c 2

Forecasting Fundamentals

Chapter Outline

2.1 Fundamental Principles of Forecasting

2.2 Major Categories of Forecasts
2.3 Forecast Errors
2.4 Computer Assistance

ntroduction - The starting point for virtually all planning systems is the ac-
tual or expected customer demand. In most cases, however, the time it
takes to produce and deliver the product or service will exceed the cus-
tomer expectation of delivery time. When that occurs, as is usually the case,
then production will have to begin before the actual demand from the cus-
tomer is known. That production will have to start from expected demand,
which is generally a forecast of the demand. This chapter discusses some of the
fundamental principles and approaches to forecasting for planning and con-
trol systems

2.1 fUNDAMENTAL PRiNCIPLES Of fORECASTING

First, we begin with a basic definition of forecasting:

Forecasting is a technique for using past experiences to project expectations

for the future.

Note that in this definition forecasting is not really a prediction, but a struc-
tured projection of past knowledge. There are several types of forecasts, used
for different purposes and systems. Some are long-range, aggregated models
used for long-range planning such as overall capacity needs, developing strate-
gic plans, and making long-term strategic purchasing decisions. Others are
short-range forecasts for particular product demand, used for scheduling and

11
18 CHAPTER 2 FORECASTING FUNDAMENTALS

launching production prior to actual customer order recognition. Regardless of

the purpose or system for which the forecast will be used, there are some fun-
damental characteristics that are very important to understand:
• Forecasts are almost always wrong. The issue is almost never about
whether a forecast is correct or not, but instead the focus should be on
"how wrong do we expect it to be" and on the issue of "how do we plan to
accommodate the potential error in the forecast." Much of the discussion
of buffer capacity and/or buffer stock the firm may use is directly related
to the size of the forecast error.
II Forecasts are more accurate for groups or families of items. It is usually
easier to develop a good forecast for a product line than it is for an indi-
vidual product, as individual product forecasting errors tend to cancel
each other out as they are aggregated. It is generally more accurate, for
example, to forecast the demand for all family sedans than to forecast the
demand for one particular model of sedan.
• Forecasts are more accurate for shorter time periods. In general, there
are fewer potential disruptions in the near future to impact product de-
mand. Demand for extended time periods far into the future are generally
less reliable.
II Every forecast should include an estimate of error. The first principle
indicated the importance to answer the question, "How wrong is the fore-
cast?" Therefore, an important number that should accompany the fore-
cast is an estimate of the forecast error. To be complete, a good forecast
has both the forecast estimate and the estimate of the error.
• Forecasts are no substitute for calculated demand If you have actual
demand data for a given time period, you should never make calculations
based on the forecast for that same time period. Always use the real data,
when available.

2.2 MAJOR CATEGORIES Of fORECASTS

There are two basic types of forecasting: qualitative and quantitative. Under
the quantitative types, there are two subcategories: time series and causal.
While this chapter provides basic descriptions of many of the common types
of forecasts in all the categories, the primary focus is discussing quantitative
time series forecasts.

Qualitative forecasting
Qualitative forecasting, as the name implies, are forecasts that are generated
from information that does not have a well-defined analytic structure. They
can be especially useful when no past data is available, such as when a product
is new and has no sales history. To be more specific, some of the key character-
istics of qualitative forecasting data include:
 CHAPTER 2 Forecasting Fundamentals 19

It The forecast is usually based on personal judgment or some external qual-

itative data.
I) The forecast tends to be subjective and, since they tend to be developed
from the experience of the people involved, will often be biased based on
the potentially optimistic or pessimistic position of those people.
I) An advantage is that this method often does allow for some fairly rapid
results.
It In some cases, qualitative forecasts are especially important as they may

be the only method available.

• These methods are usually used for individual products or product fami-
lies, seldom for entire markets.
Some of the more common methods of qualitative forecasting include
market surveys, Delphi or panel consensus, life-cycle analogies, and informed
judgment.
Market surveys are generally structured questionnaires submitted to po-
tential customers in the market. They solicit opinions about products or poten-
tial products, and also often attempt to get an understanding of the likelihood
of customer demand for products or services. If structured well, administered
to a good representative sample of the defined population, and analyzed cor-
rectly, they can be quite effective, especially for the short term. A major draw-
back is they are fairly expensive and time-consuming if done correctly.
Delphi or panel consensus forecasting uses panels of defined experts in
the market or area for which the survey is being developed. The experts at-
tempt to bring their individual knowledge of the factors that affect demand
into the analysis, interacting with each other to attempt to develop a consen-
sus as to the demand forecast for the products or product families in question.
The major difference between the two methods is one of process. While panel
forecasting tends to bring the experts together in a meeting format for the dis-
cussion, the Delphi method allows for a series of individual forecasts to be
made by each expert. Each expert develops their own forecast with their own
defined reasons. This collective set is then shared with all the experts, allowing
each to then modify their forecast based on information from the other ex-
perts. Through a series of these steps, the idea is to obtain a consensus as to the
forecast.
As one could probably imagine from the description of the process, these
methods tend to be quite expensive, primarily due to the time requirements
from a group of experts in the field. Such experts often charge fairly high fees
for their time and observations. The advantage is that when done correctly,
they do tend to be quite accurate.
Life cycle analogy forecasting is a rather special application used when the
product or service is new. The concept is fairly simple. It is based on the fact
that mQst products and services have a fairly well-defined life cycle. There is
generally growth during the early stage after introduction to the market. At
20 CHAPTER 2 FORECASTING FUNDAMENTALS

