HyeH veupusy aySimoour METHODS AND APPLICATIONS CAN You PREDICT THE FUTURE BY LOOKING AT THE PAST? j THIRD EDITIO SNOLLWOMddy ONY SaOHLaW 1S8N 0-471-53233-7 I i oases 030.27 wuMethods and Applications Third Edition Spyros Makridakis European Institute of Business Administration (INSEAD) Steven C. Wheelwright Harvard University, Graduate ‘School of Business Administration Rob J. Hyndman Monash University, Department of Mathematics and Statistics ® ITESM ‘CAMPUS CIUDAD DE MEXICO ‘BIBLIOTECA John Wiley & Sons. ine. New York * Chichester * Weinheim * Brisbane * Singapore * Toronto “FORECASTING‘Acasstions Edtor Beth Golde Maretng Maager lesle ines Prodcion Eaton ely Tres Senor eslger ea Boucher ‘This bck nas stn Computer Modern Roma by oe of he authors and printed and bound Malley Uhogaphing The cover was rte by Lehigh ress. “isbeok pried on ace pape. ‘hs nok dead to Moa Yorgi eos aA Margret Luan, Noo, Tnoty, David an Abel whose patience {nd support mate sch a projet posse a! worth Copy 95 John Wey & Son, nc Al igh resend eda ee Sina om ry ay nes, Etonic mater. pcxopng eae Bat Pana tec abe felipe cree nyorearemete is eee emcee copes ‘Clearance Center, 222 Rosewood Drive, Danvers, MA 01923, (S08) 750- meSraan masamee mance Se ce tcarare icra rte ht asin ee ae fare ¢ st avant bp tin scars we it get crete 'SBN 0-471533-9 (cloth: pape) L Forecaning Stil metho 1. Wheelright Stove C. 43h year. Rob The HIT NO 988 ue Basen—dal ca ‘Pak in the ned States of Arca moeresas2 ‘PREFACE ‘The field of organizational forecasting, born in the 1950s, is reaching ‘maturity. Significant theoretical developments in estimation and Prediction, powerful and inexpensive computers coupled with appro- riate software, several large scale empirical studies investigating the accuracy of the major forecasting methods, and, most importantly, the considerable experience gained through the actual application of such methods (in business and non-profit organizations) have: contributed toward achieving this maturity. Today, the feld of (or sanizational) forecasting rests on solid theoretical foundations while also having a realistic, practical base that increases its relevance and usefulness to practicing managers, ‘The preparation ofthis third edition, lke the previous two, is based fon the authors view that the book should: (1) cover the full range of major forecasting methods, (2) provide a complete description of their essential characteristics, (3) present the stepe neaded for their Practical application, (4) avoid getting bogged down in the theoret. ical details that are aot essential to understanding how the various methods work, (5) provide systematic comparison of the advantages and drawbacks of various methods so that the most appropriate method can be selactad for each forecasting situation, and (6) cover 8 comprehensive set of forecasting horizons (from the immediate to the long-term) and approaches (time series, explanatory, mixed) to forecasting. New in this While meeting the above criteria, this thd edition includes major revisions ofall the chapters and the addition of several completely ‘new chapters. Our purpose has not been to merely revise the second edition, but rewrite it to include the contributions ofthe latest the-os, wile peering the al davdopmants and pact concer te ‘Get nen ene ang and king, We at dt ma ths eon ote compete aly update, as wl ms heal Pca een for those who wast co pl freasting pectic, Some of the new material covered includes the XIZ-ARIMA and th STL methods of tine series decom «Teal reson soothing, best subs regresion and reste sor with ine sen ero : the ane of Aksin's Information Citesion (AIC) for model veiction evra networks and non-lnaeforecating State space modling sed vector astoregresion : 1 Nader aptonch to foecetng the long teem based on meg ends, aloge aad maton Sov ln rem si and judges rca xpeence gine from forecasting competion including the fant Mae Conse treat roar on foes aca. {he teatere of sno fvecanting packages forecasting ents onthe itera ue features ‘is bok itn fort atantion to practal oeatng iets, tscomptnste sonar bth atl mot a one ‘them in practice within a modern business environme Wetton of any een developments a oeesting roeach In pert There are donee ofl data examples and» numberof exam ples from the authors’ consulting expesience. All data set the book ae aalable onthe internet (te bw). '* We emphasise graphical methods and using graphs to help understand the analyses + Our perspective is that forecasting is much more tha fitting ‘models to historical data. While explaining the past is inser tant itis not adequate for accurately predicting the fete ‘+ Much ofthe modern research on forecasting acetracy, sed on surveys of forecast users, is summarized, Many recent developments in forecasting methodology and im lementation are included, Chapter outline ‘The book is divided into twelve chapters and three appendices. Chapter 1: The forecasting perspective. This chapter provides ' conceptual framework for understanding existing methodolo. ses of forecasting and the tasks for which they ean be used. I also provides an overview of the forecasting task and gives 8 useful structure for studying varios forecasting methods (Chapter 2: Basic forecasting tools. Chapter 2 outlines the basic ‘quantitative foundations for the remainder of the book. It fives an overview of the notations and computations required toapply quantitative methods. Furthermore, it summaries the ‘ations measures commonly ued in evaluating and compasing the forecasts of diffrent methods Chapter 3: Time series decomposition. Chapter 3 looks at the smoothing and decomposition of time series. These are ‘ot strictly forecasting methods, but they are wsefal tools for understanding time series better, and therefore they assist in forecasting, - ‘Chapter 4: Exponential smoothing methods. Chapters 4 to 8 focus on quantitative (time series and explanatory) forecasting ‘methods. In Chapter 4 exponential smoothing methods are examined, particularly single exponential smoothing, Holt's ‘method and Holt-Winters’ method. 2 Chapter 5: Simple regression. The explanatory regresion met ods (simple regression, multiple regression and econometric models) are discussed in Chapters $ and 6. Chapter 3 deals with the case where there is only one explanatory variable.Chapter 6: Multiple regression. Regression methods involving ‘more than one explanatory variable are diseussd in Chapter 6. Chapter 7: The Box-Jenkins methodology for ARIMA mod- els. In Chapter 7, the modeling approach of Box and Jenkins 's described and extended using more cecent methods such a8 the AIC. Chapter 8: Advanced forecasting models. A collection of more advanced time series methods is discussed in Chapter $ in cluding regression with ARIMA errors, dynamic regression (ot transfer function) models, intervention analysis, state space ‘models and neural networks. ‘Chapter 0: Forecasting the long-term. This chapter discusses the problems and challenges when making long-term forecats, It also provides three approaches (mega trends, analogies and sconatios) for ariving at such forecasts Chapter 10: Judgmental forecasting and adjustments. Chap- ter 10 describes the various biases and limitations that affect four judgment, 5 they relate to foreeatting, and proposes ways, to avoid or minimize them. Chapter 11; The use of forecasting methods in practice. This chapter presents information about the usage of the various forecasting methods in business organizations as well as emp ical findings about the accuracy of such methods. In addition it discussos the value of combining forecaata and the ability to Jimprove the accuracy of the reaulting forecasts through such combining, Chapter 12: Implementing forecasting: its uses, advantages ‘and limitations. The final chapter is concerned with practi- ‘al implementation issues while also dealing with the usages, ‘dvantages.and limitations of forecasting, Appendix I: Forecasting resources. Several resources available to assist in the parsut of forecast are discussed in Appendix L These include software, journals, associations and useful Tternet sites, Appendix Il: Glossary of forecasting terms. Appendix Il is & slossary of forecasting terms covering the techaiques, concepts, “ Relationships between the chapters, and tools that are the essential components of forecasting. This slossary can serve asa dictionary and reerence for terms that may be new or not clearly understood by readers Appendix IIT: Statistical tables. Appendix III contaias stati cal tables forthe various tests of significance used in evaluating ‘quantitative forecasts and methods, _ These chapters need not be covered in the order given. ‘The diagram above shows the relationships between the chapters and the necessary prerequisites for studying the material in each chapter. Supplementary resources ‘There ‘web page forthe book at [worw. maths monash edu-au/-hyndman Forecasting] ‘This contains a number of resources including all the data sets used throughout the book. Over 500 other data sets may also be obtaised “from this site as well as the 3003 series used in the Intest M3-1JF Competition ‘An instructor's manual is also available from the publisher. This includes course outlines, teaching suggestions, solutions to all exe cises, and some suggestions for case studies and projects Acknowledgments We would like to shank a number of reviewers for carefully reading our manuscript and providing many helpful suggestions. Leonard E. Ross (California State Polytecknic University) and Elaine L. Tatham (Cniversity of Kansas) reviewed the entice book, Gary Grunwald (Waiversity of Colorado) and Mindi Nath (Monash University) re, viewed Chapters 1 to 8, and Brian Monsell (U.S. Bureau of the Consus) reviewed Chapter 3, particularly the K-12-ARIMA method- ‘ology. Gary Grunwald also provided some ideas for exercises atthe ‘ends ofthe chapters, Victoria Briscoe, Betsy Brink and Linda Mayer provided excellent administrative aasistance for which we are also srateful Spyros Makridakis Fontainebleau, France Steven Wheelwright Boston, Massachusetts Rob Hyndman Melbourne, Austria CONTENTS 1/ THE FORECASTING u wn PERSPECTIVE 1 Why forecast? 2 ‘An overview of forecasting techniques 6 The base tp na eas task 13 oe . References and selected bibliography 17 Berises 2 / BASIC FORECASTING an ” TOOLS 20 Time series and cross-sectional daa 21 Graphical summaries 23 221 Tine pots and tine sves m2 20 Beseral ps 26 221 Seaterpns 7 ‘Numerical summaries 28 YM Unvarte satsies 29 272 Brarate sates 34 121 Adtocoreiston 38 ‘Measuring forecas accuracy 4 260 Sunde sat meses 42 WA? Outotsonpe scars imeasenert 35 248 Comparing oes metods 46 244 Thels Usatste 48 25 ACF of feast err $0 Prediction intervals 52 Least squares estimates S# 261 Discovering and descbing reatensipe 59 gS 27 Transformations and adjustments 63 YA Metbenates talons 63 277 Galena asin 67 207 Adsense tion sod Population changes 70 Appendices 71 2A. Notation or utatne forecatng 71 28 Sormaton sgn 72 References and selected bibliography 74 Brercises 76 3/ TIME SERIES DECOMPOSITION 81 VI Principles of decomposition 84 M1 Decomposton nade 4 2 Dewmposten gop “87 3s Sct gutter bs YA Moving erages 89 SAN Sinpe moms erage: 89 Yan Cotect mtg nets 4 BOA Gout nonng Sage” 9 32h Wegred aang eres 38 33 Loa eegesion smoothing. 101 war ween . VA. Classical decompostion 106 “VA Adve decompostion_ 7 342 Hated empeto )Mw 47 EXPONENTIAL SMOOTHING an n “4 YAS Varstont on etl (ecopostion 2 Census Bureau methods SN Frat tein le Yar (deren is 5B Beso io XARIMA STL decomposition 121 V6 everiop 12 Men One ws ws xan Forecasting he ST panes 24 ‘STL and ae decomposition 125 References and selected bibliography 27 Bvercises BO METHODS The forecasting scenario 138 ‘Averaging methods 141 41 Thenean Hl 122 Mong nerages M2 Exponential smoothing, methods “an 432 Sige apt 493 Hotstnea mebod 58 {4 Heimer end nd “7 Bs ‘Sige ponent smroang. 7 1B 8 ponent smoothing ‘preach ae : Staten ‘A compatson of methods General aspects of smoothing methods 74 m 45) tonateation 74 452 Optmzaton 76 45% Prediction wtenate 77 References and selected bibliography 79 Exercises 181 S// SIMPLE REGRESSION 185 a 2 S21 Coston nang coon 95 S20 Simple genie ad te covesten cote D8 S21 Resa thers nd Inert ebseraons 203 715 Conetion nd cxsion 208 re a! Kren th simple regression 208 SOM Regesion x st tng” 209 ‘in Tre Fen roe s 2 ” 55 Ferecaing ge le regesion model 28 5/4 Non-inear relationships 221 41 Nontineaty Inthe pumeters 222 ‘42 Using oars to orm ear modes 226, 49 Leal egpesion 224 Appendives 228 ‘5A. Determining the vals of mde ne References and selected bibliography 230 Brercises 231 6/ MULTIPLE REGRESSION 240 a os Introduction to multiple linear regression 241 611 Mukple regression mode thea and practice "248 Sina Soi ote ‘oetiene 613 Mutple reyes andthe cient of determination 251 (68 The Ete for oral siglence 252 (65 nal coef: condence lnenalsand Hest 255 (616 The assumptions bend mule Near ejesion models "259 Regression with time series 263 G20 Cec cot wast at 22 Timertedeplty waibie 25 Selecting vaisbles 274 620 Te long ist 278 2 Te sont 27 2 bes sss egpesion 279 044 Sep reyeson 285 Mutiolineasty 287 6401 Matlin nen tere re Seoreyeson” 189 ‘402 Ntslinesy when here ae tore tan ore 29 Maltpe regression and forecasting 291 651 angie set Iegesion sn ocang, 22 (32 Clap tine snes refesion tndtoccasing 54 6/6 Econometric models 299 6 The bass of easomere rrodeing “299 (62 The ahanages and crv of ecnometnemetbode 301 Appendixes. 303 (6A TheDutbititon sate 303 References and selected bibliography 305 Bsercises 306 7/ THE BOX-JENKINS METHODOLOGY FOR ARIMA MODELS 31 7A Examining covelations in ines sales dat 3B 7M The atocameson faction 30 Tia Avie nose mae? 1703 The sanping erent accor 37 714 Permanente it 73 The prt acon 2 71 Recogncingsexonaty ina tine sone 2707 Bape: Pigs doghered 322 772 Examining stationary of ine sevies data 324 720 Renovng rons na tie sense 100 rato wa mode! 19 12h Testorasaiowny 3D 7214 Seana deering_ 31 7083 Becht not 72. ARIMA models for times sees daa 335 TAN An stores noel ot serene 374 78 16 mm 78 712 Ang erage del ol rer we 8 17a Mpa same Pea aa 7014 Pg ode moves, eae isi 2 75 Mes: ARIMA modes 44 ive Mistce: ARIMA modes 45 TV? Sera and ARMA oes He Identification 347 TA Example: A on seasonal tne ener) 170 Brame 2: A sesso ine fener 32 7s Ecole A sexo ie srs eed arwormation 5 7A Recapsaton 357 Estimating the parameters 358, Identicsonrevisted 360 26) kop nar ge 62 78d Exar Sacro pt! wig Fae 32 Dingncste checking 364 Forecasting with ARIMA rode 366. 7A Poe occas 66 Ji Owls aweastag 70 J este ot oengon teeass 71 114 REA oe edn txts scmpestion "YE 1195 Except ‘ong nee 73 References an¢ selected bibliography 374 Bercises 377 ADVANCED FORECASTING References and selected 8 bibliography 440 MODELS 388 eee 8A Regression with ARIMA eros 390 S11 Neseg cess 30 91/ FORECASTING THE LONG- {112 Exagl’ parese motor angle rs TERM 451 ow Bape Ses pea nt 9A Cycles versus long-tem tends v4 Feeesing 400 forecasting copper prices 452 SON Foecasing IBM ales 457 CU poke parolee te 972 Longeterm mega econom Gh agtenie wo 21 The base lors ofthe amie 90. Cycle of aos eats a tose 407 ins 461 4 Sect te ese rer 409 922 tnpesin of enrapiaing 25 Forecasting. 40 longterm tends dot 1726 Cale Mowing sas 45 a rele AGE {85 tevention analysis 418 90M. The losin verte 200 Step-bsedtenentons 49 Inautt Rowuln 07+ 22 Poe bse erento 21 9/2 Fre maj meron the {03 Fate resding 422 Indstt Reoiton ad ther {14 leteenton noel and desloge 463 ferceasing, 23 9/4 Scenario building 472 84 Multarate autoregressive AN Suese: ping andor models 423 alnaiing ompeone Ssonages we 85 State space models 429 suns noe ie ne 475 1851 Some forecasting mode in sate 2940 Ppa resus leet fpxe form 49 ete ad peed of sg Rape ont 01 secepance 476 ew References and selected bibliography 478 16 Nonlinear models 433, Sines 87 Neural network forecasting 435 ae 10 /JUDGMENTAL FORECASTING AND ADJUSTMENTS 482 11 The accuragy of judgmental forecasts 483, 102 The nature of judgmental bises aed limitations 482 1021 gre sen ocean) 1072 sig ene Siero \oascowenrl widen 52 10% Combining statistical and Judgment forecasts 303 1081 Ang loc a budget meeting 503 ce 10/4 Conchsion 508 References and selected bibliography 509 erie S12 117 THE USE OF FORECASTING METHODS IN PRACTICE su IW) Surveys among forecast tes 35 IW Fay and satan wt ime bree metods6 twa The we oan ecg ota 50 WW2 Post-sample acura: empical fedings 525 WO Foctors influencing method selection 532 W4 The combination of forecasts 537 IWAN Faces tat contre to ming combing work 3H 42 Aa example of combing 539References and selected bibliography 543 Bvercises 547 12 / IMPLEMENTING FORECASTING: ITS USES, ADVANTAGES, AND LIMITATIONS 549. [2A What can and cannot be predicted SSI QM Shorterm predictions $53 12/12 Medise-term predictions 54 WV Long term predictions 557 122 Organizational aspects of forecasting S58 ne ute saline Sed 1243 Extrapolatie predictions versus creative insights S67 2A Hindogh ven foresght 569 12/4 Forecasting in the future 571 AN Data. formation, an forecns 57 972 Coleane pene se feceasng, References and selected bibliography 575, Exercises 576 APPENDIX | / FORECASTING RESOURCES 577 1 Forecasting software 578 Vi Spreadbeets 578 Vr stasis pacages 578 15 Specialy ecsing pages 57 4 Sc a oceasing pciage 82 2 Forecasting associations 583 3 Forecasting conferences S85 4 Forecasting joumals and newsleters 585 5 Forecasting on the Intemet 586 References and selected bibliography 588 APPENDIX II / GLOSSARY OF FORECASTING TERMS $89 APPENDIX III / STATISTICAL TABLES 949 Normal probabilities 620 Critical values for t-statistic 621 Critical values for Fstatistic 622 : Inverse normal table 628 Critical values for 73 statistic 629 Values ofthe Durbin-Watson statistic 630 G: Normally distributed obsenations 632 rm goee AUTHOR INDEX 633 SUBJECT INDEX 636 le THE FORECASTING PERSPECTIVE s Why forecast? 