You are on page 1of 18


Journalof the AmericanStatisticalAssociation

March 1970, Volume 65, Number329

ForecastingDaily Hog Pricesand Quantities:A Studyof




Relativelyfewinvestigationstreatthe problemof short-runor daily price and

quantityvariations.Thisstudyconcernsforecastingdaily hogpricesand quantities
in selectedterminalmarketsin the U. S. by using various causal and noncausal
models.The purposeis to examinethe economicand mathematicalcharacteristics
of the timeseriesdata, and then comparethe developedmodelsin termsof their
forecastingability. The performanceof each model is evaluated using the Theil
Large daily priceand quantityfluctuationscharacterizemany ofthe agricul-
tural commoditiessold throughterminalmarketsin the United States. Most
related research is confinedto investigationsof recurrentcommodityprice
cycles,trends,and patterns,whilerelativelyfewinvestigationstreat the prob-
lem ofshort-runor daily priceand quantityvariations.This researchvoid may
stem fromthe difficulties in obtainingand analyzingdaily time series data.
In this study we attemptto forecastdaily hog pricesand daily quantitiesof
hogs suppliedby means of several alternativeforecastingtechniques.Our pur-
pose is to examinethe economicand mathematicalcharacteristicsof the time
seriesdata, then comparethe developed models as to theirforecastingability.
The hog industrywas chosen because (a) the daily fluctuationsin price and
quantity of hogs typifymany of the agriculturalcommoditiessold through
terminalmarkets,(b) hog productionand processingis a multimilliondollar
industryin the United States, and (c) data are available on daily hog prices
and quantitiessupplied foreach of the marketsstudied. To restrictour anal-
ysis, we forecastdaily hog prices in the eight largest United States terminal
hog markets,whichnot onlyrepresentthe major concentrationpointsofbuyers
and sellers but also act as a guide in the pricingof hogs in other markets
[9, 11]. The marketsselected foranalysis are Chicago, Indianapolis, Kansas
City,Omaha, South St. Joseph,East St. Louis, South St. Paul, and Sioux City.
However,in makingdaily quantitysuppliedforecasts,we aggregatethe above
and view them as a singlemarket.
To forecastdaily hog pricesand quantities,we employboth an econometric
model and recentdevelopmentsby Box and Jenkinsin time series modeling
[1, 2].1 The main distinctionis that the econometricmodel attemptsto identify
* R. M. Leutholdis assistantprofessorof agriculturaleconomics,Universityof Illinois.A. J. A. MacCormick
is researchassistant,Departmentof Statistics,Universityof Wisconsin,Madison. A. Schmitzis assistantprofessor
of agriculturaleconomies,assistantagriculturaleconomistin the ExperimentStationand is on the GianniniFoun-
dation,Universityof California,Berkeley.D. G. Watts is associateprofessorofstatistics,Universityof Wisconsin,
Madison. A large part of thii study was financedby the Economic Research Serviceof the U. S. Departmentof
Agricultureand the Universityof Wisconsin.The econometricmodel used is essentiallythat in 181.
1 The originaldraftof 121was issuedin 1966and 1967as TechnicalReportsNos. 72, 77, 79, 94, 95, 99, 103, 104,
116, 121,and 122 of the Departmentof Statistics,Universityof Wisconsin,Madison; and simultaneously as Tech-

DailyHog Pricesand Quantities
Forecasting 91
and measure both economic and noneconomicvariables affectingprice and
quantity,while the time series approach identifiesstochastic componentsin
each timeseries; that is, autoregressiveand moving average components.For
both approaches,1964 data on daily priceand quantityare used forestimation.
It has become conventionalto classifyforecastsas being eitherex post or
ex ante. In the latter,truevalues of exogenousvariables are not knownat the
timethe forecastis made, hence theymustbe estimated.Observedforecasting
errorsmay, therefore,be due to errorsin the estimatesof the predetermined
variables or incorrectspecificationof the model. In this study, the forecasts
presentedare of an ex post nature. These use the actual values of the prede-
terminedvariables to predictvalues forthe endogenousvariables; any errors
are, therefore,attributableonly to the model itself.
To test the predictiveaccuracy of the forecastsobtained by both the econo-
metric model and the parametric modeling approach, 1965 data are used.
Therefore,the models are testedusing data beyond the period of fit.Also, the
predictiveaccuracy of the differentforecastingapproaches is compared by
using Theil's inequalitycoefficient[13].
The coefficient used is:

