CHAPTER 5

Forecasting

LEARNING OBJECTIVES
After completing this chapter, students will be able to:

1. Understand and know when to use various families 4. Understand Delphi and other qualitative decision-
of forecasting models. making approaches.
2. Compare moving averages, exponential smoothing, 5. Compute a variety of error measures.
and other time-series models.
3. Seasonally adjust data.

CHAPTER OUTLINE
5.1 Introduction 5.4 Measures of Forecast Accuracy
5.2 Types of Forecasts 5.5 Time-Series Forecasting Models
5.3 Scatter Diagrams and Time Series 5.6 Monitoring and Controlling Forecasts

Summary • Glossary • Key Equations • Solved Problems • Self-Test • Discussion Questions and Problems • Internet
Homework Problems • Case Study: Forecasting Attendance at SWU Football Games • Case Study: Forecasting Monthly
Sales • Internet Case Study • Bibliography
Appendix 5.1: Forecasting with QM for Windows

153
154 CHAPTER 5 • FORECASTING

5.1 Introduction
Every day, managers make decisions without knowing what will happen in the future. Inventory
is ordered though no one knows what sales will be, new equipment is purchased though no one
knows the demand for products, and investments are made though no one knows what profits
will be. Managers are always trying to reduce this uncertainty and to make better estimates of
what will happen in the future. Accomplishing this is the main purpose of forecasting.
There are many ways to forecast the future. In numerous firms (especially smaller ones),
the entire process is subjective, involving seat-of-the-pants methods, intuition, and years of ex-
perience. There are also many quantitative forecasting models, such as moving averages, expo-
nential smoothing, trend projections, and least squares regression analysis.
The following steps can help in the development of a forecasting system. While steps 5 and
6 may not be as relevant if a qualitative model is selected in step 4, data are certainly necessary
for the quantitative forecasting models presented in this chapter.

Eight Steps to Forecasting

1. Determine the use of the forecast—what objective are we trying to obtain?
2. Select the items or quantities that are to be forecasted.
3. Determine the time horizon of the forecast—is it 1 to 30 days (short term), 1 month to
1 year (medium term), or more than 1 year (long term)?
4. Select the forecasting model or models.
5. Gather the data or information needed to make the forecast.
6. Validate the forecasting model.
7. Make the forecast.
8. Implement the results.

These steps present a systematic way of initiating, designing, and implementing a forecast-
ing system. When the forecasting system is to be used to generate forecasts regularly over time,
data must be collected routinely, and the actual computations or procedures used to make the
forecast can be done automatically.
No single method is superior. There is seldom a single superior forecasting method. One organization may find regression
Whatever works best should be effective, another firm may use several approaches, and a third may combine both quantitative
used. and subjective techniques. Whatever tool works best for a firm is the one that should be used.

5.2 Types of Forecasts

The three categories of models In this chapter we consider forecasting models that can be classified into one of three categories:
are time series, causal, and time-series models, causal models, and qualitative models (see Figure 5.1).
qualitative.
Time-Series Models
Time-series models attempt to predict the future by using historical data. These models make
the assumption that what happens in the future is a function of what has happened in the past. In
other words, time-series models look at what has happened over a period of time and use a se-
ries of past data to make a forecast. Thus, if we are forecasting weekly sales for lawn mowers,
we use the past weekly sales for lawn mowers in making the forecast.
The time-series models we examine in this chapter are moving average, exponential smooth-
ing, trend projections, and decomposition. Regression analysis can be used in trend projections and
in one type of decomposition model. The primary emphasis of this chapter is time series forecasting.

Causal Models
Causal models incorporate the variables or factors that might influence the quantity being fore-
casted into the forecasting model. For example, daily sales of a cola drink might depend on the
season, the average temperature, the average humidity, whether it is a weekend or a weekday,
and so on. Thus, a causal model would attempt to include factors for temperature, humidity,
season, day of the week, and so on. Causal models may also include past sales data as time-
series models do, but they include other factors as well.
 5.2 TYPES OF FORECASTS 155

FIGURE 5.1
Forecasting
Forecasting Models Techniques

Qualitative Time-Series Causal

Models Methods Methods

Delphi Moving Regression

Method Averages Analysis

Jury of Executive Exponential Multiple

Opinion Smoothing Regression

Sales Force Trend

Composite Projections

Consumer
Decomposition
Market Survey

Our job as quantitative analysts is to develop the best statistical relationship between sales
or the variable being forecast and the set of independent variables. The most common quantita-
tive causal model is regression analysis, which was presented in Chapter 4. The examples in
Sections 4.8 and 4.9 illustrate how a regression model can be used in forecasting. Specifically,
they demonstrate how to predict the selling price of a house based on characteristics such as
size, age, and condition of the house. Other causal models do exist, and many of them are based
on regression analysis.

Qualitative Models
Whereas time-series and causal models rely on quantitative data, qualitative models attempt to
incorporate judgmental or subjective factors into the forecasting model. Opinions by experts, in-
dividual experiences and judgments, and other subjective factors may be considered. Qualitative
models are especially useful when subjective factors are expected to be very important or when
accurate quantitative data are difficult to obtain.
Here is a brief overview of four different qualitative forecasting techniques:
1. Delphi method. This iterative group process allows experts, who may be located in differ-
ent places, to make forecasts. There are three different types of participants in the Delphi
process: decision makers, staff personnel, and respondents. The decision making group
Overview of four qualitative or usually consists of 5 to 10 experts who will be making the actual forecast. The staff person-
judgmental approaches: Delphi, nel assist the decision makers by preparing, distributing, collecting, and summarizing a
jury of executive opinion, sales series of questionnaires and survey results. The respondents are a group of people whose
force composite, and consumer judgments are valued and are being sought. This group provides inputs to the decision
market survey. makers before the forecast is made.
In the Delphi method, when the results of the first questionnaire are obtained, the
results are summarized and the questionnaire is modified. Both the summary of the results
and the new questionnaire are then sent to the same respondents for a new round of
responses. The respondents, upon seeing the results from the first questionnaire, may view
things differently and may modify their original responses. This process is repeated with
the hope that a consensus is reached.
2. Jury of executive opinion. This method takes the opinions of a small group of high-level man-
agers, often in combination with statistical models, and results in a group estimate of demand.
3. Sales force composite. In this approach, each salesperson estimates what sales will be in
his or her region; these forecasts are reviewed to ensure that they are realistic and are then
combined at the district and national levels to reach an overall forecast.
4. Consumer market survey. This method solicits input from customers or potential customers
regarding their future purchasing plans. It can help not only in preparing a forecast but also
in improving product design and planning for new products.
156 CHAPTER 5 • FORECASTING

Hurricane Landfall Location Forecasts and

IN ACTION the Mean Absolute Deviation

S cientists at the National Hurricane Center (NHC) of the Na-

tional Weather Service have the very difficult job of predicting
recorded when a hurricane is 72 hours, 48 hours, 36 hours, 24
hours, and 12 hours away from actually reaching land. Once the
hurricane has come ashore, these forecasts are compared to the
where the eye of a hurricane will hit land. Accurate forecasts are actual landfall location, and the error (in miles) is recorded.
extremely important to coastal businesses and residents who At the end of the hurricane season, the errors for all the hurricanes
need to prepare for a storm or perhaps even evacuate. They are in that year are used to calculate the MAD for each type of fore-
also important to local government officials, law enforcement cast (12 hours away, 24 hour away, etc.). The graph below shows
agencies, and other emergency responders who will provide help how the landfall location forecast has improved since 1989. Dur-
once a storm has passed. Over the years, the NHC has tremen- ing the early 1990s, the landfall forecast when the hurricane was
dously improved the forecast accuracy (measured by the mean 48 hours away had an MAD close to 200 miles; in 2009, this num-
absolute deviation [MAD]) in predicting the actual landfall loca- ber was down to about 75 miles. Clearly, there has been vast
tion for hurricanes that originate in the Atlantic Ocean. improvement in forecast accuracy, and this trend is continuing.
The NHC provides forecasts and periodic updates of where the
hurricane eye will hit land. Such landfall location predictions are Source: Based on National Hurricane Center, http://www.nhc.noaa.gov.

5.3 Scatter Diagrams and Time Series

A scatter diagram helps obtain As with regression models, scatter diagrams are very helpful when forecasting time series. A scat-
ideas about a relationship. ter diagram for a time series may be plotted on a two-dimensional graph with the horizontal axis
representing the time period. The variable to be forecast (such as sales) is placed on the vertical axis.
Let us consider the example of a firm that needs to forecast sales for three different products.
Wacker Distributors notes that annual sales for three of its products—television sets,
radios, and compact disc players—over the past 10 years are as shown in Table 5.1. One sim-
ple way to examine these historical data, and perhaps to use them to establish a forecast, is to
draw a scatter diagram for each product (Figure 5.2). This picture, showing the relationship be-
tween sales of a product and time, is useful in spotting trends or cycles. An exact mathematical
model that describes the situation can then be developed if it appears reasonable to do so.
 5.3 SCATTER DIAGRAMS AND TIME SERIES 157

TABLE 5.1
YEAR TELEVISION SETS RADIOS COMPACT DISC PLAYERS
Annual Sales of
1 250 300 110
Three Products
2 250 310 100
3 250 320 120
4 250 330 140
5 250 340 170
6 250 350 150
7 250 360 160
8 250 370 190
9 250 380 200
10 250 390 190

FIGURE 5.2
Scatter Diagram for Sales (a)
Sales appear to be constant over time.
Annual Sales of Televisions

300
This horizontal line could be described by
250 the equation

200 Sales = 250

150 That is, no matter what year (1, 2, 3, and

so on) we insert into the equation, sales
100 will not change. A good estimate of future
sales (in year 11) is 250 televisions!
50

0 1 2 3 4 5 6 7 8 9 10
Time (Years)

