Cybernetics and Systems Analysis, Vol. 48, No.

2, March, 2012

SAFETY-FIRST PORTFOLIO SELECTION

V. I. Norkina and S. V. Boykob UDC 519.865.5

Abstract. A. D. Roy’s safety-first (SF) approach to financial portfolio optimization is improved. Safety
first means the minimization of the probability of negative returns. The improvement concerns a better
estimation of the negative return probabilities by means of mean excess return risk functions. The
search for the optimal SF-portfolio is similar to Roy’s geometric method but the efficient frontier is
different. In case of a finite number of scenarios, the SF-portfolio selection problem is reduced to
a mixed linear Boolean programming problem.

Keywords: financial portfolio, optimization, return, downside risk, Roy safety, probability estimate,
efficient frontier.

INTRODUCTION

In the present paper, we will improve A. D. Roy’s approach [1] to the safety optimization of a financial portfolio,
which minimizes the estimate of the probability that portfolio’s return falls critically, the average return being constrained
from below. The improvement implies a more accurate estimate of the probability of negative returns by threshold risk
functions. Searching for the optimal safety portfolio reduces to the development and analysis of a new efficient frontier of
portfolios non-dominated in risk functions. For the case of a finite set of return scenarios, the problem of choosing an
SF-portfolio reduces to a mixed Boolean linear programming problem.
The Markowitz model (1952) [2] of the optimization of a financial portfolio by risk–return criteria, where the
variance (or standard deviation) of return is used as a risk measure, is widely known. The essence of the Markowitz model is
developing an efficient frontier of portfolios with the minimum variance of return for the given average return and choosing
an efficient portfolio that maximizes some utility function [3]. A related study of A. D. Roy, a British economist [1], was
published independently in 1952, where he proposed to optimize a portfolio based on the criterion that the probability of the
portfolio’s return falling below a minimum desired threshold is minimized (the safety-first criterion or SF-criterion). Since
optimizing a probability function was a difficult computational problem in the 1950s, Roy proposed to minimize its
upper-bound estimate obtained from the Chebyshev inequality. For an approximate problem, Roy gave a simple and elegant
geometrical solution lying on the efficient frontier of non-dominated portfolios in the average return–risk (variance of return)
plane. Though this work had not become widely popular, Roy’s ideas continued to develop, for example, in [4] (average
return is maximized under a constrained probability of deriving an income less than a prescribed threshold), [5] (some
quantile of return is optimized), [6] (income is maximized and the probability that income is less than a prescribed thereshold
is minimized), [7] (lexicographic SF-principle is used), [8] (multiperiod SF-optimization is considered), [9] (the estimate of
the probability of receiving less income with the use of extremum distributions is employed), [10] (the probability of
exceeding a objective level of income is maximized and the probability of not exceeding a given level of income is
minimized), [11, 12] (applications of the SF-optimization to portfolios with high-risk assets are considered). In [13–19],
A. I. Kibzun, Yu. S. Kan, et al. analyze optimal control problems for a financial portfolio with a quantile objective criterion
closely related to the probability of not exceeding a prescribed return.
Roy’s study has been in the shadow of Markowitz’s theory for a long time; however, the situation will probably
change. Let us quote from the paper [20].
a
V. M. Glushkov Institute of Cybernetics, National Academy of Sciences of Ukraine, Kyiv, Ukraine,
[email protected]. bEuropean University, Kyiv, Ukraine, [email protected]. Translated from Kibernetika i Sistemnyi
Analiz, No. 2, March–April, 2012, pp. 29–41. Original article submitted March 2, 2010.

©
180 1060-0396/12/4802-0180 2012 Springer Science+Business Media, Inc.
 “In autumn 2001, Bruno Latour, a famous French sociologist, wrote a preface to the French translation of the book of
a no less famous German sociologist Ulrich Beck Risikogesellschaft—Auf dem Weg in eine andere Moderne (Risk Society:
Towards a New Modernity [21]) and emphasized a strange coincidence. Beck’s book was published in Germany right after
the Chernobyl disaster. Its French translation appeared right after the 9/11 tragedy. The society starts recognizing the
dangers it faces, including those resulted from its development. The time of a good hope has passed; the history of science
will not have a happy end. In the postmodern era, which Latour, following Beck, proposes to call a risk era, the science is
necessary for another purpose: to minimize inevitable losses.”
In view of the recent global financial crises, the quoted words of the sociologists seem even more prophetical.
Anyway, it seems that portfolios of pension and other social insurance funds should be optimized based on the SF-criterion.
In the present paper, Roy’s SF-approach to the optimization of a financial portfolio will be improved as follows. The
original Roy’s formulation is supplemented with a constraint for the minimum average return of the portfolio as in the
Markowitz approach. Thus, both approaches are shown to be completely similar. One of the shortcomings of Roy’s (and
Markowitz’s) model (symmetry about the mean variance as a risk measure) is eliminated by replacing the variance with an
average downside deviation of return from a given threshold.
Unlike Roy’s original approach, the Chebyshev inequality employs not variance but average downside deviation of
return from an arbitrary given threshold (downside risk) for the estimate of probability of a critical drop of return. Then this
threshold was optimally selected. As a result, more accurate estimates of probability are obtained. The efficient frontier of
portfolios is constructed not in the variance–return plane bur in another plane, downside-risk–return.
Thus, in the present paper, portfolio SF-optimization consists in the approximate minimization of the probability
measure of risk (the probability that return is less than a given critical level), the return being constrained from below. The
solution of the approximate problem is to develop and analyze the efficient frontier of portfolios.
Moreover, for a finite set of scenarios of return, the optimization problem for the probability of the critical drop of
return can be equivalently reduced to a mixed Boolean programming problem, which can be solved by the branch and bound
method. The exact solution allows estimating the degree of the error of Roy’s approach and of the modifications proposed in
the present paper.

