cash flow forecast for south african firms

Their reported cash flow forecast model had adjusted R2 of 43%,althoughtheirsamplewassmallercomparedtopre-1989 studieswhenSFASNo.95hadnotbeenpublished.Orpurtand Zang’spaperexaminedwhethe

ScienceDirect

Review of Development Finance xxx (2014) xxx–xxx

Yun Lia, Luiz Moutinhob, Kwaku K Opongb, Yang Pangb,∗

aSchool of Engineering, University of Glasgow, University Avenue, G12 8QQ, United Kingdom

bAdam Smith Business School, University of Glasgow, University Avenue, G12 8QQ, United Kingdom

Abstract

ThispaperappliesmodelsintheextantliteraturethathavebeenusedtoforecastoperatingcashflowstopredictthecashflowsofSouthAfrican firmslistedontheJohannesburgStockExchange.Out-of-sampleperformanceisexaminedforeachmodelandcomparedbetweenthem.The reportedresultsshowthatsomeaccrualterms,i.e.depreciationandchangesininventorydonotenhancecashflowpredictionfortheaverage SouthAfricanfirmincontrasttothereportedresultsofstudiesinUSAandAustralia.Inclusionofmoreexplanatoryvariablesdoesnotnecessarily improvethemodels,accordingtotheout-of-sampleresults.Thepaperproposestheapplicationofmovingaveragemodelinpaneldata,andvector regressivemodelformulti-period-aheadpredictionofcashflowsforSouthAfricafirms

©2014AfricagrowthInstitute.ProductionandhostingbyElsevierB.V.Allrightsreserved

JEL classification: G300; M410; M490

Keywords:Cash flow; Moving average; Predictive models; Accruals; Vector autoregressive model; Depreciation

1 Introduction

Given that cashflow is the life-blood of a firm, accurate

determinationof cash flowsenables firmstomake important

financial decisionsthat relatetowhether the firmsurvives or

goesbankrupt.Asameasureofafirm’sprofitabilityand

finan-cial health, cash flows could provide potential clues about

the source company’s ability to pay divided and thus attract

investors’interesttoo.Therearethreecategoriesofcashflows

recordedinstatementofcashflow,i.e.cashflowsfromoperation,

∗Correspondingauthor.

E-mail addresses:Yun.Li@glasgow.ac.uk (Y Li),

Luiz.Moutinho@glasgow.ac.uk (L Moutinho), Kwaku.Opong@Glasgow.ac.uk

(K.K Opong), Y.Pang.1@research.gla.ac.uk (Y Pang).

Peer review under responsibility of Africagrowth Institute.

1879-9337 © 2014 Africagrowth Institute Production and hosting by Elsevier

B.V All rights reserved.

http://dx.doi.org/10.1016/j.rdf.2014.11.001

financingandinvestment,ofwhichoperatingcashflow, reflect-ing theability ofthe firmtoengage inday-to-day operations andits continuityinbusiness, isof the mostimportance.For themanagersoffirms,investorsoranalysts,predictionoffuture cashflowsareofextremeusefulnessandvalue

Thereare two issuestobe consideredwhen attemptingto predict afirm’scashflows.First, thevariablesthoseare use-fulandinformativetocashforecastneedtobe identifiedand incorporated into the forecast model Secondly, the typeand structure of modelstobeemployedinthe forecast shouldbe carefully chosen toprovide a moreaccurate prediction This study shed light onbothissues, intending todemonstrate the procedure of choosing variables and models for more accu-rate prediction There are anumber of difficulties withcash flowprediction.Generallyspeaking,cashflowismorevolatile thanearningsandthusharder topredict Thereisnouniform cashflowgeneratingprocessforthewholebusinessworldand different companies provide distinct patterns of cash flows Besides,duetothepopularityofcredittrade,afirm’srevenue andexpensesarenotequaltocashinflowandoutflowandthis compoundstheproblemofaccuratecashflowprediction Aca-demicstudiesoncashflowpredictionrelyonpublicinformation