(b)
420 Sales appear to be increasing at a
constant rate of 10 radios each year.
400 If the line is extended left to the vertical
Annual Sales of Radios

380 axis, we see that sales would be 290

in year 0. The equation
360
Sales = 290 + 10(Year )
340 best describes this relationship between
320 sales and time. A reasonable estimate
of radio sales in year 11 is 400,
300 in year 12, 410 radios.
280

0 1 2 3 4 5 6 7 8 9 10
Time (Years)

This trend line may not be perfectly

(c) accurate because of variation each
Annual Sales of CD Players

200 year. But CD sales do appear to

have been increasing over the past
180
10 years. If we had to forecast future
160 sales, we would probably pick a
larger figure each year.
140

120
100

0 1 2 3 4 5 6 7 8 9 10
Time (Years)
158 CHAPTER 5 • FORECASTING

5.4 Measures of Forecast Accuracy

We discuss several different forecasting models in this chapter. To see how well one model
works, or to compare that model with other models, the forecasted values are compared with the
actual or observed values. The forecast error (or deviation) is defined as follows:
Forecast error = Actual value - Forecast value
One measure of accuracy is the mean absolute deviation (MAD). This is computed by tak-
ing the sum of the absolute values of the individual forecast errors and dividing by the numbers
of errors (n):
g ƒ forecast error ƒ
MAD = (5-1)
n
Consider the Wacker Distributors sales of CD players shown in Table 5.1. Suppose that in the
past, Wacker had forecast sales for each year to be the sales that were actually achieved in the
The naïve forecast for the next previous year. This is sometimes called a naïve model. Table 5.2 gives these forecasts as well as
period is the actual value the absolute value of the errors. In forecasting for the next time period (year 11), the forecast
observed in the current period. would be 190. Notice that there is no error computed for year 1 since there was no forecast for
this year, and there is no error for year 11 since the actual value of this is not yet known. Thus,
the number of errors (n) is 9.
From this, we see the following:

a ƒ forecast error ƒ 160

MAD = = = 17.8
n 9
This means that on the average, each forecast missed the actual value by 17.8 units.
Other measures of the accuracy of historical errors in forecasting are sometimes used
besides the MAD. One of the most common is the mean squared error (MSE), which is the
average of the squared errors:*

a 1error2
2
MSE = (5-2)
n

TABLE 5.2
ACTUAL ABSOLUTE VALUE OF
Computing the Mean SALES OF CD FORECAST ERRORS (DEVIATION).
Absolute Deviation YEAR PLAYERS SALES |ACTUAL–FORECAST|
(MAD)
1 110 — —
2 100 110 |100  110|  10
3 120 100 |120  100|  20
4 140 120 |140  120|  20
5 170 140 |170  140|  30
6 150 170 |150  170|  20
7 160 150 |160  150|  10
8 190 160 |190  160|  30
9 200 190 |200  190|  10
10 190 200 |190  200|  10
11 — 190 —

Sum of |errors|  160

MAD  160/9  17.8

*In regression analysis, the MSE formula is usually adjusted to provide an unbiased estimator of the error variance.

Throughout this chapter, we will use the formula provided here.

 5.4 MEASURES OF FORECAST ACCURACY 159

Forecasting at Tupperware
MODELING IN THE REAL WORLD International

Defining
Defining the Problem
the Problem To drive production at each of Tupperware’s 15 plants in the United States, Latin America, Africa, Europe,
and Asia, the firm needs accurate forecasts of demand for its products.

Developing Developing a Model

a Model A variety of statistical models are used, including moving averages, exponential smoothing, and regression
analysis. Qualitative analysis is also employed in the process.

Acquiring
Acquiring Input Data
Input Data At world headquarters in Orlando, Florida, huge databases are maintained that map the sales of each
product, the test market results of each new product (since 20% of the firm’s sales come from products
less than 2 years old), and where each product falls in its own life cycle.

Developing
Developing a Solution
a Solution Each of Tupperware’s 50 profit centers worldwide develops computerized monthly, quarterly, and
12-month sales projections. These are aggregated by region and then globally.

Testing the
Testing the Solution
Solution Reviews of these forecasts take place in sales, marketing, finance, and production departments.

Analyzing
Analyzing the Results
the Results Participating managers analyze forecasts with Tupperware’s version of a “jury of executive opinion.”

Implementing
Implementing the Results
the Results Forecasts are used to schedule materials, equipment, and personnel at each plant.

Source: Interviews by the authors with Tupperware executives.

Besides the MAD and MSE, the mean absolute percent error (MAPE) is sometimes used.
The MAPE is the average of the absolute values of the errors expressed as percentages of the
actual values. This is computed as follows:

g` `
error
actual
MAPE = 100% (5-3)
n
Three common measures of error There is another common term associated with error in forecasting. Bias is the average er-
are MAD, MSE, and MAPE. ror and tells whether the forecast tends to be too high or too low and by how much. Thus, bias
Bias gives the average error and may be negative or positive. It is not a good measure of the actual size of the errors because the
may be positive or negative. negative errors can cancel out the positive errors.
160 CHAPTER 5 • FORECASTING

5.5 Time-Series Forecasting Models

A time series is based on a sequence of evenly spaced (weekly, monthly, quarterly, and so on)
data points. Examples include weekly sales of HP personal computers, quarterly earnings re-
ports of Microsoft Corporation, daily shipments of Eveready batteries, and annual U.S. con-
sumer price indices. Forecasting time-series data implies that future values are predicted only
from past values of that variable (such as we saw in Table 5.1) and that other variables, no mat-
ter how potentially valuable, are ignored.

Components of a Time Series

Analyzing time series means breaking down past data into components and then projecting them
Four components of a time series forward. A time series typically has four components:
are trend, seasonality, cycles, and
random variations. 1. Trend (T) is the gradual upward or downward movement of the data over time.
2. Seasonality (S) is a pattern of the demand fluctuation above or below the trend line that
repeats at regular intervals.
3. Cycles (C) are patterns in annual data that occur every several years. They are usually tied
into the business cycle.
4. Random variations (R) are “blips” in the data caused by chance and unusual situations;
they follow no discernible pattern.
Figure 5.3 shows a time series and its components.
There are two general forms of time-series models in statistics. The first is a multiplicative
model, which assumes that demand is the product of the four components. It is stated as follows:
Demand = T * S * C * R
An additive model adds the components together to provide an estimate. Multiple regres-
sion is often used to develop additive models. This additive relationship is stated as follows:
Demand = T + S + C + R
There are other models that may be a combination of these. For example, one of the components
(such as trend) might be additive while another (such as seasonality) could be multiplicative.
Understanding the components of a time series will help in selecting an appropriate fore-
casting technique to use. If all variations in a time series are due to random variations, with no
trend, seasonal, or cyclical component, some type of averaging or smoothing model would be
appropriate. The averaging techniques in this chapter are moving average, weighted moving
average, and exponential smoothing. These methods will smooth out the forecasts and not be

FIGURE 5.3
Product Demand Charted
over 4 Years, with Trend Trend
and Seasonality
Demand for Product or Service

Component
Indicated
Seasonal Peaks

Actual
Demand
Line
Average Demand
over 4 Years

Year Year Year Year

1 2 3 4
Time
 5.5 TIME-SERIES FORECASTING MODELS 161

too heavily influenced by random variations. However, if there is a trend or seasonal pattern
present in the data, then a technique which incorporates that particular component into the fore-
cast should be used. Two such techniques are exponential smoothing with trend and trend pro-
jections. If there is a seasonal pattern present in the data, then a seasonal index may be
developed and used with any of the averaging methods. If both trend and seasonal components
are present, then a method such as the decomposition method should be used.

Moving Averages
Moving averages smooth out Moving averages are useful if we can assume that market demands will stay fairly steady over
variations when forecasting time. For example, a four-month moving average is found simply by summing the demand dur-
demands are fairly steady. ing the past four months and dividing by 4. With each passing month, the most recent month’s
data are added to the sum of the previous three months’ data, and the earliest month is dropped.
This tends to smooth out short-term irregularities in the data series.
An n-period moving average forecast, which serves as an estimate of the next period’s
demand, is expressed as follows:
Sum of demands in previous n periods
Moving average forecast = (5-4)
n
Mathematically, this is written as
Yt + Yt - 1 + Á + Yt - n + 1
Ft + 1 = (5-5)
n
where
Ft + 1 = forecast for time period t + 1
Yt = actual value in time period t
n = number of periods to average
A 4-month moving average has n = 4; a 5-month moving average has n = 5.
WALLACE GARDEN SUPPLY EXAMPLE Storage shed sales at Wallace Garden Supply are shown
in the middle column of Table 5.3. A 3-month moving average is indicated on the right. The
forecast for the next January, using this technique, is 16. Were we simply asked to find a fore-
cast for next January, we would only have to make this one calculation. The other forecasts are
necessary only if we wish to compute the MAD or another measure of accuracy.
Weights can be used to put more WEIGHTED MOVING AVERAGE A simple moving average gives the same weight 11>n2 to each
emphasis on recent periods. of the past observations being used to develop the forecast. On the other hand, a weighted
moving average allows different weights to be assigned to the previous observations. As the

TABLE 5.3
MONTH ACTUAL SHED SALES 3-MONTH MOVING AVERAGE
Wallace Garden
Supply Shed Sales January 10
February 12
March 13
April 16 (10 + 12 + 13)/3 = 11.67
May 19 (12 + 13 + 16)/3 = 13.67
June 23 (13 + 16 + 19)/3 = 16.00
July 26 (16 + 19 + 23)/3 = 19.33
August 30 (19 + 23 + 26)/3 = 22.67
September 28 (23 + 26 + 30)/3 = 26.33
October 18 (26 + 30 + 28)/3 = 28.00
November 16 (30 + 28 + 18)/3 = 25.33
December 14 (28 + 18 + 16)/3 = 20.67
January — (18 + 16 + 14)/3 = 16.00
162 CHAPTER 5 • FORECASTING

weighted moving average method typically assigns greater weight to more recent observations,
this forecast is more responsive to changes in the pattern of the data that occur. However, this is
also a potential drawback to this method because the heavier weight would also respond just as
quickly to random fluctuations.
A weighted moving average may be expressed as

a 1Weight in period i21Actual value in period i2

Ft + 1 =
a 1Weights2
(5-6)

