This action might not be possible to undo. Are you sure you want to continue?
Moments
*
GustavoM.deAthaydeandRenatoG.FlôresJr.
BancoItaúS.A.,SãoPauloEPGE/FundaçãoGetulioVargas,Rio
ABSTRACT
Wediscussgeometricpropertiesrelatedtotheminimisationofaportfoliokurtosisgiven
itsfirsttwooddmoments,consideringarisklessassetandallowingforshortsales.The
findingsaregeneralisedfortheminimisationofanygivenevenportfoliomomentwith
fixed excess return and skewness,and thenfor the case in which only excess return is
constrained.Anexamplewithtworiskyassetsprovidesabetterinsightontheproblems
relatedtothesolutions.Theimportanceofthegeometricpropertiesandtheiruseinthe
highermomentsportfoliochoicecontextishighlighted.
JELclassification:C49;C61;C63.
(Thisversion:August5,2002)
*
This is a paper specially prepared for the MultiMoment Capital Asset Pricing Models and
Related Topics Workshop held in Paris on April 29, 2002. We thank all the participants for
comments and lively related discussions and, particularly, Emmanuel Jurczenko and Bertrand
Maillet, the workshop organisers on behalf of the FinancesurSeine Association. All ideas,
affirmativesandmistakesareourown.
1. Introduction.
Portfolio optimisation taking into account more than the first two moments has been
receivingrenewedinterestinthepastyears.Beitonthetheoreticalside–includingits
links with the CAPM extensions , or on what relates to econometric tests or updates
based on higher conditional moments, works like Adcock (2002), Adcock and Shutes
(1999), Athayde and Flôres (1997, 2001, 2002), Jurczenko and Maillet (2001, 2002),
Pedersen and Satchell (1998), or Athayde and Flôres (2000), BaroneAdesi (1985),
Harvey and Siddique (1999, 2000), Hwang and Satchell (1999) and Pedersen and
Satchell (2000), far from exhausting the full list of contributions, pay good witness to
thegrowingawarenessoftheimportanceofhighermomentsinbothlinesofresearch.
Since Athayde and Flôres (1997), we have developed a systematic way to treat
the key optimisation problems posed to anyone dealing with higher moments in
portfolio theory. The approach uses a new notation to represent any moments tensor
related to a multivariate random vector of asset returns, and can be used either in a
utility maximising context or if optimal portfolios are defined by preference relations.
The new notation seemed necessary in order to treat the problem in an absolutely
general setting, which means both in the maximum order p of portfolio moments of
interestandinthepossiblepatternsoftheskewnessorhigherordertensors.Thelatteris
crucial as many works generalising the subject consider only the marginal higher
moments of the returnsvector, plainly disregarding any comoment of the same order.
Thoughthefullsetofcomomentscanquicklybecomeprohibitive–what,beyondother
questions,mayposeseriouseconometricestimationproblemsfortheapplications,and
simplifyinghypothesesonitspatternwillusuallybeimposedinpractice,itisimportant
to have a way to study the general solution to the problem, irrespective of further
assumptionsthatmightbeimposed.
The utility function approach, given its more rigid theoretical constraints, and
the debates involving any nonnormalityimplying (utility) function proposed, seems
more suitable for theoretical developments related, for instance, to the CAPM.
Preferenceorderingofportfolios,maderigorousbyScottandHorvath(1980),canlead
tomoreinterestingresultsinthestrictportfoliooptimisationcontext.
In this paper, we discuss an interesting geometric structure that arises when
optimisinganevenmomentsubjecttooddmomentsconstraints.Asusual,agents“like”
oddmomentsand“dislike”evenones.
Thestructurestudied–nottheonlyrelevantoneinthehighermomentscontext
–bearsimportantconsequencesand shedslightonthegeometryofefficientportfolios
sets in moments space. We believe that its implications have not been fully exploited
yet.Moreover,finaltestingofthegainsbroughtoutbyusinghighermomentsreliesin
extensive practical applications of the new results. These, in turn, require proper
software tools for solving the nonlinear systems and optimisation problems involved.
Better knowledge of the surfaces (or manifolds) related to them may greatly improve
thesoftwaredesign.
This paper is organised as follows. The next section discusses the optimisation
ofvariance,andthenkurtosis,giventhefirstandthirddesiredportfoliomoments;while
sectionthreediscusseshowtheseresultscouldstillbegeneralised.Sectionfourdraws,
through an example, a few more properties and analyses the sensitivity of certain
solutions. The final section concludes by explaining how the resultscan be useful ina
duality context and sets a few lines of further research. An Appendix provides a brief
explanationofthenotationused.
2. Minimalvariancesandkurtosessubjecttothefirsttwooddmoments.
Even moments, being always nonnegative, are duly associated with spread, and both
varianceandkurtosisareusedassimplenumericalsummariesofthedispersionofaset
of observations. For fixed portfolio return and skewness, the latter should perhaps be
more used in practice as an alternative objective function, given the frequency with
whichthefattailedeffectinstockreturnshasbeendetected.Ifweminimisethefourth
moment, we shall be directly attacking the heavy extremes of the density, the ultimate
culpritsofthehighvolatilityanduncertaintyofreturns.Mostmeasuresofriskfocused
ontheworstscenarios,liketheVaR,wouldprobablybemoresensitivetovariationsin
the fourth moment rather than in variance. This sort of behaviour will be further
examined in the example in section 4. The material in this section draws on parts of
AthaydeandFlôres(2001)–whereacompletesolutiontothethreemomentsportfolio
problemisfound–andAthaydeandFlôres(2002),forthedevelopmentsrelatedtothe
kurtosis;proofsomittedherecanbefoundinthesepapers.
2.1.Homotheticpropertiesoftheminimumvarianceset.
Minimising the variance, for a given mean return and skewness, amounts to find the
solutiontotheproblem:
[ ] )) ( ( )] ) 1 1 ( ( ) ( [
3
,
2
,
1
,
1 2
,
3
α α α σ λ α α λ α α
α
⊗ − + − + − + = M r M r E M L Min
p
f p
,(1)
whereM
1
, M
2
and M
3
are, resp., the matrices related to the first, second and third
moments tensors
1
, α is the vector of n portfolio weights – where short sales are
allowed,r
f
istherisklessrateofreturn,[1]standsforanx1vectorof1’s,thelambdas
are Lagrange multipliers and the two remaining symbols are the αportfolio (given)
meanreturnandskewness.
Callingx=M
1
–[1]r
f
,thevectorofmeanexcessreturns,and
R=E(r
p
)–r
f
theset(excess)portfolioreturn,
thesolutionto(1)isfoundbysolvingthenequationsnonlinearsystem,
) (
) ( ) (
3
2
2 4 0
2 0
2
2 4 0
2 4
2
3 3
α α
σ σ
α ⊗
−
−
+
−
−
= M
A A A
R A A
x
A A A
A R A
M
p p
.(2)
wherethescalars:
x M x A
1
2
,
0
−
= ,
) (
3
1
2
,
2
α α ⊗ =
−
M M x A ,(3)
) ( ) (
3
1
2 3
,
4
α α α α ⊗ ⊗ =
−
M M M A ,
havesubscriptscorrespondingtotheirdegreeofhomogeneityas(real)functionsofthe
vector α. A
0
and A
4
, in particular, are positive because the inverse of the covariance
matrixispositivedefinite.
Premultiplying(2)bytheverysolutions
,
α ,givestheoptimalvariance(s):
2
2 4 0
2
0 2
2
4
) (
) ( 2
3 3
2
A A A
A R A R A
p p
p
−
+ −
=
σ σ
σ ,(4)
anexpressionwhereboththenumeratoranddenominatorarepositive.
Thefollowingpropositionisfundamental:
Proposition1:Foragivenk,let α definetheminimumvarianceportfoliowhenR=1
and k y
p
= =
3
3
3
σ , and
2
p
σ be the corresponding minimum variance, THEN for all
optimal portfolios related to return and skewness pairs (R,
3
p
σ ) such that
3 3
3
R k
p
= σ ,
or kR y =
3
,thesolutionto(1)willbe R α α = ,withcorrespondingminimumvariance
2
2 2
R
p p
σ σ = .
The above result implies that along the direction defined in the returns x
skewnessplanby kR y =
3
,theoptimalvarianceasafunctionoftheexcessreturnwill
be a parabola. Taking now the three dimensional (3D) space where the standard
deviation
2
2
2
y
p
= σ axisisadded,inthehalfplaneformedbyaspecificdirectionkin
Rxy
3
space
2
andthepositivepartofthestandarddeviationaxis,theoptimalportfolio
surfacewillbereducedtothestraightline 0 ,
1
2
2 2
≥
+
= u
k
u
y y
p p
3
.As
2
p
y differs
withk,theanglethatthislinemakeswiththestandarddeviationaxisvariesalsowithk,.
ThePropositionhasthenafarreachingconsequence:theoptimalsurfaceinthe
positive standard deviation (sd) half of 3D space bears a homothetic property from
whateverstandpointoneassumes.Slicingthesurfacebyasequenceofplanesparallelto
thetwooddmomentsaxeswillgenerateasequenceofcurvesstartingattheoriginand
whoseexpansionratiowillbeequaltothatoftherespective(constant)variancevalues.
Of course, slicing it by planes parallel to the sd and (standardised) skewness axes will
1
SeetheAppendixforafurtherexplanationonthenotationused.
