainty. One basic
implication of most models in portfolio theory is that investors are well diversified in
their holdings of securities. However, this implication of portfolio theory does not
accord with the observed holdings of securities by investors, who typically hold a small
numberof securities.'While there are alternative explanations of this phenomenon, an important factor is
the presence of transaction costs, broadly defined to include both direct costs such as
brokeragecommissions as well as the costs of analyzing securities, which are ignored in
much of the theory. Though some attention has been focussed on the question of
variable transaction costs, (see for example, Pogue [16], and Chen, Jen, and Zionts [3]),the question of undiversified portfolios remains unresolved. Variable transaction costsrender individual securities less desirable but do not inhibit the addition of more
securities to a portfolio. Fixed transaction costs such as odd-lot commissions do
provide an explanation for the phenomenon of reduced portfolio size but this market
imperfection has not been modelled in the literature, by and large. Mao [12], Jacob [9],and more recently, Levy [10] examine the fixed transaction costs problem indirectly byplacing restrictions on the number of securities in the optimal portfolios. However,
they do not explicitly relate the optimal number of securities to the transaction costs,and provide only a general rationale for the restriction.
The only detailed analysis of the fixed transaction costs problem in the portfolioselection literature is by Brennan [2].2 Assuming that the structure of the Sharpe [18],
*Acceptedby MartinJ. Gruber; eceivedMarch1980.Thispaperhas beenwiththeauthorI month or I
revision.1i ndianInstitute f Management,Ahmedabad.INew YorkUniversity.'In a study by Blume, Crockett and Friend [1] for the tax year 1971, it was foulld that investors hold
relatively undiversified portfolios. Based on a sample of individual income tax returns, it was computed that,
among dividend paying stocks, 34 per cent held only one stock, 50 per cent held no more thanitwo anid only
I I per cent listed more thani ten stocks.
2Stapleton and Subrahmanyam [19] provide an analysis of the effect of fixed transactioni costs on
equilibriulll prices. However, they are concerned more with asset prices with this market imperfection and
do not deal with the optimal portfolio selection problem directly. They therefore use an enumilerationi ethod
to solve the portfolio problem in order to determine equilibrium prices.
303
0025-1909/82/2803/0303$0 1.25Copyrigiht 0) 1982, The Itistittite of Management Scietnces
8/2/2019 A Simple Algorithim for Capturing Portfolio
Lintner [11] and Mossin [15] capital asset pricing model holds even under fixed
transactions costs, Brennan presents a model for determining the optimal number ofsecurities in the portfolio. Initially, he assumes that all securities have the samesystematic risk and the same residual variance, and hence, in equilibrium, the same
expected return, and derives the optimal number of securities. Subsequently, he relaxes
the assumptions of identical residual variance, and through the capital asset pricing
model, identical systematic risk.There are several difficulties with the Brennan formulation. Firstly, he needs to
assume that the capital asset pricing model holds. While this may be a reasonable
approximation, this pricing relationship is itself affected when several investors operate
under fixed transaction costs as shown by Levy [10]. More importantly, this begs the
question of superior access to information and consequently different estimates of the
parameters of security returns, which imply that the standard capital asset pricing
model may not be relevant to that investor. Indeed, if the capital asset pricing modelwere always valid, portfolio selection models are of dubious merit, except for the
residual variance component which may not be significant, even with a relatively small
number of securities. Secondly, the Brennan model focusses on the number of
securities rather than on individual candidate securities themselves. Hence, the attrac-tiveness or otherwise of an individual security cannot be examined except in the trivial
sense that securities with lower residual variance are preferred, ceterisparibus. Finally,
the computation of the optimal number of securities is quite complicated, an unattrac-
tive feature of many portfolio selection models from the viewpoint of practicing
portfolio managers.The general optimal portfolio selection problem with fixed transaction costs is a
complex mathematical programming problem. But, by placing reasonable restrictionson the variance covariance matrix of returns, it is possible to simplify the problem.Interestingly, the solution of the problem can be reduced to a fairly simple algorithm
involving very little computational effort. In ?2, the model is discussed and a
simplified form of the objective function is presented. The next section derives the
algorithmand proves that the simple rules derived are optimal. A numerical example is
worked out and explained in ?4. ?5 deals with the case where changes in the
information set and/or the wealth constraint necessitate a revision in the portfolio. In
this case, the relationship of additional securities to the existing portfolio has to be
taken into account.