2;(X - X,_1) 2

whereXi denotes an observationmade at time i and XXdenotes a predicted

observation.It can be seen that U representsa comparisonofthesum ofsquares
of the one-step-aheadforecasterrorsof a model with those of a randomwalk
model. Thus fora randomwalk model, U = 1 always.
The econometricmodel, which we estimate by ordinaryleast squares, is
composedof both a demand and a supply equation forlive hogs. The quantity
demandedis a deriveddemand as hogs are the major raw materialin producing
pork. Assumingprofitmaximization,the quantity of an input demanded de-
pends upon its price,the price of otherinputs,and output price. However, at
the major markets the supply of hogs generallyarrives beforethe markets
open,and the marketsare thenclearedeach day. The daily quantitydemanded
by the buyerson any particularday is predeterminedby the quantity avail-
able, and price is the adjusting mechanism.2Therefore,the demand function
expressesinput price as a functionof quantity,output price,plus othervari-
ables, such as storage and day of the week.
For the priceof output,we assume that Pt*= Pt-, wherePt* is the price at
time t which the packer expects to receivefor wholesale pork cuts afterpro-
cessing;and Pt-, is the actual price receivedthe previousday. As the previous

nical ReportsNos. 1, 2, 3, 4, 6, 7, 8, 9, 10, 11, and 13 of the Departmentof SystemsEngineering,Universityof

2 Of course,quantityvaries fromday to day and fromweek to week. However,tradingat the terminalhog
marketsi8 conductedin such a fashionthat sellershave virtuallyno chance to respondto the pricewithinthe day.
This tradingcharacteristicis not necessarilytrueforcountrybuyers,dealers,and auctionmarkets,but theyare not
analyzedin thisstudy.
92 Journalof the AmericanStatisticalAssociation,March 1970
day's wholesale price increases,packer price bids for slaughterhogs will in-
crease the next day because of the expectationof higherwholesalepricesin the
It is also hypothesizedthat the price packers offerforhogs depends in part
upon the numberof hogs they have on hand and the amount of pork in cold
storage. Eleven models were tested using various combinations of lagged
terminaland interiormarket quantities to find a proxy to representthese
storageitems.3The most appropriateproxyis the aggregatequantitysupplied
by the eight terminalmarketsforeach of the previous two days. Thus these
two quantitiesare includedin the demand functionto representstorage.
Additionalfactorscause packers to vary theirprice bids on hogs according
to the day of the week. One is the packers' attemptto filla specificnumberof
hours per week in theirlabor contract,requiringthem to increasetheirprice
bids at the beginningof the week so that theycan purchaselarge quantitiesof
hogs and be assured of fulfilling the contractsand maintaininga uniformrate
of slaughterthroughoutthe week. This and otherinfluencesare representedby
dummyvariables whichshiftthe demand functionaccordingto the day of the
The supply functionforslaughterhogs is derived fromthe classical theory
of the firm.Hog producersare assumed to operatein a purelycompetitivemar-
ket and attemptto maximizeprofits.One shortcomingof the theory,however,
is that it fails to distinguishbetweenmarketing(short-run)decisionsand pro-
duction(long-run)decisions.Since our concernis withthe short-run, the classi-
cal theoryis modifiedsomewhat to exclude past productioncosts and most
The quantity of hogs supplied at the eight terminalmarketsarrivesbefore
the marketsopen and thereis virtuallyno opportunityforproducersto adjust
their shipmentsin response to price withinthe day at these same markets.
Therefore,it is assumed that producersoperate on the basis of an expected
pricefortheirhogs,whichis based on informationgeneratedpriorto the mar-
ketingday. Eleven modelsusing various weightedand simple movingaverage
schemes of past prices to representthis expectationwere tested. The model
whichfitsthe data the best is one in whichthe expectedpriceis the same as the
previous market day's price, afteradjusting for the day of the week. Thus,
price lagged one period is included as an explanatoryvariable in the supply
function,withthe expectationthat as this priceincreases,producermarketing
responseswill increasethe next day and conversely.
Previous studies have shown the day of the week to be an importantinflu-
ence on quantity supplied [4, 12]. Farmers tend to marketmore hogs at the
beginningof the week, especially Monday, because they expect higherprices
then, and they have more time on weekends to prepare their livestock for
market.This phenomenonis representedby dummyvariables whichshiftthe
supply functionaccordingto the day of the week.
Hog farrowingfollowsseasonal patternsand this subsequentlyaffectsthe
numberofhogsto be marketedat latertimesafterthe necessaryfeedingperiod.