2
Weshall,fromnowon,usetheangularcoefficientktonamethecorrespondingline/directioninthefirst
quadrantoftheRxy
3
plane.
produce a sequence of homothetic curves whose expansion ratio will be that of the
(excess) returns associated to each plan. Finally, inspection of formula (4) easily
convinces that for the last combination, i.e. planes parallel to the sd and mean returns
axes, the same will apply, as Proposition 1 is also true if the role of returns and
skewnessarereversed.
Thepropositionbelowisadirectconsequenceofthisimportantfact:
Proposition2:Foragivenlevelofy
2
(orR,ory
3
),cuttheoptimalsurfacewithaplane
orthogonal to the sd (or returns, or standardised skewness) axis and project the
intersectioncurveinthe‘returnsxskewness(orsdxskewness,orreturnsxsd)plane’,
THEN
iftheyexist,thedirectionsintheRxy
3
(ory
2
xy
3
,orRxy
2
)halfplanerelatedtothe
highestandlowestvalue,ineachaxis,ofthecurveareinvariantwithy
2
(orR,ory
3
).
The qualification if they exist is important as, specially in the case of cuts
paralleltothesdaxis,atleastpartofthecurvemaygotoinfinity.Forconstantvariance
cuts, it may be shown that closed curves will be produced
4
. Indeed, for this case, the
highestandlowestdirectionsareparticularlynoteworthy,asdemonstratedby
Proposition3:ThedirectionintheRxy
3
halfplanethatgivesthehighestRforallthe
minimumvarianceportfolioswiththesamestandarddeviationy
2
isuniqueandrelated
to the celebrated (Markowitz’s) Capital Market Line (CML). Moreover, in this
direction, the skewness constraint to programme (1) is not binding. As regards
3
Thevariableustandsforthecoordinatesalongtheaxisdefinedbythe“directionk”.
4
Theproofisrathertechnicaltobeincludedinthistext.
skewness, though there may be more than one “highest” (and “lowest”) direction, the
constraintpropertyalsoapplies.
This means that the unique solution to the minimum variance portfolio, for a
givenmeanreturn:
x M
A
R
1
2
0
−
= α ,(5)
that defines the famous Capital Market Line in mean x variance space, relating the
optimalvariancetothegivenR,
0
2
2
A
R
R
= σ ,(6)
also defines the (unique) direction that will pass through all the points, in each curve,
yielding the maximum mean return. In other words, in the R x y
3
half plane, this
direction is the geometric locus of all the tangency points between each (projected)
curveandastraightline,paralleltotheskewnessaxis,whichcutsthemeanreturnaxis
inthemaximummeanreturnportfoliovaluerelatedtothesetvariance(thatdefinesthe
cut). This last statement is ensured by the wellknown duality result in Markowitz
world.
Skewnessesandakcanalsobeassociatedtotheseoptimalportfolios,being
evident that they are independent of the given y
2
. It can be proved that the k – the
angularcoefficientofthelinerelatedtotheextrememeans–willbeequalto:
0
3
3
,
) ( 3
A
w w M w
R
y
k
R
R
⊗
= = ,where x M w
1
2
−
= . (7)
Hence,k
R
isindeedaninvariantandallmaximummeanreturnsforgivenvarianceslie
inthesamedirectioninmeanxskewnessspace.
Contrary to the previous, mean returns case, the optimal weights for the
skewness extremes are implicitly defined by a nonlinear system like (2). When
1
3
=
p
σ ,wehaveasolutionportfolio
s
α suchthat:
) (
1
3
1
2
4
S S s
M M
A
α α α ⊗ =
−
.(8)
Thehomothecyimpliesthat
3 3
3
s
s
s
s s
y α σ α α = = isasolutionto(2),ensuingan
optimalvariance
2
) (
3 2 2
s s s
y σ σ = .Acorresponding(excess)returnandadirection,both
independentofthevariancelevel,canbefoundas:
3
p
s s
y R R = ,
s s
R k / 1 = ,(9)
implyingthatalltheseoptimalportfolioslieinthesamedirection.
Combining both results gives a rectangular envelope that circumscribes, in the
first quadrant of the mean x skewness plane, the corresponding part of the constant
variancecurve.
2.2.Theminimumkurtosiscase.
Theinitialstepnowisminimisingkurtosisforagivenskewnessandexpectedreturn:
Min
α
[ ] 3
, , , ,
4 1 3
( ) [ ( ) ( (1 1 ) )] ( ( ))
p f
p
M E r M r M α α α α λ α α γ σ α α α ⊗ ⊗ + − + − + − ⊗ .(10)
Thefirstorderconditionsare:
4 3
4 ( ) 3 ( ) M x M α α α λ γ α α ⊗ ⊗ = + ⊗
[ ]
, ,
1
( ) ( 1 )
p f f
R E r r M r x α α = − = − = (11)
3
,
3
( )
p
M σ α α α = ⊗
Defining
[ ]
1
,
( 2) 4
( ) B x M I x α α
−
−
= ⊗ ⊗ ,
[ ]
1
,
0 4 3
( ) ( ) B x M I M α α α α
−
= ⊗ ⊗ ⊗ and
[ ]
1 , ,
2 3 4 3
( ) ( ) ( ) B M M I M α α α α α α
−
= ⊗ ⊗ ⊗ ⊗ ,
with the subscripts chosen according to the degree of homogeneity of the term with
respecttothevectorα,onecanfindthevaluesofλandγandarriveatthenonlinear
systemthatcharacterisesthesolutionto(10):
3 3
2 0 ( 2) 0
4 3 2 2
( 2) 2 0 ( 2) 2 0
( ) ( )
( ) ( )
p p
B R B B B R
M x M
B B B B B B
σ σ
α α α α α
−
− −
− −
⊗ ⊗ = + ⊗
− −
.(12)
Theoptimalkurtosiswillbegivenby:
3 3
4
2 2
2 0 ( 2)
2
( 2) 2 0
2 ( )
( )
p p
p
B R B R B
B B B
σ σ
σ
−
−
− +
=
−
.(13)
Noticing that B
(2)
and B
2
are positive, because the matrix in their middleis the
inverse of a positive definite matrix, it can be proved that both the numerator and the
denominatoroftheexpressionabovearepositive.
It is important to remark the similarities between the pairs offormulas (2)(12)
and(4)(13),astheyareattheheartofthesimilardevelopmentsthatfollow.Thefirstis
akeyproposition,closetoProposition1:
Proposition1*: Foragivenk,alltheminimumkurtosisportfoliosrelatedtoexpected
returns, skewnesses pairs (R,
3
p
σ ) such that
3 3
3
R k
p
= σ , or kR y =
3
, are given by
R α α = , where α defines
4
p
σ ,the(minimum)kurtosis oftheoptimalportfoliowhen
R=1and k y =
3
.Moreover,theminimumkurtosisforanypairofconstraintsinthek
linewillbe
4 4
4
p p
R σ σ = ,or R y y
p p
4 4
= .
The consequence of the above proposition is that exactly the same homothecy
appliesin3Dspacedefinedbythestandardisedkurtosisaxisandthetwooddmoments
axes. The results in Proposition 2 are then easily translated to the present context and
thefollowingisvalidaswell:
Proposition3*:ThedirectionintheRxy
3
halfplanethatgivesthehighestRforallthe
minimum kurtosis portfolios with the same standardised kurtosis y
4
is unique.
Moreover, in this direction, the skewness constraint to programme (10) is not binding.
As regards skewness, there is at least one direction giving the maximum skewness,
wheretheconstraintpropertyapplies.
Thesolutiontotheproblemofminimisingkurtosisforagivenexcessreturnis:
4
( 2)
( )
R R R
R
M x
B
α α α
−
⊗ ⊗ = ,(14)
which,whenR=1,becomes:
4
( 2)
1
( )
R R R
M x
B
α α α
−
⊗ ⊗ = .(15)
The systems of weights defined by R
R R
α α = are solutions to (12); thus one
only needs to find one portfolio
R
α to generate the whole set of minimum kurtosis
portfoliosforagivenR.Theskewnesscorrespondingto
R
α isgivenby:
3
3 0 0
( 2) ( 2)
R
B R B
R
B B
σ
− −
= = ;(16)
sothattheangularcoefficient
1/ 3
0
( 2)
R
B
k
B
−
 
=


\ .
(17)
defines a direction in the expected returns x skewness plane which is the “maximum
meanreturnsline”foragiven(minimum)kurtosis.
The “maximum mean returns line” divides the minimum isokurtosis curves in
two parts; since agents want the highest possible skewness, they will probably work
withtheupperhalfofthecurve.Incontrasttotheclassicalcaseofminimisingvariance
foragivenreturn,thereisnoclosedformfortheportfolioweights
R
α ,ascanbeseen
from (15). However, it is possible to show that this function is strictly convex in its
entiredomain,thereforeimplyingthatthesolutionisunique.
The highest/lowest skewness directions, as in the case of variance, will be the
onesassociatedtothesolutionoftheproblemoffindingthelowestkurtosissubjecttoa
givenskewness.Callingtheseportfolios
s
α ,theyareimplicitlydefinedbysystem
[ ]
3
4 3
2
( )
s
s s s s s
M M
B
σ
α α α α α ⊗ ⊗ = ⊗ ,(18)
theportfoliothatsolvestheproblemwhen 1
3
=
p
σ isnaturallydefinedby:
[ ]
4 3
2
1
( )
s s s s s
M M
B
α α α α α ⊗ ⊗ = ⊗ .(19)
Itcanalsobeeasilyverifiedthatthemeanreturnrelatedtothesolutionof(18)is
3
0
2
s
R B
B
σ
= , (20)
sothatthedirectionsaredefinedby
3 / 1
2
0
−


.

