Interactive forecasting: univariate and multivariate methods PDF

Preview

Summary

Interactive Forecasting: Univariate and Multivariate Methods. by Spyros Makridakis; Steven C. Wheelwright Review by: Dean W. Wichern Journal of the American Statistical Association, Vol. 74, No. 365 (Mar., 1979), pp. 246-247 Published by: Taylor & Francis, Ltd. on behalf of the American Statisti...

Description

Interactive Forecasting: Univariate and Multivariate Methods. by Spyros Makridakis; Steven C. Wheelwright Review by: Dean W. Wichern Journal of the American Statistical Association, Vol. 74, No. 365 (Mar., 1979), pp. 246-247 Published by: Taylor & Francis, Ltd. on behalf of the American Statistical Association Stable URL: http://www.jstor.org/stable/2286765 . Accessed: 05/01/2015 13:35 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 13:35:36 PM All use subject to JSTOR Terms and Conditions

246

of the American Journal StatisticalAssociation, March1979

author'sapproachis quite pedagogical,as evinced by a rather This book is one of the firstof whatI expectwillbe a longline personalexpository styleand the use of manyexamplesand his- of textsdesignedto complement interactive forecasting programs. toricalreferences throughout the text. It is a well-written-even The authorsclaimthattheirbookwill: readable-book that could serve as a text for a one-semester, 1. Minimizestartup costs for the user desiringto obtain second-or third-year graduatecourse. a forecast. The authorcarefullydelineateshis intentin the prefaceand 2. Encourageand supportself-directed learningabout foreincludesa detaileddescriptionof mathematicalprerequisites for casting,especiallyin bridgingthe gap betweentheoryand each chapter.These include a workingknowledgeof real and practice. complexanalysis,elementary measuretheory,and Hilbertspaces 3. Providea completeset of materialsforthe managementand probability, as well as some familiarity with Markov chains. orientedforecasterwho may have a limitedbackground Chapter1 presentsa briefintroduction to the notionsof stoin mathematicsand statisticsbut who clearlyneeds syschasticprocessesand covariancekernels,by way of orientation. tematicforecasting tools. A morevigoroustreatmentof stochasticprocesseshaving finite 4. Serve as a user'sguideforthe set of interactive forecasting secondmomentsis givenin Chapters2 and 3. The generaltheories programsknown as SIBYL/RUNNER (available from of differentiation and integrationof a randomfunction,characS. Wheelwright). terizationof the covariancefunctionof a second-order stochastic 5. Provide a set of practice-oriented teachingmaterialsfor process,construction of stationaryprocessesvia unitaryoperators, thefacultymemberteachingforecasting. and spectraltheoryall proceed in the (natural) Hilbert-space setting.Chapter4 developsthe theoryof spectraldecompositions As the readermay inferfromthe objectiveslisted above, the and the Wold decomposition and applies these to the respective bookis aimedat managersand studentswithlimitedmathematical problemsof interpolationand predictionof a randomprocess. backgrounds whowouldliketo produceforecastswithoutworrying Chapter5 deals withergodictheoryand measure-preserving trans- too muchabout the underlying "theory"or rationaleforthe parin particular, formations; thestronglaw oflargenumbers, Poincare's ticularmethod(s)employed.I feelthis aim is responsibleforthe recurrence theorem, and the ergodictheoremare proved. book's shortcomings and makes it a potentiallylethal weaponin The theoryof Markovprocessesis introducedat an elementary the handsoftheuninitiated. level in Chapter6, with a brieftreatmentof periodicity and reMethodscoveredinclude: exponentialsmoothingand most of currenceof Markov chains. This motivatesthe more rigorous its variants,trendand growthcurve analysis,adaptive filtering, treatment begunin Chapter7, whichcontainsa thoroughdevelop- classical decomposition,univariate and multivariate(transfer mentof the requisitesemigrouptheory.Markov transitionfunc- function)Box-Jenkins techniques,and multipleregression. In adtions and Markov semigroupsare definedin this context,and ditionthereare 24 cases forclass discussionand severalchapters a formalconstruction of Markovprocessesis begunin Chapter8. dealing with overviewsof time series analysis and forecasting, Chapters8 and 9 deal withsamplepath properties, holdingtimes, choosinga forecasting method,preparingthe