Title: Introduction to Multiple Criteria Portfolio Selection Part 1: Conventional Portfolio Selection
1Introduction to Multiple Criteria Portfolio
SelectionPart 1 Conventional Portfolio
Selection
- Ralph E. Steuer (University of Georgia)
- Yue York Qi (University of Georgia)
- Markus Hirschberger (University of
Eichstätt-Ingolstadt)
2Overview
- Conventional portfolio selection.
- Solve mean-variance formulation for all efficient
points. - Plot efficient frontier.
- Show to investor.
- Investor selects most preferred point on frontier
as optimal portfolio. - What is called the efficient set will be called
the nondominated set. - Developments in multiple criteria portfolio
selection.
Because conventional portfolio selection is
discussed to form a basis for multiple criteria
portfolio selection, it will be motivated a bit
differently.
3Conventional Portfolio Selection
- Assume
- A sum to be invested
- Beginning of holding period, end of holding
period - n securities in which it is possible to invest
- x (x1,, xn) is an investment proportion vector
(also called a portfolio) - As for the number of securities n
- many large institutions have approved lists
where n is anywhere from several hundred to a
thousand - when attempting to form a portfolio to mimic a
large broad based index (like SP500, EAFE,
Wilshire 5000), n can be up to several thousand.
4- Securities have attributes over the holding
period - ri is the percent return of i-th security
- Other attributes could be
- di is dividends of i-th security
- si is social responsibility of i-th security
- qi is liquidity of i-th security
- gi is growth in sales of i-th security
- ai is amount invested in RD of i-th security
- etc.
- Conventional portfolio selection hardly pays any
attention to attributes other than ri . - Note that all of the above-mentioned attributes
are stochastic (which we will assume are random
variables that come from distributions)
5- Trying to make as much money as possible and
picking from the n securities, investors problem
is to solve - where
- rp is percent return on a portfolio over the
holding period. - S, the feasible region in decision space, is
often as simple as - What makes portfolio selection interesting is
that the quantity rp to be maximized in (Orig) is
a random variable (as rp is a function of the
individual-security ri random variables). - Hence, (Orig) is a stochastic programming problem
-- and how to solve a stochastic programming
problem is not well-defined.
(Orig)
or
6- Strategy is to convert the stochastic program
- to an equivalent deterministic problem.
-
- But to do so, decisions must be made.
- Those made in conventional portfolio selection
include - that the means mi , variances sii and
covariances sij of the ri are known (this is a
strong assumption but is standardly done in the
theory of conventional portfolio selection) at
the beginning of the holding period - that maximizing expected utility leads to a
optimal solution - Also, conventional portfolio selection assumes
that - it suffices for expected utility to be given by
EU(rp) Erp l Varrp - which means that optimal portfolio involves can
be identified by a trade-off between the mean and
variance of random variable rp - each investors l ? 0 may be different and is not
known beforehand - l can be viewed as a risk tolerance parameter
(Orig)
7- Assuming U is Taylor series expandable, theory
guarantees the validity of EU(rp) Erp l
Varrp whenever - U is increasing and quadratic, or
- r (r1,, rn) follows the multinormal
distribution - First condition is proved quite easily. Proof of
second condition is more difficult. - To maximize expected utility EU(rp) Erp l
Varrp , the problem is to solve - (1) is to be solved for all l ? 0 because the
true value of l to use only becomes revealed
after all x-solutions have been processed.
(1)
8- Processing begins by forming the (Stdrp, Erp)
combination for each of the x-solutions generated
by (1). - In portfolio optimization, theory and computation
are generally conducted in terms of variance
while solutions are displayed to investors in
terms of standard deviation. - In multiple criteria optimization terms, a
(Stdrp, Erp) combination is a criterion
vector - The set of all (Stdrp, Erp) criterion vectors
resulting from (1) form precisely the set of all
such criterion vectors that are nondominated in
(Stdrp, Erp) space. - A combination is nondominated iff there exists no
x ? S whose combination is the same or better in
all components, and strictly better in at least
one.
9- Plotting all (Stdrp, Erp) images of all of
the x-solutions of (1) results in the efficient
frontier -- but what we will call the
nondominated frontier. - After viewing the nondominated frontier, the
investor selects his or her most preferred
criterion vector on the curve. - Taking an inverse image of the selected (Stdrp,
Erp) criterion vector, we obtain an x ? S that
specifies the investors optimal portfolio. - People are sometimes interested in the minimum
variance boundary of the set of all (Stdrp,
Erp) criterion vectors.
Erp
nondominated portion of the set of all (Stdrp,
Erp) criterion vectors
Stdrp
10- Plotting all (Stdrp, Erp) images of all of
the x-solutions of (1) results in the efficient
frontier -- but what we will call the
nondominated frontier. - After viewing the nondominated frontier, the
investor selects his or her most preferred
criterion vector on the curve. - Taking an inverse image of the selected (Stdrp,
Erp) criterion vector, we obtain an x ? S that
specifies the investors optimal portfolio. - People are sometimes interested in the minimum
variance boundary of the set of all (Stdrp,
Erp) criterion vectors.