\

=
B
B
k
s
.(21)
Unfortunately, in this case, there can be more than one solution, and
consequentlymorethanonedirectionwithalocalmaximumskewnessforagivenlevel
of kurtosis. Notwithstanding, the projection of each isokurtosis curve will also be
enveloped,inthefirstquadrant,bythetwoaxesandtwotangentlinesparalleltothem.
3. Generalisingforhigherevenmoments.
We now consider the general case of minimising an even moment given the two first
oddmoments.Thelagrangianoftheproblemwillbe:
3
, ( 1) , , 2
3
( ) ( )
p
p
p
M R x M α α λ α γ σ α α
⊗ − ⊗
+ − + − ,(22)
givingthefirstorderconditions:
( 1) 2
3
3
p
p
pM x M α λ γ α
⊗ − ⊗
= +
,
R x α = (23)
3
, 2
3
p
M σ α α
⊗
= .
Noticing that
( 1) ( 2)
( )
p p
p p n
M M I α α α
⊗ − ⊗ −
= ⊗ , and that matrix
( 2)
( )
p
p n
M I α
⊗ −
⊗ is
symmetricandpositivedefinite,thefollowingsystemcanbeformedfrom(23)togive
thevaluesofthemultipliers:
( 2) 1 ( 2) 2
3
'( ) 3 '( )
p p
p n p n
pR x M I x x M I M λ α γ α α
⊗ − − ⊗ − ⊗
= ⊗ + ⊗
3
2 ( 2) 1 2 ( 2) 2
3 3 3
( ) '( ) 3 ( ) '( )
p p
p n p n
p
p M M I x M M I M σ λ α α γ α α α
⊗ ⊗ − − ⊗ ⊗ − ⊗
= ⊗ + ⊗ .(24)
Defining
1
, ( 2)
2
( )
p
p p n
B x M I x α
−
⊗ −
−
( = ⊗
¸ ¸
,
1
, ( 2) 2
4 3
( )
p
p p n
B x M I M α α
−
⊗ − ⊗
−
( = ⊗
¸ ¸
and
1
, 2 , ( 2) 2
6 3 3
( ) ( )
p
p p n
B M M I M α α α
−
⊗ ⊗ − ⊗
−
( = ⊗
¸ ¸
,
withthesubscriptscorrespondingtothegeneraliseddegreeofhomogeneitywithrespect
tothevectorofweights,thefinalsolutioncomesfromthesystem:
3 3
2 ( 1) 2
2 6 4 6 4 4 2 3
( ) ( ) ( )
p
p p p p p p p p
p p
B B B M B R B x B R B M α σ σ α
⊗ − ⊗
− − − − − − −
− = − + − ;(25)
theoptimalportfoliopthmomentbeing:
3 3
2 2
6 4 2
2
2 6 4
2 ( )
p
p p p
p p
p
p p p
B R B R B
B B B
σ σ
σ
− − −
− − −
− +
=
−
.(26)
Again,thesimilarities(2)(12)(25)and(4)(13)(26)shouldbestressed.
Thefollowingresultsummarisesallthepropertiesofthesolutionsset:
Theorem:Foragivenp=2,4,...,considerin(R,y
3
,y
p
)spaceofstandardisedmomentsa
isopthmomentcurveΓofsolutionsto(22)
THEN
i) the optimal portfolios set is contained in the cone {O}∗Γ , where O=(0,0,0) is
theoriginof(R,y
3
,y
p
)space;
ii) theprojectionofΓintheRxy
3
planeisacurve:a)symmetrictotheoriginandb)
inscribed in a rectangle whose sides are parallel to the axes; the vertical and
horizontal sides correspond, resp., to the highest (and lowest) R and y
3
values
whichproduceasolutioninΓ.
Proof(weoutlinethestepsoftheproof):Forprovingi)onefirstfollowsstepssimilar
tothoseinPropositions1and1*,showingthatoneachlinepassingthroughtheorigin
anda general point (R,y
3
), the solutions to (22) increase linearly either with R – if the
solution to (1, y
3
/R) is taken as the fundamental one – or with y
3
– if the solution to
(R/y
3
, 1) is the one fixed. As the origin O=(0,0,0) solves (22), this is sufficient to
demonstratethatanysolutionwillbeinthecone.Inthecaseofii),thesymmetryisseen
by the fact that reverting to the pair (R,y
3
) does not change either (25) or (26). As
regardsthetangents,areasoningsimilartotheonesintheprevioussectiondetermines
thepointsrelativetothehighestRandy
3,
bysymmetrythepointsofthelowestRandy
3
areobtainedandtherectanglecanbetraced.
This basic result isimportant infinding the efficient portfolios set for the three
moments at stake. It is easy to convince oneself that not all points in the cone will
characteriseanefficientportfolio,though,ofcourse,theefficientsetwillbecontained
inthecone(seeAthaydeandFlôres(2001)).Moreover,onecouldbetemptedtoderive
thefollowing
(false)Corollary:Ifproblem(22)hasasolutionTHENtheoptimalvalueisunique.
Indeed,bytheTheorem,if(22)hasasolutionthentheoptimalpthmomentsmustliein
the cone. They will be found in the intersection of a vertical line through the point
definedbythegivenoddmomentsintheRxy
3
planeandthecone.Simplepropertiesof
aconeinfinitedimensionalEuclidianspacesensurethatthisintersectionisunique.
Thisnicepropertywouldmeanthatknowledgeofthegeometricstructureofthe
optimalportfoliossethadallowedasimpleandelegantproofofuniqueness.However,
such an argument would be circular, as the curve Γ used to characterise the cone is
supposedly the curve formed already by the minimum pth moments, related to the
optimalsolutionsof(25).Itisworthremindingthatsystem(25),asitsspecialcases(2)
and (12), implicitly defines the optimal weights, and may as well have more than one
solution. These others either will be local, not global optima or it might even happen
that different optimal vectors α yield the same optimal pth moment in (26).
Propositions1to 3(and1*and3*) arevalidforanyofthesesolutions–thismeaning
that even different “solution cones” may exist; but the Theorem considers, by
hypothesis,the“optimalcone”,andsotheCorollaryissenseless.Unfortunately,atthe
present stage, we do not have a general, deeper knowledge of the structure of the
solutions set. Moreover, the hypothesis also requires the existence of a solution;
rigorous conditions for guaranteeing this, as regards system (25), are still an open
question.
An interesting special case of (22) is when only a mean return restriction is
imposed, the skewness constraint being disregarded. Without much difficulty one sees
thatthefirstorderconditionsbecome:
( 1) p
p
pM x α λ
⊗ −
=
,
R x α = .(27)
Sothattheoptimalweightsmustsolvethesystem
( 1)
2
( )
p
p p
B M Rx α
⊗ −
−
= ;(28)
and the corresponding pth moment bears the following relationship with the given
return:
1
2 2
( )
p
p
p
B
R
σ
−
−
= .(29)
In this case, the homothecy property implies that only one system needs to be solved,
namely,theoneobtainedbysettingR=1in(28).
4. Furtherpropertiesandextensions.
Inordertogiveafurtherinsightbothonthegeometricaspectsdiscussedaswellason
thedifficultiesinvolvedinthesolutionofsystem(28),weconsiderthespecialproblem
of minimising kurtosis given expected return, in the case of two assets and setting to
zeroallcokurtoseswhereanassetappearsonlyonce.Thisleavesuswiththreedistinct
nonzeroelementsinthekurtosistensor,andtheM
4
matrix–shown,inthegeneralcase,
intheAppendix–becomes:
1 12 12 12
12 12 12 2
0 0 0 0
0 0 0 0
σ σ σ σ
σ σ σ σ
(
(
¸ ¸
.
The simplified notation used for the subscripts, suppressing repetition of identical
indexes,stressestheidenticalvaluesandshouldcausenoconfusion.Noticethat,unless
theassetsdistributionsaresingular,allentriesarestrictlypositive.