some point the product or service matures, implying little or no additional

growth, until eventually the demand declines to the point where it is no longer
offered. The major questions that arise with this life cycle include:
" What is the time frame? How long will growth and maturity last?
• How rapid will the growth be? How rapid will the decline be?
• How large will the overall demand be, especially during the mature
phase?
One potentially effective method to answer these questions will be to link
the demand for the new product or service to one in the past that is expected
to be similar. This will be especially effective if the new product or service is
essentially replacing another in the market, targeted to the same population.
In that case the method assumes that the life cycle for the new product or ser-
vice will essentially be roughly the same as for the old product or service it is
replacing.
This method may not be particularly accurate, but may be a good starting
point when no product demand history is available.
Informed judgment is among the most common forecasting methods used,
but unfortunately is also among the worst methods to use. One common ap-
proach used is for a sales manager to ask each salesperson to develop a pro-
jection of sales for their area for some defined time period in the future. The
sales manager then combines the individual sales projections into an overall
sales forecast for the company.
Why does this method tend to be so poor? There are several things that
will potentially alter the judgment of the individual salespeople, sometimes
without them being consciously aware. For example:
• Sometimes salespeople will use the forecast as an opportunity for an opti-
mistic goaL For example, if they feel they will really sell 5,000 of a certain
product over the time period, they may give the forecast as 6,000 as a goaL
This may also sometimes be prompted by their concern about the com-
pany planning to have adequate resources to produce enough product for
them to sell. If they really give a forecast of 5,000 and the company makes
only that amount, some salespeople feel they will be at a disadvantage if
the market potential is really larger than they thought. Most salespeople
want more than anything to have product available when a potential mar-
ket for that product exists.
• On the other hand, some salespeople are fearful that the sales forecast
will be used as a quota. If, for example, they feel they will really sell 5,000
in a given time period, they may give a much lower figure as a forecast. If,
for example, they give a forecast of 4,000 and they really sell 5,000, they
feel the sales above the forecast will make them look better as a salesper-
son. If they really do sell only 4,000, they can always say "See -:- I told you
so!"
 CHAPTER 2 Forecasting Fundamentals 21

• Many salespeople really want to give the right figure, but are subcon-
sciously impacted by recent events. If, for example, they had a very bad
week of sales just prior to submitting the forecast, they may be pessimistic
and lower the projection. The opposite can happen if they have a really
good week.

ANECDOTAL EXAMPLE 2.1

The following is based on an actual situation recently experienced by a produc-
tion control manager:
Joe, the production control manager, had just gotten the sales forecast for
all the major products for the next year from Frank, the sales manager (the com-
pany does not use Sales and Operation planning). As Joe was developing his ini-
tiallong-range production plans to meet the forecast, he noticed something that
bothered him. For the past several years the company had sold roughly 10,000
product X per year. Product X was sold to a small group of about six companies
that used it to make another product. The forecast for product X for next year
was given as 16,000. Joe called Frank and the following conversation took place:

JOE: "Frank, we need to talk about product X. You gave me a forecast of

16,000 for next year. Why is that?"
FRANK: "Because that's what we project to sell."
JOE: "Do you have any new customers for the product, or do you expect
to?"
FRANK: "No."
JOE: "Do any of your existing customers have new uses for the prod-
uct?"
FRANK: "Not that I know of."
JOE: "Do any of your existing customers have expansion plans or expect
to grow?"
FRANK: "Again, not that I know of."
JOE: "Do you or any of the customers for product X have any plans for
new markets?"
FRANK: "I know we don't, and I don't think any of the customers do."
JOE: "Then I don't understand. Why the forecast of 16,000?"
FRANK: "Because that's what we say we will sell!"
Now Joe is faced with a major problem. How many should he plan to make?
Product X uses some very expensive specialty steel that has a very long lead
time. He needs to place an order for that steel soon to meet the needs for next
year. He sees there are four scenarios, two of which are bad and two good:

.. He makes 16,000 and the demand is 16,000-that's good

.. He makes 16,000 and the demand is 10,000-that's bad, because a lot of ex-
pensive inventory is sitting around
• He makes 10,000 and the demand is 10,000- that's good
22 CHAPTER 2 FORECASTING FUNDAMENTALS

• He makes 10,000 and the demand is 16,000-that's bad, for a lot of obvious
reasons.
What does he do? Sometimes people who hear this story say he should
make 13,000 (hit the average), but that is likely bad for any of the four scenarios
above.
The correct answer is, of course, to plan to make 10,000. Why? Go back to
the conversation. What Joe was in fact doing is developing a forecast by asking
the questions that Frank should have been using to make the forecast in the first
place. Given the answers to the questions, Joe felt it was highly unlikely that the
demand would be greater than it had been in the past. A year later Joe was
proved correct, as the sales for product X roughly equaled 10,000 units.