on cold storageand hogs on hand, are not readilyavailable; so previousquantities

3 Daily data, specifically
suppliedare used as approximationsof storage.
4 These and othervariables whichgo to make up the demand model used are definedmorespecifically later.
DailyHog Pricesand Quantities 93
A variable whichshiftsthe supply functionaccordingto this seasonal adjust-
mentis includedas an explanatoryvariable.
The demand and supplyfunctionsas specifiedabove definea recursivemodel
of the cobweb type where firstthe quantity supplied is determinedthrough
the supply function,and then this quantity is sold in the market at a price
determinedthroughthe demand function.That is, the quantity demanded is
a predetermined variable and the errortermsof the demand and supplyfunc-
tions are assumed to be independent.Thus, with the normal assumptionson
stochastic regressors,ordinaryleast-squares regressionwill yield consistent
and asymptoticallyunbiased estimatorsof the parametersof each function.
The double (natural) logarithmicfunctionalformis used for both the de-
mand and supplyfunctionsbased on the assumptionthat buyersand producers
respond to percentage (or relative) changes in the market. Because of the
interestin explainingdaily changes ratherthan absolute amounts, plus the
need to reduce autocorrelationin the residuals and multicollinearityin the
originalvalues, the economic data are transformedinto firstdifferencesof
logarithms.However, the dummy variables are not transformedinto first
differencesand theyare specifiedto act as shiftersof the functions.The entire
statisticalmodel can be representedas follows:
Demand: A logPt = /11AlogWht-1 + 312AlogQt + f13AlogQe-1
+ 314AlogQt-2 + 115 logM + 116logT
+ j17 logW + f18log F + Aui(t)
Supply: A logQt = j321AlogPt-1 + f22 logM + /23 logT
+ 324logW + I25log F +f26 logS + Au2(t)
Pt= average price in dollars per hundredweight
of all barrows and
gilts sold at the eightmarketsin day t
Pt-, = representationforexpectedpriceby producers,definedthe same
as Pt but forthe previousday
Qt= total number of head of barrows and gilts sold at the eight
Qt-s, Qt-2 = "proxy?'representations forstorage,definedthe same as Qt but
indicatingquantityon previousdays
Wht-,= representation of the priceof the output,measuredas a weighted
average (based on a percentof total live animal weight) of the
prices in dollars per hundredweightof major wholesale pork
cuts, Chicago
M 1.0000 otherwise

T= 1.0000 otherwise
W 1.0000 otherwise
94 of theAmerican
Journal March1970

F =2.71828ifFriday
= 1.0000 otherwise5

S = 2.71828ifMarch,
August, October,
September, November
= 1.0000 otherwise
t= time in days
A = first differences,such that A log Qt= log Qt-log Qt-i, and
A log Qi-1= log Qt_l-log Qt2, etc.
Demandand SupplyEstimates
The above model was estimated,each equation separately,by ordinaryleast
squares regressionusing daily 1964 data for the aggregated eight markets.6
These resultsare presentedin Table 1. Estimates were also obtained foreach
of the eight marketsindividuallywith the only model alternationbeing the
inclusionof a weightvariable and a proxyforinformationflowsin the demand
equation. These resultsare not shown because (a) the size and statisticalsig-
nificanceof the coefficientsforthe individualmarketsare verysimilarto those
illustratedin Table 1 and technicallyadd nothingnew, and (b) our primary
interestis to see how well this basic model forecastsrelativeto the parametric
modelingapproach presentedlater.Also, because of thisinterestin forecasting,
the coefficients in Table 1 are not specificallyinterpretedfor their economic
implicationsexceptto note that the signsof the coefficients are consistentwith
the theoreticalframework ofthe model,that the coefficients are approximations
of elasticitiesor percentageresponses,and that the dummyvariables act as
demand and supply equation shifters.
Priceand Quantity
Using the precedingestimations,price and quantity forecastscan now be
determined.Predictionsof priceand quantityare made one successiveday at a
time for 1965 (data not used in estimation). The procedureis to insert the
knownvalues of the independentvariables into the proper functionand ob-
tain an estimate of the dependentvariable. This estimate is then compared
with the observedvalue of the dependentvariable forthat particularday.
For 1965 thereare nearly250 observationsforwhichprice and quantityare
predicted,and these resultsare summarizedby computingthe Theil coefficient
whichwas outlinedearlier.These coefficients,indicatingthe forecastingability
of the econometricmodel, are discussed later when we compare and evaluate
the alternativeapproaches to forecasting.
Given past daily hog pricesand quantities,the problemis to forecastthese
for1, 2, 3, * * *, n days in thefuture.The approachused is timeseriesmodeling
5Dummy variablescannotbe specifiedforall fivedays of the weekas thiscreatesa linearcombinationin the
columnsof the raw momentmatrixand makesthe matrixsingular,so Thursdayis chosenas a base. The terminal
marketsare not open on Saturdayor Sunday.
6 Essentially,onlythe quantityvariableswereaggregated;the pricesforthe eightmarketswereaveraged.
DailyHog Pricesand Quantities 95

Demand equation Supply equation

Independent and
Coefficient Independent and
variable standarderror" variable standarderror5