1. MODELING AND OPTIMIZATION OF A FINANCIAL PORTFOLIO

A financial portfolio is described by a vector x = ( x1 ,... , x n ) of costs x i and by a vector of random returns
ì n ü
w = ( w1 , K, w n ) Î W of assets of the form i = 1,... , n over a fixed period of time. Denote by X = í x Î R n : å x i £ 1, x i ³ c i ³ -¥ ý
î i =1 þ
the set of feasible portfolios with the unit total initial cost of the portfolio, c i are the lower constraints for the components of
the portfolio (constraints for the loan of assets). The definition of the set X of feasible portfolios involves the inequality
n n
å xi £ 1, which means that idle assets x 0 = 1- å x i have zero return. The portfolio is characterized by the total random
i =1 i =1
n n
return over the time interval under study f ( x, w ) = w¢ x = å w i x i , average return m ( x ) = E w f ( x, w ) = å x i E w w i , and
i =1 i =1
variance of return s ( x ) = E w ( f ( x, w ) -- E w f ( x, w )) , where E w is the expectation with respect to the distribution of the
2 2

random variable w, and w¢ is the transposed column vector w.

According to G. Markowitz [2, 3], the portfolio is optimized with respect to two criteria: average return m( x ) and
standard deviation of return s( x ) , ( m ( x ), - s ( x )) ® maxxÎX . The set of images of non-dominated portfolios
G = ( y Î R 1 , s * ( y )) such that [ s * ( y )] 2 = min s 2 ( x ) "y > u is called the efficient frontier. The optimal portfolio is
xÎ X , m( x) ³ y
selected from the efficient set, based on the optimization of some utility function F( s , m ) , defined in the risk–return plane ( s , m ) .
1.1. The SF-Approach to the Financial Portfolio Selection. According to Roy [1], portfolio optimization should be
carried out according to the safety-first criterion and consists in the minimization of the probability

Pu ( x ) = Prw { f ( x, w ) £ u} ® min (1)

xÎ X

181
that the return of the portfolio will appear less than some specified critical level u > 0, for example, u = 1. Since
optimizing a probability is a challenge, Roy proposed to minimize its upper-bound estimate obtained from the
Chebyshev inequality,
Pu ( x ) = Prw { f ( x, w ) £ u} = Pw {- f ( x, w ) ³ -u} = Prw {m ( x ) - f ( x, w ) ³ m( x ) - u}

s 2 (x )
£ = Pu ( x ) ® min . (2)
( m( x ) - u ) 2 xÎ X

As Roy noted, in the plane ( s , m ) , s ( x ) / ( m ( x ) - u ) is the cotangent of the angle formed by the line passing through
the points ( 0, u ) and ( s ( x ), m ( x )) , with the horizontal axis s. The maximum such angle corresponds to the minimum
value of Pu ( x ). Thus, the solution of the approximating problem (2) is subdivided into the following subproblems.
Subproblem 1. Constructing an efficient frontier Gu = ( y > u, s * ( y )) such that

[ s * ( y )] 2 = min s 2 ( x ) "y > u. (3)

xÎ X , m( x) ³ y

Subproblem 2. Finding the tangent to the efficient frontier Gu , passing through the point ( 0, u ) ; the tangent point
corresponds to the optimal portfolio x * ( u ) .
Note that within the framework of Roy’s approach, it is possible to consider also the lower bound of the average
return, m( x ) ³ z, where z is a given constant, z > u. To this end, it is sufficient to consider points of the efficient frontier
above the level of z, i.e., points on G z = ( y ³ z , s * ( y )) .
Thus, as a matter of fact, the difference between Roy’s approximation approach (2) and Markowitz’s approach is only
in the way how the optimal portfolio is selected from the set of efficient portfolios G.
1.2. Exact Solution of the Probability Minimization Problem for a Finite Set of Scenarios. Let the random vector
w in (1) take a finite set of values W with probabilities p w , w ÎW. Consider a problem

Pu (x ) = å p w I { f (x, w )<u} ® xÎX min

, m( x) ³ y
, (4)
wÎW
where I { f ( x, w )<u} is the indicator function of event { f ( x, w ) < u}, i.e., I { f ( x, w ) < u} = 1 if f ( x, w ) < u, otherwise
I { f ( x, w )<u} = 0 . It can be easily seen that for functions f ( x, w ) continuous in x, discontinuous functions
I { f ( x, w )<u} (and, hence, P u ( x )) are lower semicontinuous in x ; therefore, for the compact set of feasible portfolios,
problem (4) has a solution. Note that the functions Pu ( x ) from (1) and P u ( x ) from (4) are related by P u ( x ) £ Pu ( x ) ;
therefore, any upper-bound estimates for Pu ( x ) are also true for P u ( x ) . The continuity conditions for the functions
Pu ( x ) and P u ( x ) are available in [13, Sec. 2.2.2; 22; 23, Sec. 7.2.4].
Problem (4) can be solved by the stochastic branch and bound method [24]. We will reduce it to a mixed-integer
(Boolean) programming problem. Let inf f ( x, w ) ³ -M > -¥ and M + u > 0 , where M is some constant. Let us
xÎ X , wÎW
associate each w ÎW with a binary variable z w Î{0, 1} . Consider the problem