Quantitative Forecasting-Causal
The first of the two types of quantitative forecasting methods we examine are
called causal. Some of the key characteristics of these methods include:
• This method is based on the concept of relationship between variables, or
the assumption that one measurable variable "causes" the other to change
in a predictable fashion.
• There is an important assumption of causality and that the causal variable
can be accurately measured. The measured variable that causes the other
to change is frequently called a "leading indicator." As an example, new
housing starts is often used as a leading indicator for developing forecasts
for many sectors of the economy.
• If there are good leading indicators developed, these methods often bring
excellent forecasting results.
• As somewhat of a side benefit, the process of developing the models will
often allow the developers of the model to gain additional significant mar-
ket knowledge. For example, if you are developing a causal model of vaca-
tion travel based on the leading indicator of gasoline prices, there is a
good chance you will gain knowledge about both the mechanisms that
control gasoline prices as well as the patterns of typical vacation travel.
• These methods are seldom used for product, but more commonly used for
entire markets or industries.
• The methods are often time-consuming and very expensive to develop,
primarily because of developing the relationships and obtaining the causal
data.
Some of the more common methods of causal forecasting are given as:
Input-output models. These can be very large and complex models, as they
examine the flow of goods and services throughout the entire economy. As
such, they require a substantial quantity of data, making them expensive and
time-consuming to develop. They are generally used to project needs for en-
tire markets or segments of the economy, and not for specific products.
 CHAPTER 2 Forecasting Fundamentals 23

Econometric models. These models involve a statistical analysis of various

sectors of the economy. Their use is similar to the input-output models.

Simulation models. Simulating sectors of the economy on computers are

growing in popularity and use with the development of ever more powerful
and less expensive computers and computer simulation models. They can be
used for individual products, but once again gathering the data tends to be ex-
pensive and time-consuming. The real value of these models is that they are
fast and economical to use once the data has "populated" the model.
Regression. A statistical method to develop a defined analytic relationship
between two or more variables. The assumption, as with other causal models, is
that one variable "causes" the other to move. Often the independent, or causal,
variable is called a leading indicator. A common example is when the news re-
ports on housing starts, since that is often a leading indicator of the amount of
economic activity in several related markets (e.g., the lumber industry).
Since they are based on external data, causal forecasting methods are
sometimes called extrinsic forecasts.

Quantitative forecasting-Time Series

Time-series forecasts are among the most commonly used for forecasting
packages linked to product demand forecasts. They all essentially have one
common assumption. That assumption is that past demand follows some pat-
tern, and that if that pattern can be analyzed it can be used to develop projec-
tions for future demand, assuming the pattern continues in roughly the same
manner. Ultimately that implies the assumption that the only real indepen-
dent variable in the time series forecast is time. Since they are based on inter-
nal data (sales), they are sometimes called intrinsic forecasts.
Time series are also the most commonly used by operations managers when
they find they need to forecast in order to make reasonable production plans.
The reason is simple: The other two major categories of forecasting (qualitative
and causal) both require some knowledge of the external market and/or envi-
ronment. Such knowledge is seldom easily available for an operations manager,
who typically spends most of his or her attention focused internally. Previous de-
mand is, however, often readily available for the operations manager.
Most time series forecasting models attempt to mathematically capture
the underlying patterns of past demand. One is a random pattern - under the
assumption that demand always has a random element. This implies what
most people inherently know: the customers who demand goods and services
from a company do not demand those goods and services in a completely uni-
form and predictable manner.
The second pattern is a trend pattern. The trends can either be increasing
or decreasing, and they can be either linear or nonlinear in nature. Some ex-
amples of trends are illustrated in Figure 2.2.
24 CHAPTER 2 FORECASTING FUNDAMENTALS

Demand

Random demand
pattern

Time

The third major pattern is a cyclical pattern, of which a special but very
common case is a seasonal pattern (see Figure 2.3). Even though called sea-
sonal (since for many companies the most common pattern of this type fol-
lows the seasons of the year), these patterns are actually cyclical patterns,
which mayor may not be linked to the yearly seasons. Cyclical patterns then
are demand patterns that follow some cycle of rising and falling demand. In
the special case where the pattern follows the seasons of the year, the cyclical
pattern is usually called seasonal.
If we were to put a random pattern together with a trend and a seasonal
pattern, we could obtain a demand pattern that would look similar to the pat-

Demand Demand
Linear increasing Linear decreasing
trend trend

' - - - - - - - - - - - - - - Time '----------------Time

Demand Demand

Nonlinear
decreasing trend Nonlinear
increasing trend

' - - - - - - - - - - - - - - Time '---------------Time

 CHAPTER 2 Forecasting Fundamentals 25

Demand
Sample seasonal
(cyclical) pattern

Time

tern experienced by many companies for their products or services. For exam-
ple, a random, seasonal pattern with a linear increasing trend might look
something like Figure 2.4.
Now that the basic patterns are developed, we can examine some of the
simpler time series methods that have been developed to forecast demand
knowing these patterns exist. The first set of forecasting methods includes sim-
ple methods that are used to attempt to smooth the random demand patterns,
assuming no trend or seasonal patterns exist. If no seasonal or trend patterns
exist in the demand, one might be tempted to use the actual demand from the
last period as the forecast for the next period. The problem with this approach
is the organization would continually be increasing or decreasing production
to accommodate the random pattern, and because of the randomness they
would seldom be correct. For that reason the smoothing methods attempt to,
as the name implies, smooth the ragged demand pattern.
There is an important trade-off in these approaches that needs to be real-
ized. If the smoothing approach is minimal (allowing most of the randomness
to remain), then there is little stability gained from the approach. On the other
hand, if too much smoothing is done, then real potential changes in the de-
mand are not captured in the forecast.

