STEPHEN

M. GOLDFELD
Princeton University

The

Demand

for

Money
Revisited

of virtuallyall theories
that explainthe evolutionof aggregateeconomicactivity.More particularly, an accurateunderstandingand portrayalof this marketis essential
both to the analysisof past monetarypoliciesand to the formulationof
appropriatecontemporarypolicy.This paperfocuseson one aspectof the
money market,the demandside, and providesan extensivereviewof the
currentstate of the art concerningthe demandfor money.The emphasis
will be unabashedlyempirical,withconcentrationon the shortterm,taken
hereto be quarterly,sincethis horizonappearsto be the most relevantto
policypurposes.'
Therehas beena substantialamountof pastresearchon the demandfor
moneyand severalsurveypiecesas well.2Nevertheless,a numberof good
reasonsarguefor embarkingon anotherbroadempiricaleffort.In the first
THE MONEY MARKET IS A CRITICAL COMPONENT

1. The recentinterestwithin the FederalReserveSystemin monthlyand even weekly

models suggeststhat an even shorter-runfocus might be appropriate.
2. See, for example,David E. W. Laidler,TheDemandforMoney:Theoriesand Evidence(InternationalTextbook, 1969),and John T. Boorman,"The Evidenceon the Demand for Money: TheoreticalFormulationsand EmpiricalResults," in John T. Boorman and Thomas M. Havrilesky,Money Supply,Money Demand,and Macroeconomic
Models(Allyn and Bacon, 1972).
577

578

Brookings Papers on Economic Activity, 3:1973

instance,until recentlyresearchwith quarterlydata had not been that

extensive.Consequently,most of the receivedwisdomon the subjectstems
from empiricalwork with long-termannual data3 whose relevancefor
short-termpurposesis questionable.4
A secondreasonfor undertakinga broadempiricaleffortis thatmuchof
the existingevidencestems from the work of researcherswho have each
method,and estimatingtechuseda differentsampleperiod,measurement
nique.Thereis muchto be saidfor attackingthe substantialrangeof issues
that I wish to examinein a homogeneousand consistentmanner.This
procedureseemsall the more desirablesince it will permitme to use the
latest data uniformly,which seems importantin view of the variedbehaviorof moneyand interestratesin recentyears.5
A final motivationfor this paperis that recentevents have raisedthe
question,in both the popularandthe professionalpress,as to whetherthe
conventionalmoneydemandformulationis adequateto explainthe monetary experienceof the seventies.For example,from earlyto mid-1971the
moneystockroserapidlybut so did short-terminterestrates.Overroughly
the nexthalf-yearmoneygrewat a meager1 percentratebut interestrates
fell belowtheirearly1971lows. Both duringthis periodand subsequently,
observersquestionedwhetherthe economy had experiencedshort-run
shiftsin the demandfor money. More recently,the firsthalf of 1973saw
sharplyrisinginterestrates.Butwhilethe moneystockroseonlymarginally
in the firstquarter,it spurtedaheadat the annualrate of 11 percentin the
secondquarter.Onceagainthe presshas referredto the puzzlingbehavior
of the demandfor money.The basicissueis whetherthe demandfunction
for moneycan be assumedby the policy makerto be essentiallystablein
the shortrun. This issue, whichhas not been examinedpreviouslyin any
greatdetail,will receiveparticularemphasisin this paper.
3. This is certainlytrue of the researchreviewed,for example, in Laidler, Demand
for Money.
4. In fact, much of the short-termanalysisseemsto contradictmany aspectsof the received wisdom. For example,the evidencefrom the annual data tends to favor M2 over
Ml, long-termover short-terminterest rates, and wealth over currentincome. Practitioners workingwith quarterlydata tend to the opposite.
5. Whateverproblemsit may have caused money holdersand policy makers,the behaviorof interestratesin the last four years-historic peaksat the end of 1969,followed
by pronouncedcyclicalbehaviorand endingwith currentnear-recordlevels-is an econometrician'sdelight.

StephenM. Goldfeld

579

Outline

Theplanof the paperis as follows.Thenextsectionbrieflyspellsout the

conventionalstory on the originsand generalnatureof the demandfor
money and then reportsestimatesof one simpleand commonversionof
the moneydemandfunction.Theestimatesarethen analyzedwithprimary
focus on the followingtwo questions:
1. Is thereany evidenceof economiesof scalein aggregatemoneyholdings? Is there any indication,as previouslyhas been suggested,that the
incomeelasticityis difficultto pin down from quarterlydata?
2. Has the demandfunctionfor moneyremainedstableoverthe postwar
period?Putanotherway,is thereanyevidenceof eithersystematiclong-run
shifts or markedshort-runinstabilitiesthat make historicallyestimated
relationshipsunsuitablefor forecastingpurposes?
The resultsof that sectionwill serveas a roughstandardfor considering
other importantissues on the properspecificationof the money demand
functionthat are takenup in the thirdsection:
3. Whatdegreeof aggregationis appropriatewith respectto currency,
demanddeposits,and time deposits?
4. Whatsorts of lags appearto be presentin the adjustmentof money
holdingsand what rationalecan be offeredto explaintheselags?
5. Is thereany evidencethat expectedratesof inflationmeasuredeither
directlyor indirectlyinfluencethe demandfor money?
6. Shouldincome, or wealth,or perhapsboth, be used in the demand
function?
In the fourth section a numberof more technicalissues are explored:
7. Whichinterestratesworkbest in explainingthe demandfor money?
8. Are estimateddemand-for-money
functionssensitiveto the time unit
usedto constructthe aggregatedata?
9. How importantare the problemsof serial correlationand simultaneousequationsbias in the demandfor money?
10. Is the demandfor money homogeneouswith respectto prices or
population?
in somewhat
The fifth sectionexaminesthe problemsof disaggregation
more detail, using the flow of funds data on holdingsby type of holder
(businessand consumersand the rest). The basic question is whether
separateanalysisof more homogeneousgroupsof moneyholderscan im-

580

Brookings Papers on Economic Activity, 3:1973

prove understandingof the money demand process and the ability to

forecastthe demandfor money.The paperconcludeswith a summaryof
the mainresultsand an attemptto drawsome lessonsfromthem.
As the outlinesuggests,I shallcovera fairlybroadrangeof issueson the
specificationand propertiesof the demand-for-moneyfunction. While
thesequestionsareclearlyinterrelated,simultaneousconsiderationof all of
them would be a strategicand expositionalmonstrosity.Consequently,
exceptwhereit seemsparticularlywarranted,I shalltry to avoida flood of
permutationsand alternativespecifications.Evenso, some mayregardthe
outputas a "juniorencyclopedia"if not the full-fledgedthing.
SomeUnderpinnings
The conventionaltextbookformulationof the demandfor moneytypicallyrelatesthe demandfor realmoneybalances-m = M/P, assumedto
be noninterestbearing6-to "the" interestrate, r, and some measureof
economic activity such as real GNP-y = Y/P, where M = money holdings, P = the price level, and Y = gross national product. Thus

(1)
m-f(r, y).
A varietyof storiescan explainthe originsof equation(1). Perhapsthe
most satisfyingis the transactionsview, in whichthe demandfor money
evolvesfroma lack of synchronization
betweenreceiptsandpaymentsand
the existenceof a transactionscost in exchangingmoneyfor interest-bearing assets(usuallytakento be shortterm).
Oneexampleof this approachis the well-knownBaumol-Tobinformulation which readilyleads to an equationof the form of (1). Its simplest
versionis the so-calledsquareroot law of moneyholdings,7
(1')

m = ky-r?,

6. Although interest paymentson demand deposits have been prohibited,the existence of servicechargesmay producean implicityield on demanddeposits.Some writers
have used service chargesas a measure of negative interest payment but this practice
suffersfrom ratherseriousconceptualproblems.Recently, Barroand Santomerohave
constructedan explicitmarginalreturnon depositsbasedon remissionof servicecharges.
Unfortunately,the series is annual and stops in 1968. It does, however,vary substantially in the late 1960s,suggestingthat this may be an importantomittedvariablein demand-for-moneyequations.See Robert J. Barroand Anthony M. Santomero,"Household Money Holdings and the Demand Deposit Rate," Journalof Money, Creditand
Banking,Vol. 4 (May 1972), pp. 397-413.
7. One assumptionnecessaryto produce(1') is that real transactionscosts have remained essentiallyconstant. This is an assumption of doubtful validity and also may

StephenM. GoldJeld

581

wherek is relatedto the transactionscost. This impliesthat the income

elasticityof the demandfor moneyis 1/2whilethe interestelasticityis -1/2.
Other analysesof the demandfor money emphasizespeculative,precautionary,or utilityconsiderationsin additionto the transactionsmotive.8
These tend to blur the specificpredictionsof income and interestrate
elasticitiesthat emergefrom the simpletransactionsapproach,but they
are broadlyconsistentwith the generalform of equation(1).9
At an empiricallevel such an equationhas underpinnedestimationin a
numberof studiesof the demandfor money. This has typicallybeeinthe
case where annual data are involved. With quarterlydata, empirical
workershave generallyresortedto a more complicatedversionof (1) involvinglaggedas well as currentvariables.At least two motivations-not
necessarilyconflicting-havebeenofferedfor modifying(1) in this way,the
partialadjustmentmechanismandexpectationsformation.For the present
only the formerjustificationis explored,but the expectationallag will be
consideredmore extensivelybelow.
Theubiquitouspartialadjustmentassumptionusuallyproceedsby interpreting(1) as settinga "desired"valuefor moneyholdings,say m*, as in
systematicallybias standardestimatesof the demand for money. For one lighthearted
attemptto correctfor this bias, see SaschbaTelphlluch,"A Remarkon the Transactions
Demand for Money," CORE Discussion Paper 7034 (Catholic Universityof Louvain,
Belgium, 1970; processed).
8. See, for example, J. Tobin, "LiquidityPreferenceas Behavior Towards Risk,"
Review of EconomicStudies, Vol. 25 (February1958), pp. 65-86; and Don Patinkin,
Money, Interest,and Prices: An Integrationof Monetaryand Value Theory(2nd ed.,
Harperand Row, 1965).
9. The ideal would be a theory that simultaneouslytreatsthe variousconsiderations
cited above. Such a fully generaltheoryhas yet to be producedbut a numberof promising startshave been made. For example,Ando and Shell have recentlyanalyzeda model
in which risk and transactionscosts are handled simultaneously.They consider three
assets: equities,savingdeposits,and money. The rate of returnon equitiesand the rate
of change of the price level were consideredto be random variableswhile the nominal
ratesof returnon savingdepositsand moneyweretakenas knownwith certainty.Adopting an expectedutility frameworkbut allowing for transactionscosts, they were able to
show that the demandfor money becomesa function of the volume of transactionsand
the interestrate differentialbetweensaving deposits and money. Assumingthe latter is
zero leads to a formulationlike (1). In particular,money holdingsdo not dependon an
expectedreturnon equities,on wealth, or on anticipatedinflation.I shall returnto this
below. See Albert Ando and Karl Shell, "Demand for Money in a General Portfolio
Modelin the Presenceof an Assetthat DominatesMoney,"appendixto a paperpresented
to a Brookingsconferenceon model building, 1972 (June 1972; processed).

Brookings Papers on Economic Activity, 3:1973

582

m*= f(r, y).

(2)

Portfolioadjustmentcosts, both pecuniaryand nonpecuniary,are then

assumedto preventa full,immediate,adjustmentof actualmoneyholdings
to desiredlevels.Dependinguponthe functionalformof (2), actualmoney
holdingsare assumedto adjustlinearlyor logarithmicallyto the gap betweendesiredholdingsand last period'sholdings;that is,
MtMt-,

(3)

y7(m*-

y(ln m* - ln m,-),

01

or
In m, - ln mt-

(3')

wherey is the coefficientof adjustment.While,as demonstratedbelow,the

partialadjustmentmodelis not withoutits shortcomings,it seems,in view
of its widespreaduse, a convenientstartingpoint for empiricalwork.
A CONVENTIONAL
EQUATION

The firststep is estimatingan equationfollowingthe formatof (3') and

(1') above.Detaileddefinitionsof the variablesare foundin the appendix
but a few wordson the matterare in orderhere.The narrowmoneystock
(currencyplus demanddeposits,M1)is used as the dependentvariable;it
is measuredas a quarterlyaverageof monthlydata and deflatedby the
implicitGNP deflator.Incomewas definedas real GNP and the interest
rate was measuredin two ways-by the rate on commercialpaper(RCP)
andby the rateon timedeposits(RTD). The resultsobtainedwithordinary
least squares,using the Cochrane-Orcutt
techniqueto adjust for serial
numbers
in
parentheseshere and in folcorrelation,are given below (the
lowingequationsare t-statistics):
(4) lnm

0.271 +0.193Iny+0.717Inmm_
(2.2)

(5.3)

(11.5)
-

0.019 ln RCP

(6.0)
R2 =

0.045 ln RTD.

(4.0)

0.995; p = 0.414; standarderror = 0.0043; Durbin-Watsonstatistic = 1.73.

Sampleperiod = 1952:2-1972:4.10

10. This sample period was used in most of the equationsthat follow, primarilyfor
ease of comparisonwith equationsbased on the flow of funds data, which are available

Stephen M. Goldfeld

583

At firstglancethisequationseemsquitereasonable.Boththe commercial
paper rate and the time deposit rate are significant,with long-runelasticities of 0.07 and 0.16, respectively.The coefficientof adjustment-that
is, 7 in (3')-is 0.283(= 1 - 0.717);whilethisis not dramatically
rapid,it is
certainlymore plausiblethan the slow 0-10 percentestimatesthat some
writershave reported.'1The point estimateof the long-runincome elasticityis 0.68 and a 95 percentconfidenceintervalfor the incomeelasticity,
derivedby a methoddue to Fieller,12turnsout to be (0.60, 0.82). Consequently,the incomeelasticityappearsto be significantlyless than unity.'3
Besidesyieldingplausibleparametervalues, equation(4) also fits the
data quite well. This can be seen in Figure 1, which depictsthe actual
valuesof the real money stock along with the valuespredictedby equation (4).
INCOME ELASTICITY: A CLOSER LOOK

While equation(4) seems to be a satisfactoryfirst approximationto a

moneydemandfunction,the resultsneed closerscrutiny.One aspectthat
only from 1952. Equation(4), run over the longer sample period, 1949:2 to 1973:2, resulted in the following:
ln m = 0.286 + 0.179 ln y + 0.731 ln m, - 0.020 ln RCP - 0.040 ln RTD.
(4.6)
(3.6)
(3.2)
(4.9)
(12.0)
R2 = 0.988; p = 0.217; standarderror = 0.0073; Durbin-Watsonstatistic = 2.08.
The point estimatesin the above equation and in (4) are quite similar although there
are some indicationsof a differencewith respectto the errorstructure(for example,the
estimatedp and standarderror).
11. See, for example, the logarithmic specification in Franco Modigliani, Robert
Rasche, and J. Philip Cooper, "CentralBank Policy, the Money Supply,and the ShortTerm Rate of Interest,"Journalof Money, Creditand Banking,Vol. 2 (May 1970), pp.
166-218.
12. Fieller'smethod is neededsince the long-runelasticityis a ratio derivedfrom two
estimatedcoefficients.The resultingintervalwill, in general,not be symmetricaround
the point estimate.This is true here since the midpoint of the intervalis 0.71 while the
point estimate of the elasticity is 0.68. Furthermore,in the present context, since the
underlyingestimatesare not unbiased,I have only an approximateconfidenceinterval.
For a discussion of the Fieller technique,see WayneA. Fuller, "Estimatingthe Reliability of QuantitiesDerived from EmpiricalProduction Functions," Journalof Farm
Economics,Vol. 44 (February1962), pp. 82-99.
13. This is usually an implication of the transactionsapproach to the demand for
money. A problemarises in a concrete applicationof this approach,however, because
it is not clear that real GNP is a good measureof transactionsor that real transactions
costs are constant.