å p w z w ® xÎX ,min
z Î {0 ,1}
(5)
w ÎW w

under additional constraints

m( x ) ³ y, (6)

- f ( x, w ) + u £ ( M + u ) z w , w ÎW . (7)

THEOREM 1. Let the set X be compact and W be finite; functions f (×, w ) be continuous for any w ÎW ;
inf f ( x, w ) ³ -M > -¥, and M + u > 0 . Then problems (4) and (5)–(7) are equivalent in the sense that their solutions
xÎ X , wÎW
exist and the optimal values of the objective functions coincide. And if x * is the optimal solution of problem (4), then
( x * , z w* = I { f ( x* , w ) < u} ) is the optimal solution of problem (5)–(7) and vice versa: if ( x * , z w* ) is the optimal solution of
problem (5)–(7), then x * is the optimal solution of (4).

182
 Theorem 1 is a special case of a more general Theorem 2 (see Appendix).
Problem (5)–(7) can be solved numerically by the standard branch and bound method. Branching is carried out with
respect to the Boolean variables. The solution consists in the enumeration on the tree of problems (5)–(7), where
z w Î Z w Í {0, 1}, with the use of the lower- and upper-bound estimates of the optimal values of the objective function (5).
The lower-bound estimates can be found by the continuous relaxation 0 £ z w £ 1 of free Boolean constraints z w Î Z w = {0, 1},
and the upper-bound estimates by fixing the free Boolean variables at the upper values z w = 1.
1.3. Modifying Roy’s SF-Approach with the Use of Downside Measures of Risk. As Markowitz mentioned in [3],
a shortcoming of the standard deviation s( x ) as a risk measure is that s( x ) takes into account (as risks) the events whereby
the return f ( x, w ) exceeds the mean value m ( x ) = Ef ( x, w ) , which bear no relation to risk. This shortage obviously takes
place also in Roy’s approximation approach. Therefore, our first improvement of Roy’s approach consists in replacing the
standard deviation s( x ) with other, more accurate and adequate, downside risk measures. Following Markowitz [3], we will
determine the standard downside semideviation of the random return f ( x, w ) from the mean value m( x ) :

R ( m) ( x ) = E max{0, m ( x ) - f ( x, w )}. (8)

In view of the Chebyshev inequality, the following estimate is valid for u < m( x ) :
Pu ( x ) = Prw { f ( x, w ) £ u} = Prw {- f ( x, w ) ³ - u}
= Prw {m ( x ) - f ( x, w) ³ m ( x ) - u}
£ Prw {max{0, m ( x ) - f ( x, w )} ³ m ( x ) - u}

E| m ( x ) - f ( x, w ) |
r
E max{0, m ( x ) - f ( x, w )}r
£ £ , r > 0. (9)
( m( x ) - u ) r ( m( x ) - u ) r
Thus,
R ( m) ( x )
Pu ( x ) £ = Pu- ( x ) , (10)
m( x ) - u

and in view of (9), the inequality Pu- ( x ) £ Pu ( x ) holds, i.e., estimate (10) is more accurate than (2).
Following Roy’s idea [1], instead of the probability Pu ( x ) , it is possible to minimize its upper-bound estimate (10).
The last problem reduces to the following subproblems.
Subproblem 3. Creating the efficient frontier G- = ( y > u, R (*m) ( y )) such that

R (*m) ( y ) = min R ( m) ( x ) "y > u. (11)

xÎ X , m( x) ³ y

Subproblem 4. Finding the tangent to the efficient frontier G- passing through the point ( 0, u ) .
The computing problem (11) differs from (3) in that the functional R ( m) ( x ) in (11) is not expressed in a closed form in
terms of coefficients of the covariance of return on assets and that R ( m) ( x ) is a nonsmooth function. Problem (11) belongs to
the class of convex stochastic programming problems [25].
Let us consider one more modification of the Roy approach. We introduce an a-quantile q a ( x ) = max{q :
Pr{ f ( x, w ) £ q} £ a} and a risk measure
E max{0, q a ( x ) - f ( x, w )}
R qa ( x ) = w ,
Pr{ f ( x, w ) £ q a ( x )}
which is called average downside deviation from the a-quantile q a ( x ) . For q a ( x ) > u, in view of the Chebyshev
inequality, the inequality
E max{0, q a ( x ) - f ( x, w )}
Pu ( x ) £ w (12)
qa (x ) - u
holds similarly to (9). This estimate can be rearranged as

Pr{ f ( x, w ) £ q a ( x )} E w max{0, q a ( x ) - f ( x, w )} aR qa ( x )
Pu ( x ) £ £ . (13)
qa (x ) - u Pr{ f ( x, w ) £ q a ( x )} qa (x ) - u

183
The median estimate for a = 1/ 2 is a special case of this estimate. The problem is now reduced to the minimization of the
ì ü
right-hand side of (13) which, in turn, is reduced to constructing the efficient frontier í r( y ) = min R qa ( x ), y > uý
î xÎ X , q a ( x ) ³ y þ
and the tangent to it passing through the point ( 0, u ) .
The Roy approximation approach described above and its modifications show how portfolio SF-optimization is
related to traditional and some new downside risk measures. Let us generalize the modifications of Roy’s approximation
approach described above.