Demand

Time
26 CHAPTER 2 FORECASTING FUNDAMENTALS

Simple moving averages are, as the name implies, nothing more than the
mathematical average of the last several periods of actual demand. They take
the form:

F =A t-ll
+At-17+1 + ... +At-I
t n

Where: F is the forecast

t is the current time period, meaning F t is the forecast for the
current time period
At is the actual demand in period t, and
n is the number of periods being used.

The concept is much easier to see with an example. Suppose we are using a
three-period moving average. The forecast for any time period then becomes
the average of the actual demand for the three previous periods.
The calculations for the table are fairly easy. To get the forecast for period
4, take the actual demand for the three previous periods (periods 1 through 3)
and find the average: (24+26+22)/3 = 24. The forecast for period 5 comes from
the average of the demand for periods 2 through 4: (26+22+25)/3 = 24.3. The
process is called moving average because as time progresses you always move
to use the latest demand periods available. Graphically, the process looks as il-
lustrated in Figure 2.5.

Three-period moving
Period Demand average forecast
1 24
2 26
3 22
4 25 24.0
5 19 24.3
6 31 22.0
7 26 25.0
8 18 25.3
9 29 25.0
10 24 24.3
11 30 23.7
12 23 27.7
13 25.7
 CHAPTER 2 Forecasting Fundamentals 27

40 T--~'" -,,-,----'.--

30+---------~~--,.--~m_~
tn
'2::;) 20 +-----"---~-"'----"'w'_--------__! 3-Period moving
10+------------------------4 average

O+-~._._r_r_,_,_~_._._.~

Periods

There are two important points that need to be made concerning the
graph and the moving average method as well.
• First, it is fairly obvious to see that the forecast line is smoother than the
demand line, showing the impact of taking an average. The more periods
used in computing the moving average, the smoother this effect will be.
The reason is that with more periods being used in the average, anyone of
the demand points will have less overall influence.
• Second, the forecast will always lag behind any actual demand. That is not
so obvious in this graph, but suppose we use the same method to graph a
demand pattern with an upward trend, as in Table 2.2.
The graph in Figure 2.6, shows clearly how the forecast is constantly lag-
ging behind the trend in the data.
The implication of this lagging effect is that models such as simple moving
averages should normally not be used to forecast demand when the data

Three-period moving
Period Demand average forecast
1 13
2 15
3 18
4 22 15.3
5 27 18.3
6 31 22.3
7 36 26.7
8 41 31.3
9 45 36.0
10 52 40.7
11 57 46.0
12 51.3
28 CHAPTER 2 FORECASTING FUNDAMENTALS

Moving Average of a Trend

Period

clearly follows any type of trend or regular cyclical pattern. It is important to

note that forecasting methods should not be arbitrarily selected, but instead
should be selected and developed to fit the existing data as closely as possible.
Weighted moving averages are basically the same as simple moving aver-
ages, with one major exception. With weighted moving averages the weight as-
signed to each past demand point used in the calculation can vary. In this way
more influence can be given to some data points, typically the most recent de-
mand point. They take the basic form (the W stands for the weight):

n
~ = WjA t _ 1 + W2 At _ 2 + ... + ¥Y;,At - n where I W; = 1
i=l

In simpler terms, each of the weights is less than one, but the total of all
the weights must add to equal 1. Taking the same data points as in the first ex-
ample (the three-period moving average data point from Table 2.1), we will
apply a weighted moving average, with weights of 0.5, 0.3, and 0.2 (with the 0.5
weight applied to the most recent demand data) (see Table 2.3).
The calculations are again fairly simple. For example, the period 4 forecast is
calculated as 0.2(24) + 0.3(26) + 0.5(22) = 23.6. Notice this value is smaller than
the corresponding period 4 forecast using a simple moving average. The reason
is, of course, that a larger weight is being put on the latest demand figure, which
also happens to be the smallest of the three demand points being used.
Graphically, the data in the table appears in Figure 2.7.
As before, it is obvious that the forecast is smoothed, but also lags actual
demand changes.
Simple exponential smoothing is another method used to smooth the ran-
dom fluctuations in the demand pattern. There are two commonly used (math-
ematically equivalent) formulas:
 CHAPTER 2 Forecasting Fundamentals 29

Three-period weighted
Period Demand moving average forecast
1 24
2 26
3 22
4 25 23.6
5 19 24.3
6 31 21.4
7 26 26.2
8 18 26.1
9 29 23
10 24 25.1
11 30 24.3
12 23 28
13 25.3

The second form shows that the exponential smoothed forecast incorpo-
rated a weighted average of past history [(1 - ex)Ftl Since data from several
periods early is still contained in the forecast, and was weighted numerous
times as the forecast is developed period by period, one could consider it
weighted exponentially, thus the name. The first form, however, is easiest to
explain from the perspective of what the method does from a logical perspec-
tive. Essentially the forecast is found by taking the previous period's forecast
(Ft - i ) and adding a portion of the previous period's forecast error. The
forecast error is, of course, the difference between the actual demand for any
period and the forecast for that same period (A t - i - Ft - i ). The portion of the
error term is found by multiplication by ex, which is the Greek letter alpha, and
is called a smoothing constant. The alpha value is always between zero and

3-Period Weighted Average

40
30+-~~~~~~~~~~~~ _Demand
~ 20+I~~~~~~~~~~~~
:;:) ---{liI-- 3-Period Weighted
10+-~~~~~~~~~~~~ Ave.
O+-~~.-~~-,-,-.-.-.-+~

