Business For Casting

8/6/2019 Business For Casting

1/17


2/17

VICTOR ZARNOWITZ is Professor of Finance in theGraduate School of Business of the University of

Chicago. His main research interests are in the areas

of macroeconomic theory, business fluctuations, and

time-series analysis. He received his Ph.D., summa

cum laude, from the University of Heidelberg in

1951. A native of Poland, he acquired Americancitizenship in 1957. He has been a member of the

research staff of the National Bureau of Economic

Research for the past dozen years. In 1953-54, he

held a Social Science Research Council postdoctoral

fellowship at HarvardUniversity; lectured at Co-

lumbia Universitybetween 1956 and 1959; and was awarded a Ford Foundation faculty research fellow-

ship in 1963-64, to pursue studies of the variability

of investment demand. He joined the faculty of the

Graduate School of Business in 1959.

. Mr. Zarnowitz is the author of monographs and

papers on the theory of income distribution, busi-

ness cycle indicators, and the cyclical behavior of

manufacturers orders and prices. Recently, he has

been directing a study of the accuracy of short-term

economic forecasts for the National Bureau of Eco-

nomic Research. This Selected Paper is based in part upon a talk he delivered at the 6th Annual

Meeting of the National Association of Business

Economists in Philadelphia on September 29,1964.


3/17

II Forecasting

Business Conditions:

! A Critical ViewSHORT-TERM economic forecasting, as widely

practiced today, is largely an art or game ruledat best by experienced and disciplined judg-

ment, at worst by sheer luck. It probably willalways contain much of these elements, but atrend is to be expected toward increased appli-cation of scientific methods of evaluating the

evidence and drawing inferences from it.In recent years, forecasts of the nations eco-

nomic fortunes have become much more abun-j dant and ambitious than ever before. It is now possible to assemble a fair-sized collection of

continuous forecasts from well-reputed sourcesfor several major economic aggregates and in-dexes, as we have done in a study currentlyunderway at the National Bureau of Economic

Research. As one tries to extend the record1 back to the early postwar and the prewar years,sources quickly dry out.That forecasts grew in boldness as well as inquantity can be seen in the fact that many arenow expressed in specific numbers. Vague,hedged, or purely qualitative predictions ofwhats ahead for business are still quite com-

mon, but they no longer dominate. Also, at-tempts are increasingly made to predict the

course of the economy over a sequence of shortperiods-say, the four quarters of the yearahead-and this represents a particularly am-bitious, dynamic type of forecasting.

These developments reflect increased de-mand for forecasts of economic conditions.

I Business management clearly has a very largeshare in that demand, and its preference is forunconditional, specific, numerical predictions.However, the demand for forecasts is diversi-


4/17

fied as well as large. For example, the forecastsrequired by government policy makers (an-

other large source of demand) differ from thosesought by the business and financial commu-nity in that forecasts conditional upon alterna-tive policy courses are precisely what is re-quired by the former.

Changes contributing to the growth and spe-

cialization of forecasters output occurred alsoon the supply side of the market for new eco-

nomic intelligence. The amazingly rapid de-velopment of electronic computer technologyaccelerated greatly the rate at which economicdata (the raw materials for the forecaster) arecompiled and processed. It also had some moredirect effects-without the computer, the large-scale econometric models could not have beenproduced, hence output of forecasts of theeconometric variety would have been limited.However, work with such models is still essen-tially in the domain of academic economists.The great majority of forecasts are producedby business economists, who have thus far ap-peared to make very little use of formal econo-metric models.

Whatever services the forecasts are expectedto render to the user, they vary, and are noteasily defined by an outside observer. How-

ever, the usefulness of forecasts is surely in thefirst place a function of their accuracy.

With the growth of public interest in the

expanding activity of economic forecasting,there is increasing need for objective and com-prehensive evaluations of the forecasts. Thisneed is as yet largely unsatisfied. It is surpris-ing to note how little systematic testing hasbeen done in this area-despite the widespread

use of business activity forecasts, their costs,the potential rewards of good predictions andpenalties of bad ones, and the consequent im-portance of the ability to discriminate amongthe available sources and methods.