Constant - .0026d Constant .0108

term (.0014) term (.0252)
A log Qg - .0542b A log Pt-I 8.2771b
(.0035) (.8409)
A log Qt-s - .0417b log M 3.5647b
(.0040) (.3326)
A log Qg2 - .0155b log T -1.2155b
(.0036) (.3276)
A log Wht-, .3536b log W -1.2955b
(.0471) (.3255)
log M .1609b log F _3.1978b
(.0218) (.3291)
log T .0768b log S .5219?
(.0242) (.2096)
log W - .0199 R2 .7147
log F - .0804b d (Durbin-Watson) 2.3516
R2 .5487
d (Durbin-Watson) 1.5990
a Figuresin parenthesesdenotestandarderror.
b Significant
at the 99 percentlevel of confidence.
c Significant
at the 95 percentlevel of confidence.
d Significant
at the 90 percentlevel of confidence.
NOTE: Dependentvariable (demand): A log Pt.
Dependentvariable (supply): A log Qt.

in which futurevalues of the economic time series are determinedby the

stochasticmodel whichdescribesthat series.
For the purposesof this article,two generaltimeseriesmodels are discussed
-stationary and nonstationary.An economic time series is stationaryif its
propertiesare unaffectedby a shiftin the timeorigin.In particular,a stationary
seriesvaries about some fixedmean u.
Models for StationaryTime Series. We denote the values of a series at
equally spaced times t, t-1, t-2, *- , by Pt, Pt-1, Pt-2, * . - . Let Xt, xti1,
Xt-2,***,be a "whitenoise" seriesconsistingof uncorrelatedrandomnormal
deviates,each with zero mean and variance ox..
96 Journalof the AmericanStatisticalAssociation,March 1970
Let ft = Pt- u, whereAtis thedeviationfromthe mean u. If this deviationis
linearlydependenton xi and on one or moreof the previousx's, thenwe have a
movingaverage (ma) model. For example,the firstorderma model is
Pt = Xt Oixt-1.(i

In general,a qth orderma processis

Pt = Xt - Gixt-i- O*Xt-q.
- (2)

If, instead,Pt is linearlydependenton previousdeviations,Pt-,, etc., and on

xj, we have an autoregressive (ar) model. Thus,
Pt Xt + lPt-i (3)
is a firstorderar model. In general,a pth orderar model is
Pt = xt + q lPt-i + * * * + qPpt-P. (4)
In the above equations,Oqand op representthe parametersof the model which
are to be estimated.
Certaineconomictime seriesmay combineboth of the above models. Thus,
a generalmixed (ar-ma) model of order (p, q) can be writtenas
Pt - - **- pt-P = Xt - *-i * -OqXt-q. (5)

To simplifythe notation,it is convenientto use a shiftoperatorwhose opera-

tionis definedby
BiPt = Pt-j. (6)

Usingthis operator,Equation 5 can be rewrittenas

Op(B)Pt = 0q(B)xt (7)

wherecp(B) and 0,(B) denotepolynomialsin B of orderp and q, respectively;

that is, in Equation 5
op (B) = (I1- 0B- 02B2- * * BP_
O,(B) = (1 - 01B- 02B2- o

In the precedingmodels,the Pt's are expressedin termsof deviationsfrom

the mean. If, instead, the Pt's themselvesare used, the general formof the
modelrepresentedby Equation 5 may be writtenas
Op(B)Pt =o + Oq(B)xt (8)

G0= (1-41-02- * p*-&,)U.

If the roots of the ar and ma componentssatisfycertainstabilityconditions

[2], the models (2), (4), (5) and (8) may be used to representstationarytime
series.If the roots do not satisfythe stabilityconditions,the models may be
used to representcertainnonstationarytime seriesas describedbelow.
ModelsforNonstationaryTime Series [3, 6, 7, 14]. A simpletype of nonsta-
tionaryseries is one which is homogeneouslynonstationaryin level; that is,
ForecastingDaily Hog Pricesand Quantities 97
apart froma verticaltranslation,one part ofthe serieslooks muchlike another.
Such a serieshas the propertythat its firstdifference,
APt = Pt - Pt-, = (1 -B)Pt,
is a stationarytimeseries.The operatorA is called the difference
its operationcan be definedby Ad= (1- B)d. A serieswhichhas neithera fixed
level nor a fixedslope but is otherwisehomogeneoushas the propertythat its
second difference, definedby
A'Pt = Pt- 2Pt_j + Pt2 = (1 -B)2Pt

is a stationarytime series.
Finally, to incorporatenonstationarityinto the previous models, let Yt be
the variable whose behavior we wish to represent.It is assumed that its dth
differenceAdYt=Pt can be representedby the stationary model given in
Equation 8. Since Ad= (1 -B)d, the model for Yt becomes
p(B) (I1- B)dYt = 0o + Gq(B)xt. (9)

