Chapter 2

Linear Models for Portfolio Optimization

2.1 Introduction

Nowadays, Quadratic Programming (QP) models, like Markowitz model, are not
hard to solve, thanks to technological and algorithmic progress. Nevertheless, Linear
Programming (LP) models remain much more attractive from a computational point
of view for several reasons. The design and development of commercial software for
the solution of LP models is more advanced than for QP models. As a consequence,
several commercial LP solvers are available and, in general, LP solvers tend to be
more reliable than QP solvers. On average, LP solvers can solve in small time (the
order of seconds) instances of much larger size than QP solvers.
Is it possible to have linear models for portfolio optimization? How can we
measure the risk or safety in order to have a linear model? A first observation is
that, in order to guarantee that a portfolio takes advantage of diversification, no risk
or safety measure can be a linear function of the shares of the assets in the portfolio,
that is of the variables xj , j D 1; : : : ; n. Linear models, however, can be obtained
through discretization of the return random variables or, equivalently, through the
concept of scenarios.

2.2 Scenarios and LP Computability

We have indicated by Rj the random variable representing the rate of return of asset
j, j D 1; : : : ; n, at the target time.
Now we change the way we look at the uncertainty of the rates of return of
the assets at the target time and introduce the concept of scenario. A scenario is,
informally, a possible situation that can happen at the target time, in our case a
possible realization of the rates of return of the assets at the target time. Depending

Springer International Publishing Switzerland 2015 19

R. Mansini et al., Linear and Mixed Integer Programming for Portfolio
Optimization, EURO Advanced Tutorials on Operational Research,
DOI 10.1007/978-3-319-18482-1_2
20 2 Linear Models for Portfolio Optimization

Table 2.1 Scenarios: An example

Asset Scenario 1 (%) Scenario 2 (%) Scenario 3 (%) Mean return rate (%)
1 r11 D 3:1 r12 D 2:7 r13 D 1:60 1 D 0:67
2 r21 D 2:3 r22 D 2:3 r23 D 1:30 2 D 0:43
3 r31 D 4:2 r32 D 3:1 r33 D 0:2 3 D 0:43
4 r41 D 1:5 r42 D 2:0 r43 D 0:1 4 D 0:2

on what will happen between the investment time and the target time, any of several
different scenarios may become true. The scenarios may also be less or more likely
to happen. More formally, a scenario is a realization of the multivariate random
variable representing the rates of return of all the assets.
We now suppose that, on the basis of a careful preliminary analysis, T different
scenarios have been identified as possible at the target time. The P probability that
scenario t, t D 1; : : : ; T, will happen is indicated by pt , with TtD1 pt D 1. We
assume that for each random variable Rj , j D 1; : : : ; n, its realization rjt under
scenario t is known. The set of the returns of all the assets frjt ; j D 1; : : : ; ng
defines thePscenario t. The expected return of asset j, j D 1; : : : ; n, is calculated
T
as j D tD1 pt rjt . The concept of scenario captures the correlation among the
rates of return of the assets.
In Table 2.1, we show an example of n D 4 assets and T D 3 scenarios. The
table shows the rates of return of the assets in the different scenarios. Scenario 1
is an optimistic scenario: all rates of return are positive. Scenario 2 is negative,
whereas scenario 3 is positive for assets 1 and 2 and negative for assets 3 and 4.
The averages are computed under the assumptions that the scenarios are equally
probable (pt D 1=3, t D 1; 2; 3).
Identifying the scenarios, their probabilities and estimating the values of the
rate of return rjt of each asset j under each scenario t is crucial. To be statistically
significant, the number of scenarios has to be sufficiently large. Pn
Each portfolio x defines a corresponding random variable Rx D jD1 Rj xj
that represents the portfolio rate of return. The step-wise cumulative distribution
function (cdf) of fRx g is defined as

Fx ./ D P.Rx  /: (2.1)

The return yt of a portfolio x in scenario t can be computed as

X
n
yt D rjt xj ; (2.2)
jD1
2.3 Basic LP Computable Risk Measures 21

and the expected return of the portfolio .x/ can be computed as a linear function
of x

X
T X
T X
n X
n X
T X
n
.x/ D EfRx g D p t yt D pt . rjt xj / D xj pt rjt D j xj :
tD1 tD1 jD1 jD1 tD1 jD1
(2.3)

We have defined a scenario as a realization of the multivariate random variable

representing the rates of return of the assets. We may look at the set of scenarios as
a discretization of the multivariate random variable.
We will say that the returns are discretized when they are defined by their
realizations under the specified scenarios, that is by the set of values frjt W j D
1; : : : ; n; t D 1; : : : ; Tg. We will say that a risk or a safety measure is LP computable
if the portfolio optimization model takes a linear form in the case of discretized
returns.

2.3 Basic LP Computable Risk Measures

The variance is the classical statistical quantity used to measure the dispersion of
a random variable around its mean. There are, however, other ways to measure the
dispersion of a random variable. The random variable, we are interested in, is the
portfolio return Rx .
The Mean Absolute Deviation (MAD) is a dispersion measure that is defined as

X
n X
n
.x/ D EfjRx  EfRx gjg D Efj Rj xj  Ef Rj xj gjg: (2.4)
jD1 jD1

The MAD measures the average of the absolute value of the difference between
the random variable and its expected value. With respect to the variance, the MAD
considers absolute values instead of squared values. We show in the following that,
when the returns are discretized, the MAD is LP computable. Recalling that the
expected return of the portfolio can be calculated as (2.3), the MAD can be written as

X
T X
n X
n
.x/ D pt .j rjt xj  j xj j/: (2.5)
tD1 jD1 jD1
22 2 Linear Models for Portfolio Optimization

The portfolio optimization problem then becomes

X
T X
n
min .x/ D pt .j rjt xj  j/ (2.6a)
tD1 jD1

X
n
D j xj (2.6b)
jD1

  0 (2.6c)
x 2 Q; (2.6d)

where 0 is the lower bound on the portfolio expected return required by the
investor, and Q denotes the system of constraints defining the set of feasible
portfolios as described in Chap. 1.
This form is not linear in the variables xj but can be transformed into a linear
form. Using (2.2) for the return of the portfolio in scenario t, yt , .x/ can also be
written as