ON Iz

~ ON~~~~~~~~~~~0

Po~~~~~~

~~~~~~~~~~~~~~~~~-

P0

0'

*~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Stephen M. Goldfeld

585

deservesadditionalattentionis the estimateof the long-runincomeelasticity. Judgedby the size of the confidenceintervalreportedabove, the
estimateof this importantparameterappearsto be fairlyprecise.On the
other hand, WilliamPoole has suggestedthat the income elasticityestimated from quarterlypostwardata reallycannot be pinned down accurately.14Sinceit will shedsomefurtherlight on the qualityof the estimates
in (4), a briefexplorationof Poole'sargumentwill be worthwhile.
Supposean estimatingequationtakesthe form
(5)

lnm, = a+bIny,+

clnr +dlnm,-1.

Theshort-runincomeelasticityis b whilethelong-runelasticityis b/(l - d).

Supposethe long-runelasticityis constrainedto be somenumbere. Equation (5) then becomes
(6)

In m,-e

In Yt= a + c In r, + d(n

m,_

-e

In Y),

whichfor a given e could then be simplyestimated.A comparisonof the

propertiesof (6) for alternativevalues of e could then be made. Poole
triedvaluesof e rangingfrom0.5 to 3.0 and emphasizedtwo propertiesof
the resultingestimates.He found that the estimatedinterest elasticity
steadilyincreasedwith e, risingto 2.5-2.7 for e = 3.0; and the R2 of the
estimatedequationwas essentiallyflat for valuesof e from 1 to 3. It was
this latterfindingthat led Pooleto suggestthe impossibilityof obtaininga
firmestimateof the incomeelasticity.
The equationPoole primarilyfocusedon had one interestrate variable
and no laggeddependentvariable;it was, that is, like (6) withd = 0. It is
consequentlyof someinterestto see how equation(4) behavesfor alternativevaluesof e. Table1 reportsthe relevantresults,givinglong-runinterest
parameter,-y, and
elasticitiesfor RTD and RCP, the speed-of-adjustment
the R2andstandarderror.The interestelasticitiesdisplaya cleartendency
to increasewithe, butthe riseis not nearlyas pronouncedas Poolefound.'5
The tablealso showsa systematicdeclinein the speedof adjustmentas e
increases.

As for the relativeexplanatorypowerof the equationas e increases,the

table pointsto uniformlyhigh R2s,whichrise steadilywith e. Thatthis is
14. WilliamPoole, "WhitherMoney Demand?"BrookingsPapers on EconomicActivity (3:1970), pp. 485-500.

15. Although it is not indicated,the t-statistic for RTD declined to about 0.5 as e
increased.

orc

00t

(ON

C.

C4 Cmb4

Zs

~ ~~~~~~~~~~~1
Oooo.oC
O
C CC))o

4)~

Y~~~~~~~~k
on c,

,-

'0

rA~~~~~~~~~~~~~~~~~~~~~~~~~~4
0
8 ?
0

oC

oc

r\4)

>

t-

*s=

oo

oo

O
cr)
ONE

ON

4)

l- ON

.)

cr aN cr cONr
(ON
cr
P (Ocr
O
ON O ON ON

aN

'0
-4)

Not?XoOOW

C) C

'4)

0 ) O

E 00000000)

5oy~~~e

C)

'

4)

4)0

4z -

E~~~~~~~e
4)~~4,
W
o
rE

C3.

o
u

4)
X

i'

co

vo

00

ON

oD 00

00000000
~~~~~~~~~~~

tWWXoOmOeUO

n CE a OOcN
:4)

enr
;

'0

1l)~~~~~~~~~~~00
,,

t
I

4)
O )

Stephen M. Goldfeld

587

misleading,however,is plain in the second row of the table, which was

obtainedby constraininge to be the value impliedby equation(4). This
procedurenaturallyreproducedthe resultsof that equationexceptfor the
R2.The troubleis that the dependentvariablein (6) changesas e changes
and consequentlythe R2 is not strictlycomparableacross rows of the
table.'6The standarderrorof the regression,whichis comparable,tells a
differentstory.It clearlyis lowestfor the equationreportedin the second
row, as it shouldbe. As e rises so does the standarderror,althoughthe
deteriorationis mild.
Another,perhapsmore useful, way of looking at the overallperformanceof equation(4) for alternativevaluesof e relieson dynamicsimulation. In a dynamicsimulationthe laggedvaluesof the dependentvariable
that arefed into the equationarethose that aregeneratedby the equation
itself,not the historicalvalues.'7Thisis in generala morestringenttest of
an estimatedequationthan somethinglike the R2, and indeedis probably
a morerelevanttest froma forecastingpointof view.In this vein,I dynamically simulatedthe basic equationover the full sampleperiod for each
error(RMSE)of the
valueof e. Table1 also reportsthe root mean-squared
simulatedaroundthe true values.The firstRMSE columnis in the same
unitsas the standarderrorwhilethe secondconvertsthe logarithmicequation to dollar levels so that the units are in billions of 1958 dollars.18
Equation(4) (the secondrow of Table1) yieldedan RMSEof $1.1billion.
Alternativevalues of e led to a deteriorationof the RMSE much more
markedthan the correspondingworseningof the standarderror of the
natureof this technique.'9
regression,pointingup the morediscriminating
An even more vivid illustrationof this point arises from the ex post
performanceof the basicequation.The last columnof Table 1 reportsfor
alternativevalues of e the root mean-squarederrorsobtainedfrom esti16. Poole's resultspartlyreflectthis R2illusion but he has a numberof specifications
that do not sufferfrom this difficulty(for example,the one using the interestrate as the
dependentvariable).
17. Dynamicsimulationsin the presenceof seriallycorrelatederrorsalso involve an
additionalcorrectionfor the lagged disturbanceterm.
18. The simulatedvaluesof the level wereobtainedsimplyby takingantilogs.In fact,
this is not the best way to obtain them, but roughcalculationssuggestedthat the proper
correctionwas small. On this see ArthurS. Goldberger,"The Interpretationand Estimation of Cobb-Douglas Functions,"Econometrica,Vol. 36 (July-October1968), pp.
464-72.
19. The RMSE in row 2 of Table I is roughly twice the standarderror but rises to
over threetimes for e - 2.0.

4?

oo To oo
~~~~~~~~~~~enX

.w

v~~~~~~~~~~~~~~~~~~~~~~~~1

+r
0
~~~~~~~~~~~~~~~~~~~~~~~
00 CT,C7o-

8~~~~~~~o

5
oCo

Xo 00N8WS

ce @
~~~~~~~~o
O
OOOOO
X~~~~~~~~~~~~~~~~~~~~~~~~~~~~~4

X~~~~~~~~~~~~ xt.o oo0 ooonNNn oo

,b
X~~~~~~~~~~~~C

Y~~~~~~~~&
g

S_

1;~~~~~~~O
s
Po

kn
m

o
o

o o o
o o
o~~~~~~~~~

o)

CD
)

ot

W)

;Y

00
D C>

>

dF t

0!
>s a

ot

co co o

.r':~~~~~~~~~~~~~2

~~~~~~~~~~~~~~~~~~~~~o
0

000

cr)

oo ogo

O.
m~~~~~0

O > N _ N CN tH NGNa)rrfw

C>U 10

000000

*?~~~~~~~~~~~~~~~~~~~~~~~~~a
F

mo

o?0

Cd

!.
UJ

oo

=0,

en

r- r- r- ?

0
oo
CD

30

04

o o o o
o~~~~~~~

o S

??as
o

C)

StephenM. Goldfeld

589

matingthe equationthrough1961and extrapolatingforwardby dynamic

simulationto the end of 1972.The qualityof theseextrapolationsdeterioratesdramaticallyfor high valuesof e.20
On balance,then, the specificestimatesof equation(4) still seem satisfactory,both in terms of absoluteperformanceand relativeto the equations obtainedfor alternativeincome elasticities.Taken as a whole, the
resultsseem to suggestthat the relevantincomeelasticitycan be pinned
downwithina reasonablerangeof accuracy,and thatit is significantlyless
than unity,reflectingeconomiesof scale.
SHORT-TERM INSTABILITIES?

The tentativeconclusionjust reached-that an equationlike (4) does a

satisfactory
job of trackingmoneydemand-was basedon summarystatistics derivedfrom the within-sampleperformanceof the equation.However,one of the primaryconcernsis the potentialfor short-runinstability
in the demandfunctionfor money. This problemcan be attackedin a
way is to ascertainthe qualityof
varietyof ways,but one straightforward
the short-termex post forecastsgeneratedby this specification.To do this
the specificationin (4) was estimatedover twelve sampleperiods,each
startingin 1952:2 and differingin that the terminalpoint was systematically moved from the end of 1961 to the end of 1972, in steps of four
quarters.Based on the estimatesobtainedfor each sample period, the
equationwas dynamicallysimulatedfor the next four quarters.
A numberof featuresof the estimatedequationsare containedin columns 1 through6 of Table 2. Columns 1 through4 list the individual
coefficientestimates,whichon casualinspectiondo appearto shiftaround
somewhat.Columns5 and 6 give the standarderrorof the regressionand
the RMSE(in billionsof dollars)fromwithin-sample
dynamicsimulations.
Both these numberstend to rise as the end point is extended,in part
becausethe mean of the dependentvariableis also increasing.
Columns 7 and 8 assess the out-of-sampleforecastingperformance,
givingboth the RMSE of a four-quarterforecastand the meanerror.The
20. The estimateof the long-runincome elasticityobtainedfrom data through 1961
is lower than the full-sampleestimateof 0.68. Consequently,a more realisticestimateof
an attainableRMSE is higherthan the $1.65 billion reportedin Table 1 (see Table 2
below). Nevertheless,the more realisticestimate of roughly $2 billion to $5 billion in
Table 2 is distinctlylower than all the high e entriesin Table 1.

590

Brookings Papers on Economic Activity, 3:1973

data underlyingthese calculationsare plotted in Figure 1. The fourquarterforecastis for the yearfollowing the end pointfor a particularrow.
For example,the worstforecastingerroroccurredin 1966with an RMSE
of $2.3 billionand this appearsin the 1965row. In fiveof the twelveyears
the ex post forecastwas no worsethan the within-sampleRMSE, which
seemsa creditableperformance.Furthermore,this was truein 1971,a year
reputedto be one of instability,21
as well as in 1972.Theforecastsfor 1973
appearto be a bit wideof the markbut thisjudgmentis basedon only two
observations-of preliminarydata, at that-so one shouldnot make too
muchof it.
On the whole, the money demandfunction does not exhibit marked
short-runinstability.However,this is only one chapterof the short-term
forecastingstory. For one thing, the analysishas assumedboth known
interestratesandrealGNP. In addition,it explainsmoneydemandin real
terms so that to forecastnominalmoney demandwould requirea price
forecast,whichwouldintroducefurthererror.22Giventhesecaveats,however,it is reassuringto find a reasonabledegreeof short-runstability.
LONG-TERM STABILITY

Thecompanionquestionto the onejust consideredis whetherthe money

demandfunction is stable in the long run. This questionis usually addressedwith annualdata, often coveringa span of seventyor so years;
sometimesthe focus is on whetherthe samemoneydemandfunctionheld
both in the 1930sandin the rest of the period.23The concernhereis solely
withwhetherquarterlydata fromthe postwarperiodcan be usedhomogeneously in face of a numberof institutionaldevelopments(such as the
certificateof deposit and Eurodollarmarkets)that at least suggest the
possibilityof shiftsin the demandfor money.24
Long-runstabilitycan be examinedin a varietyof ways.Thedatasample
21. See, for example,the discussionof this issue in MichaelJ. Hamburger,"The Demand for Money in 1971: Was There a Shift?"Journalof Money, Creditand Banking,
Vol. 5 (May 1973), pp. 720-25.
22. One othertechnicalpoint should be noted. Table 2 is based on estimateswith the
latest and thereforefully reviseddata (exceptfor 1973).In practice,these data would not
be available.
23. See, for example,Laidler,Demandfor Money.
24. Slovin and Sushkahave reportedsome evidencethat the period 1955:1to 1962:1
may be differentfrom 1962:2to 1968:4.This intervalroughlycoincideswith the start of
the marketfor certificatesof deposit. See M. B. Slovin and M. E. Sushka,"A Financial

Stephen M. Goldfeld

591

can be split up at a priorichosenpoints25and the resultingestimatesfor

the subperiodscan be compared,either formally-say, via the Chow
test-or informally.Oneusefulinformalcomparisonis to simulatedynamically the equationbased on the firstpart of the period over the second
part,thusextendingthe techniqueusedin the previoussectionto a longer
forecastingperiod.
The last columnof Table2 reportsthe root mean-squared
errorsfor a
numberof such simulations.In each case, the money demandequation
was estimatedthroughthe indicatedend point and simulatedfrom the
followingquarterthroughthe end of 1972.The RMSEsarethus basedon
observationsover varyingperiods,the longest being forty-fourquarters.
As could be expected,these RMSEs are generallylargerthan the fourquarterRMSEs,althoughmarkedlyso only for the equationsreportedin
the firsttwo rows of the table. Moreover,these equationsdisplaycoefficients that differ substantiallyfrom subsequententries.This in turn is
consistentwith the Slovin-Sushkafindingcited earlierand arguesfor a
more carefulexaminationof the pre- and post-1961periods.Equations
(4') and (4") reportthe estimatesof equation(4) obtainedby breakingthe
sampleat the end of 1961.
(4') ln m = 0.699 + 0.216In y + 0.604In mi
(1.9) (4.6)
(6.4)
0.019 ln RCP
(5.4)

0.060 ln RTD
(4.1)

0.978; standarderror = 0.0036.

Sampleperiod: 1952:2-1961:4.

R2 =

(4") ln m

0.657 + 0.191 ln y + 0.632 ln m1

(1.8)

(3.3)

(4.8)
-

0.014 ln RCP

(2.4)

0.010 ln RTD.

(0.3)

R2 = 0.992; standarderror = 0.0050.

Sampleperiod: 1962:1-1972:4.
MarketApproachto the Demandfor Money and the Implicationsfor MonetaryPolicy"
(Board of Governorsof the FederalReserveSystem, 1972; processed).
25. Rather than split the sample at some given point, one may use techniquesfor
testing the hypothesisthat a split occurredat some arbitrarypoint in the period. A
number of these techniques are described in Stephen M. Goldfeld and Richard E.
Quandt,NonlinearMethodsin Econometrics(North-Holland,1972),Chap. 9.

592

Brookings Papers on Economic Activity, 3:1973

The biggestdifferencebetweenthesetwo equationsappearsin the coefficient of RTD and it is largelyattributableto the sizablejumpin RTD that
occurredpreciselyat the breakingpoint.26A formaltest of stability,carried
out by applyinga Chowtest to this samplesplit, resultedin an F statistic
of 0.84, whichdoes not allow one to rejectthe hypothesisof stability.27
On balance,then, the evidencedoes not seem to suggestany need to
estimatethe money demandequation over separatesubsamplesof the
postwarperiod.

AlternativeSpecificationsof the BasicEquation

Up to this point I have analyzedextensivelythe propertiesof essentially
one specification-thatembodiedin equation(4). As the firstsectionmade
clear, however,many questionsconcerningspecificationcan only be resolvedempirically.Thepurposeof the presentsectionis to shedsomelight
on these issues.
AGGREGATION AND DISAGGREGATION IN THE DEFINITION OF MONEY

Aggregation.To this point I have used the most commondefinitionof

money-M1, which is the sum of currencyand demanddeposits.Other
writers,however,have preferreda broaderdefinition,such as M2, which
includestimedepositsat commercialbanks.Thischoiceseemsquestionable
on a varietyof groundssinceit constrainsthe specification,includingthe
adjustmentpattern,of M1 andtime depositsto be the same.Furthermore,
sinceRTD shouldpositivelyaffecttime depositholdingsand shouldnegatively influenceM1 holdings,aggregationmay badly muddyinterestrate
26. The time deposit rate, RTD, jumped from 2.9 to 3.5 percentat this point. This
was the largestquarterlychange in the sampleand obviously is an importantinfluence
both on the varianceof RTD and, consequently,on the precisionwith which its coefficient can be estimated.Extendingthe sample period in (4') to include this observation
reduces the RMSE correspondingto the last column in Table 2 to 2.8. Furthermore,
includingthis observationin (4") as well makesthe coefficientsof RTD in the two equations considerablymore alike.
27. The Chow test is, strictlyspeaking,not quite valid here becauseof the use of the
lagged dependentvariablesand the serial correlationcorrection. A more appropriate
test, at least asymptotically,is the likelihood ratio test. This yielded a X2statisticof 9.3.
The appropriatecriticalvalue is 12.6 so this gives the same result as the F test.

Stephen M. Goldfeld

593

effects. On the other hand, an argument sometimes advanced in favor of

M2 is that it yields a more stable demand function.28In fact, according to
evidence developed later, this is definitely not the case.
The tabulation below contains the results of estimating equation (4) with
the M2 definition and with time deposits alone, and, for comparison, repeats the equation (4) estimates:

Definition
of money

Money
variable
Income lagged

RTD

RCP

R2

Standard
error

M2

0.119
(2.6)

0.948
(33.4)

0.006
(0.8)

-0.030
(7.7)

0.9987

0.0044

Time deposits

0.255
(3.0)

0.847
(18.7)

0.062
(4.7)

-0.051
(7.2)

0.9997

0.0075

ml

0.193
(5.3)

0.717
(11.5)

-0.045
(4.0)

-0.019
(6.0)

0.9953

0.0043

It is evident from these numbers that the use of M2 produces an equation

with properties quite differentfrom those of either of the component equations. First, the speed of adjustment is an unreasonably slow 5 percent per
quarter as compared with 15 percent for time deposits and 28 percent for
Ml. Second, RTD, as expected, has a negligible and insignificantimpact on
M2, reflecting the offsetting effects of the component equations. Finally,
the long-run income elasticity of M2 is a huge 2.3, which exceeds both the
1.7 for time deposits and the 0.7 for M1.
The only redeeming feature of the M2 equation is that its standard error
is only a smidgeon more than the M1 equation, while that of the time
deposit equation alone is substantially higher. This, however, is illusory as
can be seen by dynamic simulations. Table 3 reports the results of both
four-quarterex post forecasts and longer-term ex post forecasts obtained
by systematically changing the sample period as before. These are in columns 2 and 3 while the within-sample RMSE appears in column 1. Judged
on the basis of these results, the equation for M2 is extremely inadequate.
As compared with the results in Table 2, the RMSEs of the four-quarter
ex post forecast are both large and variable-ludicrously so in the longerrun extrapolations. From these results one would expect the equation for
28. See Laidler,Demandfor Money, p. 108.

594

Brookings Papers on Economic Activity, 3:1973

Table 3. Root Mean-SquaredErrors for M2 and Time Deposits,

and Income Coefficients,AlternativeSample Periods Ending with
1961 through1971
error
Root mean-squared
Currencyplus demandand
timedeposits,M2

Timedeposits

Ex post
End
pointa
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971

Sample
period
(1)
1.74
1.70
2.12
2.39
2.80
2.78
2.76
3.02
3.38
4.14
4.81

Fourquarter
(2)
1.54
5.05
2.30
3.56
3.25
7.84
3.09
7.00
9.81
5.11
1.10

Full
(3)
71.46
71.55
54.64
39.57
18.71
34.98
18.83
6.81
24.08
4.94
1.10

Ex post
Sample
period
(4)
0.94
0.96
1.30
1.60
2.27
2.35
2.52
2.32
2.98
2.29
2.18

Fourquarter
(5)
1.52
3.22
2.52
5.04
0.57
4.18
1.38
3.99
2.93
2.02
2.60

Full
(6)
46.76
43.37
37.61
28.58
10.86
10.80
4.87
3.68
6.64
1.73
2.60

Incomecoefficient
(7)
-0.011
0.026
0.039
0.072
0.156
0.154
0.169
0.177
0.191b

0.248b
0.267b

Source: Same as Table 2.

a. See Table 2, note a.
b. Coefficient significant at 5 percent level.

