AR R R RN RRARRRRRRRRRARARRRRRRRRRRRRRRRRDY

Reprinted from the 2005 APICS International Conference Proceedings.

Selecting the Appropriate Forecasting Method:

The Toolbox Approach
Charles W. Chase, Jr.

Today, most companies seem to use simple methods better understand the methods available along with their
that are easy to comprehend and mostly those that advantages and disadvantages.
involve judgment by company employees. On
the other hand, most forecast practitioners generally use
forecasting methods with which their decision makers feel
Available Methodologies
comfortable, even though these methods may not be the Most forecasting methods fall into two broad categories:
most effective ones. One method widely used results in qualitative (also known as “judgmental”) methods—those
goal setting rather than forecasting. Here, companies begin that rely on the subjective assessments of a person or
their planning process with a corporate goal to increase group of persons—and quantitative (also known as
sales by some percentage. This target often comes down “mathematical” or “objective”) methods—those that rely
directly from senior management as an edict. Everyone on past sales history alone or those that arc built ona
then proceeds to back into their targets based on what relationship between past sales and some other variable(s).
each business unit manager thinks they can deliver. If they Although one may have a firm grasp on these two
don’t meet their prospective targeted goal when totaled, categories of methods, it is important to realize that some
senior management either assigns individual targets to subjective assessment is usually involved in all types of
sales forecasts. Subjectively derived forecasts use intuitive
each business unit or puts a financial plug in place hoping
or gut feelings based on the experience and savvy of people
someone will over-deliver.
who understand not only what is presently occurring
in the marketplace but also what is likely to occur. The
Underlying Methodology Assumption most widely used judgmental techniques are independent
judgment, committee judgment, sales force estimates (also
The basic assumption underlying the application of known as “sales force composites”), and juries of executive
any forecasting method (statistical or judgmental) is opinion.
that the actual outcome observed will follow some
pattern associated with seasonality, trend, and/or causal Judgmental methods are often perceived as “last resort”
techniques (i.e., “We don’t have the hard data needed to use
relationships plus some random influences. This is
some mathematical technique, so we are forced to make a
algebraically written as: Actual Outcome = Pattern +
rough estimate.”). Often, judgmental methods provide very
Randomness. This simple equation is really saying that even
when the average patterns of the underlying data have been accurate forecasts. The major advantages of judgmental
methods are their low cost to develop (there’s no need for
identified, some deviation will exist between the forecast
and the actual. Our purpose as practitioners is to minimize
expensive computer hardware/software); executives usually
these deviations (errors) in the forecast by selecting the have a solid understanding of the broad-based factors and
how they affect sales demand; and sales forecasts can
appropriate method.
be developed fairly quickly. But, they are always biased
As forecast practitioners, we believe our primary toward the user group that develops them; they are not
responsibility is to provide senior management with only consistently accurate over time due to the subjective nature
accurate point estimates. As such, we tend to rely on the of development; some executives may not really understand
one-methodology -fits-all-situations forecasting mentality the firm’s sales situation since they are too far removed
when, in fact, we should consider ourselves business from the actual marketplace; and they are generally not
analysts who use an array of methods to predict the future well suited for firms with a large number of products (i.e.,
depending on each individual situation. Like a medical stockkeeping units [SKUs]).
doctor, we should carry a toolbox of generic methods and,
There are two broad segments within the quantitative
depending on the ailment, choose the technique (tool)
category: time series (which I refer to as reactive or one-
that best fits the situation. Unfortunately, before we can
dimensional methods) and causal (which I refer to as
apply the toolbox approach to forecasting, we need to
multidimensional or proactive methods). Time series are
(P22

Master Planning of Resources

RN
R R R R

DATA
Product Portfolio DATA
Product Portfolio
Stable Stable Demand
Pl
= Census X-11 = Census X-11
RR R R R

= Box-Jenkins m Multple Regression m Box-Jenkins = Mulple Regression

m Winters = Simple Regression m Winters w Simple Regression
Incomplete Complete Incomplete Complete

