110
How to choose
the best techniqueor combination of
techniquesto help solve your particular
forecasting dilemma
Manager^s
guide to
forecasting
David M. Georgoff and
Robert G. Murdick
One thing may be more certain for managers today than anything else: they have almost too much to
think about and keep in mind in trying to assess problems
realistically and solve them. Some respond by developing
prejudices against any new idea because they don't have
enough time to learn the new concepts properly. Others
throw up their hands and admit they can't judge the idea
with everything else they are handling.
Looking at forecasting at a time when they
may need good forecasts more than ever, many managers
are downplaying their importance. One reason may be
that-like many other things-when forecasts are right,
you don't hear about them. But when they're wrong....
In response to this problem, the authors
have compiled a chart that profiles the 20 most common
forecasting techniques and rates their attributes against 16
important evaluative dimensions. The result is a practical
guide that will help executives sort out their priorities
when choosing a technique and enable them to combine
methods to achieve the best possible results.
Mr Georgoff is professor of marketing at
Florida Atlantic University and chairman of the Department of Management, Marketing, and International Business. He has published articles and worked as a consultant
to large corporations in the areas of new product marketing, marketing planning, market research, and forecasting.
Mr. Murdick is professor of management at
Florida Atlantic University. Previously he worked at the
General Electric Company for 14 years. Well known in the
field of management information systems, he is the author
or coauthor of 18 books on management and marketing,
the most recent of which is MIS: Concepts and Design
(Prentice-Hall, second edition, 1986).
Early in 1984, the Houston-based
COMPAQ Computer Corporation, manufacturer of
IBM-compatible microcomputers, faced a decision that
would profoundly affect its future. Recognizing that
IBM would soon introduce its version of the portable
computer and threaten COMPAQ'S dominance in this
profitable market, the company had two options. It
could elect to specialize in this product line and continue to market its highly regarded portables aggressively, or it could expand market offerings to include
desktop microcomputers. The latter move would force
the year-old company to confront IBM on its home
ground. Moreover, COMPAQ would have to make a
substantial investment in product development and
working capital and expand its organization and manufacturing capacity.
COMPAQ'S management faced several
important unknowns, including the potential market's
size, structure, and competitive intensity. Management
recognized that the company's vitality might seriously
erode if it did not expand its product line. If the expansion were successful, COMPAQ might enjoy economies of scale that could help ensure its survival in a
dynamic and very competitive industry. If COMPAQ'S
market assumptions were incorrect, however, its future might be bleak.
Many of today's managers face similar
new market realities and uncertainties. Continually
confronted with issues critical to their companies'
competitive future, they must deal with novel and rapidly changing environments. In short, they must judge
a broad range of dissimilar influences.
For more than a decade, new forecasting techniques have theoretically helped managers
evaluate these varied factors. Much of the promise of
Authors' note: We thank Steven C.
Wheelwright for his valuable assistance
in the preparation of this article.Guide to forecasting 111
these techniques has been unrealized, however, even
as a quickening succession of related advances have
been overwhelming decision makers with new alternatives.
As the number of techniques proliferates, management also realizes that some of its crucial
assumptions and projections about the economy have
become quite tenuous. Equipped only with a little history, meager and questionable data, and frail and changing theoretical tools, the forecaster must nevertheless
make critical decisions about altered futures.
As an example, COMPAQ Computer's
quandary was further complicated because new technologies, competitors, and products were already transforming a market that had been only recently established. COMPAQ'S forecast of the size, direction, and
price trends of the 1984 microcomputer market was
confounded by uncertainties about the market's response to several vital factors:
The entry of IBM's new portable
computer.
IBM's 23% price cut in June 1984 and its
potential erosion of margins.
The entry of lap portables introduced by
Hewlett-Packard and Data Ceneral.
The launch of IBM's new PC AT, complicated by unexpected delivery delays
and compatibility problems.
The introduction of desktop computers
by Sperry, NCR, ITT, and AT&T.
Eventually, COMPAQ entered the
desktop segment of the market, even though 1984 was
unforgiving and rampageous. Several large competitors
restricted their programs,- many smaller companies
went into-or to the edge of-receivership. Financially
and competitively, COMPAQ succeeded. During 1984,
sales rose from $111 million to $329 million and earnings increased from $4.7 million to $12.8 million.
The market's dynamics, however, make
such results increasingly difficult to achieve,- positive
and negative events-both expected and unforeseenhave a decisive effect. Even when managers anticipate
outcomes, grave uncertainties about timing, form, and
impact persist.