Erp
minimum variance boundary of the set of all
(Stdrp, Erp) criterion vectors
Stdrp
11- There are several ways to solve
- where
- S is the n ? n covariance matrix (of the sij)
- m the n-vector of the mi
- One is use use Markowitzs critical line
algorithm, which is a form of parametric
quadratic programming. Many people find this
algorithm difficult to learn. - Other than for Markowitzs team, finance and
operations research have largely taken a pass on
this algorithm.
12- Another is to view (1) as a multi-criteria
optimization program
feasible region in criterion space
feasible region in decision space
where the endeavor is to compute all efficient
points in S in order to obtain the nondominated
portion of Z.
13- S and Z when
- and n 2 securities
- feasible region
Erp
Stdrp
- Extent to which Z bows toward the vertical axis
depends upon how uncorrelated the returns of the
two securities are. - Z is a hyperbola (Stdrp, Erp) space, or a
parabola in (Varrp, Erp) space.
14- S and Z when
- and n 2 securities
- feasible region
Erp
Stdrp
- Dont need mathematical programming whenever
- Everything can be computed by formulas.
15Let
Then, for instance,
and
16- S and Z when
- and n 2 securities
- feasible region
Erp
Stdrp
- where z1 (Stdr1, Er1) plots the standard
deviation and expected value of the return of the
1-st security. Same for z2 but for the 2-nd
security.
17- S and Z when
- and n 2 securities
- feasible region
Erp
Stdrp
- Need mathematical programming whenever S is more
complicated than
18- S and Z when
- and n 3 securities
- feasible region
Erp
Stdrp
19- S and Z when
- and n 3 securities
- feasible region
Erp
Stdrp
20- S and Z when
- and n 3 securities
- feasible region
Erp
Stdrp
- Efficient set is piecewise linear, while
nondominated set is piecewise hyperbolic. - All portfolios on the same nondominated
hyperbolic segment (or on the same efficient
linear segment) involve the same securities, just
in different proportions.
21Introduction to Multiple Criteria Portfolio
SelectionPart 2 Traditional Dotted
Representations
- Ralph E. Steuer (University of Georgia)
- Yue York Qi (University of Georgia)
- Markus Hirschberger (University of
Eichstätt-Ingolstadt)
22- However, the most popular approach is to attempt
a solution of
(1)
by systematically varying r across the interval
a, b in the e-constraint formulation
(2)
where a, b is a large enough interval to
generate the nondominated frontier.
- That is, solve (2) for about 10 to 20
different values of r from a, b. - With regard a, b, b is easy to obtain but a
can be more difficult. - For convenience, we will call (2) the
r-constraint formulation.
23- Solving (2) repetitively results in a dotted
representation of the nondominated frontier - By enlarging a, b, minimum variance boundary
can be obtained
Erp
Stdrp
Erp
bottom part requires r-constraint to be
Stdrp
24- Why have people opted for the r-constraint
formulation? - Because
- process of obtaining nondominated frontier is,
perhaps, more transparent. - by using (2), software from Solver, Matlab,
Cplex, can be employed. - Double duty can be obtained from (2). By
converting the r-constraint in (2) to and
assuming S x ? Rn 1Tx 1, everything about
the nondominated frontier and minimum variance
boundary can be computed in closed form. - parametric quadratic programming, which involves
advanced mathematical programming, can be avoided.
(2)
25r-Constraint Approach
26.16
27.14
28.12
.12
29.10
.10
Can stop because Erp of the point computed is
less than r-value used.
30.08
.08
31connecting the dots
32Tangency Solution Approach
33.10
34.10
35.10
36.07
37.07
38.07
39-.05
40-.05
41-.05
42Cant get nose, though, because that would
require a c ?
43- But there are disadvantages of the two procedures
- each repetition winds up being solved from
scratch - while each dot produced is on the nondominated
frontier, the full - nondominated frontier is not computed
- in tangency solution approach, difficult to
control for dots equally spaced by expected
return - Furthermore, the r-constraint approach of (2)
- does not reveal the interesting structure of the
nondominated frontier - is not easily generalizable to multiple criteria
portfolio selection
44Introduction to Multiple Criteria Portfolio
SelectionPart 3 Structure of Nondominated Sets
- Ralph E. Steuer (University of Georgia)
- Yue York Qi (University of Georgia)
- Markus Hirschberger (University of
Eichstätt-Ingolstadt)
45- Actual nondominated frontiers are piecewise
hyperbolic. The dots on the nondominated
frontier below are turning points, where one
hyperbolic segment ends an another begins. - When S is more complicated than x ? Rn 1Tx
1, nondominated frontiers tend not to be as
perfectly smooth as one gets the impression from
textbooks.