2. GENERALIZING THE ROY SF-APPROACH TO THE SAFETY-FIRST OPTIMIZATION

OF A FINANCIAL PORTFOLIO

Let us determine R ( y) ( x ) = E max{0, y - f ( x, w )}. Similarly to (9), for any y > u the estimate
E w max{0, y - f ( x, w )} R ( y) ( x )
Pu ( x ) £ = (14)
y-u y-u
holds; therefore, R ( y) ( x )
Pu ( x ) £ min . (15)
{ y: y>u} y-u
Obviously, estimate (15) is more accurate than estimates (10) and (12), which are obtained from the weaker estimate
(14) by substituting y = m( x ) and y = q a ( x ) , respectively. The lower-bound estimate for Pu ( x ) has the form
u - f ( x, w )
Pu ( x ) = E w I { f ( x, w ) £u} = E w I { f ( x, w ) £u}
u - f ( x, w )

E w ( u - f ( x, w ))I f ( x, w ) £u E w max{0, u - f ( x, w )}
³ = .
u- min f ( x, w ) u- min f ( x, w )
{w ÎW : f ( x, w ) £ u} {w ÎW : f ( x, w ) £ u}

Following Roy [1], instead of Pu ( x ) , we will minimize the upper-bound estimate (15) with respect to x Î X and,
moreover, superimpose the lower constraint z £ m( x ) on the average return m( x ) of the portfolio x. Thus, we obtain problem
with parameters ( u, z ) : R ( y) ( x )
® min . (16)
y-u {( x, y ): xÎ X , m( x) ³ z , y >u}

As a matter of fact, in this formulation we approximately minimize the probability that the portfolio return falls
critically, its average return being constrained from below. Obviously,
min R ( y) ( x )
R ( y) ( x ) xÎ X , m( x) ³ z r( y, z )
min = min = min ,
{( x, y ): xÎ X , m( x) ³ z , y>u} y-u y>u y-u y>u y-u

where the risk function has the form

r( y, z ) = min R ( y) ( x ) = min [ E w max{0, y - f ( x, w )}] .

xÎ X , m( x) ³ z xÎ X , m( x) ³ z
For a fixed z, let us construct the efficient frontier

ì ü
G z = í ( r ( y, z ), y ) Î R 2 , y ³ min f ( x, w )ý .
î xÎ X , wÎW þ

As follows from geometrical considerations, minimum in problem (16) is attained at the point ( r * = r( y * , z ), y * )
where the efficient frontier G z is tangent to the straight line passing through the point ( 0, u ) and having the minimum tilt
angle with the vertical axis y (Fig. 1).

184
 y 0.2
0.1
0
-0.1
-0.2
-0.3 Efficient frontier
Tangent line for u = – 0.07
-0.4
-0.5 r
0 0.02 0.04 0.06 0.08 0.1

Fig. 1. Risk function r( y, z ) for z = 015

. .

Note that if the function of random return f ( x, w ) is concave in x on the convex set X , then max{0, y - f ( x, w )} is
convex in the variables ( x, y ) , and the minimum function rz ( y ) = min R ( y) ( x ) is convex and monotonically increases
xÎ X , m( x) ³ z
in y. Therefore, the efficient frontier is upward concave with respect to the horizontal axis r and is monotonic in r. Thus, if there
exists a portfolio x such that with positive probability its return f ( x, w ) is greater than u, then the problem has a unique solution.
Finding the efficient frontier r ( y ) = min R ( y) ( x ) , y > u, is a convex stochastic programming problem, which
xÎ X , m( x) ³ z
can be solved, for example, by the stochastic quasigradient method [25]. The stochastic subgradient of the function R ( y) (× )
has the form
ì - w, y - wT x > 0,
g ( x, w ) = í
î w, y - w x £ 0,
T

with E w g ( x, w ) Î ¶R ( y) ( x ) , where ¶R ( y) ( x ) is a subdifferential of the function R ( m) (× ) at the point x.

Another approach to the solution of stochastic programming problems is to approximate the distribution of the vector
random variable w with the discrete distribution with the values w s and probabilities p s [26]. Then we can approximate the
expectations in R ( y) ( x ) = E w max{0, y - f ( x, w )} with the mean value
S
R (Sy) ( x ) = å p s max{0, y - f ( x, w s )}
s=1
and solve problems
R (Sy) ( x ) ® min "y > u
xÎ X , m( x) ³ z
instead of finding r( y, z ) = min R ( y) ( x ) .
xÎ X , m( x) ³ z
In the case of empirical approximation, random points {w s } are independent equally distributed vectors (with the same
distribution as the vector w has), and p s = 1/ S .
Note that for linear functions f (×, w ) the minimization of the piecewise linear function R (Sy) ( x ) reduces to linear
programming problems.

3. OTHER APPROACHES TO THE SAFETY-FIRST PORTFOLIO OPTIMIZATION

3.1. Maximizing the Return of the Portfolio under the Constraint for the Probability of the Return Falling
below a Minimum Desired Threshold. Telser [4] considered the following problem:
m ( x ) = Ef ( x, w ) ® maxxÎX (17)
under the constraint
P( u ) ( x ) = Pw { f ( x, w ) £ u} £ a, (18)

185
where f ( x, w ) is the return of portfolio x with the random vector of returns w. Note that the complicated probability
constraint (18) can be replaced with the equivalent rather complicated quantile constraint q a ( x ) ³ u. For the
approximate solution of problem (17), (18), we apply the upper approximation of the probability P( u ) ( x ) and replace
the probability constraint (18) with a stronger constraint:
R ( m) ( x ) aR qa ( x ) R ( y) ( x )
(a) £ a, (b) £ a, (c) min £ a. (19)
m( x ) - u qa (x ) - u y>u y-u

