Practical Business Forecasting - UNTAG

ACKNOWLEDGMENTSI would first like to thank Donald P. Jacobs, the “legendary Dean” at the KelloggGraduate School of Management at Northwestern University, who providedthe opportunity for me to teach this course at Kellogg. I would also like to thankDipak Jain, who was Professor of Marketing and has now replaced Don Jacobsas Dean, for many helpful comments and suggestions.Al Bruckner, the former Executive Editor at Blackwell Publishing, has beenmost helpful in guiding me through the writing and publishing process, sticking with me “through thick and thin” to develop this textbook in its presentform. In a day and age when editors often seem to change at the drop of a hat,it has indeed been a pleasure to have someone of Al’s experience and expertiseto work on the book from start to finish.This book has also benefited from numerous conversations with my son,David, who has built several credit scoring models for a major financial institution – which, I am proud to report, had continued increases in earningsthroughout the recent recession even as competitors posted sharply lower earnings or actual losses.It has become pro forma in these pages to thank one’s wife, but in this case,I owe an extraordinary debt of gratitude to my wife, Susan. For more than 30of the past 40 years, she has worked with me in my various companies, hasheard the good, the bad, and the crazy, and has always served as an invaluablesounding board. Without her encouragement this book never would have beencompleted.

PREFACEPractical Business Forecasting is designed to appeal to a wide range of academic,corporate, and consulting economists who have interest or responsibilities inforecasting at the macroeconomic, industry, or individual company level.The text first discusses various methods of forecasting and alternative goalsthat might be desired. It then turns to a discussion of econometric and regression analysis, followed by material on time-series forecasting, and concludeswith a wide variety of practical forecasting applications for both singleequation and simultaneous-equation models.The material in this text is written at a level for those who have some familiarity with basic statistics, but is not designed to be a theoretical treatment.Proofs are not given in the text, and no matrix algebra is used, although references are provided for those with further interest.The emphasis of this book, as shown by the title, is in building practical forecasting models that produce optimal results. Sometimes, robust theories cannotbe empirically implemented because of data limitations. Often, standard testswill tell the practitioner little or nothing about whether the equation can forecast accurately. The text is not intended to be an encyclopedia or long reviewarticle. Hence references are quoted only when clearly relevant. Practical hintsabout how to build a forecasting equation that works are given much higherpriority than extended discussion about properties of normal distributions.In the 1960s and the early 1970s, most economists and practical businessforecasters alike thought that econometric models would provide the mostaccurate forecasts, both at the macroeconomic level and for individual industries and product lines. However, the results were particularly disappointing inthe late 1970s and the early 1980s. As a result, several other methods of forecasting moved to the forefront: statistical methods that did not involve economictheory but were based on lagged variables and previous error terms; surveys ofconsumer and business sentiment; consensus forecasts; and informed judgment. Several economists suggested that forecasting accuracy could be significantly improved by combining these methods. Currently, there is no general

PREFACExviiagreement among economists and forecasters about which methods are mostlikely to generate forecasts with the smallest error.Interpreting forecast records will always be a subjective undertaking. Unlike,say, the stock market, where a 20% increase or decrease in the S&P 500 indexis an immutable record once it happens, data on which forecasts are based areoften subject to revision.There is also the question of being “right for the wrongreasons”; if someone forecasts continued growth next year because they missthe signs of emerging weakness, but the economy begins to soften and themonetary authorities manage to avert a downturn by timely easing, is that agood forecast or not? At the company level, who is the better forecaster: theanalyst who predicts that sales and earnings will rise next quarter, which is whatthe company initially reports, or the analyst who predicts that sales and earnings will drop, which is what finally emerges after the SEC investigates thecompany for fraudulent accounting practices?For these reasons, it is unlikely that anyone will ever establish a definitivemethod for benchmarking true ex ante forecasts. Nonetheless, it is usuallypossible to determine, for any given time frame, whether forecaster A providedmore accurate results than forecaster B, and to identify the methods used byeach forecaster. Over time, as the evidence accumulates, it should be possibleto identify which methods work better than others.This author has been generating macroeconomic, industry, and companyforecasts for almost 40 years; obviously not all of these forecasts have been successful. Nonetheless, this experience has provided a clear picture of what worksand what does not work. A fairly extensive track record of alternative methodsof forecasting at the macro level also provides additional information aboutwhich methods work best.Based on this information, this text takes the following position. Most ofthe time, the best forecasts are generated using properly specified and properlyestimated econometric models, adjusted for recent developments and structuralshifts in the underlying functions. That may appear to be a broad generalization, but in fact leads to some very specific conclusions. Mechanical methodsof forecasting, such as correlating the current variable with its own lagged valuesand previous error terms, generate inferior forecasts. At the macro level, surveysof consumer and business sentiment usually generate larger forecastingerrors than do econometric models. Combining various forecasting methodsthat individually provide inferior results does not usually improve the forecasting record.There are a few exceptions, notably short-term forecasting in the financialsector, where econometric models have so far proven to be inferior. Becausethe impact of human interaction and reaction is so great in these areas, shortterm forecasts based on econometric models are no better than naive modelsthat assume the change next period will be the same as the change this period,or the average change in the past. Over the longer term, though, underlyingeconomic relationships become more important and can be helpful in gaugingchanges in market behavior.

