American Finance Association

Portfolio Selection
Author(s): Harry Markowitz
Reviewed work(s):
Source: The Journal of Finance, Vol. 7, No. 1 (Mar., 1952), pp. 77-91
Published by: Blackwell Publishing for the American Finance Association
Stable URL: http://www.jstor.org/stable/2975974 .
Accessed: 06/06/2012 12:59

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Blackwell Publishing and American Finance Association are collaborating with JSTOR to digitize, preserve
and extend access to The Journal of Finance.

http://www.jstor.org
 PORTFOLIO SELECTION*
HARRYMARKOWITZ
The Rand Corporation

THE PROCESS OFSELECTING a portfoliomaybe dividedintotwostages.

The firststagestartswithobservation and experience and endswith
beliefsabout the futureperformances of available securities.The
secondstagestartswiththerelevantbeliefsaboutfuture performances
and endswiththechoiceofportfolio. Thispaperis concerned withthe
secondstage.We first considertherulethattheinvestor does(orshould)
maximizediscounted expected,or anticipated, returns. This ruleis re-
jectedbothas a hypothesis to explain,and as a maximumto guidein-
vestment behavior.We nextconsidertherulethattheinvestor does(or
should)considerexpectedreturna desirablethingand varianceofre-
turnan undesirable thing.This rulehas manysoundpoints,bothas a
maximfor,and hypothesis about,investment behavior.We illustrate
geometrically relationsbetweenbeliefsand choiceofportfolio accord-
ing to the "expected returns-variance of returns" rule.
One typeofruleconcerning choiceofportfolio is thattheinvestor
does (or should) maximizethe discounted(or capitalized)value of
futurereturns.1 Sincethefutureis notknownwithcertainty, it must
be "expected"or "anticipated"returns whichwe discount.Variations
of thistypeof rulecan be suggested.FollowingHicks,we could let
"anticipated"returns includean allowanceforrisk.2Or,we couldlet
the rate at which capitalizethe returnsfromparticularsecurities
we
varywith risk.
The hypothesis(or maxim) that the investordoes (or should)
nmaximize discounted returnmustbe rejected.If we ignoremarketim-
perfections the foregoing ruleneverimpliesthatthereis a diversified
portfolio whichis preferable to all non-diversified Diversi-
portfolios.
ficationis both observed and sensible;a ruleof behavior which does
of
notimplythesuperiority diversification must be rejectedboth as a
hypothesis and as a maxim.
* This paperis basedonworkdonebytheauthorwhileat theCowlesCommission for
Researchin Economicsand withthefinancial assistanceoftheSocialScienceResearch
as CowlesCommission
Council.It willbe reprinted Paper,NewSeries,No. 60.
1. See,forexample, J.B. Williams, TheTheory ofInvestmentValue(Cambridge,Mass.:
HarvardUniversity Press,1938),pp. 55-75.
2. J.R. Hicks,ValueandCapital(NewYork:Oxford Press,1939),p. 126.
University
Hicksappliestheruleto a firmrather thana portfolio.
77
 78 The Journalof Finance
The foregoing rulefailsto implydiversificationno matterhowthe
anticipatedreturns are formed;whether thesameordifferent discount
ratesare used fordifferent securities;no matterhow thesediscount
ratesare decideduponor how theyvaryovertime.3The hypothesis
impliesthatthe investorplaces all his fundsin the securitywiththe
greatestdiscounted value.If twoormoresecurities have,thesameval-
ue, thenany of theseor any combination of theseis as good as any
other.
We can see thisanalytically: supposethereareN securities; letrBbe
theanticipatedreturn(howeverdecidedupon) at timet perdollarin-
vestedin securityi; let d be therate at whichthe returnon theOh
security at timet is discountedbackto thepresent;let Xi be therela-
tiveamountinvestedinsecurity i. We exdudeshortsales,thusXi O 0
forall i. Then thediscountedanticipatedreturnoftheportfolio is
oo N
R=3 E di,t itX
t=1 i=1