Despite the difficulty, the vice president
of marketing and the CEO-the two executives most
directly involved with the decision-demonstrated
what can be done. They used an extended series of consumer and dealer surveys coupled with periodic evaluations of the technology to assess the future market
and to guide the development of products and programs
to accommodate the industry's fluid and rapidly evolving needs. Managers can use forecasting techniques to
help them reach important decisions. A large and fastgrowing body of research deals with the development,
refinement, and evaluation of forecast techniques.
Managers also have greater access to both internal and
external data and can benefit from a multitude of computer software programs on the market, as well as
easier access to computer capabilities for analyzing
these data.
Forecaster's chart
while each technique has strengths and
weaknesses, every forecasting situation is limited by
constraints like time, funds, competencies, or data. Balancing the advantages and disadvantages of techniques
with regard to a situation's limitations and requirements is a formidable but important management task.
We have developed a chart to help executives decide which technique will be appropriate to
a particular situation; tbe chart groups and profiles a
diverse list of 20 common forecasting approaches and
arrays them against 16 important evaluative dimensions. We list techniques in columns and dimensions
of evaluation in rows. Individual row-column intersections (cells) reflect our view of a technique's characteristics as they apply to each dimension. Brief
descriptions of the forecasting methods are given on
the chart.
We have used different colors to show
which dimensions represent a strength for a particular
technique and which represent its weaknesses. The
strengths are highlighted in color; weaknesses are indicated by a gray cell. Naive extrapolation, for example,
is strong in internal consistency in that it easily reflects changes in management decisions. It is weak,
however, in forecast form. It is important to keep these
distinctions in mind when you are using the chart.
The chart is useful in two ways. The
first is in deciding which technique will suit your particular needs as a forecaster. The second is in deciding
how to combine techniques to further improve the result. In this section, we discuss the simpler approach;
we talk more about combining methods later.
To use the chart, look at the 16 questions listed in the first column after the dimensions.
They are the most common questions a manager will
ask when deciding to use a certain forecast. The first
question sets out the various time spans a forecast
would have to cover. Everyone who uses the chart will
have to answer question 1. But each of the following
questions can be answered with a yes or no. If you an-112 Harvard Business Review January-February 1986
swer no to a question, you don't have to look across
that row.
In responding to question 1, make note
of those techniques whose time span matches your
needs. We have found it easiest for forecasters to write
down the technique's column letter. The row number
of each dimension and the column letter of each technique are written along the horizontal and vertical axes.
With regard to question 1, for example, if your forecast
horizon is short-term, you can write down the cell letters for naive extrapolation (A), sales-force composite
(B), jury of executive opinion (C), and so forth. But
you would ignore the letters for scenario methods (D),
Delphi technique (E), historical analogy (F), and so on.
The columns you have now listed represent techniques that are qualified for further consideration. Next read down the column of each of these
techniques and note any gray cells. If these gray cells
are associated with questions to which you have answered yes, then the dimension either precludes use of
the technique or the technique can be used but it has
difficulty accommodating that dimension. Such precautions will help you determine whether you mustor wish to-eliminate certain techniques from further
consideration. An arrow in a cell indicates that its evaluation is the same as the cell to its left.
After you have answered all the questions and have a list of surviving techniques, note the
cells that are highlighted in color. Those cells represent
specific strengths of a technique and can guide you in
making a final selection.
In the course of the exercise, you may
have eliminated a technique that you like, have heard
about, or routinely use. You can go back to that one and
compare its strengths and weaknesses with those of
the methods that the chart has indicated would be best
for you. You can then decide whether you would rather
proceed with the technique that the chart indicates
corresponds most closely to your specific requirements
or whether you can accommodate the eliminating factors in order to use the technique that you initially
favored.
Important considerations
When considering each question, you
should remember some "tricks of the trade" con-,
ceming:
Time horizon. Most managers will
want the forecast results to extend as far into the future
as possible. Too long a period, however, may make the
technique selection process even more confusing
because of the varying abilities of the techniques to accommodate different time spans. In choosing an extended time horizon, the forecaster increases the complexity, cost, and time required to develop the final
product.
You can break down the time needed to
produce a forecast into development (Dev) and execution (Ex) time. Development time includes the gathering and entry of data, the modification of programs to
the company's specific requirements, and the start-up
of the system. Execution time is the time it takes to
produce a forecast with a particular technique. Initially, of course, development time is a significant concern for the forecaster; once the forecast technique is
firmly established, however, execution time is a more
appropriate concern.