as reflectedinafirm’sfinancial statementfor cashflowdata

Amongthevariablesthat havebeen foundusefulnessincash

flowpredictionincludeearnings,accrualtermssuchaschanges

inaccountreceivableandpayableanddisaggregatedcash

com-ponents Empirical studies suggest that these variables are

usefulandinformativeinpredictingcashflows.Inthispaper,a

comparisonismadebetweendifferentsetsofvariablesas

pre-dictors.Itisexpectedthatthemorepredictorsthatareincluded,

thebetteramodelwillperformbecausemoreinclusionof

vari-ablesoftenmeansricherexploitationofinformation.Asecond

comparisonis made between modelswhich include

explana-tory variables with different lags It is expected that models

with more lagged explanatory variables could provide more

accuratepredictionwhereas thereported resultsinthispaper

suggestotherwise.Thispaperproposestwotypesofcashflow

modeling,i.e.movingaveragemodelandvectorautoregressive

(VAR)modelforcashflowprediction.Movingaveragemodel

alsomakeseconomicsenseasitmeasureshowanunexpected

cashflowshockcouldinfluencepeoplemakingfuture

predic-tion.Themovingaveragemodelisappliedtoone-period-ahead

prediction and VAR model is proposed for

multi-period-ahead prediction For this purpose VAR is more powerful

and relies less on data availability than linear regression

ThesemodelsareappliedempiricallyondataofSouthAfrica

firms

Thispaperisorganizedas follows:Section1 providesthe

introductiontothispaper.Section2reviewstheliteratureand

discussesfactorsthatinfluencethepredictionofcashflowsand

thepredictionmodelsutilizedinthestudy.Section3describes

thedataforSouthAfricanfirms.Section4reportstheresultsof

theempiricalanalysisandtheconclusionofthestudyisprovided

inSection5

2 Literature review

Cash flow forecast is of interest to investors, creditors,

employeesandratingagenciesamongothers.Investorsare

inter-ested in cash flows as input into their investment models to

enablethemtodecideonpayoffrelatingtodividendsandcapital

appreciationoftheirinvestments.Creditorsareinterestedin

sol-vencydecisionsrelatingtothefirmstheytransactbusinesswith

andemployeesareinterestedinjobsecurityandgoing-concern

issuesrelatingtofirmstheyworkfor.Ratingagenciesarealso

interestedingoing-concernandafirm’sabilitytopayitsdebts

whentheyaredue

Cashflowcanbeconsideredascomplimentaryinformation

toearningssincecombinativeanalysisofbothquantitiesmight

bringbetterresultsthananalyzingearningsonitsown.Earnings,

alsosometimesreferredtoasnetincome,arethesummationof

net cashincome andnetcreditincome,the latterof whichis

basedoncredittradeswithcustomersandisnotyetbutexpected

tobesettledbycashinalaterperiod.Theamountofcreditgiven

tocustomerscouldpotentiallybeoverlookedwithoutcashflow

informationandthismaymisleadinvestorsabouttherisk

relat-ingtoshortageofcashinthefirm.Inaddition,cashflowdirectly

measurestheoperationalabilityofthefirmtomeetits

day-to-dayfinancialcommitments.Inconventionalfinancetheory,the

worthofafirmistheoreticallyequaltothediscountedvalueof

allcashflowsgeneratedduringthefirm’slifeassumingthatall thecashflowsarepaidoutasdividend.Asaresult,newsabout cash flow canpotentially have significant impacton afirm’s marketprice.Alongwithearningsforecast,analystsare increas-inglyincludingcashflowforecastintotheiranalysisandreports

ofcashflowwhichtheyderivedfromsalesandreacheda conclu-sionthatcurrentearningsarethebestforecastoffuturecashflow Earningsequaltocashflowplusaccrualsthatincludechanges

inaccountpayable,changesinaccountreceivable, changesin inventory,depreciationandamortizationandothers.IntheDKW model,accruals,forsimplicity,includeonlychangesinaccount receivable,changesininventoryandchangesinaccountpayable, whichareequivalenttochangesinworkingcapitalwhilelong termaccrualssuchasdepreciationarenotconsidered.TheDKW model makes several strict assumptions about sales process andworkingcapitalcomponentsandtheirderivedmodelrelies heavilyonthoseassumptions.Barth,Cram,andNelson(2001)

disaggregatetheaccrualsintocomponents,anticipatingthemto havedifferentpersistenceinpredictingfuturecashflow.Lorek

BCNmodel,intime-seriesandcross-sectionalanalysis respec-tively,andfoundthattime-seriesmodelgeneratesmoreaccurate result.Thisresultisnotsurprisingsincecross-sectional estima-tion treats all firmsas homogeneous, whichis hardly truein reality.TheDKWandBCNmodelsusetheindirectmethodto measurecashflow,i.e.theycalculatecashflowcomponentfrom netincomeandadjusttheresultswithaccrualterms.IntheUSA, statement of FinancialAccounting Standards (SFAS)No 95 issuedin1988allowedthedisclosureofdirectmethodcashflow statement.Therefore,cashflowsafter1988aredirectlyavailable fromthecashflowstatement.ChengandHollie(2008)partition cashflowcomponentsintocoreandnon-coreones,andanalyzed theirpersistenceforfuturecashflowdetermination.Thestudy defines corecashflowcomponentsas cashflowsfrom:sales, costofgoodssold,andoperatingandadministrativeexpenses Thenon-corecashflowcomponentsareinterest,taxes,and oth-ers.Whentheseregressorsareappliedinapredictionmodel,the adjustedR2isslightlygreaterthantheBCNversion.Similarly,

cashflowstatementenhancescashflowmodelingisexamined Their reported cash flow forecast model had adjusted R2 of 43%,althoughtheirsamplewassmaller(comparedtopre-1989 studieswhenSFASNo.95hadnotbeenpublished).Orpurtand Zang’spaperexaminedwhetheritmakesadifferencein estimat-ingcashcomponentsusingindirectmethodcomparedtousing discloseditemsdirectlyfromthestatementofcashflow.Theydo notdirectlycomparetheaccuracyofforecastmodels.Instead, theyexaminethestatisticalsignificanceofarticulationerrorthat

isdefinedasthedifferencebetweenestimatedcashcomponents and disclosed ones in their regression model Their reported results suggestthatthecoefficientsof articulationerrorterms arestatisticallysignificantandthusthatarticulationerrorshave incrementalinformationforcashflowforecast.Ithenceimplies thatthedirectmethodforcashflowsdisclosureismore informa-tiveinpredictingfuturecashflowthanindirectmethodstatement

of cashflow The direct method is not applicable for South

Africanfirms;thereforeinthispapercashdisaggregationwill

bebasedontheindirectmethoddescribedinOrpurtandZang’s

paper

Severalresearchershaveexaminedfactorsthatinfluencethe

accuracyofcashflow prediction.Forexample,Brochet etal

tocash flow shocks and discretionary accruals mightreduce

accrualscontributiontocashflowprediction.They arguethat

positiveaccrualsand/orhighcashflowvolatilityareassociated

withhigherpredictiveabilityofaccruals.Someresearchers

sug-gestafirms’past poor performancetendstobeimprovedby

managers,whichcouldleadtohigherpredictedfuturecashflow

betterprivateinformation, sotheycanprovide moreaccurate

predictionthanothers(PaeandYoon,2012)