M2 to fail any formaltest for stabilityand, indeed,it does. Splittingthe

sampleat the end of 1961and applyinga Chowtest yieldsan F statisticof
3.53; the correspondinglikelihoodratio test yields a x2 of 18.6. Both of
these are significantat the 1 percentlevel, allowingone easilyto rejectthe
hypothesisthat the equationfor M2 is stableoverthe sampleperiod.
Sincethe M1 equationwaspreviouslyfoundto be stable,the suspicionis
that the difficultylies with the time depositcomponent,becausethat component is itself unstableor becauseof the aggregationprocess or both.
Superficially,
the time depositequationbased on the full sampleappears
quitereasonable.When subjectedto the kind of dynamicsimulationtests
just described,however,this equationalso appearsquestionable.The results are reportedin columns4 through7 of Table 3. The three sets of
RMSEs for time depositsare superiorto the correspondingRMSEs for
M2. Whenjudged by an absolutestandard,the within-sampleand fourquarterRMSEsmightbe acceptablebutthe full-periodRMSEsremaindistinctlyunreasonable.The sourceof the difficultyis indicatedin the last col-

Stephen M. Goldfeld

595

umn of Table 3, which reports the estimated income coefficient for alternative sample periods. That coefficient rises steadily over the period and does
not achieve statistical significance until the sample period runs through
1969. One would expect, as with M2, that the time deposit equation would
fail a formal stability test. The appropriate Chow F statistic is 4.25 and the
correspondingx2 is 22. 1, allowing one to reject stability by either test at the
1 percent level.
This finding suggests, at the very least, that the simple specification used
for M1 will not work for time deposits and therefore should not be implicitly so used by estimating a similar equation for M2.29The situation is,
however, worse than that, since even given the questionable time deposit
equation, the ex post forecasts of M2 obtained from the aggregate equation
are inferior to those obtained from adding together the separate component forecasts, thus suggesting that aggregation is inflicting some positive
harm in the present context.30
In summary, for both theoretical and empirical reasons, aggregation to
the level of M2 seems to be a distinctly inferior procedure.
Disaggregation. Although these findings confirm that greater aggregation in the estimation of the demand for money is not called for, there
remains the question of whether some disaggregation would be appropriate. The most obvious type of disaggregation would be to estimate separate
equations for currency and demand deposits,31 as is done in many macroeconometric models for a variety of reasons. For one, disaggregation permits greater flexibility in the choice of variables and specification of adjustment patterns. Second, and perhaps of more practical importance, currency
is needed as an endogenous variable for analyzing monetary policy. In
particular,a means of splitting up high-powered money (a variant of which
is usually taken as a policy instrument) into reserves and currency may be
needed to trace out the money supply mechanism. In any event, there are
good precedents for attempting to explain currency and demand deposits
separately.
29. I brieflyexperimentedwith several other interestrates in both the time deposit
and M2 equationsbut these never achievedstatisticalsignificance.
30. I spare the reader the additional numbers.However, the remarkin the text is
basedon addingtogetherthe separateextrapolationsfor Ml and time depositsand then
comparingthe RMSEs with those in Table 3.
31. Disaggregatingby type of holderis consideredbelow. To some extent, separation
into currencyand demanddeposits is also a partialstep in this direction.

596

Brookings Papers on Economic Activity, 3:1973

The tabulation below reports the results of estimating separate equations

for currency and demand deposits along the lines of equation (4).
Con- Money
sumer variDependent
expen- able
variable Income ditures lagged

RTD

RCP

R2

Standard
error

Demand
0.181
deposits (5.2)

...

0.693
(9.9)

-0.040
(3.7)

-0.021
(6.0)

0.992

0.0049

Currency

0.190
(5.3)

...

0.804
(19.0)

-0.046
(3.9)

-0.007
(2.0)

0.998

0.0042

Currency

...

0.591
(8.3)

-0.025
(1.7)

-0.001
(0.2)

0.998

0.0043

0.279
(6.2)

The first two equations use exactly the same specification and sample period
as equation (4). Both seem relatively satisfactory, and surprisinglyenough,
both interest rate variables show up in the currencyequation. The long-run
income elasticity of the demand deposit equation is 0.59, while that of the
currency equation is 0.97. These bracket the 0.68 elasticity found for M1.
The speed-of-adjustment coefficients also bracket the M1 result with demand deposits adjusting somewhat more rapidly than currency.
The final row of the tabulation contains the results of one minor modification in the currencyequation, the substitution of consumer expenditures
for GNP as the transactions variable and the corresponding use of the
consumption deflator.32 This procedure has pronounced effects on the
equation: first, it renders both interest variables statistically insignificant;
and second, it considerablyspeeds up the adjustmentof currencyholdings.33
How do the component equations stand up when subjected to dynamic
simulation? The relevant results are reported in Table 4. The two versions
of the currency equation perform comparably on the four-quartersimulations, producing only small forecasting errors. The demand deposit equation, as expected, yields smaller RMSEs than the aggregate equation
32. This, for example,was used in Modigliani,Rasche, and Cooper, "CentralBank
Policy."
33. There is some question, however,about the generalityof this second finding.In
particular,the laggedstock coefficientswith GNP and consumptionwere virtuallyidentical for all the equationsunderlyingTable 4 below. Only when 1971(or 1971and 1972)
were includedin the sample did the differencecited in the text emerge.

597

Stephen M. Goldfeld

Table 4. Root Mean-SquaredErrors for Extrapolations of

Demand Deposit and CurrencyEquations, Sample Periods Ending with
1961 through 1971
Full ex post extrapolation

Four-quarter
extrapolation

Currency,by
transactionsvariable

Currency,by
transactionsvariable
End
point"
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971

Demand
deposits
0.59
0.44
0.42
0.84
2.06
0.92
1.40
0.75
1.43
0.78
0.97

Consumer Gross
expendi- national
tures
product
0.10
0.61
0.30
0.23
0.34
0.13
0.17
0.25
0.09
0.19
0.35

0.09
0.68
0.05
0.22
0.45
0.25
0.34
0.10
0.23
0.22
0.37

Demand
deposits
2.07
1.25
1.94
2.27
1.78
2.21
1.79
0.90
1.15
1.14
0.97

Consumer Gross
expendi- national
tuires product
2.64
0.61
3.83
1.26
1.37
0.47
0.23
0.36
0.14
0.16
0.35

0.45
1.59
1.17
0.77
0.87
0.69
0.84
0.43
0.29
0.20
0.37

Source: Same as Table 2.

a. See Table 2, note a.

althoughit still makesa sizableerrorin forecasting1966.The long-term

extrapolationsfor all threeequationsalso performcreditably.As between
specificationsof the currencyequation,the GNP formulationdoes better
in the early part of the period but the consumptionspecificationdoes
betterat the end of the period.
Comparingthe RMSEsin Table 4 with those in Table 2 suggeststhat
extrapolationof M1 might be accomplishedbetter with the component
equations,especiallysinceany offsettingerrorsin the componentequations
shouldhelpin forecasting.To assessthis possibility,I summedthe separate
forecastsfor currencyand demanddepositsandthen computedthe appropriateRMSEs. These are reportedin Table 5, which also includes for
conveniencethe correspondingresultsfromTable2 (labeled"aggregate").
On the whole the ex post forecastsfrom the componentequationsdo
extremelywell. In particular,they improvemarkedlythe extrapolationsof
M1relativelyfarinto the future(seethe firsttwo rowsof the table).Overall,
the most successfulformulationis that whichusedconsumerexpenditures
as the transactionsvariablein the currencyequation.In the elevencases

Brookings Papers on Economic Activity, 3:1973

598

Table 5. Root Mean-SquaredErrors for Aggregate and Disaggregate

Forecasts of M1, Sample Periods Ending with 1961 through 1971
Four-quarterforecast

End
point"
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971

Full ex post forecast

Disaggregate, by
transactions variable

Disaggregate, by
transactions variable

Gross
national
product

Consumer
expenditures

Aggregate

Gross
national
product

Consumer
expenditures

Aggregate

0.64
1.13
0.41
1.04
2.49
1.10
1.72
0.73
1.64
0.88
1.12

0.64
1.05
0.53
1.06
2.53
0.97
1.54
0.77
1.40
0.87
1.09

1.42
1.60
0.65
0.88
2.33
1.14
1.48
0.70
2.05
0.86
1.10

1.87
1.56
2.97
2.95
2.48
2.82
2.50
1.04
1.28
1.15
1.12

4.48
1.37
5.47
3.35
2.84
1.93
1.75
1.02
1.19
1.15
1.09

5.22
4.08
1.24
2.57
2.19
2.71
2.21
1.13
1.53
1.23
1.10

Source: Aggregate columns are from Table 2; disaggregate data are derived from separate forecasts for
currency and demand deposits, the components of Ml.
a. See Table 2, note a.

consideredit yields an RMSE lower than the aggregateequationeight

times for long-termextrapolationsand six times for short-periodprojections. Thisevidenceprovidessomeindependentsupportfor modelbuilders
who choose to use separatecurrencyand demanddepositequationsand
who includeconsumptionin the currencyequation.34
On balance,the messageof this section should be clear: as far as the
moneydemandequationis concerned,moreratherthanless disaggregation
appearsto be desirable.
PARTIAL ADJUSTMENT, EXPECTATIONS, AND LAGS

So far, the analysishas reliedon a verysimpleform of dynamicadjustment,a Koyck-typeequationthat uses a singlelaggeddependentvariable.

Whilethis is a convenientspecification,it has the questionablefeatureof
restrictingthe adjustmentpatternof money holdingsto be the same with
34. This is the strategyfollowed in the FMP model. See Modigliani, Rasche, and
Cooper, "CentralBank Policy."

Stephen M. Goldfeld

599

respectto both income and interestrates. Carefulconsiderationof the

sourceof laggedadjustmentsin moneyholdingsis thus in order.
Thejustificationofferedabovefor the form of equation(4) restedon a
vagueappealto the partialadjustmentmechanism.Despitethe superficial
plausibilityof this mechanism,its theoreticalfoundationin the contextof
the demandfor money is unclear.For capital stock accumulationthe
mechanismis satisfactory,but the analogybetweenmoney holdingsand
capitalequipmentis far from perfectfor many reasons.One is that the
exactnatureof the costs involvedis muchless clearin adjustingfinancial
portfoliosthan in the case of adjustingstocks of machineryand plant.
Second,the lags that resultstatisticallyfor moneyadjustmentappeartoo
longto explainon groundsof adjustmentcosts.Finally,evenif the analogy
is granted,it does not necessarilyimply the simpleformulationof (3) or
(3') and indeeddoes so only undervery specialassumptions.35
This unsatisfactorystate of affairscan be partiallyremediedby reliance
on a differentrationalefor the laggedadjustment.Pushedback one step,
the adjustmentcan be conceivedas a slow responseof desiredstock itself
to actualcurrentvaluesof incomeand interestrates,ratherthana gradual
shift in moneyholdingsto meet a promptlyadoptednew level of desired
holdings.The responsecould be slow becauseof inertiaor becauseindividuals respondto expectedvalues that are in turn a function of past
values.36Of course,expectationaland partialadjustmentlags may existin
combination.
The workingsof a pure expectationsinfluencemay be examinedin a
demandfunctionof the form
(7)

m = a + bye + cre,

35. On this, see J. P. Gould, "AdjustmentCosts in the Theory of Investmentof the

Firm,"Reviewof EconomicStudies,Vol. 35 (January1968),pp. 47-55. Anotherproblem
with the partial adjustmentmechanismin the presentcontext is that the transactions
approachmay easily lead to "corner"solutions for an individual.That is, he may not
respond at all unless some critical condition is met (say, the interest rate changes by
more than a certain amount). This suggests the need to pay considerableattention to
the details of aggregatingover individualsto obtain a macro equation.
For a discussionof this point, see WilliamBreen,"A Note on the Demand for Cash
Balancesand the Stock-AdjustmentHypothesis,"InternationalEconomicReview,Vol.
12 (February1971), pp. 147-51.
36. On this, see Franco Modigliani, "The Dynamics of Portfolio Adjustmentand
the Flow of SavingsThrough FinancialIntermediaries,"in Edward M. Gramlichand
Dwight M. Jaffee (eds.), SavingsDeposits, Mortgages,and Housing(Heath, 1972).

600

RrookingsPaperson EconomicActivity,3:1973

whereye and r' areexpected(or, if one prefers,"permanent")

measures.37
Sinceye and re are unobservable,they mustbe replacedby measuredvariables.Onecommondevicefor doing so is to assumethat expectationsare
"adaptive,"that is,
(8)

Y,-1 = XCi,- Yt_i)

Y-

(9)

= X(rt -rte)

re-_e

Thisdeviceimpliesthatyt is a geometricdistributedlag of currentandpast

valuesof y; that is,
co

YXt
-p2(1G X -

(8')

i=o

iyti.

Equations(8) and (9) may then be combinedwith (7) to yield

Mt = a + bXyt + cXr,+ (1

(10)

X)mt-1.

Equation(10) obviouslyhas the same form as the equationsestimated

above,suchas (4), but Xhas a differentinterpretation.38
Equations(8) and
(9) have the same X,implyingthe restrictiveassumptionthat expectations
of y andr areformedanalogously.A morenaturalspecificationin placeof
(9) wouldbe
e

rt-r _r

(11)

= 6(rt -rt_),

wherea may be differentfrom X.Combining(7), (8), and (11) producesa

considerablymore complicatedestimatingequation:
(10')

mt =

ClYt +

Co +

C2yVt1 +

C3rA +

C4rt1

C5mt-1

C6mt-2,

wherethe cs are nonlinearfunctionsof the respectiveoriginalparameters.

This versionof the adaptiveexpectationsmodelleadsto a considerably
richerlag structure.In fact, even greatergeneralitymay be obtainedby
allowingexpectationsto adjustin differentproportionsto two or more of
the previouslyobservedforecastingerrors,as in39
(12)

-(

Yt-_) ?

X2(Yt-1

Yt_2)

37. To simplify notation I have omitted "In," although the specificationcontinues

to be logarithmic.
38. The disturbanceterm in (10), which is not shown, is actuallyof a differentform
from the one implicitin (4).
39. This has been suggestedin J. A. Carlsonand M. Parkin,"InflationExpectations"
(Purdue Universityand University of Manchester,May 1973; processed).Using (12)
instead of (8) and a correspondingreplacementfor (9) yields a version of (10') with
three lags for y and r and four lags for m.

StephenM. Goldfeld

601

Anotherextensionof the formulationis accomplishedby combiningthe

adaptiveexpectationsandpartialadjustmentmodels.Thisprocedureintroduces anotherlag in all the variablesand some furthernonlinearrestrictions.40

Table 6 reportsthe resultsof estimatinga relativelysimpleversionof

these alternatives,equation(10'), as well as two modifiedversionsthat
eitheromit the second-orderlag in the dependentvariableor the lagged
variablesfor income and interestrates. For comparison,equation(4) is
reportedas regressionA in Table6.
Severalfeaturesof the resultsare worthnoting. First, the long-runincome and interestrate elasticitiesare virtuallyidenticalfor all four equations.Thereare,however,differencesin the timingof the responsesamong
the fourequations.Thesedifferencesareillustratedfortwo of the equations
in Table 7, whichgives the fractionof the total responseto a changein
incomeor interestratesthat has occurredaftera givennumberof quarters.
For the simpleKoyck equationthis responseis identicalfor all variables,
but this is clearlynot the case for the secondequationin Table6.
A secondfeatureis thatthe threelaggedvariablesfor incomeandinterest
ratesarecollectivelysignificantwhenusedwithoutm laggedtwicebut not
when it is included.4'Finally,nmlaggedtwice appearssignificantwhether
or not these othervariablesare included.42
In my judgment,these results leave open the question of whethera
specificationmore complicatedthan the originalKoyck model is appropriate. Clearly,however,satisfactoryestimationof equation(10') is impededby pronouncedmulticollinearity.
Consequently,unlessthe nonlinear
restrictionsunderlyingsuch an equationare taken into accountproperly,
it seemspointlessto estimatea moresophisticatedversion.43
An alternative
andpotentiallymorepromisingrouteis to rely on Almondistributedlags.
40. It also permitsa test of the hypothesisthat eitherthe expectationsmechanismor
the partialadjustmentmechanismis absent. See, for example,EdgarL. Feige, "Expectations and Adjustmentsin the MonetarySector,"in AmericanEconomic Association,
PapersandProceedingsof the Seventy-ninthAnnualMeeting, 1966 (AmericanEconomic
Review,Vol. 57, May 1967), pp. 462-73.
41. The relevantF statisticfor these variablesis 3.5, which is significantat the 5 percent level, when the comparisonis between B and A in Table 6. The correspondingF
statisticfor comparingD and C is an insignificant0.4.
42. The equationsreportedin Table 6 were all estimatedassumingfirst-orderserial
correlation.Allowing for second-ordereffects did not qualitativelychange the results.
43. It should be noted that I have ignoredsuch restrictionsin estimating(10'). Basically, my energydeterioratedat this point.

o~

o~

so~
O

C
?0

0_

.o

.
~0

~~~~Eio

.o

.o~o

oc

0 * 0 * o
C4~~~~~~~~*
0

EF

'0

:xDoo

~
(J~~~~~~~~~~~0
~
00

O~~~0

r
i

(0,

"I.~~~~
00

~~~~~

.ZI .

