See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/52008212

Forecasting: Methods and Applications

Chapter  in  Journal of the Operational Research Society · January 1984

DOI: 10.2307/2581936

CITATIONS READS

1,512 21,481

3 authors, including:

Spyros Makridakis Rob J Hyndman

University of Nicosia Monash University (Australia)
187 PUBLICATIONS   8,605 CITATIONS    239 PUBLICATIONS   13,049 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Stress-testing algorithms: generating new test instances to elicit insights View project

predicting medical risks View project

All content following this page was uploaded by Spyros Makridakis on 07 January 2015.

The user has requested enhancement of the downloaded file.

Forecasting: Methods and Applications by Spyros Makridakis; Steven C. Wheelwright; Rob J.
Hyndman; Elements of Forecasting by Francis X. Diebold
Review by: William M. Briggs
Journal of the American Statistical Association, Vol. 94, No. 445 (Mar., 1999), pp. 345-346
Published by: Taylor & Francis, Ltd. on behalf of the American Statistical Association
Stable URL: http://www.jstor.org/stable/2669717 .
Accessed: 05/01/2015 13:40

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

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

Taylor & Francis, Ltd. and American Statistical Association are collaborating with JSTOR to digitize, preserve
and extend access to Journal of the American Statistical Association.

http://www.jstor.org

This content downloaded from 194.221.86.126 on Mon, 5 Jan 2015 13:40:57 PM

All use subject to JSTOR Terms and Conditions
Book Reviews 345

the variance-covariance matrixof the errorsis completelyknown,which given by each. Both textsdo a nice job of detailingthe organizational
is nevertruein practice.It is possibleto give a moredetailedtreatment of difficulties and practicalitiesof businessforecasting(forone thing,telling
asymptotictheoryand its usefulnesswithoutdivinginto measuretheory thestudentto keep in mindthatthegoal of theforecastis to aid decision
and regularity conditions.But the book covers the centrallimittheorem making,notto overwhelmsomeonewithclevermathematical models).Fi-
(CLT) onlyin the Appendixof Chapter2. The proofof theCLT belongs nally,bothtextspresentnumerouspracticalhomeworkproblems,although
in the Appendix,but a discussionof it shouldbe moreprominent. MWH has manymore.DB has an inordinatenumberof problemswhose
A second exampleof thebook's shortcomings is in the motivationand onlytask is to "discuss" (makingthesedifficult to grade).
presentation of the F statistic.The F statisticfirstappears in Chapter4, To highlightthe differences, DB has additionalmaterialon forecast
whereit is introducedwithoutany motivation.Where does thatformula graphics,and a moreextensivediscussionof unitrootsand Dickey-Fuller
come from?The answerarrivesin Chapter7 withinthecontextof Wald, tests.It also givesmorespace to ARCH and GARCH models,increasingly
LM, and LR tests.The F statisticis revisitedas an alternative to theWald, important toolsthatare onlymentionedin MWH. DB includesa fullerdis-
LM, and LR tests,when in fact it is proportionalto the Wald test!No cussion of forecastaccuracymeasures,devotingan entirechapterto this
mentionis ever made thatthe numerator of the F statisticappears again important topic. (Afterall, if a forecastis to be used in a decisionit is a
and again in the derivationof theWald, LM, and LR tests.Making these good idea to knowhow well it has performed.) This is whereDB is at its
kindsof linksis important forconveyingunderstanding to students. best (no doubtbecause theauthorhas done originalresearchin thisarea).
The bottomline is this.Thereis definitely a usefulrole foran introduc- MWH coversmoretopicsthanDB, suchas an extensiveoverviewof de-
torygraduateeconometricstextthatis less technicaland moreaccessible compositionmethodsincludingthecensusX-12-ARIMA,Loess, and STL
to studentswho do nothave extensiveundergraduate econometricsback- techniques.It also providesa fullerexplanationof regressiontechniques,
ground.It is essentialthatsuch a textavoid being purelyempiricaland includingmulticollinearity, econometricmodels, multipleregression,re-
cookbook in style.Econometricsrepresentsa step in thatdirection,but gressionwithARIMA errors,dynamicregression, and evennonparametric
further stepsare needed,includingsome morecarefulediting. local regression.It presentssectionson stationarity and Portmanteau tests.
MWH also includessectionson intervention analysis,state-spacemodels,
TimothyJ.VOGELSANG nonlinearmodels (otherthanneuralnets),and multivariate autoregressive
Cornell University models.A chapteris devotedto thespecial difficulties facedin long-range
forecasting, and a nice chaptercoversjudgmentalforecasting, includinga
REFERENCES discussionof how to combinejudgmentaland statisticalforecasts.Finally,
Davidson,R., and MacKinnon,J. G. (1993), Estimationand Inferencein an entirechapteris devotedto theresultsfromtheM competitions.(This