Discretionaryaccrualsmaybeincomeincreasingorincome

decreasing,depending on managers motivations Badertscher

man-agement, including discretionary accruals and real operation

managementon cashflowprediction They arguethat

oppor-tunisticbehavioron thepart of managerscanlead totheuse

ofdiscretionary accrualstohidethefirm’s truesituation as a

meansformanagement tomaintain theirequity value

There-fore,manipulatedstatementsusedasinputswouldreducetheir

abilityforcashflowprediction.Ontheotherhand,management

mayusediscretionaryaccrualstosignaltheirtrueviewofthe

firm’sfuture.Insuchasituation,amanagement’sstatementsof

discretionaryaccrualscouldhaveabetterpredictiveabilityof

futurecashflow.Therefore,whetherdiscretionaryaccrualscan

enhanceoradverselyaffectcashflowforecastdependsonthe

motivationbehindthemanipulation,ifany,bymanagers

When managers are motivated to manipulate

non-discretionary information to suit their own selfish ends,

non-discretionary information could be revealing or masked,

makingthe variablesmoreor less informativeinforecasting

Anumberof researchershave suggestedmethods toidentify

discretionaryaccruals[seeJones(1991),Dechowetal.(1995)

dis-cretionaryproblem,accrualwouldresultinvaryingcoefficient

onrelevantvariables.DechowandGe(2006)showthatunder

differentaccruallevel, firmscashflowshavedifferent

persis-tenceandthecorrelationsbetweencashflowsandaccrualsalso

dependsonaccruallevel.AnotheroptionforManagerstodistort

financialstatementsinordertomeetreportinggoalsisthrough

realactivitiessuchasthosediscussedinRoychowdhury(2006)

Suchoperations includeprice discounts, reduction of

discre-tionaryexpendituresandsoon

Thisstudymakesanumberofcontributions.First,thescant

numberofstudiesoncashflowstudiesindevelopingeconomies,

suggest the need for more studies in the area as the results

elsewherearenotdirectlytransferableduetodifferentbusiness

andoperationalenvironments.Secondly,applicationofmodels

basedonstudieselsewheremaynotsuitadevelopingcountry

setting.Thirdly,giventhenatureandimportanceofcashflows

forfirms,investors,creditorsandpolicymakers,amongothers,

thecurrentstudyisamajoradditiontotheextantliterature

basicrelationship:

whereCFdenotescashflowfromoperation,SALEdenotessales,

ARdenoteschangesinaccountreceivable,Pdenotespurchase andAPdenoteschangesinaccountpayable.Termsinthefirst parenthesisrepresentscashinflowwhilethesecondparenthesis representingcashoutflow.Withseveralassumptionsonfirm’s operatingactivities,DKWmodelsuggeststhattheexpectation

offuturecashflowsiscurrentearnings:

useddirectlyasapredictor,shouldbedisaggregatedintocash flowandseveralaccrualtermsasthecomponentsprovide dif-ferentpredictivepowerforfuturecashflowsuggestedbytheir theoreticalmodelbelow:

whereINVdenoteschangesininventoryandothertermsareas definedpreviously.ComparedtoDKWmodel,theparameters

onthecomponentsofearningsareallowedtodiffer.AsBCN modelislessrestrictive,itoutperformstheDKWmodelincash flowprediction(seeBarthetal.,2001)

Beyond theseaccrualterms,Cheng andHollie (2008)and

non-core components in orderto examine the persistence of thesecomponentsincashflowprediction.InChengandHollie

wherethetermsinthefirstparenthesisarecashreceivedfrom sales,cashfromcostof goodssold,cashrelatedtooperating andadministrativeexpenses,cashusedtopaytax,cashfor inter-estpayment,andothercashcomponents,respectively.Thefirst threearedefinedascorecashcomponentsandthelatterthreeare non-core components.The cashflow disaggregationexplores additionalinformationthat could beobtained from cashflow statement under direct method disclosure Cheng andHollie thususedisaggregatedcashcomponentsandalsoaccrualterms definedinBCNmodelaspredictorsforcashflow.Theempirical resultssuggestthatsuchcashflowdisaggregationdoesimprove the accuracy of BCN model,however the effectis minor.In ChengandHollie’spaper,model(4)hasareportedR 2of39.83% andBCNmodel(withcashflowun-disaggregated)hasreported

R2of38.49%.InU.S.,beforethepublicationofSFASNo.95,

operationneededtobeestimatedusingbalancesheetandincome

statement.OrpurtandZang(2009)alsosuggestthatdirect

dis-closureofcashcomponentsprovidesincrementaleffectoncash

flowprediction.Theyintroduceaconceptofarticulationerror,

whichis defined as the difference between cashcomponents

estimatedusingbalancesheetandincomestatementandones

thatareactuallydisclosed(directmethod)instatementofcash

flow.Somestudiesreportthatthedirectmethodcashflow

state-mentismoreinformativeinpredictingcashflow(seeArthurand

setting).Forfirmsthatdonotprovide directmethodcash

dis-closure,suchincrementalbenefitcannotbeexploitedbutcash

disaggregationcanstillbedone.Inthispaper,wemodelSouth

AfricancashflowsbydisaggregatingcashflowfollowingOrpurt

CF t=CFIN t−CFOUT t−INT t−TAX t

= (SALE t−AR t) − (COG t+OE t+INV t−AP t)

−INT t−TAX t+AE t ,

(5)

wherenetoperatingcashflowisequaltocashinflow(CFIN)

minus cashoutflow (CFOUT whichis paid to suppliers and

employees)andinterest andtaxcashpayment (INT andTAX

respectively).Thefirstandsecondbracketsinthesecondlineof

Eq.(5)measurestheestimatedcashinflowandoutflow

respec-tively,usinginformationinbalancesheetandincomestatement;

interestandtaxpayment incasharealsoincluded, whichare

directlydisclosedindependentofthereportformat.The

differ-ence between disclosed actual cashflow from operation and

these four components is termed as articulation error (AE),

shownasthefinaltermintheequation

Thispaperstartswithacomparisonof twoempirical cash

flowpredictionmodelsfollowingOrpurtandZang(2009):

ModelI: CF t+1=β0+β1CF t+β2AR t+β3INV t

+β4AP t+β5DDA t

+β6INT t+β7TAX t+ε t+1 , (6)