Time Series Modeling.It is assumed that there exists a model of the form
Equation 9 which describesthe random propertiesof the available historical
data-hog pricesat Kansas City,forexample.The problemis to findvalues for
p, d, q and estimatesforthe parameterscp and Es such that the model "fits"
the data as closely as possible. Once a model has been satisfactorilyfittedto
the historicaldata, it may be used forforecastingfutureevents.
Every autoregressivetime serieshas an autocorrelationfunction(acf) which
is characterizedby an exponentialdecay, and it is possible by examinationof
the acfto identifythe order-that is, the value ofp-for an unknownar series.
Every ma time series,on the otherhand, has an acf whichhas only q non-zero
values. Afterlag q, all autocorrelationcoefficientsare zero. Inspection of the
sample acf computedfroma data seriesreveals *-+ to an experiencedobserver
whetherthe time series fromwhich it was computed is ar or ma or mixed
ar-ma and also the ordersp and q of the ar and ma components.
The sample autocorrelationcoefficient rk at lag k is computedfroma data
series P1, P2, * * , P. as follows:
rk(P) =CkP
Ce (P)

1 n-k
Ck(P) (Pt - P)(Pt+k -5)
n t=l

P= E P,
n t=i

The nextstep in fittinga nmodel

is to obtain estimatesforthe parameters,cp,
i=l, . . ., p; and OA,i-1, . ., q. A brief outline of one method for doing
thisis given in the appendix.
98 Journal
of theAmerican
Statistical March1970



V Firstdifference


2 4 6 8 10 12 14 16



The thirdstep in fittinga model is to check that the model providesan ade-
quate fit,and this is done by examiningthe residuals.The residualsof a well-
fittedmodel will be very similarto white noise. In otherwords,if the fitted
model describesthe structureof the data sequence, therewill be no structure
leftin the residuals.We therefore examinethe aefof the residualsto detectany
structurewhich remainsand to findout how to alter our model to provide a
Example 1 (Price). The acf's computedfromthe originaldata and fromthe
firstdifference of the originaldata are plottedin Figure 1 fordaily hog prices
at Kansas City. The behavior exhibitedby the firstacf is characteristicof
nonstationarity-that is, it indicates that differencing is necessary. On the
otherhand, the acf of the firstdifference is approximatelyzero except at lag
zero. This behavior is characteristicof the acf of white noise. Therefore,the
implicationis that the operationof takingthe firstdifference of daily hog price
data at Kansas City produceswhitenoise. Thus the model
(1 - B)Pt = xt (10)


Ptnn Pt-1 + Xi
DailyHog Pricesand Quantities 99



2 4 6 8 10 12 14 16



representsthe behavior of hog price data at this market.7It was found that
the same behaviorapproximatelycharacterizesdaily hog pricedata at Chicago,
Indianapolis, Omaha, St. Joseph,St. Louis, St. Paul, and Sioux City.
The significanceof Equation 10 can be seen when it is used to forecasta
futurevalue. The estimatemade at time t (today) fora futureevent occurring
at time t+1 (tomorrow)is denoted by Pt+j and is obtained by takingthe ex-
pectation at time t+1 conditional on the informationavailable at time t.
That is,
Pt+1 = E(Pt+1) = E(Pt + Xt+l) = E(Pt) + E(xt+i).
But Pt is knownsince it is today's price,and E(xt+1) is zero since xt denotes
whitenoise with zero mean. Thus,
Pt+j = Pt. (11)
That is, the best estimateof tomorrow'sprice is today's price.
It is perhapsworthnotingthat this resultis not a defectin the model or the
model-fitting procedure,nor is it a defectwhichmay be rectifiedby more ad-
vanced noncausal models developed in the subsequent section. The result is
7 The process,model (10), sometimescalled a randomwalk model,was also foundto characterizestockmarket
prices.See [5, 10].
100 Journal
of theAmerican



-0.5 2 A 6 8 10 12 14 16

Lag -

inherentin the data and tells us that the historicaldata alone do not contain
information which will enable us to forecastfuturepricesmore accurately.
Example 2 (Quantity).In Example 1 it is shownthat daily hog pricesforma
sequence whichfollowsa veryloose structure.Daily hog quantitysupplied,on
the other hand, has a quite definitestructure.For this example we use the
aggregateddaily hog deliveriesto the eightmarketsstudied. The autocorrela-
tionof the total quantityof hogs supplied is plottedin Figure 2. Inspectionof
theacfreveals (a) a strongfive-dayweeklyeffectand (b) possiblenonstationar-
ity over days or over weeks or over both. To investigate(b), one-day or five-
day differencing may be called for.
Figures 3 and 4 show the acf of the first Pt= (1-B)yt and weekly
P (1-B5)yt difference,respectively.Inspection of Figure 3 suggests the
a. The high peaks at lag = 5, 10, 15, * indicate a weekly firstorder ar
b. The peaks are of almostequal heightand so the weeklyar parameterwill
be quite large, probablyclose to 1.
c. The peaks are depressed-that is, a line drawnthroughthe peaks at lags
15, 10, and 5 would be nearly horizontaland would not intersectthe
verticalaxis close to 1.0. This suggeststhe presenceof a weeldyfirstorder
ma component.
DailyHog Pricesand Quantities
Forecasting 101