X
T X
n
.x/ D pt .jyt  j xj j/:
tD1 jD1

X
n
We now define the deviation in scenario t as dt , that is dt D jyt  j xj j. Then,
jD1
the portfolio optimization problem is

X
T
min pt dt (2.7a)
tD1

dt D jyt  j t D 1; : : : ; T (2.7b)
X
n
yt D rjt xj t D 1; : : : ; T (2.7c)
jD1

X
n
D j xj (2.7d)
jD1

  0 (2.7e)
x 2 Q: (2.7f)
2.3 Basic LP Computable Risk Measures 23

Since jyt  j D maxf.yt  /I .yt  /g, the problem can be written in the
following equivalent linear form
X
T
min pt dt (2.8a)
tD1

d t  yt   t D 1; : : : ; T (2.8b)
dt  .yt  / t D 1; : : : ; T (2.8c)
X
n
yt D rjt xj t D 1; : : : ; T (2.8d)
jD1

X
n
D j xj (2.8e)
jD1

  0 (2.8f)
dt  0 t D 1; : : : ; T (2.8g)
x 2 Q: (2.8h)

The equivalence comes from observing that if yt    0 constraints (2.8c)

are
PT redundant. In this case constraints (2.8b), combined with the minimization of
tD1 pt dt in (2.8a) that pushes the value of each dt to the minimum value allowed
P constraints, impose that dt D yt   D jyt  j. If, on the contrary
by the
yt  njD1 j xj  0, constraints (2.8b) are redundant. In this case, constraints (2.8c),
P
combined with the objective function, impose that dt D .yt  njD1 j xj / D
Pn
jyt  jD1 j xj j. Thus, in conclusion, the optimization model (2.8) is a linear
programming model for the optimization of a portfolio where the risk is measured
through the MAD of the return of the portfolio.
In Fig. 2.1, we represent the calculation of the MAD measure. In other words, we
assume that the values of the shares xj are given. We represent over the horizontal
axis the scenarios t D 1; : : : ; T and over the vertical axis the values yt of the return

Fig. 2.1 The MAD measure

24 2 Linear Models for Portfolio Optimization

of the portfolio under the various scenarios

P t. The thickPhorizontal line identifies
the expected return of the portfolio  D njD1 j xj D TtD1 pt yt . The length of a
vertical segment is the absolute value of the deviation dt (in Fig. 2.1, the deviation
d15 corresponding to the scenario t D 15 is drawn as example). The MAD model
aims at minimizing the average absolute deviation.
In the case the rates of return are a multivariate normally distributed random
variable, the rate of return of the portfolio is normally distributed. Then, the
proportionality relation between
q the mean absolute deviation and the standard
2
deviation occurs .x/ D  .x/. As a consequence, minimizing the MAD is
equivalent to minimizing the standard deviation, which means, in this specific case,
the equivalence of the associated optimization problems. However, the MAD model
does not require any specification of the return distribution.
The MAD accounts for all deviations of the rate of return of the portfolio from
its expected value, both below and above the expected value. However, one may
sensibly think that any rational investor would consider real risk only the deviations
below the expected value. In other words, the variability of the portfolio rate of
return above the mean should not be penalized since the investors are concerned
with under-performance rather than over-performance of a portfolio. In terms of
scenarios, the risky scenarios are those where the rate of return of the portfolio is
below its expected value. We can modify the definition of the MAD in order to
consider only the deviations below the expected value. We define the Semi Mean
Absolute Deviation (Semi-MAD)
X
n X
n
N
.x/ D Efmaxf0; Ef R j xj g  Rj xj gg; (2.9)
jD1 jD1

where the deviations above the expected value are not calculated. The portfolio
optimization problem (2.8) presented for the MAD can be adapted to the Semi-
MAD as follows:

X
T
min pt dt (2.10a)
tD1

d t    yt t D 1; : : : ; T (2.10b)
X
n
yt D rjt xj t D 1; : : : ; T (2.10c)
jD1

X
n
D j xj (2.10d)
jD1

  0 (2.10e)
dt  0 t D 1; : : : ; T (2.10f)
x 2 Q: (2.10g)
2.3 Basic LP Computable Risk Measures 25

The formulation for the Semi-MAD is the formulation of the MAD, from which
inequalities (2.8b) have been dropped. If, for a given scenario t,   yt > 0, this
means that under scenario t the rate of return of the portfolio yt is below the expected
value. In this case dt in the optimum will be the difference   yt . If instead   yt 
0, constraint (2.10b) becomes redundant and in the optimum dt D 0. Thus, the
deviations above the expected value are not calculated in the objective function.
The Semi-MAD seems to be a very attractive measure, focusing on the downside
risk only. However, it can be seen that it is equivalent to the MAD as the corre-
sponding optimization models generate the same optimal portfolio. The intuition
behind the equivalence, that is somehow surprising, is that the MAD is the sum
of the deviations above and below the expected value. By definition of expected
value, the sum of the deviations above the expected value is equal to the sum of
the deviations below the expected value. Thus, the Semi-MAD is half the MAD.
Minimizing the downside deviations is equivalent to minimizing the total deviations
and equivalent to minimizing the deviations above the expected value as well. We
make this equivalence formal.
Theorem 2.1 Minimizing the MAD is equivalent to minimizing the Semi-MAD as
N
.x/ D 2.x/.
Proof We first write the mean deviation of the portfolio rate of the return from its
expected value and show that it is equal to 0:

EfRx  EfRx gg D EfRx g  EfRx g D 0

From this it immediately follows that the average positive deviation (yt  .x/ > 0
implies the rate of return of the portfolio in scenario t is above its expected value) is
equal to the opposite of the average negative deviation (yt  .x/ < 0 implies the
rate of return of the portfolio in scenario t is below its expected value). The absolute
value of the average positive deviation is thus equal to the absolute value of average
negative deviation, from which it follows that the MAD is twice the Semi-MAD.
t
u
Although the MAD has become a very popular risk measure, a different LP
computable risk measure was earlier proposed, namely the Ginis mean difference.
The variability of the portfolio return is captured here by the differences of the
portfolio returns in different scenarios. For a discrete random variable represented
by its realizations yt , the Ginis mean difference (GMD) considers as risk the average
absolute value of the differences of the portfolio returns yt in different scenarios:

1XX
T T
.x/ D jyt0  yt00 jpt0 pt00 : (2.11)
2 0 00
t D1 t D1

The risk function .x/, to be minimized, is LP computable.