ModelII: CF t+1=β0+β1CFIN t+β2CFOUT t+β3AE t

+β4DDA t+β5AR t+β6INV t

+β7AP t +β8INT t+β9TAX t+ε t+1 ,

(7) Allthevariablesarethesameaspreviouslydefined.βsare

parameterstobeestimated.ModelIistheempiricalBCNmodel

with the addition of interest and tax payment as predictors

ModelIIfurtherdisaggregatesthecashflow terminmodelI

CashinflowandcashoutflowareestimatedbasedonEq.(5)

ModelIIinprincipleissuperiortomodelIintwoways:

disaggre-gatingcashflowincorporatesinformationof otheraccounting

itemssuchassalesandallowscashcomponentstohavedifferent

persistence

Apartfromabovementionedlinearregressionmodels,this

paperalsointroduceautoregressivemovingaveragemodelwith

exogenous variables(ARMAX)developedby Whittle(1951)

(see Boxetal.,2008for details).Thisisatimeseries model andthispaperattempttoadaptittopaneldataapplication.The generalmodelARMA(1,1)withexogenousvariablesiswritten as:

whereεdenotesthemovingaverage(MA)termsandXdenotes exogenousvariables.Themovingaveragetermsdependsonthe model parameters,thereforetraditionalOLS methoddoesnot apply andwe usemaximumlikelihood estimators (MLE) for parameterestimation

In practicalapplication, amulti-period-aheadpredictionof cashflow maybeof greater importancethanasingleperiod Theabovementionedmodelsaresimpletoadapttosuch require-ment.Theusualwayistoadjustthelaglengthbetweentarget variableanditspredictivevariables.Forone-period-ahead pre-diction,asdiscussedabove,theexplanatoryvariablestakeone lag values.Thus,consideringtwo-period-aheadpredictionfor example,theexplanatoryvariablesshouldtakevaluesof their second lag.Thedisadvantageof suchmethodappearscritical for long term forecast where data covering a long period is demanded but notalways available Moreover,when the his-torical datais too outdated, the informationthey contain for predictionpurposesmaybelimited.Todealwiththisproblem,

weapplyvectorautoregressive(VAR)model(Sims,1980).VAR modelispresentedas(takeoriginalBCNmodelasexample):

⎡

⎢

⎢

⎢

⎣

⎤

⎥

⎥

⎥

⎦

=A+B

⎡

⎢

⎢

⎢

⎣

⎤

⎥

⎥

⎥

⎦ +et , (9)

whereAandeare 4×1 vectorsofconstantsanderrorterms respectively,Bisa4×4parametermatrix.ThereforeVARcould capturetheevolvementofnotonlycashflowbutalsoother pre-dictorsthroughtime.OncetheparametersAandBareestimated,

wecouldapplytheVARmodeloflagonetoforecastasmany periodsaheadasneeded.Theforecastequationiswrittenas:

⎡

⎢

⎢

⎢

⎣

1

⎤

⎥

⎥

⎥

⎦

=



1 0

A B

p

⎡

⎢

⎢

⎢

⎣

1

⎤

⎥

⎥

⎥

⎦

where0isa1×4vectorofzeros;pisthenumberofperiods aheadtobeforecast

In the laterempirical section, we willstart from compari-sonbetweenmodelIandmodelIIexaminingthe incremental powerofcashflowdisaggregationincashflowprediction.Also examined isthe effectof including twolags of predictorson models withinclusionofonlyonelagof predictors.Thenwe introduce ARMAXmodel.Withinclusionof movingaverage (MA)term,wecouldinthefirstplaceseewhetherthismodel providemoreaccurateforecastandsecondlyshowthedirection

Finally,we applyVARmodeltopredict two-year-ahead cash

flow,incomparisonwithregressionmodelwithcashflowasthe

onlydependentvariable

3 Data

ThisstudyusesdataforSouthAfricanfirmscollectedfrom

Datastream.Allfirmslisted onJohannesburgStockExchange

areincludedinthestudy.Cashflowdatabefore1994isnot

avail-able,sothesamplespansfrom1994to2012.Thefollowingitems

aregatheredforeachfirm:revenuefromsales(SALE),costof

goodssold(COST),selling,generalandadministrativeexpenses

(SGA),depreciation,depletionandamortization(DDA),account

payable (AP), account receivable (AR), inventory (INV), net

operatingcashflow(CF),interestpaidincash(INT),andtax

paidincash(TAX),11 variablesintotal Changes inaccount

payables,receivablesandinventory, cashinflow,cashoutflow

andarticulationerrorarenotprovideddirectlysotheyare

cal-culatedmanually.Selling,generalandadministrativeexpenses

(SGA)areusedtorepresentotherexpenses(OE)inEq.(5).All

variablesaredeflated byaverage totalassetsfor eachfirm in

ordertosmooththepotentialeffectoffirmsizes.Firmsthatdo

nothaveavailabledataforallthesevariablesareexcluded.The

wholesamplecontains192firmswith1021firm-year

observa-tions.Thepanelisunbalanced,i.e.firmshaveunequalnumber

ofobservations.Forregressionanalysis,alltheexplanatory

vari-ablesarelagged;thereforethefinalsampleisfurtherreduced

Formodels withonelag periodof explanatory variables, the

samplecontains791 firm-yearobservationsintotalandthere

are623firm-yearobservationsformodelestimation withtwo

lagperiodsofexplanatoryvariables

4 Empirical results

Itisworthnotingthataveragesalesis1.857timesaveragetotal

assets.Thestandarddeviationofsalesis1.748,suggestingthat

thesampleishighlydispersed.Netoperatingcashflowdeflated

by averagetotal assetshas a mean of 0.127 Eq (5) is used

toestimatecashinflowandoutflowandarticulationerror.The

estimatednetoperatingcashflowhasarelativelylowcorrelation

withthe disclosedactual value andthus cashflow estimated

indirectlymightcausesomenoiseincashflowforecast

Inthisstudy,allfirmsaretreatedas homogeneousandthe

modelsdescribedinprevioussectionareestimatedusingpooled

methods.Thesampleispartitionedintoestimationperiodfrom

1995to2009,andout-of-sampletestperiodfrom2010to2012

The estimation sample contains 579 firm-year observations

andthetestsample 212.The parametersestimatedin-sample

areappliedout-of-samplefor performancecomparison.R2 of

Table 1 Descriptive statistics.