Ibchnical sophistication. Experience
shows that computer and mathematical sophistication
is integral to many techniques. Although many executives have improved their skills in this area, not all
have sharpened their quantitative skills enough to be
comfortable with some of the forecast results a computer will spill out.
Cost. The cost of any technique is generally more important at the beginning when it is being developed and installed; after that, any technique's
potential value to a decision maker usually exceeds the
expense of generating an updated forecast.
Data availability. Before choosing a
technique, the forecaster must consider the extensiveness, currency, accuracy, and representativeness of the
available data. More data tend to improve accuracy, and
detailed data are more valuable than those presented in
the aggregate. Because a technique's ability to handle
fluctuations is important to a forecast's success, the
manager must match the sensitivity and stability of a
technique to the random and systematic variability
components of a data series.
Variability and consistency of data. Beyond changes that might occur in the company's structure or its environment, the manager must look at the
kind of stable relationships assumed among a model's
independent variables (represented by the "external
stability" dimension). For example, while most historically oriented quantitative forecasts might use expected levels of automobile production as a basis for
deterrnining demand for steel, the forecast model may
not reflect changes over time in the average amount of
steel used in automobiles. These relationships sometimes do change, but any variation is usually so gradual
that it will not aiffect a short-term forecast. When the
forecasts are long-term, however, or when the company
expects a substantial change in a vital relationship, the
forecaster should either apply judgment in a quantitative technique or use a qualitative method.Manager's guide to forecastingBrief descriptions of methods
Judgment methods Counting methods Time series methods
Naive extrapolation:
the application of a simple
assumption about the economic
outcome of the next time period,
or a simple, if subjective, extension of the results of current
events.
Sales-force composite:
a compilation of estimates by
salespeople (or dealers) of
expected sales in their territories, adjusted for presumed
biases and expected changes.
Jury of executive opinion:
the consensus of a group of
"experts," often from a variety
of functional areas within a
company
Scenario methods:
smoothly unfolding narratives
that describe an assumed future
expressed through a sequence
of time frames or snapshots.
Delphi technique:
a successive series of estimates
independently developed by a
group of "experts" each member of which, at each step in the
process, uses a summary of
the group's previous results to
formulate new estimates.
Historical analogy:
predictions based on elements
of past events that are analogous to the pre.sent situation.
Market testing:
representative buyers' responses to new offerings, tested
and extrapolated to estimate
the products' future prospects.
Consumer market survey:
attitudinal and purchase
intentions data gathered from
representative buyers.
Industrial market survey:
data similar to consumer surveys but fewer, more knowledgeable subjects sampled,
resulting in more informed
evaluations.
Moving averages:
recent values of the forecast
variables averaged to predict
future outcomes.
Exponential smoothing:
an estimate for the coming
period based on a constantly
weighted combination of
the forecast estimate for the
previous period and the most
recent outcome.
Adaptive filtering:
a derivation of a weighted
combination of actual and estimated outcomes, systematically
altered to reflect data pattern
changes.
Time series extrapolation:
a prediction of outcomes derived
from the future extension of a
least squares function fitted to a
data series that uses time as an
independent variable.Association or causal
methods
Time series decomposition:
a prediction of expected outcomes from trend, seasonal,
cyclical, and random components, which are isolated from a
data series.
Box-Jenkins:
a complex, computer-based
iterative procedure that
produces an autoregressive,
integrated moving average
model, adjusts for seasonal and
trend factors, estimates appropriate weighting parameters,
tests the model, and repeats the
cycle as appropriate.
Correlation methods:
predictions of vaiues based on
historic patterns of covariation
between variables
Regression models:
estimates produced from a predictive equation derived by
minimizing the residual variance
of one or more predictor (independent) variable.
Leading indicators:
forecasts generated from one or
more preceding variable that is
systematically related to the
variable to be predicted.
Econometric models:
outcomes forecast from an integrated system of simultaneous
equations that represent relationships among elements of the
national economy derived from
combining history and economic
theory.
Input-output models:
a matrix model that indicates
how demand changes in one industry can directly and cumulatively affect other industries.