= XiEdit rit)
i=l t=l

RX= E di ritis thediscounted therefore

oftheith security,
return
t=1

R = YXiRi whereRi is independent of Xi. Since Xi 3 0 forall i

and 2Xi = 1, R is a weightedaverage RXwiththeXi as non-nega-
of
tiveweights.To maximizeR, we let Xi = 1 fori withmaximumRi.
If severalRaa, a = 1, ..., K are maximumthenany allocationwith
K
=
Ea=1 Xaa 1

maximizesR. In no case is a diversified to all non-

preferred
portfolio
diversified
portfolios.
It willbe convenientat thispointto considera staticmodel.In-
stead of speakingof the timeseriesof returnsfromthe ih security
(ril, ri2,... , rit,. . .) we will speak of "the flow of returns" (ri) from
as a wholeis
the iP security.The flowof returnsfromthe portfolio
3. The resultsdependon theassumption thatthe anticipated and discount
returns
ratesareindependent oftheparticularinvestor's
portfolio.
amountof moneywouldbe placedin the
4. If shortsaleswereallowed,an infinite
security withhighestr.
 PortfolioSelection 79
R = 2X,Xr.As in thedynamiccase iftheinvestor wishedto maximize
"anticipated"returnfromtheportfolio he wouldplace all his fundsin
thatsecurity withmaximumanticipatedreturns.
Thereis a rulewhichimpliesboththattheinvestor shoulddiversify
and thathe shouldmaximizeexpectedreturn. The rulestatesthatthe
investor does (orshould)diversify hisfundsamongall thosesecurities
whichgivemaximumexpectedreturn.The law oflargenumberswill
insurethattheactualyieldoftheportfolio willbe almostthesameas
theexpectedyield.5Thisruleis a specialcase oftheexpectedreturns-
varianceofreturns rule(to be presented below).It assumesthatthere
whichgivesbothmaximum
is a portfolio expectedreturn andminimum
variance,and it commends thisportfolio to theinvestor.
This presumption, thatthelaw of largenumbersappliesto a port-
cannotbe accepted.The returnsfromsecurities
folioof securities, are
Diversification
too intercorrelated. cannoteliminateall variance.
The portfoliowithmaximumexpectedreturnis notnecessarily the
onewithminimum variance.Thereis a rateat whichtheinvestorcan
gainexpectedreturn by takingon variance,orreducevariancebygiv-
ingup expectedreturn.
We saw thattheexpectedreturnsor anticipatedreturnsruleis in-
adequate.Let us now considerthe expectedreturns-variance of re-
turns(E-V) rule.It willbe necessary to first present a few elementary
conceptsand resultsof mathematical statistics.We will thenshow
someimplications oftheE-V rule. After thiswe willdiscussitsplausi-
bility.
In ourpresentation we tryto avoidcomplicated mathematical state-
mentsandproofs.As a consequence a priceis paid in termsofrigorand
generality. The chieflimitations fromthissourceare (1) we do not
deriveour resultsanalyticallyforthe n-security case; instead,we
presentthemgeometrically forthe3 and4 security cases; (2) weassume
staticprobability beliefs.In a generalpresentation we mustrecognize
thattheprobability distribution ofyieldsofthevarioussecurities is a
function oftime.The writerintendsto present,in thefuture, thegen-
eral,mathematical treatment whichremovestheselimitations.

We will need the followingelementaryconceptsand resultsof

mathematical statistics:
Let Y be a randomvariable,i.e.,a variablewhosevalueis decidedby
chance.Suppose,forsimplicity that Y can take on a
of exposition,
finitenumberofvaluesyi, y2, .. , YN. Let theprobability that Y =
5. Williams,
op. cit.,pp. 68, 69.
80 The Journalof Finance

yi, be Pi; thatY = y2be p2 etc. The expectedvalue (ormean)of Y is

definedto be
E=Plyl+p2y2+. * *+PNYN

The varianceof Y is definedto be

V = p1(y1-E) 2+ P2(y2-E) 2+. . +PN (YN-E) 2

V is theaveragesquareddeviationof Y fromitsexpectedvalue. V is a
commonly used measureof dispersion.Othermeasuresof dispersion,
closelyrelatedto V are the standarddeviation,o-= -VV and theco-
ofvariation,alE.
efficient
Suppose we have a numberof randomvariables: R1, . . . , R,. If R is
a weightedsum (linearcombination) oftheRi
aLiRi+ a2R2+. . . + a.R.