Mathematically, this is
w1Yt + w2Yt - 1 + Á + wnYt - n + 1
Ft + 1 = (5-7)
w1 + w2 + Á + wn
where
wi = weight for ith observation
Wallace Garden Supply decides to use a 3-month weighted moving average forecast with
weights of 3 for the most recent observation, 2 for the next observation, and 1 for the most dis-
tant observation. This would be implemented as follows:

WEIGHTS APPLIED PERIOD

3 Last month
2 2 months ago
1 3 months ago
3 Sales last month + 2 Sales 2 months ago  1 Sales 3 months ago

6
Sum of the weights

The results of the Wallace Garden Supply weighted average forecast are shown in Table 5.4. In
this particular forecasting situation, you can see that weighting the latest month more heavily
provides a much more accurate projection, and calculating the MAD for each of these would
verify this.
Choosing the weights obviously has an important impact on the forecasts. One way to
choose weights is to try various combinations of weights, calculate the MAD for each, and
select the set of weights that results in the lowest MAD. Some forecasting software has an op-
tion to search for the best set of weights, and forecasts using these weights are then provided.
The best set of weights can also be found by using nonlinear programming, as will be seen in a
later chapter.
Some software packages require that the weights add to 1, and this would simplify Equa-
tion 5-7 because the denominator would be 1. Forcing the weights to sum to 1 is easily achieved
by dividing each of the weights by the sum of the weights. In the Wallace Garden Supply exam-
ple in Table 5.4, the weights are 3, 2, and 1, which add to 6. These weights could be revised to
the new weights 3/6, 2/6, and 1/6, which add to 1. Using these weights gives the same forecasts
shown in Table 5.4.
Both simple and weighted moving averages are effective in smoothing out sudden fluctua-
tions in the demand pattern in order to provide stable estimates. Moving averages do, however,
Moving averages have two have two problems. First, increasing the size of n (the number of periods averaged) does smooth
problems: the larger number of out fluctuations better, but it makes the method less sensitive to real changes in the data should
periods may smooth out real they occur. Second, moving averages cannot pick up trends very well. Because they are aver-
changes, and they don’t pick up ages, they will always stay within past levels and will not predict a change to either a higher or a
trend. lower level.
USING EXCEL AND EXCEL QM IN FORECASTING Excel and spreadsheets in general are frequently
used in forecasting. Many forecasting techniques are supported by built-in Excel functions. You
can also use Excel QM’s forecasting module, which has several components. To access Excel
 5.5 TIME-SERIES FORECASTING MODELS 163

TABLE 5.4
MONTH ACTUAL SHED SALES 3-MONTH MOVING AVERAGE
Weighted Moving
Average Forecast January 10
for Wallace Garden February 12
Supply March 13
April 16 [(3 13)  (2 12)  (10)]>6  12.17
May 19 [(3 16)  (2 13)  (12)]>6  14.33
June 23 [(3 19)  (2 16)  (13)]>6  17.00
July 26 [(3 23)  (2 19)  (16)]>6  20.5
August 30 [(3 26)  (2 23)  (19)]>6  23.83
September 28 [(3 30)  (2 26)  (23)]>6  27.5
October 18 [(3 28)  (2 30)  (26)]>6  28.33
November 16 [(3 18)  (2 28)  (30)]>6  23.33
December 14 [(3 16)  (2 18)  (28)]>6  18.67
January — [(3 14)  (2 16)  (18)]>6  15.33

QM after it has been installed in Excel 2010 or Excel 2007 (see Appendix F for information
about installing Excel QM), go to the Add-Ins tab and select Excel QM and then select
Forecasting. If you click on a technique such as Moving Average, Weighted Moving Average, or
Exponential Smoothing, an input window will open. To use Excel QM for the Wallace Garden
Supply weighted moving average forecast, select Forecasting—Weighted Moving Average, as
shown in Program 5.1A. Enter the number of past periods of data and the number of periods to
be averaged, as shown in Program 5.1B. Click OK when finished, and a spreadsheet will be ini-
tialized. Simply enter the past observations and any parameters, such as the number of periods
to be averaged, and the output will automatically appear because the formulas are automatically
generated by Excel QM. Program 5.1C provides the results. To display the formulas in Excel,
simply press Ctrl + (grave accent). Pressing this again returns the display to the values instead
of the formulas.

PROGRAM 5.1A
Selecting the Forecasting
Module in Excel QM
From the Add-Ins tab, select Excel QM.

Put the cursor over Forecasting.

Click on the method that appears on the right.

164 CHAPTER 5 • FORECASTING

PROGRAM 5.1B
Input the title.
Initialization Screen for Input the number of past observations.
Weighted Moving
Average

You may select to see a graph of the data.

Input the number of periods to average.

Click OK.

PROGRAM 5.1C
Weighted Moving
Input the past observations.
Average in Excel QM for
Wallace Garden Supply
The names of the periods
Past forecasts, errors, and measures of accuracy are shown.
can be changed.

Input the weight. Note that

the highest weight is for the
most recent observation.

The forecast for the next period is here.

Exponential Smoothing
Exponential smoothing is a forecasting method that is easy to use and is handled efficiently by
computers. Although it is a type of moving average technique, it involves little record keeping
of past data. The basic exponential smoothing formula can be shown as follows:
New forecast = Last period’s forecast (5-8)
+ a1Last period’s actual demand - Last period’s forecast2
where  is a weight (or smoothing constant) that has a value between 0 and 1, inclusive.
 5.5 TIME-SERIES FORECASTING MODELS 165

Equation 5-8 can also be written mathematically as

Ft + 1 = Ft + 1Yt - Ft2 (5-9)
where
Ft + 1 = new forecast (for time period t + 1)
Ft = previous forecast (for time period t)
 = smoothing constant 10 …  … 12
Yt = previous period’s actual demand
The concept here is not complex. The latest estimate of demand is equal to the old estimate
adjusted by a fraction of the error (last period’s actual demand minus the old estimate).
The smoothing constant, ␣, The smoothing constant, , can be changed to give more weight to recent data when the
allows managers to assign weight value is high or more weight to past data when it is low. For example, when  = 0.5, it can be
to recent data. shown mathematically that the new forecast is based almost entirely on demand in the past three
periods. When  = 0.1, the forecast places little weight on any single period, even the most
recent, and it takes many periods (about 19) of historic values into account.*
For example, in January, a demand for 142 of a certain car model for February was pre-
dicted by a dealer. Actual February demand was 153 autos. Using a smoothing constant of
 = 0.20, we can forecast the March demand using the exponential smoothing model. Substi-
tuting into the formula, we obtain
New forecast 1for March demand2 = 142 + 0.21153 - 1422
= 144.2
Thus, the demand forecast for the cars in March is 144.
Suppose that actual demand for the cars in March was 136. A forecast for the demand in
April, using the exponential smoothing model with a constant of  = 0.20, can be made:
New forecast 1for April demand2 = 144.2 + 0.21136 - 144.22
= 142.6, or 143 autos
SELECTING THE SMOOTHING CONSTANT The exponential smoothing approach is easy to use
and has been applied successfully by banks, manufacturing companies, wholesalers, and
other organizations. The appropriate value of the smoothing constant, , however, can make
the difference between an accurate forecast and an inaccurate forecast. In picking a value for
the smoothing constant, the objective is to obtain the most accurate forecast. Several values
of the smoothing constant may be tried, and the one with the lowest MAD could be selected.
This is analogous to how weights are selected for a weighted moving average forecast. Some
forecasting software will automatically select the best smoothing constant. QM for Windows
will display the MAD that would be obtained with values of  ranging from 0 to 1 in incre-
ments of 0.01.
PORT OF BALTIMORE EXAMPLE Let us apply this concept with a trial-and-error testing of
two values of  in an example. The port of Baltimore has unloaded large quantities of grain
from ships during the past eight quarters. The port’s operations manager wants to test the use
of exponential smoothing to see how well the technique works in predicting tonnage unloaded.
He assumes that the forecast of grain unloaded in the first quarter was 175 tons. Two values
of  are examined:  = 0.10 and  = .50. Table 5.5 shows the detailed calculations for
 = 0.10 only.