Note:All variables are deflated by average total assets of each firm.CFINdenotes Cash inflow,CFOUTfor cash outflow, and AE is short for articulation error These three terms are calculated by Eq (6) Note there are three extra‘AE’s.

Due to the fact that not all firms report interest and/or tax paid in cash in their operative cash flow section, Eq (6) is not accurate for some firms I exam-ined three more calculations of net operating cash flow: 1.CFIN–CFOUT;2.

CFIN–CFOUT–INT;3.CFIN–CFOUT–TAX,assuming every firm must lie in one of the four groups Estimated cash flow using method 1 has the highest correlation with actual value (0.69) whereas the one deducting both interest and tax cash payment has the lowest correlation among the four (0.65) However, note that the mean of articulation error is much lower when Eq (6) is adopted Therefore, Eq (6) is still the default formula for articulation error.

differentmodelsarethencomparedtodeterminethebest pre-dictingmodel

ModelIissetasthebenchmarkmodel.ModelII disaggre-gatescashintodifferentcomponents.Theresultsareshownin

signifi-cance.Numbersinparenthesisaret-statisticsforeachparameter With regards to the bench mark model I, lagged cash flow, changesinaccountreceivable,changesinaccountpayable,and taxcashpaymentaswellastheintercepttermsarestatistically significant,whereaslaggeddepreciation,depletionand amorti-zation(DDA),changesininventory,andinterestcashpayment arenot statisticallysignificant.Thisisnotconsistentwiththe reportedresultsofBCNmodelusingU.SdataandFarshadfar

is no direct explanationwhy DDA should be usefulin cash flow forecast Changes ininventory andinterestpayment are itemsthatdirectlyrelatetocashflowbecausetheymeasure pro-portionsofcashoutflow.Forthesethreeitems,SouthAfrican datashowsmuchweakerpredictivepowerthanstudiesforother countries.R2 is0.54 for model I,whichismuchhigher than that reportedin studieson U.S data Forexample, in Cheng

morethan40% Ittherefore impliesthat SouthAfricanfirms have less variable cash flows than U.S firms Model II has higher adjusted R2 as cashdisaggregation incorporates extra information frombalance sheetand incomestatement How-ever,the out-of-sampletestshowsthat modelIIprovidesless accurate predictionthan model I The additionalinformation

Table 2

Regression results for one period lag models.

Lagged variables

Estimation period

(18.5)

( −18.37)

(6.33)

( −6.34) ( −5.95) ( −5.99) ( −8.41) ( −5.59)

( −5.51) ( −5.15) ( −5.17) ( −6.58) ( −4.81)

( −12.50) ( −13.85) ( −7.01) ( −10.61) ( −7.23) ( −6.72) ( −10.89)

Test period

Note:Asterisks indicates significance at levels of 5% Numbers in parentheses are corresponding t statistics.

Model I:CF t=β0 +β1CF t−1 +β2ΔAR t−1 +β3ΔINV t−1 +β4ΔAP t−1 +β5DDA t−1 +β6INX t−1 +β7TAX t−1 +ε t.

Model II:CF t=β0 +β1CFIN t−1 +β2CFOUT t−1 +β3AE t−1 +β4DDA t−1 +β5ΔAR t−1 +β6ΔINV t−1 +β7ΔAP t−1 +β8INX t−1 +β9TAX t−1 +ε t.

Model III:CF t=β0 +β1CF t−1 +β2SALE t−1 +β3COST t−1 +β4SGA t−1 +β5DDA t−1 +β6AR t−1 +β7INV t−1 +β8AP t−1 +β9INX t−1 +β10TAX t−1 +ε t Model IV:CF t=β0 +β1CF t−1 +β2SALE t−1 +β3COST t−1 +β4SGA t−1 +β5DDA t−1 +β6AR t−1 +β7AP t−1 +β8INX t−1 +β9TAX t−1 +ε t.

Model V:CF t=β0 +β1CF t−1 +β2SALE t−1 +β3COST t−1 +β4SGA t−1 +β5DDA t−1 +β6AR t−1 +β7AP t−1 +β8INV t−1 +β9TAX t−1 +ε t.

Model VI:CF t=β0 +β1CF t−1 +β2SALE t−1 +β3COST t−1 +β4SGA t−1 +β5DDA t−1 +β6AR t−1 +β7AP t−1 +β8INV t−1 +β9INX t−1 +ε t.

Model VII:CF t=β0 +β1CF t−1 +β2SALE t−1 +β3COST t−1 +β4SGA t−1 +β5DDA t−1 +β6AR t−1 +β7AP t−1 +β8TAX t−1 +ε t.

doesnotimprovethemodel’spredictivepower.Secondly,itis

noteworthythatchangesininventoryandinterestpaymentare

significantwhiletaxpaymentisinsignificantinmodelII.Model

IIappearstoprovidedifferentinferencesthanmodelI.To

fur-therexaminethisproblem,wedevelopmodelIIIbyreplacing

thecashdisaggregatedcomponentswithrevenuefromsales,cost

ofgoodssoldandselling,generalandadministrativeexpenses

(SGA):

ModelIII: CF t+1=β0+β1CF t+β2SALE t+β3COST t

+β4SGA t+β5DDA t+β6AR t

+β7INV t+β8AP t+β9INT t

+β10TAX t+ε t+1 (11)