thenR is also a randomvariable.(ForexampleR1,maybe thenumber

whichturnsup on one die; R2,thatof anotherdie,and R the sumof
thesenumbers. In thiscase n = 2, a, = a2 = 1).
It willbe important forus to knowhow the expectedvalue and
varianceof the weightedsum (R) are relatedto the probability dis-
tributionof the R1, . . . , R,n.We state these relationsbelow; we refer
thereaderto anystandardtextforproof.6
The expectedvalue of a weightedsum is theweightedsum of the
expected values. I.e., E(R) = alE(Ri) + a2E(R2) + . . . + anE(Rn)
The varianceofa weighted sumis notas simple.To expressit we must
define"covariance."The covarianceofR1and R2 is
o12=E I [R1-E (R1)I [R2-E (R2) I I
i.e., theexpectedvalue of [(thedeviationof R1 fromits mean) times
(thedeviationofR2 fromits mean)].In generalwe definethecovari-
ance betweenRi and Rj as
crijE I [R -E (Ri) ] [R -E (Rj) I }
O'i may be expressedin termsof the familiarcorrelation coefficient
betweenRi and Rj is equal to [(theircorrelation)
(pi2). The covariance
times(thestandarddeviationofRj) times(thestandarddeviationof
Rj)]T

6. E.g.,J.V. Uspensky, Introduction (NewYork:McGraw-

toMathematicalProbability
Hill,1937),chapter9, pp. 161-81.
 PortfolioSelection 8i
The varianceofa weightedsumis
N N
?
NI
V (R) = a2 V (Xi) +2 aia jaij
i=1 i=1 t>1

If we use thefactthatthevarianceofRi is oi then

V (R) = aiajxij

Let Auibe the expectedvalue

Let Ri be thereturnon theith security.
ofRi; oij,be thecovariancebetweenRi and Rj (thusoii is thevariance
ofRi). Let Xi be thepercentage oftheinvestor'sassetswhichare al-
The yield(R) on theportfolio
locatedto thei1k security. as a whoieis

R-=RiXi

The Ri (and consequentlyR) are consideredto be randomvariables.7

The Xi are notrandomvariables,but are fixedby theinvestor.Since
theXi are percentages we have 2Xi= 1. In our analysiswe willex-
cludenegativevaluesoftheXi (i.e.,shortsales); thereforeXi I 0 for
all i.
as a wholeis a weightedsumofran-
The return(R) on theportfolio
domvariables(wheretheinvestorcan choosetheweights).Fromour
discussionof suchweightedsumswe see that the expectedreturnE
fromtheportfolio as a wholeis
N

E = 1

and thevarianceis

V= oEijxix
i-1 j-1

7. I.e., we assumethattheinvestordoes (and should) act as ifhe had probabilitybeliefs

concerningthesevariables.In generalwe would expectthat the investorcould tell us, for
any two events(A and B), whetherhe personallyconsideredA morelikelythan B, B more
likelythanA, or both equally likely.If the investorwereconsistentin his opinionson such
matters,he would possess a systemof probabilitybeliefs.We cannot expect the investor
to be consistentin every detail. We can, however,expect his probabilitybeliefsto be
roughlyconsistenton importantmattersthat have been carefullyconsidered.We should
also expectthat he will base his actionsupon theseprobabilitybeliefs-even thoughthey
be in part subjective.
This paper does not considerthe difficultquestionof how investorsdo (or should) form
theirprobabilitybeliefs.
82 The Journalof Finance
For fixedprobabilitybeliefs(pi, aij) theinvestorhas a choiceofvari-
ous combinations of E and V dependingon his choice of portfolio
X1,... , XN. Supposethattheset of all obtainable(E, V) combina-
tionswereas in Figure1. The E- V rulestatesthattheinvestor would
(or should)wantto selectoneofthose portfolioswhich give riseto the
(E, V) combinationsindicatedas efficient in the i.e.,
figure; thosewith
minimumV forgivenE or moreand maximum E forgiven V or less.
Thereare techniquesby whichwe can compute the set of efficient
and efficient
portfolios (E, V) combinations associatedwithgivenAi