*The term exponential smoothing is used because the weight of any one period’s demand in a forecast decreases expo-
nentially over time. See an advanced forecasting book for algebraic proof.
166 CHAPTER 5 • FORECASTING

TABLE 5.5
ACTUAL TONNAGE FORECAST FORECAST
Port of Baltimore QUARTER UNLOADED USING ␣ = 0.10 USING ␣ = 0.50
Exponential
1 180 175 175
Smoothing Forecasts
for ␣ = 0.10 2 168 175.5  175.00  0.10(180  175) 177.5
and ␣ = 0.50 3 159 174.75  175.50  0.10(168  175.50) 172.75
4 175 173.18  174.75  0.10(159  174.75) 165.88
5 190 173.36  173.18  0.10(175  173.18) 170.44
6 205 175.02  173.36  0.10(190  173.36) 180.22
7 180 178.02  175.02  0.10(205  175.02) 192.61
8 182 178.22  178.02  0.10(180  178.02) 186.30
9 ? 178.60  178.22  0.10(182  178.22) 184.15

To evaluate the accuracy of each smoothing constant, we can compute the absolute devia-
tions and MADs (see Table 5.6). Based on this analysis, a smoothing constant of  = 0.10 is
preferred to  = 0.50 because its MAD is smaller.
USING EXCEL QM FOR EXPONENTIAL SMOOTHING Program 5.2 illustrates how Excel QM han-
dles exponential smoothing with the port of Baltimore example.
EXPONENTIAL SMOOTHING WITH TREND ADJUSTMENT The averaging or smoothing forecast-
ing techniques are useful when a time series has only a random component, but these techniques
fail to respond to trends. If there is trend present in the data, a forecasting model that explicitly
incorporates this into the forecast should be used. One such technique is the exponential smooth-
ing with trend model. The idea is to develop an exponential smoothing forecast and then adjust
this for trend. Two smoothing constants,  and , are used in this model, and both of these val-
ues must be between 0 and 1. The level of the forecast is adjusted by multiplying the first
Two smoothing constants are smoothing constant, , by the most recent forecast error and adding it to the previous forecast.
used. The trend is adjusted by multiplying the second smoothing constant, , by the most recent error
or excess amount in the trend. A higher value gives more weight to recent observations and thus
responds more quickly to changes in the patterns.
As with simple exponential smoothing, the first time a forecast is developed, a previous
forecast 1Ft2 must be given or estimated. If none is available, often the initial forecast is as-

TABLE 5.6 ACTUAL ABSOLUTE ABSOLUTE

Absolute Deviations TONNAGE FORECAST DEVIATIONS FORECAST DEVIATIONS
and MADs for the QUARTER UNLOADED WITH ␣ = 0.10 FOR ␣ = 0.10 WITH ␣ = 0.50 FOR ␣ = 0.50
Port of Baltimore 1 180 175 5 175 5
Example
2 168 175.5 7.5 177.5 9.5
3 159 174.75 15.75 172.75 13.75
4 175 173.18 1.82 165.88 9.12
5 190 173.36 16.64 170.44 19.56
6 205 175.02 29.98 180.22 24.78
7 180 178.02 1.98 192.61 12.61
8 182 178.22 3.78 186.30 4.3
Sum of absolute deviations 82.45 98.63

g|deviation|
MAD = = 10.31 MAD = 12.33
n
 5.5 TIME-SERIES FORECASTING MODELS 167

PROGRAM 5.2
Port of Baltimore If initial forecast is given, enter it here. If you do not want
Exponential Smoothing to include the error for this initial forecast, cells E10:H10.
Example in Excel QM

Enter the data and alpha.

The forecast for quarter 9 is here.

sumed to be perfect. In addition, a previous trend 1Tt2 must be given or estimated. This is often
estimated using other past data, if available, or by using subjective means, or by calculating the
Estimate or assume initial values increase (or decrease) observed during the first few time periods of the data available. Without
for Ft and Tt such an estimate available, the trend is sometimes assumed to be 0 initially, although this may
lead to poor forecasts if the trend is large and  is small. Once these initial conditions have been
set, the exponential smoothing forecast including trend 1FITt2 is developed using three steps:

Step 1. Compute the smoothed forecast 1Ft + 12 for time period t + 1 using the equation

Smoothed forecast = Previous forecast including trend + 1Last error2

Ft + 1 = FITt + 1Yt - FITt2 (5-10)

Step 2. Update the trend 1Tt + 12 using the equation

Smoothed trend = Previous trend + 1Error or excess in trend2

Tt + 1 = Tt + 1Ft + 1 - FITt2 (5-11)

Step 3. Calculate the trend-adjusted exponential smoothing forecast 1FITt + 12 using the
equation
Forecast including trend 1FITt + 12 = Smoothed forecast 1Ft + 12 + Smoothed trend 1Tt + 12
FITt + 1 = Ft + 1 + Tt + 1 (5-12)
where
Tt = smoothed trend for time period t
Ft = smoothed forecast for time period t
FITt = forecast including trend for time period t
 = smoothing constant for forecasts
 = smoothing constant for trend
168 CHAPTER 5 • FORECASTING

TABLE 5.7
YEAR ELECTRICAL GENERATORS SOLD
Midwestern
2004 74
Manufacturing’s
Demand 2005 79
2006 80
2007 90
2008 105
2009 142
2010 122

Consider the case of Midwestern Manufacturing Company, which has a demand for electri-
cal generators over the period 2004 to 2010 as shown in Table 5.7. To use the trend-adjusted
exponential smoothing method, first set initial conditions (previous values for F and T) and
choose  and . Assuming that F1 is perfect and T1 is 0, and picking 0.3 and 0.4 for the smooth-
ing constants, we have
F1 = 74 T1 = 0  = 0.3  = 0.4
This results in
FIT1 = F1 + T1 = 74 + 0 = 74
Following the three steps to get the forecast for 2005 (time period 2), we have

Step 1. Compute Ft + 1 using the equation

Ft + 1 = FITt + 1Yt - FITt2
F2 = FIT1 + 0.31Y1 - FIT12 = 74 + 0.3174 - 742 = 74

Step 2. Update the trend 1Tt + 12 using the equation

Tt + 1 = Tt + 1Ft + 1 - FITt2
T2 = T1 + 0.41F2 - FIT12 = 0 + 0.4174 - 742 = 0

Step 3. Calculate the trend-adjusted exponential smoothing forecast 1FITt + 12 using the
equation
FIT2 = F2 + T2 = 74 + 0 = 74
For 2006 (time period 3) we have

Step 1.
F3 = FIT2 + 0.31Y2 - FIT22 = 74 + 0.3179 - 742 = 75.5

Step 2.
T3 = T2 + 0.41F3 - FIT22 = 0 + 0.4175.5 - 742 = 0.6

Step 3.
FIT3 = F3 + T3 = 75.5 + 0.6 = 76.1
The other results are shown in Table 5.8. The forecast for 2011 would be about 131.35.
 5.5 TIME-SERIES FORECASTING MODELS 169

Using Excel QM for Trend-Adjusted Exponential Smoothing

Program 5.3 shows how Excel QM can be used for the exponential smoothing with trend forecasts.

TABLE 5.8 Midwestern Manufacturing Exponential Smoothing with Trend Forecasts

TIME DEMAND
(t) 1Yt2 Ftⴙ1 ⴝ FITt ⴙ 0.31Yt ⴚ FITt2 Ttⴙ1 ⴝ Tt ⴙ 0.41Ftⴙ1 ⴚ FITt2 FITtⴙ1 ⴝ Ftⴙ1 ⴙ Ttⴙ1
1 74 74 0 74
2 79 74 = 74 + 0.3174 - 742 0 = 0 + 0.4174 - 742 74 = 74 + 0
3 80 75.5 = 74 + 0.3179 - 742 0.6 = 0 + 0.4175.5 - 742 76.1 = 75.5 + 0.6
4 90 77.270 1.068 78.338 = 77.270 + 1.068
= 76.1 + 0.3180 - 76.12 = 0.6 + 0.4177.27 - 76.12
5 105 81.837 2.468 84.305 = 81.837 + 2.468
= 78.338 + 0.3190 - 78.3382 = 1.068 + 0.4181.837 - 78.3382
6 142 90.514 4.952 95.466 = 90.514 + 4.952
= 84.305 + 0.31105 - 84.3052 = 2.468 + 0.4190.514 - 84.3052
7 122 109.426 10.536 119.962 = 109.426 + 10.536
= 95.466 + 0.31142 - 95.4662 = 4.952 + 0.41109.426 - 95.4662
8 120.573 10.780 131.353 = 120.573 + 10.780
= 119.962 + 0.31122 - 119.9622 = 10.536 + 0.41120.573 - 119.9622

PROGRAM 5.3
Midwestern Input values for the smoothing constants.
Manufacturing Trend-
Adjusted Exponential
Smoothing in Excel QM

You may enter initial values for F1 and T1.

Enter the past observations.

The forecast for the next year is given here.

Trend Projections
A trend line is a regression Another method for forecasting time series with trend is called trend projection. This technique
equation with time as the fits a trend line to a series of historical data points and then projects the line into the future for
independent variable. medium- to long-range forecasts. There are several mathematical trend equations that can be
170 CHAPTER 5 • FORECASTING

developed (e.g., exponential and quadratic), but in this section we look at linear (straight line)
trends only. A trend line is simply a linear regression equation in which the independent vari-
able (X) is the time period. The form of this is
YN = b0 + b1X
where
YN = predicted value
b0 = intercept
b1 = slope of the line
X = time period (i.e., X = 1, 2, 3, Á , n)
The least squares regression method may be applied to find the coefficients that minimize
the sum of the squared errors, thereby also minimizing the mean squared error (MSE). Chapter 4
provides detailed explanation of least squares regression, and formulas to calculate the coeffi-
cients by hand are in Section 4.3. In this section, we will rely on Excel and Excel QM to per-
form the calculations.
MIDWESTERN MANUFACTURING COMPANY EXAMPLE Let us consider the case of Midwestern
Manufacturing Company. That firm’s demand for electrical generators over the period
2004–2010 was shown in Table 5.7. A trend line to predict demand (Y ) based on the time pe-
riod can be developed using a regression model. If we let 2004 be time period 1 1X = 12, then
2005 is time period 2 1X = 22, and so forth. The regression line can be developed using Excel
2010 (see Chapter 4 for details) by going to the Data tab and selecting Data Analysis—Regression
and entering the information as shown in Program 5.4A. The results are shown in Program 5.4B.
From this we get
YN = 56.71 + 10.54X
To project demand in 2011, we first denote the year 2011 in our new coding system as
X = 8:
1sales in 20112 = 56.71 + 10.54182
= 141.03, or 141 generators
We can estimate demand for 2012 by inserting X = 9 in the same equation:
1sales in 20122 = 56.71 + 10.54192
= 151.57, or 152 generators

PROGRAM 5.4A
Excel Input Screen for
Midwestern
Manufacturing Trend
Line
 5.5 TIME-SERIES FORECASTING MODELS 171

PROGRAM 5.4B
Excel Output for
Midwestern
Manufacturing Trend
Line

The next year will be time period 8.