2 B The basi steps in aforecasting V2 An overview of forecasting tsk teciges 6 Referees nd selected 120 Spay eres ible” van Gn eng 2 Berces 9Chapter I. "The Forecasting Perspective 1/1 Why forecast? Frequently there is a time lag between awareness of an impending event of need and occurrence of that event. This lead time is the main reason for planning and forecasting. If the lead time is zero for very small, there is no need for planning. If the lead time is Jong, and the outcome of the final even is conditional on ideatifable factors, planaing can perform an important role. In such situations, forecasting is needed to determine whea an event will ccur or a need arise, so that appropriate actions can be taken. ‘In management and administrative situations the need for plansing is great because the lead time for decision making ranges from several years (For the case of capital investments) to a few days or hours ((or transportation or production schedules) to a few saconds (for telecommunication routing or electrical utility loading). Forecasting is an important aid in effective and efficient planning. Opinions on forecasting ate probably ax diverse ax viaws on any set of scientific methods used by decision makers. The layperson may question the validity and efficacy ofa discipline aimed at predicting an lancertain future. However, it should be recognized that substantial progress has been made in forecasting over the past several centuries, ‘There are a large number of phenomena whose outcomes can tow bbe predicted easily. The sunrise can be predicted, as can the speed of a falling object, the trajectory ofa satellite, rainy weather, and a ‘mytiad of other events. However, that was not always the case. ‘The evolution of science has increased the understanding of various aspects of the environment and consequently the predictability of ‘many events. For example when the Ptolemaic aystem of astronomy was developed almost 1900 years ago, it could predict the movement of any star with an accuracy unbeard of before that time. Even then, however, systematic errors were common, Then came the emergence ‘of Copetnican astronomy, which was much more accurate than its Ptolemaic predecessor and could predict the movement of the stars to within hundredths of a second. Today, modera astronomy is more accurate than Copernican astronomy. ‘The same increase in accuracy is showa in the theory of motion, which Aristotle, Galileo, Newton, and Einstein each improved The trend to be able to more accurately predict a wider variety 1/L Why forecast? of events, particularly those in the economic/business environment, will continue to provide a better base from which to plan. Formal forecasting methods are the means by which this improvement is occurting, Regardless of these improvements, two important comments mast bbe kept in view. The fist is that successful forecasting is aat always Girectly useful to managers and others. More than 100 years ago, Jules Verne correctly predicted such developments as submarines, ‘clear energy, and travel to the mooa, Similarly, in the mid.1800s, Charles Babbage not only predicted the need for computers, but also ‘roposed the design and did the actual construction for one. In spite Of the accuracy of these forecasts, they were of litle value in helping ‘organizations to profit from such forecasts or achieve greater success. A second important point isthe distinction between uncontrollable external events (originating with the national economy, governments, ‘customers, and competitors) and controllable internal vente (wich as mmarkoting oF maaufactutiag decisions within the frm). ‘The success ‘of acompaay depends on both types of events, but forecasting apples ly to the former, while decision making applies diecty #o latter, Planning isthe ink that integrates both For the important areas of sales forecasting, planning, and decision ‘making, these relationships are shown in Figuve I-L. Recognizing the ‘ole of forecasting in its organizational and managerial context is usually as important as selecting the forecasting method iteetf, and thus it wil be addressed throughout this book A wide varity of forecasting methods are aallabe to management (see, for example, Makridakis and Wheelwright, 1989) These range from the most naive methods, suck as use of the most recent ob servation as a forecast, to highly complex approaches such as neural ets and econometric systems of simultaneous equations. In addition, the widespread introduction of computers has led to readily avalable software for applying forecasting tachniques. Complementing such software and hardware has been the availabilty of data Gescibing the state of economic events (GNP, consumption, etc) and natucal Phenomena (temperature, rainfall, ete). These data in conjunction ‘with organizational statistics (sales, prices, advertising, etc) and technological know-how provide the base of past information needed forthe various forecasting methods.a ©) Chapter 1. ‘The Forecasting Perspective: = ie = je ag {Bena lee Figure 1-1: Information flows insoles fressting and business planning, (Adapted {from Lippit, 1969. Used by permission) As suggested above, forecasting is an integral part of the decision ‘making activities of management. An organization establishes goals land objectives, seeks to predict environmental factors, then selects ‘actions that it hopes will result in attainment of these goals and objectives. The need for forecasting is increasing as management attempts to decrease its dependence on chance and becomes more scientific in dealing with its environment, Since each area of an organization is related to all others, a good or bad forecast can affect the entire organization. Some ofthe area in which forecasting currently plays an important role are: 1. Scheduling: Efficient use of resources requires the scheduling of production, transportation, cash, personel, and so on. Forecasts of the level of demand for product, material, labor, financing, or service are an eseatal input to such scheduling Acquiring resources: The lead time for acquiring raw materials, hiring personnel, or buying machinery and equipment can vary from a tew days to several years, Forecasting is required to determine future resource requitements Determining resource requirements: all ogasizations must de termine what resources they want to have in the long-term. Such decisions depend on market opportunities, environmen. tal factors, and the internal development of financial, human, product, and technological resources. ‘These determinations all cequre good forecasts and managers who can interpret the predictions and make appropriate decision, Although there are many different areas requiting forecasts, the Preceding three categories are typical of the short, medium, and long-term forecasting requirements of today’s organisations. This range of needs requires that a company develop multiple approaches, 10 predicting uncertain events and build up a system for forecasting. ‘his, in turn, requires that an organization possess knowigdge and skills covering at least four areas: identification and definition of forecasting problems; application of a range of forecasting methods; procedures for selecting the appropriate methods for a specif sitt- ation; and organizational support for applying and using formalized forecasting methods.(Chapter 1, The Forecasting Parspectve! A forecasting system must establish linkages among forecasts made by different management areas. There is a high dogree of interde- pendence among the forecasts of various divisions or departments, which cannot be ignored if forecasting is to be successful. For ‘example, errors in sales projections can trigger 2 series of reactions alfecting budget forecasts, operating expenses, cash fows, inventory levels, pricing, and so on. Similarly, budgeting errors in projecting the amount of money available to each division will affect product development, modernization of equipment, hiring of personael, and advertising expenditures. This, in ture, will influence, ‘mine, the level of sales, operating costs, and cash flows. Clearly there is a strong interdependence among the diferent forecasting areas in ‘an organization. ‘A major aim of this book is not only to examine the techniques available for meeting an organization's forecasting cequirements, but also to consider the interdependence of needs in areas such as pur chasing, production, marketing, finance, and general management a a ti a | | Bet 1/2 An overview of forecasting techniques N i @ Forecasting situations vary widely in thelr time horizons, factors | . ee eens ul : ©. Sales of product © 4, Australian clay brick producti aspects, Figure 1-2 shows graphs of four variables for which forecasts right be required. Figure 1-24 Monthly Australian eletity produciin from March 1056 vo August 1995/(Source: Australian Bureau of Statistics) Note the increasing send, increasing variation each year, and the strong seasonal pate tat i sowty chagiog in shape ‘Thee strong histonal patterns maiw this vadable an exay ies eaealere cae eens Coe tee one to forecast. Because of the changing easonal patteras, eral quaton is whether this downward teed ttaly vo tome ofthe exty data may oot be uefa in constructing & ae y todel, Forecast sciportant for fetare planing ofelectity production facilities and for ensuring existing facies can meet Figure 1-2: Historical data on four variables for which freasts might be reuired” Figure 1-2 Sales of “product C* from a major oil company. This peak demands Droduct was a lubricant sold only in anite of large volume. To forecast a variable of this nature, i ie necessary to investi Figure 1-26 US. Treasury Bill contracts on the Chicago matket for eaaeauat a - aan 100 consecutive trading days in 1981. The downward trend is thee nae ig oduct market, who is buying it and what thee future needs are likely to interesting, but it may only be a short dowaward movement a eaeaga The Forecasting Pepe Figure 1-24 Austeaian monthly clay brick production from March 1956 to September 1994, (Source: Australian Bureau of Statistics.) Clearly, the market is seasonal and quite volatile Accurate forecasts ate very diffcult unless the cause of the fluctuations can be identified ‘To deal with such diverse applications, several techniques have. been developed. These fall into two major categories: quantitative and qualitative methods. Table 1-1 summarizes this categorization ‘scheme and provides examples of situations that might be addressed by forecasting methods in these categories. JUANTITATIVE: Safin yanttaive oration aa Vix vran reitng te cntnnien oft puters ee a the gov elo gon aol pode « Beglnstory, Undentacing how epastny vin euch a pr ff vray ae le Little or no quantitative information is avail able, but sufficient qualitative knowledge exists. ~ '» Prodicsing the speed of telecommunications around the year 2020, «Forecasting how a large increase in oll prices wil affect the consumption of ol. QUALITATIVE: UNPREDICTABLE: Little or no information is avilable ‘« Proicting the effects of interplanetary travel + Predicting the discovery of a new, very cheap form of energy that produces no pellution, “Table 1: Categories of forecasting methods and examples oftheir application, 1/2 An overview of forecasting techniques Quantitative forecasting can be applied when three conditions 1. Information about the past is available 2, This information can be quantified in the form of numerical data. 3. It can be assumed that some aspects of the past pattern will continue into the Future. This last condition is known as the assumption of continuity, it is an underlying premise of all quantitative and many qualitative forecasting methods, no matter how sophisticated they may be Quantitative forecasting techniques vary considerably, having been developed by diverse disciplines for different purposes. Each has its ‘own properties, accuracies, and costs that must be considered in choosing a specific method. Quantitative forecasting procedures fall of a continuum between two extremes: intuitive or ad hoe methods, ‘aod foal quantieative methods based on stalistcal principles. The frst type is based on empirical experience that varies widely from business to business, product to product, and forecaster to forecaster, Intuitive methods are simple and easy to use but not always as accurate as formal quantitative methods. Also, they usually give lctle oF no information about the accuracy of the forecast. Because of these limitations, their use has declined as formal methods have sained in popularity: Many businesses sil use these methods, either ‘because they do not know about simple formal methods or because they prefer a judgmental approach to forecasting instead of more objective approactes. Formal statistical methods can also involve extrapolation, but itis done in a standard way using a systematic approach that attempts to ‘minimize the forecasting errors. There are several formal methods, ‘often requiring limited historical data, that ace inexpensive and easy to use and that can be applied in a mechanical manner (aoe ‘Chapter 4). These methods are useful when forecasts are Beeded for large numberof items and when forecasting errors on a single item will nt be exteemely costly Persons unfamiliar with quantitative forecasting. methods often think that the past cannot describe the future accurately because ‘everything is constantly changing. After some familiaity with data foveatingeplanatory models 3 Chapter 1, The Forecsting Pecieetve) and forecasting techniques, however it becomes clear that although ‘nothing remains exactly the same, some aspects of history do repeat themselves in a sense. Application of the right method can often Identify the relationship. between the variable to be forecasted and time itself (or several other variables), making improved forecasting possible 1/2/1 Explanatory versus time series forecasting An additional dimension for clasifying quantitative forecasting meth- ‘ods is to consider the underlying model involved. There are two major types of forecasting models: time series and explanatory models. Explanatory models assume that the variable to be forecasted ‘exhibits an explanatory relationship with one or more independent vaciables. For example, GNP = f(monetary and fiscal policies, infation, (1.1) ‘capital spending, imports, exports, erro) Notice that the relationship is not exact. There will always be ‘changes in GNP that ean not be accounted for by the variable in the ‘model, and thus some part of GNP changes will emain unpredictable ‘Therefore, we include the “error” term on the right which represents ‘random effects, beyond the variables in the model, that affect the GNP figures. Explanatory models can be applied to many systems—a national ‘economy, a company’s market, or a household. The purpose of the ‘explanatory model is to discover the form of the relationship and tse it to forecast future values of the forecast variable. According to explanatory forecasting, any change in inputs will affect the out- but of the system in a predictable way, assuming the explanatory ‘relationship will not change (assumption of continuity), ‘The procedure for selecting an appropriate functional form of ‘equation (I.1) and estimating its parameters will be discussed in detail later on. At this point it should be emphasized that according to (1.1), GNP depeads upon, or is explained by, the factors on the right-hand side of the equation. As these factors change, GNP will vary in the manner specifed by (1.1). alike explanatory forecasting, time series forecasting treats the YR. Be enon of oeatng Crigies TS system as a black box and makes no attempt to discover the factors affecting its behavior. Therefore, prediction of the fature is based on, past values of a variable and/or past errors, but not on explanatory variables which may affect the system, ‘The objective of such time seties forecasting methods is to discover the pattera inthe historical data series and extrapolate that pattern into the future, ‘There are two main reasons for wanting to treat a system asa black box. Fit, the aystem may not be uaderstod, and even i were understood may be extremely dial to measure he relationships assumed to govern is behasor. Second, the main concer may be aly to predict what will happen and aot to know why i happens During the eighteenth, nineteenth, and frst part of the twenth ‘centuries, for example, there were several people concerned with the magnitude of sunspots, These was litle Known at that tine as to the reasons fr the sunspots o the sources of energy of the sun. This lack of kaowiadge, however, did not hinder many Investigators who collected and analyzed the frequency of sunspot. Schuster (1006) tound that there was a rogular pater in the magnitude of sump, and he and several others were able to predict their continuation through time series analysis If the only purpose is to forecast future values of GNP without concern as to why a certain level of GNP will be realized, a time series approach would be appropriate. [tis known that the magaitude of GNP does not change drastically from one month to another, ot ven from one year to another. Thus the GNP of nex: month will depend upon the GNP of the previous month and possibly that ofthe ‘months before, Based on this observation, GNP might be expressed as follows GNPeai = f(GNP,,GNP,-, GN} GNPtasy---sett0r), (1.2) where ¢ is the present month, ¢+1 is the next month, {1 is the last ‘month, {~2 is two months ago, and s0 on. Equation (1.2) is similar to (1.1) except that the factoés on the fight-hand side are previous values ofthe left-hand side. This makes ‘the job of forecasting easier once (1.2) ie known, since it requis no ‘special input values as (1.1) does. However, a requirement with both ‘equations (1.1) and (1.2)is thatthe relationship between the left- and Fighthand sides of the equations must be discovered and measured,usitatve fevecating Chapter 1, The Forecasting Perspective Both time series and explanatory models have advantages in ce. tain situations, Tlie series models can often be used more easly to forecast, whereas explanatory models can be used with greater success for policy and decision making. Whenever the necessary data are availabe, a forecasting celationship can be hypothesized either as a function of time of as function of explanatory variables, and tested AAs demonstrated by the GNP example, quite often itis possible to forecast by using either explanatory or time series approaches. [tis also possible to combine the two approaches. Models which involve both time series and explanatory features ate discussed in Chapter 8, 1/2/2 Qualitative forecasting Qualitative forecasting methods, on the other hand, do not require data in the same manner as