ModelsIIIusesthesameinformationasmodel II,so their in-sample fitness and out-of-sample performance are almost the same Models III nonetheless suggests that tax payment

is significant,withchanges ininventory andinterest payment also significant From model I toIII,we seethat parameters

ofchangesininventory,interestpaymentandtaxpaymentare not consistent.This couldbe dueto collinearitybetween the explanatoryvariablesincludedinthemodels.ModelsIVtoVI thatinturnremovechangesininventory,interestpayment,tax paymentfrommodelIIIoneatatimearethenbuilttoexamine theeffectsoftheseexcludedvariables.Whenchangesin inven-toryareremovedasinmodelIV,theothertwovariablesareboth significant.Theout-of-sampleperformanceishighlyenhanced Therefore,changes ininventoryactually deterioratethe mod-elswhenweincludevariablesfromincomestatement.Thecase

Table 3

Comparison of sum squared errors of model I and VII with different division of

sample.

a

b

Note:the numbers in the table are sum squared errors of out-of-sample prediction

for model I and VII Bold numbers denotes they are the smaller of the two models

during different test year, suggesting the model is better than the other in that

particular test year.

issimilarforinterestpayment,whichisremovedinmodel V

TheR2intestperiodalsoincrease,thoughnotasremarkably

asmodelIV.Whentaxpaymentisremoved,modelVIreports

declinedout-of-sampleR2,suggestingthattaxpaymentcouldbe

actuallyprovidingincrementalinformation,withoutwhichthe

model’spredictiveperformanceisharmed.Inaddition,changes

ininventoryinmodelVIbecomeinsignificantbeingconsistent

withmodelI.Fromthecomparisonofthese6models,itcanbe

concludedthat changesininventory andinterestpayment are

notveryinformativeincashflowforecasting,andevenworse,

wemaysignificantlyreducethepowerofthemodelsbyadding

extravariablesfrom incomestatement Asaresult, for

mod-elsincludingsales,cost,andotherexpenses,itisbeneficialto

removechangesininventoryandinterestpaymentfromtheset

ofexplanatoryvariables.ThisgivesmodelVII:

ModelVII: CF t+1 =β0+β1CF t+β2SALE t+β3COST t

+β4SGA t+β5DDA t+β6AR t

+β7AP t+β8TAX t+ε t+1 (12) TheestimationresultsarealsoreportedinTable2.Thesigns

ofparametersareconsistentwithexpectationforcost,expenses,

andchangesinaccountpayablesastheyarenegativelyrelated

withnetcashflow.Notethatselling,generalandadministrative

expenses’effecton one-period-aheadcashflow isof slightly

lowermagnitudethancostofgoodssold.ModelVIIhashigher

out-of-sampleR2thanmodelIItoVI,validatingthepointthat

removingchangesininventoryandinterestpaymentwhensales,

cost,andSGAarepresentprovidessuperiormodel.Howeverthe

R2isstilllowerthanthatofmodelI,despitetheadditionofextra

information.Thetestperiodsamplecontainsthreeyearsofdata,

2010,2011and2012.ModelVII,duetoitsinclusionofmore

variables,shouldhavestrongerpredictivepower,atleastduring

theyearthatimmediatelyfollowstheendofestimationsample

Therefore,modelVIIisexpectedtooutperformmodelIatleast

during2010.Toexaminethispoint,sumsquarederrors(SSE)

arecalculatedforeachyearintestperiod,whichareshownin

out-of-sampleprediction LowerSSEmeanshigher accuracy,

andaredisplayedinbold.TheresultsshowthatmodelVII

pro-videsabetterperformanceatthestartofthepredictionperiod

butperformlesswellinlateryears.Thisphenomenoncouldbe duetothechangingenvironmentthroughtime,whichmayresult

inlittlepersistenceoftheestimatedparameters.Ifthiswerethe case,theparameterswouldbebettertobeupdatedyearbyyear

Inattempttoarguethispoint,thesampleisthenfurther parti-tionedwiththeperiod1995–2010asestimationperiodand2011 and2012astestperiod.TheSSEfor2011and2012are recal-culatedandtheresultsareshowninTable3sectionb.Similar

tothereportedresultsinpanela,modelVIIdoesperformbetter

in2011butworsein2012.ThebenchmarkmodelIissimpler andmorepersistentlyeffectivethanmodelVIIwhilemodelVII could provide betterone-period-aheadpredictionifestimated withupdateddata

Priorstudiesfocus mainlyononelagperiodofpredictors, butitisnaturaltoaskwhetherfurtherpastexplanatoryvariables containpotentialinformationthatareincrementaltocashflow prediction.ModelsIandVIIareseparatelyestimatedtwice,the firsttimewithonelagofpredictorsandthesecond timewith twolags,sowecouldcomparethedifferencemadebyinclusion

ofonemorelagperiodofpredictors.Itisexpectedthat mod-elswithtwo lagsofindependentvariablesshouldperformno worsethantheirone-lagcounterparts.Theresultsareshownin

dataarerequired.Table4showssomesignificantdifferencesto thereportedresultsinTable2.Forexample,foronelagversion

ofmodelI,changesininventoryandinterestpaymentbecome significant.ItcanbeseenthatmodelVIIthatexcludesthetwo variableshaslowerin-samplefitbasedonR2fortheonelag ver-sion.AdditionalvariablesaddedtomodelVIIfromtheincome statement becomeinsignificant.The out-of-sampleR2 for the fourmodels areclose ModelVIIunderperformsmodel I,no matterhowmanylagsareincluded.However,itisnoteworthy thatformodelI,out-of-sampleperformanceoftwo-lagmodel

isslightlyworsethanone-lagmodelwhileformodelVII, two-lagmodelismarginallybetterthanone-lagmodel.Sumsquared errorsfordifferentyearsarealsocomparedinTable5.The differ-encebetweenmodelIandVIIhasdeclinedremarkably.Model VIIoutperforms model Ifor 2010,thestart of theprediction period,whichisconsistentwiththeresultsreportedinTable3

butunder-performsforthefollowingtwoyears.Inaddition,for modelI,two-lagmodelisbetterthanone-lagmodelin2010,but

itistheoppositefor2011and2012,whileformodelVII, the two-lagmodelisgenerallybetterthanone-lagmodelexceptin