Periods
30 CHAPTER 2 FORECASTING FUNDAMENTALS

one, since if it equals zero you add no part of the error and your forecast is al-
ways the same number, while if equal to one you add the entire forecast error
and do no smoothing at alL As you might expect, the larger the alpha value,
the more of the forecast error is added. It makes the forecast more responsive
to actual changes in the demand, but also can equate to more reaction (and
disruption) in the organization as it constantly strives to react to a more er-
ratic forecast. The impact of the alpha value on the forecast can clearly be seen
by taking the same data set used earlier and finding exponentially smoothed
forecasts using alpha values first of 0.2, then 0.5, and finally O.S. The table uses
simple moving average for the first two periods to develop an initial forecast
of 25 units for period 3, after which exponential smoothing can be used to cal-
culate the remaining forecasts.
Notice that exponential smoothing assumes you have a forecast quantity
(Ft_I ). When you initially start developing the forecast, however, you do not
typically have such an initial forecast. This implies that you must start the
process using another forecasting method, after which the forecast from that
method can be used as the initial Ft - I .
The resulting graph showing the demand data and the forecast data is
shown in Figure 2.S.
As can be seen, with such a small alpha, there is very little change in the
forecast graph line. More responsiveness can be seen when alpha equals 0.5
(see Table 2.5).

Exponential smoothing
Period Demand with alpha =0.2
1 24
2 26
3 22
4 25 24.4
5 19 24.5
6 31 23.4
7 26 24.9
8 18 25.1
9 29 23.7
10 24 24.8
11 30 24.6
12 23 25.7
13 25.2
 CHAPTER 2 Forecasting Fundamentals 31

Exponential Smoothing with 0.2 Smoothing Constant

---+- Demand

-l!i- Exponential
Smoothing

Period

VSing .~::;:O.5withSal'rleDelT!andD~1:a
Exponential smoothing
Period Demand with alpha =0.5
1 24
2 26
3 22
4 25 23.5
5 19 24.3
6 31 21.6
7 26 26.3
8 18 26.2
9 29 22.1
10 24 25.5
11 30 24.8
12 23 27.4
13 25.2

The graph line for the forecast is obviously more responsive than it was for
an alpha of 0.2, but shows even more responsiveness when the alpha is
changed to 0.8, as in Table 2.6.

Exponential Smoothing with 0.5 Smoothing Constant

40
__ Demand
30 +---,-,--~""
.l!l
'2 20-t-1.:::..........:~5~r::.
-Iii- Exponential
~ 10+--~~-----_---~
Smoothing
O+-.-,,-.-.~~.-~-,-.~

Period
32 CHAPTER 2 FORECASTING FUNDAMENTALS

Exponential smoothing
Period Demand with alpha =0.8
1 24
2 26
3 22
4 25 22.6
5 19 24.5
6 31 20.1
7 26 28.8
8 18 26.6
9 29 19.7
10 24 27.1
11 30 24.6
12 23 28.9
13 24.2

Exponential Smoothing with 0.8 Smoothing Constant

40
__ Demand

.l!l
'2: 20-t-I~~~~~
__ Exponential
~ 10~~~~~~~~~~~
Smoothing
O~~~~~~~~=r~~ L -_ _ _ _ _ _ _ _~

Period

Regression has sometimes been called the "line of best fit." It is a statisti-
cal technique to try to fit a line from a set of points by using the smallest total
squared error between the actual points and the points on the line. A particu-
lar value for regression is to determine trend line equations. The best way to
show how it can be used is to illustrate with an example. In Table 2.7, we also
add a seasonal aspect to the data so that we can illustrate an approach to deal
with seasonal data using the same data set. We start with a data set that con-
tains 2 years of demand data, listed by quarters. Notice that quarters 1 and 5
represent the same season, as do quarters 2 and 6, and so forth.
Placing the demand history in Microsoft Excel (or any of the multiple sta-
tistical packages that can calculate regressions) and applying the regression
analysis, it was found the data had an intercept of 268.3 with an X-variable co-
efficient of 18.8. The general form of the regression equation is Y = aX + b,
where 'a' is the slope of the line and 'b' is the X-intercept. This means the re-
 CHAPTER 2 Forecasting Fundamentals 33

Quarter Demand
1 256
2 312
3 426
4 278
5 298
6 387
7 517
8 349

Quarter Demand Regression forecast

1 256 287.1
2 312 305.9
3 426 324.7
4 278 343.5
5 298 362.3
6 387 381.1
7 517 399.9
8 349 418.7
9 437.5

gression line has an equation of Y = 18.8(Quarter) + 268.3. Applying this for-

mula, we obtain Table 2.8, including the regression forecast.
As you might expect, a straight-line forecast calculated by using a linear
regression model does not show any of the seasonality of the data, as is clearly
shown in Figure 2.1l.
In order to pick up the seasonality for the forecast, we need to develop
seasonal mUltipliers for each quarter. To do this we first find the ratio of the
actual demand to the regression forecast in Table 2.9.
For example, in the first quarter the 0.89 comes from 256/287.1. Now an
average for corresponding quarters is calculated. This means that for the first
quarter of the year (as represented by quarter 1 and quarter 5) the seasonal
multiplier is (0.89 +0.82)/2, which equals 0.86. Other multipliers are listed in
Table 2.10.
Now the seasonal multipliers can be applied to the basic regression fore-
cast to develop a seasonally adjusted regression forecast by simply multiplying
the seasonal multipliers by the regression forecast, as in Table 2.11.
34 CHAPTER 2 FORECASTING FUNDAMENTALS