quantitative forecasting methods. ‘The inputs required depend on the specific method and are mainly the product of judgment and accumulated knowledge. (See Table 1-1.) Qualitative approaches often roquire inputs from & number of spe- ally trained people. ‘As with their quantitative counterparts, qualitative techniques vary widely in cost, complexity, and value. They can be used separately but are more often used in combination with each other ‘or in conjunction with quantitative methods. It is more dificult to measure the usefulness of qualitative fore ‘casts. ‘They are used mainly to provide hints, to ald the planner, and to supplement quantitative forecats, rather than to provide a specific numerical forecast. Because of their nature and cost, they ate used almost exclusively for mediam- and long-range situations such as formulating strategy, developing new produets and tachnologies, and developing long-range plans. Although doubts are often expressed about the value of qualitative forecasting, it frequently provides ‘useful information for managers. It is a premise of the authors that qualitative methods can be used successfully in conjunction with quantitative methods in such areas as product development, capital expenditures, goal and strategy formulation, and mergers, by even medium and small organizations, Whatever the shortcomings of qualitative methods, frequently the only alternative is no forecast atl. 1/3) The basic steps in a forecasting tase ‘The forecaster has a wide range of methods wailable that vary in accuracy, seope, time horizon, and cost. Key tasks are deciding which ‘method to apply in each situation, how much reliance to place on the method itself, and how much modification is requiced to incorporate personal judgment before predictions are usad ata bass for plaaning future actions, These issues will be addressod throughout this book, 1/3 The basic steps in a forecasting task ‘There are five basic steps in any forecasting task for which quantita. tive data are available. Step 1: Problem definition ‘The definition of the problem is sometimes the most dificult ‘aspect of the forecaster’ task. It iavolves developing a deep understanding of how the forecasts will be used, who requires the forecasts, and how the forecasting function fits within the organization. [tis worth spending time talking to everyone who will be involved in collecting data, maintaining databases, and using the foreeasts for future planning. Consider the following statement by the manager of a paper products nianufacturing compan ‘We havea computerized inventory control system and we can get daly, wookly, and monthly reports at the drop ofa hat. But our inventory situation is bad, We have far too much inventory at the factories, in the ‘warehouses, and in the pipeline. Can we get better forecasts of future production and demand so we can ‘reduce our inventory and save storage costs? ‘A forecaster has agreat deal of work todo to properly define the forecasting problem, before any answers can be provided. For example, we need to know exactly what products ate stored, who uses them, how long it takes to produce each item, what level of unsatisfied demand the company is prepared to bear, and so on. problem dein aStep 2: Gathering information gathering infernaton preliminary analysis forecasting models Step 3: Preliminary (exploratory) analy eee ot 1. The Forecating Perspective ‘There ate always at least two kinds of information available: (a) statitical (usually numerical) data, and (b) the accumu- lated judgment and expertise of key personnel. Both kinds of Information must be tapped. Iti necessary to collect historical data of the items of interest. We use the historical data to construct a model which can be used for forecasting. In the case of the paper products inventory, the data collected may consist of monthly demand and production for each item of interest over the previous tree years. Other relevant data such as the timing and length of any significant production downtime due to equipment failure oF industrial disputes may also need to be collected. What do the data tell us? We start by graphing the data for fal inspection. Then we compute some simple descriptive Statistics (eg, mean, standard deviation, minimum, maximum, percentiles) associated with each set of data. Where more than one series of historical data is available and relevant, we ‘an produce scatter plots of each pair of series and related descriptive statistics (e.g.,cortelations). ‘These graphical and ‘numerical summaries are discusted in Chapter 2. Another useful tool is decomposition analysis (Chapter 3) to chock the ralative strengths of trend, seasonality, cycles, and to identity ‘unusual data points ‘The purpose in all cases at this stage is to ge a feel for the data ‘Ase there consistent patterns? Is there a significant trend? Is ‘seasonality importan:? Is there evidence of the presence of business cycles? Are there any outliers (extreme points) in the data that need to be explained by thove with expert knowledge? How strong are the relationships among the variables available for analysis? Such preliminary analyses will help suggest a class of quantita: tive models that might be useful ia the forecasting assignment. Step hosing and fitting models ‘This step involves choosing and fitting several quantitative forecasting models. In this book we will be discussing many types of quantitative forecasting models and will explain the 1/3. The basic steps in a forecasting task technical details with completely worked-out examples. For ‘ow, we merely mention that the preliminary analysis (Step 3 bore) serves to limit the search for an appropriate forecasting ‘model and we would pursue one or two leading contenders for subsequent analysis Bach model is itself an artificial construct. It is based on a set of assumptions (explicit and implicit) and asually involves fone ot more parameters which must be “fitted” using the known historical data. We will discuss exponential smoothing methods (Chapter 4), regression modes (Chapters 5 and 6), Box-Jenkins ARIMA models (Chapter 7), and a variety of other topics in cluding non-linear models, regression with ARIMA error, inter vention models, transfer function models, multivariate ARMA, models, and state space models (Chapter 8) ‘When forecasting the long-term, a less formal approach is often better. ‘This can involve identifying and extrapolating mega trends going back in time, using analogies, and constructing swenarion UD consider future possiblities," These issues ate discussed in Chapter 9 Step 5: Using and evaluating a forecasting model Once a model has been selacted judiciously and ite parameters estimated appropriately, the model is to be used to make forecasts, and the users of the forecasts will be evaluating the ros and cons of the model as time progresses. A forecasting assignment is not complete when the model has been fited to the known data. The performance of the model can only be properly evaluated after the data for the forecast period have become available In this book, we have made a clear distinction between “fitting cervors” and “forecasting errors.” We will examine a variety of accuracy measures for both fitting and forecasting (in Chap- ter 2) and we wll emphasize that, in practice, the models fore- casts are seldom used without modifcation. Expert judgment is invariably brought to bear on the use of the forecasts, The incorporation of expert judgment is addressed in Chapter 10. [tis important to be aware of how each forecasting method has performed in practice in other forecasting contexts, There has now been quite a lot of research on this issue looking at 15= Chapkse 1 The Forecasting Perspective Users’ preferences and experiences with a range of forecasting methods. This cesnarch is summarized in Chapter 11 ln addition, the accuracy of future forecasts is not the only criterion for assessing the success of a forecasting assignment. A ‘successful forecasting assignment will usually also be a stimulus to action within the organization. If the forecasts suggest a loomy picture ahead, then management will do its best to ly to change the scenario so that the gloomy forecast will not come true. Ifthe Forecasts suggest a positive future, then the management will work hard to make that come true. In seoeral, forecasts act as new information and management must incorporate such information into its basic objactive to enhance the likelihood of a favorable outcome. Implementing forecast. ing is often at least as important as the forecasts themselves ‘Chapter 12 addresses this important subject, References and selected bibliography: References and selected bibliography AnuistRoNc, J.S, (1978) Long-range forecasting: from crystal ball to computer, New York: John Wiley & Sons AsteR, W. (1978) Forecasting: an appraisal for policy makers and planners, Baltimore: Johas Hopkins University Baits, D.G. and L.C. Psprer (1993) Business fluctuations: fore- casting techniques and applications, 2nd ed., Englewood Clif, Na: Prentice-Hall BaIGHT, J.R. (1978) Practical technological foreasting, Austin, ‘Texas: Industrial Management Center. Clamoens, J.C., S.K. Muuuick, and D.D. Sut (1971) How to choose the right forecasting technique, Harvard Business Review, July-August, 45-57 (1974) An executive's guide to forecasting, New York ‘John Wiley & Sars. CLARK, A.C. (1971) Profiles of the future, New York: Bantam Books. CuBany, J.P. and H. Levewsactt (1982) The professional fore caster: the forecasting process through data analyis, Belmont Cal: Lifetime Leaening Publications. CuirtoN, P., H. NovyeN, and S. Nutr (1991) Market research ‘sing forecasting in business, Stoneham, Mass: Butterworth and Heinemann, Kwowa, K.K., C. Lt, V. SimUNEK, and C.L, JAIN (1995) Bibli- ‘ography on forecasting and planning, Flushing, N.Y.: Graceway Publishing Company. Liperet, V.G. (1969) Statistical sales forecasting, New York: Finan ial Executives Intute, MAKRIDAKIS, S. and S.C. WHEBLWRIGHT (1979) “Forecasting,” TIMS studies in forecasting, Vol. 12, Amsterdam: North-Holland. (1981) “Forecasting an organization's futures,” in Handbook of organizational design, Vo. 1, eds. P.C. Nystrom and W.HL Starbuck, Oxford: Oxford University Press, 122-188, (1989) Forecasting methods for management, Sth ed, New York: Joha Wiley & Sons,SE Chapter AE The Focecastng Perspective ScwusreR, R. (1906) On the periodicity of sunspots, Philosophical Transactions, Series A, 208, 69-100. Steiwen, G.A. (1979) Strategic planning, New York: The Free Press ‘Twomoroutos, N.T. (1980) Applied forecasting methods, Engle: ‘wood Clifs, NJ.: Prentice Hal. Ererciee Exercises 1.4 Several approaches have been suggested by those attempting to pradict stock market movements. Theee of them are de- scribed briely below. How does each relate to the different Approaches 0 forecasting decribed in this chapter? (a) Dow Theory: There tend to be support levels (lower bounds) and cesistance levels (upper bounds) for stack prices both for the overall market and for individual stocks. These levels ean be found by plotting prices of the market or stock overtime, (b) Random Walk Theory: There is no way to predict future ‘movements in the stock market or individual stocks, since all avallable information is quickly assimilated by the investors and moves market prices in the appropriate direction. (©) The prices of individual stocks or of the market in general are largely determined by eacnings. 1.2 You ace asked to provide sales forecasts of several products for large biscuit manufacturing company. Define the five steps of forecasting in the context ofthis project.Chapter 2. Basic Forecasting Tools a. i ‘To develop an understanding ofthe eld of quantitative forecasting requires some basie notation and terminology. This chapter presents such fundamentals In Append 2A the actation wed throeghout i the book is preseatad, and inthe Body of tis chapter the allowieg TOOLS topics are discussed: graphical methods fo visualizing data tobe wed foe foreasting(Sation 2/2), the moat important sutumary staan ime series and cross-sectional WS Prediction intervals 52 1 (ection 2/3), and the various measures of forecasting accuracy that ble or eieebtecaccan ae used to help judge the appropriateness ofa model (Seton 2/3), ts 2 Least sures catinales 54 calculation of prediction intervals (Section 2/5), the least squares Geaehleal senate 52) Se ere ee, procedure for estimating parameters of a model (Section 2/6), and a ee ‘the use of transformations to simplify data patterns (Section 2/7). patterns 24 2/7 Transformations and yy tom a 2 Madera stomsins 63 iin cart se acctlonial data 2 Numeral summaries 28 Jon eee 2/1 Time series and cross-sectional dat 2A Una sats 29 7h Naatnet orton ae Tn Brana ae Popston changes 70 Throughout this chapter, we wil use two dat sets to ilusteate ideas sathtmanmamalie ‘Appendices 71 2A Mexsuing forecast sceuragy 41 TA. Nealon fo cunts + Price ($US), muleage (mpg), and country of orga for 48 autor 2A" Sundud taal nents 2 focatng Fr mobiles from Consumer Reports, April 1990, pp: 235-25. 24a Osta ery 28 Suman sgt 72 + Monthly Australian beer production (megaliters, Ml from Jan- Ws Capt ec metads 4 References and selected vary 1991-August 1995. 244 Tels Usatatie 48 DAS ACF of ees eroe 50 ‘These data are given in Tables 2-1 and 22, Often our historical data will consist of a sequence of observations over time. We call such & sequence a time series. For example, time sates ‘monthly sales figures, daily stock prices, weekly interest rates, yearly profits, daily maximum temperatures, anaual crop production, and ‘ectrocardiograph measurements areal time series. In forecasting, we are trying to estimate how the sequence of ‘observations will continue into the future. To make things simple, we will assume that the times of observation are equally spaced. ‘This, fs not a great restriction because most business series are measured ally, monthly, quartery, or yearly and so will be equally spaced OF the twoexamples above, the beer data form a time series as they are monthly figures over a period of time. However, the automobile data do not form a time series. They are cross-sectional data; all cro-setona data observations are from the same ti 20= (Chater 2. Basie Forecasting Tools 2/2 Graphical summisies Month vo0i_1we3 1800 1woa 1005 Taaiary 10a Ta — 190 Tot Fibraty 81331138 March” 152163150 16e 188 Sek ine Al dat) ast that oe a My = 55d aBL jae tee i Jee has Se ike | jy) on jg Blea Auge Ta feat Gor eave Ein toe imi i ia a8 New tase oe Ocaber 100 168 Sk) Name Spe ve so of oa November 19218117 Si te een hee December 19218816482 Mate ee Be Table 22: Monti Austrian ter prtin: Jewry LTA 198 Stiaeh Sa Sornne Ve SA ae Sanaece” aft Rugs at See ene oe 8 ise Ene Pre Ue m= 8 iS 2/2. Graphical summaries ete a 2 is Riou ine vs ee aie es sararr = 3 ie The singe most important thing to do when fst exploring the data Siete Ca a 2 oe i to visualize the data through graphs. The bu Tare ofthe data Prd Thaanderbied VE USA s = ‘data including patterns and tal observatic st Fo Tandeind BA Bee 3 Patets and unesalosecratons te met cay sne Balt Le Sobe'Ve we ies through graphs. Sometimes graphs also suggest posible explanations fart jen te. fer some of the vacation fa the dat, Ete mt RS Ene tous see ee For example, indostial daputes wil often afte: tine strinof So bane ve ag ube troduction; changes in government il aft sanomic une etn exc jee Bi anges in define may oul in enable change neces Breath Le a eee patterns. Grapha are the mot effective nay of dent ying the elec Seat Darn mF we of such events Inthe data. Whete pouble, thee events shuld be Sees. bg Seneca eee Fed Pits Se Stim ‘The type of data will determine whch type of graph i ost ae cre) appropriate. Figures 2-1, 2-2, and 2-3 show three plots that provide Rwemes te = 3 JSe el Information for freasing. ‘These graphical forms abel te Eau Selous ees eee } routinely ted in forecasting profs and til be ailnedtrsognn Sia ia ie is there ofthe books oman eae ‘able 21: Pre, mean ory of iin for eto fom Conner Reports, April 1990, pp. 235-255. |"Chapter 2. Basic Forecasting Toole igure 2-1: Time plat of monthly Australian bee production (megaliters, MA) from January 1991- August 1985, time plot time seis patton 2/2/1 Time plots and time series patterns For time series, the most obvious graphical form is a time plot in which the data are plotted over time. Figure 1-2 (p. 7) shows some examples. A time plot immediately reveals any trends over time, ‘any regular seasonal behavior, and other systematic features of the data. These need to be identifed so they can be incorporated into the statistical model. Figure 2-1 shows a time plot of the beer data. This reveals the range of the data and the time at which peaks occur, It also shows the relative size of the peaks compared with the rest of the series and the randomness in the series since the data pattern is not perfect. An important step in selecting an appropriate forecasting method 's to consider the types of data patterns, so that the methods most appropriate to those patterns can be utilized. Four types of time se- Fes data patterns can be distinguished: horizontal, seasonal, cyclical, and trend. 2/2 Graphical summa 1. A horizontal (H) pattern exists when the data values fuctuate around a constant mean. (Suck a series is called “stationary” in its mean.) A product whoee sales do not increase or decrease over time would be ofthis type. Similarly, a quality contol sit- uation involving sampling from a continuous production process that theoceticaly does not change would also show a horizontal pattern, 2A seasonal (S) pattern exists when a series is influenced by Seasonal factors (eg. the quarter of the yeas, the month, of day of the week). Sales of products such as soft drinks, iee creams, and household electricity consumption all exhibit this type of pattern. The beer data show seasonality with a peak in production in November and December (in preparation for (Christmas) each yeae. Seasonal series are sometimes also called “periodic” although they do not exactly repeat themselves over each period 3. A cyclical (C) pattern exists when the data exhibit siee and falls that ace not of a fized period. For economic series, these are usualy due to economic fluctuations such as those asso. ciated with the business cycle. The sales af products such as automobiles, steel, and major appliances exhibit this type of pattera. The clay brick production shown in Figure 1-24 (0. 