Asset allocation models

in discrete variable

Marius Rădulescu
Institute of Mathematical Statistics and Applied Mathematics, Casa Academiei Române, Calea 13
Septembrie nr.13, RO-050711, Bucharest 5, ROMANIA,
e-mail: [email protected]

Constanţa Zoie Rădulescu

National Institute for Research and Development in Informatics, 8 -10 Averescu Avenue, 011455,
Bucharest 1, ROMANIA, e-mail: [email protected],

Gheorghita Zbăganu
Faculty of Mathematics and Computer Science, University of Bucharest, Academiei 14, Bucharest,
RO-010014, ROMANIA, email: [email protected]

Abstract: In the classical portfolio selection theory the value of the assets is considered infinitely divis ible. In the real
portfolio selection models one should consider only finitely divisible assets. This is because the investors purchase only
a finite number of shares or minimum transaction lots. We present several asset allocation models in discrete variable
and we make an analysis of the results. Our models are closer to reality but they are more difficult t o be solved.

Keywords: portfolio selection, asset allocation, finitely divisible assets, minimum transaction lots, integer
programming model

Dr. Marius Radulescu graduated Faculty of Mathematics in 1977. He took his Ph.D. degree in 1985 at Centre of
Mathematical Statistics “Gheorghe Mihoc” in Bucharest. In 1991 he was awarded a prize of the Romanian Academy.
At present he is a senior research worker at Institute of Mathematical Statistics and Applied Mathematics Gheorghe
Mihoc-Caius Iacob in Bucharest. He published several books and research papers in the area: nonlinear functional
analysis and its applications to boundary value problems for different ial equations, real analysis, numerical analysis,
optimization theory, approximation theory, mathematical modeling (operations research) , portfolio theory. He is a
member of the Editorial Board of Journal of Applied Sciences , Advanced Modeling and Optimization and Arhimede.

Dr. Constanta Zoie Radulescu graduated Faculty of Mathematics in 1977. She is a senior research worker at National
Institute for Research and Development in Informati cs in Bucharest. She has got a Ph.D. degree at Centre of
Mathematical Statistics in Bucharest. Her scientific interests are: DSS for the management of financial investments,
multicriteria decision analysis, risk analysis, mathematical modeling (operations research).

Prof. dr. Gheorghita Zbăganu graduated Faculty of Mathematics in 1975. He took his Ph.D. degree in 1986 at Centre
of Mathematical Statistics “Gheorghe Mihoc” in Bucharest. In 1999 he was awarded a prize of the Romanian Academy.
At present he is a senior research worker at Institute of Mathematical Statistics and Applied Mathematics Gheorghe
Mihoc - Caius Iacob in Bucharest and a professor at the Faculty of Mathematics and Computer Science, University of
Bucharest He published several books a nd research papers in the area: probability theory, ruin theory, actuarial science,
information theory, combinatorics, optimization theory, functional analysis .

1. Introduction

The original Markowitz model of portfolio selection has received a widespread theor etical
acceptance and it has been the basis for various portfolio selection techniques. The model is known