Econometrics,Oxford,U.K.: OxfordUniversity Press. is not surprising, since Makridakisarrangedthese;for those unawareof
Greene,W. H. (1997), EconometricAnalysis(3rd ed.), Saddle River,NJ: thesecompetitions, theywere an organizedeffortto investigatehow well
PrenticeHall. varioustimeseries methodsperformin practice.)
J.,and DiNardo,J.(1997), EconometricMethocds
Johnston, (4thed.), New DB has manyof the typicalflawsof a firstedition,whichsometimes
York:McGraw-Hill. make the book difficult to follow.For example,thereare a numberof
typos,some of which could be misleadingto students.There are also
misinterpretations that should have been caught.For example,on page
Forecasting: Methods and Applications (3rd Ed.). 25 adjustedR2 is said to compensateforthe possible overfitting caused
"(when)a varietyof righthand-sidevariablesare triedand the'bestmodel'
Spyros MAKRIDAKIS, Steven C. WHEELWRIGHT,and Rob J.
HYNDMAN. New York: Wiley, 1998. ISBN 0-471-53233-9. xiv + is selected."
642 pp. $87.95. Two stylisticcriticismsare the choice of fontused for the equations
and theoveruseof footnotes.The eye-straining mathematical fontis rem-
iniscentof thatfoundin some commercialword-processing programs.At
Elements of Forecasting. times,it is confusingor impossibleto understandor interpret the given
equation.(On p. 22, forexample,thereadercannottell whethertheT - 1
Francis X. DIEBOLD. Cincinnati,OH: South-Western College Pub-
belongs inside or outside of the radical sign in the definitionof sample
lishing,1998. ISBN 0-538-86244-0.xiii + 392 pp. $76.95.
standarddeviation.)The footnotesinterrupt the flowof reading,and it is
These two textbooksare aimed at identical audiences: (advanced) not always clear whythe information in themis in a note ratherthanin-
undergraduate- or masters-levelbusiness studentswho have an interest cluded in themaintext.Moreover,theauthoroftenrepeatshimselfin the
in or will be usingforecastingtechniquesas partof theirjobs. The book footnotes(and in thetext),givingthesensethatthebook was notsubjected
by MakridakisWheelwright and Hyndman(MWH) is in its thirdedition, to close editing.The overallreadabilityand organizationof the textgive
whereasthatof Dielbold (DB) is a newcomer. the same feeling.
Both books reviewthe major highlightsof businessforecastingfrom A majorcriticismis thatDB oftenintroducestermsand phraseswithout
a modelingand organizationalperspective.Both intentionally leave aside definition or explanation.On page 163, thereaderis told thatall rootsof
theoreticaldevelopmentsand concentrateon practicalaspects. There is a certainpolynomialshouldbe "inside the unitcircle."The readernever
a greatdeal of overlapbetweenthe two books and one should consider learns what a unit circle is and how a root mightlie there(recall that
buyingone or theother,butnotboth.Here I outlinethematerialcovered the intendedaudience for thisbook is businessstudents).On page 197,
in bothbooks first,thenprovideby individualcritiquesand a discussion the operatorP is used withoutexplanation.The experiencedreaderwill
of unique materialpresentedin each. understand through thecontextthatP is meantas a projectionoperator, but
Both books assume verylittlebackgroundknowledgeof the student. theuninitiated mighteasily confuseit witha probabilitystatement. Later,
However,at least one priorcourse in statisticswould benefitthe student on page 296, anotheruse forP is given(to indicatea polynomial).Finally,
greatly.DB has more introductory materialthanMWH, such as expla- on page 331, in the contextof introducingneural nets,the "squashing
nationsof means and standarddeviations.Both books presentsmoothing function"(normallycalled theactivationfunction)is introducedbutnever
methods,althoughhere (and mosteverywhere else) MWH's explanations explained(neitherexplainedis thenew and odd terminology).
are of greaterdepthand clarity.Regressiontechniquesforforecastingare In its favor,DB is the firstbook I have seen with an entirechapter
covered,as are model and variableselectionand testing.Of course,both devotedto theimportant area of forecastinggraphics.The author'sinspi-
books cover ARIMA modelingin some detail. DB builds up to a full rationis drawnmainlyfromtheworksby Tufte.DB also bringsin theuse
ARIMA model more slowly,takingseveral chapters.MWH succinctly of cumsumchartsforregressionmodel parameters.However,one is left
coversARIMA and seasonal ARIMA. Both do an adequatejob withstan- wonderingwhytheauthordid nottakehis own advice,because thebook's
dard forecastinggraphics(such as ACF plots),althoughDB goes further latterhalfcontainsmorethana dozen examplesof less thaninspirational
(but has some problems).Neural nets are mentionedin both texts,with graphics.Justto cite one example,on page 138 the readeris treatedto
MWH givingthe fullertreatment. The interestingtopic of ex-postfore- the ACF and PACF of a whitenoise process (to imaginewhatthislooks
cast performance is also takenup by bothbooks,withsome uniquetwists like,draw a graphwithnothingon it). This phenomenacould have better