In Fig. 2.2, the values of the rate of return for a given portfolio under T D 25
scenarios are shown. The length of the vertical segment is the absolute value of the
26 2 Linear Models for Portfolio Optimization

Fig. 2.2 The GMD measure

difference between the portfolio returns under scenarios 15 and 20, i.e. d15;20 D
jy15  y20 j.
The portfolio optimization model based on the GMD risk measure can be written
as follows:

X
T X
min pt0 pt00 dt0 t00 (2.12a)
t0 D1 t00 t0

dt0 t00  yt0  yt00 t0 ; t00 D 1; : : : ; TI t00 t0 (2.12b)

X
n
yt D rjt xj t D 1; : : : ; T (2.12c)
jD1

X
n
D j xj (2.12d)
jD1

  0 (2.12e)
dt0 t00  0 t0 ; t00 D 1; : : : ; TI t00 t0 (2.12f)
x 2 Q: (2.12g)

The model contains T.T1/ non-negative variables dt0 t00 and T.T1/ inequalities
to define them. The symmetry property dt0 t00 D dt00 t0 is here ignored. However,
variables dt0 t00 are associated with the singleton coefficient columns. Hence, while
solving the dual instead of the original primal problem, the corresponding dual
constraints take the form of simple upper bounds which are handled implicitly by
the simplex method. In other words, the dual problem contains T.T  1/ variables
but the number of constraints is then proportional to T. Such a dual approach may
dramatically improve the required computational time in the case of large number
of scenarios.
2.4 Basic LP Computable Safety Measures 27

Similarly to MAD, in the case when the rates of return are multivariate normally
distributed, the proportionality relation .x/ D p2 .x/ between the Ginis mean
difference and the standard deviation occurs. As a consequence, minimizing the
GMD is equivalent to minimizing the standard deviation, which means, in this
specific case, the equivalence of the associated optimization problems. Albeit, the
GMD model does not require any specific type of return distribution.

2.4 Basic LP Computable Safety Measures

In the previous chapter and in the previous section of this chapter, we have seen
some specific risk measures, the variance (Markowitz model), the mean absolute
deviation (MAD), the Ginis mean difference (GMD). These measures capture, in
different ways, the variability of the rate of return of the portfolio. Given a required
expected return of the portfolio 0 , the investor may wish to reduce the variability
of the portfolio rate of return, that is to minimize any of these risk measures. We
analyze here different ways to measure the quality of a portfolio and define some
specific safety measures, to be maximized. We do not consider the variability of the
portfolio rate of return, neither the deviations from its expected value. In fact, we
ignore the expected rate of return and try instead to protect the investor from the
worst scenarios.
An appealing safety measure is the worst realization of the portfolio rate of
return. We aim at maximizing the worst realization of the portfolio rate of return.
The worst realization is defined as

X
n
M.x/ D min yt D min rjt xj ; (2.13)
tD1;:::;T tD1;:::;T
jD1

and is LP computable. The portfolio optimization model with the worst realization
as safety measure (the Minimax model) can be formulated as:

max y (2.14a)
X
n
rjt xj  y t D 1; : : : ; T (2.14b)
jD1

X
n
D j xj (2.14c)
jD1

  0 (2.14d)
x 2 Q: (2.14e)
28 2 Linear Models for Portfolio Optimization

The variable y is an artificial variable that in the optimum takes the value of the
portfolio rate of return in the worst scenario. In Fig. 2.3, the rates of return for a
given portfolio over 25 scenarios are drawn. and the worst realization of the portfolio
rate of return is emphasized.
Suppose that, among the feasible portfolios of the Minimax model, there are the
two portfolios shown in Table 2.2. Suppose that the required expected rate of return
is 0 D 2 %. Both portfolios x0 and x00 guarantee an expected rate of return not
worse than 2 %. Whereas portfolio x0 has a larger expected rate of return, the model
would prefer portfolio x00 to portfolio x0 because portfolio x00 has the rate of return in
the worst scenario, 2 %, larger than the worst rate of return of portfolio x0 , 1:8 %. The
maximization of the worst realization somehow pushes all the realizations toward
larger and thus better values, but at the same time focuses on the worst scenario
only.
A natural generalization of the measure M.x/ is the statistical concept of quantile.
In general, for given 2 0; 1, the -quantile of a random variable R is the value q
such that

PfR < qg   PfR  qg:

For 2 .0; 1/, it is known that the set of such -quantiles is a closed interval
(see Embrechts et al. 1997). Given a value of , in order to formalize the quantile

Fig. 2.3 The worst realization measure

Table 2.2 Optimal portfolio Rates of return

for the worst realization
Scenario Probability Portfolio x0 (%) Portfolio x00 (%)
safety measure: An example
1 0.2 4.9 2.0
2 0.5 4.0 3.0
3 0.2 2.2 2.0
4 0.1 1.8 2.0
Mean .x/ 3.6 2.5
Worst M.x/ 1.8 2.0
2.4 Basic LP Computable Safety Measures 29

1 Fx ( )

q (x)