Erp
from a problem with n 12
Stdrp
46- Lets step back and look at the process we have
been following.
to only make money
overall focus
random variable to be maximized
expected utility assumption
equivalent deterministic objectives
- How is it that Varrp and Erp are
deterministic? Recall that all mi , sii and
sij are assumed known at the beginning of the
holding period - Note that U has only one argument that being rp
47- One might wonder why not to incorporate multiple
objectives into portfolio selection using
constraints? - For example, liquidity ?
- However, RHSs like are essentially judgment or
guesses. - Since adding extra constraints only reduces the
size of the feasible region, new nondominated
frontier can only be lower than the old one. - Since nondominated set is a surface, what is our
chance of guessing a line across the nondominated
surface that hits an optimal criterion vector of
an investor?
Erp
Stdrp
48- One might wonder why not to incorporate multiple
objectives into portfolio selection using
constraints? - For example, liquidity ?
- However, RHSs like are essentially judgment or
guesses. - Since adding extra constraints only reduces the
size of the feasible region, new nondominated
frontier can only be lower than the old one. - Since nondominated set is a surface, what is our
chance of guessing a line across the nondominated
surface that hits an optimal criterion vector of
an investor?
Erp
Stdrp
49Multiple Objective Portfolio Selection
- Say that in addition to
- investor has additional stochastic
objectives such as
max rp portfolio return
max dp dividends max qp liquidity max
gp growth in sales max sp social
responsibility max tp portfolio return over
some benchmark
50- With multiple stochastic original objectives, the
process becomes
build an optimal well rounded portfolio
overall focus
random variables to be maximized
expected utility assumption
equivalent deterministic objectives
- Distinguishing characteristic here is that U has
multiple arguments.
51- Conventional (1-quadratic 1-linear)
Erp
piece-wise hyperbolic
Stdrp
efficient set is a path of connected line
segments in S
- Multiple criteria (1-quadratic 2-linear)
Erp
platelet-wise hyperboloidic
Edp
Stdrp
efficient set is a set of connected polyhedral
subsets in S
52- To compute all hyperboloidic platelets of a
1-quadratic q-linear program, q 2, we have been
writing a computer code. Intended for the
academic public domain (not ready yet), it is
called CIOPS. - Some results for the 1-quadratic 1-linear case
(when the covariance matrices are 100 dense and
diagonal) are
- Times in seconds on Dell 2.13GHz Centrino laptop
(sample size 10 for all cells) - Until early 90s, because of CPU time limitations,
about the only portfolio problems larger than n
500 that could be solved were those that
possessed covariance matrices sparse in nonzero
elements.
53Introduction to Multiple Criteria Portfolio
SelectionPart 4 Emerging Research
- Ralph E. Steuer (University of Georgia)
- Yue York Qi (University of Georgia)
- Markus Hirschberger (University of
Eichstätt-Ingolstadt)
54- One might ask where did all of the covariance
matrices come from for the tests. - From a random covariance matrix generator.
- Distributions of variances and covariances in a
typical real portfolio-selection covariance
matrix.
55?
Assume investor is interested in optimizing both
portfolio return and dividends
?
Under appropriate assumptions, expected utility
is given by
?
Because of two unknown parameters, equivalent
deterministic problem is
?
Erp
- actual nondominated surface of
- 111 platelets
- 130 platelet corner points
- for an n 50 problem
Stdrp
E(dp)
56- With multiple stochastic original objectives, the
process becomes
build an optimal well rounded portfolio
overall focus
random variables to be maximized
expected utility assumption
equivalent deterministic objectives
- Distinguishing characteristic here is that U has
multiple arguments.
57- For 1-quadratic 2-linear problems, we have
- Are able to compute all platelet corner points
(which are the images of the polyhedral subsets
of S corresponding to the platelets. - Number of unique platelet corner points is nearly
the same as number of nondominated platelets.
58- Consider the 2-quadratic 1-linear
- Strategy for generating discrete representations
of the nondominated set -
- When S1 and S2 are positive definite, any convex
combination -
- is positive definite.
- Thus, nondominated set of (b) is a subset of
nondominated set of (a) -
- Solve (b) for a hundred different convex
combinations of S1 and S2
(a)
(b)
59- Consider the 2-quadratic 2-linear
- Strategy for generating discrete representations
of the nondominated set -
- Let be a convex combination of S1 and S2
-
- Then, nondominated set of (d) is a subset of
nondominated set of (c) -
(c)
(d)
60Future Research
- We think strategy can be applied to 2-quadratic
2-linear, and beyond.
- We believe hundreds of thousands of points can be
searched using
versions of the following interactive procedures
from multiple criteria optimization
- projected line search (Korhonen and Karaivanova)
- dispersed probing (Tchebycheff Method)
- classification (Miettinen)
- interactive decision maps (Lotov)
61- Other original objectives that are deterministic,
but harder to implement than stochastic
objectives. Note their portfolio-as-a-whole
nature.
min pa deviations from given asset allocation
percentages min pn number of securities in
portfolio min pt turnover min pw
maximum investment proportion weight min ps
amount of short selling min ph number
securities sold short
62The End