In the case (a), the solution reduces to finding the crosspoint (farthest from the origin of coordinates) of the efficient
frontier G- = ( R (*m) ( y ), y > u ) and the straight line passing through the point ( 0, u ) and having the tangent 1/ a of the tilt
angle with the horizontal axis.
For the finite set of scenarios W, problem (17), (18) can be reduced to a mixed-integer (Boolean) programming
problem. For equiprobable scenarios, constraint (18) means that no more than a | W | constraints f ( x, w ) £ u should be
satisfied in the feasible solution x, where | W | is the total number of scenarios. In discrete programming, there is a method of
reducing such problems to mixed-integer programming problems [27, Ch.2, Sec. 4]. Let inf f ( x, w ) ³ -M > - ¥. We
xÎ X , wÎW
associate each w ÎW with a binary variable z w Î{0, 1}. Consider the problem

Ef ( x, w ) ® maxxÎX , {zw Î{0 ,1}} (20)

under the constraints
å p w z w £ a, (21)
wÎW

- f ( x, w ) + u £ ( M + u ) z w , w ÎW . (22)

The equivalence of problems (17), (18) and (20)–(22) follows from Theorem 2 in Appendix.
3.2. Minimizing the Monetary Reserve (VaR) That Ensures the Satisfaction of the Probability Constraint for
Total Money Resources. Kataoka [5] considered the following problem:

v ® min ,
xÎ X , v

Pu ,v ( x ) = Pw {v + f ( x, w ) £ u} £ a. (23)

Obviously, this problem is equivalent to the maximization of the quantile

q a ( x ) = max{q : Pr{ f ( x, w ) £ q} £ a} ® max .

xÎ X
Similarly to (19), the probability constraint (23) can be replaced with a stronger one by using the upper-bound
estimates for the probability Pu ,v ( x ) . For example, consider the problem
v ® min ,
xÎ X , v

E w max{0, y - f ( x, w )}
min £a.
y>u - v y - (u - v )

As follows from geometrical considerations, the optimum in the last problem is attained on the portfolio x *
corresponding to the point ( r * , y * ) of the efficient frontier G 0 = ìí ( r, y ) Î R 2 : r( y ) = min R ( y) ( x ), y > uüý such that the
î xÎ X þ
tangent of the angle between the horizontal axis and the tangent line to the boundary G 0 is 1/ a, the optimal value of
the objective function being defined as v * = u + r * / a - y * .

186
 TABLE 1

Annual returns on securities

Year
Am.T. A.T. &T. U.S.S. G.M. A.T.& Sfe C.C. Bdn. Frstn. S.S.
1937 -0.305 -0.173 -0.318 -0.477 -0.457 -0.065 -0.319 -0.400 -0.435
1938 0.513 0.098 0.285 0.714 0.107 0.238 0.076 0.336 0.238
1939 0.055 0.200 -0.047 0.165 -0.424 -0.078 0.381 -0.093 -0.295
1940 -0.126 0.030 0.104 -0.043 -0.189 -0.077 -0.051 -0.090 -0.036
1941 -0.280 -0.183 -0.171 -0.277 0.637 -0.187 0.087 -0.400 -0.240
1942 -0.003 0.067 -0.039 0.476 0.865 0.156 0.262 1.113 0.126
1543 0.428 0.300 0.149 0.225 0.313 0.351 0.341 0.580 0.639
1944 0.192 0.103 0.260 0.290 0.637 0.233 0.227 0.473 0.282
1945 0.446 0.216 0.419 0.216 0.373 0.349 0.352 0.229 0.578
1946 -0.088 -0.046 -0.078 -0.272 -0.037 -0.209 0.153 -0.126 0.289
1947 -0.127 -0.071 0.169 0.144 0.026 0.355 -0.099 0.009 0.184
1948 -0.015 0.056 -0.035 0.107 0.153 -0.231 0.038 0.000 0.114
1949 0.305 0.038 0.133 0.321 0.067 0.246 0.273 0.223 -0.222
1950 -0.096 0.089 0.732 0.305 0.579 -0.248 0.091 0.650 0.327
1951 0.016 0.090 0.021 0.195 0.040 -0.064 0.054 -0.131 0.333
1952 0.128 0.083 0.131 0.390 0.434 0.079 0.109 0.175 0.062
1953 -0.010 0.035 0.006 -0.072 -0.027 0.067 0.210 -0.084 -0.048
1954 0.154 0.176 0.908 0.715 0.469 0.077 0.112 0.756 0.185

3.3. Portfolio Optimization as a Lexicographic Optimization Problem. The following lexicographic optimization
problem is considered in [7]:
( p u , a ( x ), m ( x )) ® Lex maxxÎX ,
where

ïì 1- Pr{ f ( x, w ) £ u}, Pr{ f ( x, w ) £ u} > a,

p u, a (x ) = í
ïî 1, Pr{ f ( x, w ) £ u} £ a.
In such an approach, first, the problem
Pr{ f ( x, w ) £ u} ® min
xÎ X
is solved and if it has a feasible solution x * such that Pr{ f ( x * , w ) £ u} £ a , the problem
m( x ) ® min ,
xÎ X
Pr{ f ( x, w ) £ u} £ a

is then solved. In this approach, the probabilities Pr{ f ( x, w ) £ u} can also be replaced with their upper-bound estimates
(10), (12), and (15).
See [28] for the further discussion of the financial portfolio optimization based on safety-first criteria.