C4I

tr)

C)

C
0DtC)
m

04C

.0
?y

tr)
c
U

)c
:

sC

0~~~~~~~~~~

Po

~~~~~~~~~~~~~~~~~~0

603

StephenM. Goldfeld
Table 7. Comparisonof CumulativePercentage Responses of
Regressions A and B of Table 6 after Selected Numbers of Quarters
RegressionB (text equation10')
Numberof
quarters

RegressionA
(text equation4)

Income
24.7
50.3
67.2
78.2
93.2

28.3
48.6
63.0
73.5
90.2

1
2
3
4
7

Interestrate Interestrate
on commercial
on time
paper
deposits
15.7
45.7
65.7
78.6
96.6

18.1
45.6
63.8
75.6
91.3

Source: Same as Table 6.

The basic estimatingequationfor this techniqueis givenby

n2

nj

(13) In m, =c + E
i=O

Wi l:nYt-i +

M3

"l

E w' In RTD,_
i=O

CtX

i=O

Thisequationcan be rationalizedin a numberof ways.44For example,the

form of its compositevariablesis simplya generalizationof equation(8')
with a finitehorizon.Alternatively,one maysimplyregard(13) as a convenientand flexibleequationfor approximatinga rathercomplicatedunderlying process.
A numberof a prioriexpectationssurroundthe coefficientsin (13). The
the lag distributionfor income,shouldall be positiveand
wis,representing
should probablydeclinemonotonically.The correspondinginterestrate
coefficientsshouldbe negative;they mightwell exhibita humpedpattern,
especiallyfor RTD, becauseRCP is likely to affectprimarilylarge transactors,who are less subjectto a learningdelay.45
Equation(13) was estimatedover the same sampleperiodas equation
(4)-1952:2 through1972:4-by the Almontechnique,withan adjustment
for serialcorrelation.The individuallag coefficientswere assumedto lie
on a third-degreepolynomialand no end-pointconstraintswereimposed.
Thelengthof eachlag (nl, n2,andn3)was determinedempiricallywith the
44. See, for example,Harold D. Dickson and Dennis R. Starleaf, "PolynomialDistributedLag Structuresin the Demand Function for Money,"Journalof Finance,Vol.
27 (December1972), pp. 1035-43, or Modigliani,Rasche, and Cooper, "CentralBank
Policy."
45. On this, see ibid.

604

Brookings Papers on Economic Activity, 3:1973

Table 8. Estimates of Income and Interest Elasticity Coefficients

Using Almon DistributedLags in the Demand-for-MoneyEquationsa
Dependent
Cuirrencyplus demanddeposits, Ml

Demand deposits

Income

RTD

RCP

Income

RTD

RCP

0.146
(3.6)
0.119
(4.8)
0.094
(6.9)
0.073
(6.5)
0.056
(3.7)
0.041
(2.2)
0.030
(1.5)

-0.028
(2.1)
-0.033
(4.9)
-0.034
(4.7)
-0.030
(3.7)
-0.021
(3.2)
-0.009
(1.3)
0.009
(0.6)

-0.014
(3.7)
-0.014
(6.7)
-0.012
(5.7)
-0.011
(3.9)
-0.009
(3.0)
-0.006
(2.0)
-0.003
(0.7)

0.131
(3.1)
0.105
(4.3)
0.082
(6.5)
0.063
(5.9)
0.047
(3.0)
0.034
(1.7)
0.024
(1.2)

-0.024
(1.67)
-0.031
(4.7)
-0.032
(4.2)
-0.028
(3.2)
-0.019
(2.6)
-0.004
(0.6)
0.017
(1. 1)

-0.014
(3.4)
-0.014
(7.2)
-0.013
(6.1)
-0.012
(4.2)
-0.010
(3.3)
-0.007
(2.2)
-0.003
(0.7)

0.022

...

...

0.018

...

...

0.015

...

...

(0.9)
0.015
(1.2)

...

...

0.019

...

...

0.025
(0.9)

(1.2)
0.017

(1.1)
0.016

(1.2)
0.018

(1.1)
0.023
(0.9)

2 = 0.656

...

...

(0.9)

(1.2)

2 =-0.145

2 =-0.068

(17.3)
(8.8)
(5.4)
R2 = 0.995, standard error = 0.0046, p = 0.82

2 = 0.577

2 =-0.121

2 =-0.073

(19.4)
(10.3)
R2 = 0.992, standard error = 0.0051, p

(5.8)
0.69

Source: Derived from text equation (13). The sample period is 1952:2 through 1972:4. For data sources
and definitions, see appendix.
RTD and RCP are the interest rates on time deposits and commercial paper, respectively.
a. The summations are calculated from data before rounding.

roughaid of the informationon speedof adjustmentfromthe stockadjustmentequations.

The resultsreportedin Table 8 agree remarkablywell with those obtainedearlier.For the long-runincomeand interestelasticities,whichare
reportedin Table9, the Koyck andAlmonestimatesdo not differby more
than 0.02. The equationsdo differ, of course, in the pattern of lagged
response,and on this score,the resultsof equation(13) seemsensible.The
length of the lag on incomeis substantiallylongerthan that of the correspondinglag for interestrates,a findingthat is roughlysupportedby some

605

StephenM. Goldfeld

variable
Currencywith income as transactionsvariable
Income

RTD

RCP

0.153

-0.020

-0.0045

(5.4)

(1.3)

(1.2)

0.126
(6.2)
0.103
(7.2)
0.082
(7.4)
0.065

-0.018
(2.0)
-0.017
(1.5)
-0.016
(1.8)
-0.016

-0.0047
(2.1)
-0.0047
(2.0)
-0.0044
(2.1)
-0.0039

(6. 1)
0.051

(1 .0)
...

(1 .1)
...

Currencywith consumerexpendituresas
transactionsvariable
Consumer
expenditures
0.247
(3.8)

0.261
(5.4)
0.191
(3.9)
0.036
(0.5)

RTD

RCP

-0.028

-0.0050

(1.6)

-0.012
(0.9)
-0.009
(0.7)
-0.020
(1.1)

(1.1)

-0.0036
(1.3)
-0.0023
(0.8)
-0.0010
(0.2)

.......

(4.4)
0.041

...

.....

...

...

...

.....

...

...

...

...

(3.2)
0.033

...

(2.5)
0.029

(2.3)
0.028

...

...

...

(2.3)
0.031

...

...

.....

...

...

...

......

...

...

...

...

(2.6)
0.037

(2.8)
0.046

.....

(2.6)
0.058..........

(2.4)
2 =-0.022
2 =0.883
% = -0.086
(13.4)
(2.8)
(2.7)
R2 = 0.998, standard error = 0.0045, p = 0.97

Z
2 = 0.734
2c -0.069
-0.012
(12.5)
(2.4)
(1.5)
R2 = 0.997, standard error = 0.0048, p = 0.97

previouswork.46Furthermore,the peak impactof RTD occursaftertwo

quarters,so the interestrate responsedoes exhibit the humpedpattern
positedabove.
More details on the exact timing of responsesare given in Table 10,
whichreportsthe fractionof the total responseto changesin incomeand
interestrates that has occurredafter a given numberof quarters.47The
46. See, for example,A. A. Shapiro,"Inflation,Lags, and the Demand for Money,"
IzternationalEconomicReview,Vol. 14 (February1973),pp. 81-96, and Feige, "Expectations and Adjustments."
47. It should be recalled that the dependentvariable is measuredin logarithms,

606

Brookings Papers on Economic Activity, 3:1973

Table 9. Comparisonof Long-runIncome and Interest Elasticities from

Koyck and Almon Estimates, for Money and Components
Koyck
Dependent
variable

Income

Money, M1
Demand deposits
Currency
Currencyb

0.68
0.59
0.97
0.68

Almon

RTD&

RCPa

0.16
0.13
0.23
0.06

0.07
0.07
0.04
0.00

Income
0.66
0.58
0.88
0.73

RTD&

RCPa

0.15
0.12
0.09
0.07

0.07
0.07
0.02
0.01

Sources: Koyck. equation (4); Almon, equation (13). RTD and RCP are the interest rates on time deposits and commercial paper, respectively.
a. All interest elasticities are negative.
b. Currencyequation using consumer expendituresas a transactions variable.

Table 10. Comparisonof CumulativePercentage Responses, after

Selected Numbers of Quarters, of Koyck and Almon Equations,
for Money and Components
Dependent variable
Demand deposits

Currencyplus demanddeposits, Ml
Number

Almon

Almon

of

quarters

Koyck

Income

RTD

RCP

Koyck

Income

RTD

RCP

1
2
3
4
7
10

28.3
48.6
63.0
73.5
90.2
96.3

22.2
40.3
54.6
65.7
85.1
93.5

19.3
42.1
65.5
86.2
100.0
100.0

20.5
41.2
58.8
75.0
100.0
100.0

30.7
51.9
66.7
76.9
92.3
97.3

22.7
40.9
55.1
66.0
84.2
92.5

19.8
45.4
71.9
95.0
100.0
100.0

19.2
38.4
56.2
72.6
100.0
100.0

Dependent variable
Currencywith income as transactionsvariable

Currencywith consumerexpendituresas
transactionsvariable
Almon

Almon

1
2
3
4
7
10

Koyck

Income

RTD

RCP

19.6
35.3
48.0
58.2
78.2
88.7

17.3
31.6
43.3
52.6
70.4
80.6

23.2
44.2
70.0
82.6
100.0
100.0

20.5
41.8
62.2
82.2
100.0
100.0

Consumer
Koyck expenditures
40.9
65.1
79.4
87.4
97.5
99.5

33.7
69.2
95.2
100.0
100.0
100.0

RTD
40.6
58.0
71.0
100.0
100.0
100.0

RCP
41.7
71.7
90.9
100.0
100.0
100.0

Sources: Same as Table 9. RTD and RCP are the interest rates on time deposits and commercial paper,
respectively.

Stephen M. Goldfeld

607

Koyckversion-equation(4)-necessarilyhas only one patternof response

while the Almon equationhas three separatepatterns.The Almon responsesto incomechangesareuniformlyslowerthanthe Koyckresponses.
For interestrates, the Almon responseis slowerfor severalquartersbut
then overtakesthe Koyck response.Evidently,constrainingall the responsesto the sameshapein the Koyckversionproducesan inappropriate
averageresponsewhichmasksindividualdifferences.
In additionto the resultsfor M1,Tables8, 9, and 10 reportthe findings
of Almon versionsof separateequationsfor estimateddemanddeposits
and currency.The resultsfor demanddepositsare, not surprisingly,quite
comparableto those for M1 both in termsof absoluteperformanceand in
comparisonwith the Koyck version presentedon page 596. Somewhat
largerdifferencesemergebetweenthe Koyck and Almon versionsof the
two currencyequations;but on the whole the Almon currencyequation
performscreditably.
In summary,a modestamountof evidencesuggeststhat the Koyck formulationof equation(4) is a bit too restrictive.48
The pricepaid for this
simplificationdoes not seem severebut it deservesadditionalresearchfor example,to examinethe comparativeperformanceof alternativelag
structuresin such simulationexperimentsas those reportedearlier.
INFLATIONARY EXPECTATIONS

The discussionof lags and expectationformationin the previoussection

was restrictedto incomeand interestrate variables.This sectionexplores
anothervariable-prices-and particularlyinvestigateswhetherinflationaryexpectationshavean independentroleto playin the demand-for-money
function.
Even at the theoreticallevel, this questionis controversial.On a strict
transactionsview of the demandfor money, a variablemeasuringanticipatedinflationseemsto haveno place.49On the otherhand,in theoretical
48. There appears to be some serial correlationleft in the Almon equations even
after correctingfor first-ordercorrelation.However, Dickson and Starleaf,in "Polynomial DistributedLag Structures,"performa second-ordercorrectionin a somewhat
analogousM1equationand get essentiallythe same kind of resultsas I did. For example,
their income elasticityis identicalto the one in Table 9 and the interestelasticitiesare
quite close.
49. Under suitable assumptionsthis can be formallyshown, as in Ando and Shell,
"Demand for Money." Inflationaryexpectations will be reflectedto some extent in
nominalinterestrates and thus will indirectlyaffect the demandfor money.

608

Brookings Papers on Economic Activity, 3:1973

writingson demand-for-money
functionsin the Chicagotradition,money
servesas an alternativefor physicalgoods, and the expectedrate of price
changeis given a prominentrole.50This approachhas beenbuttressedby
empiricalevidencefrom hyperinflationsabroad.In view of these latter
findings,HarryJohnsoncalls the absenceof "Americanevidencethat the
expectedrateof changeof pricesentersthe demandfor moneyfunction...
somethingof a puzzle."'51
He tentativelyattributesit to the relativemildness of U.S. inflationsand to the possiblepresenceof thresholdeffects.52
In the spiritof empiricismof this paperandin light of the divergenceof
opinionjust cited,the performanceof expectedinflationvariablesin money
demandequationswill be given a brieflook. Followingone of manypossible routes, I shall modify equation(7) to include an expectedrate of
inflation,pe:
(7T)

m = a + bye+ cre +dpe.

If the expectedrate of inflationis definedby an adaptiveexpectations

mechanismas in (8) or (9), the resultingequationtakesthe form53
(14) Inm, = a+ blny,+

cInm,_1 +dln

RTDt
+ e ln RCPt + f ln (P1/P,_i).

The resultsof estimatingequation(14) are given in the firstrow of Table 11. The pricevariableis quite significantlynegativeand its inclusion
raisesthe elasticityfor incomeandlowersthe speedof adjustmentas comparedwith equation(4). The elasticitiesfor the interestrate variablesremain virtuallythe same. (The measuresin rows 2 and 3 are considered
afterthe discussionof Table 12.)
50. See, for example, the various studies in Milton Friedman (ed.), Studies in the
QuantityTheoryof Money(Universityof Chicago Press, 1956).
51. HarryG. Johnson, Macroeconomicsand MonetaryTheory(Aldine, 1972), p. 127.
52. Also relevanthere is the notion of Allais that people will pay more attention to
currentand less to past events the more rapidlythe currentsituationis changing.This
suggeststhat one needs more than a simple distributedlag of past rates of inflationto
measureexpectedinflation.See MauriceAllais, "A Restatementof the QuantityTheory
of Money," AmericanEconomicReview,Vol. 56 (December 1966), pp. 1123-57.
53. The functionalform for the expectedinflationterm in (14) is equivalentto using
APt/Pi-, directlywithout logarithms.This is so since
ln (P/Pt1)

In

APt/Ptgi.

The regressionsreportedbelow were,in fact, estimatedboth as shown and with APg/Ptas a variableand the resultswere identicalto three decimal places.

609

Stephen M. Goldfeld

Table 11. Coefficientsof Variables in Demand-for-MoneyEquations,for

Three Measures of Price Expectations
Interest rate
Income

Money
lagged

Tiune Commercial
paper
deposits

Price
variable

R2

Equation (14)

0.166
(4.9)

0.782
(13.1)

-0.038
(3.6)

-0.015
(5.0)

-0.657
(4.2)

0.996

0.46

de Menil I

0.200
(5.6)

0.698
(11.3)

-0.046
(4.1)

-0.016
(4.9)

-0.143
(1.9)

0.996

0.41

de Menil II

0.200
(5.6)

0.693
(11. 1)

-0.044
(4.0)

-0.016
(4.8)

-0.211
(1.8)

0.996

0.41

Measure

Sources: Row 1 gives the results of estimating equation (14) as derived in the text, defining the expected
rate of inflation by an adaptive expectations mechanism. In rows 2 and 3, direct measures of price expectations from series constructed from surveys of expected price performanceare substituted in equation (14).
The series are from G. de Menil, "Rationality in Popular Price Expectations"(Princeton University, August
1973; processed). For other data sources, see appendix.

The impactof the pricevariableon the moneydemandequationcan be

assessedby a simpleconceptualexperiment.In an equilibriumsituation
thathaspersistedlong enough,andin whichinterestratesareconstant,real
incomeis growingat 4 percentand the actualrateof inflationis 2 percent,
equation(14)statesthatrealmoneystockshouldgrowat 3 percentandthe
change
nominalmoneystockat 5 percent.Now imaginea once-and-for-all
2
to
in the rate of inflationfrom percent 6 percentthat leavesinterestrates
and the rate of growthof real GNP unchanged.In the long run, the rate
of growthof the realmoneystockwill remain3 percent,thoughthe nominal money stock will grow at 9 percent.In the short run, however,substantialdeviationsfromtheseratesof growthwill occurif incomegrowth
and interestratesare to remainunchanged.I used the estimatesof equation (14) to computethese short-rundeviations,with the resultsreported
in Table 12.
Thelargesteffectoccursin the initialquarterandaftereightquartersthe
growthrates have nearlyreachedtheir equilibriumvalues.At that point
the realmoneystockis 21/2percentbelowwhereit wouldhavebeenhadthe
rate of inflationremainedunchanged.The nominalmoney stock (given
behaviorof the Federal Reserve)is
the assumedsuper-accommodating
51/2percent higher.

It is, of course,unrealisticto assumethat nominalinterestrateswill be

unchangedin the face of this higherrate of inflation.For illustrativepurposes, assumethat RTD wouldrise from 5 percentto 6 percentand RCP
from6 percentto 9 percentas a resultof the higherinflation.The resulting

610

Brookings Papers on Economic Activity, 3:1973

Table12. Short-runRatesof Growthof the MoneyStock in Transition

from2 Percentto 6 PercentInflation,withFixedand
VariableInterestRates
Percent
Interestratesfixed
Quarter

Real

1
2
3
4
5
6
7
8

0.4
1.0
1.4
1.8
2.1
2.3
2.5
2.6

Interestrates variable

Nominial
6.4
7.0
7.4
7.8
8.1
8.3
8.5
8.6

Real

Nominal

-5.7
-3.8
-2.3
-1.1
-0.2
0.5
1.0
1.5

0.3
2.2
3.7
4.9
5.8
6.5
7.0
7.5

Source: Computed from estimates of equation (14). For the variable interest rate colunis, the interest
rate on time deposits is assumed to rise from 5 percent to 6 percent, and that on commercial paper from 6
perceht to 9 percent.

money growthrates are given in the final two columnsof Table 12, and
reveal more dramaticvariations.At the end of eight quartersthe real
moneystockis about 81/2percentlowerthan it otherwisewouldhave been
and the nominal money stock is 1/2percent lower.

While the specificinflationaryassumptionsand calculationsare unrealistic,the resultsin Table 12 indicatethat substantialshort-runvariations in the growth of money demandmay accompanychangesin inflationaryexpectationsand these in turn may immenselycomplicatethe job
of the monetaryauthorities.
It is also possibleto interpretequation(14) as arisingfrom a partial
adjustmentmodelratherthan fromexpectationallags. To do this requires
modifyingthe equationdefiningthe desiredstock of money-for example,
(2) above-to includethe anticipatedrate of inflation:
(2')

m=

7 p

Under this interpretation,however,equation(14) resultsonly in the unlikely event that expectationsare perfectlyaccurate-that is, only if pe
APt/Pt-1.Fortunately,somealternativemeasuresfor pe yield a more satisfactoryinterpretation.In particular,Georgede Menilhas constructedtwo
series of expectedprice performancefrom the annual surveys of inflationaryexpectationsconductedby the SurveyResearchCenterof the Universityof Michigan,that can be used to give a directmeasureof expecta-

StephenM. Goldfeld

611

tions.54 Substituting these in equation (14) leads to substantially smaller

price coefficients(see Table 11), which barely border on statisticalsignificanceand do not providestrongsupportfor the anticipatedinflation
variable.
In fact,an alternativeview of the stock adjustmentprocesssuggeststhat
(14) is misspecifiedregardlessof how pe is measured.In particular,it may
be more plausibleto combine (2) or (2') with an adjustmentequation
specifiedin nominalterms:
(3")

ln M,

In M,zl

y (In M*

In M1-),

whereM,*= P,m*.If this is done the followingequationresults:

(15) In(MI/P,) = a + b In y, + c In(M,-,/P')
+ d In RTD, + e1lnRCPt +fpe.
The majordifferencebetween(14) and (15) is the deflatorfor the lagged
nominalstock of money. Equation(15) uses the currentprice level while
(14)usesthe laggedpricelevel.Withinthe contextof the stock adjustment
model,equation(14)thusimpliesthat anyreductionof the realvalueof the
laggednominalmoney stock due to risingpricesis subjectto immediate
adjustment,while equation(15) views it as subjectto partial or lagged
adjustment.
When(15)wasestimatedwith each of the threepossiblemeasuresfor pe,
it neveryieldeda significantcoefficientfor the pe.55At leastunderthe stock
of equaadjustment
interpretation,then,this suggeststhat misspecification
tion (14) led to a spuriouseffect of pe.56 Under the expectationallag hypothesis,(14) is the properspecification.57
The expectationalversioncan be investigatedfurtherwith Almon distributedlags. Among otherthings,this techniquehas the virtueof getting
54. The series are denoted de Menil I and de Menil II here, and their construction
is describedin G. de Menil, "Rationality in Popular Price Expectations"(Princeton
University,August 1973; processed).
55. Equation(15) without pe yielded essentiallythe same results as equation (4).
56. If the correct hypothesisis (15) withf = 0 and one estimates(14), one would expect to find the coefficientf in (14) to be roughly equal and opposite in magnitudeto
the coefficientc. This is so since c ln Mt-I/Pt c In MtI/Pt, - c ln Pt/Pt-1. Row 1 in
Table 11 suggeststhat this is indeed the case.
57. One findingthat is invariantto whether(14) or (15) is correct is the result containedin Table 12. Whilethe numbersare slightlydifferentfor (15), the basic story told
by that table holds.

612

Brookings Papers on Economic Activity, 3:1973

Table 13. CoefficientsShowingEffect on Equation(16), of Three Alternative

Measures of Price Expectations
Interest rate

Measure

Income

Time
deposits

Commercial
paper

Equation (17)

0.693
(16.7)

-0.157
(8.9)

-0.062
(4.8)

de Menil

0.652
(17.6)

-0.144
(9.1)

de Menil II

0.641
(17.9)

-0.138
(8.9)

Price
level

R2

Standard
error

-1.911&
(2.1)

0.996

0.0044

0.84

-0.066
(5.1)

-0.088
(1.1)

0.995

0.0046

0.81

-0.064
(5.2)

-0.257
(2.1)

0.995

0.0045

0.80

Sources: Row 1 gives the results of estimating equation (16), as derivedin the text, with Almon distributed
lags (equation 17). In rows 2 and 3, direct measures of price expectations from de Menil (cited in Table 11),
are substituted in equation (16). For other data sources and definitions, see appendix.
a. Individual coefficients are as follows:
Lag

Coefficient

-0.607
(3.2)

-0.440
(2.2)

-0.311
(1.4)

-0.222
(1.0)

-0.172
(1.0)

-0.160
(1.1)

the lagged money stock out of the equation and removingthe possible
statisticalartifactjust cited. The relevantestimatingequationis
n2

ni

(16) ln m =k +

wi In y,-i +
i=0

E
i3O

'

ln RTDt-i
ns

+ E w'' ln RCPt-i + bpe,

i=O

in which expectedinflationcan be expressedby either of the de Menil

measuresor by
n4

(17)

pe

wi"' ln (Pt_t/Pt--1).

i=o

The variousresultsare givenin Table 13.58Only one of the two equations

using the direct measureshas a statisticallysignificantprice effect, and
even that effectis much smallerthan that yieldedby the distributedlag
proxy-that is, equation(17). This latter variableseemsto work reasonably well; it producesa sensibledynamicadjustmentpattern(shown in
58. To conservespace, except for the price variablefrom (17), I have reportedonly
the sum of the lag coefficients.The individualcoefficients,however,wereextremelyclose
to those reportedin Table 8. The same lag lengths and polynomialdegreeswere used
in Tables 8 and 13.

StephenM. Goldfe(d

613

Table 13, note a) in whichthe lengthof the lag for past ratesof inflation
is slightlyshorterthanthat for interestratesandconsiderablyshorterthan
the incomelag.59
Takentogether,these resultsare a mixedbag. Under the expectations
view, some case emergesfor includinga measureof expectedinflationin
the demandfor money.On the otherhandthe partialadjustmentview, at
least as amended,suggeststhat this case may rest merelyon a statistical
curiosity.The readershouldfeel free to indulgehis own prejudices.

THE APPROPRIATE SCALE VARIABLE: INCOME OR WEALTH?

An issuethat has been extensivelyexaminedin the literatureis whether

incomeor wealth(or perhapspermanentincome)is the appropriatescale
variable.Laidlerhas reviewedthis literatureand concludesthat the evidencefavorswealth.Citingworkof Meltzer,he suggeststhat once wealth
is included,incomehas little to explain.Furthermore,he reportswork of
Brunnerand Meltzerthat suggeststhat the wealth variablehas superior
predictiveability.60Nonetheless,numerouswriterscontinueto follow the
transactionsapproach,whichfocuseson incomeas the primaryscalevariable.
The evidencecited by Laidleris based on long-termannualdata while
recentwritingsfollowingthe transactionsapproachhavetendedto be concernedwitha shorterterm.Whateverthe meritsof Laidler'sevidencein the
long-termcontext,the conclusionsdo not necessarilyapplyin explaining
the short-rundemandfor moneywith quarterlydata, andtheirrobustness
shouldbe examined.
Whilethe transactionsapproachemphasizesincome,it allowsroom for
a wealth variablesince some transactionsare obviouslyassociatedwith
portfolioshiftsrelatedto total wealth.Unfortunately,a good measureof
suchtransactionsis difficultto obtain.An attemptto use the valueof stock
(equity)transactionshad only limitedsuccess.61Anotherpossibilityis to
add the changein net worth to the variablesin the demandfor money,
59. At least one vaguelysimilarequationthat has been reportedin the literaturehas
the same feature.See Shapiro,"Inflation,Lags, and the Demand for Money."
60. See Laidler, Demandfor Money, Chap. 8, and the referencescited therein, pp.
121, 123.
61. See Modigliani,Rasche, and Cooper, "CentralBank Policy."

614

Brookings Papers on Economic Activity, 3:1973

Table 14. Estimates of the Money Demand Equation with Alternative

Wealth .nd Income Variables
Interest rate
Equation
(4) variant

Money
lagged

Time Conmmercial
deposits
paper
Income

0.920
(25.4)

-0.027
(2.5)

-0.015
(4.2)

...

0.104
(3.9)

0.986
(30.7)

-0.005
(0.5)

-0.010
(2.9)

...

0.040
(1.5)

0.801
(12.5)

-0.031
(2.7)

-0.014
(4.1)

0.139
(3.6)

0.729
(11.4)

-0.049
(4.2)

-0.018
(5.7)

0.165
(3.8)

0.801
(12.5)

-0.031
(2.5)

-0.014
(4.1)

0.140
(3.3)

Wealth

Changein
wealth

0.52

0.995

0.201
(3.2)

0.39

0.995

...

0.160
(2.9)

0.35

0.996

0.032
(1. 1)

...

0.43

0.995

0.161
(2.7)

0.35

0.996

-0.001
(0.04)

...

Source: Derived from variants of the basic money demand equation (4). For data sources and definitions,
see appendix.

thus allowingmoney holdings to absorb an arbitraryfraction of initial

allocationsof new wealth.62
Table 14 reportsthe resultsof estimatingseveralvariantsof the basic
equation.The firstsubstitutesa measureof net worthfor the incomevariable while the second uses both net worth and its change.The next two
equationsuse the incomevariableand one of the net worthmeasureswhile
the last equationutilizesall three.Severalfindingsare worthemphasizing.
First, withoutan incomevariablethe speedof adjustmentbecomesunreasonablylow. Second,incomeand the changein net worthboth achieve
statisticalsignificancewhenthey appearin the same equation,suggesting
thattransactionson wealthaccountmaywellbe important.Finally,unlike
the resultscited above, the level of net worth is unimportantwhen used
with incomealone whilethe latterretainsits significance.When all three
variablesare used,the level effectof net worthis obliterated.
The predictiveabilityof the variousequationsis reportedin Table 15,
62. This suggestionhas been made by WilliamC. Brainardand James Tobin, "Pitfalls in Financial Model Building," in American Economic Association, Papers and
Proceedingsof the EightiethAnnualMeeting,1967 (AmericanEconomicReview,Vol. 58,
May 1968),pp. 99-122. Brainardand Tobin actuallyspecifya demand-for-moneyequation in the context of a complete balancesheet, so they expressmoney as a fraction of
wealth as a function of interestrates and income. Furthermore,they suggestdecomposing the change in net worth into new saving and capital gains since the source of the
change in wealth may affect asset choice.

615

Stephen M. Goldfeld

Table 15. Root Mean-SquaredErrors for Extrapolationswith

Wealth Variables, AlternativeSample Periods Endingwith 1961 through
1971
Incomeand changein wealth
Ex post

Wealthonly
Ex post

End
pointa

Full

Fourquarter

Sample
period

Full

Fourquarter

Sample
period

1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971

9.84
5.89
2.68
1.53
1.67
1.95
2.35
1.67
1.81
1.81
1.62

2.75
0.96
0.84
0.72
1.10
0.41
0.94
1.04
2.35
0.90
1.62

0.49
0.85
0.95
0.99
0.94
1.02
1.03
1.00
1.03
1.08
1.10

17.57
16.38
11.00
8.21
5.47
7.10
5.77
3.49
4.67
1.57
1.30

3.53
2.42
1.17
0.65
0.73
2.36
2.16
0.50
4.57
0.81
1.30

0.93
1.40
1.95
2.16
2.24
2.19
2.32
2.34
2.38
2.50
2.53

Source: Equations 1 and 3 of Table 14, dynamically simulated.

a. See Table 2, note a.

reflectingthe resultsof dynamicsimulationsof the specificationsembodied

in equations(1) and(3) of Table14.Theresultsof usingthe wealthvariable
alone in level form are distinctlyinferiorto the originalequation(4) (see
Table2) both for extrapolationsand withinthe sampleperiod.Whenthe
variablereflectingchangein wealthis addedin equation(4), the resultsare
somewhatmoremixed,but the originalequationis still to be preferredon
its ex post performance.
On balance,then, at least for quarterlydata, use of an incomevariable
in the demand-for-money
equationseems eminentlysensible.A variable
reflectingthe changein wealthslightlyimprovesthe explanatorypowerof
the equationbut slightlyworsensits predictiveability.63
63. This conclusion should be temperedfor two reasons.For one, the quality of the
quarterlynet worth data is suspect. In addition, as defined,net worth includescapital
gains on equitiesthat should probablybe excludedor at least separatedout. Along these
lines Bosworthand Duesenberryhave successfullyused a variabledefinedas net acquisition of financialassets in equationsexplaininghouseholdliquid assets of varioustypes.
See BarryBosworthand James Duesenberry,"A Flow-of-FundsModel and Its Implications,"in FederalReserveBank of Boston, Proceedingsof the MonetaryConference,
1973 (FRBB, 1973).

tn

tn

(N

ON

ti

IR
t

Iti

IRT

o
Q

. 0

lto

r"

%O
*0

*o