7) shows cycles of several years in addition to the quarterly seasonal pattern. The major distinction between « seasonal and a cyclical pattern is that the former is of a coastaat length and recurs on a regular periodic base, while the latter vari in length. Moreover, the average length of a cycle is usually longer than that of seasonality and the magnitade of a cycle is usually more variable than that of seasonality. 4. A tend (T) pattern exists when there is long-term increase ot decrease in the data. The sales of many companies, the gross ‘ational product (GNP), and many other business or economic indicators follow a trend pattern in their movement oye time. ‘Tho electricity production data shown in Figure 1-2a (p. 7) ‘exhibit a strong trend in addition to the monthly seasonality. ‘The beer data in Figure 2-1 show no trend. Many data series include combinations of the preceding patterns. For example, Figuce 1-24 (p.7) shows trend, seasonality, and cyclical aE stations, pede trend2/2 Graphical summaries seasonal pot Figure 22: A seasonal plot of the Australion ber production data. Note that prox ‘tution aks in November and December in preparation forthe southera hemisphere ‘summer ond i east in winter behavior. One of the things that makes forecasting interesting and challenging is the huge variety of patterns that commonly occur in teal time series data, Forecasting methods that are capable of distinguishing each of the patterns must be employed if a separation ‘of the component pattaras is needed. Similaly, alternative methods ‘of forecasting can be used to identity the pattern and to best fit the data so that future values can be forecasted, 2/2/2 Seasonal plots For time series data that are seasonal, it is often useful to also ‘produce a seasonal plot. Figure 2-2 shows a seasonal plot of the beer data. This graph consists of the data plotted against the individual “seasons” in which the data were observed. (In this case a “season” is a month.) This is something like a time plot except that the data from each season are overlapped. A seasonal plot enables the undeciying seasonal pattern to be seen more clearly, and also allows any substantial departures {fom the seasonal pattern to be easly igure 2-8: A scaterplo of price versus mileage for the eulomobile data ‘identified. For example, there is one year (1991) ia which the October beer production was higher than the pattera evident in the other years, 2/2/3 Scatterplats ‘The automobile data of Table 2-1 are not a time series making time or seasonal plots inappropriate for these data. However, these data, are wall suited to a scatterplot (see Figure 2-3) such as that of price scatteplot Against mileage. In Figure 23 we have plotted the variable we wish to forecast (price) against one of the explanatory variables (mileage) Each point on the graph represents one type of vehicle. The plot shows the relationship between price and mileage: vehicles with high mileage per gallon are generally cheaper than less fuel-efficient vehicles, (Both price and fuel-efficiency are related to the vehicle and engine size.) Vehicles with low mileage pe gallon are generally priced ‘over a range from around $12,000 to $18,000, with three vehicles ‘much more expensive than other vehicles of comparable efficiency, ‘The acatterplot helps is visualize the relationship and suggests that 8 forecasting model must include mileage as an explanatory variable,Figure 24: A scottepot showing price, mileage, ond the country of origin forthe ‘avtomobie data, (Chapter 2 Base Forecasting Tools ESSE ees Prcamage relationship for 4S autorobie With the automobile data there isan additional explanatory var able, country of origin, which ia a categorical (qualitative) variable rather than a numerical variable (its values are categories rather ‘than numbers). Thus we cannot plot price against country in the same way. However, we can augment our scatterplt of price against mileage 10 also show the country of origin information. This is displayed In Figure 2-4, Here we have a scatterplot showing three variables. It shows that the tvo most efficient automobiles and the ‘three expensive automobiles are all Japanese. It also shows that ‘overall U.S.A. automobiles may be more eficint than the Japanese ones. A forecasting model might also include the country of origin as an explanatory variable, but that itis probably going to be less fective than mileage in giving accurate predictions 2/3 Numerical summaries In addition to graphics, it i also helpful to provide numerical sum- rmacies. A summary number for a data set is called a statistic. 2/3 Numerical summariés 0: For a single data set (univariate dsta) ot a single time series, the most common descriptive statistics are the mean, the standard deviation, and the variance. Section 2/3/1 deals with these univariate statistics, For a pair of random vaciables (bivariate data) it is of interest to describe how the two data sets relate to each other. The most widely used summary qumbers (statistics) for this purpose are the covariance and the correlation, and these will be defined in Section 2/3/2. ‘Then for a single time series, it ie very useful to compare the ‘observation at one time period with the observation at another time petiod. The two most common statistics hete are the autocovarance fand the autocorrelation, which are defined in Section 2/3/3. These ‘measures will be used extensively in Chapter 7. 2/3/1. Univariate statistics Consider the mileage of the 19 Japanese automobiles given in Table 21, reproduced in Table 2-8. The vehicles have been numbered from 110 19 for easy ceference Using the letter M to denote mileage and a subscript i (i 1,2,3,...,19) to denote the ith vabicle, the mean mileage can be written! (M+ Ma + Ma +4 Ma) /19 1 = 5hM 469/19 = 24.58 mpg. ‘The mean should not be confused with the median, whichis the ‘middle observation. So for the preceding 19 vehicles, the median is, the mileage ofthe tenth vehicle when they are arranged in icreasing order as shown in Table 23. That is, the median of the mileage data is 23. Both the mean and the median are designed to provide a ‘wumerical measure ofthe center of the dataset "The wommation station, sed here explained in Append 2. nvr data ivarite data edi‘Chapter 2. Basic Forecasting Tools ‘Make Vaada MPV V6 Nissan Van 4 Acura Legend VB Mitsubishi Wagon 4 Nissan Axxess 4 Mitsubishi Sigma V6 Nissan Stanza 4 Mazda 929 V6 Nissan Maxima V6 ‘Teyoca Cressida 6 Nissan 2408K 4 Subaru Loyale 4 Mitsubishi Galant 4 Honda Prelude $i sWS 4 | | | | | Subaru XT 4 Masda Protege « ow Honda Civic CRX Si 4 uw Stbaru Justy 3 BH Toyota Tercel 4 35, ‘Table 2-8: Mileage of Jopnene automobiles listed in Table 21 ‘As well as measuring the center of a data set, it is also valuable to measure the spread of the data. That is, we want a numerical ‘measure indicating ifthe data are tightly bunched together or spread across a wide range. To develop a measure of spread, we first need to calealate for each vehicle how far its mileage is from the mean mileage. The mean if 's subtracted from each M; to give the ith deviation from the mean, eviton from mein (Ms ~ 8). sean absolute deviation ‘The sum of the deviations will always equal zero (as shown under column 3 of Table 2-4). Therefore, to develop a useful descriptive statistic from these deviations, they ate ether squared (a8 in columa 5 of Table 2-4), or, occasionally, the absolute value i taken (a8 in column 4). The mean of the absolute daviations is denoted MAD, 2/3. Nuretiealsumiates Se, ‘nd for the mileage data Tee pode MaD= 4° 1M— sz ‘The mean of the squared deviations is designated MSD: MSD = 5 So(04 — si? = sna.ii9= 27a. Closely related to the mean squared deviation (MSD) is the vari ance, which is defined as the sum of squared deviations divided by one less than the total numberof observations.” For the mileage data the variance of age $7 = EDM — ait = si./ie = 286, Note that since the variance formsla uses 18 in the denominator nd MSD formula uses 18,5? larger than MSD. The variance 3! 1s ls intuitive than MSD but it has some desirable mathematical properties. The deviations (M; ~ Nf) are defined in units of miles per gallon. ‘Therefore, the squared deviations ate in units of squared mpg and 0 the mean squared deviation, MSD, and the variance, S®, are also defined in units of squared mpg. By taking the square root ofthese ‘wo summary numbers, we get summary statistics inthe same units 8 the data, In particular, we wil define the standard deviation as S=V3 = VBE = 4. Both the MAD and the standard deviation $ provide measures, of spread. They (roughly) measure the average deviation of the "The sum of oqared devine i divided bythe “degrees of fede” which, ‘am be defined asthe number of data poits min the mambee af parame ‘etimated, When caluating a variance, we bve fa emate the meas weg te st, so the degrns of feedom ison lee tha the total amber of sbcrudane ‘tn statics « dictions made betwen based eat nd an sabe ‘timatr, For sample data the MSD i sed eimeta of population manaore and the variance 5" is an ubieedetimator ofthe population vatance Ses ace (1998), 19, fra deiion of wnbasednese mean squared deviation standard deviation assures of spreadmedian mean absolute Seviton sean squares Seviation standard deviation Chapter 2. Basie Forecs ing Tools ‘observations from their mean. If the observations are spread out, they will tend to be far from the mean, both above and below. Some deviations will be large positive numbers, and some will be large negative numbers..But the squared deviations (or the absolute deviations) will be all positive. So both MAD and 5 will be large when the data are spread out, and small when the data are close together. Both MAD and S have the same units as the observations For many data sets the following useful rules of thumb hold * approximately two-thirds of the observations lie within 1 stan- dard deviation of the mean; and © approximately 95% of the observations lie within 2 standard deviations of the mean. ‘To.summarize, suppose there ate n obeervations and the individual observations are denoted by Y; for i= 1,....n. Then the univariate statistics (summary numbers) that wil be used inthis text are defined (generally) as follows: v=ipy (ay mile observation imo; Madian = { Seeai (22) oberaton if een, Map = LDMm-¥1 (2) Msp = 2(-¥) 4) x= pwr ey) s= vf yur es) All summation ae over the index i from 1 through a Of course, with much statistical software readily accessible, there ‘is usually no need to compute these statistics by hand. However, it is important to understand the formulae behind them in order to ‘understand what they mean. 2/3) Numerical summaries 8 ———_—_—_——_ Qa @) @ o Vehicle Mileage Deviation Me (Mi T 2 3 4 8 7 8 o 10 uw 2 13 4 15 6 7 B 19 Sum Mean sing (21) Median sing (2.2) Mean Absolute Deviation col 4 sum)/19 = 4.37 sing (2.3) Mean Squared Deviation MSD = (col 5 sum)/19= 27.1 ving (2.4) Variance sing (2.5) ‘Standard Deviation : sing (2.8) Ese asee ice ceeeeasaa dan dncncce cea eas tas eceeegaentg oases scat ‘Table 24: Computation of the univariate statistics forthe mileage of Japanese automedilenChapter 2. Base Focdéastng Toole 2/3. Numerical sammarie ‘Table 2-5: Price and milage forthe Jepanese automobiles listed in Table 2 egstive/positive 2/3/2_ Bivariate statistics ‘Table 2-5 shows the price and mileage for the Japanese automobiles sven in Table 2, To prevent the computations from becoming. ‘cumbersome, we will dal withthe price variable in units of thousands of dollars When these data ace plotted, as in Figure 2-5, it can be seen that a negative relationship exists betwoon these two variables. By nogative relationship we mean that as mileage increases, price tends to decrease. (A positive relationship would be similar to the height versus weight relationship—as height increases weight increases too.) Whenever we are dealing with two paired observations (e.g, price and mileage, height and weight, price and demand), itis of interest to examine and measure the extent of the relationship between the ‘wo variables, Suppose we denote the two variables by X and Y. A statistic Make Wahi Milage Peee Pricelmileage relationship for Japanese automobiles Wards MPV VE T 1 Tau Nasa Van 4 2 1a Acura Legend V6 a 24780 a Mitaubioms Wagon 4 2 4920 g [Niwan Axaes 4 5 2 ow : itaubiai Sige V6 8 a n879 ; Nissan Sanaa 4 1 11880 Mazda 920 V6 8 1 23300 i [Nissan Maxima VS 4 2 Tae i ‘Toyota Cremida 8 mB a8 - Nissan 2408X 4 no M1320 Subaca Loyale 4 2% 9800 Mitaubah Galant 4 | B 10888 Honda Prelude SiWS4- MT 13.945 Sebara XT 4 BB BOTL Magda Protege 4 6 888 Honde Give CRXSi4 ITB LO Swhars Junty 3 5.866 ‘Toyota Tec 4 sass wiich indicates how two variables “co-vary” i called the covariance covariance and is defined as follows: Covey = Su —K-7) en where and fare the means of X and ¥, respectively, and isthe ‘number of observations on each variable. For the price and mileage data in Table 25, the computations necessary for determining the covacianc (Cov) between pice (P) and mileage (Mf) are shown in Table 24, Fist, the mean pce (P) and the mean mileage (M) are computed using columns 2 and 3, respectively. Then deviations frm the mean are cleolated fa olamas ¢ and 5, and column 8 gives the prodact of thee two deviation. Summing the deviation product (column 8) and dividing by the dogres of redom, n~ I= 1, yee the desired covasiace, Cove =36 aT m0 4700 2» 499 2 109 2 ITT 2 11850 2 23.300 2 17809 2 21498 M1120 2% 9509 25 10989 T1396 bon Rm 4599 Bm 9410 M5865 355.488, “doa ‘Msn mileage Mean price 068 032 om 20 a 732 832 932 1032 Standard deviation of Mf Standaed deviation of P Covariance between Pand M- Cove Correlation between P and M Chapter 2, Basic Forecasting Tools Ca) PP (ahs ~ Tor 3st oss 323i tos2 2198 099 21D ool 2198 so 3ST 29 0 133T 938133 3980 72k 15 2M os oat “4M 010 29 00 oo 5.38 os 1090 TM 5352 245305 Tot 8678 Hr45__ 106.42 asia z a, (a= PP Tor on una 098 0.00 680 5 8785, 1500 srs. O48 1333 370 0.00 53.85 250 5.18 55.50 Mf 460/19 = 24.88 mpg P = 254.823/19 = 13.938 thousands of dolls Su = VEIT = S: mays. -aT3.09/18. Cove, SS 334 540 21.00 21.00 OE = Mi = N(R ~ PD =o n439 50.89 4.64 0.08 =1452 843 3480 =1085, Sia73 oat -13t 093 002 53.69 cars, a1520 “on ‘Table 2-8: Computations for determining the covariance ond the correlation ofthe rice ord mileage data of Tale 2-5. 2/3. Numetcal summaries Note that the units of covariance are problematical. Iti dffult to interpret thousands of dolar-miles per gallon. Hence the value of ‘computing the correlation coeficien decribed below. Note that the covariance between price and mileage is negative, bt the magnitude of Covpy clearly depends on the units involved. Ifthe mileage figures were converted to km per liter and the prices to dollars, the plot (Figure 25) woul! lok he same but the cnatiance would be quite The corelain eof, dig is «spa ovine rear tat als cao th see pole ut ene ihe Cxrrince(Covg) indie yt tro sear det (Bsa Svhvthn te unin the eats snd th donot ce ot Ivins deans meme, nich th conan esse, betwen X and. Ths sweet blows DOG DM. Ex Covy y rYE% es) SS y ‘The efect of this sealing (dividing Covgy by Sy and Sy) is to restrict the range of ry to the interval ~1 to +1. No matter what the units of measurement for X and ¥ the coreelation coefficient, ‘xr i always restricted tole within that interval. More information about corcelation is given in Sections 5/2/2 and 5/2/3, For the data in Table 25 the computations involved ns involved in geting to the colton coin are incede in Table 26. Coluas 8 and are the squared deviations for height and weight, reapectvely, and ‘an be used to determine the standard deviations Sy and Su, taing ‘equation (2.6). Then the covasiance between P and M can be divided by Sp and Suy to yield the correlation between price and mileage, =21.00 rem = Saya 7-078 (29) ‘This summary number is readily interpretable, There is a corelation of -0.73 between price and mileage, which is negative and substantial There isa strong negative assocation between price and mileage. Covariance and expecially correlation are the basic statistics for bivariate data sets, and for more extensive multivariate data sets (Care should be taken, however, to remember that these ate mensu covration 3738 tetocoreation Iagsed variable Chapter 2. Basic Forecasting Tools of linear association between two variables, so that it is not appro- priate (meaningfl) to apply the correlation measure when there I a pronounced curvilinear relationship between the two variables. This Point is amplified in Chapter 5. In summary, the two “vital” statistics for bivariate data sets are Covey Low 1% vm = OME (%- HW -¥) SSS YE YEN - FP In practice these calculations are done by computer. ) 2/3/3. Autocorrelation ‘The covariance and cortelation cooficient are statistics (eummary measures) that measure the extent of the linear relationship between two variables. As such, they can be used to identify explanatory relationships. Autocovariance and autocorrelation are comparable ‘measures that serve the same purpose for a single time series For example, if we compare Y; (the observation at time!) with Yi=1 (the observation at time t~1), then we see how consecutive obsarva- tions are related. The observation ¥;- is described as “lagged” by ‘one period. Similarly, itis possible to compare observations lagged by two periods, three periods, and so on Table 2-7 shows the beer series, whichis a single time series over ‘56 months from January 1901 to August 1995. The observations ¥;, Yay vy Yap are observed at time periods 1, 2, ..., 58, respectively. If we lag the series by one period, as shown in column 3, thea there will be 55 pars of observations to compare. For these 58 overlapping ‘observations we can compute the covariance ard the correlation as if they were two separate series. However, since they are one and the same series (witha lag of one period) the summary measures are called autocovariance and autocorralation Because the two series are almost the same, rather than use silent db. 2/3. Numerical summaries the equas ns (2.7) and (2.8) to compute the astocovariance and autocorrelation, we normally use simpler formulas which give almost the same answers. We denote the autocovariance at lag & by cy and the autocorrelation