Dev =
Ex =
Indicates
strength
Development
time
Execution
time
Indicates
weaknessDimensi
Time
Resource
requirements
Input
Output
Dns
Span
Urgency
Frequency
Mathematical
sophistication
Computer
Financial
Antecedent
Variability
Internal
consistency
External
consistency
External
stability
Detail
Accuracy
Capability for
reflecting
direction
changes
Capability for
detecting direction changes
Form
Questions
Is the forecast
period a:
Present need,
or Short-,
Medium-, or
Long-term
projection?
Is the forecast
needed
immediately?
Are frequent
forecast
updates
needed?
Are quantitative skills
limited?
Are computer
capabilities
limited?
Are only limited
financial
resources
available?
Are only limited
past data
available?
Does the
primary series
fluctuate
substantially?
Are significant
changes in
management
decisions
expected?
Are significant
environmental
changes
expected?
Are significant
shifts expected
among variable
relationships?
Are component
forecasts
required?
s a high level
of accuracy
critical?
Should turning
)oints be
reflected
)romptly?
Should turning
)oints be idenified early?
s an interval or
)robabilistrc
orecast
critical?
Judgment methods
Naive
extrapolation
Present need to
Medium
Rapid results
are a strong
advantage of
this technique.*
Dev Short
Ex Short
Can easily
accommodate
frequent
updates.
Minimal
quantitative
capabilities
are required.
Computer
capabilities are
not essential.
Very inexpensive to
implement and
maintain.
Some past data
are required,
but extended
history is not
essential.
Has difficulty
adequately
handling wide
fluctuations.
Can reflect
changes.
Can reflect
chanqes. but
quality can
also vary
substanlialiy.
Often insensitive to shifts.
Focus can be
readily
restricted.
Often provides
a limited practical level of
accuracy.
Can be very
responsive to
shifts.
Apt to miss
urning points.
Provides point
orecast with
crude estimated
ange.
Sales-force
composite
Short or
Medium
Forecast can be
assembled,
combined, and
adjusted relatively quickly.
Dev Short
Ex Moderate
Forecast
can be quickly
compiled, but
data collection
restricts rapidity.
Nominal
processing
does not require
a computer.
Inexpensive to
implement and
maintain.
Past data
are helpful but
not always
essential.
Significant
changes are
frequently not
transmitted and;
or realistically
reflected.
Generally has
difficulty realistically reflecting
changes.
Can often
)rovide useful
xeakdowns.
Can be very
accurafeorsubecttosubstaniai bias.
Can only
)rovide crude,
subjectively
determined
irobabilistic
orecast.
Jury of
executive
opinion
Short or
Medium
In-house group
forecasts are
quicker than
outside experts'.
Dev Short
Ex Short to
Moderate
Can accomplish
quickly.
Financial
requirements
are nominal for
executive
groups; they
may be higher
for outside
experts.
Does not handle
fluctuations well
but can accommodate them if
the panel meets
frequently.
If changes come
from an internal
corporate
group, technique can readly reflect them.
Reflects
changes well;
;echnique combines a range of
expertise.
Usually aware
of shifts and can
reflect them in
the forecast.
Can reflect
component
orecasts, but is
generally concerned with
aggregate
orecasts.
^ay be most
accurate
under dynamic
conditions.
Early turning
)oint identificaion can be a
strength under
dynamic
conditions.
Only subjecively determined
approximate
ange or frequency distribuion is possible.
Qualitative
Scenario
methods
Medium or
Long
Urgency
seriously compromises quality
Dev Moderate
to Long
Ex Moderate
Frequency need
is moderate;
updates are
generally provided as need
arises.
Usually expensive for thorough efforts.
^
Technique's
extended view
dampens
impact of shortrun influences
and random
variability.
Can readily
reflect internal
changes.
Reflects
changes well.
Adapts well to
shifts.
Generally confined to aggregate forecasts.
Mot particularly
accurate, but
usually most
accurate when
lorizons are
extended and
conditions are
dynamic.
Can readily
adjust if recognized, but long
ime horizon
often precludes
he need.
Can detect
cyclical turning
points early
under dynamic
conditions, but
ong time horizon often precludes the need.
Delphi
technique
Medium or
Long
Urgency
seriously compromises quality
Dev Moderate
Ex Moderate
to Long
Usually used for
one-time forecasts, but they
can be revised
as new informafion becomes
available.
Expense
depends on
makeup and
affiliation ot
participants.
Can accommodate changes,
but ease of
retlecting them
depends on
group's background.
—Pw
Technique is
subjective, but
distributions are
an inherent part
of technique.