/ a~~~~~ffa;'nable
E,V combinations

\ / ~~~~~~~~~~~~~fficient
\ /"~ E,V combinations

E
FIG. 1

and aio.We willnotpresentthesetechniqueshere.We will,however,

illustrategeometricallythe natureof the efficient
surfacesforcases
in whichN (thenumberofavailablesecurities) is small.
surfacesmightpossiblybe of practical
The calculationof efficient
use. Perhapsthereare ways,by combining statisticaltechniquesand
the judgmentof experts,to formreasonableprobabilitybeliefs(ii,
oij). We could use thesebeliefsto computethe attainableefficient
combinations beinginformed
of (E, V). The investor, ofwhat(E, V)
combinations wereattainable,couldstatewhichhe desired.We could
thenfindtheportfolio whichgave thisdesiredcombination.
 PortfolioSelection 83
Two conditions-atleast-mustbe satisfied beforeit wouldbe prac-
surfaces
ticalto use efficient in themanner described above. First,the
investor must desireto act accordingto the E-V maxim. Second,we
mustbe able to arriveat reasonablepi and o,j. We willreturnto these
matterslater.
Let us considerthecase ofthreesecurities.In thethreesecurity case
ourmodelreducesto

1) E =EXil4

3 3
2) V =iXjaij
i=l j=l

3
3) EXi=1
i-1

4) XiO for i=1,2,3.

From (3) we get

3') X3=1-X1-X2

Ifwesubstitute(3') inequation(1) and (2) wegetE and V as functions

ofX1 and X2. For examplewe find
It) E = 3 +Xl (Al -A3) + X2 (A2 - A)

The exactformulas here(thatof V is givenbe-

are not too important
low).8We can simplywrite
a) E -E (X1, X2)
b) V = V(XI, X2)

C) XI,>0, X2)(), 1iI- X1- X2)>'0

By usingrelations(a), (b), (c), we can workwithtwo dimensional

geometry.
The attainableset of portfoliosconsistsof all portfolioswhich
satisfyconstraints(c) and (3') (or equivalently(3) and (4)). The at-
tainablecombinations ofX1,X2 are represented by the triangleabcin
Figure2. Anypointto theleftoftheX2 axis is notattainablebecause
it violatesthe conditionthatX1 ) 0. AnypointbelowtheX1 axis is
not attainablebecause it violatesthe conditionthat X2 0 0. Any
8. V =X (oi - 2I1a + a8) + X2(o22 - 2o2a + 8T3) + 2XtX2(ff,2 - I13 -28 + o8B)
+ 2X1 (ff18 - ag) + 2X2(9a2a - a,3) + Ta8
84 The JournalofFinance
pointabove the line (1 - X - X2 0) is not attainablebecauseit
violatesthe conditionthatX3 = 1 - X -X2 0.
We define an isomneancurveto be the set of all points(portfolios)
witha givenexpectedreturn.Similarly an isovariance to
lineis defined
witha givenvarianceofreturn.
be theset ofall points(portfolios)
An examination oftheformulae forE and V tellsus theshapesofthe
isomeanand isovariancecurves.Specifically theytellus thattypically9
theisomeancurvesare a systemofparallelstraight lines;theisovari-
ancecurvesarea systemofconcentric ellipses(see Fig. 2). Forexample,
if 2 ; A3 equation1' can be writtenin thefamiliar formX2 = a +
bX1;specifically(1)

X2=E-3 AlTA X1.