2. The Model
The portfolio selection problem of the investor in a mean-variance context with
fixed transaction costs can be written as:
11 s1 17 n
Max xiRi + xoR- ai x x1sis1p1- t Yii= Z1 i1 j=1 i =1 (1)
subject o xi+?xO=Wi=l
where the xi are unrestricted in sign, yi = 0 or 1, and xi # 0 only if yi = 1. The
definition of the variables are as follows:= 1 if security i is part of the portfolio; i = 1,2, . .. , n.
= 0 otherwise.
xi = amount invested in (risky) security i; i = 1,2, ... , n.
xo= amount invested in the riskless security.
Ri= 1?
expected rate of return on security i; i =1, 2, . . ., n.
8/2/2019 A Simple Algorithim for Capturing Portfolio
OPTIMAL PORTFOLIO SELECTION WITH FIXED TRANSACTION COSTS 305
R = 1 + rate of return on the riskless security.
a = risk aversion factor of the investor (a > 0).
Si= standard deviation of the rate of return on security i; i = 1,2, . , n.
pij = correlation coefficient between the returns on securities i and j; i, j = 1,
2, . . .,n.
t = fixed transaction cost incurred for each risky security included in the portfolio.
This cost is assumed to be incurred at the end of the period. Equivalently, the fixed
transaction cost per security at the beginning of the period is equal to t/R.
W = initial wealth of the investor.The first term refers to the mean end of period cash flow to the investor from
holdings of risky assets. The next term relates to the return from the riskless asset. The
third term is the variance of end of period cash flow, while the last term refers to the
fixed transaction cost. The above formulation appears to be similar in form to the
standard Markowitz model with a riskless asset, except for the term involving the fixedtransaction cost.3 However, there is a basic difference; in this model, the computation
of the mean return and the variance of the return of the portfolio takes into account
only those securities that are in the portfolio rather than all the securities available in
the market. This aspect of the problem complicates it considerably, since it becomes a
mixed-integer quadratic programming problem for which no simple solution method
exists. At this point, there are two possibilities. The first would be to simplify the
structure of the model by placing some restrictions on the parameters. The other
method would be to obtain an approximate solution to the problem as formulated. We
have chosen the former alternative.One possible restriction on the structure of returns is that all pairwise correlation
coefficients between security returns are equal. While this assumption is difficult to
justify on any conceptual grounds, it has been shown to be a reasonable assumption,
empirically. In a study of alternative methods of forecasting the correlation structure
of returns, Elton and Gruber [4] demonstrate that this assumption produces better
estimates of the future correlation structure of security returns than do correlation
coefficients based on historical returns which may differ across pairs of securities, the
most general specification possible. Surprisingly,another intuitively appealing simplifi-
cation, the single index approximation of Sharpe [17] does not perform as well as the
constant correlation assumption. In the Elton and Gruber study, the assumption of aconstant correlation coefficient produced forecasts that were more accurate than nine
other specifications that were tested. In a more recent study by Elton, Gruber, and
Urich, [8], the forecasting performance of the constant correlation coefficient assump-
tion is tested using a variety of statistical procedures. The performance of this
assumption is consistently superior to that of all the alternatives in each of the
statistical tests performed. The constant correlation coefficient assumption will be usedin the rest of this paper.4
Suppose we assume that
pUj= p for all i, j; i #/ j (p > 0).
3In this formulation, we have chosen to model the fixed transaction cost as leading to a reductioll in the
mean end of period caslhflow. An alternative formulation would be to treat it as a reduction in the resources
available for investment, with the cost appearing in the wealth constraint. As will be clear in the analysis
later on, the two approaches are formally equivalent.
4The fixed transaction costs problem using the simplified structure suggested by the single index model of
Sharpe [17] will be analyzed in a subsequent paper. However, it turns out that the solutioni technique for this
case is more complex than the case studied here. The structureof the single-index model does not permit the
decomposition analyzed in this paper, since the interactions due to the integer constrainitsbecome more
difficult to handle.
8/2/2019 A Simple Algorithim for Capturing Portfolio
OPTIMAL PORTFOLIO SELECTION WITH FIXED TRANSACTION COSTS 307
where cj = (R - R)/sj represents the excess return of securityj over the riskless rate
of return as a ratio of the standard deviation of the return of the security.