4. NUMERICAL EXPERIMENTS

Let us use the numerical example from [3] to compare the solutions of the problem of SF-optimization of a financial
portfolio by the original (Sec. 1.1) and modified (Sec. 2) Roy’s method with the exact solution obtained by the branch and
bound method (Sec. 1.2). In the example, portfolio consists of nine securities of nine large American companies with annual
returns for 18 years (1937–1954) presented in [3, Table 1, p. 13]. Thus, the vectors x and w have nine components each, and
the set of (equiprobable) scenarios W contains 18 elements (row vectors from Table 1).
For the average return z = 015. , Fig. 1. presents the efficient frontier G0.15 = {( r( y, 015
. ), y ) Î R 2 , y ³ -0.477} from
Sec. 2 and also the tangent line to it passing through the point ( 0, u = - 0.07) and corresponding to the optimal portfolio with
u = - 0.07. The tangency is at the point with the coordinates ( 0.0252, 0.0352) to which the optimal portfolio
x * = ( 0, 0, 0.3951, 0, 0.2151, 0, 0.3846, 0, 0.0052) corresponds.

187
 P
Roy’s estimate
Refined estimate
Probability Ï(u)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 u
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2

Fig. 2. Probability P( u ) and its

estimates for z = 01. .

TABLE 2

Structure of the optimal portfolios

Technique
A.T. A.T. &
Am.T. U.S.S. G.M. C.C. Bdn. Frstn. S.S.
&T. Sfe

Markowitz 0 0 0.1152 0.0226 0.084 0 0.4907 0 0

Roy 0 0 0.1154 0.0226 0.0841 0 0.4913 0 0
Approximate SF 0 0 0.3534 0 0.0769 0.0015 0.2590 0 0
Exact SF 0.0582 0.0708 0.0417 0.0869 0.1354 0.0577 0.1843 0.2154 0.0943

For the average return z = 01

. , Fig. 2 presents the graphs of the exact probability function
P( u ) = min Pr{ f ( x, w ) < u},
xÎ X , m( x) ³ z
its Roy’s approximation
P( u ) = min [ s 2 ( x ) / ( m( x ) - u ) 2 ] ,
xÎ X , m( x) ³ z
and the refined estimate from Sec. 2
~
P( u ) = min[ r( y, z ) / ( y - u )] ,
y >u
where
r( y, z ) = min E w max{0, y - f ( x, w )}.
xÎ X , m( x) ³ z
For other values of the average return z the situation is similar; as the parameter z decreases, the curves shift to the
right, as it increases, they shift to the left. The example shows that Roy’s approximation P( u ) is a poor estimate of the
~
function P( u ) in the whole range of values of u. The refined estimate P( u ) is much more accurate; however, it can be
applied only to negative values of u. For the critical level u = -0. 1 and average return z = 0. 1 , Table 2 presents the structure
of the optimal portfolios found by the Markowitz technique (for the average return z = 0. 1 ), Roy technique (Sec. 1.1), and
approximate (Sec. 2) and exact SF-technique (Sec. 1.2).
n
The characteristics of the optimal portfolios from Table 2 are presented in Table 3, where x 0 = 1- å x i , P-0.1 (× ) and
i =1
P0 (× ) are the probabilities of portfolio returns smaller than -0. 1 and 0, respectively, m(× ) is the average return, and s(× ) is the
standard deviation of return.
Tables 2 and 3 allow drawing the following conclusions. The first two (Markowitz and Roy) portfolios differ little. The
structure of the third approximate SF-optimal portfolio is similar to the structure of the Markowitz and Roy portfolios, but this
portfolio is safer since the probability P0 = 0.0556 of negative return is less for it. The fourth SF-optimal portfolio considerably
differs from the others: it is more diversified, safer, has large average return, but is less focused around the average return.

188
 TABLE 3

Characteristics of the optimal portfolios

Technique
~
x0 P P P P-0. 1 (×) P0 (×) m(×) s ( ×)

Markowitz 0.2875 – – – 0.0556 0.1667 0.1 0.1174

Roy 0.2867 0.3448 – – 0.0556 0.1667 0.1001 0.1175
Approximate SF 0 – 0.122 – 0.0556 0.0556 0.1 0.1362
Exact SF 0.0552 – – 0.0556 0.0556 0.0556 0.1323 0.2044

CONCLUSIONS

In the paper, we have improved the Roy approach [1] to the safety-first optimization of financial portfolios. In this
approach, the estimate of the probability of the portfolio’s return falling below a minimum desired threshold is minimized,
the average return being constrained from below. The approaches of Roy and Markowitz to the optimization of a financial
portfolio are rather close and differ only in the way of choosing a point on the same efficient frontier in the risk–return plane
(the variance of return). In the paper, first, we have improved the estimate of the risk of negative returns of the portfolio
using downside threshold measures of risk. Searching for the optimal portfolio reduces to the construction and analysis of
a new efficient frontier of portfolios in the plane threshold–threshold-risk-measure. Constructing this efficient frontier
consists in solving convex stochastic programming problems of the minimization of threshold measures of risk. The optimal
safety-first portfolio can be found by a simple and elegant geometrical Roy’s method. Second, for the case of a finite set of
scenarios of return, the problem of safety-first portfolio selection reduces to a linear mixed-integer Boolean programming
problem where continuous variables correspond to portfolio components and the number of Boolean components coincides
with the number of scenarios. The problem can be solved by the standard branch and bound method. The theoretical
derivations and conclusions are illustrated by a numerical example of searching for a safety-first portfolio with nine
components and 18 scenarios of return.