In attempting to survey the field we must


5/17

3

ask ourselves: How is the accuracy of economicforecasts to be assessed? How useful are such

assessments as can be made, in terms of a quan-titative description of forecasting errors-theirmagnitudes, types, and structure? What infer-ences can be drawn about the dependability andusefulness of the forecasts? What about thefeasibility of improvements? Though research

on these problems has lagged badly, we haverecently begun to make some real headway inattacking these questions.

This is not the place to review the literatureon the subject, but one recent study is so im-portant that it must be mentioned: The workof Henri Theil on the methodology of forecastevaluation and the accuracy analysis of certainEuropean forecasts both of the business surveyand the econometric-model variety.1 And thecurrent National Bureau study of U.S. mate-rials, in which I am engaged, already has pro-duced some interesting findings about the ac-curacy and other characteristics of aggregativeshort-term forecasts. Let me now turn to somecentral issues in economic forecasting, as re-vealed by these recent explorations.

The Hazards of Economic ForecastingThat economic forecasting is a hazardous

art is common knowledge. The precise reasonsfor this are not so well understood.

It is of some help here to consider the typicaleconomic time series to be predicted as a com-posite of four factors: trend, cyclical, seasonal,and purely irregular or random movements.

Trend fittings and projections are often rela-tively successful in application to long-termforecasts, but in many cases they can play onlya subordinate role in the short-run context.(However, for some series trends are impor-tant even over short periods, and will often bewell approximated by simple methods, which

1 Henri Theil, Economic Forecasts and Policy, North-

Holland Publishing Company, Amsterdam, 1958.


6/17

4

should facilitate the forecasters task consid-erably.)

Strictly periodic, repetitive fluctuationsshould also be rather easy to predict: stableseasonal movements would often be more orless of this type, but forecasters usually workaround them, trying to forecast the season-

ally adjusted series.This leaves the cyclical and irregular com-ponents as the main sources of trouble for theshort-term forecaster. Looking forward, it isanything but easy to distinguish the cyclicalfrom the random element in the movement of

an economic time series, though retrospective-ly it is usually possible to do so with fair re-sults.2 The forecasting errors that are directlytraceable to very short random movementsmust really be accepted as unavoidable. Theforecaster can hardly be expected to predict an

event generally regarded as unforeseeable suchas an outbreak of a war (e.g., Korea, 1950) ora strike started without advance warning.

Though such shocks cannot themselves bepredicted by the techniques of economics, theirm o r esignificant effects on the economy are, of

course, the proper concern of the forecaster.The requirement of a good forecast is that it

predict well the systematic movements, trendsand cycles-not that it predict perfectly theactual values of the variables concerned (itcould not do that, except by accident, for eco-

nomic series-which, as a rule, contain randomelements). And of the systematic movementsit is the cyclical fluctuations, not the longertrends, that produce the greatest difficulties inshort-term forecasting. These fluctuations arerecurrent but nonperiodic; they vary greatly in

duration and amplitude; calling them cyclical2 The practice used most frequently for this purpose

is to pass a moving average of intermediate lengththrough the seasonally adjusted series and get the de-viations of the smoothed from the unsmoothed values

as estimates of the irregular component (which can thenbe tested for their randomness properties).


7/17

5

should convey neither more nor less than that

they reflect mainly the participation of thegiven economic factor in the business cycle.

I shall illustrate later the importance of thebusiness cycle as a source of forecasting errors.Meanwhile, some related trouble-making fac-tors must be noted. One is the lack of accurate

information about the conditions prevailing atthe time the forecast is made. The initial levelor base from which the predicted change ismeasured must itself be predicted; and al-though they are estimated at a close range, thebase figures often contain significant errors.For example, in predicting the level of GNPnext year, errors made in estimating the baseoften contribute as much as 30-40 per cent tothe total forecast error.

Better estimates of current position could

improve substantially the forecasts themselves.Moreover, such common targets of forecasters

as the nations aggregate output are exceed-ingly difficult to measure or even to define.The forecasts also frequently involve factors ofpresumed importance which are very elusive,

such as the state of business confidence.Where measurement is difficult, and estimatescan have substantial errors, prediction seemsparticularly hazardous.

Size and Direction of Errors

Let us now take a closer look at the accuracyof forecasts of business conditions.