2 4 V 6 8 10 12 14 16

Lag -

d. There appears to be a firstor second order ar componentwithinweeks.

e. The large magnitudeof the acf at lag= 1 suggeststhat firstdifferencing
is too stronga conditionto impose on the model.
In summary,Figure 3 suggeststhe model
(1 - VIB5)kv(B)Yt (1 -eB5)0q(B)xt (12)

where+Vis close to 1.0, and the structureof 4,(B) and Oq(B) is not yet clear.
Inspectionof Figure 4 suggeststhe following:
a. The deep troughat lag= 5 confirmsthe existenceof the weeklyfirstorder
ma component(1- EB5)in Equation 12.
b. The oscillatorynatureof the acfapart fromthe troughat lag= 5 suggests
a second orderar componentwithinweeks.
c. The oscillationsare quite small in amplitude and this may suggest the
presenceof a second orderma componentwithinweeks interactingwith
a second orderar componentwithinweeks.
In summary,Figures 3 and 4 suggestthe model
(1 -AtB5)(1 - -1B-02B2)Yt = (1 - eB5)(1 - 01B- 02B2)xt. (13)

was imposed on the model, the

In the actual fittingprocess no differencing
parametersbeingleftto findtheirown level. Several simplermodelsweretried,
and also a more complexmodel was triedin orderto check that the assumed
102 of theAmerican
Journal Statistical March1970






2 4 6 8 10 12 14 16

-0.2 Log -

modelwas neithertoo complexnor too simple; however,the residualsfromthe

above model displayed an acf closer to that of white noise, and a sum of
squares smallerthan all alternativemodels. The finalparametervalues are
6= .90 ki= 1.44 Oi = 1.52
e= .70 42 =-0.47 02= --0.66

and the model is

(1 - O.90B5) (1 - 1.44B + 0.47B2) Yt = (1 - 0.70B5)
(1 - 1.52B + 0.66B2)xt.

The acf of the residualsfromthe finalmodel is shown in Figure 5.

Case 1 (Price)
As shown previously,the best noncausal forecastof daily prices is yester-
day's price; that is, PAt=Pt-j.That is not to say, however,that this random
walk model forforecastingpricesis the best possible. This is demonstratedin
Table 2 in whichthe Theil coefficientis calculated fromforecastson daily 1965
DailyHog Pricesand Quantities 103

Theil coefficient
Type ofmodel
of Kansas City St. Paul Omaha
Full model .7978 .7655 .8556 .8234
Economic variables .8888 .7806 .9599 .8255
Dummyvariables 1.0288 1.0329 .9847 1.0400
Random walk model 1.0000 1.0000 1.0000 1.0000
Mean model 18.2342 14.8747 14.6248 15.8638

price data for the average price of the eight markets and three individual
markets,8Kansas City, Omaha, and St. Paul, for (a) the econometricmodel
presentedat the outset, (b) the randomwalk model,and (c) a mean model. In
order to observe the influenceof the dummy variables, coefficientsfor the
econometricmodel are calculated for the full model (Table 1), the economic
variables alone, and the dummyvariables alone. (New regressionswere run in
each of the latter two cases.) The mean model simply uses the mean ofthe
data ti, . .. , tnto forecasttn+l
The Theil coefficients in Table 2 suggestthat the fulleconometricmodel does
a betterjob offorecastingdaily marketpricesthan does eitherthe randomwalk
or mean models in each of the fourcases presented.However, the differences
betweenthe Theil coefficients forthe econometricand randomwalk modelsare
extremelysmall when comparedwiththose forthe mean model. This suggests
that ifone chose to forecastpricesnoncausally,a randomwalk model would be
superiorto a simplemean model. Also, the coefficients indicate that using the
economicvariables alone will providemoreaccurate forecaststhan the random
walk model, while the dummy variables alone provide slightlyless accurate
forecasts,except for St. Paul. In each case the coefficient for the full econo-
metricmodel is smallerthan the respectivecoefficientwhen using economic
variables only,suggestingthe inclusionof the dummyvariables does provide
some increasein forecastingaccuracy.

Case 2 (Quantity)
The supply equation previouslyestimatedusing time seriesanalysis is
(1 - 0.90B5)(1 - 1.44B + 0.47B2)Yt = (1 - 0.70B5)(1 - 1.52B + 0.66B2)xt.