~~~~~~~~~~~~~~

;:

t i

~~~C

;~~~~~
{ |

gP,

a~~~~~C

0 -

~c

c;

o.
'

>

E~~N00

*;~~~~~~~~~C
t}

m W m

tFoN=NWX

~N

el)e

"TC

en
***

O "t O)

t
o

.wl

Xo

m 4i%

C'

So

c N c'1

C) WI

oN^@o
tORXmW

~~~~~~~~~~~~~~~~~~~~~~~0

Oo~~~~~~7

N R m WOWWoto

StephenM. Goldfeld

617

SomeEconometricIssues
The previoustwo sectionsconsidereda numberof basicproblemsin the
specificationof the money demandfunction.The presentsection focuses
on a somewhatnarrowerandmoretechnicalset of issuesand considersin
sequencequestions(7) through(10) posed at the beginningof the paper.
ALTERNATIVE INTEREST RATES

The originaldebateoverinterestratesinitiallycenteredon whetherany

interestrate mattered.In more recent years, with this questionsettled,
discussionhas turnedto the appropriaterate or rates to includein the
money demandfunction.64The majordisputehas concernedshort rates
(on commercialpaper,Treasurybills, and the like) versuslonger rates65
(on corporatebonds, U.S. governmentobligations,or even equities),althoughthe importanceof varioustypes of savingdepositrates(at savings
and loan associations,mutualsavingsbanks,and commercialbanks)has
also been an issue.Most researchersdo not confrontthe questiondirectly,
however;they simplyuse whateverset of interestratesis consistentwith
the rationaleofferedfor the demandfor money. In the context of the
transactionsapproachsuch a set typicallymeans somethinglike the two
rates (RCP and RTD) used in equation(4) but there are other choices.
Table 16 reportsthe resultsof some alternativespecifications.
The ratesconsidered,in additionto RCP and RTD, werethe Treasury
bill rate,RTB,a weighted-average
savingrate,RAVG(combiningRTD, a
savingsand loan, and a mutualsavingsbankrate),and, for completeness,
the corporatebond rate(RCB).Generallyspeaking,RCPandRTBappear
as do RTDandRAVG.Althoughthe resultsarenot shown
interchangeable
in the table,a weightedaverageof the savingsandloan andmutualsavings
bank rates,and a separaterate on certificatesof deposit,were also tried.
64. See, for example, Laidler, Demandfor Money; Tong H. Lee, "AlternativeInterestRates and the Demandfor Money: The EmpiricalEvidence,"AmericanEconomic
Review,Vol. 57 (December 1967), pp. 1168-81; and comments by Harvey Galper and
Michael J. Hamburger,AmericanEconomicReview, Vol. 59 (June 1969), pp. 401-07,
and 407-12, respectively.
65. Hamburger,in "Demand for Money in 1971," has been one main proponentof
the longerrates,while Laidler,in DemandforMoney,has suggestedthat the appropriate
rate may depend on the definitionof money.

00

%D
en

N
re4

o
C;

^
q

o
X

^
?

w0

cs

D
oy

X
^~~~~~OC

ooO

o~~~C

oy

N
oy
oy14

IW

N _

ON m o

oto,to^

t,C'

-C

E~~~~C

It

C;

*~~~~~%
IQ 00

s t.g

V-

-C

-0~

ci

St b.o
0F
o

'4-

4-i
Q

.>

ko

o
.