executiveforecast, stoppingtimes,and strongMarkovprocesses.The book ends with and programfilesand data handling.Apartfromthe introductory brief a but interesting introduction to martingales. chapters,the format,in general,is of the form:methods(with The theoreticaldevelopment is made accessibleby the author's examples)-cases-methods-cases-. . . . Most of the cases are strategicplacementof problemsand examplesthroughout the text, interesting and mightbe the best part of the book. In fact they ratherthan at the end of each chapter.The variousresultsare could probablybe collectedin a separatevolumeand marketed explainedas well as derived,and the relegationof certainproofs as supplementary materialforforecasting courses. and derivationsto the status of problemsservesboth to involve Because the bookcoversa lot ofground,is gearedto the mathethe reader(student)and to maintainthe flowof exposition.Many maticallynaive user,and is tied to a particularsuiteof computer of the problemsare essentialto a completeunderstanding of the programs,it is necessarilyterse and "cookbookish."In a given theoryand examples,and it seemsadvisablethattheseriousstudent methodology chapter,formulasare presentedvirtuallyout of the workthemall whileproceeding thetext. through blue,simplecalculationsare performed to illustratethe procedure, thegenerality Although ofthetitleofthisbookmaybe somewhat and photographs of computerprintoutfortest cases are provided. misleading,the author achieves his stated purpose,namely to Rules of thumbare abundant.An example: In theirdiscussion survey"some" of the moderntheoryand to preparethe student of "generalizedadaptive filtering," the authorsdeclare: "As a forresearchin morespecializedareas. One weaknessof the book, practicalrule,one can choosethe degreeof the AR process-the insofar as it is designedas a text,is theratherlimitedbibliography. numberof weights-to equal the time lag corresponding to the The references seem to be historicalratherthan complete,with largestpositiveautocorrelation aftera time lag of 3." coefficient the omissionof several moremodernworks,such as Blumenthal There are othersuspiciousstatements.For one of the subroutines and Getoor's(1968) MarkovProcessesand PotentialTheory.The designedto do a univariateBox-Jenkins analysis,Makridakisand authormanages to eitherspecifyor provide the mathematical Wheelwright note, "the user has the choice of any of fourteen foundation necessaryforrigorous ofeach ofhis results different development models which,for all practical purposes,are flexible withoutwandering farafieldof the probabilistic topicsthemselves. enoughto fitany type of data." In fairnessto the authors,howThis is exemplified most markedlyby the detaileddevelopment ever,thereis anothersubroutinewhichwill fitany ARMA (p, q) of semigroup theorygivenin Chapter7, all of whichis necessary modeland provideforecasts. to the subsequentdiscussionof Markov processes.In this sense There are occasionswherequantitiesappear in the (sample) thebookservesas a good exampleof appliedmathematics, insofar computeroutputwhichwerenot mentionedin the precedingdeas one cannotapply that mathematics whichone does not know. scriptionof the forecasting method.For example,a constantor The preface provides a good general orientation,especially meantermis includedin themodelsfortheend-of-chapter examples regarding certain notions of applied science,and this reviewer illustrating but it does not appear the Box-Jenkins methodology, highly recommends thatit notbe skippedby thereader. in theinitialformulation oftheARMA models.Alsoin thischapter, the Box-Piercex2 statisticis suggestedas a tool fordetermining PETER F. THALL theappropriate levelofdifferencing. University of Texas at Dallas The book is decidedlyforecasting, as opposedto modelbuilding, oriented.Consequentlythereis virtuallyno mentionof residual REFERENCES section.This is unfortunate. analysisexcept in the Box-Jenkins intervals, Blumenthal,R.M., and Getoor, R.K. (1968), Markov Processes and Potential Point estimatesof futurevalues are stressed.Prediction again apart fromthe Box-Jenkinssection,are dismissedwith New York: Academic Press. Theory, Lamperti,J. (1966), Probability:A Surveyof theMathematicalTheory,New York: "the meansquarederrorcan be used to estimateconfidence limits. W.A. Benjamin,Inc. These limitsare equal to the forecast,plus or minustwice the square root of the mean squared error. .