To use this equation forforecasting,multiplythe factorsto give

(1 - 1.44B + 0.47B 2- 0.90B5 + (0.90 X 1.44)B6 -(0.90 X 0.47)B7) Yt
= (1 - 1.52B + 0.66B2 - 0.70B5 + (0.70 X 1.52)B6 - (0.70 X 0.66)B7)xt.
Then, the forecastingequation is
2 Theother
fiveindividual arenotincluded
markets to thethreeshown.
as theirresults
104 of theAmerican
Journal Statistical March1970
Yt = 1.44Yt1- 0.47 Ye-2 + 0.90 Yt5 - (0.90 X 1.44)Ye-e
+ (0.90 X 0.47)Yt-7 + Xt - 1.52xt-,+ 0.66xt-2-0.70xt-5
+ (0.70 X 1.52)xt-6- (0.70 X 0.66)xt.-7
Thus, the one-step-aheadforecastis
Yt+1 1.44 Y - 0.47 Yt-1 + 0.90 Y-4 - (0.90 X 1.44) Yt_5
+ (0.90 X 0.47) Yt-6 + E(xt+1)- 1.52xt+ 0.66xt_-- 0.70xt-4
+ (0.70 X 1.52)xts5- (0.70 X 0.66)xt-6
where Yt, ** , Yt-6 and xt, * * *, xt-6 have already been observed,but xt
has not. Since x representszero-meanwhite noise, E(xt+) =0.9


Type ofmodel Theil coefficient

Full model .65
Economicvariables .95
Dummy variables .75
Random walk model 1.00
Mean model .96
Time seriesmodel .70

Table 3 gives the Theil coefficients for the above time series forecasting
model and three other models-(1) the econometricmodel presentedat the
outset and then divided into two subgroupsin the same fashionas described
whenforecastingdaily prices,(2) a randomwalk model,and (3) a mean model.
The latter two were also discussedin the previoussection. In contrastto the
mean model forprices,the mean modelforsupplydoes a betterjob offorecast-
ing than does the randomwalk model. However, neitherof these can forecast
as well as the fulleconometricor timeseriesmodels. Comparingthe lattertwo,
the fulleconometricmodel does a betterjob of forecastingas the Theil coeffi-
cientis .70 forthe timeseriesmodel and .65 forthe econometricmodel. On the
otherhand, the timeseriesmodel is moreaccurate than the econometricmodel
if eitherthe dummyor economicvariables are used alone. Using the dummy
variables alone (.75) is more accurate than using the economicvariable alone
(.95). This latter relationshipmightbe expected when consideringthe strong
weeklyeffectdemonstratedby the time seriesmodel.
This study is concernedwith forecastingdaily hog prices and quantities in
selectedterminalmarketsin the United States. Forecasts are made usingvari-
9For example,the three.step-aheadforecastis
Y9+3 = 1.44 Y,+2 - 0.47 Yt+? + 0.90 Yt2 - (0.90X1.44) Yg.3
+ (0.90X0.47) Y_4 + 0.66 E(xt+i) - 0.70 Xti-
+ (0.70 X 1.52) xt-3- (0.70X0.66) xtg_
whereE (xt+l) -0.
DailyHog Pricesand Quantities
Forecasting 105
ous causal and noncausal models,includinglinear stochasticmodels. The per-
formanceof each model is evaluated using the Theil coefficient.
The findingssuggestthat daily hog prices exhibita randomwalk behavior;
similarfindingswere previouslydiscovered with daily stock market prices.
However,our resultsindicate that this does not necessarilymean that a ran-
dom walk modelprovidesthe best forecastofprices.As shown,our econometric
model yields more refinedprice forecasts,even when only the economicvari-
ables are used and the dummyvariables are excluded. When the dummyvari-
ables are used alone, generallythe resultsare less accurate than the random
walk model. Our resultsalso suggestthat in forecastingdaily quantities the
econometricmodel yields slightlysuperiorresultsthan the time series model.
However,the time seriesmodel is more accurate than if the dummyvariables
are used alone in the econometricmodel.
It should be emphasized that the Theil coefficient for comparingforecasts
does not take into account the costs involved in constructingalternativefore-
castingmodels. One advantage of timeseriesmodelingis that only data on the
variable to be forecast are needed. However, in constructingeconometric
models,data are needed on both the regressandand regressors.Therefore,even
thoughthe forecastsusingthe econometricmodel are slightlybetterthan those
using a stochasticnoncausal framework,the cost of makinga slightlygreater
errorusing the latter could well be less than the additional cost involved in
settingup an econometricmodel and collectingthe data. Unfortunately, we do
not have these cost figures.
The stochasticmodels and the methods for fittingthem describedin this
articleare due to the workof Box and Jenkins[2]. The fittingof a model is an
iterativeprocess which consistsof three steps: identification,estimation,and
It is conventionalforstatisticalproblemsto begin with an assumed model
but, in practice,a sequence of observationsdoes not come with a label tied to
it indicatingwhich model to assume. Thus, "identification"is the procedure
for obtainingan approximateand intuitiveidea of the structurewhich may
reasonablybe assumed to describethe observedbehavioruntilfurtherevidence
is obtained indicatinghow the model may be modifiedto provide a closer fit
to the data. We have brieflyillustratedthe identification of a model fora lag
quantitymarketeddata sequence.
Autoregressiveparametersin the model may easily be estimatedby least
squares, but moving average parameterscause difficulty. There are several
ways to overcome this difficulty
[21. We used a gridsearch method as follows:
Suppose we have identifieda model of the form
Pt - -lPt-1 2Pt-2 = Xt - GlXt-1 - O2Xt-2 (1)