The errors of the annual forecasts of thegross national product (GNP) in current dol-lars (both over- and underestimates) averagedabout 7-11 billion dollars in the years 1953-

1965. The change in GNP during this periodaveraged about $22 billion per year.3 The er-rors, then, tended to be approximately one-third to one-half the size of errors that wouldbe produced by a naive model which as-

3 Computed without regard to sign.


8/17

6

sumed the predicted years level of GNP to

be the same as the previous years.Compared with average levels of GNP, these

errors would amount to no more than 1 1/2 - 2 1/2per cent. But short-period changes in GNP aregenerally small relative to already attainedlevels (their order of magnitude here is 4-5 per

cent). The recent recessions were all short andrelatively mild. A margin of plus-minus twoper cent of GNP can mark the difference be-tween a good and a bad year.

Forecasts of average levels of economic ac-tivity in the coming year (expressed in termsof GNP and its major components and the in-dustrial production index) are in general moreaccurate than mere mechanical extrapolationsof the past. The forecasts examined provedsuperior not only to the simplest naive mod-

el extrapolations of the last known level orchange, but also to some much more demand-ing standards of trend extrapolations and au-toregression models.4 In beating the predic-

tions produced on the computer by weighting,averaging, and extrapolating past values of the

given series, the annual and shorter forecastsby economists scored what must be regardedas a significant success-even though the bestof the mechanical yardsticks against which theywere measured leave something to be desired,and occasionally the margins of success have

been slight.Business forecasters have been called cau-

tious or conservative because of their tendencyto underestimate changes. The data analyzed

in the current National Bureau study confirmthis observation emphatically. Underestima-

tion applies to both increases and decreases,and is evident in forecasts relating to differentvariables. Underestimation of increases usuallyresults in underestimation of the ensuing lev-els. Underestimation of decreases, analogously,

4 Statistical models of the relationship between pres-ent value and past values of a given economic series.


9/17

7

tends to result in over-estimation of levels. In

series with upward trends, such as GNP orindustrial production, increases are more fre-quent than decreases so that the levels are un-derstated most of the time.

To the extent that underestimation reflectedmerely a smoothing out of the random com-

ponent of the actual values in the forecastingprocess, it could not be objected to as a typeof systematic error. There is ample evidence,however, that the observable tendency goes farbeyond that and constitutes indeed a true bias.

The average changes in actual values gener-ally exceed those in predicted values, whethertaken with or without regard to sign.

The tendency to underestimate changes ison the whole stronger in simple mechanicalextrapolations than in the forecasts proper

over short spans of time. But forecasts andextrapolations probably have much in com-mon on this point. The primary dependenceon the data of the recent situation itself canbe a basic source of the bias. Forecasters neces-sarily rely on the stability of some relation-

ships observed in the past, but these in fact areundergoing changes. As a result, stability isexaggerated; that is, change is understated.

We find that forecasts of rather good quality

often have been made for the very near future-the next quarter or six months. (The averageannual forecasts can be viewed as having meanspans of little more than six months, too.)However, with further extension of their reachinto the future, short-term forecasts deterioraterapidly.

A few examples will be sufficient to demon-strate the regularity and pervasiveness of thisrelation. In a semi-annual forecast of GNP for1955-63, representing an average of a fairly

large group of individual predictions, themean absolute errors of the relative change


10/17

are: for a six-month span 1.5 per cent, for a

twelve-month span 2.6 per cent. In a quarterlyforecast by the staff of a large company for thesame years, the mean absolute errors computedanalogously in per cent are: for one quarter1.1, for two quarters 1.8, for three, four, five,and six quarters 2.5, 2.9, 3.4, and 3.8, respec-

tively. Again, in an semi-annual forecast of theFederal Reserve index of industrial production

for 1947-63, the mean absolute errors are 2.8per cent for six, 5.8 per cent for twelve, and9.5 per cent for eighteen months!5

Why should the accuracy of short-term fore-

casts be a sharply decreasing function of thespan of the forecast? Let us think of these fore-casts as consisting of any or all of the followingingredients:

1. Extrapolation, of some kind, of the past

behavior of the given series.2. Relation of the series to be predicted to

known or estimated values of some othervariables.

3. Any other external information consid-ered relevant, e.g., a survey of investmentintentions or a government budget esti-mate.

4. The judgment of the forecaster.

Now it can be argued that each of these po-

tential sources of the forecast is subject to adeterioration with the lengthening of the pre-dictive span.