Callingα=(α
1
,α
2
)’thevectorofweights,andnoticingthat:
i)
3 2
3 1 1 1 2 12
4
2 3
1 2 12 2 2
3
3
M
α σ αα σ
α
α α σ α σ
⊗
( +
=
(
+
¸ ¸
;
ii)matrix
1
2
4 2
( ) M I α
−
⊗
( ⊗
¸ ¸
willbeequalto:
2 2
1 1 12 2 2 1 2 12
2 2
1 2 12 1 1 2 12
2
2
α σ α σ α α σ
α α σ α σ α σ
−
( + −
∆
(
− +
¸ ¸
where
4 2 2 2 4
1 1 12 1 2 1 2 12 2 12 2
( 3 ) α σ σ α α σ σ σ α σ σ ∆ = + − + is the determinant of the direct
matrix;
oneisreadytobuildupsystem(28).Ofcourse,assaidintheprevioussection,onlyone
solution matters, namely that which considers R=1. We shall, however, impose the
additional assumptions that the marginal kurtoses are equal (i.e., σ
1
=σ
2
=σ) and that
excess returnsfor both assetsarealso equal (toa common valuex). With this, we can
finallywritesystem(30):
5 2 4 3 2 2 2 3 2 4 2
1 12 1 2 12 1 2 12 12 1 2 12 1 2 12
4 2 2 2 2 4
1 12 1 2 12 2 12
[ ( 2 ) 4 (3 7 ) 12 3 ]
( 3 )
x α σσ σ α α σσ α α σ σσ α α σ α α σ
α σ σ α α σ σ α σ σ
+ − + + − + =
= + − +
4 2 3 2 2 2 3 2 2 4 5
1 2 12 12 1 2 12 1 2 12 12 1 2 12 2 12
4 2 2 2 2 4
1 12 1 2 12 2 12
[3 (2 ) 12 (3 2 ) 4 ]
( 3 )
x α α σσ σ α α σ α α σ σσ σ α α σσ α σσ
α σ σ α α σ σ α σ σ
+ − + + + − + =
= + − +
Given the symmetry of the parameter values, the optimal weights will be
identical,beingeasytoseethattheircommonvalueis:
x 2
1
= α .(31)
These weights, however, can be related to either maxima or minima. For the
latter,theborderedHessiansufficientcondition
5
amounts,inthiscase,tocheckwhether
matrix
(
(
(
¸
(
¸
− −
− +
− +
0
) ( 12 24
24 ) ( 12
12
2
12
2
12
2
12
2
x x
x
x
σ σ α σ α
σ α σ σ α
(32)
hasanegativedeterminant.Replacingαbyitsvaluein(31),theconditionbecomes:
σ σ σ σ < < −
12 12
0 ) ( 6 or .(33)
The symmetric weights solution produces a minimum only if the nonnull co
kurtosisissmallerthanthecommonmarginalkurtosis.
Thisrathersimpleexamplemayserveasanillustrationofhowfarintuitioncan
help when considering higher moments, as well as of the impact of simplifications in
thehighermomentstensors.Thefinalsolutionisindependentofthemarginalkurtoses
andoftheevencokurtosis.Indeed,astheriskmeasureshaveacompletelysymmetric
structureasregards thetwo(risky) assets,theidenticalweightscanbefoundbydirect
solution of the excess return constraint. The higher the identical return, obviously the
lesswillbepurchasedofeachriskyasset–astheportfolioexcessreturnisfixedin1–
andmorewillbeputintherisklessasset
6
.
Given the similar roles played by kurtosis and variance, we could then expect
that the same would apply for the identical weights that result when equal marginal
variancesareusedinsteadofkurtosis.Infact,(31)isexactlythesolutionto (5)inthis
case,the(common)variancesandcovariancesplayingnoroleatall.Moreover,useof
theborderedHessianconditionshowsthataminimumexistsonlyif
σ σ <
12
.(34)
Though“identical”to(33),(34)willbealwaysvalidiftheassetscovarianceisnegative,
what cannot happen in the case of the even cokurtosis. Indeed, in our simplified
kurtosiscontext,thereisnoroomfordiversification.
Absentfrom(31)–initstwoversions/solutions,theriskmeasuresdohowever
playarole.Beyonddeterminingwhetheraminimumhasbeenachieved,theyexplicitly
5
See,forinstance,Theorem9.9,page202,inPanik(1976).
6
Asymptotically,alltheweightwillgototherisklessasset.
appear in the shadow price of the restrictions, given by the value of the Lagrange
multipliers.Theseareequalto
2
12
2x
σ σ
λ
+
= inthevariancecase,andto
2
12
2x
σ σ
λ
+
=
in that of kurtosis
7
. The formal identity of the two values hides different behaviours.
Again, in the case of the second moment, a negative covariance may substantially
decrease“thecost”oftheunitreturnrestriction.Ontheotherhand,both(nonnegative)
kurtosesaddup,penalisingmoreheavilyanincreaseinthefixedreturn.
Summing up, the example shows that the choice to minimise either kurtosis or
variance (in this very simple, symmetric case) has, in spite of producing exactly the
samesolutionweights,fairlydifferentimplications.Moreover,radicalsimplificationsin
the moments tensor may produce rather particular solutions. A small change in the
example, like allowing for different marginal kurtoses, would completely alter the
abovediscussion.Informallyspeaking,introducinghighermomentsinportfoliochoice
makesita“morenonlinear”problemand,consequently,muchmoresensibletosmall
changesintheinitialconditions.
5. Concludingremarks.
The availability of a general method to treat portfolio choice in a higher
moments context seems an unquestionable advantage. We outlined in the previous
sectionsonesuchmethod,thatallowsforacompact,analyticaltreatmentofallformulas
involvedintheoptimisationproblem.Thankstothis,powerfulgeometricinsightscould
begained.
Nevertheless, the task before anyone interested in the subject is still nearly
formidable. A basic existence result and more insights on the solutions set would be
welcome. Final characterisation of the efficient portfolios set requires more than the
techniques here discussed, duality methods being needed to completely identify the
efficient points. We solved this up to the fourth moment, Athayde and Flôres (2001,
2002), but a general method seems possible. Moving from static to dynamic
optimisation frameworks generates additional, rather difficult theoretical and
computationalproblems
8
.
7
Thereadershouldkeepinmindthatbothσandσ
12
havedifferentmeaningsinthetwoformulas.
8
WorkinthisdirectionhasbeeninitiatedwithBerçRustem(ImperialCollege,London).
Last,butnotleast,assection4glimpsedinto,thenumberofdifferentsituations
inthehighermomentscaseisextremelylarge,agreatprobabilityexistingofsenseless
or unattractive special formulations. These can only be sorted out through a
combination of more theoretical findings with several examples and applied
experiments. The notation developed, and its corresponding algebra, may help in
designingmanyoftheseexperiments.
Appendix:Thematrixnotationforthehighermomentsarrays.
Dealingwithhighermomentscaneasilybecomealgebraicallycumbersome.Givenan
dimensional random vector, the set of its pth order moments is, as a mathematical
object,atensor.Thesecondmomentstensoristhepopularnxncovariancematrix,while
the third moments one can be visualised as a nxnxn cube in threedimensional space.
However,thetensornotation,whichissousefulinphysics,geometryandsomeareasof
statistics (see, for instance, McCullagh (1987)), did not appear so convenient to deal
withtheportfoliochoiceproblem.Wethendevelopedaspecialnotation,whichallows
performing all the needed operations within the realm of matrix calculus. The
advantages of this are manifold. Beyond having a synthetic way to treat complicated
expressions, the mathematical tools required are standard linear algebra results and,
withthehelpofEuler’stheorem–asmostrealfunctionsinvolvedarehomogeneousin
the vector of portfolio weights , a differential calculus easily ensues. Moreover, the
differentformulaeand systems arrived atare written inacompactand straightforward
way,easilytranslatedintoformalprogramminglanguages.
Before presenting the notation, we remind that, throughout the paper we deal
with all the possible pmoments of a given ndimensional random vector of asset
returns.Undoubtedly,thedifficultyinmanipulatingallthesevaluessimultaneouslyhas
been a deterrent to tackle the problem in its full generality. Thinking of skewness and
kurtosis,forinstance,therespectivethreeandfourdimensional“cubes”,whereseveral
identicalvaluesarefound,haven
3
andn
4
elements.Ofcourse,inpractice,gatheringall
thesevaluesmayquicklybecomeaformidabletask.Indeed,asanexample,thenumber
of different kurtoses is in principle


.

\
 +
4
3 n
, what, in the case of five assets, gives
already70valuestobecomputed.Itisthenverylikelythat,ineachpracticalproblem,
either a significant number of comoments will be set a priori to zero or another
simplifyingassumptionwillbeused,andveryseldomonewillworkwiththefullsetof
cross moments. However, as said in the introduction, the great variety of possible
assumptionsisanextraargumentforageneraltreatmentoftheproblem.
Our notation transforms the full pth moments tensor, with n
p
elements, into a
matrix of order nxn
p1
obtained by slicing all bidimensional nxn
p2
layers defined by
fixing one asset and then taking all the moments in which it figures at least once and
pastingthem,inthesameorder,sideways.Rowi’ofthematrixlayercorrespondingto
have fixed the ith assetgives – in a preestablished order – all the moments in which
assets i and i’ appear at least once. Of course, assets must be ordered once and for all
and this order respected in the sequencing of the layers and in the numbering of the
rowsofeachlayer.Accordingly,aconformalorderingmustbechosen,andthoroughly
used,forthecombinations(withrepetitions)ofnelementsintogroupsofp2thatdefine
thecolumnsofeachmatrixlayer.
In the case of kurtosis, for instance, two indices/variables/coordinates must be
heldconstant.Callingσ
ijkl
ageneral(co)kurtosis,whenn=2,theresulting2x8matrix
willbe:
1111 1112 1121 1122 1211 1212 1221 1222
2111 2112 2121 2122 2211 2212 2221 2222
σ σ σ σ σ σ σ σ
σ σ σ σ σ σ σ σ
(
(
¸ ¸
where,asexpected,manyentriesareidentical.
Nowsupposethatavectorofweightsα∈R
n
isgiven,andM
1
,M
2,
M
3,...
andM
p
stand for the matrices containing the expected (excess) returns, (co)variances,
skewnesses ... and pmoments of a random vector of n assets. The mean return,
variance, skewness ... and pth moment of the portfolio with these weights will be,
respectively:α’M
1
,α’M
2
α,α’M
3
(α⊗α)...andα’M
p
(α⊗α⊗α...⊗α)≡α’M
p
α
⊗p1
where‘⊗’standsfortheKroneckerproduct.