This content downloaded from 194.221.86.126 on Mon, 5 Jan 2015 13:40:57 PM

All use subject to JSTOR Terms and Conditions
346 Journalof the AmericanStatisticalAssociation,March 1999

explainedthroughwordsthanthrough pictures.And severaltimes(e.g.,p. havebegunto covertheseissues.Unfortunately, mostbooks considerthese

238), he plots a historicaltimeseriesand forecaston one figure,followed issues onlybriefly.As researchon unitroottestsand cointegration reaches
immediately by an absolutelyidenticalfigurebutwiththeinclusionof the maturity, books devotedexclusivelyto these two topics are now being
realizationsplottednextto the forecasts.These could have all been col- published.TimeSeries, UnitRoots, atndCointegrationis one such book
lapsed to one figure.DB does includea disk containingall of thedata for (withmoreemphasison cointegration thanunitroottests).
the examplesin thebook. Otherbooks cover unitroots and cointegration extensively(Hamilton
MWH is a comprehensiveoverviewof the fieldof practicalbusiness 1994; Reinsel 1997). One of the distinctivefeaturesof Dhrymes' book
forecasting.As this book is in its thirdedition,most typos have been is its emphasison theoreticaldevelopmentand probabilisticfoundations.
eliminated,and theflowis well designedfora textbook.Like DB, it pro- In its theoreticaldevelopment,Dhrymesfocuses on a narrowerclass of
vides data,butin theformof a web siteinsteadof a disk.There are some nonstationary models. For example,unitroot testsfocus mainlyon the
minorflaws,butnothingthatwouldproveespeciallyproblematicfora stu- case of iid errors,and cointegration concentrateson vectorautoregres-
dent.Most mistakesare confinedto usingtermswithoutexplanation(e.g.,
sions and integratedmovingaverages.However,the in-depthdiscussion
on p. 280, Cp and AIC are used withoutdefinition, althoughtheirpurpose
d of a narrowlyfocusedclass of models makes it easier to understandthe
is clear, and on p. 210, = is includedwithno explanation).Something
fundamentalsof these two topics. It also helps preparereadersfor the
thatdoes not actuallyqualifyas a problemis the chapteron long-range
studyof a widerclass of models,includingthose withdependenterrors.
forecasting. MWH offerssome interesting insights,mainlysuggestingthat
forthelong If one wantsa book withsound fundamentals and theoreticalaspects of
the"usual" methodologiesdo notworkwell whenforecasting
term.It is interesting thatlong-rangetime series data are just data, so cointegration and unitroottests,thisis thatbook.
thatthestandardmethodologiesshouldwork.(Afterall, an ARMA model Dhrymesbegins with an introduction to time series, estimation,and
cannot"sense" thetimeperiodsinvolved.)Why theydo not,of course,is prediction.These topics are preliminaryto cointegrationand unit root
thatthestandardassumptions(like stationarity) are no longervalid forthe processes.Unitrootprocessesand relatedasymptoticpropertiesare then
longterm.This shouldemphasizeto thestudentthatall timeseriesmodels presentedin a systematicmannerin Chapter3 by makinga distinction be-
are only approximations and not explanatorymodels. It shouldalso give tweentrueand operationalmodels,especiallyinvolvinga constantterm.
the studentpause when applyingshort-term models,suggestingthatthe Althoughthe treatment is thoroughfor the case of iid errors,some dis-
resultsare notnecessarilyto be trusted.Perhapslost on thestudentis that cussion about autocorrelatederrorswould be desirable.Various aspects
most long-rangeforecastsare pure speculation(includingsome curious of cointegration are discussed in the next threechapters,whichpresent
forecastsprovidedby MWH itselfon pp. 474-476). Anyonedoubtingthis rigorouscharacterization and implicationsof cointegration. Constrained
shouldtake noteof WilliamSherden'snew book (Sherden1998) review- estimationof thecointegrating vectorsbased on workof Ahn and Reinsel
ing thebusinessof economicforecasting, whicharguesthatmostbusiness (1990) and Johansen(1988) is discussed,and a connectionto canonical
and economicforecastsare of littleor no value. variablesis made. A conformity test for cointegrationbased on an un-
To continuethe discussionof long-rangeforecasts,MWH includes a constrainedestimatoris discussedextensively. Probabilitytopics such as
chapteron judgmentalforecasting and adjustments thatprovidesan excel- Brownianmotion,stochasticintegration, and invarianceprinciplesare es-
lentreviewof all thatcan go wrongwhenhumanbiases influence forecasts. sential to understanding cointegration and unit root processes. The last
And,as mentioned earlier,MWH includesan entirechapteron resultsfrom threechaptersprovidea good overviewof these topics and containma-
M competitions. Not much surprising here,but thisis a greatchapterto terialpertinentto cointegration and unitroot processes. These chapters
enable a studentto put all previouslylearnedtheoryintoperspective.The come in handyreadingChapters3-6.
studentmay be shocked to learn thatthe more complex methods(like
It is worthcomparingDhrymes'book to the recentbooks by Hamil-
ARIMA models) are oftenoutperformed by simplemodels (like smooth-
ton(1994) and Reinsel (1997), whichalso covercointegration extensively.
ing). It is fascinatingthat,althoughtheresearchcommunity is well aware
Hamiltonprovidesa "grocerystore" of time series analysis,coveringa
of thesefacts,thishas notsloweddowntheeffort expendedbuildingnewer
vast rangeof topicswithbreadth,whereasReinsel centerson analysisof
and morecomplexmodels withno real evidencethatperformance is im-
multivariate timeseries.Tine Series, UnitRoots,and Cointegration com-
proving.
MWH does notdevotemuchmaterialto unitroottests,whichis surpris- plementsthese two books, and vice versa. As notedin the Preface,this
ing consideringtheinordinateamountof timeand researchthateconome- book can be used as a secondarytextbookor a referencein an advanced
triciansdevoteto thistopic.MWH does a nicejob buildingand explaining graduate-leveltimeseries or econometricscourse. It can also be used as
in simplelanguagethecompleteseasonal ARIMA model,a techniquethat a primarytextbookin a courseon unitrootsand cointegration. This book
will surelyfindpracticaluse. shouldbe includedin a requiredreadinglist for those who studymulti-
Overall,Forecasting:Methodsand Applicationsis a maturetextwith variatetimeseriesor econometrics.
thoroughcoverageof essentialand important techniques,containsa good
mix of problemsets, and should serve well as a textbook.Elementsof Sung K. AHN
Forecastingis novel and interesting because of the new ideas in graphics State University
Washington
and practicaladvice thatit presents.