.

. For more than one

periodahead forecaststhe limitsof the confidenceintervalswill be a littlewider,buttheirmagnitude can stillbe roughlyassessed." How? There are a fair numberof misprints;most of them will be SpyrosMakridakisand StevenC. Wheelwright. San Francisco: caughtby experiencedreaders.One I can't resistthe temptation Holden-Day,1977.xxiii+ 650 pp. $25.00.

Interactive Forecasting: Univariate and Multivariate Methods, 2ndEdition.

This content downloaded from 194.221.86.126 on Mon, 5 Jan 2015 13:35:36 PM All use subject to JSTOR Terms and Conditions

BookReviews

247

on page 564. In the accompanying to correctoccursin thefirstlineofthefourthreference table, neitherof the marginalfrequencies is It should read: Naylor, T.H., T.G. Seaks and D.W. Wichern, equal to one-half. "Box-Jenkins Methods:An Alternative to.... To summarize,this book, and the programsit complements, (2) When o is knownthe researchercan usually assume that the theoretical samplingdistributionof means is normaland can referto the normal diseffort to makea varietyofforecasting are theresultofa prodigious tributiontable formakingprobabilitystatements.(p. 106) methodsreadilyavailable to the people who will be teachingor It is only partiallysuccessful.In the hands of doingforecasting. (3) On theuse ofthestandarderroroftheestimateforprediction: people who alreadypossesssome knowledgeof the methodsdisFor the regressionequation Y = -17.26 + 0.70X we can say that, for cussed,theprograms, and to a lesserextentthe book,shouldprove any value of X, the chances are two out of three (or moreexactly 68 in 100) useful.On the otherhand, as a textbookforeitherthe classroom that the true value of Y will fall in a band that is the predictedscore plus or self-study, thereis too much missing.For the neophyteforeor minusone standarderrorof estimate. (p. 212) caster,the littleknowledgecontainedhere may be a dangerous A diagramis includedshowingtwo confidence thing. limitsforpredicting Y-each parallelto the estimatedregression line. DEAN W. WICHERN (4) Testingthe significance ofthe point-biserial correlation: The Universityof Wisconsin-Madison authorsadvocate using the normal theorytest for correlations "since the point-biserial is a Pearson's r" (p. 224), afterstating explicitlythat odd behaviorcan be expectedfromthis statistic: "If the continuousvariable is platykurticor rectangular,coeffiLogic, and Statistical ofGambling TheTheory cientsin the 0.80's may be expected."Assumptions underlying the Edition. Revised t-testforHo: p = 0 are nevermentioned. Richard A. Epstein. New York: Academic Press, 1977. xv (5) On curvilinear data: + 450 pp. $19.50. With most of the variables used in the social sciences,a linear relationship