This is clear for the first ingredient: for ex-ample, a prediction that the level of industrialproduction in April will be the same as in

March is likely to be less in error than the pre-diction that the level next September is goingto equal that of March. What is here illus-

5 The last figure refers to the period 1947-55 (the

strictly corresponding averages for the six- and twelve-month forecasts covering only those earlier years are 2.8

and 7.9 per cent, respectively).


11/17

9

trated on the simplest case is less obvious but

still basically true for more sophisticated ex-trapolative or autoregressive models.

Informed judgments and estimates willprobably also serve best over a #relatively shorttime range. The forecasting relations betweentime series involve lags of various lengths, but

typically from some point on the relationsweaken as the lags are increased. In the caseof the so-called leading indicators, whichtend to precede the turning points in generalbusiness activity, several factors combine to re-duce their effective forecasting lead. The earli-est presumed signals of reversal must usually be

confirmed by subsequent behavior of the sameand other series. The evidence of a group of in-dicators is more reliable than that of individualmembers. Brief erratic variations often obscure

movements of cyclical significance. Smoothinghelps to bring out the major movements andturns in the indicators, and to reduce the num-ber of false warnings, but it also cuts down thelength of the effective forecasting lead.

While the forecasts with short spans are gen-

erally superior to extrapolations, those withlonger spans (say of 12 to 18 months) are oftenworse than the more sophisticated types of ex-trapolation. For example, several of the recentforecasts of GNP came out poorer than theresults of autoregression methods or simple

trend estimates (such as projections of the aver-age historical change in the series). Signifi-cantly, the failures included forecasts of GNPfor the end of the next year, but not those cov-ering the year as a whole (including its earlierparts), nor shorter-span forecasts from the

same sources.

Forecasting and the Business Cycle

Cyclical movements are persistently under-valued by most forecasters. When forecast er-rors are averaged separately for different stagesof the business cycle, it turns out that the levels


12/17

10

of major aggregates (GNP, industrial produc-

tion) are understated most in the first year ofan expansion. Later in the expansion, whenthese aggregates usually grow slower, their lev-els are underestimated much less. (Occasional-ly, they are overestimated, as in the unexpectedretardation of 1962.) In contractions, forecasts

as a rule exceed the actual levels, either be-cause the downturn is missed or because thedecline turns out to be larger than predicted.Such errors can be observed in forecasts withdifferent spans, except for the longest ones (ex-ceeding one year) where the errors are very

large throughout and the differences within thecycle are statistically not significant.

Failures to recognize the turning point con-stitute another important category of cycli-cal forecasting errors. To appraise these er-rors, one must ask two questions: How oftendo turning points occur which have not beenpredicted? How often do predicted turns actu-ally occur?

In annual forecasts of aggregates which tendto grow most of the time, false signals ofturning points are understandably infrequent.Few reversals of direction will be here foreseenfrom one year to another, but rather increaseswill as a rule be expected. In forecasts thatrelate to shorter intervals and are issued morefrequently, however, the false warnings are

likely to be more troublesome. Given the spanand frequency of the forecasts, errors of thiskind will occur more often for variables withweaker trends but stronger cyclical and irregu-lar movements than for the smoothly growingseries.

Forecasters often failed to predict 50 percent or more of the turning points that didoccur. The proportion of these errors does notseem to depend systematically on the length ofthe predictive span. The hit-miss record of, say,fifty-fifty may appear worse than it actually is.

One reason for this is that the record includes


13/17

1 1

some errors that are largely excusable. This

clearly applies to the errors connected with theoutbreak of the Korean War and its early eco-nomic consequences, an event described earlieras an externally caused shock. Another largeconcentration of errors occurred in 194748,

when an early postwar depression was widely

anticipated. This was a grave misjudgment ofthe situation which cannot be exculpated justbecause it was very common, but one must alsoconsider that the disruption of economic re-lationships caused by the war made the earlypostwar forecasts particularly vulnerable.