1
in the literature also, as the mean -variance portfolio selection model. Generally, in the classical
mean-variance portfolio selection approach severa l realistic features are not taken into account.
Among these ”forgotten” aspects, one of particular interest is the not infinite divisibility of the
financial asset to select, i.e. the obligation to buy/sell only integer quantities of asset lots which
contain a predetermined number of shares.
A fundamental question in the mathematical finance is how risk should be measured properly. The
mean-variance models behave well in the case when the distribution of the random vector of returns
is close to a multivariate normal distribution, and not so well in other cases. An important problem
is to build portfolio selection models based on risk measures that capture risk adequately. In 1952,
Roy [14] suggested that investors are interested in selecting a portfolio so as to maximize the
probability of achieving at least a given return. The idea is that if return falls below the threshold
there will be some bad consequence. This model of investor behavior is called safety- first. Drawing
the efficient frontier in standar d deviation - expected return space, the portfolio which maximizes
the probability of realized return being greater than the threshold, can be found by identifying the
straight line passing through the expected return axis at the threshold that is tangent to efficient
frontier. The portfolio at the point of tangency is the desired portfolio.
The development of the theory of stochastic dominance had stimulated the research for
asymmetric risk measures (downside risk measures) like: shortfall probability, expectations of loss,
semi-variance and lower partial moments. An important class of risk measures considered in the
literature is the coherent risk measures [1] , [13].
In order to consider such a feature, we build several nonlinear integer programming pr oblems. The
present paper continues the ideas from [12]. The risk measure considered in this paper is the lower
partial moment of the first order . We propose a formulation of this problem in terms of quantities,
i.e. integer numbers of asset lots to buy, i nstead of starting capital percentages to invest. We give
necessary and sufficient conditions for the existence of feasible solution(s) with a sequence of steps
to be followed by the investor when he wants to make investment decisions in the presence of
minimum transaction lots. A numerical example illustrating the previous points is presented.
The portfolio selection problems with discrete constraints (such as buy-in thresholds, cardinality
constraints and transaction roundlot restrictions) were studied in the literature mainly in the last
decade. Integer programming approaches in the mean-risk models were studied in [2]-[13], [15],
[16]. Mansini and Speranza in [ 8] consider the constraint stating that assets can be traded only in
indivisible lots of fixed size. In this case, the problem is formulated in terms of integer -valued
variables — as opposed to real-valued ones — that represent, for each asset, the number of
purchased lots, instead of the real -valued ones.
Given that assets are normally composed by units, this constraint is certainly meaningful; its
practical importance however depends on the ratio between the size of the minimum trading lots
and the size of the shares involved in the portfolio. In [3] Corazza and Favaretto study the existence
problem for the solutions of a discrete mean–variance portfolio selection model.
Mansini and Speranza in [ 9] consider a single-period mean-safety portfolio selection problem with
transaction costs and integer constraints on the quantities selected for the secu rities (rounds). They
propose an exact approach based on the partition of the initial problem into two sub -problems and
the use of a simple local search heuristic to obtain an initial solution.
In [5], [8], [10] and [16] are presented heuristic algorithms for the portfolio selection problem with
minimum transaction lots.

2
In [2] is presented an approach of the mixed-integer nonlinear mean-variance portfolio selection.
The authors proposed an algorithm (which is based on the proposed conditions) for finding a
“good” feasible solution and proved its convergence.
In [7] the classic mean-variance framework is extend to a broad class of investment decisions under
risk where investors select optimal portfolios of risky assets that include perfectly divisible as we ll
as perfectly indivisible assets. The author develops an algorithm for solving the associated mixed -
integer nonlinear program and report on the results of a computational study.
In [4] are examined the effects of applying buy -in thresholds, cardinality constraints and transaction
roundlot restrictions to the portfolio selection problem. Such discrete constraints are of practical
importance but make the efficient frontier discontinuous. The resulting quadratic mixed -integer
(QMIP) problems are NP -hard and therefore computing the entire efficient frontier is
computationally challenging. The authors proposed alternative approaches for computing this
frontier and provided insight into its discontinuous structure.

2. Mean-risk models with minimum lot constrain ts

Denote by hi the price for the minimum lot of asset i. In case that there are no minimum lots, then
hi is the price of a share of asset i . The integer variable xi represents the number of the minimum
lots for the asset i. Let a particular portfolio be defined by a vector x  ( x1 , x 2 ,..., x n ) T  Z n
where the integer variable xi represents the number of the minimum lots invested in the asset i.
Let the assets returns be represented by a vector of random variables   (1 ,  2 ,...,  n ) T with
means   ( 1 ,  2 ,...,  n ) T . Denote H=diag(h), that is H is the diagonal matrix whose diagonal is
equal to vector h. Consider the vector d= d1 , d 2 ,  , d n  , d i  hi  i for all 1  i  n . Denote
T

x i  hi x i , 1  i  n and consider the vector x   x1 , x 2 ,  , x n  . Then xi is the sum invested by

T

the investor in asset i.