The slope of the trend line is 10.54.

A plot of historical demand and the trend line is provided in Figure 5.4. In this case, we may
wish to be cautious and try to understand the 2009–2010 swings in demand.
USING EXCEL QM IN TREND ANALYSIS Regression can also be performed in Excel QM. Go to
the Add-Ins tab in Excel 2010 and select Excel QM—Forecasting—Regression/Trend Analysis.
Enter the number of periods of data (7 in this example), enter a title and name for the time peri-
ods (e.g., week, month, year) if desired, and then click OK. When the initialized spreadsheet
appears, enter the past data and the time periods, as shown in Program 5.5.

Seasonal Variations
Time-series forecasting such as that in the example of Midwestern Manufacturing involves look-
ing at the trend of data over a series of time observations. Sometimes, however, recurring varia-
tions at certain seasons of the year make a seasonal adjustment in the trend line forecast

FIGURE 5.4
Electrical Generators and 160
the Computed Trend Line
150
140
Trend Line
130 Yˆ  56.71  10.54X
Generator Demand

120
110
100
90
80
Actual Demand Line
70
60
50

2004 2005 2006 2007 2008 2009 2010 2011 2012

Year
172 CHAPTER 5 • FORECASTING

PROGRAM 5.5
Excel QM Trend Past forecasts and errors are shown here.
Projection Model

Input the past data and the time periods.

The intercept (b0) and slope (b1) are found here.

To obtain a forecast for a future period, enter the time period here.

necessary. Demand for coal and fuel oil, for example, usually peaks during cold winter months.
Demand for golf clubs or suntan lotion may be highest in summer. Analyzing data in monthly or
quarterly terms usually makes it easy to spot seasonal patterns. A seasonal index is often used in
multiplicative time series forecasting models to make an adjustment in the forecast when a sea-
sonal component exists. An alternative is to use an additive model such as a regression model
that will be introduced in a later section.
A seasonal index indicates how a particular season (e.g., month or quarter) compares with
an average season. When no trend is present, the index can be found by dividing the average
An average season has an index value for a particular season by the average of all the data. Thus, an index of 1 means the season
of 1. is average. For example, if the average sales in January were 120 and the average sales in all
months were 200, the seasonal index for January would be 120>200 = 0.60, so January is be-
low average. The next example illustrates how to compute seasonal indices from historical data
and to use these in forecasting future values.
Monthly sales of one brand of telephone answering machine at Eichler Supplies are shown
in Table 5.9, for the two most recent years. The average demand in each month is computed, and
these values are divided by the overall average (94) to find the seasonal index for each month.
We then use the seasonal indices from Table 5.9 to adjust future forecasts. For example, suppose
we expected the third year’s annual demand for answering machines to be 1,200 units, which is
100 per month. We would not forecast each month to have a demand of 100, but we would ad-
just these based on the seasonal indices as follows:
1,200 1,200
Jan. * 0.957 = 96 July * 1.117 = 112
12 12
1,200 1,200
Feb. * 0.851 = 85 Aug. * 1.064 = 106
12 12
1,200 1,200
Mar. * 0.904 = 90 Sept. * 0.957 = 96
12 12
1,200 1,200
Apr. * 1.064 = 106 Oct. * 0.851 = 85
12 12
1,200 1,200
May * 1.309 = 131 Nov. * 0.851 = 85
12 12
1,200 1,200
June * 1.223 = 122 Dec. * 0.851 = 85
12 12
 5.5 TIME-SERIES FORECASTING MODELS 173

TABLE 5.9 AVERAGE

Answering Machine SALES DEMAND
AVERAGE 2- MONTHLY SEASONAL
Sales and Seasonal MONTH YEAR 1 YEAR 2 YEAR DEMAND DEMANDa INDEXb
Indices
January 80 100 90 94 0.957
February 85 75 80 94 0.851
March 80 90 85 94 0.904
April 110 90 100 94 1.064
May 115 131 123 94 1.309
June 120 110 115 94 1.223
July 100 110 105 94 1.117
August 110 90 100 94 1.064
September 85 95 90 94 0.957
October 75 85 80 94 0.851
November 85 75 80 94 0.851
December 80 80 80 94 0.851
Total average demand = 1,128

1,128 Average 2 year demand

aAverage monthly demand = = 94 bSeasonal index =
12 months Average monthly demand

Seasonal Variations with Trend

When both trend and seasonal components are present in a time series, a change from one month
to the next could be due to a trend, to a seasonal variation, or simply to random fluctuations. To
help with this problem, the seasonal indices should be computed using a centered moving
Centered moving averages are average (CMA) approach whenever trend is present. Using this approach prevents a variation
used to compute seasonal indices due to trend from being incorrectly interpreted as a variation due to the season. Consider the fol-
when there is trend. lowing example.
Quarterly sales figures for Turner Industries are shown in Table 5.10. Notice that there is a
definite trend as the total each year is increasing, and there is an increase for each quarter from
one year to the next as well. The seasonal component is obvious as there is a definite drop from
the fourth quarter of one year to the first quarter of the next. A similar pattern is observed in
comparing the third quarters to the fourth quarters immediately following.
If a seasonal index for quarter 1 were computed using the overall average, the index would
be too low and misleading, since this quarter has less trend than any of the others in the sample.
If the first quarter of year 1 were omitted and replaced by the first quarter of year 4 (if it were
available), the average for quarter 1 (and consequently the seasonal index for quarter 1) would
be considerably higher. To derive an accurate seasonal index, we should use a CMA.
Consider quarter 3 of year 1 for the Turner Industries example. The actual sales in that quar-
ter were 150. To determine the magnitude of the seasonal variation, we should compare this with
an average quarter centered at that time period. Thus, we should have a total of four quarters (1
year of data) with an equal number of quarters before and after quarter 3 so the trend is averaged
out. Thus, we need 1.5 quarters before quarter 3 and 1.5 quarters after it. To obtain the CMA,

TABLE 5.10
QUARTER YEAR 1 YEAR 2 YEAR 3 AVERAGE
Quarterly Sales
($1,000,000s) for 1 108 116 123 115.67
Turner Industries 2 125 134 142 133.67
3 150 159 168 159.00
4 141 152 165 152.67
Average 131.00 140.25 149.50 140.25
174 CHAPTER 5 • FORECASTING

TABLE 5.11
YEAR QUARTER SALES ($1,000,000s) CMA SEASONAL RATIO
Centered Moving
Averages and 1 1 108
Seasonal Ratios for 2 125
Turner Industries 3 150 132.000 1.136
4 141 134.125 1.051
2 1 116 136.375 0.851
2 134 138.875 0.965
3 159 141.125 1.127
4 152 143.000 1.063
3 1 123 145.125 0.848
2 142 147.875 0.960
3 168
4 165

we take quarters 2, 3, and 4 of year 1, plus one-half of quarter 1 for year 1 and one-half of quar-
ter 1 for year 2. The average will be
0.511082 + 125 + 150 + 141 + 0.511162
CMA 1quarter 3 of year 12 = = 132.00
4
We compare the actual sales in this quarter to the CMA and we have the following seasonal
ratio:
Sales in quarter 3 150
Seasonal ratio = = = 1.136
CMA 132.00
Thus, sales in quarter 3 of year 1 are about 13.6% higher than an average quarter at this time.
All of the CMAs and the seasonal ratios are shown in Table 5.11.
Since there are two seasonal ratios for each quarter, we average these to get the seasonal
index. Thus,
Index for quarter 1 = I1 = 10.851 + 0.8482>2 = 0.85
Index for quarter 2 = I2 = 10.965 + 0.9602>2 = 0.96
Index for quarter 3 = I3 = 11.136 + 1.1272>2 = 1.13
Index for quarter 4 = I4 = 11.051 + 1.0632>2 = 1.06
The sum of these indices should be the number of seasons (4) since an average season should
have an index of 1. In this example, the sum is 4. If the sum were not 4, an adjustment would be
made. We would multiply each index by 4 and divide this by the sum of the indices.

Steps Used to Compute Seasonal Indices Based on CMAs

1. Compute a CMA for each observation (where possible).
2. Compute seasonal ratio = Observation/CMA for that observation.
3. Average seasonal ratios to get seasonal indices.
4. If seasonal indices do not add to the number of seasons, multiply each index by (Number
of seasons)/(Sum of the indices).

Figure 5.5 provides a scatterplot of the Turner Industries data and the CMAs. Notice that
the plot of the CMAs is much smoother than the original data. A definite trend is apparent in the
data.
 5.5 TIME-SERIES FORECASTING MODELS 175

FIGURE 5.5
200 CMA
Scatterplot of Turner
Industries Sales and 150
Centered Moving

Sales
Average 100

50 Original Sales Figures

0
1 2 3 4 5 6 7 8 9 10 11 12
Time Period

The Decomposition Method of Forecasting with Trend

and Seasonal Components
The process of isolating linear trend and seasonal factors to develop more accurate forecasts is
called decomposition. The first step is to compute seasonal indices for each season as we have
done with the Turner Industries data. Then, the data are deseasonalized by dividing each num-
ber by its seasonal index, as shown in Table 5.12.
A trend line is then found using the deseasonalized data. Using computer software with
this data, we have*
b1 = 2.34
b0 = 124.78
The trend equation is
YN = 124.78 + 2.34X
where
X = time
This equation is used to develop the forecast based on trend, and the result is multiplied by
the appropriate seasonal index to make a seasonal adjustment. For the Turner Industries data,

TABLE 5.12
SALES SEASONAL DESEASONALIZED
Deseasonalized Data ($1,000,000s) INDEX SALES ($1,000,000s)
for Turner Industries
108 0.85 127.059
125 0.96 130.208
150 1.13 132.743
141 1.06 133.019
116 0.85 136.471
134 0.96 139.583
159 1.13 140.708
152 1.06 143.396
123 0.85 144.706
142 0.96 147.917
168 1.13 148.673
165 1.06 155.660

*If you do the calculations by hand, the numbers may differ slightly from these due to rounding.
176 CHAPTER 5 • FORECASTING

the forecast for the first quarter of year 4 (time period X = 13 and seasonal index I1 = 0.85)
would be found as follows:
YN = 124.78 + 2.34X
= 124.78 + 2.341132
= 155.2 1forecast before adjustment for seasonality2
We multiply this by the seasonal index for quarter 1 and we get
YN * I1 = 155.2 * 0.85 = 131.92
Using this same procedure, we find the forecasts for quarters 2, 3, and 4 of the next year to
be 151.24, 180.66, and 171.95, respectively.