Fig. 2.4 VaR measure q .x/ and the cdf of portfolio returns

measures in the case of non-unique quantile, the left end of the entire interval of
quantiles is taken. In our case, we denote by q .x/ the value of the -quantile, that
is the value of the rate of return defined as

q .x/ D inf f
W Fx .
/  g for 0 <  1; (2.15)

where Fx ./ is the cumulative distribution function defined in (2.1) (see Fig. 2.4).
In finance and banking literature, this quantile is usually called Value-at-Risk
or simply VaR measure. Actually, for a given portfolio x, its VaR depicts the worst
(maximum) loss within a given confidence interval (see Jorion 2006). However, with
a change of sign (losses as negative returns Rx ), it is equivalent to the quantile
q .x/.
Due to possible discontinuity of the cdf, the VaR measure is, generally, not an
LP computable measure. The corresponding portfolio optimization model can be
formulated as a MILP problem:

max y (2.16a)
X
n
rjt xj  y  Mzt t D 1; : : : ; T (2.16b)
jD1

X
T
pt zt   ; zt 2 f0; 1g t D 1; : : : ; T (2.16c)
tD1

X
n
D j xj (2.16d)
jD1

  0 (2.16e)
x 2 Q; (2.16f)

where M is an arbitrary large constant (larger than any possible rate of return) while
 is an arbitrary small positive constant ( < pt ; t D 1; : : : ; T). Note that, due to
30 2 Linear Models for Portfolio Optimization

inequality (2.16b), binary variable zt takes value 1 whenever

P variable y is greater
than the portfolio return under scenario t (y > yt D njD1 rjt xj ). Inequality (2.16c)
guarantees that the probability of all scenarios such that y > yt is less than .
Therefore, the optimal value of the maximized variable y represents the optimal
-quantile value q .x/.
Recently, risk measures based on averaged quantiles have been introduced in
different ways. The tail mean or worst conditional expectation M .x/, defined as
the mean return of the portfolio taken over a given tolerance level (percentage) 0 <
 1 of the worst scenarios probability is a natural generalization of the measure
M.x/. In finance literature, the tail mean quantity is usually called Tail VaR, Average
VaR or Conditional VaR (where VaR reads after Value-at-Risk). Actually, the name
CVaR is now the most commonly used and we adopt it.
For the simplest case of equally probable scenarios (pt D 1=T) and proportional
D k=T, the CVaR measure M .x/ is defined as average of the k worst realizations

1X
k
M k .x/ D yt ; (2.17)
T k sD1 s

where yt1 ; yt2 ; : : : ; ytk are the k worst realizations for the portfolio rate of return.
In Fig. 2.5, we show an example of a portfolio whose CVaR value has been
computed for k D 3 and T D 25.
For any probability pt and tolerance level , due to the finite number of scenarios,
the CVaR measure M .x/ is well defined by the following optimization

1X X
T T
M .x/ D minf yt u t W ut D ; 0  ut  pt t D 1; : : : ; Tg; (2.18)
ut tD1 tD1

where at optimality ut is the percentage of the t-th worst return in M .x/. More
precisely, ut D 0 for any scenario t not included in the worst scenarios, ut D pt

Fig. 2.5 The CVaR model M Tk .x/

2.4 Basic LP Computable Safety Measures 31

Table 2.3 Optimal portfolios Rates of return

for the CVaR measure: An
Scenario Probability Portfolio x0 (%) Portfolio x00 (%)
example
1 0.2 4.9 2.0
2 0.5 4.0 3.0
3 0.2 2.2 2.0
4 0.1 1.8 2.0
Worst M.x/ 1.8 2.0
CVaR M0:05 .x/ 1.8 2.0
CVaR M0:1 .x/ 1.8 2.0
CVaR M0:2 .x/ 2.0 2.0
CVaR M0:3 .x/ 2.07 2.0
CVaR M0:5 .x/ 2.84 2.0
CVaR M0:8 .x/ 3.28 2.38
CVaR M1:0 .x/ 3.6 2.5
Mean .x/ 3.6 2.5

for any scenario t totally included in the worst scenarios, and 0 < ut < pt for one
scenario t only.
When parameter approaches 0 and becomes smaller than or equal to the
minimal scenario probability (  mint pt ), the measure becomes the worst return
M.x/ D lim!0C M .x/. On the other hand, for D 1 the corresponding CVaR
becomes the mean (M1 .x/ D .x/).
Recall the case of two portfolios shown in Table 2.2. In Table 2.3, we show
their CVaR values for various tolerance levels. For D 0:05 and D 0:1 the
CVaR values are equal to the corresponding return in the worst scenario, M.x0 / D
1:8 % and M.x00 / D 2 %, respectively. For D 0:2 one gets equal CVaR values
M0:2 .x0 / D M0:2 .x00 / D 2 %, while for D 0:3 one has M0:3 .x0 / D 2:07 % greater
than M0:3 .x00 / D 2 %. The difference becomes larger for tolerance levels D 0:5
and D 0:8. Obviously, for D 1 one gets the corresponding means as CVaR
values.
Problem (2.18) is a linear program for a given portfolio x, while it becomes non-
linear when the yt are variables in the portfolio optimization problem. It turns out
that this difficulty can be overcome by taking advantage of the LP dual problem
to (2.18) leading to an equivalent LP dual formulation of the CVaR model that
allows one to implement the optimization problem with auxiliary linear
P inequalities.
Indeed, introducing dual variable
corresponding to the equation TtD1 ut D and
variables dt corresponding to upper bounds on ut one gets the LP dual problem:

1X  
T
M .x/ D max f
 pt dt W dt 
 yt ; dt  0 t D 1; : : : ; Tg: (2.19)


;dt tD1

Due to the duality theory, for any given vector yt the measure M .x/ can be found as
the optimal value of the LP problem (2.19). Thus, the CVaR is a safety measure that
32 2 Linear Models for Portfolio Optimization

is LP computable. The portfolio optimization model can be formulated as follows:

1X 
T
max .
 pt dt / (2.20a)
tD1

dt 
 yt t D 1; : : : ; T (2.20b)
X
n
yt D rjt xj t D 1; : : : ; T (2.20c)
jD1

X
n
D j xj (2.20d)
jD1

  0 (2.20e)
dt 0 t D 1; : : : ; T (2.20f)
x 2 Q; (2.20g)
where
is an auxiliary (unbounded) variable that in the optimal solution will take
the value of the -quantile.
In the case of P fRx  q .x/g D , one gets M .x/ D E fRx jRx  q .x/g.
This represents the original concept of the CVaR measure. Although valid for many
continuous distributions this formula, in general, cannot serve as a definition of the
CVaR measure because a value  such that P fRx  g D may not exist. In
general, P fRx  q .x/g D 0  and M .x/  M0 .x/ D E fRx jRx  q .x/g.