Multiplying (3) by Xj,adding over j E S, and rearranging,we get
SiS ~~~(RjR)p 2 sjs~xix~?j+I -p) s]xj2=i~~Sj~~S j~s, jx S 2a 7.(8)e Sje s y es ye s
Substituting for x0 in the objective function, using the wealth constraint, and using (8),
the objective function becomes
n
2, (Rj - R)xjl2- tm + WR. (9)j=1
Substituting for xj from (7) the objective function expressed in terms of y. becomes
4a(1l_-p) 1 -' Yj- [(m- ? I - tm - WR. (10)
For fixed m, since a, p, t, w, R are constants, our problem is reduced to
Max cj2y1
j=ll [(in-I)p + 1](J
subject to yj =0,l and yj = m.j=l
We assume that the securities have been ranked so that cl > C2 . .. > c,7- Note if
two securities have the same ci, they may be treated as a single security in our analysis,
so that the strict inequalities above do not imply any loss of generality.
3. The Algorithm
We are now in a position to derive a simple algorithm for portfolio selection and
show that it is optimal.
THEOREM 1. Let S be the set of optimal securities for a given number of securitieschosen, m, m >.2 (i.e., ISj = m and S = (k IYk = 1}). Then there is noj E S such that
both] - 1 andj + 1 X S.
We argue from contradiction. Suppose there is such a j. Let S' = S - (j}. Let
OFV ( Q) denote the objective function value for any set of indices Q.
OFV (S)= 1.c7 - Ec1)S iES [(m-I)p + 1(is /
- Ec - ( ( E ci) (11)
-S[(i l)p 1][ [ilM )p? 1] cj* (12)
OFV(S'U{]-1})= cI
2pcki v [(m-2)p + 13
8/2/2019 A Simple Algorithim for Capturing Portfolio
[ (m-1)p + 1 (jP -lcj) iE c - [ (m -2)p + I ] (CJ_lc /-) > O
or
(ci- I - c.) t2p ci - [(m - 2)p + I](cj-, + cj)) > O (14)
and since cj- l - cj > 0, the expression within the curly brackets above (call it (a)) must
be greater than or equal to zero.
A similar argument using OFV(S) - OFV(S'U{j ? 1}) > 0 leads to
(cj+I-cj){2p E ci-[(m-2)p + (cI + c) > 0
i E-sI
and since cj+I - cj < 0, the expression within the curly brackets above, (call it (b)),
must be less than or equal to zero.
Subtracting (b) from (a) we have
[ (m -2)p + I I(Ci+ I- ci- ) >1 .
For m > 2, the first factor above > 0 which implies cj+I > cl which is a contradic-
tion. Q.E.D.We conclude from Theorem 1 that for a given m, the optimal S is of the form
{1, 2, ... , u, 1, + 1, . . ., n} where u + n-I + = m. In fact, an optimal portfoliocan be represented by the single number u, since 1 = n - m + u + 1. We shall
therefore say that an optimal portfolio is u* to mean that the securities in the portfolio
are {1,2, . . ., u*,l*,l* + 1, . . ., n} where l*= n-m + u* + 1; if u* = m, l* = n + 1
and the portfolio represented in this case will be {1,2, . .. , m}.
THEOREM . Let u,*, U*1+ be optimalportfoliosof sizes m and m + 1 respectively.Let
1 = n-rn +?u* + 1 and i+H = n (m + 1) + u*1 + 1 =n-rm + u* + . Then u* +
1 u,+ > U,* if p > 0 and c1 > 0.
(i) We first show u,*7 1 > u,s*+. We argue from contradiction; suppose u,*+?
? u* ?2.
Since u,* is optimal, increasing 1/*7and u,* both by 1 cannot increase the objective
The analysis presented in the previous section deals with the case where an investoroptimizes his portfolio holdings of risky securities with fixed transaction costs under
the assumption that the investor's initial wealth is in the form of cash. While thisaccurately describes the initial portfolio problem of the investor, it does not properlyreflect the costs of changing the optimal holdings over time. Specificially, there may be
a reluctance on the part of the investor to decrease the holding of a particular securityin light of adverse information on, say, its expected return since there are fixed
transaction costs incurred in selling this security as well as investing the proceeds in
one or more other securities. It may be reasonably hypothesized that the optimalholdings of securities in the portfolio after revision will be biased in favour of securities
that are initially included in the portfolio.