Steps to Develop a Forecast Using the Decomposition Method

1. Compute seasonal indices using CMAs.
2. Deseasonalize the data by dividing each number by its seasonal index.
3. Find the equation of a trend line using the deseasonalized data.
4. Forecast for future periods using the trend line.
5. Multiply the trend line forecast by the appropriate seasonal index.

Most forecasting software, including Excel QM and QM for Windows, includes the decomposi-
tion method as one of the available techniques. This will automatically compute the CMAs, de-
seasonalize the data, develop the trend line, make the forecast using the trend equation, and
adjust the final forecast for seasonality.
The following example provides another application of this process. The seasonal indices
and trend line have already been computed using the decomposition process.
SAN DIEGO HOSPITAL EXAMPLE A San Diego hospital used 66 months of adult inpatient hospi-
tal days to reach the following equation:
YN = 8,091 + 21.5X
where
YN = forecast patient days
X = time, in months
Based on this model, the hospital forecasts patient days for the next month (period 67) to be
Patient days = 8,091 + 121.521672 = 9,532 1trend only2
As well as this model recognized the slight upward trend line in the demand for inpatient serv-
ices, it ignored the seasonality that the administration knew to be present. Table 5.13 provides
seasonal indices based on the same 66 months. Such seasonal data, by the way, were found to be
typical of hospitals nationwide. Note that January, March, July, and August seem to exhibit

TABLE 5.13
MONTH SEASONALITY INDEX MONTH SEASONALITY INDEX
Seasonal Indices for
Adult Inpatient Days January 1.0436 July 1.0302
at San Diego Hospital February 0.9669 August 1.0405
March 1.0203 September 0.9653
April 1.0087 October 1.0048
May 0.9935 November 0.9598
June 0.9906 December 0.9805

Source: W. E. Sterk and E. G. Shryock. “Modern Methods Improve Hospital Forecasting,”

Healthcare Financial Management (March 1987): 97. Reprinted with permission of author.
 5.5 TIME-SERIES FORECASTING MODELS 177

significantly higher patient days on average, while February, September, November, and
December experience lower patient days.
To correct the time-series extrapolation for seasonality, the hospital multiplied the monthly
forecast by the appropriate seasonality index. Thus, for period 67, which was a January,
Patient days = 19,532211.04362 = 9,948 1trend and seasonal2
Using this method, patient days were forecasted for January through June (periods 67 through
72) as 9,948, 9,236, 9,768, 9,678, 9,554, and 9,547. This study led to better forecasts as well as
to more accurate forecast budgets.
USING EXCEL QM FOR DECOMPOSITION In Excel QM, to access the decomposition procedure,
go to the Add-Ins tab and click Excel QM—Forecasting—Decomposition, and the initialization
window will open. Input the relevant information, as illustrated in Program 5.6A, and the
spreadsheet will be initialized for the size of problem specified. Enter the past periods of data,
as shown in Program 5.6B, and the results will appear.
USING QM FOR WINDOWS FOR DECOMPOSITION QM for Windows can also be used for the de-
composition method of forecasting. See Appendix 5.1 for details.

Using Regression with Trend and Seasonal Components

Multiple regression can be used Multiple regression may be used to forecast with both trend and seasonal components present in
to develop an additive a time series. One independent variable is time, and other independent variables are dummy
decomposition model. variables to indicate the season. If we forecast quarterly data, there are four categories (quarters)
so we would use three dummy variables. The basic model is an additive decomposition model
and is expressed as follows:
YN = a + b1X1 + b2X2 + b3X3 + b4X4
where
X1 = time period
X2 = 1 if quarter 2
= 0 otherwise
X3 = 1 if quarter 3
= 0 otherwise
X4 = 1 if quarter 4
= 0 otherwise

PROGRAM 5.6A
Initialization Screen for
the Decomposition
Method in Excel QM

Specify that a centered moving

average should be used.
Input a title, the number of past periods,
and the number of seasons.
Click OK.
178 CHAPTER 5 • FORECASTING

PROGRAM 5.6B
Turner Industries
Seasonal indices are based on CMAs.
Forecast Using the
Decomposition Method Input the past demand. The CMAs are here.
in Excel QM

The intercept and slope are here.

If X2 = X3 = X4 = 0, then the quarter would be quarter 1. It is an arbitrary choice as to which

of the quarters would not have a specific dummy variable associated with it. The forecasts will
be the same regardless of which quarter does not have a specific dummy variable.
Program 5.7A provides the Excel input, and Program 5.7B provides the Excel output for
the Turner Industries example. You can see how the data is input, and the regression equation
(with coefficients rounded) is
YN = 104.1 + 2.3X1 + 15.7X2 + 38.7X3 + 30.1X4
If this is used to forecast sales in the first quarter of the next year, we get
YN = 104.1 + 2.31132 + 15.7102 + 38.7102 + 30.1102 = 134
For quarter 2 of the next year, we get
YN = 104.1 + 2.31142 + 15.7112 + 38.7102 + 30.1102 = 152
Notice these are not the same values we obtained using the multiplicative decomposition
method. We could compare the MAD or MSE for each method and choose the one that is better.

PROGRAM 5.7A
Excel Input for the
Turner Industries
Example Using Multiple
Regression
 5.6 MONITORING AND CONTROLLING FORECASTS 179

PROGRAM 5.7B
Excel Output for the
Turner Industries
Example Using Multiple
Regression

Quarter 1 is indicated by
letting X2 = X3 = X4 = 0.

p-value

IN ACTION Forecasting at Disney World

W hen the Disney chairman receives a daily report from his

main theme parks in Orlando, Florida, the report contains only
With 20% of Disney World’s customers coming from out-
side the United States, its econometric model includes such
variables as consumer confidence and the gross domestic prod-
two numbers: the forecast of yesterday’s attendance at the parks uct of seven countries. Disney also surveys one million people
(Magic Kingdom, Epcot, Fort Wilderness, Hollywood Studios (for- each year to examine their future travel plans and their experi-
merly MGM Studios), Animal Kingdom, Typhoon Lagoon, and ences at the parks. This helps forecast not only attendance, but
Blizzard Beach) and the actual attendance. An error close to zero behavior at each ride (how long people will wait and how
(using MAPE as the measure) is expected. The chairman takes his many times they will ride). Inputs to the monthly forecasting
forecasts very seriously. model include airline specials, speeches by the chair of the Fed-
The forecasting team at Disney World doesn’t just do a daily eral reserve, and Wall Street trends. Disney even monitors
prediction, however, and the chairman is not its only customer. It 3,000 school districts inside and outside the United States for
also provides daily, weekly, monthly, annual, and 5-year forecasts holiday/vacation schedules.
to the labor management, maintenance, operations, finance, and
park scheduling departments. It uses judgmental models, econo- Source: Based on J. Newkirk and M. Haskell. “Forecasting in the Service Sec-
metric models, moving average models, and regression analysis. tor,” presentation at the 12th Annual Meeting of the Production and Operations
The team’s annual forecast of total volume, conducted in 1999 Management Society. April 1, 2001, Orlando, FL.
for the year 2000, resulted in a MAPE of 0.

5.6 Monitoring and Controlling Forecasts

After a forecast has been completed, it is important that it not be forgotten. No manager wants to
be reminded when his or her forecast is horribly inaccurate, but a firm needs to determine why the
actual demand (or whatever variable is being examined) differed significantly from that projected.*

*If the forecaster is accurate, he or she usually makes sure that everyone is aware of his or her talents. Very seldom does

one read articles in Fortune, Forbes, or the Wall Street Journal, however, about money managers who are consistently
off by 25% in their stock market forecasts.
180 CHAPTER 5 • FORECASTING

One way to monitor forecasts to ensure that they are performing well is to employ a
A tracking signal measures how tracking signal. A tracking signal is a measurement of how well the forecast is predicting ac-
well predictions fit actual data. tual values. As forecasts are updated every week, month, or quarter, the newly available demand
data are compared to the forecast values.
The tracking signal is computed as the running sum of the forecast errors (RSFE)
divided by the mean absolute deviation:
RSFE
Tracking signal = (5-13)
MAD
g1forecast error2
=
MAD
where
g ƒ forecast error ƒ
MAD =
n
as seen earlier in Equation 5-1.
Positive tracking signals indicate that demand is greater than the forecast. Negative signals
mean that demand is less than forecast. A good tracking signal—that is, one with a low RSFE—
has about as much positive error as it has negative error. In other words, small deviations are
okay, but the positive and negative deviations should balance so that the tracking signal centers
closely around zero.
Setting tracking limits is a matter When tracking signals are calculated, they are compared with predetermined control limits.
of setting reasonable values for When a tracking signal exceeds an upper or lower limit, a signal is tripped. This means that there
upper and lower limits. is a problem with the forecasting method, and management may want to reevaluate the way it
forecasts demand. Figure 5.6 shows the graph of a tracking signal that is exceeding the range of
acceptable variation. If the model being used is exponential smoothing, perhaps the smoothing
constant needs to be readjusted.
How do firms decide what the upper and lower tracking limits should be? There is no
single answer, but they try to find reasonable values—in other words, limits not so low as
to be triggered with every small forecast error and not so high as to allow bad forecasts to
be regularly overlooked. George Plossl and Oliver Wight, two inventory control experts,
suggested using maximums of ;4 MADs for high-volume stock items and ;8 MADs for
lower-volume items.*
Other forecasters suggest slightly lower ranges. One MAD is equivalent to approximately
0.8 standard deviation, so that ; 2 MADs = 1.6 standard deviations, ; 3 MADs = 2.4 standard
deviations, and ;4 MADs = 3.2 standard deviations. This suggests that for a forecast to be “in

FIGURE 5.6
Plot of Tracking Signals Signal Tripped
Upper Control Limit Tracking Signal


Acceptable
0 MADs
Range


Lower Control Limit

Time

*See G. W. Plossl and O. W. Wight. Production and Inventory Control. Upper Saddle River, NJ: Prentice Hall, 1967.
 SUMMARY 181

control,” 89% of the errors are expected to fall within ;2 MADs, 98% within ;3 MADs, or
99.9% within ;4 MADs whenever the errors are approximately normally distributed.*
KIMBALL’S BAKERY EXAMPLE Here is an example that shows how the tracking signal and RSFE
can be computed. Kimball’s Bakery’s quarterly sales of croissants (in thousands), as well as
forecast demand and error computations, are in the following table. The objective is to compute
the tracking signal and determine whether forecasts are performing adequately.