2.5 The Complete Set of Basic Linear Models

As shown in the previous sections several LP computable risk measures have been
considered for portfolio optimization. Some of them were originally introduced
rather as safety measures in our classification (e.g., CVaR measures). Nevertheless,
all of them can be represented with positively homogeneous and shift independent
risk measures % of classical Markowitz type model. Simple as well as more
complicated LP computable risk measures %.x/ can be defined as
%.x/ D minfaT v W Av D Bx; v  0; x 2 Qg; (2.21)
where v is a vector of auxiliary variables while the portfolio vector x, apart from
original portfolio constraints x 2 Q, only defines a parametric right hand side vector
b D Bx. Obviously, the corresponding safety measures are given by a similar LP
formula

X
n
.x/  %.x/ D maxf j xj  aT v W Av D Bx; v  0; x 2 Qg: (2.22)
jD1
2.5 The Complete Set of Basic Linear Models 33

For each model of type (2.21), the mean-risk bounding approach (1.10) leads to
the LP problem

X
n
minfaT v W Av D Bx; v  0; j xj  0 ; x 2 Qg; (2.23)
x;v
jD1

while the mean-safety bounding approach (1.12) applied to (2.22) results in

X
n X
n
maxf j xj  aT v W Av D Bx; v  0; j xj  0 ; x 2 Qg: (2.24)
x;v
jD1 jD1

Similarly, the trade-off analysis approach (1.13) results in the LP model

X
n
maxf j xj  aT v W Av D Bx; v  0; x 2 Qg: (2.25)
x;v
jD1

2.5.1 Risk Measures from Safety Measures

Recall that, for a discrete random variable represented by its realizations yt , the
worst realization M.x/ D mintD1;:::;T fyt g is an appealing LP computable safety
measure (see (2.13)). The corresponding (dispersion) risk measure .x/ D .x/ 
M.x/, the maximum (downside) semideviation, is LP computable as

X
n
.x/ D minfv W v  .j  rjt /xj ; t D 1; : : : ; Tg: (2.26)
jD1

The portfolio optimization model with the maximum semideviation as risk measure
can be formulated as:

min v (2.27a)
X
n
 rjt xj  v t D 1; : : : ; T (2.27b)
jD1

X
n
D j xj (2.27c)
jD1

  0 (2.27d)
x 2 Q: (2.27e)
34 2 Linear Models for Portfolio Optimization

The variable v is an auxiliary variable that in the optimum will take the value of the
maximum downside deviation of the portfolio rate of return from the mean return.
Similarly, the CVaR measure is a safety measure. The corresponding risk
measure  .x/ D .x/  M .x/ is called the (worst) conditional semideviation
or conditional drawdown measure. For a discrete random variable represented by
its realizations, due to (2.19), the conditional semideviations may be computed as
the corresponding differences from the mean:

X
n
1X 
T
 .x/ D minf j xj 
C d pt W dt 
 yt ; dt  0; t D 1; : : : ; Tg;
jD1
tD1 t
(2.28)

or, equivalently, setting dtC D dt 

C yt , as:

XT
1 
 .x/ D minf .dtC C dt /pt W dt  dtC D
 yt ; dt ; dtC  0; t D 1; : : : ; Tg;
tD1

(2.29)

where
is an auxiliary (unbounded) variable that in the optimal solution will take
the value of the -quantile q .x/.
Thus, the conditional semideviation is an LP computable risk measure and the
corresponding portfolio optimization model can be formulated as follows:

XT
1 
min .dtC C d /pt (2.30a)
tD1
t

dt  dtC D
 yt ; dt ; dtC  0 t D 1; : : : ; T (2.30b)
X
n
yt D rjt xj t D 1; : : : ; T (2.30c)
jD1

X
n
D  j xj (2.30d)
jD1

  0 (2.30e)
x 2 Q: (2.30f)

Note that for D 0:5 one has .1  /= D 1. Hence, 0:5 .x/ represents the
mean absolute deviation from the median q0:5 .x/. The LP problem for computing
2.5 The Complete Set of Basic Linear Models 35

this measure can be expressed in the form:

X
T
0:5 .x/ D minf dt pt W dt 
 yt ; dt  yt 
; dt  0 t D 1; : : : ; Tg:
tD1

One may notice that the above model differs from the classical MAD formulation
(2.8) only due to replacement of .x/ with (unrestricted) variable
.

2.5.2 Safety Measures from Risk Measures

Symmetrically, a safety measure can be obtained from a positively homogeneous

and shift independent (deviation type) risk measure. For the Semi-MAD (2.9) the
corresponding safety measure can be expressed as

N
.x/  .x/ D Ef.x/  maxf.x/  Rx ; 0gg D EfminfRx ; .x/gg; (2.31)

thus representing the mean downside underachievement. The corresponding portfolio

optimization problem can be written as follows:

X
T
max pt vt (2.32a)
tD1

X
n
vt  rjt xj t D 1; : : : ; T (2.32b)
jD1

vt   t D 1; : : : ; T (2.32c)
X
n
D  j xj (2.32d)
jD1

  0 (2.32e)
x 2 Q: (2.32f)

The Ginis mean difference (2.11) has the corresponding safety measure defined as

X
T X
T
.x/  .x/ D minfyt0 ; yt00 gpt0 pt00 ; (2.33)
t0 D1 t00 D1

where the latter expression is obtained through algebraic calculations. Hence, (2.33)
is the expectation of the minimum of two independent identically distributed random
variables, thus representing the mean worse return.
36 2 Linear Models for Portfolio Optimization

This leads to the following LP formula

X
T X
T
.x/  .x/ D maxf vt0 t00 pt0 pt00 W
t0 D1 t00 D1 (2.34)
X
n X
n
0 00
vt0 t00  rjt0 xj ; vt0 t00  rjt00 xj ; t ; t D 1; : : : ; Tg:
jD1 jD1

The portfolio optimization model can be written as follows:

X
T X
T
max pt0 pt00 vt0 t00 (2.35a)
t0 D1 t00 D1

vt0 t00  yt0 t0 ; t00 D 1; : : : ; T (2.35b)

vt0 t00  yt00 t0 ; t00 D 1; : : : ; T (2.35c)
X
n
yt D rjt xj t D 1; : : : ; T (2.35d)
jD1

X
n
D  j xj (2.35e)
jD1

  0 (2.35f)
x 2 Q: (2.35g)

2.5.3 Ratio Measures from Risk Measures

As mentioned in Chap. 1, an alternative approach to the bicriteria mean-risk

approach to portfolio selection is based on maximizing the ratio ..x/  r0 /=%.x/.
The corresponding ratio optimization problem (1.14) can be converted into an LP
form by the following transformation: introduce an auxiliary variable z D 1=%.x/,
then replace the original variables x and v with xQ D zx and vQ D zv, respectively,
getting the linear criterion and an LP feasible set. For risk measure %.x/ defined by
(2.21) one gets the following LP formulation of the corresponding ratio model

X
n
maxf j xQ j  r0 z W cT vQ D z; AQv D bQx; vQ  0;
xQ ;Qv;z
jD1
X
n (2.36)
xQ j D z; xQ j  0; j D 1; : : : ; ng;
jD1
2.5 The Complete Set of Basic Linear Models 37

where the second line constraints correspond to the simplest definition of set
Pn
Q D fx W jD1 xj D 1; xj  0; j D 1; : : : ; ng and can be accordingly formulated
for any other LP set. Once the transformed problem is solved, the values of the
portfolio variables xj can be found by dividing xQ j by the optimal value of z. Thus,
the LP computable portfolio optimization models, we consider, remain within LP
environment even in the case of ratio criterion used to define the tangency portfolio.
For the Semi-MAD model (2.10) with risk measure %.x/ D .x/ N , the ratio
optimization model can be written as
( )
  r0
max PT W (2.10b)(2.10g) :
tD1 pt dt

P
Introducing variables z D 1= TtD1 pt dt and vQ D z we get the linear criterion vQ  r0 z.
Further, we multiply all the constraints by z and make the substitutions: dQ t D zdt ,
yQ t D zyt , for t D 1; : : : ; T , as well as xQ j D zxj , for j D 1; : : : ; n. Finally, we get the
following LP formulation:

max vQ  r0 z (2.37a)
X
T
pt dQ t D 1 (2.37b)
tD1

dQ t  vQ  yQ t ; dQ t  0 t D 1; : : : ; T (2.37c)
X
n
yQ t D rjt xQ j t D 1; : : : ; T (2.37d)
jD1

X
n
vQ D j xQ j (2.37e)
jD1

X
n
xQ j D z; xQ j  0 j D 1; : : : ; n; (2.37f)
jD1

where the last constraints correspond to the simplest definition of set Q.

Clear identification of dispersion type risk measures for all the LP computable
safety measures allows us to define tangency portfolio optimization for all the
models.
For the CVaR model with conditional semideviation as risk measure %.x/ D
 .x/ (2.30) the ratio optimization model can be written as
( )
  r0
max PT C 1 
W (2.30b)(2.30f) :
tD1 .dt C
dt /pt
38 2 Linear Models for Portfolio Optimization

P
Introducing variables z D 1= TtD1 .dtC C 1
dt /pt and vQ D z we get the linear
criterion vQ  r0 z. Further, we multiply all the constraints by z and make the
substitutions: dQ tC D zdtC , dQ t D zdt , yQ t D zyt for t D 1; : : : ; T , as well as xQ j D zxj , for
j D 1; : : : ; n. Then, we get the following LP formulation:

max vQ  r0 z (2.38a)
XT
1  Q
.dQ tC C dt /pt D 1 (2.38b)
tD1

dt  dtC D
 yt ; dt ; dtC  0 t D 1; : : : ; T (2.38c)
X
n
yQ t D rjt xQ j t D 1; : : : ; T (2.38d)
jD1

X
n
vQ D j xQ j (2.38e)
jD1

X
n
xQ j D z; xQ j  0 j D 1; : : : ; n: (2.38f)
jD1

2.6 Advanced LP Computable Measures

The LP computable risk measures may be further extended to enhance the risk
aversion modeling capabilities. First of all, the measures may be combined in a
weighted sum which allows the generation of various mixed measures. Consider a
set of, say m, risk measures %k .x/ and their linear combination with weights wk :

X
m X
m
%.m/
w .x/ D wk %k .x/; wk  1; wk  0 for k D 1; : : : ; m: (2.39)
kD1 kD1

Note that
X
m
.x/  %.m/
w .x/ D w0 .x/ C wk ..x/  %k .x//;
kD1

Pm
where w0 D 1  kD1 wk  0.
2.6 Advanced LP Computable Measures 39

In particular, one may build a multiple CVaR measure by considering, say m,

tolerance levels 0 < 1 < 2 < : : : < m  1 and using the weighted sum of the
conditional semideviations k .x/ as a new risk measure

X
m X
m
.m/
w .x/ D wk k .x/; wk D 1; wk > 0 for k D 1; : : : ; m; (2.40)
kD1 kD1

with the corresponding safety measure

X
m
Mw.m/ .x/ D .x/  .m/
w .x/ D wk Mk .x/: (2.41)
kD1

The resulting Weighted CVaR (WCVaR) models use multiple CVaR measures, thus
allowing for more detailed risk aversion modeling. The WCVaR risk measure is
obviously LP computable as

.m/
X
m
1 X 
T
Mw .x/ D maxf wk .
k  dkt pt / W dkt  0;
kD1
k tD1
X
n (2.42)

dkt 
k  rjt xj ; t D 1; : : : ; TI k D 1; : : : ; mg:
jD1

The corresponding portfolio optimization model can be formulated as follows:

X
m
1 X 
T
max wk .
k  d pt / (2.43a)
kD1
k tD1 kt

dkt 
k  yt ; dkt  0 t D 1; : : : ; TI k D 1; : : : ; m (2.43b)
X
n
yt D rjt xj t D 1; : : : ; T (2.43c)
jD1