The above expressions provide a clue on the mentioned advantages of the
notation. The fact that the tensors were transformed into matrices allows the use of
matrix algebra – and differential calculus  in all expressions and derivations, giving
waytocompactandelegantformulas.Itisimmediatetoseethat,asrealfunctionsofα,
thefourexpressionsabovearehomogenousfunctionsofthesamedegreeastheorderof
the corresponding moment. This means that Euler’s theorem can be easily used in the
neededderivations.
As an example, the derivative of the portfolio kurtosis with respect to the
weightswillbe:
) ( 4 )] ( [
4 4
α α α α α α α
α
⊗ ⊗ = ⊗ ⊗ ′
∂
∂
M M =4M
4
α
⊗3
.
References
Adcock, C. J. 2002. Asset pricing and portfolio selection based on the multivariate
skewstudent.Thisvolume.
Adcock,C.J.andK.Shutes.1999.Portfolioselectionbasedonthe multivariateskew
normal distribution, in A. Skulimowski, ed., Financial Modelling, Progress and
BusinessPublishers,Krakow.
Athayde,G.M.andR.G.FlôresJr.1997.ACAPMwithHigherMoments:Theoryand
Econometrics.EPGE/FGV,EnsaiosEconômicosn°317,RiodeJaneiro.
Athayde, G. M. andR. G. Flôres Jr. 2000. Introducing higher moments in the CAPM:
some basic ideas, in C. L. Dunis, ed., Advances in Quantitative Asset Management,
KluwerAcademicPublishers,Norwell,Mass.
Athayde,G.M.andR.G.FlôresJr.2001.Findingamaximumskewnessportfolio–a
generalsolutiontothreemomentsportfoliochoice.JournalofEconomicDynamicsand
Control,forthcoming.
Athayde,G.M.andR.G.FlôresJr.2002.Ajourneythroughanundiscoveredcountry:
efficientportfoliosetswithfourmoments.Submitted.
BaroneAdesi, G. 1985. Arbitrage equilibrium with skewed asset returns. Journal of
FinancialandQuantitativeAnalysis20,299313.
Harvey,C.R.andA.Siddique.1999.Autoregressiveconditionalskewness.Journalof
FinancialandQuantitativeAnalysis34(4),46587.
Harvey, C. R. and A. Siddique. 2000. Conditional skewness in asset pricing tests. The
JournalofFinance60,126395.
Hwang,S.andS.Satchell.1999.Modellingemergingmarketriskpremiausinghigher
moments.InternationalJournalofFinanceandEconomics4(4),27196.
Jurczenko,E.andB.Maillet.2001.The3CAPM:Theoreticalfoundationsandanasset
pricingmodelcomparisoninaunifiedframework,inC.L.Dunis,A.TimmermanandJ.
Moody,eds.,DevelopmentsinForecastCombinationandPortfolioChoice,JohnWiley
&Sons.
Jurczenko,E.andB.Maillet.2002.Thefourmomentcapitalassetpricingmodel:some
basicresults.Thisvolume.
McCullagh,P.1987.TensorMethodsinStatistics.ChapmanandHall,London.
Panik,M.J.1976.ClassicalOptimization:FoundationsandExtensions.NorthHolland
PublishingCompany,Amsterdam.
Pedersen,C.S.andS.E.Satchell.1998.Anextendedfamilyoffinancialriskmeasures.
GenevaPapersonRiskandInsuranceTheory23,89117.
Pedersen, C. S. and S. E. Satchell. 2000. Small sample analysis of performance
measures in the asymmetric response model. Journal of Financial and Quantitative
Analysis35,42550.
Scott,R.andP.A.Horvath.1980.Onthedirectionofpreferenceformomentsofhigher
orderthanthevariance.TheJournalofFinance35,91519.
1. Introduction. Portfolio optimisation taking into account more than the first two moments has been receiving renewed interest in the past years. Be it on the theoretical side – including its links with the CAPM extensions , or on what relates to econometric tests or updates based on higher conditional moments, works like Adcock (2002), Adcock and Shutes (1999), Athayde and Flôres (1997, 2001, 2002), Jurczenko and Maillet (2001, 2002), Pedersen and Satchell (1998), or Athayde and Flôres (2000), BaroneAdesi (1985), Harvey and Siddique (1999, 2000), Hwang and Satchell (1999) and Pedersen and Satchell (2000), far from exhausting the full list of contributions, pay good witness to the growing awareness of the importance of higher moments in both lines of research. Since Athayde and Flôres (1997), we have developed a systematic way to treat the key optimisation problems posed to anyone dealing with higher moments in portfolio theory. The approach uses a new notation to represent any moments tensor related to a multivariate random vector of asset returns, and can be used either in a utility maximising context or if optimal portfolios are defined by preference relations. The new notation seemed necessary in order to treat the problem in an absolutely general setting, which means both in the maximum order p of portfolio moments of interest and in the possible patterns of the skewness or higher order tensors. The latter is crucial as many works generalising the subject consider only the marginal higher moments of the returns vector, plainly disregarding any comoment of the same order. Though the full set of comoments can quickly become prohibitive – what, beyond other questions, may pose serious econometric estimation problems for the applications , and simplifying hypotheses on its pattern will usually be imposed in practice, it is important to have a way to study the general solution to the problem, irrespective of further assumptions that might be imposed. The utility function approach, given its more rigid theoretical constraints, and the debates involving any nonnormalityimplying (utility) function proposed, seems more suitable for theoretical developments related, for instance, to the CAPM. Preference ordering of portfolios, made rigorous by Scott and Horvath (1980), can lead to more interesting results in the strict portfolio optimisation context. In this paper, we discuss an interesting geometric structure that arises when optimising an even moment subject to odd moments constraints. As usual, agents “like” odd moments and “dislike” even ones.
the ultimate culprits of the high volatility and uncertainty of returns. If we minimise the fourth moment. while section three discusses how these results could still be generalised. through an example. Minimal variances and kurtoses subject to the first two odd moments. the latter should perhaps be more used in practice as an alternative objective function. like the VaR. and then kurtosis. given the frequency with which the fattailed effect in stock returns has been detected. The final section concludes by explaining how the results can be useful in a duality context and sets a few lines of further research.The structure studied – not the only relevant one in the higher moments context – bears important consequences and sheds light on the geometry of efficient portfolios sets in moments space.1. Homothetic properties of the minimum variance set. being always nonnegative. Moreover. proofs omitted here can be found in these papers. given the first and third desired portfolio moments. Most measures of risk focused on the worst scenarios. For fixed portfolio return and skewness. require proper software tools for solving the nonlinear systems and optimisation problems involved. . Section four draws. Even moments. a few more properties and analyses the sensitivity of certain solutions. The next section discusses the optimisation of variance. These. The material in this section draws on parts of Athayde and Flôres (2001) – where a complete solution to the three moments portfolio problem is found – and Athayde and Flôres (2002). are duly associated with spread. final testing of the gains brought out by using higher moments relies in extensive practical applications of the new results. This sort of behaviour will be further examined in the example in section 4. would probably be more sensitive to variations in the fourth moment rather than in variance. 2. This paper is organised as follows. we shall be directly attacking the heavy extremes of the density. We believe that its implications have not been fully exploited yet. in turn. and both variance and kurtosis are used as simple numerical summaries of the dispersion of a set of observations. Better knowledge of the surfaces (or manifolds) related to them may greatly improve the software design. An Appendix provides a brief explanation of the notation used. for the developments related to the kurtosis. 2.
gives the optimal variance(s): σ p2 = A4 R 2 − 2 A2σ p3 R + A0 (σ p3 )2 A0 A4 − ( A2 ) 2 .Minimising the variance. second and third moments tensors1. . A0 and A4. the matrices related to the first. (3) have subscripts corresponding to their degree of homogeneity as (real) functions of the vector α. R = E(rp) – rf the vector of mean excess returns. in particular. M 3 M 2 M 3 (α ⊗ α ) −1 −1 −1 . and the set (excess) portfolio return. M 2α = A4 R − A2σ p 3 A0 A4 − ( A2 ) 2 x+ A0σ p 3 − A2 R A0 A4 − ( A2 ) 2 M 3 (α ⊗ α ) . M 2α + λ1 [ E (rp ) − (α . M 2 M 3 (α ⊗ α ) A4 = (α ⊗ α ) . M 1 + (1 − α . for a given mean return and skewness. Premultiplying (2) by the very solutions α . M 2 x A2 = x . amounts to find the solution to the problem: Minα L = α .. M 3 (α ⊗ α )) . . (4) an expression where both the numerator and denominator are positive. M2 and M3 are. are positive because the inverse of the covariance matrix is positive definite. rf is the riskless rate of return. . the solution to (1) is found by solving the nequations nonlinear system. (1) where M1 . . [1])r f )] + λ 2 (σ p 3 − α . resp. α is the vector of n portfolio weights – where short sales are allowed. Calling x = M1 – [1] rf . the lambdas are Lagrange multipliers and the two remaining symbols are the αportfolio (given) mean return and skewness. [1] stands for a nx1 vector of 1’s. (2) where the scalars: A0 = x .