TIME FORECAST ACTUAL FORECAST CUMULATIVE TRACKING

PERIOD DEMAND DEMAND ERROR RSFE ERROR ERROR MAD SIGNAL
1 100 90 10 10 10 10 10.0 1
2 100 95 5 15 5 15 7.5 2
3 100 115 +15 0 15 30 10.0 0
4 110 100 10 10 10 40 10.0 1
5 110 125 +15 +5 15 55 11.0 +0.5
6 110 140 +30 +35 30 85 14.2 +2.5

In period 6, the calculations are

a ƒ forecast error ƒ 85
MAD = =
n 6
= 14.2
RSFE 35
Tracking signal = =
MAD 14.2
= 2.5 MADs
This tracking signal is within acceptable limits. We see that it drifted from -2.0 MADs to
+ 2.5 MADs.

Adaptive Smoothing
A lot of research has been published on the subject of adaptive forecasting. This refers to com-
puter monitoring of tracking signals and self-adjustment if a signal passes its preset limit. In
exponential smoothing, the  and  coefficients are first selected based on values that minimize
error forecasts and are then adjusted accordingly whenever the computer notes an errant track-
ing signal. This is called adaptive smoothing.

Summary
Forecasts are a critical part of a manager’s function. Demand time-series models were developed. Regression and multiple
forecasts drive the production, capacity, and scheduling sys- regression models were recognized as causal models. Four
tems in a firm and affect the financial, marketing, and person- qualitative models were briefly discussed. In addition, we
nel planning functions. explained the use of scatter diagrams and measures of forecast-
In this chapter we introduced three types of forecasting ing accuracy. In future chapters you will see the usefulness
models: time series, causal, and qualitative. Moving averages, of these techniques in determining values for the various
exponential smoothing, trend projection, and decomposition decision-making models.

*To prove these three percentages to yourself, just set up a normal curve for ;1.6 standard deviations (Z values). Using
the normal table in Appendix A, you find that the area under that the curve is 0.89. This represents ;2 MADs.
Similarly, ; 3 MADs = 2.4 standard deviations encompasses 98% of the area, and so on for ;4 MADs.
182 CHAPTER 5 • FORECASTING

As we learned in this chapter, no forecasting method is its forecasts to make sure that errors do not get out of hand.
perfect under all conditions. Even when management has Forecasting can often be a very challenging but rewarding part
found a satisfactory approach, it must still monitor and control of managing.

Glossary
Adaptive Smoothing The process of automatically monitor- Mean Absolute Percent Error (MAPE) A technique for
ing and adjusting the smoothing constants in an exponential determining the accuracy of a forecasting model by taking
smoothing model. the average of the absolute errors as a percentage of the ob-
Bias A technique for determining the accuracy of a forecast- served values.
ing model by measuring the average error and its direction. Mean Squared Error (MSE) A technique for determining
Causal Models Models that forecast using variables and fac- the accuracy of a forecasting model by taking the average
tors in addition to time. of the squared error terms for a forecasting model.
Centered Moving Average An average of the values Moving Average A forecasting technique that averages past
centered at a particular point in time. This is used to values in computing the forecast.
compute seasonal indices when trend is present. Naïve Model A time-series forecasting model in which the
Decision-Making Group A group of experts in a Delphi forecast for next period is the actual value for the current
technique that has the responsibility of making the forecast. period.
Decomposition A forecasting model that decomposes a time Qualitative Models Models that forecast using judgments,
series into its seasonal and trend components. experience, and qualitative and subjective data.
Delphi A judgmental forecasting technique that uses Running Sum of Forecast Errors (RSFE) Used to develop
decision makers, staff personnel, and respondents to deter- a tracking signal for time-series forecasting models, this is
mine a forecast. a running total of the errors and may be positive or
Deseasonalized Data Time series data in which each value negative.
has been divided by its seasonal index to remove the effect Scatter Diagrams Diagrams of the variable to be forecasted,
of the seasonal component. plotted against another variable, such as time.
Deviation A term used in forecasting for error. Seasonal Index An index number that indicates how a par-
Error The difference between the actual value and the fore- ticular season compares with an average time period (with
cast value. an index of 1 indicating an average season).
Exponential Smoothing A forecasting method that is a Smoothing Constant A value between 0 and 1 that is used
combination of the last forecast and the last observed value. in an exponential smoothing forecast.
Holt’s Method An exponential smoothing model that includes Time-Series Models Models that forecast using only histori-
a trend component. This is also called a double exponential cal data.
smoothing model or a second-order smoothing model. Tracking Signal A measure of how well the forecast is pre-
Least Squares A procedure used in trend projection and dicting actual values.
regression analysis to minimize the squared distances Trend Projection The use of a trend line to forecast a time-
between the estimated straight line and the observed values. series with trend present. A linear trend line is a regression
Mean Absolute Deviation (MAD) A technique for line with time as the independent variable.
determining the accuracy of a forecasting model by taking Weighted Moving Average A moving average forecasting
the average of the absolute deviations. method that places different weights on past values.

Key Equations
g ƒ forecast error ƒ
a ` actual `
error
(5-1) MAD =
n
(5-3) MAPE = 100%
A measure of overall forecast error called mean absolute n
deviation. A measure of forecast accuracy called mean absolute
g1error22 percent error.
(5-2) MSE = Sum of demands in previous n periods
n
(5-4) Moving =
A measure of forecast accuracy called mean squared error. average forecast n
An equation for computing a moving average forecast.
 SOLVED PROBLEMS 183

Yt + Yt - 1 + Á + Yt - n + 1 (5-10) Ft + 1 = FITt + 1Yt - FITt2

Equation to update the smoothed forecast 1Ft + 12 used
(5-5) Ft + 1 =
n
A mathematical expression for a moving average forecast. in the trend adjusted exponential smoothing model.

g 1Weight in period i21Actual value in period i2 (5-11) Tt + 1 = Tt + 1Ft + 1 - FITt2

(5-6) Ft + 1 = Equation to update the smoothed trend value 1Tt + 12
g1Weights2
used in the trend adjusted exponential smoothing model.
An equation for computing a weighted moving average
forecast. (5-12) FITt + 1 = Ft + 1 + Tt + 1
Equation to develop forecast including trend (FIT ) in the
w1Yt + w2Yt - 1 + Á + wnYt - n + 1
(5-7) Ft + 1 = trend adjusted exponential smoothing model.
w1 + w2 + Á + wn
A mathematical expression for a weighted moving aver- RSFE
(5-13) Tracking signal =
age forecast. MAD
(5-8) New forecast = Last period’s forecast + 1Last period’s g1forecast error2
=
actual demand - Last period’s forecast2 MAD
An equation for computing an exponential smoothing An equation for monitoring forecasts with a tracking
forecast. signal.
(5-9) Ft + 1 = Ft + 1Yt - Ft2
Equation 5-8 rewritten mathematically.

Solved Problems

Solved Problem 5-1

Demand for patient surgery at Washington General Hospital has increased steadily in the past few years,
as seen in the following table:

YEAR OUTPATIENT SURGERIES PERFORMED

1 45
2 50
3 52
4 56
5 58
6 —

The director of medical services predicted six years ago that demand in year 1 would be 42 surger-
ies. Using exponential smoothing with a weight of  = 0.20, develop forecasts for years 2 through 6.
What is the MAD?
Solution

YEAR ACTUAL FORECAST (SMOOTHED) ERROR |ERROR|

1 45 42 +3 3
2 50 42.6 = 42 + 0.2(45 - 42) +7.4 7.4
3 52 44.1 = 42.6 + 0.2(50 - 42.6) +7.9 7.9
4 56 45.7 = 44.1 + 0.2(52 - 44.1) +10.3 10.3
5 58 47.7 = 45.7 + 0.2(56 - 45.7) +10.3 10.3
6 — 49.8 = 47.7 + 0.2(58 - 47.7) — —

g ƒ errors ƒ 38.9
MAD = = = 7.78 38.9
n 5
184 CHAPTER 5 • FORECASTING

Solved Problem 5-2

Quarterly demand for Jaguar XJ8’s at a New York auto dealership is forecast with the equation
YN = 10 + 3X
where
X = time period (quarter): quarter 1 of last year = 0
quarter 2 of last year = 1
quarter 3 of last year = 2
quarter 4 of last year = 3
quarter 1 of this year = 4, and so on
and
YN = predicted quarterly demand
The demand for luxury sedans is seasonal, and the indices for quarters 1, 2, 3, and 4 are 0.80, 1.00, 1.30,
and 0.90, respectively. Using the trend equation, forecast the demand for each quarter of next year. Then
adjust each forecast to adjust for seasonal (quarterly) variations.
Solution
Quarter 2 of this year is coded X = 5; quarter 3 of this year, X = 6; and quarter 4 of this year, X = 7.
Hence, quarter 1 of next year is coded X = 8; quarter 2, X = 9; and so on.
YN 1next year quarter 12 = 10 + 132182 = 34 Adjusted forecast = 10.8021342 = 27.2
YN 1next year quarter 22 = 10 + 132192 = 37 Adjusted forecast = 11.0021372 = 37
YN 1next year quarter 32 = 10 + 1321102 = 40 Adjusted forecast = 11.3021402 = 52
YN 1next year quarter 42 = 10 + 1321112 = 43 Adjusted forecast = 10.9021432 = 38.7

Self-Test
䊉 Before taking the self-test, refer to the learning objectives at the beginning of the chapter, the notes in the margins, and the
glossary at the end of the chapter.
䊉 Use the key at the back of the book to correct your answers.
䊉 Restudy pages that correspond to any questions that you answered incorrectly or material you feel uncertain about.