Furthermore, it is important to recognizethat the ability to predict correctly at leastsome of the turning points, which the forecasts

reviewed demonstrably have, is an advantageover the extrapolations, which in general can-

not signalize turns at all. (The turns in extrap-olations will as a rule Zag those in the actualvalues; the strength of a good projection liesalmost entirely in that it may predict well the

longer-term trends.) An accuracy score of 40-60 per cent in predicting reversals, such as is

found for many short-term aggregative fore-casts, may be far from good but it is surelymuch better than zero. It is true that forecastsmay signal many false or extra turns, whichextrapolations could avoid. But this disad-vantage may be outweighed by the advantage

of the correct turning-point predictions, andthere is evidence that it frequently is.

It appears that the forecasters have on the

whole a better record in predicting upturnsthan in predicting downturns. In a contractionof the recent postwar variety, it should indeed

be reasonable to start watching out for an up-turn in, say, the third quarter of the move-ment, as pressures for an effective counterreces-sionary action will have mounted and forcesworking for a recovery will have gainedstrength by that time. An analogous argumentcould be made for an expansion, where one


14/17


15/17

1 3

duction, employment, and income-occur at

approximately the same time, but there aresome series that reach their turns earlier andothers that tend to lag. Some indicators wouldbe expected to lead others and do so: for ex-ample, new orders predict production of du-rable goods; housing starts predict residential

construction; investment commitments (appro-priations, contracts) predict plant and equip-ment expenditures. Since such indicators an-ticipate the movements of the coincidentaggregates, their tendency to turn ahead of the

general business recessions and revivals has a

strong logical as well as empirical foundation.Their early timing provides the forecasterwith an advantage for which there is no goodsubstitute. They can, together with other an-ticipatory data (e.g., surveys of investmentintentions) help rather efficiently at least toreduce the lag in recognizing the cyclical po-sition of the economy. Thus a business reces-

sion may be identified at about the time ofits occurrence or shortly thereafter, which isno mean achievement, considering the neces-

sity to compensate for the delays in the col-lection and processing of the data and thefact that historically such events had a demon-strably long recognition lag.

We know already that at times better resultsyet are achieved: a turning point is predicted

correctly ahead of the event. Evidence of earlyindicators and other anticipatory data mayhelp produce such forecasts over short spansof time. However, good judgment must prob-

ably be given a large share of the credit (aspoor judgment must be blamed on other oc-

casions for unduly long lags of recognition).The same factor (if not simply the forecastersluck) would certainly have to be credited forany successful turning-point predictions withlonger spans, where the indicators could nothave provided a sufficiently early signal. This

is so because it is very difficult to find series


16/17

1 4

which would have effective leads in relationto the early cyclical indicators themselves.

Observat ions

Far-reaching changes in the structure andfunctioning of the economy (involving bothits public and its private sector and their in-

teraction) have made business fluctuationsmilder in the postwar period than they usedto be in earlier times. The interest of academiceconomists in business cycles as a subject ofsystematic research has correspondingly weak-ened. However, despite important changes in

some of their aspects, business cycles continuedto be a recurring feature of the Americaneconomy.

It appears that economists have of late un-

derestimated the importance of cyclical proc-esses. An economic theorist may choose his

subjects or view them in such a way as to avoidthe troublesome business cycle problems, buta practical analyst and forecaster of businessconditions cannot very well do so. However,he too tends to underrate the cyclical move-ments of the economy, as attested by the types

of forecast errors I have discussed. There maybe a useful lesson in the fact that the recentbusiness fluctuations, even though mild, werenot quite as minor as many had apparentlyanticipated.

Forecasts of business conditions in the near

future are, in large part, considerably moreaccurate than several types of extrapolation,which is encouraging. It would be importantto know how much of the success of a fore-caster was due to his method or model andhow much to his pure judgment. In general,however, little can be said about this, sincelittle is known about the business forecastersassumptions and techniques. From some com-

parisons between business condition forecastsand more formal forecasts based on statistically

estimated economic relationships, it appears


17/17

1 5

that judgment often does make a net positivecontribution to the predictive performance.Yet it is also important to realize that it is theimprovement in that is necessary ifforecasting is to become more dependable; su-perior individual judgment is not replicableand not readily communicable to others. The

best examples of the forecasting art may wellsurpass the results of a scientific forecastingfrom explicit models-the patterns observed inthe past. But to the extent progress in this areais possible, it must nevertheless depend mainlyon the development of the scientific rather

than the artistic component of forecasting.

Business For Casting

Documents