at lag & by ry. Then define a= iS o-nme-n am | oe By way of illustration, consider the beet data and the calulations of autocovariance and autocorrlation in Table 27, The mens of al the data points in column 2 & ¥ = 149.30, and the deviations in columns 4 and 5 are deviations from ¥. Column 6 is the squated deviations from column 4 and the sum of these squares is the de- nominator in equation (2.11). Column 7 is the columa of deviation products (columa 4 times column 5). The calculations, using (2.10) and (2.11), are given at the bottom of Table 2-7. Using exactly similar procedures, the autocrrlations frags wo, three and beyond ean be obtained, The rats fo the beet data ae 1s folows oat ost ~0.080 Toss -oast “oan 038 Te Notice thatthe autocorrelation at lg 12s higher thas forthe other lng, Thi due to the seasonal pattern in the data: the peaks tend to be 12 months apart andthe troughs tnd to be 12 months apart Silat, the autocorclation at lag 6 i more negative than for he ater lags because troughs tend to be 6 monte behind peaks, nm n re0 fantion corlogram 54130 55119 56__153 Same S36T 2.1. Computing the aulocovariance and the autocorrelation wing equations (2:10) and (2.11), and 20g of one period Autocovasiance lag 1 Autocorrelation lag Lr: Chapter 2. Basic Forecasting Tools a oO. Ist 1990.7 372.63, 130 -1930 91831 9 =30.30 13.6, HIS5.84 8361 Set = 19.90 8895.51 36 8898.51 238.88 Mean ¥ eee = 1588 = 0.21 ‘Together, the autocortelations at lags 1, 2, autocorrelation function or ACE. Rather than scanning a list of ‘aumbers, it is much easier to plot the autocorrelations against the lag. Such a plot is known as a correlogram and helps us visualize the ACF quickly and easily. Figure 26 shows the ACE forthe beer data. + make up the Here the seasonal pattern is sen very clearly. ‘A plot of the ACF is a standard tool in exploring a time ser before forecasting, {t provides a usaul check for seasonality, cycles, and other time series patterns. In the exercises at the end of this chapter are some data sets that display various kinds of patt (trend, seasonality, and cycles) and the autocorrelations for these 2/4 Measuring forecast accuracy i ACF of Beer Production sea Figure 2.6: The corelogram (or ACF pot) for the brer production data. series willbe very holpful in verifying the pattern ‘The ACF also helps us identify if previous values of the series, contain much information about the next value, or whether there is litle relationship between one observation and the next, ‘To sum up, much is to be learned about a single time series by examining the autocorrelations of the series with ied, lagged one Period, two periods, and soon. The ACF plays a very importaat role in time seties forecasting, 2/4 Measuring forecast accuracy ‘We now turn to another fundamental concern—how to meisure the suitability of a particular forecasting method for a given data set In mest forecasting situations, accuracy is treated as the overriding criterion for selecting a forecasting method. In many instances, the word “accuracy” refecs to “goodness of fi,” which in turn refers to how well the forecasting model is able to reproduce the data that are2 one-step forecat Period Observation Forecast ‘ % Ro T 13s Tw. 2 136 130.50 3 132 15725 4 at 143.30 5 151 138.00 6 130 127350 7 ug 13825, 8 153 14150 9 = 140.50 10 167.25 ‘Table 2-8: The lst eight beer production figures and forests obtined by taking ‘the average ofeach month oer the pst foor pears Chapter 2. Basic Forecasting Tools already known. ‘To the consumer of forecasts, itis Uhe accuracy of the future forecast that is most important. In this section, a vaiety of measures of forecasting (or modeling) accuracy will be defined ‘and in subsequent chapters these measures will be used in the context of worked examples To ilsteate the computations involved, we will refer to the Aus tralian monthly beer production. Table 2-8 shows the last 8 months of observations (Janvary-August 1995). The second column shows orecasts for these values, obtained using a very simple method, by taking the average of each month over the past four years. So, for ‘example, the forecast for January 1995 i the average production for January 1901, January 1992, January 1999, and January 1994, 2/4/1. Standard statistical measures ILY; is the actual observation for time period t and F; isthe forecast for the same period, then the error is defined as aah (2.2) Usually, Fy is caleulated using data Yi,...,¥in1. It ia one-step forecast because itis forecasting one period ahead of the last obnet- haan 2/4 Measuring forecast accuracy ‘ation used in the calculation. ‘Therefore, we describe ey a8 a one-step forecast error. Ibis the difference between the observation ¥, and the forecast made using all the observations up to but not including ¥,, If there are observations and forecasts for n time periods, thea there will be m ecror terms, and the following standard statistical measures can be defined: Table 2-9 illustrates the computation of these standard statistical Equation (2.12) can be used to compute the error for each petid. ‘These can then be averaged as in equation (2.13) to give the mean srror. However, the ME is likely to be small since postive and negative ecrors tend to offset one another. In fact, the ME wil nly tell you if there is systematic under- or overforecasting, called the forecast bias. It does not give much indication as to the size of the ‘typical errors. ‘Therefore, the MAB is defined by first making each error positive by taking its absolute value, and then averaging the results A similar iia is behind the definition of MSE. Here the errors are made positive by squaring each one, then the squared errors are averaged. The MAE has the advantage of being more interpretable and is easier to explain to non-specialists. ‘The MSE has the advantage of being sasiec to handle mathematically (and 20 it is often used in statistical optimization). Each of these statistics deals with measures of accuracy whose size depends on the scale of the data. Therefore, they do not facilitate comparison across different time series and for diferent time intervals. An error-of 10 MI when forecasting monthly beet Production is quite diferent from an eror of 10 Ml when forecasting a mean sbalte er ean susted ereChapter 2. Basic Forecasting Tools 2/4 Measuring forecast accuracy elation percentage ‘Absolute Squared 1140.20, | Period Observation Forecast Error Error Error | t Ms A, =F Wv-Fil (i- AP as 3.50 5.5 1650-27225 1300 169.00 2.50 6.35 1925 370.56 115013225 | using equation (2.13) using equation (2.14) using equation (2.15) ‘Table 2-9: Computations of the standart measures forthe beer data. annual beer production or an error of 10 MI when forecasting the ‘water consumption of a city. To make comparisons like these, we need to work with relative or percentage error measures lest we need to define a relative or percentage error as oer ven a)« (2.16) e~ (25-2) cam, a1 ‘Ther the following two relative measures ate frequently wed: mean percentage ee = LDPE (2.7) ean absolute LS pe, aa frcniage ere yy Equation (2.16) can be used to compute the percentage eror for any time petiod. These can then be averaged asin equation (2.17) to give the mean percentage atrot. Howover, as with the ME, the MPE is likely to be small since positive and negative PEs tend to offset one 45 Percent Absolute Period Observation Forecast Error Error Percent Error ‘ A (Hp) 002 100 T Toa —i2a5— — = 3g 2 13950-35026 26 3 18725525 as 35 4 M350 1650-130 bo 5 133001006 86 5 1750250 19 19 7 13825-1925 162 162 8 Miso 180 75 13 Total En) we using equation (2.17) using equation (2.18) ‘Table 2-10: Computations ofthe percentage measures for the beer dat. nother. Hence the MAPE is defined using absolute values of PE in ‘equation (2.18). Table 2-10 shows how to compute the PE, MPE, and MAPE measures, From the point of view ofthe ultimate user of forecasting, knowing. that the MAPE of a method is 5% means a grest deal more than simply knowing that the MSE is 183. However, the MAPE is only meaningful if the scale has a meaningful origin. For example, one would not use MAPE for assessing the accuracy of temperature forecasting since the common temperature scales (Fahrenhelt and (Celsius) have fairly arbitrary zero points. Difficulties also arise when the time series contains zeros, since the percentage error (2.16) cannot then be computed. (When the time series values ace very close to ‘ero, the computations involving PE can be mesaingless.) 2/4/2 Out-of-sample accuracy measurement ‘The summary statistics described so far measure the goodness oft of the model to historical data. Such fitting does not necessarily imply ‘od forecasting. An MSE or MAPE of zero can always be obtainedoe UP ABTIOLECY ¢ 6 Chapter 2. Basic Forecasting Tols 2/4. Measuring frees accuracy “ im the fing phase by asing a polynomial of suficienty high order (i NH Abate Abe omnsitng Ovecttng a model toa data series, which i equivalent to including end One en eee eee eae ] randomaes as part ofthe generating process, i as bad as falling to | i a ee | ‘deatty the systematic patters in the data : a it easel | A second drawback of these measures of accuracy is that that 2 5 18 2 13 l diferent methods use diferent procedures inthe fting phase. For : 3 32 ise 2 is ‘example, smoothing methods (Chapter 4) ae highly dependent upon i 4 7 12 25 137 initial frecating estimates; decomposition methods (Chapter 3} i 5 1s rr By 139 : include the tend-ycle in the ting phase as though it were known; i 6 130 i a 162 ‘ regression methods (Chapter §-8) minimize the MSE by giving equal 7 iia 30 a ae ; weight to all observations; and Box-Jenkins methods (Chapter 5 153, uo a 22 ; minimize the MSE by 2 non-linear optimization procedure. Thus, Tout 7 ae comparison of such methods on a single eitvion i of lenited value oe bles we by iste i MAE ‘using equation (2.17) : se problems can be overcome by measuring true out-of sample Ae forecast accuracy. That the total data are divided into an “inital: MAPE = 127.1/8 = 159% sing equation (2.18) test set ization” set and & “est” st or *oldost” set. Thea, the aitaiation hold ot se is und to etinave any parameters and to initialize the method. Table 2-14: Computations o he percentage measures forthe NEY free f the rely E a comparing freast Forecasts are made forthe test set. Since the tast set was not used in wer dats ‘the model fitting, these forecasts are genuine forecasts made without using the values of the observations for these times. The accuracy measures are computed for the errors in the test set only. 2/4/3 Comparing forecast methods [None of these measures give a good basis of comparison a5 tothe gains in accuracy made by applying a specfc forecasting method. Does a MSE of 5 or a MAPE of 3.2% indicate a good or bad forecast performance? One basis for making such a comparison isto define some very simple naive methods against which the performance of ‘more sophisticated methods can be compared. We have found it useful to define two different naive methods of forecasting for use as a basis in evaluating other methods in a given situation, The fist is refered to as Naive Forecast 1 ot NFI ‘This method uses the most recent observation availabe asa forecast. ‘Table 2-11 shows NFL used to forecast the monthly beer production Each forecast is produced by taking the value of the previous month's production. So the forecast for January 1995 isthe production figure from December 1994, the forecast for February 1096 i the production figure from January 1995, and so on. The diference between the MAE or MAPE obtained from a more sophisticatad method of forecasting and that obtained using NFI provides a measure of the improvement attainable through use of that more sophisticated forecasting method. This type of comparison is much more useful than simply computing the MAPE or MAE of the first method, since it provides a bass for evaluating the celative accuracy of thove results. In this case, the fist forecasting method achieved a MAPE of 7.8% compared to about twice that for NF and ‘8 MAE of 10.5 Ml compared to 22.1 Ml for NFI. Clearly the fst method provides much better forecasts, [A second naive method of forecasting has also been found to be extremely useful as a basis for evaluating more formal forecasting ‘methods. This method is referred to as Naive Forecast 2 or NF2 and. goes beyond NPI in that it considers the possibility of seasonality la the series, Since seasonality often accounts for a substantial percentage ofthe fuctuation in a series, this method can frequently do much better than NFI and yet is till very simple straightforward Naive Forecast 243 ‘Chapter 2. Basic Forecasting Tools ero The proce to cove sonal fom the ofa Period Observation Forecast Numerator Denominator datain order to obtain seasonally adjusted data, Once the seasonality fru)? (Haar)? has been removed, NF2is comparable to NPI in that it uses the most u - prea re) i recent seasonally adjusted value as a forecast forthe next seasonally a Bs 150-25 C0008 0.0002 adjusted value, In practice, NF2 allows one to decide whether or 2 el pe os 0.0138 ‘ot the improvement obtained from going beyond a simple seasonal ; al 182 13725 00118 oz adjustment of the data is worth the time and cast involved. 4 : = ame 010s 0.0367 a 4 8 151 13800 0.0008 0.0193 | 6 130 verso 0029 oor 2/4/4 Thei’s U-statistic 7 u9 13825 0.0093 o.0si6 8 153 14150 = = ‘The relative measures in the previous section all give equal weight to Total 560 OE all erorsin contrast to the MSE, which squares the errors and thereby ‘emphasizes large errors. It would be helpful to have a measure that : considers both the disproportionate cost of large errors and provides ‘Theil U = Va Tpag = 9580 2 relative basie for comparison with naive methods, One measure that has these characteristics is the U-statste developed by Theil (900 ‘Table 2-12: Computations involved in determining Theis U-statstic for the beer orcats This statistic allows a relative comparison of formal forecasting ‘methods with naive approaches and also squares the errors involved to that large errs are given ach move weg than small ros by simplifying eto the frm shown in (220). When the valves of ‘The positive characteristic that is given up in moving to Theil’s U: PPErs1 and APEt,1 are substituted into equation (210), the result statistic as a measure of accuracy is that of intuitive interpretation, ‘This difficulty will become more apparent as the computation ofthis statistic and its application are examined. Mathematically, Theil's U-statistic is defined as Serta - APE)? Theis Uatiie v = a9) (22 aaaaaae Fare? Comparing the numerator of equation (2.20) with equation (2.18) mt shows that it is similar to what was defined previously ae the MAPE ofa given forecasting method. The denominator i equivalent to the sumerator with Fi replaced by ¥. So tis simi tothe MAPE of [NFL. Thus, the U-statistic isan accuracy measure that incorporates You =¥ both concept. . where FP Exes (forecast relative change) and APE: (actual relative change) = i a ‘Table 212 shows how to compute Theil's U-statstic forthe beer ata, Equation (219) is actualy very straightforward, as can be seen ‘Theil Ucatatnic can be better undertood by examining itsChapter 2. Basic Forecasting Tools interpretation. The value of the U-statistic given by equation (2.19) will be Oonly if FPEv4y = APEras fort = 1,2,...,2—1. That in turn ‘occurs only when the forecasts are exact (give 0 error). Alternatively, the U-statistic will have a value oft only when FPE;41 is equal to 0 ‘That would be the case only if the errors in the forecasting method were the same as those that would be obtained by forecasting no ‘change at all in the actual values. That is comparable to assuming ‘an NF1 approach. If FPE:4: isin the opposite direction of APE,4:, the U-statiatic will be greater than unity since the numerator will be larger than the denominator. The ranges of the U-statistic can thus be summarized as follows: U =: the naive method is as good asthe forecasting technique being evaluated, U < 1s the forecasting technique being used is better than the naive method. The smaller the U-statistc, the better the forecasting ‘technique is relative to the naive method, > 1: there is no point in using a formal forecasting method, since ‘sing a naive method will produce better results 2/4/5 ACF of forecast error One other tool for analyzing forecast etrot needs to be mentioned. ‘The autocorrelation function of the one-step forecast ertors is very useful in determining if there is any remaining pattern in the errors (or residuals) after a forecasting model has been applied. This is not ‘a measure of accuracy per se, but rather can be used to indicate if ‘the forecasting method could be improved. For example, suppose the naive forecast method NF was used for the beer data. We analyze the forecast errors by a time plot and ACF plot as shown in Figure 2-7, [Note that there isa patcetaremsining in these errors—the January value is particulary low each year. Clearly, there is not what would be called a random set of errors. The autocorrelation statistics are sensitive to such patterns. ‘The ACF plot in the lower panel of Figure 2-7 tells a similar story. The autocorrelation at lag 12 is much larger than the other 2/4 Measuting forecast accuracy st Errors from NF1 forecasts, Figure 2-7: Top: Forecast errors obained by applying the NF method tothe ber | data "Bottom: The ACF ofthe forecst errors autocorrelations. This shows there is some seasonal pattern in the cervor series (the low January values are 12 months apart) {tis important not to read too much into the other autocorcelations shown in Figure 2-7. With random series, no autocortelations wil be exactly zero, even if the sores is entirely random, These small fluctuations around zero are quite acceptable and do not indicate52 ACF erica waive diction intervals Chapter 2. Basic Forecasting Tools there s information in the seies which is not being captured by the forecasts, 1 flo ave cht trina atsrin issignfcanly large. A nimple rule isto only consider antocorrela- tion that are larger than the erin salve of 2/7 in magnitude (Why ths work wil be discussed in Chapter 7.) or the ber data, f= 3600 the eral values are at 2/V56 =O Figure 27 shows the boundary as wo horizontal ines a 807, Autocorrelation that fall within se boundades ean be enlay ignored, Autocorelatons hat il outside the boundaros suggest there may be some adional information inthe series whichis not being captured by the forecat inated In this example, only the autocorrelation at lag 12 falls outside the critical values, This is another indication that there is some seasonality in the forecast errors, and that the forecast method could be improved. This is not surprising since the forecast method NFL does not use the seasonal pattern in producing forecasts 2/5. Prediction intervals 1 soy din o ov nl ove an a a onpanying uncertainty statement, sully inthe form of prediction interaal, This wsfl because provides the wer of the forecast ith “wont or “be” cae estates and wth a snee f how Aependable the frear sand becae i protects the forecaster from the etc thatthe frecana ate “wroog.” Forecast eannot be expected tobe perfect and Intervals emphasc this Prediction intervals are usually based on the MSE because it provides an estimate ofthe variance ofthe one-step forecast error, So the square oot ofthe MSE'Ss an estimate ofthe standard deviation of| the forecast error. The usual assumption for constructing prediction intervals is thatthe forecast etrors ate normally distributed with zero ‘mean. Under this assumption, an approximate prediction interval for the next observation is Fas # 2VISE. ‘The value of z determines the width and probability of the prediction interval, For example, 2 = 1.96 gives a 95% prediction interval. That 2/5. Prediction intervals is, the interval has probability of 95% of containing the true value, as yet unknown, For other percentages, different values of = ean be used. ‘The table below gives the most common values of =. 