2 2 The above result implies that along the direction defined in the returns x skewness plan by y 3 = kR .. . and σ p 2 be the corresponding minimum variance. Slicing the surface by a sequence of planes parallel to the two oddmoments axes will generate a sequence of curves starting at the origin and whose expansion ratio will be equal to that of the respective (constant) variance values. the optimal portfolio surface will be reduced to the straight line y p 2 = y p 2 u k +1 2 . slicing it by planes parallel to the sd and (standardised) skewness axes will 1 2 See the Appendix for a further explanation on the notation used. with corresponding minimum variance σ p = σ p R2 .The following proposition is fundamental: Proposition 1: For a given k. from now on. As y p 2 differs with k. u ≥ 0 3. THEN for all optimal portfolios related to return and skewness pairs (R. use the angular coefficient k to name the corresponding line/direction in the first quadrant of the R x y3 plane. Taking now the three dimensional (3D) space where the standard deviation 2 σ p = y 2 axis is added. the angle that this line makes with the standard deviation axis varies also with k. the solution to (1) will be α = α R . let α define the minimum variance portfolio when R=1 and y 3 = 3 σ p 3 = k . σ p 3 ) such that σ p3 = k 3R3 . The Proposition has then a far reaching consequence: the optimal surface in the positive standard deviation (sd) half of 3D space bears a homothetic property from whatever standpoint one assumes. or y3 = kR . We shall. Of course. the optimal variance as a function of the excess return will be a parabola. in the halfplane formed by a specific direction k in 2 R x y3 space2 and the positive part of the standard deviation axis.
in this direction. of the curve are invariant with y2 (or R. as Proposition 1 is also true if the role of returns and skewness are reversed. cut the optimal surface with a plane orthogonal to the sd (or returns. at least part of the curve may go to infinity. the same will apply. The qualification if they exist is important as. For constant variance cuts. or R x y2) half plane related to the highest and lowest value. it may be shown that closed curves will be produced4. i. Indeed. the directions in the R x y3 (or y2 x y3. in each axis. or y3). inspection of formula (4) easily convinces that for the last combination. THEN if they exist. . for this case. the highest and lowest directions are particularly noteworthy. or standardised skewness) axis and project the intersection curve in the ‘returns x skewness (or sd x skewness. Moreover. or returns x sd) plane’. The proposition below is a direct consequence of this important fact: Proposition 2: For a given level of y2 (or R.e. Finally. as demonstrated by Proposition 3: The direction in the R x y3 half plane that gives the highest R for all the minimum variance portfolios with the same standard deviation y2 is unique and related to the celebrated (Markowitz’s) Capital Market Line (CML). The proof is rather technical to be included in this text. planes parallel to the sd and mean returns axes. specially in the case of cuts parallel to the sd axis.produce a sequence of homothetic curves whose expansion ratio will be that of the (excess) returns associated to each plan. the skewness constraint to programme (1) is not binding. As regards 3 4 The variable u stands for the coordinates along the axis defined by the “direction k”. or y3).
in the R x y3 half plane. M 3 ( w ⊗ w) A0 . Skewnesses .skewness. It can be proved that the k – the angular coefficient of the line related to the extreme means – will be equal to: kR = y R3 R = 3 w . which cuts the mean return axis in the maximum mean return portfolio value related to the set variance (that defines the cut). where w = M 2 x . σR = 2 R2 A0 . parallel to the skewness axis. this direction is the geometric locus of all the tangency points between each (projected) curve and a straight line. for a given mean return: α= R − M 21x A0 . being evident that they are independent of the given y2. . though there may be more than one “highest” (and “lowest”) direction. in each curve.and a k . In other words. relating the optimal variance to the given R.can also be associated to these optimal portfolios. This last statement is ensured by the wellknown duality result in Markowitz world. (5) that defines the famous Capital Market Line in mean x variance space. kR is indeed an invariant and all maximum mean returns for given variances lie in the same direction in mean x skewness space. the constraint property also applies. This means that the unique solution to the minimum variance portfolio. (6) also defines the (unique) direction that will pass through all the points. yielding the maximum mean return. −1 (7) Hence.
The minimum kurtosis case. k s = 1 / Rs . the corresponding part of the constant variance curve. both independent of the variance level. M 1 + (1 − α . When σ p = 1 . mean returns case. [1])rf )] + γ (σ p3 − α . Combining both results gives a rectangular envelope that circumscribes. x (11) σ p = α . (10) The first order conditions are: 4 M 4 (α ⊗ α ⊗ α ) = λ x + 3γ M 3 (α ⊗ α ) R = E (rp ) − rf = α .2. (8) The homothecy implies that α s = α s 3 σ s3 = α s y s3 is a solution to (2).Contrary to the previous. The initial step now is minimising kurtosis for a given skewness and expected return: Minα α . M 4 (α ⊗ α ⊗ α ) + λ[ E (rp ) − (α . M 3 (α ⊗ α ) 3 Defining . in the first quadrant of the mean x skewness plane. ( M 1 − rf [1]) = α . can be found as: Rs = Rs y p3 . A corresponding (excess) return and a direction. we have a solution portfolio α s such that: 3 αs = 1 −1 M 2 M 3 (α S ⊗ α S ) A4 . ensuing an optimal variance σ s 2 = σ s2 ( y s3 ) 2 . M 3 (α ⊗ α )) . 2. (9) implying that all these optimal portfolios lie in the same direction. the optimal weights for the skewness extremes are implicitly defined by a nonlinear system like (2).
because the matrix in their middle is the inverse of a positive definite matrix. it can be proved that both the numerator and the denominator of the expression above are positive. the (minimum) kurtosis of the optimal portfolio when R=1 and y 3 = k . It is important to remark the similarities between the pairs of formulas (2)(12) and (4)(13). the minimum kurtosis for any pair of constraints in the kline will be σ p4 = σ p4 R 4 .B( −2) = x . or y p4 =y p4 R . [ M 4 (α ⊗ α ⊗ I )] x −1 . close to Proposition 1: Proposition 1*: For a given k. as they are at the heart of the similar developments that follow. one can find the values of λ and γ and arrive at the nonlinear system that characterises the solution to (10): M 4 (α ⊗ α ⊗ α ) = B2 R − B0σ p3 B( −2) B2 − ( B0 ) 2 x+ B( −2)σ p3 − B0 R B( −2) B2 − ( B0 )2 M 3 (α ⊗ α ) . (12) The optimal kurtosis will be given by: B2 R 2 − 2 B0 Rσ p3 + B( −2) (σ p3 ) 2 B( −2) B2 − ( B0 )2 σp = 4 . The consequence of the above proposition is that exactly the same homothecy applies in 3D space defined by the standardised kurtosis axis and the two oddmoments . σ p 3 ) such that σ 4 p3 = k 3 R 3 . [ M 4 (α ⊗ α ⊗ I ) ] M 3 (α ⊗ α ) −1 B2 = (α ⊗ α ). M 3 [ M 4 (α ⊗ α ⊗ I )] M 3 (α ⊗ α ) . skewnesses pairs (R. (13) Noticing that B(2) and B2 are positive. and . where α defines σ p . The first is a key proposition. are given by α = α R . all the minimum kurtosis portfolios related to expected returns. −1 with the subscripts chosen according to the degree of homogeneity of the term with respect to the vector α . or y3 = kR . B0 = x. Moreover.
there is at least one direction giving the maximum skewness. The results in Proposition 2 are then easily translated to the present context and the following is valid as well: Proposition 3*: The direction in the R x y3 half plane that gives the highest R for all the minimum kurtosis portfolios with the same standardised kurtosis y4 is unique. where the constraint property applies. thus one only needs to find one portfolio α R to generate the whole set of minimum kurtosis portfolios for a given R. as can be seen from (15). since agents want the highest possible skewness. . (16) so that the angular coefficient B0 kR = B( −2) 1/ 3 (17) defines a direction in the expected returns x skewness plane which is the “maximum mean returns line” for a given (minimum) kurtosis. the skewness constraint to programme (10) is not binding. However. when R = 1. therefore implying that the solution is unique. The solution to the problem of minimising kurtosis for a given excess return is: M 4 (α R ⊗ α R ⊗ α R ) = R x B( −2) . in this direction. (15) The systems of weights defined by α R = α R R are solutions to (12).axes. there is no closed form for the portfolio weights α R . they will probably work with the upper half of the curve. In contrast to the classical case of minimising variance for a given return. The skewness corresponding to α R is given by: σR = 3 B0 R B = 0 R3 B( −2) B( −2) . becomes: M 4 (α R ⊗ α R ⊗ α R ) = 1 x B( −2) . Moreover. As regards skewness. (14) which. it is possible to show that this function is strictly convex in its entire domain. The “maximum mean returns line” divides the minimum isokurtosis curves in two parts.