Historical
analogy
Medium or
Long
Forecast can be
computed
quickly if data
are available;
data gathering
may cause delay.
Dev Moderate
Ex Moderate
Sophistication
level is variable,
but some quantitative skill is
desirable.
A computer may
be helpful.
If data are readily available,
out-of-pocket
costs are
minimal.
Extended history is essential.
Can crudely
reflect changes
at best.
Can handle
changes, but
forecast quality
can vary
substantially.
Can accommodate shifts
crudely.
nferential relationships are
often tenuous;
predictions are
suspect.
Can only predict
Toncyclical
points very
crudely.
n limited situaions, only an
approximate
range can be
urnished.
A B DCounting methods
Market testing
Medium
Substantial lag
Is involved.
Dev Moderate
Ex Long to
Extended
Extended, basically used for
one-time forecasts.
Technical competencies are
generally
needed.
A computer is
generally
needed for data
analysis.
Generally very
expensive.
Past data are
useful but not
essential.
Substantial fluctuations limit the
accuracy of
projections.
Can reflect
changes well if
they are incorporated into
original
research
design.
Seriously weak
in handling
changes.
Seriously weak
in accommodating shifts.
Handles detail
but scope can
be limited.
Provides highest accuracy in
new product
and limited data
conditions.
Responsive, but
this is not one of
its purposes.
Early turning
point identification is not a
purpose or
capability of this
technique.
Can provide
interval
estimates.
G
Market survey
Consumer
market survey
Medium
Method of gathering data may
cause a substantial time lag.
Dev Moderate
Ex Long to
Extended
Depending on
methodology,
frequent updates are possible, but updates
are generally
provided at extended intervais.
Generally
expensive for
good controls.
Handles fluctuations poorly,
but tracking
improves
performance.
Generally
cannot validly
reflect changes.
Ease of handling changes
depends on
consumers'
awareness and
interpretation.
Seldom reflects
significant
shifts.
Has limited predictability with
durables, somewhat better with
nondurables.
Often highly
responsive to
demand shifts.
Can be responsive to turning
points but
usually cannot
anticipate
them.
With probability
sampling,
accommodates
any desired
fornn.
H
Industrial
market survey
Medium or Long
Moderately
expensive,
depending on
controis.
Past data very
helpful but not
essential.
Wide fluctuations are
frequently
a significant
concern.
If changes are
recognized,
adjustments
can be made.
Reflects
changes
indirectly; it is
frequently very
sensitive to
them.
If carefully controlled, can handle shifts weii.
Can be most
accurate
approach in
special cases.
Can be very
sensitive to
turning points.
1
Time series
methods
Moving
averages
Short. Medium,
or Long
Rapid resuits
are a strong
advantage of
this technique.
Dev Short
Ex Short
Forecast can be
systematicaily
updated easily.
Minimal
quantitative
capabilities
are required.
A computer is
helpful for repetitive updating.
If data are
readily available, out-ofpocket costs
are minimal.
Past history is
essential.
Can accommodate fluctuations with
appropriate
averaging
period.
Cannot validly
reflect changes.
Cannot validly
reflect changes.
Cannot validly
reflect shifts.
Focus can be
readily
restricted.
Accurate under
stable
condition.s.
Variable lags
always exist.
Cannot
anticipate
turning points.
Confidence limits can be easily
derived based
on variability of
data series
J
Exponential
smoothing
Present need to
Short or
Medium
W
Only recent
forecasts and
current data are
required once
alpha is
determined.
Can accommodate fluctuations with
suitable alpha.
Can only moderately reflect
changes with
prior trend.
Can only moderately reflect
shifts with prior
trend.
Generally rates
high in accuracy
for short-term
forecasts.
Depending
on alpha value,
can be very
responsive.
Generally only
provides point
forecast.
K
Adaptive
filtering
Short or
Medium
Forecast can be
produced
quickly once
programmed
and past data
are available.
Dev Moderate
Ex Siiort
A fundamental
competency
level is required.
A computer is
essential.
Forecast is
moderately
expensive to
develop.
Past history is
essential
although detail
and extent vary.
Absorbs
random fiuctuations and
adjusts to systematic shifts.
Deals very well
with systematic
shifts in
variables.
L
Time series
extrapolation
Short, Medium,
or Long
Computation is
quick if data are
available; data
gathering can
cause delays.