X,
_

oo14

X~~~~~~~~~~~~~~~,
.E

P-"jXg

8
io
g~~C

c
._;

StephenM. Goldfeld

619

The formerworked,althoughit did not do as well as RTD or RA VG and

wasnot significantwhenusedin conjunctionwithRTD. Thecertificaterate
was quiteinsignificant.
Table 16 makes clear that includinga saving deposit rate of any sort
increasesthe speed of adjustment,from much less than 10 percentper
quarterto about20 percent.Thecorporatebondratedoes not worknearly
as well as these others,neverachievingstatisticalsignificanceand in some
unreportedcombinationsactuallyyieldinga positivecoefficient.
On balancethen, the specificationin (4) seemsto workabout as well as
any other. One potentialproblemwith this for extrapolationpurposesis
that RTD (or RA VG) has becomemoredifficultto measurein view of the
widespreadimportanceof consumer-typecertificates.
TIME UNIT OF MEASUREMENT

The quarterlymoneyseriesused thus far was obtainedby averagingthe

officiallyreportedmonthlydatafor the threemonthsof the quarter.These
monthlydata are in turn producedby averagingdaily data. Gibson has
the behavior
arguedthat this procedureis the properway of characterizing
of the moneyseriesovera quarter,and that it providesa reasonablecorrespondencewith the GNP data from the nationalincomeaccounts.66But
the moneystockserieshas beenmeasuredin manyotherwaysin empirical
researchon the demandfor money: by an averageof two months'data
centeredon the end of the quarter,by data for the last month of the
quarter,and by end-of-quarterpoint estimates(for example,from call
reportdata).
Wouldsubstitutingone of these definitionschangeany of the basicresults?Thisquestionis of particularinterest,becauseGibsonhas foundthat,
for the early postwarperiod, the time unit of measurementmay have a
pronouncedimpacton the coefficientof the speedof adjustment.67
The resultsobtainedfromestimatingequation(4) with each of the three
alternativemeasuresjust noted are reportedin Table 17. (The point estimateof the moneystockis takenfromthe flow of fundsdata,whichwillbe
utilizedmoreextensivelybelow.)
66. W. E. Gibson, "Demandand Supply Functionsfor Money in the United States:
Theoryand Measurement,"Econometrica,Vol. 40 (March 1972), pp. 361-70.
67. Ibid.

O.

PC

(f*e0~
E

0O

-- '.
o~*

Eo40

00

O~~~O\

cq

1~) 06

i,

00)

r,

,O.

0r

.
0~~~~~~~~~~~~~~~~~~~~~~~~~u

U,

Ot.

'.4-

X~~~~~~~

,
i

Su

0k

'

B~~
J0I

StephenM. Goldfeld

621

The three resultsreportedin Table 17 and the results of the original

equation(4) areobviouslyall quitesimilarto one another.68Consequently,
Gibson'sfindingthat use of quarterlyaveragedata led to a much more
rapidspeed of adjustmentis not borne out when the sample period is
extendedto the later part of the postwarperiod.
SERIAL CORRELATION AND SIMULTANEITY

All of the estimatesreportedso far havebeen obtainedby applyingthe

Cochrane-Orcutt
techniquefor correctionof serialcorrelationin conjunction with ordinaryleast squares.Thus,problemsof simultaneousequation
biashavebeen ignored.In the absenceof a completemodel,the choice of
meansto carry out simultaneousequationestimationis somewhatarbitrary.Moreover,a casual interpretationof the evidencesuggestedthat
simultaneitybias was not likelyto be importantbut that serialcorrelation
was.Theseroughimpressionswerecheckedby choosinga plausibleset of
instrumentsand reestimatingequation (4) by ordinaryleast squares
(OLSQ)andby two-stageleastsquares,both corrected(TSCORC)andnot
corrected(TSLS)for serialcorrelation.69
The resultsarereportedin Table
18.
Theresultsobtainedby OLSQand TSLSarefairlysimilarto each other
and to the estimatesgiven in equation (4). Correctingfor both simultaneityandserialcorrelation(TSCORC)yieldsa considerablyfasterspeed
of adjustmentbut the long-runelasticitiesare essentiallythe same as in
equation(4).70 To see whetherthis fasterspeed of adjustmentwould improvethe trackingabilityof the equation,I performedthe standardset of
dynamicsimulationsdescribedabove. I also computedthese simulations
based on the estimatesobtainedby OLSQ, with the resultsreportedin
Table19.
68. The only differenceof any note is that the residualsbased on the point estimate
definitiondo not seem to be seriallycorrelated(all the other estimatesof p are statistically significantlydifferentfrom zero).
69. In carryingout the two-stageprocedures,income and both interestrate variables
were treatedas endogenous.For TSLS the instrumentsused were population, the discountrate,stateand local governmentspending,and the laggedmoney stock. To ensure
consistencyfor TSCORC,four additionalinstrumentswere used-income lagged, both
interestrateslagged, and money lagged twice.
70. The original estimates of elasticitiesfor income and for interest rates on time
depositsand commercialpaper were 0.68, 0.16, and 0.07, respectively,while they are
0.66, 0.18, and 0.04 for the TSCORCequation.

622

Brookings Papers on Economic Activity, 3:1973

Table 19. Root Mean-SquaredErrors for Ordinary and Two-Stage

CorrectedLeast Squares Estimating Techniques,AlternativeSample
Periods Ending with 1961 through 1971
Ordinaryleast squares

Two-stagecorrectedleast squares

Ex post

Ex post

End

Sample

Four-

Sample

Four-

points'

period

quarter

Full

period

quarter

Full

1.12
1.19
1.35
1.46
1.44
1.50
1.47
1.46
1.46
1.48
1.52

1.33
1.98
1.86
0.90
1.92
0.60
1.85
1.36
1.48
1.21
0.60

6.37
5.39
2.12
1.61
1.48
2.33
2.61
1.84
1.77
0.91
0.60

0.81
0.86
0.91
0.92
0.85
1.11
0.99
0.99
1.03
1.11
1.11

1.87
1.21
0.45
1.54
2.16
1.20
1.44
0.79
2.18
1.08
1.38

3.32
1.57
2.61
3.25
1.95
2.93
2.22
1.32
1.61
1.44
1.38

1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971

Source: Derived from dynamic simulations using the techniques of Table 18.
a. See Table 2, note a.

For OLSQ,the within-sampleRMSEs are all about 40 percentlarger

thanthe correspondingresultsin Table2, thus pointingup the benefitsof
correctingfor serialcorrelation.71
Theex post resultsalsofavorthe original
estimates,butby a smallermargin.In six out of the elevencases,the withinsampleresultswith TSCORCare actuallybetterthan the original.However,the ex post extrapolationsdistinctlyfavor the originalestimateson
balance.
An alternativespecificationof the money demandfunctionalso sheds
somelight on the simultaneityquestion.The moneydemandfunctioncan
be invertedto put an interestrate on the left-handside, relegatingthe
money variableto the right-handside.72Among the interestrates used
71. A comparisonof equation(4) to the OLSQ result in Table 18 indicatesthat the
standarderrors tend to be understatedif serial correlation is ignored. The original
standarderrorsare themselvessomewhat understatedsince I have not accounted for
the presenceof the laggeddependentvariable.See J. P. Cooper,"AsymptoticCovariance
Matrixof Proceduresfor LinearRegressionin the Presenceof First-OrderAutoregressive
Disturbances,"Econometrica,Vol. 40 (March 1972),pp. 305-10. Cooper presentssome
formulasfor makingthe appropriatecorrection,which for equation(4) yields roughlya
30 percentincreasein all standarderrors.
72. This procedurewas used in Poole, "WhitherMoney Demand?"

623

Stephen M. Goldfeld

above,the commercialpaperrateseemsto be the morenaturalcandidate.

The result of invertingequation (4) and estimatingby ordinaryleast
squarescorrectedfor serialcorrelationis as follows:
(18) ln RCP = -6.75-

(1.1)

9.128 ln m + 7.319 lnm_

(3.5)

(3.0)
+ 2.761 ln y
(2.5)

0.280 In RTD.
(0.8)

R2 = 0.928; standard error = 0.116; p = 0.82.

If one reinverts(18), an equationratherdifferentfrom (4) results.73In

particular,the elasticitiesare 1.53for income,0.55 for RCP, and 0.15 for
RTD.Onlythe last estimateis evenclose to whatwas previouslyobtained.
Surprisinglyenough, given the results of Table 18, the source of the
discrepancyturns out to be the existenceof ratherstrong simultaneous
equationsbiasin (18).Thatthis is the casecan be seenby reestimating(18)
by TSCORC,whichyields74
(19) ln RCP = 17.77 - 30.338 ln m + 12.329 In mi

(2.1)

(4.7)

(2.2)
+ 12.197 n y

(3.7)
R2

3.135 In RTD.

(3.0)

= 0.853; standard error = 0.165; p = 0.54.

The resultsof (19) are dramaticallydifferentfrom (18) and in fact much

morein line with (4). The impliedelasticitiesare 0.68 for income,0.17 for
RTD, and 0.06 for RCP,virtuallyidenticalto those obtainedinitially.The
majordifferencebetween(19) and (4) is that the speed of adjustmentof
(19) is considerablyfaster-roughly 60 percentper quarter,whichis even
fasterthan the correspondingresultin Table 18.
On balance,it appearsimportantto correctfor serialcorrelationand
probablyfor simultaneousequationsbias as well, especiallyif an interest
rateis the dependentvariable.Onevirtueof the TSCORCestimatesis that
they producesubstantiallyfaster speeds of adjustment.On a partialadjustmentview,this resultseemsdesirable,but it did not appearparticularly
to improvethe trackingabilityof the equation.It might,however,make a
greaterdifferencein the contextof a completeeconometricmodel.Finally,
73. Poole, ibid., found a similardiscrepancy.
74. The instrumentsare the same as describedin note 69 above.

Brookings Papers on Economic Activity, 3:1973

624

while it might have been desirableto use simultaneousequationstechniquesthroughoutthis analysis,the generallycomparableperformanceof

the originaland TSCORCestimatessuggeststhat the resultswouldnot be
qualitativelyaffectedby such a procedure.
HOMOGENEITY WITH RESPECT TO PRICES AND POPULATION

In the demandfunctionsconsideredthroughoutthis paperreal money

holdingshavebeenassumedto be a functionof realGNP. Althoughsome
writershaveusednominalmagnitudes,the specificationin realtermsis the
most commonformusedin empiricalresearchandis the one suggestedby
economictheory.For example,underthe simplestBaumol-Tobinformulation, moneyholdingsare given by
(20)

(k Y/2r),

wherek is a fixedchargepertransaction.Dividingboth sidesof (20)by the

pricelevel yields
(21)

-k
[p

p/2r],
1

or
(22)

(k'y/2r) is

wherek' is a transactionscost in realterms.Assumingk' is constantyields

the type of specificationemployedin this paper-that is, an equationof the
form
(23)

lnm=a+blny+clnr.

While(20) impliesthat the appropriatespecificationis in real terms,it

saysless aboutwhetherdeflationby populationis required.Indeed,strictly
speaking,one cannot aggregate(20) in any simpleway. Rather,the distributionof income needs to be taken into account,which suggeststhat
some featuresof the incomedistributionmightbe importantvariablesin
the moneydemandfunctionandthatin the aggregateeitherrealincomeor
realincomepercapitamaynot be strictlyappropriatevariables.However,
in the simplifiedsituationin whicheach individualhas the same income,
aggregationof (20) is possible,andthatcasedoesimplythatrealpercapita
moneyholdingsbecomea functionof realpercapitaincome.

StephenM. Goldfeld

625

As an empiricalmatter,the appropriatenessof each type of deflation

can be tested simply.For pricesone should estimatean equationof the
form
(24)

ln m=a

+ b ln y + c ln r + dln P,

and test the hypothesisthat d = 0. For population,one estimates

Inm = a+ blny+

(25)

clnr+

dln(POP).

If per capita deflationis appropriateone should be able to accept the

hypothesisthat d = -1b and that d is significantlydifferentfrom zero.
It shouldbe notedthat theselattertestsignorethe problemof incomedistributionand simplycomparethe merits of two approximations-using
realincome or realincomeper capita.
Versionsof equations(24) and (25) based on the detailedspecification
of equation(4) are reportedbelow.
(24') ln m

0.272 + 0.193 ln y
(1.6)
(5.3)

0.019 In RCP
(5.9)

0.045 ln RTD
(3.8)

+ 0.717ln m_- + 0.0017ln P

(11.0)
(0.008)
(25') ln m = 0.820 + 0.222 ln y

(1.8)

(5.1)

0.019 ln RCP

(6.2)
+ 0.707ln m,
(11.3)

0.033 ln RTD

(2.3)
-

0.133 ln (POP).
(1.2)

In (24')the coefficientof ln P is insignificantlydifferentfromzeroso that

one cannotrejectthe hypothesisof unitarypriceelasticity.75In equation
(25'), the coefficientof populationis insignificantlydifferentfrom zero.
Consequently,unlikedeflationby the pricelevel, deflationby population
does not seemto be calledfor.76
75. There is an alternativebut not quite identicaltest which involves estimating(24)
in nominal terms and testing whether = 1- . When this was done, the hypothesis
was accepted,confirmingthe appropriatenessof deflatingby the price level.
76. In addition to statisticalinsignificanceof the coefficientof population, one can
rejectthe hypothesisthat its coefficientis equal to unity less the coefficientof y and the
coefficientof m-1. This latter test is the analog of the test for d = 1 - ? referredto in
the text.

626

Brookings Papers on Economic Activity, 3:1973

Disaggregation Using Flow of Funds Data

A numberof resultsreportedin the last two sectionssuggestedthe desirabilityof greaterdisaggregationof moneyholdingsby type of asset(such
as currencyanddemanddeposits).I shallnow exploredisaggregationwith
respectto typeof holder,usingflow of fundsdata,compiledby the Federal
Reserve,that disaggregatemoneyholdingsinto the followingbroadcategories:households;business;state and local governments;financialsectors; rest of the world;and mail float.
Ideally,each of these componentsshouldbe analyzedin the context of
a completemodel of the determinationof assets and liabilitiesfor each
type of holder,so as to yield a clearpictureof the appropriateexplanatory
variablesand permit systematicuse of balance sheet constraints.This,
however,is a task for an armyof econometricians(one has alreadybeen
mobilized,in fact). Withinthe scope of this paper,it is possiblemerelyto
explore component money holdings with some rough and ready adhockery.
Thenatureof the ventureis clarifiedby the basicdataon moneyholdings
at the end of 1972:77

Sector
Business(includingfloat)
Household
State and local government
Financial
Rest of the world
All sectors

Dollars
(billions)

Percent
of
total

Percent
change,
1952-72

72.3
156.5
14.6
17.0
7.8

27.0
58.3
5.4
6.3
2.9

36.6
152.3
102.8
151.5
309.0

268.3

100.0

105.0

At thattime, 15percentof moneyholdingswereaccountedfor by groups

other than businessand households;for these groupsthe variablesconventionallyused in money demandequationsmay not be appropriate.
Furthermore,the compositionby sectorhas changedgreatlyin the past
77. Source: Board of Governorsof the FederalReserve System, Flow of FundsAccounts, 1945-1972 (August 1973) (see appendixat the end of this paper). Figures are
based on data beforerounding.

627

Stephen M. Goldfeld

twentyyears.In particular,the shareof businessholdingsof money has

declinedsteadilyfrom 40 percentin 1952,whilehouseholdshave steadily
increasedtheir share from 48 percent.The remainingcomponentshave
risenin the aggregatebut have also exhibitedsubstantialfluctuations.
Moneyholdingsof the differentsectorshavealso movedin diverseways
in the short run as evidencedby the very low and frequentlynegative
simplecorrelationcoefficientsfor the seasonallyadjustedquarterlyflows
of the differentsectorsoverthe period 1952:2to 1972:4:

Business

Household

-0.11

State and
local
government

-0.03
-0.14

Financial

-0.06
0.25
-0.19

Rest
of the
world

-0.09
-0.05
0.09

0.06
This result again suggeststhat disaggregatingby holder should pay off.
Disaggregationwill not be a simplematter,however.The firstproblem
lies in the qualityof the data.In recentyearsthe FederalReservehas conducteda surveyon the ownershipof demanddepositsby type of holder.78
Attemptsto reconcilethesedata with the flow of fundsdata haverevealed
a numberof discrepancies
that raiseseriousquestionsaboutthe qualityof
the flow of fundsdata in generaland the allocationbetweenbusinessand
householdsin particular.Judgingby the survey,the flow of funds data
understatebusinessholdingsand overstatehouseholdholdingsof money.
Eventakingthe dataat facevalue,a numberof otherclueswarnthatthe
analysisof sectoralmoney holdingsmay be complicated.Whenthe total
percentagegrowthin the variouscomponentsfromthe end of 1952to the
end of 1972 is comparedwith the growth in the transactionsvariables
relevantfor each sector, some markeddifferencesemerge.For example,
businesstransactionsarenearlythreetimestheir 1972level(if measuredby
businesssales) or three and one-halftimes (if measuredby businessoutput), but businessmoney holdingshave increasedby less than one-half.
78. For a good descriptionof the surveyand a reconciliationwith both the conventional money stock data and the flow of funds data, see "Surveyof Demand Deposit
Ownership,"FederalReserveBulletin,Vol. 57 (June 1971), pp. 456-67.

628

Brookings Papers on Economic Activity, 3:1973

Table 20. Coefficientsfor Household Demand for Moneya

Interest rate
Lagged
money
0.736

Time Commercial
deposits
paper
-0.055

-0.025

GNP
0.312

Consumer
expendi- Change in
tures
net worth

R2

Standard
error

...

...

0.991

0.013

-0.20

...

0.230
(2.0)

0.992

0.012

-0.27

(11.5)

(3.3)

(4.5)

(4.4)

0.784
(12.6)

-0.044
(2.7)

-0.017
(2.7)

0.251
(3.6)

0.796
(12.3)

-0.045
(2.5)

-0.021
(3.7)

...

0.249
(3.5)

...

0.992

0.013

-0 15

0.844
(13.8)

-0.033
(2.0)

-0.013
(2.1)

...

0.187
(3.7)

0.260
(2.2)

0.993

0.013

-0.24

Sources: Based on flow of funds data from the Board of Governors of the Federal Reserve System. See
the appendix for specific information on data used.
a. The period of fit is 1952:2 to 1972:4. The equations are estimated in logarithmic form by ordinary
least squares, with a correction for serial correlation, and use the Koyck lag specification.