Denote by t  the positive part of the real number t, that is:
t t
t   max t ,0  .
2
The risk of the investment in portfolio x with respect the target  is defined as the lower partial
moment of the first order of the return ξ T x , that is LPM  x   E   ξ T x   . Note that the
average return is E ξ T x   μ T x  d T x . Let W be the minimum level of the expected return desired
by the investor and M1 and M2 be the upper and lower bounds for the investment sum.
Then the purpose of the minimum risk portfolio selection problem with minimum lot constraints is
to find that value of x that will
minimize E    T x  
( P1 ) 
 M 1  h x  M 2 , d x  W , x  Z and x  0
T T n

The purpose of the maximum return portfolio selection problem with minimum lot constraints is to
find that value of x that will
maximize (d T x)
( P2 ) 
 M 1  h x  M 2 , E    x    r , x  Z and x  0
T T n

Here r is an upper bound for the risk accepted by the investor.

3
In case the risk aversion of the investor is taken into account then we can formulate the trade-off
mean-risk portfolio selection problem:
minimize (1   E    T x     d T x)
P3  
 M 1  h x  M 2 , x  Z and x  0
T n

Here   0,1 is the coefficient of risk aversion. The case when  takes small values corresponds
to a risk averter investor who is more worried about below average returns tha n attracted by the
above average gains.
The case when  takes great values corresponds to a risk -lover investor which is more worried
about the above average gains than attracted by the below average returns. By varying  from 0 to
1 all efficient portfolios can be determined.
In the problems (P 1), (P2) and (P 3) the restriction x  0 can be replaced with the more realistic
restriction 0  x  a . The upper bounding constraints exist for legal, personal or institutional
reasons. Some components of the vector a may be infinite.
Consider a time horizon composed of several moments t = 1,2,…,m. Denote by R  rti  the m  n
matrix of historical rates of retu rns of the assets. rti is the rate of return at moment t for asset i.
Consider the matrix R  rti  , R  R H . Note that rti  rti hi for every t  1,2,  , m ,
i  1,2,  , n . In the following we shall use the following estimates :
m

r ti
ci  t 1
is an estimation for the expected return  i of asset i,
m
m

r h
t 1
ti i
is an estimation for the entry d i of the vector d,
m
m n

r h x ti i i
t 1 i 1
is an estimation for the expected return E ξ T x  of the portfolio x
m
1 m  
rti xi  is an estimation for the risk E    T x   of the portfolio x
n

 
m t 1 
  
i 1 
Denote c  c1 , c 2 ,..., c n  . Taking into account the above estimations we shall associate to the
models (P1), (P2) and (P 3) the following models:
 1 m  n

( P1)
minimize

 
m t 1 
  
i 1
rti xi 

 M  h T x  M , c T x  W , x  Z n and x  0
 1 2

maximize (c T x)

M1  h x  M 2 ,
T

 m
( P2 ) 1  n

m      rti xi   r ,
 t 1  i 1 
x  Z n and x  0

4
  1 m  n

 minimize ( 1         rti xi    c T x)
P3  m t 1  i 1 
 M  h T x  M , x  Z n and x  0
 1 2

By addition of supplementary variables y t , t  1,2,  , m to the above models we can show that
these models are equivalent to the following mixed-integer linear models:
 1 m
 minimize  yt
m t 1

 n
( P1)     rti xi  y t , t  1,2,  , m
 i  1

M 1  h T x  M 2 , c T x  W ,

x  Z , y  R and x, y  0
n m

maximize (c T x)
 n
   rti xi  y t , t  1,2,  , m
 i 1


( P2 )  M 1  h T x  M 2 ,
1 m
  yt  r ,
 m t 1
x  Z n , y  R m and x, y  0

  1 m 
 minimize  1     yt   c T x
  m t 1 

P3    rti xi  yt , t  1,2,  , m
n

 i 1
M1  hTx  M 2 ,

x  Z n , y  R m and x, y  0

3. The structure of the investment p rocess for the minimum risk model

The structure of the investment process is presented in the following.