2576 0.99 Other values of = are given in Table D of Appendix II. ‘The rules of thumb given on page 32 are based on this table ‘The September 1995 forecast for the beer data can be calculated as the average production for the previous four Septembers. This sivas Fees = 140.50. Table 2-9 gives the MSE for the beer forecasts 88 192.52. So a 90% prediction interval is 140.50 & 1.645Vi42.53 = 140.50 19.64 = [120.86, 160.01 ‘That is, we can be 90% sure that the actual beer production figure maber 1995 wil lie between 120.86 and 160.01 megaliters. (In was 14 megaliters) A similar ealeulation gives the October 1095 forecast as 167.25 MA with a 90% prediction interval of 167.25 + 1.645VTN253 = 167.254 19.64 47.61, 186.89 ‘The actual figure for October was 165 Ml. Although October is two periods ahead of the last observation, these ate both one-step forecasts since the forecast method is based ply on data from previous years. The forecast method treats the ddata from each month as separate seies. That is why we could use the same MSE value in computing the prediction interval for Gctober. If we had used the NFI method, then the forecast for October would have been a two-step forecast, and the MSE would not have been ‘ald for calculating a prediction interval This procedure only works for one-step forecasts since the MSE 's based on one-step forecasts. For multi-step forecasts, we need to 354 step MSE asco aor Chapter 2. Basic Forecasting Tools modify the MSE to be based on multi-step forecasts. One approach is to define the A-step MSE as where of") is the error from making an A-step forecast of the obser vation at time ¢. Then, if we assume the h-step forecast error is normally distributed with zero mean, we have the prediction interval ——a Fra 2VIBEs Before calculating any prediction intervals in this way, the errors should be checked to ensure the assumptions of zero mean and normal distribution have been met. 2/6 Least squares cstimats In Chapter 1 we introduced two kinds of quantitative forecasting ‘models: explanatory models and time series models. An explanatory model for GNP is of the form GNP. = f(monetary and focal pois, inflaton, capital spending, imports, exports, ero) whereas a time series model is ofthe form GNP sett. Neither model can be exact. That is why the error term is included fon the right-hand sides of these equations. The error term represents, vatiatios in GNP that are not accounted for by the relationship J. (GNPs, GNP yt, GNPy=2y GNP, In both cases, what is observed as the output of the system is dependent on two things: the functional relationship governing the system (or the pattern, as it will be called from now on) and randomness (or etot). That is, data = pattern + ertor. 21) 2/6 Least squares estimates The critical task in forecasting isto separate the pattern from the error component so that the former can be used for forecasting The general procedure fo eatimating the pattern of a elatoahip, ‘whch epanory rin sar throug ing some aan form in such a way ast miniias the ever componett of eqetion (221), One form o is eimation test guar, This approach vet ld (developed ft by Gans in the 1800) ad isthe one noe ‘idly wed In clasts ‘The name feast squares is based on the fact that this estimation rocedure seeks to minimize the sum of the squared errors in equation (2.21). The example shown below illasteates the basis of the least ‘squares method. Its application to all types of functional forms (ie, linear oF nonineae) is analogous to that shown here. Suppoe that che manager ofa supermatkt wats to kw bow mmc typical canoer spends inthe sore, Th isthe sore The manager might Stat by taking a sul fay 12 clea at random besig oe raat Shown fn Tale 339 From Table 2-1, itis clear that not all customers spend the same ‘amount. Some of the variation might be explained through factors ‘such as time ofthe day, day of the week, discounts offered, meximum oF minimum amount of checks cashed, and 20 on, while part of the variation may be random or unexplainable. For purposes of this ilustration, it will be assumed that no variation can be explained through explanatory or time setiea relationships. In such a case, the store manager faced with finding an appropriate estimator to describe lest squares a (3) | T a 7 rn 2 8 5 7 3 ° ° 8 4 2 0 9 5 9 u vn 8 2 2 10 55 Client Amount Spent | Client Amount Spent56 Chapter 2. Basic Forecasting Tools 2/6 Least squares estimates 57 Brimate of Tiimaieof __Eatimateat | ¥ v=o an | e! Amount Beer Exot Exoe \ lene “pent _Eroe* squared __Evror_aquated Etro squared . Teas Teenie laa aa teeeaaD eas 1 ot 2 « 8 eee eae tee ese a] eee ee rapeer eae ey Bares See eee gece ee ce ¢ qa 58 eae a 7h sees toot 1 cai iO eeeo gees Gece % 8 ones 8% 308 1 o 9 i aE EEC gE 5 ° thee 16 toot 1 Cae ae ener Bo sity Maen yes ‘ | = SSE (eum of sunt eo) 8 a Sina aie MSE (mesn square etor) a 3 cial *Brror = umount spent ~ estimated value | ‘Table 2-14; Mean squared errors for estimates of client expenditure | the data may take a fixed value as an estimate. Having made this decision, the manager might decide to select an estimate in such a way as to minimize the mean (average) squared error. This could bbe done by trial and error. Suppose we denote the estimate by the symbol ¥. The manager tries values of ¥ = 7, ¥ = 10, and ¥ = 12. ‘The resulting mean squared errors are shown in Table 2-14 ‘From Table 2-4 it is clear that the squared error is least when the manager chocees 10 asthe estimate. However, there may be a better estimate. Figure 28 shows the resulting MSEs forall estimates from 0 through 20, and it can be seen that the MSESs form a parabola Furthermore, the minimum value on this parabola is indeed at the point where the estimate i 10. Thus, the minimum MSE will be Achieved when the value of the estimate is 10, and we say that 10 is the least squares estimate of customer spending, Figure 2-8: Relating MSE to estimate sane Because Figure 24 is mathematical function whi ie : ical function whose properties a be found exactly, it is not necessary to use trial and ertor to find the estimator that minimizes the MSE. Rather, this value can be found mathematically with the help of differentiation. The frst step 5 to renite equation (221) so a toaolate the eco on the lef-hand ‘error = data ~ pattern (222) ‘As usualy the error will be denoted by ¢, the data by Y, and the sstimate pattern by ¥. In addition, the subscript i (i= 1,2,3,-..,12) will be added to denote the ith’ customer. Using this notavion, sauation (2.22) becomes: e; = ¥; —¥. Then the sum of squared SE=) d= m-¥)? > 2) ‘nd the mean squared error is, MSE: Deeee eae pSse= GLU -¥)58 © Chapter 2. Basic Forecasting Toole Now intially the value ¥ will not be known, but the store manager wants that value of ¥ which will minimize the sum of the squared errors (or, equivalently, to minimize the MSE).* ‘This can be found by taking the derivative of (2.23), setting it ‘equal to zero, and solving for Yas follows: 4856 _ Si -¥) so that ‘which implies val y yer. (2.24) Solution (2.24) a easily recognized aa the mean of the data, and it ives a value that minimizes the sum of the squared errors. Applying (2.24) to the siore manager's data in Table 2-13 gives Fa ee Le 120 PoPop hep ‘This value is the minimum point of Figure 2-8. As a single point «stimate of the pattern of the data, the mean fits the data as closely 8 possibe, given the criterion of minimizing the MSE. While the ‘mean is a somewhat simple estimate of the data in most situations, the procedure of least squares that was used to determine a MSE. estimate can be applied no matter how complex or sophisticated the ‘estimation situation is, minimum MSE Its ofcourse possible to minimize some other criterion (e., MAE) instead of minimising the MSE. However, minimizing MAB is not ‘as ety mathematically as minimizing the MSE. Also, squasing the “Note thr minniing SSE (the sr of quar eros) tthe “eat equates” ‘roceure Dividing by (whch 17 inthe example gen) gives the MSE. Thus Pimmining the MSE san exactly equivalent procedure 2/6 Least squares estimates errors magnifies (or gives more weight to) extreme values, and this result is attractive because large ertors are lee desirable than small certors. (Many cost relationships are quadratic in nature, suggesting, the appropriateness of squaring.) 2/6/1 Discovering and describing relationships If the measurable output of a system is viewed as data that include 4 pattern and some error, major consideration in forecasting, ‘whether explanatory or time series, is to identify and fit the most appropriate pattern (functional form) so a8 to minimize the MSE. ‘The basic procedure isillusteated by comparing two diferent methods for forecasting the price of a Japanese automobile using the data for the 19 vehicles listed in ‘Table 25. The first method is simply to use the mean a8 a forecast (as was done forthe store expenditure data). Table 2-15 gives the price data along sith the Forecasts and the seaulting eros ‘The forecasting model underlying Tuble 2-15 is Fao. (225) ‘That Is, the forecasts are the same for all vehicles because we are ot sing other information about the vehicles. The value of a is ‘stimated from the data to be equal to the mean of the data. Hence, Yea=v. Using the mean as the estimate of the pattern might be acceptable it we had no other information about these vehicles. But we have already seen that price is correlated with mileage. We can ute the rileage information to form a more accurate estimate of price than ‘the mean. Figure 2.9 shows a straight line fitted to the prie-mileage felationship, Notice that the erors (vertical dstaaces froma the points to the line) are smaller in this plot than the ertore obtained using the mean as an estimate Using a straight line forecast model with mileage as the explana- tory variable means ¥ma+bx mileage (2.28) Where’ a and ate suitably chosen values representing the intercept And slope ofthe ine. The values of a and ban be chosen inthe same 59 {right ne modeChapter 2. Basic Forecisting Tools 2/6 Least squares estimates Table 2-15: The mean at an estimate of he price of Joponese automobile, a pice cee Stet Priceimileage relationship for Japanese automobiles Tata 15998 107 Tor 14709 13.938 0.88 OTA 2760 13938 1082 ITI 14929 13938 099008 13949 13938 OL 00 1787913038 15.55 11680 13.958 324 25.300 13.938, 87.65 17.809 13.938 15.69 21498 13.998, st.5 13219 13.998 048 9.590 13.038 B83 1.939 13.938 810 13.965 13.938 000 Tom 1.938 075 ‘5m 1a ae san steno aio 13.998 2050 5866 13.998 55.16 88 13.998, Total 265.823 965.525 0.00 P= mg 3.938 simple lin regreton ‘way as a was chosen in (2.25): that is, by minimizing the MSE. This procedure is known as simple linear regression and will be examined in detail in Chapter 5. The mechanics of how a and 8 are calculated ‘are not important at this point. For these data, the values of a and 6 which minimize the MSE are 32.1 and 0.735 respectively. So the forecast model is y 12.1 ~ 0.795 x mileage. (zn ‘This lin is shown in Figuee 2-9. Figure 2-9; Straight line estimate of vehicle price using mileage as an explanatory variable 1s import not contin an eatin ltetip (cha sha betten pc ad mag) with cua an Beer higher age dou aot co lover prcarcith ieped eo cas sch a the sino he capes The stage Ine ee Sodel (227) al when we tno te milage tise sd ah Credit a pe For forecasting purposes, either the mena forecasting model of the straight line forecasting model can be used to predict the price of 8 Japanese automobile not listed. For example, if a vehicle has a nileage of 23 mpg, the mean gives a forecast of $13,038 while the straight line model gives a forecast of [82.1 ~ 0.736(23)] x 1,000 $15,176. From historical data, one would expect the straight line oil tobe etter se had alr MSE ne (b00ced with 27.63), : ie A little care must be taken in comparing MSE values (or any ‘other measure of accuracy) from different models. More complicated causal reationhina (Chapter 2. Basic Focecasting Tools Squared Price Mileage Forecast Error error | ($1000) (mpg) (S'000) 9 19 Tells 31 1007 1479919 13418-3319 1.014 24760 20 i382 7378 54.420 1492920 17382-2453 6.019 139% 20 17382-3433 11.788 i879 2 weedT 12321518 1165021 16.647 4997 24971 23.3001 16.647 6653 44.261 1789922 15912 1987 3.049 ras 15.176 6322 39.962 132194 asl -1192 142 9509 35. 13708 4107 16.866 1098925 13708-2717 T38 139827 12235 17102923 | nual 28 vison UsTL 2468 650932 8559-1960 3.840 gal 33. Te231sa7 2817 5565 4 7088 1922 1493 6.43835, 6353 0.135 0.018 Sum of squared errors SSE): 764.915 (MSE = 246.913/19 = 13.00 Table 216: Straightline estimate of vehicle price using mileage ax explanatory arabe. Straight line formule: ¥ = 32.1~0.735( mileage). models generally have smaller MSE values, even if they do aot sive more accurate forecasts. This is because they measure the goodness of ft of the model to historical data, rather than true utof-sample forecasting performance. However, in this case the scatterplot (Figure 2-5) and negative correlation (Equation 2.) both indicate that mileage should be included in the forecasting model Also, the straight line model has only half the MSE of the mean model. Such a reduction is unliely to be due to the additional 2/7 Transformations and adjustments complexity of the steaight line model In general, there is no way that a statistical method can automati- cally detecmine the best pattern (functional form) to describe a given set of data. Rather, this decision must be based on judgment, Then 4 statistical method can be used to ft the specified pattern in such way as to minimize the MSE. 2/T Transformations and adjustments ‘Sometimes, adjusting the historical data will lead to a simpler and ‘more interpretable forecasting model. In this section we deal with three kinds of adjustment: mathematical transformations (such as Jogacithms and square oots) ate discussed in Section 2/7/1; adjust. ‘ments to remove data variation due to the effects of the calendar are discussed in Section 2/7/2; and adjustments due to population ‘changes and inflation are digenseed in Section 2/7/2 2/T/A_ Mathematical transformations Figure 2-10shows monthly Australia electricity production data, the same data that were plotted in Figure I-2a (p. 7). Notice that the size ofthe annual seasonal vacation increases asthe level of the series increases. At the start ofthe series the total variation throughout the ‘year was only about 800 million Kwh, but in the most recent years, When the production is very high, the total variation is over 2,500 million kwh. Clearly any forecasts for these data must take account of the obvious increasing trend, the strong seasonal pattern, and this Increasing variation with level. A mathematical transformation is & convenient method for accounting forthe increasing variation. One such transformation is the square root function. The top plot in Figure 2-11 shows the square roots of the electricity production data. The new data set was formed simply by taking the square root of each observation in the original data set. This mathematical ‘transformation has helped in reducing the variation in the size of the annual cycles, making it easier to forecast these data than those shown in Figure 2-10, Square roots are only one kind of transformation that can be rmathematis! trantormation sate rast tesntormationamass Chapter 2. Basic Forecasting Tools 2/7 Transformations and adjustments 6s Square root of electricity production “anaaansassoOS Leva | 1 30 1990 ‘a0 eeess i Cube root af electricity production yes | 99 a 1980 1 Figure 210, Mhly Anaration clus iity prauction from Janaary 1956 82 August 1988. Note the increasing variation as the level ofthe serie sncreases, Logarithm of electricity production used in this way. Many other tragsformations are possible, but in Panam Hass tn epics sn sed bp eos So Ligh, 8) fe aeaas eae Cosen hetooes nyoe cee csr Netr aerECUEEEEE in log value are relative (percent) changes on the original scale, 1980 1a 10 ry Other useful transformations are given in Table 2-17. Here we denote the original observations as Yi,...,Yq and the transformed observations as W,...,W,. Figuee 2-11 displays the electricity data transformed using some of the transformations given in Table 2-17, showing the effect ofthe increasing strength of the transformations Each ofthe transformations in Table 2-178 a member ofthe family power waneormaton of power transformations: { -¥f, P<; Negative reciprocal of electricity production log(¥), P= 0; (228) Yr, P>o. tre 2-11: Tronsformations ofthe electricity production data. Of these, either @ square rot, cube roc, oral transformation could buted to atablize the variation ( form a serie which hav variation approsimaely constant