Similarly,state and local governmentspendingis ten times what it was

twentyyearsearlier,whilemoneyholdingshavejust doubled.For households, transactionsas measuredby consumptionhave quadrupledand
money holdingsare two and one-halftimes the earlierlevel. At the very
leastthesenumberssuggestthat "income"elasticitiesaredramaticallydifferentacrosssectors.In principle,allowingfor suchdifferencesis one of the
virtuesof disaggregating.More importantly,however,this evidencesuggests that a simpletransactionsmodel(especiallyif couchedin realterms)
will have a hardtime explainingmoneyholdingby businessand by state
and local governments.With these caveats,I turnto some results.
The sampleperiodfor all the estimatesto be presentedis identicalto the
one used above-1952:2 to 1972:4.All estimateswere obtainedby ordinaryleastsquareswitha correctionforserialcorrelation,althoughthis was
not muchof a problem.The equationswereestimatedin logarithmicform
and the lag specificationwas limitedto the Koyck form.79
HOUSEHOLD SECTOR

The householdsectorhas the largestshareof total moneyholdingsand

in manywaysis the easiestto explain.Essentiallythe sametype of specificationusedfor aggregatemoneydemandworksequallywellfor the household sector.Somerepresentativeresultsare containedin Table20.
The firsteciuationis identicalin specificationto eciuation(4). Both in79. Estimationof Almon distributedlags would have requiredsacrificinga substantial numberof initialobservationsand would havemadecomparisonswith earlierresults
difficult.

StephenM. Goldfeld

629

conmeand interestelasticitiesfor the householdsectorexceedthose found

for total money holdings,with the long-runincome elasticityexceeding
unity. The change in net worth, as before, also achievesstatisticalsignificance.An equallysensibleequationis obtainedif one substitutesconsumptionfor GNP as the scalevariable.80
On the whole,the resultsfor the
householdsectorare reasonable.
BUSINESSSECTOR

The businesssectorcomesnext in the size orderof moneyholdingsbut

here, as anticipated,I met with considerablyless success.One typically
unsatisfactoryresultfollows:
(26) In mb = 0.359 + 0.010 In SALE+ 0.905ln mb,
(1.4) (0.5)
(18.9)

0.016In RCP.
(2.3)

R2 = 0.948; standarderror = 0.014; p = 0.02.

This equationis in real terms, deflatedby the businessdeflatorof the

nationalincomeaccounts;SALEis manufacturingand trade sales.
Unfortunately,the equationproducesa transactionsvariablethat is not
significantand a speedof adjustmentthat is unreasonablyslow. A number
of attemptswere made to improve this equation. In particular,I tried
a businessGNP measure,a certificateof depositrate, a measureof cash
flow, and inventoryinvestment,but none of these variablesachievedstatistical significance.I also tried a linearfunctionalform but this did not
help either.On balanceI can only concludeeitherthat the qualityof the
data makes this a futile exerciseor that considerablymore ingenuityis
neededto explainaggregatebusinessmoneyholdings.8'
FINANCIAL SECTOR

Next in importancein volume of money holdingscomes the financial

sector(savingsand loan associations,mutualsavingsbanks, and so on).
80. In this instance the consumptiondeflator is used to deflate all nominal magnitudes.
81. One directionfor possible improvementwould be to'integratethe money hold,ingand trade credit variables.Another problemof unknownproportionsis createdby
the absenceof any reliableinformationon compensatingbalances.Developmentof such
informationwould also be a step in the right direction. Finally, the mail float item is
includedwith the businesssector.Whilethis is approximatelycorrect,some furtherwork
could be done. See the articlecited-innote 78 above.

630

Brookings Papers on Economic Activity, 3:1973

Table 21. Coefficientsfor Financial Sector Demand for Moneya

Lagged
money

Treasury
bill rate

0.698

-0.014

Deposits
0.154

Change Proxy for

in
outflow of
deposits depositsb
0.514

(8.5)

(1.6)

(4.0)

(1.6)

0.659
(8.1)

-0.016
(1.9)

0.179
(4.5)

1.429
(2.9)

R2

Standard
error

...

0.995

0.018

0.12

0.066
(2.2)

0.995

0.018

0.04

a. See sources and note for Table 20. The equations are in undeflated form.
b. Ratio of the Treasury bill rate to the saving deposit rate after 1968:3, and zero before.

For this sector,the appropriatescale variableis a measureof depositactivity.The level of depositsandthe changein depositsusedjointlyworked
relativelywell.Two suchequationsusingthesevariablesaregivenin Table
21. The firstemploysthese variablesin conjunctionwith the Treasurybill
rate whilethe secondadds a variabledesignedto capturethe anticipated
Thisvariableis definedas the
outflowof depositsdueto disintermediation.
ratio of the Treasurybill rate to the savingdepositrate after 1968:3and
zero before.The higherthis variablethe morefinancialinstitutionsexpect
andthe moreliquidthey therefore
to lose fundsthroughdisintermediation
wish to be. This variableobtainsthe expectedpositivesign and is statisticallysignificantat the 5 percentlevel.Onthe whole,then,moneyholdings
by the financialsectorappearto lend themselvesto a reasonablystraightforwardexplanation.82

STATE AND LOCAL GOVERNMENT SECTOR

The finalequationto be consideredis for the state andlocal government

sector. From initial examinationof the data, this was expectedto be a
troublesomesector,andindeedit was. As with the businesssector,a numberof specificationsweretried,includingseveralinterestratevariablesand
a budgetsurplusvariable.No fully satisfactoryequationeveremerged.A
typicalequationis
(27) In msl = 0.092 + 0.946 ln msL1 + 0.011ln gsl
(0.4)
(0.6) (22.1)

R2 = 0.867; standard error = 0.047; p = -0.10.

82. The equationsin Table 21 were run in undeflatedform.

0.022 In RCP,
(1.0)

StephlenM. Goldfeld

631

Table 22. Root Mean-Squared Errors for Aggregate and

ComponentMoney Holding Equations, AlternativeSample
Periods Ending with 1961 through 1971
Aggregate equation

Component equations

Ex post

Ex post

End
point"

Full

Fourquarter

Withinsample

Full

Fourquarter

Withinsample

1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971

7.21
4.54
2.33
3.70
3.50
2.15
2.23
2.82
2.94
3.83
3.30

2.07
1.28
2.04
0.67
2.93
1.08
1.80
1.59
1.86
1.99
3.30

0.73
0.87
0.95
1.06
0.99
1.15
1.09
1.12
1.13
1.17
1.18

18.71
17.41
14.95
3.59
2.96
3.44
4.51
6.77
5.11
6.29
3.84

2.60
1.03
3.03
1.03
1.15
2.03
1.40
1.98
1.54
3.20
3.84

1.19
1.20
1.32
1.66
1.60
1.65
1.55
1.51
1.69
1.69
1.89

Sources: See sources and note for Table 20. The aggregate equation corresponds to the aggregate flow of
funds equation in Table 17. The component equations used are (26) for the business sector, (27) for the state
and local governmentsector, the first equation in Table 21 for the financial sector, and the fourth in Table 20
for the household sector. Money holdings for the rest of the world were considered exogenous.
a. See Table 2, note a.

wheregsl is state and local governmentexpenditures.83

Quite evidentlyI
am unableto provideanythingclose to a satisfactoryexplanationfor this
sector.
OVERVIEW

Takenas a whole, the batting averageon disaggregationby holder is

.500. Two of the four categories,householdsand the financialsector,are
reasonablywellexplained;two others,businessandstateandlocal government,arenot. Onewouldexpectthatthe firsttwo wouldbehavereasonably
wellin the simulationexercisesI haveperformed,and-sparing the reader
the details-this wasthe case.The remainingtwo sectors,not surprisingly,
didrelativelypoorly.Table22 summarizes
the performancein the aggregate
errorsfor total moneyholdings.The
withthe relevantroot mean-squared
83. Equation(27) is in real terms, all nominal variableshaving been deflatedby the
state and local expendituredeflator. Estimation without deflation produced slightly
betterbut still unsatisfactoryresults.

632

Brookings Papers on Economic Activity, 3:1973

left half of the table correspondsto the aggregateflow of fundsequation

reportedin Table 17, while the right half extrapolatesaggregatemoney
holdingsbasedon the equationsfortheindividualcomponents.84
Whilethe
aggregateequationdoes not uniformlydominatethe individualequations,
one plainlywouldnot forecasttotalmoneyholdingsby separateuse of this
particularset of componentequations.Nevertheless,despitethistentatively
pessimisticfinding,theresultshold enoughpromiseto warrantwithholding
a finalverdicton this issue.Furtherresearchalong theselines is clearlyin
order.

Remarks
Concluding
In the processof sequentiallyexaminingeach of the questionsset forth
at the beginningof this paper,a considerableamountof informationhas
emergedconcerningthe nature of the demandfor money. This section
enumeratesthe highlightsof the findings,attemptsto illuminatethem by
examiningvelocity,both actualand simulated,undera varietyof assumptions, and brieflyassessesthe demandfor moneythrough1974.
Perhapsmost interestingis the apparentsturdinessof a quite conventional formulationof the money demandfunction,howeverscrutinized.
Moreparticularly,sucha functionyieldssensibleinterestandincomeelasticities.Theincomeelasticityappearsto be significantlyless thanunityand
canbe pinneddownreasonablywellon the basisof quarterlydata.In addition, the conventionalequationexhibitsno markedinstabilities,in either
the shortrun or the long run. Finally,the conventionalequationyields a
reasonablespeedof adjustmentto changesin incomeor interestrates,with
patternsand magnitudesof adjustmentthat are generallysimilarin the
Koyck and Almon specifications.
Whilethe conventionalequationperformswell,it is neverthelesspossible
to improveon it in a numberof ways.In the firstinstancedisaggregationof
M1 into currencyand demanddeposits appearsdesirablefrom both a
structuraland a forecastingpoint of view.Aggregationto the level of M2,
84. The equationsused in these calculationswere (26) and (27), the first equation in
Table 21, and the fourthin Table 20. Money holdingsof the rest of the worldwere taken
as exogenous. Furthermore,the results were made comparable with those reported
earlier by reinflatingthe component forecasts(where necessary)and then deflating by
the GNP deflator.

633

StephenM. Goldfeld

Furthermore,the additionof a
however,is definitelycounterproductive.
numberof variablesappearsto improvethe performanceof the standard
formulation.Theseincludethe changein wealthand, possibly,a variable
inflationexpectations.Onthe otherhand,substitutionof wealth
measuring
for income imposes a markeddeteriorationin the performanceof the
equation.
Finally,whilethe diversesectoralpatternof movementsin moneyholdings exhibitedby the flow of funds data impliedsome payoff to greater
disaggregation,
effortsin this directionwereonly partiallysuccessful.The
tentativenatureof the resultssuggeststhat this remainsan open issue.
THE BEHAVIOR OF VELOCITY

An empiricalmoney demandfunctionhas implicationsabout the behaviorof the incomevelocityof money,v = y/m. One importantimplication, long debatedby economists,concernsthe sensitivityof v to interest
ratechanges,whichis simplythe otherside of the coin of the debateconcerningthe interestelasticityof the demandfor money. The resultshere
havereconfirmed
the importanceof interestratevariablesin explainingthe
demandfor money,andtheirimplicationsfor the behaviorof velocityhelp
to put theirimportancein perspective.
The basicmoney demandfunctionestimatedabove can be written(in
nonlogarithmic
form)as
m = Ayarb,
whichyields
v = y/m = yl-a/Arb

Thisequationimpliesthat, with a constantinterestrate, velocitywill increaseat the fraction(1 - a) of the growthrate of y. With a valueof a of
about0.7, annualgrowthin y of 4 percentwould lead to a 1.2 percent
growthin v. Since 1952,v has actuallyincreasedat about21/2percentper
year;the excessover 1.2 reflectsthe upwardtrendin interestrates.85
Whilevelocityhas trendedupward,its path has hardlybeen steady,as
theserieslabeled"actualv"in Figure2 readilydemonstratesfor the period
1968:1to 1973:2.To assessthe sensitivityof velocityto alternativepaths
85. Velocity(definedon the basis of MI) has risen from about 2% in 1952 to nearly
5 in 1973.

634

BrookingsPaperson EconomicActivity,3:1973

Figure 2. Actual and Simulated Income Velocity of Money underFour

Assumptions,Quarterly, 1968-73
Velocity rate
4.9

101~~

Smooth y,
actual r.