WilliamM. BRIGGS REFERENCES

Cornell University Ahn, S. K., and Reinsel,G. C. (1990), "Estimationfor PartiallyNonsta-
tionaryMultivariateAutoregressive Models," Journalof theAmerican
REFERENCE StatisticalAssociation,85, 813-823.
of theEstimatesfor
Dickey,D. A., and Fuller,W. A. (1979), "Distribution
Sherden,W. A. (1998), The FortuneSellers: The Big Business of Buying
Autoregressive Time Series Witha UnitRoot,"JournaloftheAmerican
and SellingPredictions,New York:Wiley.
StatisticalAssociation,74, 427-431.
(1981), "LikelihoodRatio StatisticsforAutoregressive
Time Series
Witha UnitRoot,"Econometrica,49, 1057-1072.
Engle,R. F., and Granger,C. W. J. (1987), "Cointegrationand ErrorCor-
Time Series, Unit Roots, and Cointegration. rection:Representation, Estimation,and Testing,"Economnetrica, 55,
PhoebusDHRYMES. San Diego: Academic Press, 1997. ISBN 0-12- 251-276.
214695-6. xiv + 524 pp. $89.95. Hamilton,J. D. (1994), Time Series Analysis,Princeton,NJ: Princeton
UniversityPress.
Unit root tests and cointegrationhave been two of the most heavily Johansen,5. (1988), "StatisticalAnalysisof Cointegration Vectors,"Jour-
researchedtopicsin timeseriessincetheworkof Dickey and Fuller(1979, nal ofEconomicDynamicsand Control,12,231-254.
1981) and Engle and Granger(1987). Because of theimportanceof these Reinsel, G. C. (1997),ElementsofMultivariateTimeSeriesAnalysis,New
topics in theoryand practice,recenttimeseries and econometricsbooks York: Springer-Verlag.

This content downloaded from 194.221.86.126 on Mon, 5 Jan 2015 13:40:57 PM

All use subject to JSTOR Terms and Conditions
View publication stats