A2- A3 Al2- 3

Thus theslopeoftheisomeanlineassociatedwithE = Eo is - (1 -
A3)/(2 - A3) its interceptis (E0 - A3)/(Ab2 - 3). If we changeE we
changetheintercept but not the slopeof theisomeanline.This con-
firmsthe contention theisomeanlinesforma systemofparallel
that
lines.
Similarly,by a somewhatless simpleapplicationofanalyticgeome-
try,we can confirm the contention thatthe isovariancelinesforma
family of concentricellipses. The "center" of the systemis thepoint
which minimizes V. We will label this pointX. Its expectedreturnand
variancewe willlabel E and V. Varianceincreasesas you moveaway
fromX. Moreprecisely, if one isovariancecurve,C1,lies closerto X
thananother,C2,thenC1is associatedwitha smallervariancethanC2.
Withthe aid of the foregoing geometric apparatuslet us seek the
efficientsets.
X, the centerof the systemof isovarianceellipses,may falleither
insideoroutsidetheattainableset.Figure4 illustrates a case in which
X fallsinsidetheattainableset.In thiscase: Xis efficient. Forno other
has as
portfolio a V low X; as therefore no portfolio have either
can
smallerV (withthesameor greaterE) or greaterE withthesameor
smallerV. No point (portfolio) withexpectedreturnE less than E
For we have E > E and V < V.
is efficient.
Considerall pointswitha givenexpectedreturn E; i.e.,all pointson
theisomeanlineassociatedwithE. The point theisomeanlineat
of
whichV takeson itsleastvalue thepointat whichtheisomeanline
is
9. The isomean "Ccurves"are as describedabove except when I,u = ,U2= IA3. In the
lattercase all portfolioshave the same expectedreturnand the investorchooses the one
with minimumvariance.
As to the assumptionsimplicitin our descriptionof the isovariancecurvessee footnote
12.
 PortfolioSelection 85
is tangentto an isovariancecurve.We call thispointX(E). If we let
E vary,X(E) tracesout a curve.
Algebraicconsiderations(whichweomithere)showus thatthiscurve
line.We willcallitthecriticalline1.The criticallinepasses
is a straight
through X forthispointminimizes V forall pointswithE(Xi, X2) = E.
As we go alongI in eitherdirectionfromX, V increases.The segment
of the criticallinefromX to thepointwherethe criticallinecrosses

X2 \ Direction of
increasingE*

\\ \ \ isomean lines--

~~~isovariance ~~~~ curves

\ \m \ \ portfolios
efficient

\\ attainableset
c b

ft~~~ _\\

c \b
\ X
XI
\ \\ \
of increasingE
*direction
depends on It,i,u. #3
FIG. 2

theboundary oftheattainablesetis partoftheefficient set.The restof

the set
efficient is (in the case the
illustrated) segment of the ab line
fromd to b. b is thepointofmaximum attainableE. In Figure3, X lies
outsidethe admissiblearea but the criticalline cuts the admissible
area. The efficient line beginsat the attainablepointwithminimum
variance(in this case on theab line).It movestowardb untilit inter-
sectsthe critical line,moves alongthecriticallineuntilit intersects a
boundary and finally moves along theboundary to b. The reader may
a
E
~~~~~~~~~~~~~~~increasing

-.0 ~ ~ ~ -

I--,

FIG. 3

Xw|efficient portfolios

FIG. 4
 PortfolioSelection 87
wish to constructand examinethe following othercases: (1) X lies
outsidetheattainableset and thecriticallinedoesnotcut theattain-
able set. In thiscase thereis a security whichdoes notenterintoany
efficientportfolio. (2) Two securitieshave thesame isi.In thiscase the
isomeanlinesare parallelto a boundaryline.It mayhappenthatthe
efficientportfolio withl maximum E is a diversified (3) A case
portfolio.
wherein onlyoneportfolio is efficient.
The efficientsetin the4 security case is,as in the3 security and also
theN security case,a seriesofconnected linesegments. At one endof
theefficient set is thepointofminimum variance;at theotherend is
a pointofmaximumexpectedreturn'0 (see Fig. 4).
Now thatwe have seen thenatureof theset ofefficient portfolios,
it is notdifficultto see thenatureofthesetofefficient (E, V) combina-
tions.In thethreesecurity caseE = ao + aiX, + a2X2is a plane; V =
bo+ b,Xj+ b2X2+ bl2XlX2+ b,,X + b,2X2 is a paraboloid." As
showninFigure5, thesectionoftheE-planeovertheefficient portfolio
set is a seriesof connected line segmen'ts. The sectionof theV-parab-
oloid overthe efficient portfolioset is a seriesof connectedparabola
segments.If we plottedV againstE forefficient portfolioswe would
againgeta seriesofconnected parabolasegments(see Fig. 6). This re-
sultobtainsforanynumberofsecurities.
Variousreasonsrecommend theuse oftheexpectedreturn-variance
ofreturnrule,bothas a hypothesis to explainwell-established
invest-
mentbehaviorand as a maximto guideone's own action.The rule
servesbetter,we willsee,as an explanationof,and guideto, "invest-
ment"as distinguished from"speculative"behavior.
4
10. Justas weusedtheequationE = 1 to reducethedimensionality
in thethree
i= 1
security case,we can use it to represent thefoursecurity case in 3 dimensional space.
Eliminating X4we getE = E(X,, X2,X8), V = V(X,, X2, X8). The attainablesetis rep-
resented, inthree-space, bythetetrahedron withvertices (0,0,0), (0,0, 1), (0, 1,0), (1,0,0),'
representing portfolios with,respectively,X4 = 1, X3 = 1, X2 = 1, XI = 1.
Let s12sbe the subspaceconsisting ofall pointswithX4 = 0. Similarly we can define
SaX ... , aa to be thesubspace consisting ofall pointswithXi = 0, i $ a,, . .. , aa. For
eachsubspaceSal,.... , aa we can define a criticalline la, .... aa. Thislineis thelocusof
pointsP whereP minimizes V forall pointsinsal,... Xaa withthesameE as P. If a point
is in Sal, . . . , aa andis efficient
it mustbe onla,, . . ., aa. The efficient setmaybe traced
out by starting at thepointof minimum availablevariance,movingcontinuously along
variousla, .. . , aa according to definite
rules,endingina pointwhichgivesmaximum E.
As in thetwodimensional casethepointwithminimum availablevariancemaybe in the
interior oftheavailablesetorononeofitsboundaries. Typically weproceedalonga given
criticallineuntileitherthislineintersects oneofa largersubspaceor meetsa boundary
(and simultaneously thecriticallineof a lowerdimensional subspace).In eitherof these
casestheefficient lineturnsandcontinues alongthenewline.The efficient lineterminates
whena pointwithmaximum E is reached.
11. See footnote 8.
 E