xviii PREFACEThe weight of the evidence this author has been able to accumulate stronglysuggests that the proper use of econometric models will generate the most accurate forecasts. Hence most of the book focuses on the methods that should beused to estimate these models. Some material is included on time-series models– generally known as ARIMA models – and on sentiment and judgment, butthe evidence suggests they have not improved forecasting accuracy over the pastyears and decades.Having said this, it should be emphasized that forecasting is an art, not ascience, and even the best econometric model needs to be adjusted regularly togenerate optimal forecasts. That will indeed raise the cry in some quarters thatit is really judgment, rather than econometrics, that is driving the forecast, andthose complicated multiple regression equations are merely an attempt to distract the unwary client into believing that a tremendous amount of intellectualeffort went into building this sophisticated model structure, whereas in fact theequations are largely extraneous and the real forecasts are based on judgmental factors.Once again, it is virtually impossible to provide an unbiased answer to thatcharge; the consultant will claim that “of course” the econometrics are the keyingredient, while the critic will answer “nonsense.” However, it is not necessaryto engage in a battle of histrionic proportions to provide an intelligent framework in which to evaluate this claim. Consider three alternative and competinghypotheses.1 Underlying structural relationships remain stable over time, although they aresometimes jolted by exogenous forces.2 Because people learn from past mistakes, behavioral patterns of the past willnot be repeated in the future, and hence any attempt to project the future fromthese past patterns is doomed to failure.3 No underlying structural relations are really stable, but they change slowly.If researchers and forecasters believe that (1) most accurately describes thestate of the real world, then the classical postulates of statistics and econometric would apply. If they believe that (2) is most accurate, which is essentiallythe view postulated by the rational-expectations critics of models, than econometric and statistical methodology are inept and inaccurate methods to use forforecasting. If they believe that (3) is most accurate, which is the hypothesisadvanced by Charles Bischoff in a recent article on forecasting discussed indetail in chapter 12, then one ought to proceed by estimating econometricmodels but adjusting them frequently.The underlying focus and purpose of this book is based on the hypothesis that(3) most accurately describes the underlying economic forces in the real world, andhence concentrates on this approach.The material in this text is divided into five parts. Part I provides some basicbackground information about the forecasting process. Chapter 1 discusses thegeneral concepts of forecasting accuracy and describes alternative types of

PREFACExixforecasts. Since no model is any better than its data, chapter 2 discusses whereto find useful data and how to check for possible errors. A brief statistical reviewof goodness-of-fit statistics is also included. Since the examples in this book arebased on the EViews statistical package for the PC, this program is brieflydescribed.Part II focuses on estimating and forecasting with econometric models thatconsist of a single equation. Chapter 3 provides a brief review of the standardclassical linear model; a detailed statistical treatment is not included, since thatis available in many other textbooks. Also, most of the underlying assumptionsare not applicable to practical business forecasting. Hence after setting thestage, the remainder of chapter 3 discusses how forecasters should treat autocorrelation and heteroscedasticity. Chapter 4 covers the topics of eliminating orreducing spurious trends, and determining the optimal lag structure, issues thatare not often treated within the confines of the classical linear model but are ofgreat importance for practical business forecasting. This chapter also explainsthe proper use of dummy variables and estimation of nonlinear regressions.Chapter 5 presents and analyzes a battery of tests designed to determine thestructural stability of the equation, and also discusses methods for adjusting theconstant terms in forecasting equations. Most of this material in this partfocused on the econometric approach; the contributions of judgment are notdiscussed here.Part III covers the standard material for univariate time-series modelingand forecasting. Chapter 6 shows how a time series can be decomposedinto the trend, cyclical, seasonal, and irregular components, with particularemphasis given to smoothing data and methods of seasonal adjustment.Chapter 7 presents the development and rationale for ARIMA models –regressing a variable on its lagged values and previous error terms. The issueof trend removal and integration is also discussed in this chapter. Whilethese methods can be useful for tracking models containing thousands ofequations, several examples are introduced to show that, most of the time, thesemethods do not generate very accurate forecasts compared with econometricmodels.Part IV focuses on the forecasting accuracy for single-equation models, asopposed to the underlying econometric and statistical methodology. Chapter 8discusses the situations under which combining forecasts is likely to providemore accurate results, and when it probably will not. In general, combiningforecasts will not improve forecasting accuracy if the forecasts are all based onthe same underlying data set, assumptions, or methodology. In contrast, forecasting accuracy can be significantly improved by adjusting the constant termsin the regression equations. Chapter 9 provides several examples of actual shortterm sales forecasting models, including the key steps that are important inpresenting these forecasts to management or consulting clients. Chapter 10turns to a variety of statistical and judgmental methods for long-term forecasting, where it is generally assumed that the underlying structural relationshipwill shift over time.