X
n
D  j xj (2.43d)
jD1

  0 (2.43e)
x 2 Q: (2.43f)

For appropriately defined weights the WCVaR measures may be considered

some approximations to the Ginis mean difference with the advantage of being
computationally much simpler than the GMD model itself.
The risk measures introduced in the previous section are quite different in
modeling the downside risk aversion. Definitely, the strongest in this respect is the
maximum semideviation .x/ while the conditional semideviation  .x/ (CVaR
40 2 Linear Models for Portfolio Optimization

model) allows us to extend the approach to a specified quantile of the worst returns
which results in a continuum of models evolving from the strongest downside
risk aversion ( close to 0) to the complete risk neutrality ( D 1). The mean
(downside) semideviation from the mean, used in the MAD model, is formally a
downside risk measure. However, due to the symmetry of mean semideviations
from the mean it is equally appropriate to interpret it as a measure of the upside
risk. Similarly, the Ginis mean difference, although related to all the conditional
maximum semideviations, is a symmetric risk measure (in the sense that for Rx and
Rx it has exactly the same value). For better modeling of the risk averse preferences
one may enhance the below-mean downside risk aversion in various measures. The
below-mean risk downside aversion is a concept of risk aversion assuming that the
variability of returns above the mean should not be penalized since the investors
are concerned about the under-performance rather than the over-performance of
a portfolio. This can be implemented by focusing on the distribution of downside
underachievements minfRx ; .x/g instead of the original distribution of returns Rx .
Applying the mean semideviation (2.9) to the distribution of downside under-
achievements min fRx ; .x/g one gets

N2 .x/ D EfmaxfEfminfRx ; .x/gg  Rx ; 0gg D Efmaxf.x/  .x/

N  Rx ; 0gg:

This allows us to define the enhanced risk measure for the original distribution
of returns Rx as N .2/ .x/ D .x/
N C N2 .x/ with the corresponding safety measure
N .2/ N N
.x/  .x/ D .x/  .x/  2 .x/. The above approach can be repeated recursively,
resulting in m (defined recursively) distribution dependent targets 1 .x/ D .x/,
k .x/ D EfminfRx ; k1 .x/gg for k D 2; : : : ; m, and the corresponding mean
semideviations N1 .x/ D .x/
N , Nk .x/ D Efmaxfk .x/  Rx ; 0gg for k D 1; : : : ; m. The
measure
X
m
Nw.m/ .x/ D wk Nk .x/ 1 D w1  w2  : : :  wm  0 (2.44)
kD1

gives rise to the so-called m-MAD model. Actually, the measure can be interpreted
as a single mean semideviation (from the mean) applied with a penalty function:
Nw .x/ D Efu.maxf.x/  Rx ; 0g/g, where u is an increasing and convex piecewise
.m/

linear penalty function with breakpoints bk D .x/k .x/ and slopes sk D w1 C: : :C

wk , k D 1; : : : ; m. Therefore, the measure Nw .x/ is referred to as the mean penalized
.m/

semideviation and is obviously LP computable leading to the following LP form of

2.6 Advanced LP Computable Measures 41

the m-MAD portfolio optimization model:

X
m
min wk v k (2.45a)
kD1

X
T
vk  pt dkt D 0 k D 1; : : : ; m (2.45b)
tD1

X
k1
dkt  0; dkt    yt  vi t D 1; : : : ; TI k D 1; : : : ; m (2.45c)
iD1

X
n
yt D rjt xj t D 1; : : : ; T (2.45d)
jD1

X
n
D  j xj (2.45e)
jD1

  0 (2.45f)
x 2 Q: (2.45g)

The Ginis mean difference is a symmetric measure, thus equally treating both
under and over-achievements. The enhancement technique allows us to define the
downside Ginis mean difference by applying the Ginis mean difference to the
distribution of downside underachievements minfRx ; .x/g

X
T X
T X
T
2 .x/ D minfyt ; .x/gpt  minfminfyt0 ; .x/g; minfyt00 ; .x/ggpt0 pt00 :
tD1 t0 D1 t00 D1

Hence, we get the downside Ginis mean difference defined as the enhanced risk
measure:

X
T X
T
N
d .x/ D 2 .x/ C .x/ D .x/  minfyt0 ; yt00 ; .x/gpt0 pt00 : (2.46)
t0 D1 t00 D1

The downside Ginis safety measure takes the form:

X
T X
T
.x/  d .x/ D minfyt0 ; yt00 ; .x/gpt0 pt00 ; (2.47)
t0 D1 t00 D1
42 2 Linear Models for Portfolio Optimization

which is obviously LP computable. The portfolio optimization model based on the

downside Ginis safety measure can be written as follows:

X
T X
T
max pt0 pt00 vt0 t00 (2.48a)
t0 D1 t00 D1

vt0 t00  yt0 t0 ; t00 D 1; : : : ; T (2.48b)

vt0 t00  yt00 t0 ; t00 D 1; : : : ; T (2.48c)
0 00
vt0 t00   t ; t D 1; : : : ; T (2.48d)
X
n
yt D rjt xj t D 1; : : : ; T (2.48e)
jD1

X
n
D  j xj (2.48f)
jD1

  0 (2.48g)
x 2 Q: (2.48h)

The notion of risk may be related to a possible failure of achieving some targets
instead of the mean. It was formalized by the concept of below-target risk measures
or shortfall criteria. The simplest shortfall criterion for a specific target value  is
the mean below-target deviation (first Lower Partial Moment, LPM)

N .x/ D Efmaxf  Rx ; 0gg: (2.49)

The mean below-target deviation is LP computable for returns represented by their

realizations and the corresponding portfolio optimization model can be written as
follows:

X
T
min pt dt (2.50a)
tD1

dt    yt t D 1; : : : ; T (2.50b)
X
n
yt D rjt xj t D 1; : : : ; T (2.50c)
jD1

X
n
D  j xj (2.50d)
jD1

  0 (2.50e)
dt  0 t D 1; : : : ; T (2.50f)
x 2 Q: (2.50g)
2.6 Advanced LP Computable Measures 43