3. in the first quadrant. Notwithstanding. they are implicitly defined by system M 4 (α s ⊗α s ⊗ α s ) = σs 3 B2 M 3 [α s ⊗ α s ] .x (23) . and consequently more than one direction with a local maximum skewness for a given level of kurtosis. M pα ⊗( p −1) + λ ( R − α . the projection of each isokurtosis curve will also be enveloped. Calling these portfolios α s . x) + γ (σ p − α . will be the ones associated to the solution of the problem of finding the lowest kurtosis subject to a given skewness. in this case. σ p = α . Generalising for higher even moments. (18) the portfolio that solves the problem when σ p3 = 1 is naturally defined by: M 4 (α s ⊗ α s ⊗ α s ) = 1 M 3 [α s ⊗ α s ] B2 . M 3α ⊗2 3 . there can be more than one solution.The highest/lowest skewness directions. The lagrangian of the problem will be: α . (20) so that the directions are defined by B ks = 0 B2 −1 / 3 . (21) Unfortunately. by the two axes and two tangent lines parallel to them. M 3α ⊗2 ) . as in the case of variance. (19) It can also be easily verified that the mean return related to the solution of (18) is R= σs 3 B2 B0 . We now consider the general case of minimising an even moment given the two first odd moments. 3 (22) giving the first order conditions: pM pα ⊗ ( p −1) = λ x + 3γ M 3α ⊗2 R = α.
consider in (R. with the subscripts corresponding to the generalised degree of homogeneity with respect to the vector of weights. (26) Again.yp) space. and −1 B6− p = (α ⊗2 ).y3.0) is the origin of (R.y3. (24) Defining B2− p = x. the final solution comes from the system: ( B2− p B6− p − B4− p ) M pα ⊗ ( p −1) = ( B6− p R − B4− pσ p3 ) x + ( B4− p R − B2− pσ p3 ) M 3α ⊗2 2 . −1 x M 3α ⊗2 −1 . M p (α ⊗ ( p − 2) ⊗ I n ) .yp) space of standardised moments a isopth moment curve Γ of solutions to (22) THEN i) the optimal portfolios set is contained in the cone {O}∗Γ . The following result summarises all the properties of the solutions set: Theorem: For a given p=2.. the similarities (2)(12)(25) and (4)(13)(26) should be stressed.4.0. where O=(0. and that matrix M p (α ⊗ ( p − 2) ⊗ I n ) is symmetric and positive definite.. the following system can be formed from (23) to give pR = λ x ' pα ⊗ ( p − 2) ⊗ I n ) −1 x + 3γ x ' pα ⊗ ( p − 2) ⊗ I n ) M 3α ⊗2 (M (M pσ p3 = λ ( M 3α ⊗2 ) ' pα ⊗( p − 2) ⊗ I n )−1 x + 3γ ( M 3α ⊗2 ) ' pα ⊗( p − 2) ⊗ I n ) M 3α ⊗2 (M (M .Noticing that M pα ⊗( p −1) = M p (α ⊗ ( p − 2) ⊗ I n )α the values of the multipliers: . .. . (25) the optimal portfolio pth moment being: σp = p B6− p R 2 − 2 B4− p Rσ p3 + B2− p (σ p3 ) 2 B2− p B6− p − B4− p 2 . M p (α ⊗ ( p − 2) ⊗ I n ) B4− p = x. M 3 M p (α ⊗ ( p − 2) ⊗ I n ) M 3α ⊗2 .
As the origin O=(0. Proof (we outline the steps of the proof): For proving i) one first follows steps similar to those in Propositions 1 and 1*. resp. This basic result is important in finding the efficient portfolios set for the three moments at stake. In the case of ii).. They will be found in the intersection of a vertical line through the point defined by the given odd moments in the Rxy3 plane and the cone.y3) does not change either (25) or (26). 1) is the one fixed. if (22) has a solution then the optimal pth moments must lie in the cone. the symmetry is seen by the fact that reverting to the pair (R. the vertical and horizontal sides correspond. of course. this is sufficient to demonstrate that any solution will be in the cone. As regards the tangents. Simple properties of a cone in finite dimensional Euclidian spaces ensure that this intersection is unique. such an argument would be circular. However. by the Theorem.0) solves (22). y3/R) is taken as the fundamental one – or with y3 – if the solution to (R/y3. one could be tempted to derive the following (false) Corollary: If problem (22) has a solution THEN the optimal value is unique. by symmetry the points of the lowest R and y3 are obtained and the rectangle can be traced. Indeed. It is worth reminding that system (25). to the highest (and lowest) R and y3 values which produce a solution in Γ. the efficient set will be contained in the cone (see Athayde and Flôres (2001)). a reasoning similar to the ones in the previous section determines the points relative to the highest R and y3. as its special cases (2) . though. showing that on each line passing through the origin and a general point (R.y3). This nice property would mean that knowledge of the geometric structure of the optimal portfolios set had allowed a simple and elegant proof of uniqueness.ii) the projection of Γ in the Rxy3 plane is a curve: a) symmetric to the origin and b) inscribed in a rectangle whose sides are parallel to the axes. the solutions to (22) increase linearly either with R – if the solution to (1. Moreover.0. related to the optimal solutions of (25). It is easy to convince oneself that not all points in the cone will characterise an efficient portfolio. as the curve Γ used to characterise the cone is supposedly the curve formed already by the minimum pth moments.
the hypothesis also requires the existence of a solution.and (12). are still an open question. Propositions 1 to 3 (and 1* and 3*) are valid for any of these solutions – this meaning that even different “solution cones” may exist. (28) and the corresponding pth moment bears the following relationship with the given return: σp R p 2 = ( B2− p )−1 . but the Theorem considers. (27) So that the optimal weights must solve the system ( B2− p ) M pα ⊗ ( p −1) = Rx . at the present stage. the skewness constraint being disregarded. Moreover. 4. Further properties and extensions. rigorous conditions for guaranteeing this. the “optimal cone”.x . in the case of two assets and setting to zero all cokurtoses where an asset appears only once. This leaves us with three distinct . Unfortunately. In order to give a further insight both on the geometric aspects discussed as well as on the difficulties involved in the solution of system (28). not global optima or it might even happen that different optimal vectors α yield the same optimal pth moment in (26). (29) In this case. and so the Corollary is senseless. Without much difficulty one sees that the first order conditions become: pM pα ⊗ ( p −1) = λ x R = α. implicitly defines the optimal weights. These others either will be local. the one obtained by setting R=1 in (28). and may as well have more than one solution. we consider the special problem of minimising kurtosis given expected return. An interesting special case of (22) is when only a mean return restriction is imposed. the homothecy property implies that only one system needs to be solved. namely. deeper knowledge of the structure of the solutions set. by hypothesis. we do not have a general. as regards system (25).
Calling α = (α1. in the general case. With this. Notice that. all entries are strictly positive. Of course. we can finally write system (30): 2 2 3 2 4 2 [α15 (σσ 12 + 2σ 2 ) − 4α14α 2σσ 12 + α13α 2 (3σ 12 + 7σσ 12 ) − 12α12α 2σ 12 + 3α1 α 2σ 12 ]x = 2 2 4 = α14σ σ 12 + α12α 2 (σ 2 − 3σ 12 ) + α 2 σ σ 12 2 2 2 3 2 4 5 [3α14α 2 (2σσ 12 + σ 12 ) − 12α13α 2σ 12 + α12α 2 (3σ 12 + σσ 12 + 2σ 2 ) − 4α1 α 2 σσ 12 + α 2σσ 12 ]x = 2 2 4 = α14σ σ 12 + α12α 2 (σ 2 − 3σ 12 ) + α 2 σ σ 12 Given the symmetry of the parameter values. and the M4 matrix – shown. impose the additional assumptions that the marginal kurtoses are equal (i. the optimal weights will be identical. unless the assets distributions are singular. stresses the identical values and should cause no confusion. however. namely that which considers R=1.e. being easy to see that their common value is: . suppressing repetition of identical indexes.nonzero elements in the kurtosis tensor. α2)’ the vector of weights.. 2 α12σ 12 + α 2σ 2 −2α1 α 2σ 12 2 −2α1 α 2σ 12 α12σ 1 + α 2 σ 12 2 2 4 ∆ = α14σ 1σ 12 + α12α 2 (σ 1σ 2 − 3σ 12 ) + α 2σ 12σ 2 is the determinant of the direct one is ready to build up system (28). as said in the previous section. σ1=σ2=σ) and that excess returns for both assets are also equal (to a common value x). will be equal to: M 4 (α ⊗2 ⊗ I 2 ) −1 ∆ −1 where matrix. We shall. and noticing that: i) M 4α ⊗3 = ii) matrix 2 α13σ 1 + 3α1α 2σ 12 3 3α12α 2σ 12 + α 2σ 2 . only one solution matters. in the Appendix – becomes: σ1 0 0 0 σ 12 0 0 σ 12 σ 12 0 0 0 σ 12 σ 12 σ 12 σ2 . The simplified notation used for the subscripts.
for instance. (33) The symmetric weights solution produces a minimum only if the nonnull cokurtosis is smaller than the common marginal kurtosis. they explicitly 5 6 See. all the weight will go to the riskless asset. Replacing α by its value in (31). In fact. however. Theorem 9. The final solution is independent of the marginal kurtoses and of the even cokurtosis. Absent from (31) – in its two versions/solutions . use of the bordered Hessian condition shows that a minimum exists only if σ 12 < σ . The higher the identical return. (31) is exactly the solution to (5) in this case. the (common) variances and covariances playing no role at all. Moreover. to check whether matrix 12α 2 (σ + σ 12 ) 24α σ 12 −x 2 24α 2σ 12 2 −x (32) 12α (σ + σ 12 ) − x −x 0 has a negative determinant. obviously the less will be purchased of each risky asset – as the portfolio excess return is fixed in 1 – and more will be put in the riskless asset6. Beyond determining whether a minimum has been achieved. in Panik (1976). as the risk measures have a completely symmetric structure as regards the two (risky) assets. Indeed. what cannot happen in the case of the even cokurtosis. (31) These weights. (34) Though “identical” to (33). Given the similar roles played by kurtosis and variance. page 202. as well as of the impact of simplifications in the highermoments tensors. This rather simple example may serve as an illustration of how far intuition can help when considering higher moments. the identical weights can be found by direct solution of the excess return constraint.9. the risk measures do however play a role. there is no room for diversification. . in this case. the bordered Hessian sufficient condition5 amounts.α= 1 2x . (34) will be always valid if the assets covariance is negative. Asymptotically. For the latter. can be related to either maxima or minima. Indeed. we could then expect that the same would apply for the identical weights that result when equal marginal variances are used instead of kurtosis. in our simplified kurtosis context. the condition becomes: 6(σ 12 − σ ) < 0 or σ 12 < σ .