Basic Linear Regression

__ Demand
-lim- Regression

Period

Quarter Demand Regression forecast Ratio of demand to forecast

1 256 287.1 0.89
2 312 305.9 1.02
3 426 324.7 1.31
4 278 343.5 0.81
5 298 362.3 0.82
6 387 381.1 1.02
7 517 399.9 1.29
8 349 418.7 0.83
9 437.5

Ratio of demand Seasonal

Quarter Demand Regression forecast to forecast multipliers
1 256 287.1 0.89 0.86
2 312 305.9 1.02 1.02
3 426 324.7 1.31 1.30
4 278 343.5 0.81 0.82
5 298 362.3 0.82 0.86
6 387 381.1 1.02 1.02
7 517 399.9 1.29 1.30
8 349 418.7 0.83 0.82
9 437.5 0.86
 CHAPTER 2 Forecasting Fundamentals 35

Seasonal Seasonally adjusted

Quarter Demand Regression forecast multipliers regression forecast
1 256 287.1 0.86 246.1
2 312 305.9 1.02 311.3
3 426 324.7 1.30 422.9
4 278 343.5 0.82 282.2
5 298 362.3 0.86 310.5
6 387 381.1 1.02 387.8
7 517 399.9 1.30 520.8
8 349 418.7 0.82 343.9
9 437.5 0.86 376.3

Now if we look at the graphical comparison between the actual demand

and the seasonally adjusted regression forecast in Figure 2.12, it can easily be
seen how closely they compare. In addition, the forecast for period 9 will give
us a fair confidence, given how closely other quarters track (in fact, on this
graph it is very difficult to distinguish that there are in fact two separate lines).
To show how closely they track, Table 2.12 shows the percentage error be-
tween the seasonal forecast and the actual demand.
Before leaving the topic of regression, it may help to clear up any confu-
sion that may exist because regression was classified as a forecasting method-
ology for both causal forecasting and time series. There is a fundamental
difference, even though the mathematical computation of the regression lines
is the same. The difference is that the independent variable in time series re-
gression is always time, while the independent variable in causal regression is
always some other variable, usually a leading indicator from the economy.
It also should be noted that even though the discussion of seasonal in-
dexes was presented in the context of time series regression, the concept of
developing and applying seasonal indexes can be used for virtually any of the
time series models.

Seasonally Adjusted Regression

-+-Demand
~ 400
!: - - Seasonally
~ 200~~----~--~--~~~ Adjusted
Regression

Period
36 CHAPTER 2 FORECASTING FUNDAMENTALS

Quarter Demand Forecast Error Percentage error

1 256 246.1 9.9 4%
2 312 311.3 0.7 0%
3 426 422.9 3.1 1%
4 278 282.2 -4.2 -1%
5 298 310.5 -12.5 -4%
6 387 387.8 -0.8 0%
7 517 520.8 -3.8 -1%
8 349 343.9 5.1 1%

2.3 fORECAST ERRORS

Early in the chapter it was mentioned that every forecast should contain two
numbers: the forecast and the error estimate. Since the first rule of forecasting
is that the forecast is likely to be incorrect, a critical question is "How incor-
rect is it?" This is very important in planning and control since it represents a
critical issue to run the business. Buffer inventory, buffer capacity, or other
methods may be needed to be planned to accommodate actual demand that
differs from that forecasted.
There are several important error calculations used. Among some of the
most useful are included:
Mean Forecast Error (MFE). As the name implies, this term is calculated as
the mathematical average forecast error over a specified time period. The for-
mula is:
n

LJAI-~)
MFE = ..'-1=--"-1_ _ __
n

The (At - Ft) has been encountered earlier in the chapter. It represents the
difference between the forecast and the actual demand for any given time pe-
riod, also called the forecast error. The MFE involves adding all the individual
forecast errors and dividing by the total number of errors. This number is not
as important for the actual value of the number, but instead for the sign of the
number, whether it is positive or negative in value. If positive, it implies that
over the range of numbers included, the actual demand was larger than the
forecast. Another way of putting that is that the forecasting method was bi-
ased on the low side. If negative, of course, it means the forecasts were larger
than the demand on average, implying the forecasting method was biased on
the high side. For this reason, MFE is often referred to as forecast bias.
 CHAPTER 2 Forecasting Fundamentals 37

Period Demand (A) Forecast (F) Error (A-F)

1 12 14 -2
2 15 13 2
3 13 12 1
4 16 13 3
5 14 15 -1
6 11 14 -3

There is a very good reason the MFE does not really represent the aver-
age forecast error, as can be shown in Table 2.13.

MFE=(-2+2+1+3+ -1+ -3)/6=0/6=0

Adding all the errors yields a zero, making the MFE equal zero. It is clear,
however, that forecast errors do exist, so the MFE is not a good method to
find those errors. It does show, however, that in this case the forecasting
method was not biased, in that over the full range of the demand history the
forecasting method did not over or under project the total demand.
Mean Absolute Deviation (MAD). The formula is again given as the name
of the term. It literally means the average of the mathematical absolute devia-
tions of the forecast errors (deviations). The formula is, therefore:

iiAI-r:i
MAD = -,-1=-=.1_ __
n

This represents a very important number, as it tells the average forecast

error (always positive) over the time period in question. If we use the same
basic data from the table above, we can calculate the true forecast error in
Table 2.14.