/ I bI _X
set of
efficient
a portfolios

X2

FIG. S

efficient
E,V combinations

E
FIG. 6
 PortfolioSelection 89
Earlierwe rejectedtheexpectedreturns ruleon thegroundsthatit
neverimpliedthe superiority of diversification. The expectedreturn-
on
varianceofreturnrule, theotherhand,impliesdiversification fora
of
widerange pi, o-j. This does not mean that the E-V rule never im-
pliesthesuperiority ofan undiversified portfolio. It is conceivablethat
one security mighthave an extremely higheryieldand lowervariance
thanall othersecurities;so muchso thatone particularundiversified
portfolio wouldgive maximumE and minimumV. But fora large,
presumably representative rangeofAi, rijtheE- V ruleleadstoefficient
portfolios almostall ofwhichare diversified.
Not onlydoes the E-V hypothesis implydiversification, it implies
the"rightkind"ofdiversification forthe"rightreason."The adequacy
ofdiversification is not thoughtby investorsto dependsolelyon the
numberofdifferent securities held.A portfolio withsixtydifferent rail-
waysecurities, forexample,wouldnotbe as welldiversified as thesame
size portfolio withsomerailroad,somepublicutility,mining,various
sort of manufacturing, etc. The reasonis that it is generallymore
likelyforfirms withinthesameindustry to do poorlyat thesametime
for in
than firms dissimilar industries.
Similarly in trying to makevariancesmallit is notenouglh to invest
in manysecurities. It is necessaryto avoidinvesting in securities with
highcovariancesamongthemselves. We shoulddiversify acrossindus-
triesbecausefirmsin different industries, especiallyindustrieswith
different economiccharacteristics, lowercovariancesthanfirms
have
withinan industry.
The concepts"yield" and "risk" appear frequently in financial
if
writings.Usually the term"yield" were replacedby "expected
yield"or "expectedreturn,"and "risk"by "varianceofreturn," little
changeofapparentmeaningwouldresult.
Varianceis a well-known measureofdispersion abouttheexpected.
If insteadofvariancetheinvestorwas concerned withstandarderror,
a' = VV, or withthe coefficient of dispersion, oyE, his choicewould
stilllie in theset ofefficient portfolios.
Supposean investor betweentwoportfolios
d'iversifies (i.e.,ifheputs
someofhismoneyin oneportfolio, therestofhismoneyin theother.
An exampleofdiversifying amongportfolios is thebuyingoftheshares
of twodifferent investment companies).If thetwooriginalportfolios
have equalvariancethentypicallyl2 thevarianceoftheresulting (com-
pound)portfolio willbe less thanthevarianceofeitheroriginalport-
12. In no casewillvariancebe increased. The onlycase in whichvariancewillnotbe
decreasedis ifthereturn frombothportfolios areperfectly To drawtheiso-
correlated.
variancecurvesas ellipsesitis bothnecessary toassumethatno twodistinct
andsufficient
portfolios
haveperfectly correlated returns.
go The Journalof Finance
folio.Thisis illustrated
by Figure7. To interpretFigure7 wenotethat
a portfolio (P) whichis builtoutoftwoportfolios P' = (X, X2) and
P (X1, X2) is of the formP = XP + (1 - X)P = (XX +
(- X)X7, XXI+ (1 - )X2). P is on the straightline connecting
P' andP".
The E- V principle
is moreplausibleas a ruleforinvestmentbehavior
as distinguished fromspeculativebehavior.The thirdmoment"3 M3 of