4.8

4.7

Actiialy,

I/

;._.fiactual

.7

SmZSoothy,

4.6

r
~~~conistanit

.>
-i"v

4.5
/,,,-w

contstantr

//
/

-Actital

4.4

--

1968

1969

1970

1971

1972

Sources: See appendix.

for incomeandinterestrates,I dynamicallysimulateda versionof equation

(4) estimatedthrough1973:2with fouralternativepaths.Thesefourpaths
resultedfrom combiningtwo assumptionsfor interestrates with two assumptionsfor GNP. In particular,interestrates were either assumedto
takeon theirhistoricalvalues("actualr" in Figure2) or to remainconstant
at their 1967:4 values("constantr"). Similarly,for incomeI eitherused
actualvalues("actualy") or let it grow smoothly("smoothy") over the
period(endingup at the actualvaluein 1973:2). In eachcase the dynamic
simulationwas startedin 1968:1.
In comparisonwiththe serieslabeled"actualv,"the one labeled"actual

StephenM. Goldfeld

635

y, actual r" in Figure2 indicatesthe trackingabilityof the equationwith

historicaldata.Onthe whole,that abilityis reasonablysatisfactory,butthe
patternof errorsis interesting.86When actualvelocityis 'below that predictedby the equation,someinterestratesmustbe higherand GNP lower
than would otherwisebe the case, given the money supply. The precise
impacton interestrates and GNP dependson the relationshipof investmentandconsumerdemandto interestratesandincome-the shapeof the
IS curve,a set of vital issuesoutsidethe scope of this paper.Nonetheless,
to some extent, unusuallylow velocity is .a drag on aggregateeconomic
activity,whileunusuallyhigh velocityis a stimulus,so long as the Federal
Reservedoes not fully offsetthe surpriseby changingthe stock of money.
By this standard;the low valueof velocityearlyin 1969,a periodof exinfluence.On the otherhand,
cess demand,exertedsome anti-inflationary
thelow velocityreadingsin the secondandthirdquartersof 1971couldbe a
factorin the ratherweak startof the economicrecoveryafterthe 1969-70
recession. In 1971:4 and 1972:1, velocity swung sharply upwardand
crossedits predictedvalue;thatmovementmayhavereinforcedthe acceleration of economicactivity.
The series"smoothy, constantr" demonstrates,as indicatedabove,that
velocitywill steadilyincreasewith continuedgrowthin income.Thislatter
curvecan be comparedwiththe remainingtwo curves-"smooth y, actual
r" and "actualy, constantr"-to isolate the impactof fluctuationsin interestratesand income,respectively.The historicalmovementof interest
ratesproducesa strikinglydifferentpatternfromthe constantinterestrate
assumption.Similarly,the actualpatternof incomeyieldsa velocityseries
that is markedlydifferentfrom the steadygrowthassumption.
In short,velocitycan be extremelyvariablein the shortrun(quiteapart
fromthe residualsin the moneydemandfunction)and anypolicyprescription that does not take this into accountmay be very misleading.
SOME ILLUSTRATIVE PROJECTIONS

In earlierparts of this paper,I made extensiveuse of dynamicsimulations to examinethe forecastingperformanceof variousspecifications.In

86. The residuals depend on the starting point of the dynamic simulations, and
hence the patterndiscussedhere could be differentfor other startingpoints.

. E
M46)It

,0

|6
;~~~~~~~0
0a

tQ
o6)

'

'

>

Cs~~~~~%

E-

O mN

%
s^%6

637

StephenM. Goldfeld

view of the reasonablygood performance,what can be said about the

futurebehaviorof moneydemandin 1974?One sensibleway to approach
this problemis to take as given forecastsfor real GNP and for the GNP
deflatorand examinethe behaviorof money demandfor alternativepatternsof interestrates.For this purposeI chosethe forecastproducedby the
Michiganquarterlymodel,whichforeseesrealgrowthof about21/2percent
and priceinflationof 6 percentfor the year 1974.87
Table23 sets out the annualpercentagerates of growthof the money
stock(both in nominaland in realterms)consistentwith threealternative
patternsfor the commercialpaperrate.88The firstpanelshowsa moderate
declinein interestrates,the next a verymilddeclinein shortrates,andthe
last a moresubstantialdecline.It shouldbe emphasizedthat these arenot
forecastsof actual money growth, but ratherof the rates of monetary
expansionconsistentwiththe assumedvaluesfor interestrates.89Overthe
six quarterstakenas a wholethe threeinterestratepatternsimplynominal
monetarygrowthratesof from6 to 63/4percent,withnot muchquarter-toquartervariabilityafter 1973:3.90 While this findingin no way offers a
prescriptionfor monetarypolicy,it does suggestthat extremelylow nominal growthrates are not consistentwith any plausibleexpansionof the
economythrough1974.
87. See Saul H. Hymans and H. T. Shapiro,"The EconomicOutlook at Mid-Year"
(Universityof Michigan,August 1973; processed).This forecastprecededthe Arab oil
embargo.The quarterlypatternof changes(at annualrates)producedby the model is as
follows:
Quarter
Real GNP
GNP deflator

1973:3
5.0
7.4

1973:4
2.7
7.3

1974:1
1.8
6.2

1974:2
1.3
5.1

1974:3
1.6
4.9

1974:4
3.1
4.6

88. The resultsare basedon extrapolatinga versionof equation(4) estimatedthrough

1973:2where the time deposit rate is held at its level for 1973:3 throughall succeeding
quarters.While this is a convenient assumption,it is also in the nature of a forecast
consistentwith the assumedpatternsof behaviorfor the commercialpaperrate and the
impact of interestrate ceilings.
89. The Michiganforecast of interestrates is essentiallylike the first panel in Table
23. Consequently,any major deviationsfrom this patternwill not be consistent (as far
as the Michiganmodel is concerned)with the assumed priceand output behavior.
90. Extrapolationsbased on the equation includingthe priceexpectationterm show
variaroughlythe same growth over the periodas a whole but more quarter-to-quarter
bility. For example,the growthrates of money correspondingto panel 1 in the table (in
real terms)are as follows: -3.5; -0.8; 0.8; 1.9; 2.3; and 3.0.

638

Brookings Papers on Economic Activity, 3:1973

APPENDIX

Data Sources
used in this paperare in billions and are seasonally
adjusted.The flow data are at annualrates.The interestrate variablesare
in percentagepointsand arenot seasonallyadjusted.Grossnationalproductandrelatedvariablesarebasedon the July1973revisionsof the national
income accounts(publishedin the Surveyof CurrentBusiness)while the
flow of funds data are based on the August 1973 revisions.Although
readilyavailablein publishedsources,many of the seriesused wereactually taken from the data deck of the FederalReserve-MIT-Pennsylvania
(FMP)EconometricModel.This was generouslysuppliedby JaredEnzler
of the Boardof Governors,FederalReserveSystem.Theflowof fundsdata
neededfor this studyand helpfulcommentsabouttheiruse weresupplied
by StephenT. Taylor,also of the Boardof Governors.
ALL DOLLAR DATA

Currency, demand deposits, time deposits. Taken primarily from the

February1973 issue of the FederalReserveBulletin.The time deposits

seriesexcludeslarge negotiablecertificatesof deposit.Exceptas noted in
the text these weremeasuredas quarterlyaveragesof monthlydata.
Interestrates. The rateson Treasurybills, commercialpaper,corporate
bonds,time deposits,savingsand loan deposits,and mutualsavingsbank
shareswereall taken from the FMP deck.The latterthreevariableswere
combinedinto the averagedvariableusedin Table16by weightingthe individual rates. The weights summedto one and are proportionalto the
quantityof depositsassociatedwitheachratein the previousperiod.These
quantityvariableswerealso used directlyin Table21.
Price indexes.Unless indicated,the nominalmoney stock is put in real
termsby use of the implicitGNP deflator.The exceptionsare in the cases
whereconsumptionis the scalevariable,whenthe consumptiondeflatoris
used;in equation(26), whenthe businessoutputdeflatoris used;and (27),
wherethe implicitdeflatorfor stateandlocal governmentspendingis used.
Flowoffunds. Seasonallyadjustedquarterlyflowswerecumulated(both
forwardandbackward)startingfromthe 1970stockdata,to yieldadjusted
seriesfor the variousstocksused.

Commentsand
Discussion
James Duesenberry:StephenGoldfeld has written a very fine paper, a
thoroughpiece of workthat reallymoves us aheadin the field.I do have
severalcomments,thoughthey shouldnot be classifiedas criticisms.
First, I hope that the econometriciansamong us will note Goldfeld's
methodof choosingamongequations.He does not relyprimarilyon small
differencesin R2s or on the t-statisticsof additionalvariables;instead,he
comparesthe successin forecastingmoney demandof simulationsrun on
alternateequations.WhileI know of no formaltheorythat tells us how to
assesssuchevidence,I feelthatthe use of thistechniqueis one of the merits
of Goldfeld'swork.On the whole,Goldfeldhas gone aboutas far as possible in extractinginformationfrom this body of aggregatedata, short of
takingit down to the cellarand beatingit with a rubberhose. The next
majorstepsin researchin this areashouldprobablytry to incorporateinformationfrom sourcesother than time serieson the structureof money
demand,and to employBayesianmethodsto evaluatethe time seriesdata.
Withregardto the substanceof the paper,I was somewhatdisturbedby
the resultson the businessdemandfor money.The database usedin these
disaggregatedequationsis weak: seriousmeasurementerrorsarisein the
attemptto breakdown ownershipof demanddepositsand currencyinto
householdand businesscategories.Moreover,some confusionmay arise
fromthe effectsof compensatingbalances,whichare includedin the measurementof businessholdings.I thinkthis is an areain whichmicro-level
datamightbe usedto refinethe aggregateequation.However,the successof
the demandequationwith householddata leads one to believe that the
failureof the businessdemandequationmaybe dueto a basicdifferencein
the responseof businessdemandto changesin incomeand interestrates.
Thatdifferencemay show up in the resultsof the aggregateequationand
may accountfor the estimatedlong-runincome elasticityof moneyholdings of 0.68, againsta valuenearunityfor the householdsector.
639

640

Brookings Papers on Economic Activity, 3:1973

The difference between household and business responses may involve

threshold effects in businessmen'sdecisions to economize on cash balances.
Short-run interest rates may have to rise considerably in order to induce
businesses to incur the set-up costs, in employees and perhaps computer
facilities, of working to economize on cash balances. The extent of the
business response is also dependent on the size of the money holdings (and
therefore of the potential interest gain) of any firm. Similar threshold
effects may exist on the down side: once the initial overhead cost has been
incurred, businessmen may continue to hold smaller transactions balances
even when the interest rate falls slightly. The aggregate equation is looking
for prompt responses to changes in interest rates and it is likely to miss
some of these delayed business responses. It may hence understate the
interest elasticity of the demand for money, and, since the effects of interest
and income run in opposite directions, also understate the response to
changes in income.
I found the performance of the price expectations variable in the money
demand equation puzzling, since it appearsto contradict some of the results
I have obtained in measuring the effects of expected inflation on saving.
Goldfeld obtained significant coefficients for these variables in his equations, though a comparison of the abilities of the equations with and without this variable to simulate the demand for money in future periods resulted in a tie. The belief that price expectations influence the demand for
money is based on the assumption that, if prices are expected to rise, people
will trade their money holdings for physical assets. Except perhaps in the
most recent period, saving typically responds to inflationaryexpectations in
the opposite way: people tend to save more in anticipation of higher prices
in order to prevent a deterioration in their living standards.
Since Goldfeld referred to the Bosworth-Duesenberry model in his section on the use of wealth as a scale variable, I might discuss our results
briefly. We found that the net acquisition of financial assets (from the flow
of funds data) does help to explain portfolio behavior. This variable essentially represents the change in wealth due to accumulation without regard
to the change in the price of assets, and it might be substituted for the full
change-in-wealth variable, part of which is due to capital gains.
From a theoretical viewpoint, a consideration of cross-sections suggests
that wealth should play a role in the demand equation. People with a small
net worth face a liquidity constraint because much of their wealth is tied up

Stephen M. Goldfeld

641

in physical assets. After making their downpayments on homes, cars, and

so forth, they have little liquidity. Wealthier people, having taken care of
these needs, are in a better position to trade off between more and less
liquid financial assets. But the greater the collinearity between income and
wealth, the less important the choice between the two scale variables. Moreover, threshold effects may be involved, in that some increases in wealth,
such as those resulting from a rise in the stock market, may not affect the
liquidity actions of very wealthy people on the margin.
Finally, the success of the currency equation bothers me. Regressions of
the stock of currency on interest rates frequently get significant results, but
they do not explain the reasons for the large volume of currency holdings.
The denominations of these holdings are not those needed to run a news
stand. Much of large-denomination currency must be associated with
hoarding and illegal activities. Historically, the volume of currencyjumped
suddenly in the Second World War and again during the Korean War.
Then it remained flat until a renewed rise began in the early sixties; during
the fifties, the wartime accumulation of currency gradually came out into
the open. At any rate, I tend to be suspicious of dandy equations about currency-and this makes me suspicious of dandy equations about other
things, including aggregate money holdings.
As a final point, I would like to see a few results on first differences from
the simulations. Though the correction for serial correlation probably introduces technical problems here, the first differences could throw additional light on how precisely these equations are predicting money demand.
William Poole: This paper is a very useful study of a large number of the
unsettled issues in this field. Instead of paying all of the compliments that
Goldfeld deserves, I would like to focus my discussion on the few areas in
which I have reservations about the methodology.
First, with regard to the question of pinpointing the income elasticity of
the money stock, Goldfeld is quite right in citing my error in using R2s to
compare equations with different dependent variables. However, I do not
think that a comparison of the standard errors of the equations shown in
Table 1justifies his confidence in the accuracy of his income elasticity estimate. The equation with the elasticity of income constrained to 1.0 has a
standard error only about 6 percent greater than that of the equation with
the elasticity constrained to his estimated value of 0.68. This indicates that

642

Brookings Papers on Economic Activity, 3:1973

the use of a Bayesian technique of combining the sample evidence with a

very diffuse prior distribution centered on unity would result in a flat
distribution of income elasticity estimates over a very wide range.
A second point concerns the choice of the dependent variable. Several
argumentsfavor the use of the interest rate, rather than money, on the lefthand side. The most widely recognized of these is that the interest rate is
"more endogenous" than the money stock. I would stress that the presence
of errorsin variables may provide an even more compelling reason for running the equation in this form. Aside from the obvious problem of measurement error, there may be errorsin the interest rate variables that arise from
their use as proxies for other interest rates, since nobody knows which rate
affects behavior the most. I won't care about the commercial paper rate if I
am never going to buy commercial paper, but I may respond to changes in
this variable if it is highly correlated with a rate I do care about. As a technical aside, I would point out that serious statistical problems could result
from the correction for serial correlation in the presence of measurement
errors in the variables.
I have my doubts about the inclusion of the change in prices in the money
demand equation. I would argue that the cost of holding cash balances is
equal to the real rate of interest plus the rate of inflation (or of expected inflation), and the sum of the two components is presumablymeasured by the
nominal interest rate. The only situation in which the nominal interest rate
would fail to measure this cost would be a severe deflation, when the nominal rate of interest could not become negative whatever the expected rate
of deflation.
On the choice of the appropriate interest rate for the money equation, I
would like to make one point that favors the long-term rate. Since the
difference in goodness of fit between the equations is small, and since the
R2s are very high in both, using the long rate may be preferablebecause it is
more appropriate in the context of a complete econometric model. The
long-term rate enters directly into the investment equation, so its use in the
money demand equation may eliminate the need for the model to grapple
with the term structure of interest rates.
The disaggregated equations for currency and demand deposit holdings
raise many questions in my mind. The results of the two equations with income as a scale variable show the income elasticity of currency holdings to
be much higher than that of demand deposit holdings. Given the long-run
increase in income, the currency-depositratio should therefore have risen,

Stephen M. Goldfeld

643

and indeed it did display a gradual updrift between 1920 and 1950. Thereafter, the ratio does not exhibit the continued rise that would be expected
on the basis of the strong increase in real income in the postwar period. The
equations attributethe puzzle to increases in the interest rate, since holdings
of currency show a higher interest elasticity than do demand deposits. But
that differencein interest elasticities is not a plausible result.
The separate equations for currency and demand deposits also raise an
issue about the strategy of monetary policy. They imply that the Federal
Reserve might just as well pursue its monetary policy objectives by setting
a target for the quantity of currency and then adjusting the reserve base as
necessary to provide whatever volume of demand deposits the public desired in conjunction with the targeted volume of currency. I doubt, however, that any economist would believe that such control through currency
alone would represent an adequate way for monetary policy to influence
GNP.
My final point concerns the flow of funds data. The measurement errors
in the flow of funds accounts may be quite large. Estimates of demand
deposits and currency held by the various sectors on December 31, 1972,
were about 6 percent less than the estimate of the sum of currency outstanding and the deposit liabilities of the commercial banks. The discrepancy represents an unknown combination of measurement error and mail
float. Also, the results obtained by disaggregating holdings of households,
corporations, and so forth differ considerably from the results of the new
survey of demand deposit ownership. For these reasons, I doubt that work
with the disaggregated data can prove fruitful; the data are simply not yet
robust enough to stand up to statistical regression techniques.

GeneralDiscussion
Severalparticipantscommented on the statistical problems of Goldfeld's
estimate of the demand for money that uses the stock of money as the dependent variable. William Gibson remarkedthat a complete analysis of the
behavior of the money stock should deal explicitly with responses of the
money supply through the actions of the Federal Reserve. He advocated
using simultaneous estimation techniques to separate the effect of interest
rates on money supply and demand. On some assumptions about the determination of the money supply, Robert Hall argued, the estimation of the

644

BrookingsPaperson EconomicActivity,3:1973

demand function by ordinary least squares could give seriously misleading

results. In support of Hall's contention, Franco Modigliani cited the results
he obtained from estimating the demand equation as part of a system of
simultaneous equations. These estimates show a substantially larger interest elasticity of demand and a faster speed of adjustment to changes in
income and interest rates than estimates obtained by ordinaryleast squares.
Though Goldfeld's results using two-stage least squares move in the direction of the results from simultaneous equations, Modigliani felt that the
issue had not been completely resolved.
Hall saw a possibility of obtaining reliable estimates of the relationship
among income, interest rates, and the money supply by ordinary least
squares if the money supply could indeed be considered exogenous to the
model. In that case, income or interest rates, rather than the money stock,
would appear on the left side in an equation like Goldfeld's. Hall pointed to
work by Christopher Sims which supports the view that the money supply
can indeed be considered exogenous. Thomas Sargent agreed that Sims'
exogeneity tests show that it is inappropriateto treat money as endogenous
and income as exogenous in estimating a demand schedule for money.
Lawrence Klein, on the other hand, held that tests that rely on lagging and
leading correlations of money and income are insufficient to determine the
exogeneity of the money stock. In particular, Klein argued, the actual
money supply resulting in the very short run depends not only on Federal
Reserve actions in targeting the money supply but also on income and
interest rate conditions. Stephen Magee added that international money
flows present a further argument in favor of considering the money stock
endogenous. Sargent maintained that Sims' theorems prove that his exogeneity test is valid, and that the mere fact that the Fed has pursued an
interest rate target is insufficient to show that the money stock was not
statistically exogenous, since the Fed revised its target interest rate quite
often and in a complicated way.
In response to these criticisms, Goldfeld agreed that the issue was unresolved. But he emphasized that his results with two-stage estimation using
the interest rate on the left-hand side essentially duplicated those of the
equation with money on the left side, and also agreed with the ordinary
least squares estimates using money as dependent. The ordinary least
squaresestimates with the interest rate dependent differed sharply from the
other three sets. He saw this 3-to-i "vote" as offering some support for his
approach.

StephenM. Goldfeld

645

Modigliani expressed his admiration for the thoroughness of the tests

Goldfeld made. He was reassured that the results generally confirmed sensible theoretical views-in particular, that income is the proper transactions variable, that the demand for money is stable and homogeneous in
prices, and that the income elasticity of the demand for money is less than
one. In addition, he approved the use of the short-term rate rather than the
long-term rate; in response to a suggestion by Poole that firms' decisions to
set up cash management programs might depend on anticipations of the
short rate over a longer horizon, Modigliani suggested that the most relevant rate might be a combination of the short-term rate and an interest
expectations variable.
On the question of the nature of the lags in the equation, Modigliani emphasized the need to distinguish between the partial adjustment and the expectations arguments for the use of distributed lags. He held that the lagged
adjustment to the interest rate reflects the process of learning to economize
on cash balances in the face of higher interest rates. It then takes time to
adjust cash balances to the desired level. For this reason, Modigliani did
expect the differences in lag structures and adjustment speeds for the interest rate and income variables that Goldfeld obtained in his Almon
equations.
Interpreting the lag as a partial adjustment resolved the mystery of the
large impact of price change on the demand for money, in Modigliani's
judgment. The desired stock of money is determined in real terms, but the
gradual adjustment applies to the nominal money stock. That adjustment
will take time when prices rise as well as when real incomes or interest rates
change. Temporarily-but only temporarily-rising prices can hold down
the real demand for money through the partial adjustment. This view is
confirmed by Goldfeld's finding that the deflation of lagged money by current (rather than lagged) prices eliminates the significance of the inflation
variables.
Modigliani was puzzled by the results of Goldfeld's test of the effects of
population growth on money demand. The economies of scale in the transactions model should apply to income growth per capita; that portion of
aggregate income growth that reflects just keeping up with population
should raise the demand for money proportionately. Goldfeld's empirical
finding disagrees-his only counter-theoretical result, in Modigliani's judgment. Arthur Okun suggested that the proper population variable to capture the economies of scale might be the number of households rather than

646

Brookings Papers on Economic Activity, 3:1973

of persons, and that the latter may be a poor proxy for the former during
the postwar period.
Severalparticipantscommented on the price expectations variable in the
money demand equation. Gibson did not see why the expected rate of inflation should influence the demand for money except through its influence
on the nominal rate of interest. Okun agreed with Gibson's remarks (and
with Poole's earlier comments) insofar as they applied to an Irving Fisher
world in which the full effect of anticipated inflation would be reflected in
nominal interest rates. But Okun and William Brainard suggested that direct substitution of goods for money would alter the story. As Brainardexplained, some people might be more concerned with the allocation of their
wealth between physical and financial assets than with allocation among
various forms of financial assets. If this were the case, the rate of increase of
commodity prices would affect the demand for money, whether or not the
nominal interest rate captured fully the effects of price expectations. Moreover, even with no direct substitution between money and goods, the demand for money could be influenced by price expectations if for any reason
nominal interest rates did not adjust fully to changes in price expectations.
David Fand noted that the extent to which disaggregation is desirable
depends on the reasons for wanting to know a particularmonetary total.
Though it might be possible to get a better estimate of the sum of currency
and demand deposit holdings by estimating these holdings separately,
someone interestedin a differentmonetary total might want to disaggregate
differently,or not at all. Poole questioned the conclusiveness of Goldfeld's
evidence in favor of the disaggregation of money into currencyand demand
deposits, pointing out that the disaggregatedequations using GNP showed
lower ex post mean-squared errors in only five of the eleven cases.
Brainard offered a final comment about Goldfeld's tests. He applauded
Goldfeld's use of out-of-sample forecasts, but he thought the ex post forecasts should give more weight to the accuracy of the estimates for the early
quartersafter the sample period-say, the first four-than to later ones. He
also wondered how well the equations performed in the ex post simulations
as compared to simple naive autoregressive models.