= Simple Moving = Robust = Simple Moving » Robust

Average Regression Average Regression
= Committees » Commitlees
ZA
windependent = Independent
Judgment — m Sales Force Judgment — m Sales Force
Composites
AR AR

Composites
Factory
o Unstable
Unsiable

Figure 1. Figure 2.

portfolio. Those products that have incomplete data and By reviewing the product portfolios, we found,
are unstable can only be forecasted using sales force on average, 10 percent of the products fell in the lower
composites, independent judgment, committee judgment, left-hand quadrant, 50 percent fell in the upper left-hand
and simple moving average, most of which are qualitative quadrant, 35 percent fell in the upper right-hand quadrant,
and five percent fell in the lower right-hand quadrant.
AR RN

methods or judgmental (see lower left-hand quadrant). On

the other hand, if you have very stable data but its historical Furthermore, of those 35 percent that fell into the upper
data are incomplete, then you are forced to use time series right-hand quadrant, 85 percent represented corporate unit
methods. Such methods would include Winter's, Box- and dollar volumes. See Figure 3 for details. This means
Jenkins, and Census X 11 (see upper left-hand quadrant). 1f 50 percent of your product portfolio can be forecasted
you have complete data, including causal information, and using advanced time series techniques that can be easily
the data are stable, indicating a distinct pattern, you can use automated with minimal resources. As such, resources can
simple regression, multiple regression, and/or econometric be utilized to provide more analysis and develop causal
techniques (see upper right-hand quadrant). Finally, if the
data are complete, including all causal variables, but the
data are unstable, displaying aberrations or outliers, you DATA
are forced to use robust regression (see lower right-hand Product Portfolio
AR R R

quadrant), which adjusts for outliers using a weighting Stable

approach. The approach is too complicated to explain in this
short space.
Now that we have established the dimensions by which = Census X-11

we can categorize what methods can be applied based on = BoxJenkins = Multple Regression
data availability and stability, we can now apply a third 50% mWinters mSimple Regression — 35%
dimension based on “business strategy.” See Figure 2 for
the details.
R

Incomplete Complete
For example, if the corporate business strategy changes
from a factory “push” strategy, where sales volumes are = Simple Moving mRobust
Average Regression
R R R

pushed through the channels of distribution by discounting 10%

products, to a demand “pull” strategy, where volumes are = Committees
pulled through the channels of distribution by growing = Independent
consumer demand, you are forced to use more sophisticated Judgment = Sales Force
techniques. In this situation, the company needs to model Composites
consumer behavior, which requires additional causal data,
Unstable
as well as longer historical sales volume.
AR

Figure 3.

Master Planning of Resources

AR
 |
models for 35 percent of the products falling in the upper
DATA
right-hand quadrant. All of these require more resources. Product Portfolio
In fact, 85 percent of the products in most product
portfolios can be forecasted using advanced time series and
causal methods. Unfortunately, the forecasting needs of
most companies fall in the lower left-hand quadrant. See
Figure 4 for details.

Conclusion
Whatever the method chosen, be it judgmental, time series,
or causal, all presume that the past can be drawn upon to
predict the future. Consequently, each class of methods uses
the past differently and possesses a different set of strengths
and weaknesses. Your products, goals, and constraints
should be considered when selecting the forecasting
method(s). There is always a tendency to use the one-
methodology-fits-all philosophy, because it's very appealing
from an implementation standpoint. However, we need to
realize that forecasting methods are really generic tools that
can be applied simultaneously across groups of products
based on the corporate product portfolio. This “toolbox”
approach for selecting forecast methods is much more
effective, not to mention more accurate. It provides us with
the framework to focus our resources on more sophisticated
techniques that capitalize on market opportunities, resulting
in increased customer value.