The mean below-target deviation from a specific target (2.49) represents only a
single criterion. One may consider several, say m, targets 1 > 2 > : : : > m and use
weighted sum of the shortfall criteria as a risk measure:

X
m X
m
wk Nk .x/ D wk Efmaxfk  Rx ; 0gg; (2.51)
kD1 kD1

where wk (for k D 1; : : : ; m) are positive weights which maintain the measure

LP computable (when minimized). Actually, the measure can be interpreted as a
single mean below-target deviation applied with a penalty function: Efu.maxf1 
Rx ; 0g/g, where u is increasing and convex piece-wise linear penalty function with
breakpoints bk D 1  k and slopes sN k D w1 C : : : C wk , k D 1; : : : ; m.
The below-target deviations are very useful in investment situations with clearly
defined minimum acceptable returns (e.g. bankruptcy level). Otherwise, appropriate
selection of the target value might be a difficult task. However, for portfolio
optimization they are rather rarely applied. Recently, the so-called Omega ratio
measure defined, for a given target, as the ratio of the mean over-target deviation
by the mean below-target deviation was introduced:
R1
EfmaxfRx  ; 0gg  .1  Fx .// d
 .x/ D D R : (2.52)
Efmaxf  Rx ; 0gg 1 Fx ./ d

Since   Efmaxf  Rx ; 0gg D .x/  EfmaxfRx  ; 0gg, one gets

N .x/  .  .x// .x/  

 .x/ D D1C :
N .x/ N .x/

Thus, the portfolio optimization model based on the Omega ratio maximization
is equivalent to the standard ratio (tangent portfolio) model (1.14) for the N .x/
measure with target  replacing the risk-free rate of return:
( )

max PT W (2.50b)(2.50g) :
tD1 pt dt

Similarly to the MAD ratio model it is easily transformed to an LP form. Introducing

P
variables z D 1= TtD1 pt dt and vQ D z we get the linear criterion vQ   z. Further, we
multiply all the constraints by z and make the substitutions: dQ t D zdt , yQ t D zyt for
44 2 Linear Models for Portfolio Optimization

t D 1; : : : ; T , as well as xQ j D zxj , for j D 1; : : : ; n. Finally, we get the following LP

formulation:

max vQ   z (2.53a)
X
T
pt dQ t D 1 (2.53b)
tD1

dQ t C yQ t   z; dQ t  0 t D 1; : : : ; T (2.53c)
X
n
j xQ j D vQ (2.53d)
jD1

X
n
rjt xQ j D yQ t t D 1; : : : ; T (2.53e)
jD1

X
n
xQ j D z; xQ j  0 j D 1; : : : ; n; (2.53f)
jD1

where the last constraints correspond to the set Q definition.

2.7 Notes and References

Initial attempts to have portfolio optimization models depended on the piecewise

linear approximation of the variance (see Sharpe 1971a; Stone 1973). Later, several
LP computable risk measures were introduced. Yitzhaki (1982) proposed the LP
solvable portfolio optimization mean-risk model using Ginis mean (absolute)
difference as the risk measure (the GMD model). The mean absolute deviation was
very early considered in portfolio analysis by Sharpe (1971b). The complete LP
solvable portfolio optimization model based on this risk measure (the MAD model)
was presented and analyzed by Konno and Yamazaki (1991). The MAD model
was extensively tested on various stock markets (see Konno and Yamazaki 1991;
Mansini et al. 2003a; Xidonas et al. 2010) including its application to portfolios of
mortgage-backed securities by Zenios and Kang (1993) where the distribution of
rates of return is known to be non-symmetric. The MAD model usually, similarly
to the Markowitz one, generated the portfolios with the largest returns but also
entailing the largest risk of underachievement. This model has generated interest
in LP portfolio optimization resulting in many new developments. Young (1998)
analyzed the LP solvable portfolio optimization model based on risk defined by
the worst case scenario (Minimax model), while Ogryczak (2000) introduced the
multiple criteria LP model covering all the above as special aggregation techniques.
The Semi-MAD was independently presented by Feinstein and Thapa (1993)
and Speranza (1993). The m-MAD model were introduced by Michalowski and
2.7 Notes and References 45

Ogryczak (2001), while Krzemienowski and Ogryczak (2005) introduced the

downside Ginis mean difference. The mean absolute deviation from the median
was suggested as risk measure by Sharpe (1971a).
The quantile risk measures were introduced in different ways by many authors
(see Artzner et al. 1999; Embrechts et al. 1997; Ogryczak 1999; Rockafellar and
Uryasev 2000). The tail mean or worst conditional expectation, defined as the mean
return of the portfolio taken over a given percentage of the worst scenarios is a
natural generalization of the measure due to Young (1998). In financial literature,
the tail mean quantity is usually called tail VaR, average VaR or Conditional
VaR (CVaR) (see Pflug 2000). Actually, the name CVaR, after Rockafellar and
Uryasev (2000), is now the most commonly used. The measure was studied in
several applications (see Andersson et al. 2001; Krokhmal et al. 2002; Roman et al.
2007; Topaloglou et al. 2002), and expanded in various forms (see Acerbi 2002;
Krzemienowski 2009; Mansini et al. 2007; Zhu and Fukushima 2009).
Formal classification into risk and safety measures and their complementary pairs
was introduced in Mansini et al. (2003a). The maximum semideviation measure
was introduced in Ogryczak (2000). The (deviation) risk measure corresponding
to the CVaR was considered as the (worst) conditional semideviation (Ogryczak
and Ruszczynski 2002a) or conditional drawdown measure (Chekhlov et al. 2005).
General deviation risk measures were analyzed by Rockafellar et al. (2006). Linear
formulations of the ratio optimization models for all the basic LP computable risk
measures was introduced in Mansini et al. (2003b).
The notion of risk related to a possible failure of achieving some targets was
introduced by Roy (1952) as the so-called safety-first strategy and later led to
the concept of below-target risk measures (see Fishburn 1977; Nawrocki 1992) or
shortfall criteria.
The Omega measure was introduced by Shadwick and Keating (2002), while the
first LP portfolio optimization model with this measure was shown by Mausser et al.
(2006).
http://www.springer.com/978-3-319-18481-4