Period Demand (A) Forecast (F) IA-FI

1 12 14 2
2 15 13 2
3 13 12 1
4 16 13 3
5 14 15 1
6 11 14 3
38 CHAPTER 2 FORECASTING FUNDAMENTALS

MAD = (2 + 2 + 1 + 3 + 1 + 3)/6 = 12/6 = 2

From these calculations, we now know that for these six periods used the
forecasting method was unbiased (MFE calculation) with an average forecast
error of two units (from the MAD calculation).
Tracking Signal. Similar to the concept of control limits for statistical
process control charts, the tracking signal provides a somewhat subjective
limit for the forecasting method to go "off track" before some action is taken.
It is calculated from the MFE and the MAD:

Tracking Signal = (n*MFE)/MAD

In some cases this formula is written using the "running sum of the forecast er-
rors," abbreviated as RSFE. The formula then becomes:

Tracking Signal = RSFE/MAD

This number is clearly a ratio that has no unit value-it is merely used as a
signaL A rule of thumb for use of the tracking signal is that if the value of the
tracking signal is larger than 4 or less than -4, the forecasting method may not
be effective for tracking demand over the time period in question. It merely
calls attention to investigate and adjust the forecasting method as necessary.
The tracking signal emphasizes an important trade-off: it would be time-
consuming and possibly costly to evaluate and modify the forecasting method
too frequently, but how often is too frequently? By the same token, to allow the
method to proceed too long without evaluation could produce serious deterio-
ration of the forecasts. The tracking signal, therefore, allows a systematic
method to determine when the forecasting method should be evaluated or not.

2.4 COMPUTER ASSISTANCE

The speed, reliability, and relatively low cost of today's computers makes the
use of very powerful computer packages utilizing time series formulas very at-
tractive. Some modern packages come with several time series formulas built
in with a variety of smoothing factors. Once demand data is input into the
package, the system will find the best approach based on the lowest MAD (or
some other error approach). The results from these packages can then become
direct inputs into other planning and control systems, where they can be a
great source to start the planning process.
These computer packages allow a fast and inexpensive approach to the
process that should be followed with or without the computer package. Specif-
ically, it is important to understand the use of the forecast, the past demand
patterns (when they exist), and the need to seriously attempt to find or de-
 CHAPTER 2 Forecasting Fundamentals 39

velop a forecasting method that will serve the purpose of the forecasting need
the best. Once developed, the method should be tested against past data and
modified as necessary.

SOLVED EXAMPLE
A demand pattern for 10 periods for a certain product was given as 127,113,
121,123,117,109,131,115,127, and 118. Forecast the demand for period 11
using each of the following methods: a 3-month moving average; a 3-month
weighted moving average using weights of 0.2, 0.3, and 0.5; exponential smooth-
ing with a smoothing constant of 0.3; and linear regression. Compute the MAD
for each method to determine which method would be preferable under the cir-
cumstances. Also calculate the bias in the data, if any, for all four methods, and
explain the meaning.
Solution: An Excel spreadsheet was set up to calculate the forecasts for each
method using the formulas and approaches outlined in the chapter. The follow-
ing chart shows the result from that analysis. Notice that because the overall
trend from the data was fairly "flat" and there appeared to be no seasonality or
other cyclicality, the coefficient for the period in the regression equation was
quite small (0.0182), making all the regression forecasts very close to each other.
The starting exponential smoothing forecast value of 115 was selected as the ac-
tual demand from the previous period (not shown on the table) that was 115
units.

Period Demand 3Mo.MA 3M.WMA Expon. smooth. Regression

1 127 115 120.0
2 113 118.6 120.0
3 121 116.9 120.1
4 123 120.3 119.8 118.1 120.1
5 117 119.0 120.4 119.6 120.1
6 109 120.3 119.6 118.8 120.1
7 131 116.3 114.2 115.9 120.1
8 115 119.0 121.6 120.4 120.1
9 127 118.3 118.6 118.8 120.2
10 118 124.3 124.2 121.3 120.2
11 120.0 120.1 120.3 120.2

The MADs for each of the forecasting methods were calculated using the
formula presented in the chapter. The result was as follows:
Method MAD
3-month moving average 7.1
3-month weighted moving average 7.9
Exponential smoothing 7.1
Linear regression 5.7
40 CHAPTER 2 FORECASTING FUNDAMENTALS

Given the data and information in the problem, regression should probably
be used because of the relatively small MAD compared to the other methods.
As more data is collected this could, of course, change.
Calculation of the MFE for each brought an interesting result. The first
three methods (moving average, weighted moving average, and exponential
smoothing) each brought a positive number (0.23,0.16, and 1.76, respectively).
The interpretation for those is that each of those three methods is biased, specif-
ically producing forecasts that are forecasting too low for the demand over the
range of data points given. That should not be too surprising, given that the re-
gression coefficient is slightly positive-an indication that there is a slight up-
ward slope to the data. Since it was discussed in the chapter how all three of
these methods tend to lag behind and trend in the data, it is logical that the fore-
casting method is a bit behind (biased low).
By contrast, the regression method picks up this slight upward trend-so
much so that the MFE equals zero, indicating the lack of bias in that method.