Dev Short to
Moderate
Ex Short
W
A computer is
helpful for repetitive updating.
If data are readily available,
out-of-pocket
costs are
minimal.
Wide fluctuations result in
decreased
confidence in
projected
outcomes.
Cannot validly
reflect changes.
Cannot validly
reflect shifts.
Normally accurate for trends
and stafionary
series.
Very
unresponsive.
Probability
range is easily
constructed.
M
Time series
decomposition
Short or
Medium
Program setup
and data gathering
may cause delays,
but once programmed, computation is quick.
Dev Moderate
Ex Short
Moderately
expensive to
acquire,
develop, and
modify
Past history is
essential with
some detail
required.
Can isolate and
determine the
level of component effects.
Can only moderately reflect
changes with
prior trend.
Can only moderately reflect
shifts with prior
trend.
Effectively isolates idenfifiable
components.
Generally
responds
slowly
Generally
cannot predict
turning points
unless series
lags.
N
Box-Jenkins
Short. Medium,
or Long
Operationalizing
program can
take time, but
forecast can be
produced
quickly.
Dev Long
Ex Moderate
A high level of
understanding
is required.
A computer is
essential.
Acquisition and
modification
costs are
expensive.
Past history is
essential with
detail required.
Handles
variabiiity
effectively.
Frequently the
most accurate
for short-tomedium-range
forecasts.
When points are
identified,
adjusts quickly.
A weak predictive ability IS
possible
0Association or
Causai methods
Correlation
methods
Short, Medium,
or Long
Data evaluation
may cause
delays, but forecast computation is quick.
Dev Moderate
Ex Short to
Moderate
A fundamental
competency
level is required.
A computer is
desirable.
If data are on
hand, development costs are
moderate.
w
Technique is
good if covariation is high;
otherwise it is
poor.
Insensitive
to significant
changes unless
they are correlated with
predictor
variables.
Insensitive
unless they are
related to
predictor
variabies.
Predictive accuracy is weakened if shifts
occur.
Predictive
accuracy can
vary widely.
Can adapt
quickly to turning points.
Can predict
turning points
only if a iagged
relationship
exists.
Regression
models
Short, Medium,
or Long
Model formulation takes time,
but forecast computation is gulck.
Dev Moderate
to Long
Ex Shonto
Moderate
A computer is
essential for
most oases.
May handle
large fluctuations well with
appropriate
independent
variables.
Insensitive to
changes, but
they can be
reflected among
predictor
variables.
Handles
changes well
if they are
appropriately
reflecfed in
predictor
variabies.
A restricted
focus might
substantially
compromise
technique's
predictive
accuracy.
Can be accurate
if variable relationships are
stabie and the
proportion of
explained variance is high.
Sensitive to
changes once
they are
identified.
if relationships
are stabie, can
effectively
predict turning
points.
Confidence
limits are
provided.
Leading
indicators
Short, Medium.
or Long
Data evaluation
may cause
delays, but forecast computation is quick.
Dev Moderate
Ex Short to
Moderate
Extended
history is helpful in initial
development.
Can readily
adjust to systematic and random patterns.
Insensitive to
changes unless
they are
reflected in the
indicators.
Sensitive to
changes if they
are retiected in
appropriate
indicators.
Focus can be
readily
restricted,
depending on
indicators
used.
Oniy moderately accurate
under most
conditions.
Especially
effective in forecasting cyclical
changes.
Probability
range is easily
constructed.
Econometric
models
Short, Medium,
or Long
Model building
is lengthy, but
producing forecast is quick.
Dev Long to
Extended
Ex Short to
Moderate
Forecast
can be updated
quickly if data
are available.
A high level of
understanding
is required.
A computer is
essential for all
cases.
Development
costs are substantial; operating costs are
moderate.
Highly sensitive
to relevant
changes.
Genarally confined to aggregate forecasts.
Give spotty
performances in
dynamic
env.ronments.
Confidence
iimits are
provided.
Input-output
models
Medium or Long
Original model
may require up
to a year to
develop,
Dev Extended
Ex Short to
Moderate
Extended
detailed history
is required.
Time lag
further reduces
accuracy.
Insensitive to
changes.
Can be modified
to reflect
changes.
Cannot validly
reflect shifts
without updafed
coefficients.
Effectively
reflects demand
by SIC groups.
With stable
reiationships,
predictive accuracy can be
very good.
Cannot anticipate turning
points but can
effectively predict outcomes.