Step 1. The investor chooses a set S of assets where he intends to invest his money and a sum of
money M2 to be invested in the assets.
Step 2. The investor gathers information (historical data) from the stock market and forecasts from
financial experts about the chosen set of assets. As a result he obtains a matrix R  rti  where rti is
the rate of return at moment t for asset i. He finds also the vector h of minimum transaction lots of
the assets from S.
Step 3. Starting from the matrix R and taking into account some statistical hypotheses about the
returns vector, the investor computes estimates for the vector  of mean returns.
Step 4. Starting from M2 the investor computes
M 0  maxh T x : h T x  M 2 , x  Z n , x  0

5
Step 5. The investor chooses a lower bound M1 for the sum h T x invested in the assets from S in
the range 0 , M 0  .
Step 6. The investor computes
W1  mincT x : M 1  h T x  M 2 , x  Z n , x  0
W2  maxc T x : M 1  h T x  M 2 , x  Z n , x  0
Step 7. The investor compares W2 , that is the greatest mean rate of return to the risk -less rate of
return and makes a decision.
In case he decides not to make an investment in the assets from S then he returns at step 1. and
chooses another set of assets S’.
In case he decides to make an investment in the assets from S then he goes to the next step.
Step 8. The investor chooses a lower bound W, for the return of his investment, in the range
W1 , W2  .
Step 9. The investor computes
n 
1  min rti xi : t  1,2,  , m, M 1  h T x  M 2 , c T x  W , x  Z n and x  0
 i 1 
 n

 2  max rti xi : t  1,2,  , m, M 1  h T x  M 2 , c T x  W , x  Z n and x  0
 i 1 
and chooses  in the range 1 ,  2  .
Step 10. The investor solves model ( P1 ) and find an optimal portfolio x  . In case the investor
wants to continue his research he can start again from step 1. If this is not the case then the investor
stops.

Numerical results
In the following we shall illustrate how the minimum risk model works. We shall consider n = 7
assets and m = 8 moments of time (the years 2000,2001,…,2007). Take M 2  1000000 euros. The
entries of the matrix R= rti  of historical rates of return are displayed in table 1.

Year Asset 1 Asset 2 Asset 3 Asset 4 Asset 5 Asset 6 Asset 7

2000 0.30 0.30 0.55 0.36 0.34 0.39 0.30
2001 0.54 0.26 0.44 0.31 0.30 0.26 0.44
2002 0.25 0.43 0.51 0.32 0.34 0.43 0.51
2003 0.20 0.68 0.61 0.52 0.24 0.23 0.61
2004 0.10 0.53 0.61 0.52 0.54 0.53 0.51
2005 0.13 0.44 0.71 0.68 0.66 0.25 0.21
2006 0.17 0.34 0.51 0.48 0.56 0.55 0.51
2007 0.30 0.34 0.50 0.47 0.55 0.56 0.51

Table 1 Matrix R= rti  of historical rates of return.

The prices of the minimum lot of assets, that is the components of the vector h, are h1  10 euros,
h2  7 euros, h3  5 euros, h4  2 euros, h5  12 euros, h6  4 euros, h7  13 euros.
One can easily see that M 0 = 1000000 euros. The investor chooses M 1 =500000 and computes the
range for the parameter W. He finds W1 =124380 euros and W2 =450000 euros. Then the investor

6
chooses W=230.000 euros and computes the range for the parameter  . He finds 1 = 92463 euros,
 2 =659000 euros. We divide the range 1 ,  2  in ten equal intervals i , i  1 , i = 1,2,…,10. The
composition of the optimal portfolios for the minimum risk problem P1 , for various values of
parameter  are displayed in table 2.