It turns out that the portfolio revision problem can be modelled with minor changesin the earlier analysis. Suppose xi is the amount already invested in the security i (thepresent portfolio) and di is the change in this amount (the new portfolio). The
maximization problem in equation (2) can be modified as
n fl
Max (xi +di)Ri+ (xo + dO)R t yidi i=Ii I
n n1s(afl 12 x 1(2+dyss l-)E(x js | (5
n
subject to Ed + do= W'
where W' is the additional amount available for investment and with the proviso that
Yi= 1 if di #- 0, i = 1, 2, . .. , n. It should be noted that W' could be zero, positive or
negative and does not impose any restrictions on the problem. If the problem is one of
portfolio revision merely due to changes in the information set, W' would be zero; if
there are additions to or redemptions from the portfolio, from other income or due toconsumption needs, W' would be positive or negative respectively.
The portfolio problem can be rewritten by grouping separately the terms involvingthe di's, the decision variables, in (15).
n n
Max d1Ri+ doR-t ydi= I
-a tpL (dxjsisj+ d)x1s.s1+ didsisj) + (I p) 2dixis2 + d2SJ2j + K
n
subject to d + do= W',i = I
Y= I if di::--0, i= 1,2, .. n, (16)
where
12 n 1 1 A
K xiRi+ xoR7-a p) Xi2si2
;~~~~~i_ss i _ 1) ;
8/2/2019 A Simple Algorithim for Capturing Portfolio
OPTIMAL PORTFOLIO SELECTION WITH FIXED TRANSACTION COSTS 313
which is unaffected by the portfolio revision. Equation (16) can be rewritten as
Max,1d,[R - 2aps(1 xs- 2a(1 -p)s,xij
1 r 11 11 11 A
+ doR-t fi-af p~ 2 d,dj,s1sj+ 1 -p) 2 aj2s7 9(17)
subject o E= 4j+d '1=1
y= 1 if d,i#0, i =1,2, ... , n.
This is of the same form as the original problem in (2), except that
= [R - 2apsi( x - 2a(I - A),2x1
The rest of the analysis would be similar and the same procedure as before can be used
to solve for the optimal d1's.Intuitively, the only difference between the initial problem
and the portfolio revision problem is that the interactions of each security with the
existing portfolio as reflected in the definition of R,' have to be taken into account in
the latter case. It pays to revise the existing holdings of a particular asset only if the
round trip transaction costs can be recouped by revising the portfolio.
6. CoIiclusions
This paper derives simple decision rules for optimal portfolio selection under fixed
transaction costs. Assuming that the pairwise correlation coefficient is the same across
all securities, it is shown that the optimal portfolio can be chosen without directly
solving the complex mixed-integer quadratic programming problem. The interesting
feature of the model is that securities can be ranked on the basis of a simple index, theratio of the excess returnto the standard deviation of return of the security. Thus, even
fairly large problems can be solved by hand, as convergence to the optimal solution is
rapid. Even for cases where the assumption of equal correlation coefficients is notvalid, perhaps an intuitive understanding to the solution of a fairly complex problem
can be gained. The portfolio revision problem is shown to be similar in structureto theproblem of the portfolio optimization starting with cash holdings, the difference being
that in the former case, round trip transaction costs lead to a reluctance to revise the
portfolio unless the gains are correspondingly greater.
References
1. BLUME, M. E., CROCKETT, J. AND FRIEND, I., "Stock Ownership in the United States: Characteristics
and Trenids," Surveyof ClurenitBusiness,Vol. 54, No. 11(November 1974), pp. 16-40.2. BRENNAN, M. J., "The Optimal Numllber f Securities in a Risky Asset Portfolio WheniThere Are FixedCosts of Transacting: Theory and Some Empirical Results," J. FinianicialQlanititativeAnial.,Vol. 10,No. 3 (September 1975), pp. 483-496.
3. CHEN, A. H. Y., JEN, F. C. AND ZIONTS, S., "The Optimal Portfolio RevisioniPolicy," J. Blusiniess, ol.44, No. 1 (January 1971), pp. 51-61.
4. ELTON, E. J. AND GRUBER, M. J., "Estimating the Depenidenice Structurle of Slhare Pr-ices," J. Finance,
Vol. 28, No. 5 (December 1973), pp. 1203-1232.5. AND PADBERG, M. W., "Simple Criteria for Optimal Por-tfolio Selection", J. Finance,
Vol. 31, No. 5 (December 1976), pp. 1341-1357.6. , AND , "Simiple Criteria for Optimal Portfolio Selectioni With Upper Bounlds,"
OperationsRes., Vol. 25? No. 6 (November-December 1977), pp. 952-967.
8/2/2019 A Simple Algorithim for Capturing Portfolio