Confidence
limits can be
developed.
7 8
10
11
12
13
14
15
16
Q R TGuide to forecasting 119
Amount of detail necessary. While aggregate forecasts are easy to prepare, the manager will
need specific information (including individual product classes, time periods, geographic areas, or productmarket groupings, for example) to determine quotas or
allocate resources. Since forecasts vary widely in their
ahility to handle such detail, the manager may want a
technique that can accurately predict individual components and then comhine the results into an overall
picture. Otherwise, the forecaster can use one technique to provide an overall picture and then use past
patterns or market factors to determine the component forecasts.'
Accuracy. While accuracy is a forecaster's holy grail, the maximum accuracy one can expect
from a technique must fall within a range hounded hy
the average percentage error of the random component
of a data series. Also, hecause of self-defeating and selffulfilling prophecies, accuracy must he judged in light
of the control the company has over the predicted outcome and within the time and resource constraints imposed on the forecaster.
Rememher also that accuracy alone is
not the most important criterion. The forecaster may
wish to forgo some accuracy in favor of, for example, a
technique that signals turning points or provides good
supplemental information.
Ibming points. Because turning points
represent periods of exceptional opportunity or caution, the manager will want to analyze whether a technique anticipates fundamental shifts. Some techniques
give false turn signals, so the forecaster must keep in
mind not only a technique's ahility to anticipate
changes hut also its propensity to give erroneous information.
Form. Final form varies greatly; it is
always advisahle to use a technique that provides some
kind of mean or central value and a range of possihle
1 For additional discussion,
see G. David Hughes,
"Sales Forecasting Requirements,"
in The Handbook of Forecasting:
A Manager's Guide,
ed. Spyros Makridakis and
Steven C. Wheelwright
(New York:
John Wiley & Sons, 1982|, p.l3.
2 For a discussion of examples,
see Spyros Makridakis et al.,
"The Accuracy of Extrapolation
ITime Series) Methods,"
Journal of Forecasting,
April-June 1982,p. inland '
Steven P Schnaars,
"Situational Factors Affecting
Forecast Accuracy,"
fournal of Marketing Research,
August 1984, p. 290.
3 See Essam Mahmoud,
"Accuracy in Forecasting;
A Survey,"
journal of Forecasting,
April-June 1984, p. 139:
Spyros Makridakis and
Robert L. Winkler,
"Averages of Forecasts:
Some Empirical Results,"
Management Science,
September 1983, p. 987, and
Victor Zamowitz,
"The Accuracy of Individual and Group
Forecasts from Business Outlook Surveys,"
Journal of Forecasting,
January-March 1984, p. 10.
outcomes. If even remotely accurate, such information
helps the manager determine more explicitly risk exposure, expected outcomes, and likelihood distributions.
Improving the forecast
Because no dramatic breakthroughs in
technique development have occurred during the past
several years, efforts to improve forecasts have shifted
to searching for a better approach to technique selection. In part, these attempts have explored the strengths
and performance characteristics of various techniques.^
Our chart extends this approach by helping the forecaster match different tecliniques' strengths and characteristics to the needs and constraints of the required
forecast.
Managers can improve their
projection in the following ways:
Comhining forecasts.
Simulating a range of input assumptions.
Selectively applying judgment.
Combining forecasts
The research on combining forecasts to
achieve improvements (particularly in accuracy) is
extensive, persuasive, and consistent.-^ The results of
combined forecasts greatly surpass most individual
projections, techniques, and analyses by experts. Because top-rated experts and the most popular techniques cannot consistently outperform an approach
that combines results, and because the manager cannot predetermine which experts or techniques will be
superior in any situation, combining forecasts-particularly with techniques that are dissimilar-offers the
manager an assured way of improving quality.
The forecasting chart can help the
manager select the best combination of techniques.
As the chart shows, each method has strengths and
weaknesses. By carefully matching two or more complementary techniques, the forecaster can offset any
technique's limitations with the advantages of another,
all the while retaining the strengths of the first. Simply
compare an approach's highlighted cells against those
of other qualified methods. Various techniques incorporate very different underlying notions. Not knowing
which of these will ultimately prove to be most accurate in a particular economic environment, forecasters120 Harvard Business Review January-February 1986
can add to their awareness of possible outcomes by
evaluating the range and the distribution of the projections produced hy the various methods."
part article on scenario forecasts by Pierre Wack in the
September-October 1985 and November-December
1985 issues of HBR provides a good example of this. ^
Simulating various outcomes
The manager can also establish a range
of probable outcomes hy varying the combination and
the levels of inputs of a particular technique. Such sensitivity analysis can underscore the most critical variables, the range and distribution of expected outcomes,
and the probable outcomes from different assumptions.