4 ©2010 APICS The Association for Operations Management

techniques built on the premise that future sales will mimic sales promotions, and merchandising. Therefore, once the
the pattern of past sales. In other words, time series methods nature of that association is quantified, it can be used to
rely on the identification of patterns (i.e., seasonality, trend, forecast sales. The most widely used causal methods are
and/or cyclical) within the past sales history of items being simple regression, multiple regression, and econometrics.
forecasted and assume those patterns will continue into In recent years—due to the development of more advanced
the future. The most basic time series method is called the software—a technique called robust regression has been
“naive” model, because it assumes future sales will replicate gaining wide acceptance by forecast practitioners. The
past sales. It is a little naive to assume that sales will be major advantages of causal methods are: They are available
exactly the same as the prior year given the dynamics of the in most software packages; they are inexpensive to run on
marketplace. Another time series method is called “moving computers; they are covered in most statistics courses, so
averaging.” Moving average techniques are also called they have become increasingly familiar to managers; they
“smoothing” models since they level out small, random provide accurate short- and medium-term forecasts; and
fluctuations. The most widely used time series methods they are capable of supporting “what-if” analyses. The
are called “exponential smoothing.” The basic premise of major disadvantages of causal models are: Their forecasting
exponential smoothing is that the sales volumes for the accuracy depends on a consistent relationship between
most recent periods have more impact on the forecast and independent and dependent variables; an accurate estimate
therefore should be given more weight. Among the array of the independent variable is crucial; many managers—due
of exponential smoothing techniques are Brown's double to a lack of understanding—view it as a “black box”
exponential smoothing, Holt's two-parameter exponential technique; they are more time-intensive to develop and
smoothing, and Winter's three parameter exponential require a strong understanding of statistics; they require
smoothing. larger data storage and are less easily systematized; and
A more advanced time series method is called they tend to be more expensive to build and maintain.
“decomposition.” This technique is based on the assumption
that sales are affected by four basic elements: trend, Selecting Method(s) Based on Product
seasonal influences, cyclical influences, and random
(irregular) influences. The most advanced time series Portfolio Analysis
technique is called Box-Jenkins (also known as auto A forecast practitioner developing a sales forecast can
regressive integrated moving average [ARIMA] models), choose from among all the methods discussed, but all these
which combines the key elements from both time series methods are not equally effective for any given situation.
and regression models. Here, autocorrelation coefficients The key factors to consider evolve around the completeness
identify the association between a variable at one time and stability of the data set(s) being forecast. Like a
period and with the same variable at some other time doctor, we need to assess each situation and prescribe the
period. Thus, autocorrelation is really the correlation of appropriate treatment. In this case, the practitioner should
a variable with itself. In spite of its forecasting success, use an appropriate method(s). Unfortunately, most systems
the Box-Jenkins approach is still the least used time series and most practitioners use one methodology to forecast all
method. This is because of its complexity that discourages their products, which results in poor performance, rather
many forecast practitioners and managers from using it. than apply the appropriate method(s) depending on the
The major advantages of time series methods are: situation.
They are well suited to situations where sales forecasts are 1f you consider your product portfolio as falling on
needed for a large number of products; they work very well two intersecting planes—one that is either incomplete or
for products with fairly stable sales; they can smooth out complete, and the other either unstable or stable—you can
small, random fluctuations; they are simple to understand determine which method to apply in almost any situation.
and use; they can be easily systematized and require little Incomplete data refer to having limited sales history for a
data storage; software packages for such methods are particular item, not having all the required causal variables,
readily available; and they are generally good at short- and/or not having any data, for that matter. Complete
term forecasting. The major disadvantages of time series refers to having all the required data for a particular item,
methods are: They require a large amount of historical as well as all the causal variables. Unstable refers to data
data; they adjust slowly to changes in sales; a great deal having aberrations or being random with no distinct pattern
of searching may be needed to find the weighted (alpha) associated with it. Stable refers to data that have a distinct
value; they usually fall apart when the forecast horizon is pattern associated with it, such as seasonality and trend.
long; and forecasts can result in great error because of large Understanding the strengths and limitations of the methods
fluctuations in current data. discussed, you can plot them in their corresponding
The basic premise of causal models is that future quadrants based on data availability within your product
sales of a particular product are closely associated with portfolio. See Figure 1 for the details.
changes in some other variable(s). For example, changes in With the help of this diagram, you can categorize
sales can be associated with changes in price, advertising, methods based on data availability within your product

2 02010 APICS The Association for Operations Management