over the erin For p the data alone. Choosing p = gives a square root and the transformation is simply W, = Yj; so this leaves =66 Chapter 2. Basic Forecasting Tools (oo eee weve | Cube root We=a% — Incceasing Logacithm W,=log(¥i) strength Negative reciprocal Wi=-W/% ‘able 2-17: Mathematical transformations for stabilizing variation. Dacktranforing igves the negative reciprocal. It might seem artificial to define the me transbrmation ft p= 0 tobe the logarithm, bo actly Togs thee because 1 orp coe to sar behaves much le he Toga Forp <0, the npative ofthe powe raaormation toed so tha all ranaormatione cre lacresing eneons (1, the tanformed wale acosn a Y, ican). Th parca pean be any number he data ar pani, but? must be grat Team zero ifthe data have seon If he data have angie sae, te power ansformation pombe vl they are ajstd st by ‘dg ueontan taal ln Forecuts are clesated on the tanfomed data cathe than the cvigial data, But since we te cal Intra infor of the Gril data, oot the ransomed dala wet eee the rao tmaon (or ac-rufom) toca focara onthe orignal scale For example the revere othe age rot faction the oeare fancon, th reverse othe lath fenton ithe exponential fanction, and soon, Ceneraly there power tuafrmations are given by eee {atte p=; (war, p>. For example, if we were to forecast the square root of the electricity, ata, we could obtain forecasts on the original scale by taking the square of the forecasts of the square root data Ita prefeabe to chow a simple ale fp to ie tar maton ech wt thas ve ia Table 17. Model aa frect o& Gime ves ae rlatvly ise to the value op choer-nearby taee ofp wil produce slr ota This sen In Figute 211 se 2/1 Transformations and adjustments ~ where cither p = 1/2, p = 1/3, of p = 0 could be used to stabilize the variation. Also, simple values of p such 28 0, =1, or 1/2 make the results much easier to interpret than a number like p = 0.38463, Very often i is found that no transformation (i.e, p= 1) is needed. When the data have been transformed, then prediction intervals also need to be transformed back to the original scale. The simplest way to proceed is to apply the inverse transform to the end points of the prediction interval. So, if logarithms have been used, and the forecast on the log scale is Fy, and the prediction interval is (Lagas Ung1)s then the Forecast on the original seale is ef with the prediction interval (eM, Uns). Note that these prediction intervals need not be symmetric atound the forecast. Empirical studies which have considered the merits of mathe- matical transformations have demonstrated that, for many series, transformation doesnot often have a major effect on forecast accuracy (Makridakis and Hibon, 1979; Makridakis et al, 1982; Meese and Goweke, 1984) hic ic becauce moot forecast metho place more weight on the most recent data, Therefore the small annual variation eacier in the electricity series is unlikely to influence the forecasts very much. Only when the series is rapidly changing in vaviation will mathematical transformations make a large difference to the forecast, However, the MSE (and the other measures of accuracy) gives ‘aval weight to all data and so prediction intervals will be afected by transformations. In calculating prediction intervals, it is assumed that the variation is approximately constant over the series Some of the vaciaton ina time series may be due to the variation in the number of days (or trading days) each month, It is a good idea {0 adjust for this known source of variation to allow study of other Interesting features, : Month length can have quite large effect, since length ean differ by about Ugg = 10%. If this is not cemoved, it shows up 48 a seasonal effect, which may not cause problems with forecasts ‘hough it dors make any seasonal pattern hard to interpret. is month length sdjustmant68 trading dy adjustment Chapter 2. Basi Forecasting Tools easly adjusted for: no, of days in an average month yo. of days in month £ 385,25/12 * Go. of days in month ¢ Figure 2-12 shows monthly milk production per cow over a period of U4 years. Month length wil, of course, affect the total monthly milk production of @ com. Figure 2-12 shows the milk production data adjusted for month length. The simpler pattern will lead to better forecasts and easier identification of unusual observations. ‘Trading day adjustment is similar to month length adjustment ‘buts not completely predictable. Trading days adjustments are often necessary because a given month may not have the same number of working, or trading, days in different years. In some industries such as etal sales and banks, this factor becomes very important, since it can have « significant influence on the lal of alee. This source of variation occurs in monthly data when there is also a weekly cycle, since the proportions ofthe various days in a given month vary from yeat to year. For example, March may have four or ive Sundays, and if Sunday is a non-trading day this must be accounted for. While the various proportions are complately predictable from the calendar (like month length adjustment) the effects of the various days are not predictable so this must be estimated. Inthe simplest case, days are classed as either trading or non- trading days, and all trading days are assumed to have the same effect. [a this ease the adjustment is analogous to month length: no, of trading days in an average month “go, of trading days in month? where ¥; has already been adjusted for month langth and transformed if neceseary. More complicated cases are discussed in Section 6/2/2. Ww, 2/T “Transformations and adjustments ‘Monthly milk preduction per cow ‘Agjustea monthly rk production per cow A Figure 2:12 ony mk proton per 4 pare (Su + ev oer 14 sare (Sours Cry, 1085). The scand graph ho he deta ode or he kengh of wields a simpler pattern enabling better, 2 witceton of enna ayia np Forcats ond easier identification of ensue0 _adjuting for inflation adjusting or population changes (Chapter 2. Basic Forecasting Tools 2/7/3. Adjustments for inflation and population change: One obvious source of variation that afflicts time series isthe effect of inflation or changes in population. For example, when forecasting the price of a new motor vehicle, itis essential to take into account the effect of inflation. A $15,000 vehicle this year is not the same ‘5 & $15,000 vehicle 10 years ago. The standard approach is to tse equivalent value in 1900 dollars (for instance). ‘Then the data ‘ae directly comparable and forecasts wil not be affected by this additional source of variation. Adjusting for population changes is similar. For example, when forecasting the number of public transport users in a city, it is preferable to take into account the effet of population changes. In this case, the data could be adjusted by the total numberof people in the city, Rather than forecasting the total aumber of public transport users, it will probably be more accurate to forecast the proportion ‘of people who are public teansport users. Demographic studies are needed to provide forecasts of population and these can then be used to obtain forecasts of the number of public transport users in the fature. A more refined approach would be to produce forecasts for different age groups and/or different socioaconomic groups. Appendix 2-A Notation for quantitative forecas Quantitative forecasts are based on data, or observations, that de scribe some factor of interest. I thie book a single observed value will be cepresented by ¥., (See Table 2-18.) This variable can be the actual number of units sold, the cost of production, the advertising budget, price per unit, gross national product, of aay other event of interest, as long ast can be quantified. The abjective of forecasting i to predict future values of Y. The individual forecasts will be denoted by Fy ot Yo, and the ertor by er, where the error is the difference between the actual value and the forecast value for observation i aa oo 4 ia F, In time series forecasting and in explanatory forecasting, when the data are taken at equal time intervals, a will denote the present time Petiod, m1 last period, n ~ 2 two periods agn, and so-on. A period fan be a day, a week, a month, quarter, year, and so forth, The forecasts usually will be for future time periods such as n + 1 Observed Values Forecasted Values Move % Yer Ye Period ¢ Ean a a [n+ a+? .. ntm Estimated values Yi ¥; ¥y Yat Ya | Yar Pasa Yasm ARG. Fos Fa | Fon Fae Fatn Error a8 6Appendix 2-B Summation sign © In order to simplify-the manipulation of expressions involving the adding of many numbers, itis convenient to use a summation sign, E. The uso of this sign and the elements of notation mentioned Previously can be demonstrated using the data in Table 2:19 Based on Table 219, Y¥; is the actual sales value, Yror Fis the forecast values for sales, and 4 is the error or difference ¥; ~ Yj, fone wants the sum of the errors, it can be obtained from atatotote = De or 3-3-2- -t= 3 Below and above the summation sign are the “limits” showing the ‘variable whichis indexing the sum (t) and the range of the summation (from 1 to 23). If itis obvious what the limits are, sometimes they are omitted, ‘Tae cumulative sales for the years 1985 through 1994 can be obtained from bg Yan + Yia + Yia t+ Yao ATS 4175 4176 +--+ 251 20m, n Append 2.8: Summation sign © No of Unite Sd Pete Actual Forsagt Exe ‘Table 219: Use of quantitative forecasting notation ‘The following rules apply to the use of summation signs: 1 DPW = POY, where F isthe sample mean (therfore a constant) of the variable ¥; a Bran, a De-%y Svshy: 4 Dore Sy Seedy ae 5. Dm -¥y Doe -2PK% +74 Dvn daar Srp —art =ER- dim “2Bry Chapter 2. Basic Forecasting Tools References and selected bibliography ARMSTRONG, J.S. and F. Cottory (1992) Error measures for ‘generalising about forecasting methods: empirical comparisons (sith discussion), International Journal of Forecasting 8, 69-111. BieaMax, H., C.P. BONIM, and W.H. HAUSMAN (1991) Quanti tative analysis for business decisions, 8th ed., Homewood, Il Richard D. lewis. ‘CuaTriELD, C. (1993) Calculating interval forecasts (with diseus- sion), Journal of Business and Economic Statistics, 11, 121-144, Coccea, K.O. (1979) “Time series analysis and forecasting with an absolute erior criterion,” in TIMS Studies, Vol. 12, Forecasting, SS, Makridakis and S.C. Wheelwright, eds., Amsterdam: North Holland, 89-102. Cowtey, D-L., G.S. KraHenpunt, L.N. BuRKerT, and A.L. MILLAR (1981) Physiological correlates of female road racing performance, Research Quarterly Ezereve Sport, 52,4148, Consumer reports (1990) Vol. 58, No. 4, Consumers Union, Me Vernon, NY, Caver, J.D. (1986) Time series analysis, Belmont, Cal: Duxbury Press. Fiuoss, R. (1902) The evaluation of extrapolatve forecasting meth- ‘ode (with discussion), International Journal of Forecasting, 8, ae Hottoway, C.A. (1979) Decision making under uncertainty, Engle- wood Clif, NJ: Prentice Hal Koosis, D.J. (1972) Business statistics, New York: John Wiley & ‘Sons. Locks, F.M. (1972) Business mathematics, New York: John Wiley 1 Sons. Maxrupakis, S. and M. Hisox. (1979) Accuracy of forecasting: ‘an empirical investigation (with discussion), Journal of the Royal Statistical Society A, 142, 97-145. References and sxlcted bibliography Mavaionss 8. A. ANOERSES, R. CARBONE, RFD, M. fox, R Lewsvoowant, J. Newton. B, Panzen, and R. Whe Kun (182) The acarcy of extrapolation (ime sts) metho Ses of a orcasting competition, Jowna of Foret 111-153, i: oe MeLavontm, RL (1975) The Real Record ofthe Economie Fore asta, Buinesa Boonie, 1, No. , 2898 Messe, R. and J. Gewese (1984) A comparison of axtorgrsive univdate foreeting procedure for mactoeconome tas ta Journal of Business and Economie Satin, 2, 19-200 Movrcowsny, D.C. L.A. Jowwson, and J. Ganotnen (180 Forecasting’ and tine series anin, 2nd ed New Yor McGraw-Hill. i woe Newaoio,P. and. Bos (194) Introductory business oct 2nd ed., Cincinnati, Ohio: South-Western Publishing Co. . Srums, W.A. and C-P. Bont (173) Statistica nag for busi ete deisions, revise Homer lt Richart Deve Sriece, B.M. (062) “The Eralention of Forecasts” in Handbok of Forecasting 5, Maksdain and S.C, Wheelwright its Now Yorke Joba Wyte Sons 467-08, Twi, (196) Aptiedecmomicforeting, Amsterdam: North Holland, 26-32. b a ea Rice, JA (195) Matematica! tates and data analy, 2nd Belmont, Cal.: Duxbury Press. Sa Wowsscorn, TH and RJ. Woxvacort (199) Inmdsctory statin or business an comm Sth es New Tote Se Wiley & Sons, ae 15Chapter 2. Basic Forecasting Tools Exercises 241 Table 2-20 gives average monthly temperatures in Paris (=) What is your best estimate of the average temperature in June 19987 (b) Make a time plot of the data. Is there any time pattern in the temperature readings? Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 194 7.6 71 83 15 17 172 185 197 151 89 85 85. 1995 77 69 61 105 129 ‘Table 2-20: Average monthly temperature in Pars (degrees Celis) 2.2 For each of the following series, what sort of time patterns ‘would you expect to see? (@) Monthly retail sales of computer disks for the past 10 years at your local store, (b) Hourly pulse cate ofa person for one week, (c) Dally sles at a fast-food store for six months. (a) Weekly electricity consumption for your local area over the past 10 years. 12.8 For each of the following series on the web page, make a graph of the data (using a computer package), describe the main features and, if transforming seems appropriate, do so and describe the effect. (a) Monthly total of people on unemployed benefits in Aus- tralia January 1956-July 1992). () Daily morning temperature of a cow for 75 days, (©) Number of lynx trapped annually in the McKenzie River district of northwest Canada (1821-1934) (a) Monthly total of accidental deaths in the United Scates (anuary 1973-December 1978). (e) Quarterly production of bricks (in millions of units) at Portland, Australia (March 1956-Septembec 1904). Erercies 1. Daly morning temperatur of cow n 2. Accigertal deinen USA mony) i [ a 3. ntaatona aitine passengers : VAM nf 4 Annual mink wappings (Canad) ; 8 weet hy il TIT m7 ; pees [lime &Chapter 2: Basic Forecasting Tools 8 24 In the graphs on the previous page, four time series are plotted slong with their ACFs. Which ACF goes with which time 2.8 Table 2-21 shows data on the performance of 14 trained female distance runners. The variables measured are the running. time (iinutes) in a 10 kilometer road race and the maximal aerobic power collected during the week following the run, (a) Calculate the mean, median, MAD, MSD, and standard deviation for each variable (b) Which of these statistics give a measure ofthe center of data and which give a measure of the spread of data? (©) Caleulate the correlation of the two variables and pro- duce a scatterplot of ¥ against X, (a) Why is it inappropriate to calculate the autocorrelation of these data? X Ae Ge OS TH BAe oe aN oo oo ce wn a oar abso Woo 432 azo) «ar 4999 4490 4.90 4.12 aS00 1609 ATS 4855 “Tnble 2-21: Running times (¥) and masimal aerobic capacity (X) for i female runners. Source: Conley etal. (1981) 2.6 Columa 1 on the fllowing page is the actual demand for prod- uct E15 over 20 months. Columns 2 and 3 are the one-month ahead forecasts according to two different forecasting models to be discuss ia Chapter 4. (Method 1 gives forecasts from ‘Table 4-4 and Method 2 gives forecasts from Table 46.) (a) Plot the actual demand on a graph along with the fore- caste from the two methods (b) For each method, compute the Mean Error, Mean Abso- lute Error, Mean Squared Error, Mean Percentage Error, and Mean Absolute Percentage Error using equations (2.18) through (2.18) (c) Repeat Part (b) using columns 1 and 3 below. Which forecasting method appears to be better? Exercises Oey ay Actual Method 1 Method 2 Period Demand ForecastForecast 1 139 187 T70 2 rus 162 3 mo 137 4 e182 i 5 owl M9 tos 6 et 158 7 1 oe 168, 3 et 179 2 om wr i 1 206169 180 M13 199 2 wre, 202 3 Boe a Mo R93 za 5a 22 16 aa wom 25 ww mo 22 19 em 23 2023025 243 27 Download the Dow Jones index fom the web page an Ac time plot othe sna ings camper peage (0) Caelate the change in the index fo each day by a teatng the vale for the previous day. (This owe as “differencing” the data and is discussed in Chapter 6.) (b) Forecast the change in he nde foreach ofthe net 20 ay by taking the average ofthe histo chngea (©) From thee forcats, compete forecasts fo the ogi inde foreach he 20 dae (2) Aad the forcast to the graph (0) Show thatthe grap forcast ae deta! extend ing he ine drawn Between th fst and at oeeration80 Chapter 2. Basic Forecasting Toole 198011043 1981 11180 1982 10732 1983 11112 198411465 1985 12271 1986 12250 lost 12240 12700 1326 | 1954 17 1048 10 194819 ‘Table 2-22: Japanese motor vehicle production (1947-1980) in thousands. Source World motor vehicle data, Motor Vehicle Manufocturers Association of U:S. Ine, | Detroit, 199 2.8 Japanese motor vehicle production for 1947-1989 is given in ‘Table 2-23. (@) Plot the data in a time plot, What features of the data indicate a transformation may be appropriate? (b) Transform the data using logarithms and do another time plot. (€) Caleulate forecasts for the transformed data foreach year from 1048 to 1990 using Naive Forecast 1. (4) Compute the forecast errors and calculate the MSE and =MAPE from these errors. (©) Transform your forecast for 1990 back to the original scale by find the exponential of your forecast in (e). Add the forecast to your graph. (£)- From the graphs you have made, can you suggest a better forecasting method? (g) The world motor vehicle market was greatly affected by the oll ersis in 1973-1074. How did it afect Japanese motor vehicle production? If this information could be included in the forecasts, how would it afect the values of the MSE and MAPE? TIME SERIES DECOMPOSITION n 4 Principles of decomposition 84 Wi Deanpoitin noes 4 Wa Becfaton paps 87 WS Sexeal adusisen 88 Moving werages 89 YAN Sipe mong ages 89 Var Cored mong wages 2A Sabie mony Soaps 38 2M Went matey aes 8 local regression smoothing 10! an tees Wt . Classical decompostion 106 342 ianpste emp oi) 345 Vat on cs ‘eons Hd 35 Census Bureau methods 113, {SF eration 1H 52 Laer is 5B Ensen 0 XA2-ARIMA 19 36 STL decomposition 121 YW lever oop 22 W682 Oster op 5 WA Choong ie STL puaetes 24 ‘Wo Compa STU wen XARA DA 37 Forecasting and