2012.Itcanbetoldfromthesecomparisonsthatfornearfuture prediction, model VIIisabettermodel, especiallywhen two lagsofexplanatoryvariablesareincluded

In summary, this section examines different combinations

of predictorsin one-period-ahead cashflow forecasting The benchmark model uses least variablesandits performanceis consistentinsampleandoutofsample.DDA,changesin inven-tory andinterest paymentare insignificant,whichis different fromstudiesinothercountries.Modelsincorporatingincome statementinformationseemtomakeworseout-of-sample pre-diction.However,whentwolagsofindependentvariablesare included in prediction models, results differ in several ways whichsuggeststhat variableinclusion incashflowmodeling shouldbetreatedwithcare

Table 4

Comparison of model I and VII with different lags of independent variables.

Model I, one lag Model I, two lags Model VII, one lag Model VII, two lags

Lagged variables

Estimation period

Test period

Note:Asterisks indicates significance at levels of 5% Numbers in parentheses are correspondingtstatistics.

L1 and L2 denotes the variable of first lag (timet− 1) and second lag (timet− 2) respectively Model I and model VII are compared when different periods of lagged predictors are applied.

Table 5

Sum squared errors for model I and VII with different lags of independent variables.

Note:the numbers in the table are sum squared errors of out-of-sample prediction for model I and VII Bold numbers denotes they are the smaller of the two models during different test year, suggesting the model is better than the other in that particular test year.

ARMAXmodel explores thepersistence of boththe main

variableandtheerrorofthemodelovertime.Itseemsthatthe

linearregressionmodeldescribedabovecanbetreatedasfirst

orderautoregressivemodelwithexogenousvariables,orsimply

ARX(1,1).One-period-laggedcashflowisthefirstorder

autore-gressivevariableandtheotheraccountingvariablesaretreated

asexogenousvariablesthattakeafirstorderlaggedvalue.For

modelIandmodelVII,weincludeafirstordermovingaverage

term.Thetwomodelsforcomparisonarespecifiedasbelow:

ModelIwithMA(1): CF t+1=β0+β1CF t+β2AR t+β3INV t+β4AP t+β5DDA t+β6INT t

+β7TAX t+β8ε t+ε t+1 , (13) ModelVIIwithMA(1): CF t+1 =β0+β1CF t+β2SALE t+β3COST t+β4SGA t+β5DDA t+β6AR t+β7AP t

+β8TAX t+β9ε t+ε t+1 , (14)

whereε t denotesthemoving averageterm,anditscoefficient will measureitsrelationshipwithfuturecashflow.Intuitively speaking,themovingaveragetermistheunexpectedshockfor lastperiodcashflow.Therefore,itsparametersmeasureshow shocksinfluencenextperiodforecast.Weusemaximum likeli-hoodestimation(MLE)toestimatetheparameters.Thismethod calculatesthejointlikelihoodforallobservations,andthe esti-mator finds the optimum values for the coefficients that can maximize the likelihood.Assume the conditionalexpectation

Table 6

Comparison of linear regression model and ARMAX(1,1) model for model I and model VII.

Lagged variables

Estimation period

Test period

Note:Asterisks indicates significance at levels of 5% Numbers in parentheses are correspondingtstatistics.

Model I with MA(1):CF t=β0 +β1CF t−1 +β2AR t−1 +β3INV t−1 +β4AP t−1 +β5DDA t−1 +β6INX t−1 +β7TAX t−1 +β8ε t−1 +ε t.

Model VII with MA(1):CF t=β0 +β1CF t−1 +β2SALE t−1 +β3COST t−1 +β4SGA t−1 +β5DDA t−1 +β6AR t−1 +β7AP t−1 +β8TAX t−1 +β9ε t−1 +ε t.

ofcashflowisi.i.dbyGaussiandistributionwithconstant

vari-anceσ2.ThecoefficientsobtainedarereportedinTable6.When

the moving average (MA) term is introduced, parameters in

modelI havenotchanged much,andthe MAterm isof low

valueandinsignificant.Out-of-sampletestfor modelIshows

thattheadditionofMAtermhasdonenogoodtothe

predic-tivepowerandhasratherhadanegativeeffect.TheMAtermin

modelVIIisstatisticallysignificant.Thepositivesignimplies

thatcashflowshocksofoneperiodtendtopersistinthecash

flowofthe nextperiod.Howeverthe additionofMAterm to

modelVIIprovidesaworseout-of-sampleperformanceR2of

0.43comparedto0.49withoutMAterm

Intheprevioustwosections,modelsaredevelopedfor

one-period-aheadcashflowprediction.Thefundamentalideaisto

findtherelationshipbetweencashflowandone-period-lagged

explanatoryvariables.ModelIandmodelVII,asstated

previ-ously,canbetreatedasautoregressive modelwithexogenous

variable.VARmodelsaysthatallexplanatoryvariablescanbe

consideredasavectorofendogenousvariablesthatdependon

vectors of their laggedvalues In this way,the whole vector canbeforecastrecursively.In thissection,lag-tworegression andVARformofmodelIandmodelVIIareexaminedin two-period-aheadprediction.Theestimationperiodisfrom1996to

2009,andtestperiodis2011and2012.Thereisareasonfornot testingperformancein2010.When2010cashistobeforecast, eitherlag2modelorVARmodelrequiresdatain2008asinput ForVARmodel,2008datafirstisusedtopredict2009,andthen

weusethe2009predictiontofurtherpredict2010.Recallthat datafrom2008to2009isalreadyusedinparameters estima-tionhenceVARmodelfor2010predictionisnotpurelyoutof thesample,andthecomparisonwouldhavemisleadingresults With thisconcern, year2010 areexcluded fromcomparison Themainconcernofthissection istocompareout-of-sample predictivepowerofdifferentmodels,soonlythecomparisonof sumsquarederrorsforeachyeararelistedwhereasregression resultsarenotreported

and 2012 separately The result is obvious that VAR model for model I hasoutperformed the otherthreemodels inboth years For both model I and model VII, regression model and VARmodel havevery similarresults in 2011,but VAR

Table 7

Comparison of sum squared errors for model I and VII in two- period-ahead

prediction.