On the other hand. a negative covariance may substantially decrease “the cost” of the unit return restriction. rather difficult theoretical and computational problems8. much more sensible to small changes in the initial conditions. We solved this up to the fourth moment. London). Informally speaking. fairly different implications. and to λ = σ + σ 12 2x 2 in that of kurtosis7. radical simplifications in the moments tensor may produce rather particular solutions. Nevertheless. . given by the value of the Lagrange multipliers. Again. but a general method seems possible. Work in this direction has been initiated with Berç Rustem (Imperial College. 5. both (nonnegative) kurtoses add up. Athayde and Flôres (2001. A small change in the example. powerful geometric insights could be gained. the task before anyone interested in the subject is still nearly formidable. in spite of producing exactly the same solution weights.appear in the shadow price of the restrictions. The availability of a general method to treat portfolio choice in a higher moments context seems an unquestionable advantage. analytical treatment of all formulas involved in the optimisation problem. Moving from static to dynamic optimisation frameworks generates additional. Summing up. Concluding remarks. consequently. Thanks to this. penalising more heavily an increase in the fixed return. We outlined in the previous sections one such method. introducing higher moments in portfolio choice makes it a “more nonlinear” problem and. symmetric case) has. that allows for a compact. Moreover. the example shows that the choice to minimise either kurtosis or variance (in this very simple. These are equal to λ= σ + σ 12 2x 2 in the variance case. in the case of the second moment. duality methods being needed to completely identify the efficient points. Final characterisation of the efficient portfolios set requires more than the techniques here discussed. like allowing for different marginal kurtoses. 2002). A basic existence result and more insights on the solutions set would be welcome. would completely alter the above discussion. The formal identity of the two values hides different behaviours. 7 8 The reader should keep in mind that both σ and σ12 have different meanings in the two formulas.
Before presenting the notation. in practice. These can only be sorted out through a combination of more theoretical findings with several examples and applied experiments. in each practical problem. The notation developed. the mathematical tools required are standard linear algebra results and. a great probability existing of senseless or unattractive special formulations. the respective three and fourdimensional “cubes”. which is so useful in physics. Undoubtedly. the number of different kurtoses is in principle n+3 4 . the different formulae and systems arrived at are written in a compact and straightforward way. for instance. and its corresponding algebra. McCullagh (1987)). However. as a mathematical object.Last. but not least. throughout the paper we deal with all the possible pmoments of a given ndimensional random vector of asset returns. may help in designing many of these experiments. Dealing with higher moments can easily become algebraically cumbersome. as section 4 glimpsed into. where several identical values are found. for instance. geometry and some areas of statistics (see. gives already 70 values to be computed. in the case of five assets. the set of its pth order moments is. . we remind that. have n3 and n4 elements. did not appear so convenient to deal with the portfolio choice problem. the number of different situations in the higher moments case is extremely large. a tensor. Appendix: The matrix notation for the higher moments arrays. gathering all these values may quickly become a formidable task. Beyond having a synthetic way to treat complicated expressions. the difficulty in manipulating all these values simultaneously has been a deterrent to tackle the problem in its full generality. The second moments tensor is the popular nxn covariance matrix. Thinking of skewness and kurtosis. the tensor notation. which allows performing all the needed operations within the realm of matrix calculus. The advantages of this are manifold. while the third moments one can be visualised as a nxnxn cube in threedimensional space. what. with the help of Euler’s theorem – as most real functions involved are homogeneous in the vector of portfolio weights . Moreover. We then developed a special notation. easily translated into formal programming languages. Indeed. as an example. It is then very likely that. Given a ndimensional random vector. Of course. a differential calculus easily ensues.
with np elements. M3 . skewnesses . Of course. However. Now suppose that a vector of weights α ∈ Rn is given. two indices/variables/coordinates must be held constant.. (co)variances. for the combinations (with repetitions) of n elements into groups of p2 that define the columns of each matrix layer. as expected. M2 . and pth moment of the portfolio with these weights will be. skewness . and M1. as said in the introduction. In the case of kurtosis. when n=2. The fact that the tensors were transformed into matrices allows the use of .. assets must be ordered once and for all and this order respected in the sequencing of the layers and in the numbering of the rows of each layer. and Mp stand for the matrices containing the expected (excess) returns. α’M2α ... Calling σijkl a general (co) kurtosis. and thoroughly used. α’M3 (α⊗α) . in the same order. . sideways... The mean return. variance. Row i’ of the matrix layer corresponding to have fixed the ith asset gives – in a preestablished order – all the moments in which assets i and i’ appear at least once.. Accordingly. respectively: α’M1 . the great variety of possible assumptions is an extra argument for a general treatment of the problem. a conformal ordering must be chosen. and pmoments of a random vector of n assets. for instance. the resulting 2x8 matrix will be: σ 1111 σ 1112 σ 1121 σ 1122 σ 1211 σ 1212 σ 1221 σ 1222 σ 2111 σ 2112 σ 2121 σ 2122 σ 2211 σ 2212 σ 2221 σ 2222 where. and very seldom one will work with the full set of cross moments.. many entries are identical. The above expressions provide a clue on the mentioned advantages of the notation.either a significant number of comoments will be set a priori to zero or another simplifying assumption will be used. into a matrix of order nxnp1 obtained by slicing all bidimensional nxnp2 layers defined by fixing one asset and then taking all the moments in which it figures at least once and pasting them. Our notation transforms the full pth moments tensor.. and α’Mp (α⊗α⊗α . ⊗α)≡α’Mpα⊗p1 where ‘⊗’ stands for the Kronecker product..
It is immediate to see that. Kluwer Academic Publishers. ed. Journal of Financial and Quantitative Analysis 20. G.matrix algebra – and differential calculus . 2001. Athayde. Rio de Janeiro. Introducing higher moments in the CAPM: some basic ideas. G. This means that Euler’s theorem can be easily used in the needed derivations. in A. forthcoming. A journey through an undiscovered country: efficient portfolio sets with four moments. 2002. G. 299313. Flôres Jr. M. BaroneAdesi.in all expressions and derivations. and R. giving way to compact and elegant formulas. M. Portfolio selection based on the multivariate skewnormal distribution. the derivative of the portfolio kurtosis with respect to the weights will be: ∂ [α ′M 4 (α ⊗ α ⊗ α )] = 4 M 4 (α ⊗ α ⊗ α ) = 4M4α⊗3 ∂α . 1997. C. Shutes. M.. Finding a maximum skewness portfolio – a general solution to threemoments portfolio choice.. Athayde. Submitted. G. and K. L. in C. Krakow. Financial Modelling. as real functions of α. Norwell. and R. 2000. and R. Athayde. G. Arbitrage equilibrium with skewed asset returns. Flôres Jr. Flôres Jr. the four expressions above are homogenous functions of the same degree as the order of the corresponding moment. Ensaios Econômicos n° 317. and R. G. 1999. Advances in Quantitative Asset Management. G. As an example. C. Dunis. References Adcock. J. J. . EPGE/FGV. Journal of Economic Dynamics and Control. Asset pricing and portfolio selection based on the multivariate skewstudent. Skulimowski. Athayde. A CAPM with Higher Moments: Theory and Econometrics. 1985. G. 2002. Progress and Business Publishers. Mass. Flôres Jr. Adcock. G. This volume. ed. M.
The Journal of Finance 35. and S. Dunis. M. C. 1976. and B. and B. Pedersen. Panik. 1999. The Journal of Finance 60. Moody. 126395. S. Jurczenko. Horvath. John Wiley & Sons. On the direction of preference for moments of higher order than the variance. Timmerman and J. . McCullagh. in C. 91519. 2000. Small sample analysis of performance measures in the asymmetric response model. Autoregressive conditional skewness. 46587. International Journal of Finance and Economics 4 (4). Conditional skewness in asset pricing tests. 27196. eds. Harvey. Journal of Financial and Quantitative Analysis 35. This volume. E. Hwang. 1980. Geneva Papers on Risk and Insurance Theory 23. 2001. Classical Optimization: Foundations and Extensions. The fourmoment capital asset pricing model: some basic results. and A. Siddique. R. The 3CAPM: Theoretical foundations and an assetpricing model comparison in a unified framework. Jurczenko. Chapman and Hall. 1998. E. Satchell. Scott. J. Satchell. 2002. Maillet.Harvey. 1999. Siddique. and P. 1987. Satchell. L. Modelling emerging market risk premia using higher moments. NorthHolland Publishing Company. 42550. C. London. and S. An extended family of financial risk measures. R. Pedersen. and S. Tensor Methods in Statistics. C. P. S. A. 2000. and A. E. Amsterdam. Developments in Forecast Combination and Portfolio Choice. E. Maillet. S. A.. Journal of Financial and Quantitative Analysis 34 (4). C. R. 89117.