xxPREFACEPart V covers additional forecasting issues associated with multi-equationmodels. Chapter 11 discusses the issues associated with simultaneous equationbias, which is shown to be a much smaller contributor to forecasting error thanimproperly specified equations. Chapter 12 presents the results of a small prototype macro model, and shows how forecasting accuracy is affected by adjusting the constant terms of the model, using an autoregressive adjustment, andcombining these results with surveys, consensus forecasts, and informedjudgment.Throughout the book, case studies are used to illustrate how these modelsare actually estimated and adjusted to generate true ex ante forecasts. An oldstory has it that, in the old days of secretaries and Dictaphones, one novicetranscribed “econometrics” as “economist’s tricks.” In the minds of many economists and business executives, that aroma still lingers. This book attempts todispel that concern, to erase the “black box” image, and explicitly present theactual methods and adjustment procedures that are used. It is clear that in manycases, the choice of variables that are used, the form in which they enter theequations, the lag structures, and the adjustment of the constant terms, are allmatters of judgment. However, after reading this book, it is also hoped that thereasons behind these choices will become clearer, and students and readers willhave a firmer foundation to use in estimating econometric models used forpractical business forecasting.

Practical Business ForecastingMichael K. EvansCopyright Michael K. Evans 2003PART I

Practical Business ForecastingMichael K. EvansCopyright Michael K. Evans 2003CHAPTER 1CHOOSING THE RIGHTTYPE OF FORECASTINGMODELINTRODUCTIONPractical business forecasting is both a science and an art. It is a science in thesense that correct use of sophisticated statistical tools will invariably improveforecasting accuracy. It is an art in the sense that empirical data seldom if everprovide an unequivocal answer, so the user must choose between alternativerelationships to select those equations that will provide the most accurateforecasts.There are no perfect forecasts; they always contain some error.While perhapsthat is obvious, it is nonetheless important to emphasize this fact at the outset.The point of this book is to show how to minimize forecast error, not to pretendthat it can be eliminated completely. To accomplish this goal, a variety of forecasting methods may be used. In many cases, these methods will be complementary, not competitive.Forecasts can be used for many purposes. Sometimes, predicting the direction of change is sufficient. For example, a model that could accurately predictthe direction of the stock market the following day – even without providingany information about how much it would rise or fall – would be extremelyvaluable and profitable. No such model has ever been successfully constructed,although many have tried, and the goal will presumably remain elusive. At theother extreme, a model that predicted the direction of change in the consumerprice index (CPI) the following month without forecasting the magnitudewould be virtually useless, since over the past 40 years the monthly changes inthe CPI have been negative only about 1 percent of the time.There are many ways of forecasting, not all of which are based on rigorousstatistical techniques. In some cases, informed judgment can provide the bestforecasts, such as when “insiders” have company information that is not available to anyone else. Surveys may provide useful information about forecasts forthe overall economy, specific sectors, or individual industries and firms. To theextent that these methods improve forecasting accuracy, they should be utilized.

4CHOOSING THE RIGHT TYPE OF FORECASTING MODELNonetheless, there is no rigorous way of testing how much informedjudgments or survey techniques have boosted forecast accuracy, so they arementioned only peripherally in this text. Instead, this text concentrates on illustrating how statistical and econometric methods can be used to construct forecasting models and minimize forecast errors. Initially, most economic forecastswere generated with structural equations; more recently, time-series analysishas been utilized more effectively. The benefits and shortcomings of bothmethods for generating optimal forecasts are identified.This book is not a theoretical text; the emphasis is placed on practical business forecasting. As a result, theorems and proofs, which can be found in manyother texts, will be kept to a minimum, with most of the material related toactual forecasting examples. In particular, this text will illustrate how

Useful Tools for Practical Business Forecasting 29 Introduction 29 2.1 Types and Sources of Data 30 2.1.1 Time-series,Cross-section,and Panel Data 30 2.1.2 Basic Sources of US Government Data 32 2.1.3 Major Sources of International Government Data 34 2.1.4 Principal Sources of Key Privat