Model ILag 2 Model IVAR Model VIILag 2 Model VIIVAR

Note:the numbers in the table are sum squared errors of out-of-sample prediction

for model I and VII Bold numbers denotes they are the smaller of the two models

during different test year, suggesting the model is better than the other in that

particular test year.

form is much better in year 2012 prediction It is simple

andstraightforwardtoextendthepredictiontolongerperiods,

and the result from this test is very encouraging for further

studies

5 Conclusion

Previous studieson cash flow prediction mainlyfocus on

theinformationprovidedinfinancial statements.Studiessuch

and how disaggregating accruals into its major components

enhancescashflowprediction.Thisstudyappliesanumberof

modelstocashflowdataforSouthAfricanfirms.Threemain

cashflow modelsareinvestigatedinthisstudy,mainly, linear

regression,whichhasbeenwidelyadoptedinsimilarand

rele-vantstudies,movingaveragemodel,whichismostlyappliedin

time-seriesanalysis,andvectorautoregressivemodel,thathas

beenwidelyappliedinmacroeconomicsandfinance.Thelatter

twotypesofmodelsareappliedforthefirsttimetocashflow

prediction

Theresultsreportedinthisstudycontrastthatreported

else-where Disaggregating cash flows into its major components

doesnot appeartoenhancecashflowpredictionforthe

aver-ageSouthAfricanfirmscomparedtoresultsreportedbyBarth

Australianfirms.Theresultssuggestthatimplicationsof

stud-ies conducted elsewhere cannot be extrapolated across other

countrieswithouttakingintoaccountcountrycontextand

dif-ferences.The reportedresultsinthisstudy show thatmodels

incorporatingincomestatement informationseemtoresult in

worse out-of-sample prediction However, when two lags of

independentvariablesareincludedinpredictionmodels,results

differ inseveral ways whichsuggests that variable inclusion

incashflowmodeling inSouthAfricashouldbetreatedwith

caution In addition, predictionaccuracy, as measured by R2

for South African firms are high compared to extant studies

elsewhere.Studiesoncashflowpredictionpoolallfirms’data

togetherandignoreheterogeneitythatexistsamongfirmsand

industry Thereisthereforethe needfor studiesoncashflow predictionthatfocusonindustryandindividualfirms

Acknowledgement

Theauthorsappreciatethecommentsbytheparticipantsat AfricanAccountingandFinanceAssociation3rdAnnual Con-ference,2013,UgandaandthefinancialsupportbyKelvinSmith scholarshipfromUniversityofGlasgow

References

Arthur, N., Chuang, C., 2006 Ias 7 Alternative Methods of Disclosing Cash Flow from Operations: Evidence on the Usefulness of Direct Method Cash Flow Disclosures Working Paper The University of Sydney.

Badertscher, B.A., Phillips, J.D., Pincus, M., Rego, S.O., 2009 Earnings man-agement strategies and the trade-off between tax benefits and detection risk:

to conform or not to conform? Account Rev 84, 63–97.

Barth, M., Cram, D., Nelson, K., 2001 Accruals and the prediction of future cash flows Account Rev 76, 27–58.

Box, G.E., Jenkins, G.M., Reinsel, G.C., 2008 Time Series Analysis: Forecast-ing and Control John Wiley & Sons, Hoboken, New Jersey.

Brochet, F., Nam, S., Ronen, J., 2008 The Role of Accruals in Predicting Future Cash Flows and Stock Returns Working Paper, Available at SSRN 1126022 Cheng, C., Hollie, D., 2008 Do core and non-core cash flows from operations persist differentially in predicting future cash flows? Rev Quant Finance Account 31, 29–53.

Dechow, P., Kothari, S., Watts, R.L., 1998 The relation between earnings and cash flows J Account Econ 25, 133–168.

Dechow, P.M., Ge, W., 2006 The persistence of earnings and cash flows and the role of special items: implications for the accrual anomaly Rev Account Stud 11, 253–296.

Dechow, P.M., Sloan, R.G., Sweeney, A.P., 1995 Detecting earnings manage-ment Account Rev 70, 193–225.

Farshadfar, S., Monem, R., 2013 Further evidence on the usefulness of direct method cash flow components for forecasting future cash flows Int J Account 48, 111–133.

Francis, R.N., Eason, P., 2012 Accruals and the nạve out-of-sample prediction

of operating cash flow Adv Account 28, 226–234.

Jones, J.J., 1991 Earnings management during import relief investigations J Account Res 29, 193–228.

Kothari, S.P., Leone, A.J., Wasley, C.E., 2005 Performance matched discre-tionary accrual measures J Account Econ 39, 163–197.

Lorek, K.S., Willinger, G.L., 2010 Time series versus cross-sectionally derived predictions of future cash flows Adv Account 26, 29–36.

Ogneva, M., 2012 Accrual quality, realized returns, and expected returns: the importance of controlling for cash flow shocks Account Rev 87, 1415–1444.

Orpurt, S., Zang, Y., 2009 Do direct cash flow disclosures help predict future operating cash flows and earnings? Account Rev 84, 893–935.

Pae, J., Yoon, S.-S., 2012 Determinants of analysts cash flow forecast accuracy.

J Account Audit Finance 27, 123–144.

Roychowdhury, S., 2006 Earnings management through real activities manip-ulation J Account Econ 42, 335–370.

Sims, C.A., 1980 Macroeconomics and reality Econometrica 48, 1–48 Whittle, P., 1951 Hypothesis testing in time series analysis Hafner Publishing Company, New York.