KEY TERMS

Forecasting Qualitative Forecast Quantitative Forecast

Market Surveys Delphi Method Panel Consensus
Life Cycle Analogy Informed Judgment Causal Forecasts
Input-Output Model Econometric Model Simulation
Regression Intrinsic Forecasts Extrinsic Forecasts
Random Pattern Trend Pattern Seasonality
Moving Averages Weighted Moving Average Exponential Smoothing
Forecast Error Smoothing Constant Mean Forecast Error
Tracking Signal Mean Absolute Deviation

SUMMARY

This chapter presents an overview of the tend to be primarily used for specific
some of the major characteristics of product demand, which is again useful
forecasting, and categorizes them into for the detailed product planning activi-
three major categories: qualitative, ties required of operations managers.
causal, and time series. Both qualitative A major characteristic of all fore-
and causal methods tend to require a casting methods is that they should be
great deal of information about external considered to be incorrect. The key to
markets and environments. Since much future planning methods is the issue of
of that information is not readily avail- just how incorrect they really are. For
able to the operations manager, the time this reason there should always be an
series methods (needing only past de- error estimate presented with the fore-
mand data) are appealing. Adding to cast. Some of the more common meth-
their appeal is the relative ease of calcu- ods for error calculation and use were
lation, especially with computers. They also discussed.
 CHAPTER 2 Forecasting Fundamentals 41

REFERENCES

Fogarty, D. w., J. H. Blackstone, Jr., and Willis, Raymond E., A Guide to Forecasting
T. R. Hoffmann. Production and Inven- for Planners and Managers. Englewood
tory Management. Cincinnati, OH: South- Cliffs, NJ: Prentice-Hall, 1987.
Western, 1991.

DISCUSSION QUESTIONS

1. Think of some of the leading indicators that could be used as a major input to
causal forecasts in the economy. Discuss their use.
2. Which type of forecasts would most likely be used for Sales and Operations Plan-
ning (S&OP), and why are they the most appropriate?
3. What value does it bring to an operation if a forecasting method is used that only
forecasts for families of products?
4. Think of at least three products recently introduced that would probably use the
life-cycle analogy. What products would they "copy"? Why is life-cycle appropri-
ate for those products?
5. How should a company include information for their forecast that indicates the
economy is headed for a recession? How, if at all, should that information impact
time-series forecasting information?
6. Discuss the arguments for using a large smoothing constant for exponential smooth-
ing instead of a small one. Under what conditions would each be better? Why?
7. Describe in your own words why using the MAD is better for describing the fore-
cast error than is the MFE. What is the major use of each? Should they really be
used together? Why or why not?

EXERCISES

1. Given the following data:

Period Demand
1 43
2 37
3 55
4 48
a. Calculate the three-period moving average for period 5.
b. Calculate the exponential smoothed forecast for period 5 using an alpha value
of 0.4. Assume the forecast for period 4 was the three-period moving average
of the first three periods.
c. Which method appeals the most for the data? Why?
2. Given the following demand data:
Period 1 2 3 4 5 6 7 8
Demand 17 22 18 27 14 18 20 25
a. Calculate the four-period weighted moving average for period 9 using weights
of 0.1, 0.2, 0.3, and 0.4 where the 0.4 is the weight for the most recent period.
42 CHAPTER 2 FORECASTING FUNDAMENTALS

b. Calculate the forecast for period 9 using a 3-month moving average forecasting
method.
c. Which method would you recommend using and why?
3. Given the same data for the previous problem:
Period 1 2 3 4 5 6 7 8
Demand 17 22 18 27 14 18 20 25
a. Use Excel or some other statistical computer package to calculate the regres-
sion equation for the data.
b. Use the regression equation to forecast the demand for period 9.
4. A forecasting method resulted in the following forecasts shown by the data in the
following table:

Period Demand Forecast

1 132 127
2 141 130
3 137 133
4 159 135
5 146 139
6 162 144
7 166 149
8 175 155
9 194 161
10 181 169
a. Use the data to calculate the MAD.
b. Find a regression equation for the demand data.
c. Use the regression equation to forecast demand for period II.
d. Is the regression method preferred over the method used? Why or why not?
5. The following demand data was collected over a 3 year period for one product:

Month Demand, year 1 Demand, year 2 Demand, year 3

1 72 84 97
2 67 98 119
3 85 86 138
4 99 113 124
5 87 121 143
6 135 140 162
7 127 133 157
8 131 156 178
9 102 125 136
10 96 134 141
11 88 118 122
12 79 102 120
 CHAPTER 2 Forecasting Fundamentals 43

Use the data to develop a regression-based forecast. Be sure to note that there is a
seasonal factor to the demand.
6. The following information is presented for a product:
1998 1999
Forecast Demand Forecast Demand
Quarter I 212 232 222 245
Quarter II 341 318 316 351
Quarter III 157 169 160 145
Quarter IV 263 214 251 242
a. What is the MAD for the data above?
b. Given the information above, what should the forecast be for the first quarter
of 2000 if the company switches to exponential smoothing with an alpha value
ofO.3?
7. The following information is presented for a product:
2001 2002
Forecast Demand Forecast Demand
Quarter I 200 226 210 218
Quarter II 320 310 315 333
Quarter III 145 153 140 122
Quarter IV 230 212 240 231
a. What are the seasonal indicies that should be used for each quarter?
b. What is the MAD for the data above?
8. Consider the forecast results shown below. Calculate MAD and MFE using the
data for months January through June. Does the forecast model under- or over-
forecast?
Month Actual Demand Forecast
January 1040 1055
February 990 1052
March 980 900
April 1060 1025
May 1080 1100
June 1000 1050