APPENDIX. REDUCING A PROBABILISTIC OPTIMIZATION PROBLEM WITH

A FINITE SET OF SCENARIOS TO A MIXED INTEGER PROGRAMMING PROBLEM

Let the random vector w in (1) take a finite set of values W with probabilities p w , w ÎW . Denote
Pk ( x ) = å p w I { f k (x, w )<0} , k = 1,... , K , where I { f k ( x, w )<0} is the indicator function of event { f k ( x, w ) < 0} , i.e.,
w ÎW

I { f k ( x, w ) < 0} = 1 if f k ( x, w ) < 0 , otherwise I { f k ( x, w )<0} = 0. Function Pk ( x ) has the meaning of the probability of event
{ f k ( x, w ) < 0}. Consider the problem
G0 ( x, P1 ( x ),... , PK ( x )) ® min , (24)
xÎ X

G j ( x, P1 ( x ),... , PK ( x )) £ 0, j = 1,... , m , (25)

where the functions G j ( x, p 1 , K, p K ), j = 0, 1, ... , m, are monotone (do not decrease) in the arguments
p 1 ³ 0,... , p K ³ 0 and X is a compact set. Optimization of probabilities and optimization with probability constraints
are special cases of problem (24), (25). Then we reduce problem (24), (25) to a mixed-integer programming problem.
Assume that
inf f k ( x, w ) ³ -M kw > -¥.
xÎ X

189
 Let us introduce variables z kw Î{0,1}, z k = {z kw , w Î W} and linear Boolean functions p k ( x, z k ) = å p w z kw .
Consider the problem w ÎW
G0 ( x, p1 ( x, z1 ),... , p K ( x, z K )) ® min , (26)
xÎ X , {zkw }

G j ( x, p1 ( x, z1 ),... , p K ( x, z K )) £ 0, j = 1,... , m , (27)

- f k ( x, w ) £ M kw z kw , z kw Î{0,1}, w ÎW, k = 1,... , K . (28)

THEOREM 2. Let the set X be compact and W be finite; functions f k (×, w ) be continuous for any w ÎW ; functions
G j ( x, p 1 , K, p K ) b e c o n tin u o u s in th e ir ar g u me n ts a n d mo n o to n e ( n o n d e c r e a s in g ) in p 1 , K, p K ;
inf f k ( x, w ) ³ -M kw > - ¥, and M kw > 0 . Then problems (4) and (5), (7) are equivalent in the sense that their solutions
xÎ X
exist and the optimal values of the objective functions coincide. If x * is the optimal solution of problem (24), (25), then
( x * ,{z k*w = I { f ( x* , w )<0} }) is the optimal solution of problem (26)–(28) and vice versa, if ( x * ,{z kw
*
}) is the optimal solution
k

of problem (26)–(28), then x * is the optimal solution of (24), (25).

Proof. Let us prove the equivalence of the optimization problems (24), (25) and (26)–(28) according to [29, p. 131],
i.e., let us show that optimal solutions of the problems simultaneously exist or do not exist, establish the correspondences
between the optimal solutions, and prove that if the solutions exist, the optimal values of the objective functions coincide. To
this end, it will suffice to specify for each feasible point of one problem a feasible point of the other problem where the
objective function is not greater.
Let ( x, {z kw }) be a feasible solution of problem (26)–(28). If f k ( x, w ) < 0, then it follows from (28) that z kw = 1.
Otherwise (28) does not impose constraints on z kw Î{0,1}. Thus,

p k ( x, z k ) = å p w z kw ³ å pw I{ f } = Pk ( x ) ,
k ( x , w )<0
wÎW wÎW
and since the functions G j ( x, p 1 , K, p K ) are monotone in the arguments p 1 ³ 0,... , p K ³ 0 ,

G j ( x, p1 ( x, z1 ),... , p K ( x, z K )) ³ G j ( x, P1 ( x ),... , PK ( x )) .

Thus, we associate each feasible solution ( x, {z kw }) of problem (26)–(28) with a feasible solution x of problem (24),
(25), where the value of the objective function is not greater, G0 ( x, P1 ( x ),... , PK ( x )) £ G0 ( x, p1 ( x, z1 ),... , p K ( x, z K )) .
Vice versa, let x be an arbitrary feasible solution of problem (24), (25). Calculate z kw = I { f ( x, w )<0} for all k and w.
k
Obviously,
G j ( x, P1 ( x ),... , PK ( x )) = G j ( x, p1 ( x, z1 ),... , p K ( x, z K )) .
Thus, we associate each feasible solution x of problem (24), (25) with a feasible solution ( x, {z kw = I { f k ( x, w )<0} }) of
problem (26)–(28), where the value of the objective function is not greater.
Hence, problems (24), (25) and (26)–(28) simultaneously either have or do not have feasible solutions.
Under conditions of the theorem in problem (24), (25) functions I { f k ( x, w )<0} , G j ( x, P1 ( x ),... , PK ( x )) , j = 0,1,... , m , are
lower semicontinuous and the set X is compact. Therefore, if the feasible solution of problem (24), (25) exists, its optimal
solution x * exists as well. In problem (26)–(28), functions ( f k ( x, w ) + M kw z kw ) and G j ( x, p1 ( x, z1 ),... , p K ( x, z K )) are
continuous and the set X and the set of possible values {z kw } are compact. Therefore, if the feasible solution of problem
(26)–(28) exists, its optimal solution ( x * , {z kw
*
}) exists as well. Hence, in view of the established correspondence of the
feasible solutions x « ( x, {z kw = I { f k ( x, w ) < 0} }) , if problems (24), (25) and (26)–(28) are feasible, the optimal values of
their objective functions coincide. Moreover, if x * is the optimal solution of problem (24), (25), then
( x * , z k*w = I { f ( x* , w )<0} ) is the optimal solution of problem (26)–(28). Vice versa, if ( x * , {z kw
*
}) is the optimal solution of
k

problem (26)–(28), then x * is the optimal solution of problem (24), (25). The theorem is proved.