This is essentiallyan encyclopediaof applicationsof elementary is found between the variables studied. This may be confirmedgenerally mathematicaltools to the analysisof gamblinggames. The first by notingwhethereach of the variablesis normallydistributed.(p. 234) threechaptersprovidea briefhistoryof probabilityand an introtheuse ofa scatterplot is discussed. ductionto probability,statistics,game theory,decisiontheory, Laterinthesamepage,however, and what is purportedto be a "theoryof gambling"containing, (6) Use ofthe samplemean: analogsof the optionalsampling amongothermaterial,elementary The mean is used as the average with intervalor ratio data when distribuand othergamblingresults.The remainder theoremformartingales tions are symmetricalor approximatethe shape of the normalcurve ... if furtherstatistical computationsare to be made, it is often essential that and analysesofmany ofthe bookis devotedto detaileddiscussions the mean be used as the average. This is why means are frequentlycomgames of chance and skill,fromcraps and horse racingto war puted on data that are skewedor are not equal intervalsin nature ... (p. 30) games and the stock market.Each chaptercontainsan extensive listofreferences. (7) The primaryrequirementof a sample is that it be representative of the deAn inquisitivestudent who has already been introducedto fined parent population on every relevant dimension. (p. 59, emphasis in elementaryprobability(the treatmentin the book is probably original). could findthis too terseto be of muchvalue as an introduction) ofclustersampling: The questionsraised in examiningthese games A misconception book fascinating. and sometimesnot fullyanswered. nontrivial, are oftenintriguing, With this proceduresubpopulations are definedwhich are similar to the probabilityor statistics,this book For the teacherof elementary total population and to one another with respect to relevant dimensions; providesa sourceof "real" exampleswhichcan spurclass interest in essence,each constitutesa mini-population.(p. 61) the relevanceof probabilityto everydaydecision by illustrating Obviously, approximationsto a simple random sampling plan are only as making.Althoughit wouldprobablynot serveas a text,thisbook good as the judgment of the researcherwho definesthe relevant strata or probability could well be used as a supplementin an elementary clusters.(p. 61) course. mar the book. In manyplaces the exposi- The bias in the 1936 LiteraryDigest poll is attributedto nonCertaindeficiencies tion is not as clear as it shouldbe. The text oftendoes not refer representative clusters. even when the materialand ideas presented to listed references (8) The authors'obsessionwith hypothesistestingwith preset Finally,some "facts"are givenwithout a levelsleads to the following seemto requirereference. of test results wronginterpretation whichare eitherwrong(e.g., the discussionof the dis- in Chapter6. Aftersupposingthat a nondirectional reference (two-tailed) tributionof the pari-mutuelplace pool on page 290) or highly hypothesis comparingteachingmethodsC and D is rejected,they implausible(e.g., the estimateon p. 248 of 0.05 forthe expected assertthat a statementof whichmethodis bettercannotbe jusvalue of constant-wagerblackjack with perfect strategy). tifiedbecause this "violatesthe assumptionson whichtest probaNonetheless,the book is unique in providingsuch a wide range bilitiesare based" (p. 68). Calculatinga confidenceintervalfor of applicationsof elementarytechniquesto games and gambling IAC - ifD clearlyshowsthefallacyofthislogic. problems. AlthoughI have other,moresubjective,objectionsto the style DAVID HEATH and contentof thistext,it is not worthwhile to includethemhere. CornellUniversity This bookis unacceptableas a textbecauseof errorssuch as those shownabove.

Descriptiveand InferentialStatistics.

CATHY CAMPBELL

University of Minnesota

N.M. Downie and A.R. Starry.New York: Harperand Row Publishers,1977.ix + 362 pp. $11.50. I couldwritemanypagesaboutmyreactionsto thisintroductory AppliedOperationsResearch:A Survey. statisticstext, but the conclusionhas to be that studentsshould Gary E. Whitehouseand Ben L. Wechsler.New York: John and imprecisestatements notbe exposedto the wrong,misleading, Wileyand Sons, 1976.xi + 434 pp. $14.95. that fillthe book. The followingexamplesare a small sample of Beforecommencingon a review,I must admit to an initial I objectto. thestatements prejudicethat persecutedmy judgmentthroughoutmy reading. fora 2 X 2 table: (1) On the X2 testofindependence I was naively lookingforwardto an applied book, and found to operations If the variables are in fact independent,people in favorof the centerwould myselfreviewinganotherelementaryintroduction be split in theirattitudes toward the city hall; half approvingand half dis- research.Notice how the word "another" has crept in already approving.In other words, when the attitudes are independent,knowledge as a reactionto my disappointment, with its veiled suggestion of a person'sattitude towardone constructionproject tells us nothingabout that we alreadyhave too many.No, thereis always a need for his attitudetowardthe otherproject. (pp. 87-88) moreintroductory books,especiallyonesso carefully and thought-

This content downloaded from 194.221.86.126 on Mon, 5 Jan 2015 13:35:36 PM All use subject to JSTOR Terms and Conditions View publication stats...