Das könnte Ihnen auch gefallen

• The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  

  Von Everand

  The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  Mark Manson

  Bewertung: 4 von 5 Sternen

  4/5 (5794)
• The Psychology of Money: Timeless lessons on wealth, greed, and happiness

  

  Von Everand

  The Psychology of Money: Timeless lessons on wealth, greed, and happiness

  Morgan Housel

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (804)
• The Hobbit

  

  Von Everand

  The Hobbit

  J.R.R. Tolkien

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (25134)
• • Zeitschriften
  • Podcasts
  • Noten
• Never Split the Difference: Negotiating As If Your Life Depended On It

  

  Von Everand

  Never Split the Difference: Negotiating As If Your Life Depended On It

  Chris Voss

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (3275)
• The Hobbit

  

  Von Everand

  The Hobbit

  J.R.R. Tolkien

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (24585)
• Rich Dad Poor Dad

  

  Von Everand

  Rich Dad Poor Dad

  Robert T. Kiyosaki

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (1110)
• The Handmaid's Tale

  

  Von Everand

  The Handmaid's Tale

  Margaret Atwood

  Bewertung: 4 von 5 Sternen

  4/5 (13186)
• It Starts with Self-Compassion: A Practical Road Map

  

  Von Everand

  It Starts with Self-Compassion: A Practical Road Map

  Celeste Headlee

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (188)
• The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  

  Von Everand

  The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  Mark Manson

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (20020)
• The 7 Habits of Highly Effective People: 30th Anniversary Edition

  

  Von Everand

  The 7 Habits of Highly Effective People: 30th Anniversary Edition

  Stephen R. Covey

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (41)
• Elon Musk

  

  Von Everand

  Elon Musk

  Walter Isaacson

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (437)
• American Gods: The Tenth Anniversary Edition

  

  Von Everand

  American Gods: The Tenth Anniversary Edition

  Neil Gaiman

  Bewertung: 4 von 5 Sternen

  4/5 (12945)
• The 48 Laws of Power

  

  Von Everand

  The 48 Laws of Power

  Robert Greene

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (2263)
• Orgullo y prejuicio: Clásicos de la literatura

  

  Von Everand

  Orgullo y prejuicio: Clásicos de la literatura

  Jane Austen

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (20517)
• Los secretos de la mente millonaria: Cómo dominar el juego interior de la riqueza

  

  Von Everand

  Los secretos de la mente millonaria: Cómo dominar el juego interior de la riqueza

  T. Harv Eker

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (814)
• Never Split the Difference: Negotiating As If Your Life Depended On It

  

  Von Everand

  Never Split the Difference: Negotiating As If Your Life Depended On It

  Chris Voss

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (838)
• Matar a un ruisenor (To Kill a Mockingbird - Spanish Edition)

  

  Von Everand

  Matar a un ruisenor (To Kill a Mockingbird - Spanish Edition)

  Harper Lee

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (23003)
• The Courage to Be Disliked: The Japanese Phenomenon That Shows You How to Change Your Life and Achieve Real Happiness

  

  Von Everand

  The Courage to Be Disliked: The Japanese Phenomenon That Shows You How to Change Your Life and Achieve Real Happiness

  Ichiro Kishimi

  Bewertung: 4 von 5 Sternen

  4/5 (393)
• The 5AM Club: Own Your Morning. Elevate Your Life.

  

  Von Everand

  The 5AM Club: Own Your Morning. Elevate Your Life.

  Robin Sharma

  Bewertung: 4 von 5 Sternen

  4/5 (342)
• American Gods [TV Tie-In]: A Novel

  

  Von Everand

  American Gods [TV Tie-In]: A Novel

  Neil Gaiman

  Bewertung: 4 von 5 Sternen

  4/5 (12520)
• To Kill a Mockingbird \ Matar a un ruiseñor (Spanish edition)

  

  Von Everand

  To Kill a Mockingbird \ Matar a un ruiseñor (Spanish edition)

  Harper Lee

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (22902)
• Good Omens: A Full Cast Production

  

  Von Everand

  Good Omens: A Full Cast Production

  Neil Gaiman

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (10905)
• And Then There Were None

  

  Von Everand

  And Then There Were None

  Agatha Christie

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (8934)
• Good Omens

  

  Von Everand

  Good Omens

  Neil Gaiman

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (12010)
• Pride and Prejudice: Bestsellers and famous Books

  

  Von Everand

  Pride and Prejudice: Bestsellers and famous Books

  Jane Austen

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (19653)
• The Perfect Marriage: A Completely Gripping Psychological Suspense

  

  Von Everand

  The Perfect Marriage: A Completely Gripping Psychological Suspense

  Jeneva Rose

  Bewertung: 4 von 5 Sternen

  4/5 (1107)
• It Ends with Us: A Novel

  

  Von Everand

  It Ends with Us: A Novel

  Colleen Hoover

  Bewertung: 4.5 von 5 Sternen

  4.5/5 (5754)