We want to estimate 41, k2, 01, and 02. We know that they must fall withina
stabilityregion[2] and, by refinement procedure,we have a
of theidentification
roughguess as to values forthe estimatesOl,62.We select an area containing
the coordinatesof our guess and imagine a grid superimposedupon it. Each
106 Journalof the AmericanStatisticalAssociation,March 1970
grid point representsa possible value for 6i and 62. For each grid point we
computeresidualsfromthe ma part of the model. Symbolically,we write
(1 - -1B 2B2)Pt (1 - 01B- 02B2)xt

Pt (1- 01B 02B2)(1 - -1B 2B2)-lx (2)

= (1 - 01B- 02B2)et (3)
et = (1 - 01B - 4)2B2)-1xt (4)
and we computethe residuals
et-Pt + 02et- + 62e.-2. 2 (5)
Now, usingleast squares, we computeestimates '), 42 and residualsxtand the
sum of squared residuals Ett2 from
(1 - 1B -2B2))t = t.

Thus, we obtain a sum of squared residualsforeach grid point. We are inter-

estedin minimizingthe residualsum of squares, and so we take the gridpoint
at whichthe smallest sum of squares occurs to give estimates6', 02. Starting
values foret,in Equation 5 may presenta minorproblem (see [2] fordiscus-
sion); but since our data sequence contains 257 data points, the errorintro-
duced by settinginitialvalues equal to zero is considerednegligible.
How good is our assumed model? Diagnostic checkingseeks to findout. If
theassumed model is satisfactory,it describesall the structurein the data se-
quence and the residualswill consistof whitenoise. Thus, diagnosticchecking
consistsofapplyingthe identification phase ofour investigationto the residuals
ofthe assumed model. It is to be hoped that identification will show that the
residualsdo consistof whitenoise; but ifnot, a model will be identifiedforthe
residualsaild this will be incorporatedas a modificationto the originalmodel,
parameterswill be reestimated,residualswill be rechecked,and so on until a
satisfactoryfitis obtained.
[1] Box, G. E. P., and Jenkins,G. M., "Some Recent Advances in Forecastingand
Control,"Applied Statistics,17, No. 2 (1968).
[21 --, Time SeriesAnalysisForecasting and Control,"(In preparationforpublication
by Holden-Day, San Francisco).
[31Brown,R. G., Smoothing, Forecasting, and PredictionofDiscreteTime Series,Engfle-
wood Cliffs,N. J.: Prentice-Hall,Inc., 1962.
[4] Cramer,Charles, WhytheEarly-WeekMarket?NorthCentral Regional Publication
91, MissouriArgicultural ExperimentStationBulletin712, Columbia,1958.
[5] Granger,C. W., "Some AspectsoftheRandom Walk Model ofStock MarketPrices,"
InternationalEconomicReview,9, No. 2 (June1968).
[61Holt, C. C., ForecastingTrendsand Seasonals by ExponentiallyWeightedMoving
Averages,Carnegie Instituteof Technology,O.N.R. MemorandumNo. 52, Pitts-
17] -,Modigliani, F., Muth,J. F., and Simon, H. A., Planning Production,Inven-
tories,and WorkForce,EnglewoodCliffs,N. J.: Prentice-Hall,Inc., 1960.
DailyHog Pricesand Quantities
Forecasting 107
[81 Leuthold,Raymond M., "An EconomicAnalysisof Daily Hog Price Fluctuations,"
UnipublishedPh.D. dissertation,Universityof Wisconsin,Madison, 1968.
[9] Love, Harold G. and Shuffett,D. Milton,"Short-RunPrice Effectsof a Structural
Change in a Terminal Market for Hogs," Journalof Farm Economics,47, No. 3
(August 1965).
[101 Mandelbrot,Benoit, "Some Aspects of the Random Walk Model of Stock Market
Prices:Commnent," EconomicReview,9, No. 2 (June1968).
[111 National Commissionon Food Marketing,Food FromFarmerto Consumer, Washing-
ton: U.S. GovernmentPrintingOffice,June 1966.
(121 Schneidau, R. E., Pherson,V. W., and Cox, C. B., Is Therea Best MarketDay?
Pardue UniversityArgiculturalExperimentStation, Research Bulletin 709, La-
[13] Theil, Henri,AppliedEconomicForecasting, Amsterdam:North-HollandPublishing
Co., 1966.
[141 Winters,P. R., "ForecastingSales by ExponentiallyWeightedMoving Averages,"
ManagementScience,6, No. 3 (April1960).