Using judgment
while many quantitative forecasts incorporate some subjectivity, forecasters should rely
more heavily on the output of a quantitative forecast
than on their own judgment. Forecasting research has
concluded that even simple quantitative techniques
outperform the unstructured intuitive assessments of
experts and that using judgment to adjust the values of
a quantitatively derived forecast will reduce its accuracy^ This is so because intuitive predictions are susceptible to bias and managers are limited in their ability to process information and maintain consistent
relationships among variables.'
The forecaster should incorporate subjective judgments in dynamic situations when the
quantitative models do not reflect significant intemal
and extemal changes. Even in these cases, the forecaster should incorporate the subjective adjustments
as inputs in the model rather than adjusting the model's final outcome.
When confronted with extended horizons or with novel situations that have limited data
and no historical precedent, judgment or counting
methods should he used. Applying judgment in such
situations, however, should be done on a structured
basis. The forecaster should also employ judgment to
stimulate thought and explore new relationships but,
where possible, quantitative techniques should be incorporated to test and support assumptions. The twoForecasting strategies
4 See Hillel J. Einhom and
Robin M. Hogarth,
"Prediction, Diagnosis,
and Causal Thinking,"
lournal of Forecasting,
January-March 1982, p. 23.
5 For survey articles that
address this issue,
see Mahmoud, p. 139; and
Robin M. Hogarth and
spyros Makridakis,
"Forecasting and
Planning: An Evaluation,"
Management Science,
February 1981, p. 115.
6 Lennart Sjoberg,
"Aided and Unaided Decision Making:
Improved Intuitive Judgment,"
Journal of Forecasting,
October-December 1982, p. 349.
[There are] three basic strategies of forecasting....
The deterministic strategy assumes that the present has a close causal relation to the future. This is
the strategy that would be used by a cardsharp,
who had stacked the deck of cards, to predict the
deal. In economic forecasting, the strategy would
be used to predict construction expenditures by
a knowledge of construction contract awards
already made.
The symptomatic strategy assumes that present
signs show how the future is developing; such signs
do not "determine" the future but reveal the process of change that is already taking place. Thus, a
falling barometer may reveal a coming storm, or a
rising body thermometer an incipient illness. In economic forecasting, this strategy calls for the spotting
of "leading indicators"-time series whose movements foreshadow rises or declines in general business activity.
The systematic strategy assumes that, though
changes in the real world may seem accidental or
chaotic, careful analysis can reveal certain underlying regularities (sometimes called principles,
theories, or laws). The way to find these regularities is to black out much of reality and hold only to
the abstractions that make up a system, such as a
solar system, or a nuclear system, or an economic
system.
Though the theories that result from this process of
abstraction are "unreal," they may nevertheless
possess the power to affect the real world-provided, of course, that the theories are sound. The
test of the soundness of a theory is how it measures
up when applied to reality: An atomic explosion confirms Einstein's E = me'. Similarly, a price cut that
leads to increased sales confirms the hypothetical
demand curve that no man has ever seen outside
an economics textbook.
To be sure, economic "laws" do not have the consistency of those in the physical sciences. Nevertheless, economic relations or theories, derived from
a study of the past, may be useful tools for prediction, within some acceptable range of probable error.
From
BusinessForecaslIng:
With a Guide to Sources ot
Business Data
by Leonards. Silk and
M. Louise Curtey
(New York:
Random House, 1970), p. 3.
Copyright © 1970
by Random House, inc.
Reprinted with the permission
of the publisher.Harvard Business Review Notice of Use Restrictions, May 2009
Harvard Business Review and Harvard Business Publishing Newsletter content on EBSCOhost is licensed for
the private individual use of authorized EBSCOhost users. It is not intended for use as assigned course material
in academic institutions nor as corporate learning or training materials in businesses. Academic licensees may
not use this content in electronic reserves, electronic course packs, persistent linking from syllabi or by any
other means of incorporating the content into course resources. Business licensees may not host this content on
learning management systems or use persistent linking or other means to incorporate the content into learning
management systems. Harvard Business Publishing will be pleased to grant permission to make this content
available through such means. For rates and permission, contact [email protected].