Asset 1 Asset 2 Asset 3 Asset 4 Asset 5 Asset 6 Asset 7

1 = 92463 92463 0 0 0 0 0 0

2 =149116,7 86074 0 0 0 1478 0 9347

3 =205770,4 75960 0 0 0 20033 0 0

4 =262424,1 60234 0 0 0 22660 0 9672

5 =319077,8 12809 0 0 0 0 0 67070

6 =375731,5 33548 0 80622 1 21784 0 0

7 =432385,2 23974 0 100000 6 0 36085 8916

8 =489038,9 3154 0 100000 33575 0 0 30870

9 =545692,6 0 0 100000 10376 14063 0 23884

 =602346,3
10 0 0 100000 100000 1 0 23076

 =659000
11 0 0 100000 100000 1 0 23076

Table 2. Composition of the optimal portfolio for various values of parameter 

Conclusions

We presented several nonlinear integer programming models for the portfolio selection with
minimum transaction lots. The risk in the models is measured by the lower partial moment of the
first order. We determine all the steps the investor must perform in order to make an optimal
investment for the minimum risk model . The composition of the optimal portfolio for various
values of parameter  is computed.

References

[1] Artzner, P., Delbaen, F., Eber, J.M., Heath, D., Thinking coherently, Risk, 10, (1997) 68-71.

[2] Corazza, M., Favaretto, D., Approaching mixed-integer nonlinear mean-variance portfolio
selection, Giorgi, Giorgio (ed.) et al., Generalized convexity and optimization for economic and
financial decisions. Bologna: Pitagora Editrice. (1999), 137 -154.

[3] Corazza, M., Favaretto, D., On the existence of solutions to the quadratic mixed -integer
mean–variance portfolio selection problem , European Journal of Operational Research , Vol. 176,
(3), 2007, 1947-1960.

7
[4] Jobst, N.J., Horniman, M.D., Lucas, C.A., Mitra, G., Computational aspects of alternative
portfolio selection models in the presence of discrete asset choice constraints , Quantitative
Finance 1(2001), 489-501.

[5] Kellerer, H., Mansini, R., Speranza, M.G., Selecting portfolios with fixed costs and
minimum transaction lots, Ann. Oper. Res. 99, (2000), 287 -304.

[6] Konno, H., Yamamoto, R., Integer programming approaches in mean -risk models,
Computational Management Science , Volume 2, Number 4, 2005, 339-351.

[7] Lazimy, R., Portfolio selection with divisible and indivisible assets: Mathematical
algorithm and economic analysis , Annals of Operations Research , Vol. 152, No. 1, 2007, 273-295

[8] Mansini, R., Speranza, M. G., Heuristic algorithms for the portfolio selection problem
with minimum transaction lots , European Journal of Operational Research , 114, No.2, (1999),
219-233.

[9] Mansini, R., Speranza, M.G., An exact approach for portfolio selection with transaction
costs and rounds, IIE Transactions, Vol. 37, No. 10, 2005 , pp. 919-929.

[10] Lin, C.C., Liu, Y.T., Genetic algorithms for portfolio selection problems with minimum
transaction lots, European Journal of Operational Research , Volume 185, Issue 1, 16 February
2008, Pages 393-404

[11] Mitra, G., Ellison, F., Scowcroft, A., Quadratic programming for portfolio planning:
Insights into algorithmic and computational issues Part II — Processing of portfolio planning
models with discrete constraints , J. of Asset Management, Vol. 8, No. 4, 2007, pp. 249-258

[12] Rădulescu, M., Rădulescu, S., Rădulescu, C.Z., Mean-variance portfolio selection models
with finitely divisible assets , 7th International Conf. on Informat ics in Economy IE 2005 –
Information & Knowledge Age, (2005), 1300 -1305. ISBN: 973-8360-04-8.
[13] Rădulescu, M., Rădulescu, S., Rădulescu, C.Z., Mathematical Models for Optimal Asset
Allocation, Editura Academiei, Bucuresti 2006. (Romanian).

[14] Roy, A.D., Safety first and holding of assets , Econometrica 20, 1952, 431-449.

[15] Soleimani, N., Golmakani, H.R., Salimi, M.H., Markowitz-based portfolio selection with
minimum transaction lots, cardin ality constraints and regarding sector capitalization using
genetic algorithm, Expert Systems with Applications , In Press, Corrected Proof , Available
online 14 June 2008, doi:10.1016/j.eswa.2008.06.007 .

[16] Speranza, M.G., A heuristic algorithm for a portfolio optimization model applied to the
Milan stock market, Comput. Oper. Res. 23, No.5, (1996), 433-441.