X2
a

X ;~~~~~~~~~sovarionee

c b
XI
FIG. 7

theprobabilitydistributionofreturnsfromtheportfoliomay be con-
nectedwitha propensity to gamble.For exampleiftheinvestormaxi-
mizesutility(U) whichdependson E and V(U = U(E, V), d U/1E >
0, d Ul/E < 0) he willneveracceptan actuariallyfair'4bet. But if
13. If R is a randomvariablethattakeson a finite
number ofvaluesrl,. . . , r. with
n
Pi,..., pnrespectively,
probabilities andexpectedvalueE, thenMs = pi(ri-E)3
t=1
14. Oneinwhichtheamount
gainedbywinning
thebettimestheprobability
ofwinning
 PortfolioSelection 91

U = U(E, V, M3) and if 69U/0M3$ 0 thenthereare somefairbets

whichwouldbe accepted.
Perhaps-fora greatvarietyof investinginstitutions whichcon-
sidery'ieldto be a good thing;risk,a bad thing;gambling,to be
avoided-E, V efficiency is reasonableas a working hypothesis and a
working maxim.
Two uses oftheE-V principlesuggestthemselves. We mightuse it
in theoreticalanalysesor we mightuse it in the actual selectionof
portfolios.
In theoretical analyseswe mightinquire,forexample,about the
variouseffects ofa changein the beliefsgenerally held about a firm,
or a generalchangeinpreference as to expectedreturnversusvariance
ofreturn, or a changein thesupplyofa security.In ouranalysesthe
Xi mightrepresent individualsecuritiesortheymightrepresent aggre-
gatessuchas, say,bonds,stocksand realestate."
To use theE-V rulein theselectionofsecurities we musthavepro-
ceduresforfinding reasonable,uiand aiq. Theseprocedures,I believe,
shouldcombinestatisticaltechniquesand the judgmentof practical
men.My feeling is thatthestatisticalcomputations shouldbe used to
arriveat a tentativeset of A,iand -ij.Judgment shouldthenbe used
in increasingor decreasing someoftheseAi and aoj on thebasis offac-
torsor nuancesnot takenintoaccountby the formalcomputations.
Usingthisrevisedset ofAi and aii, theset ofefficient E, V combina-
tionscouldbe computed, theinvestorcouldselectthecombination he
preferred,and theportfolio whichgave riseto thisE, V combination
couldbe found.
One suggestionas to tentativeji, aij is to use the observed Ai, ai
forsomeperiodofthepast. I believethatbettermethods,whichtake
intoaccountmoreinformation, can be found.I believethatwhatis
neededis essentially reformulation
a "probabilistic" of securityanaly-
sis. I willnotpursuethissubjecthere,forthisis "anotherstory."It is
a storyof whichI have read onlythe firstpage of the firstchapter.
In thispaperwe have considered thesecondstagein theprocessof
This
selectinga portfolio. stagestarts withtherelevantbeliefsabout
the securitiesinvolved and endswith the selectionof a portfolio.
We
have not consideredthe first
stage: the formation of the relevantbe-
liefson thebasis ofobservation.
relations
15. Caremustbeusedinusingandinterpreting Wecannot
amongaggregates.
ofaggregation.
andpitfalls
dealherewiththeproblems