190
REFERENCES

1. A.D. Roy, “Safety first and the holding of assets,” Econometrica, 20, 431–449 (1952).
2. H. Markowitz, “Portfolio selection,” The J. of Finance, 7, No. 1, 77–91 (1952).
3. H.M. Markowitz, Portfolio Selection. Efficient Diversification of Investments, John Wiley & Sons, New York;
Chapman & Hall, London (1959).
4. L.G. Telser, “Safety first and hedging,” Review of Economic Studies, 23, 1–16 (1955/56).
5. S. Kataoka, “A stochastic programming model,” Econometrica, 31, 181–196 (1963).
6. V.S. Bawa, “Safety first, stochastic dominance, and optimal portfolio choice,” J. of Financial and Quantitative
Analysis, 13, 255–271 (1978).
7. E.R. Arzac and V.S. Bawa, “Portfolio choice and equilibrium in capital markets with safety first investors,” J. of
Financial Economics, 4, 277–288 (1977)).
8. D. Li, T.F. Chan, and W.L. Ng, “Safety-first dynamic portfolio selection,” Dynamics of Continuous, Discrete and
Impulsive Systems, 4, 585–600 (1998).
9. D.W. Jansen, K.G. Koedijk, and C. G. de Vries, “Portfolio selection with limited downside risk,” J. of Empirical
Finance, 7, 247–269 (2000).
10. T. Goodall, Adequate Decision Rules for Portfolio Choice Problems, Palgrave, New York (2002).
11. M. Haque, M. K. Hassan, and O. Varela, “Safety-first portfolio optimization for US investors in emerging global,
Asian and Latin American markets,” Pacific-Basin Finance Journal, 12, 91–116 (2004).
12. M. Haque, O. Varela, and M.K. Hassan, “Safety-first and extreme value bilateral U.S.–Mexican portfolio
optimization around the peso crisis and NAFTA in 1994,” The Quarterly Review of Economics and Finance, 47,
449–469 (2007).
13. A.I. Kibzun and Y.S. Kan, Stochastic Programming Problems with Probability and Quantile Functions, John Wiley &
Sons, Chichester–New York–Brisbane–Toronto, Singapore (1996).
14. A.I. Kibzun and E.A. Kuznetsov, “Optimal control of the portfolio,” Automation and Remote Control, 62, No. 9,
1489–1501 (2001).
15. A.I. Kibzun and E.A. Kuznetsov, “Positional strategy of forming the investment portfolio,” Automation and Remote
Control, 64, No. 1, 138–152 (2003).
16. A.I. Kibzun and R. Lepp, “Discrete approximation in quantile problem of portfolio selection,” in: S. Uryasev and
P.M. Pardalos (eds.), Stochastic Optimization: Algorithms and Applications, Kluwer, Dordrecht (2001), pp. 119–133.
17. Yu.S. Kan, “Application of the quantile optimization to bond portfolio selection,” in: K. Marti (ed.), Stochast.
Optim.Techniques: Numerical Methods and Technical Appl., 513, Springer, Berlin (2002), pp. 285–308.
18. P.V. Grigor’ev and Yu.S. Kan, “Optimal control of the investment portfolio with respect to the quantile criterion,”
Automation and Remote Control, 65, No. 2, 319–336 (2004).
19. Yu.S. Kan and A.N. Krasnopol’skaya, “Selection of a fixed-income portfolio,” Automation and Remote Control, 67,
No. 4, 598–605 (2006).
20. D. Bayuk, “Russian regress. Whether a massive outflow of scientists abroad remains a great problem for Russia?”
Troitskii Variant, No. 22(41), 6–7 (2009).
21. U. Beck, Risikogesellschaft. Auf dem Weg in eine andere Moderne [in German] (Risk Society: Towards a New
Modernity), Suhrkamp, Frankfurt a.M. (1986).
22. E. Raik, “On the stochastic programming problem with the probability and quantile functionals,” Izv. AN ESSR,
Physics–Mathematics, 21, No. 2, 142–148 (1972).
23. A. Shapiro, D. Dentcheva, and A. Ruszczynski, Lectures on Stochastic Programming: Modeling and Theory, SIAM,
Philadelphia (2009).
24. V. Norkin, “Global optimization of probabilities by the stochastic branch and bound method,” in: Stochastic
Optimization: Numerical Methods and Technical Applications, Lecture Notes in Economics and Mathematical
Systems, 458, Springer, Berlin (1998), pp. 186–201.
25. Yu.M. Ermoliev, Stochastic Programming Methods [in Russian], Nauka, Moscow (1976).
26. A. Ruszczynski and A. Shapiro, Stochastic Programming, Elsevier, Amsterdam (2003).
27. A.A. Korbut and Yu.Yu. Finkel’shtein, Discrete Programming [in Russian], Nauka, Moscow (1969).
28. J.-L. Prigent, Portfolio Optimization and Performance Analysis, Chapman&Hall/CRC (2007).
29. V.S. Mikhalevich, A.M. Gupal, and V.I. Norkin, Methods of Nonconvex Optimization [in Russian], Nauka, Moscow (1987).

191