1. Qualitative forecasting models include 4. Which of the following is a time series model?
a. regression analysis. a. the Delphi model
b. Delphi. b. regression analysis
c. time-series models. c. exponential smoothing
d. trend lines. d. multiple regression
2. A forecasting model that only uses historical data for the 5. Which of the following is not a component of a time
variable being forecast is called a series?
a. time-series model. a. seasonality
b. causal model. b. causal variations
c. Delphi model. c. trend
d. variable model. d. random variations
3. One example of a causal model is 6. Which of the following may be negative?
a. exponential smoothing. a. MAD
b. trend projections. b. bias
c. moving averages. c. MAPE
d. regression analysis. d. MSE
 DISCUSSION QUESTIONS AND PROBLEMS 185

7. When comparing several forecasting models to determine 12. If the seasonal index for January is 0.80, then
which one best fits a particular set of data, the model that a. January sales tend to be 80% higher than an average
should be selected is the one month.
a. with the highest MSE. b. January sales tend to be 20% higher than an average
b. with the MAD closest to 1. month.
c. with a bias of 0. c. January sales tend to be 80% lower than an average
d. with the lowest MAD. month.
8. In exponential smoothing, if you wish to give a d. January sales tend to be 20% lower than an average
significant weight to the most recent observations, then month.
the smoothing constant should be 13. If both trend and seasonal components are present in a
a. close to 0. time-series, then the seasonal indices
b. close to 1. a. should be computed based on an overall average.
c. close to 0.5. b. should be computed based on CMAs.
d. less than the error. c. will all be greater than 1.
9. A trend equation is a regression equation in which d. should be ignored in developing the forecast.
a. there are multiple independent variables. 14. Which of the following is used to alert the user of a fore-
b. the intercept and the slope are the same. casting model that a significant error occurred in one of
c. the dependent variable is time. the periods?
d. the independent variable is time. a. a seasonal index
10. Sales for a company are typically higher in the summer b. a smoothing constant
months than in the winter months. This variation would c. a tracking signal
be called a d. a regression coefficient
a. trend. 15. If the multiplicative decomposition model is used to fore-
b. seasonal factor. cast daily sales for a retail store, how many seasons will
c. random factor. there be?
d. cyclical factor. a. 4
11. A naïve forecast for monthly sales is equivalent to b. 7
a. a one-month moving average model. c. 12
b. an exponential smoothing model with  = 0. d. 365
c. a seasonal model in which the seasonal index is 1.
d. none of the above.

Discussion Questions and Problems

Discussion Questions 5-9 Explain how the number of season is determined
when forecasting with a seasonal component.
5-1 Describe briefly the steps used to develop a forecast-
ing system. 5-10 A seasonal index may be less than one, equal to one,
or greater than one. Explain what each of these val-
5-2 What is a time-series forecasting model?
ues would mean.
5-3 What is the difference between a causal model and a
5-11 Explain what would happen if the smoothing con-
time-series model?
stant in an exponential smoothing model was equal
5-4 What is a qualitative forecasting model, and when is to zero. Explain what would happen if the smooth-
it appropriate? ing constant was equal to one.
5-5 What are some of the problems and drawbacks of 5-12 Explain when a CMA (rather than an overall aver-
the moving average forecasting model? age) should be used in computing a seasonal index.
5-6 What effect does the value of the smoothing con- Explain why this is necessary.
stant have on the weight given to the past forecast
and the past observed value? Problems
5-7 Describe briefly the Delphi technique. 5-13 Develop a four-month moving average forecast for
5-8 What is MAD, and why is it important in the selec- Wallace Garden Supply and compute the MAD.
tion and use of forecasting models?

Note: means the problem may be solved with QM for Windows; means the problem may be
solved with Excel QM; and means the problem may be solved with QM for Windows and/or Excel QM.
186 CHAPTER 5 • FORECASTING

A three-month moving average forecast was devel- The sales manager had predicted, before the busi-
oped in the section on moving averages in Table 5.3. ness started, that year 1’s sales would be 410 air con-
5-14 Using MAD, determine whether the forecast in ditioners. Using exponential smoothing with a
Problem 5-13 or the forecast in the section concern- weight of  = 0.30, develop forecasts for years 2
ing Wallace Garden Supply is more accurate. through 6.
5-15 Data collected on the yearly demand for 50-pound 5-20 Using smoothing constants of 0.6 and 0.9, develop
bags of fertilizer at Wallace Garden Supply are forecasts for the sales of Cool-Man air conditioners
shown in the following table. Develop a 3-year mov- (see Problem 5-19).
ing average to forecast sales. Then estimate demand 5-21 What effect did the smoothing constant have on the
again with a weighted moving average in which forecast for Cool-Man air conditioners? (See Prob-
sales in the most recent year are given a weight of 2 lems 5-19 and 5-20.) Which smoothing constant
and sales in the other 2 years are each given a weight gives the most accurate forecast?
of 1. Which method do you think is best? 5-22 Use a three-year moving average forecasting model
to forecast the sales of Cool-Man air conditioners
DEMAND FOR FERTILIZER (see Problem 5-19).
YEAR (1,000S OF BAGS) 5-23 Using the trend projection method, develop a fore-
1 4 casting model for the sales of Cool-Man air condi-
2 6 tioners (see Problem 5-19).
3 4
5-24 Would you use exponential smoothing with a
smoothing constant of 0.3, a 3-year moving average,
4 5 or a trend to predict the sales of Cool-Man air condi-
5 10 tioners? Refer to Problems 5-19, 5-22, and 5-23.
6 8 5-25 Sales of industrial vacuum cleaners at R. Lowenthal
7 7 Supply Co. over the past 13 months are as follows:
8 9
SALES ($1,000s) MONTH SALES ($1,000s) MONTH
9 12
11 January 14 August
10 14
14 February 17 September
11 15
16 March 12 October

5-16 Develop a trend line for the demand for fertilizer in 10 April 14 November
Problem 5-15, using any computer software. 15 May 16 December
5-17 In Problems 5-15 and 5-16, three different forecasts 17 June 11 January
were developed for the demand for fertilizer. These 11 July
three forecasts are a 3-year moving average, a
weighted moving average, and a trend line. Which
one would you use? Explain your answer. (a) Using a moving average with three periods, de-
5-18 Use exponential smoothing with a smoothing con- termine the demand for vacuum cleaners for next
stant of 0.3 to forecast the demand for fertilizer February.
given in Problem 5-15. Assume that last period’s (b) Using a weighted moving average with three pe-
forecast for year 1 is 5,000 bags to begin the proce- riods, determine the demand for vacuum clean-
dure. Would you prefer to use the exponential ers for February. Use 3, 2, and 1 for the weights
smoothing model or the weighted average model de- of the most recent, second most recent, and third
veloped in Problem 5-15? Explain your answer. most recent periods, respectively. For example,
if you were forecasting the demand for February,
5-19 Sales of Cool-Man air conditioners have grown
November would have a weight of 1, December
steadily during the past 5 years:
would have a weight of 2, and January would
have a weight of 3.
YEAR SALES (c) Evaluate the accuracy of each of these methods.
1 450 (d) What other factors might R. Lowenthal consider
2 495 in forecasting sales?
3 518
4 563
5 584
6 ?
 DISCUSSION QUESTIONS AND PROBLEMS 187

5-26 Passenger miles flown on Northeast Airlines, a com-

MONTH INCOME ($1,000S)
muter firm serving the Boston hub, are as follows
for the past 12 weeks: February 70.0
March 68.5

ACTUAL PASSENGER ACTUAL PASSENGER April 64.8

WEEK MILES (1,000S) WEEK MILES (1,000S) May 71.7
1 17 7 20 June 71.3
2 21 8 18 July 72.8
3 19 9 22
4 23 10 20 Use exponential smoothing to forecast August’s in-
5 18 11 15 come. Assume that the initial forecast for February
6 16 12 22 is $65,000. The smoothing constant selected is
 = 0.1.
5-30 Resolve Problem 5-29 with  = 0.3. Using MAD,
(a) Assuming an initial forecast for week 1 of which smoothing constant provides a better forecast?
17,000 miles, use exponential smoothing to 5-31 A major source of revenue in Texas is a state sales
compute miles for weeks 2 through 12. Use tax on certain types of goods and services. Data are
 = 0.2. compiled and the state comptroller uses them to
(b) What is the MAD for this model? project future revenues for the state budget. One par-
(c) Compute the RSFE and tracking signals. Are ticular category of goods is classified as Retail
they within acceptable limits? Trade. Four years of quarterly data (in $millions) for
5-27 Emergency calls to Winter Park, Florida’s 911 sys- one particular area of southeast Texas follow:
tem, for the past 24 weeks are as follows:

QUARTER YEAR 1 YEAR 2 YEAR 3 YEAR 4

WEEK CALLS WEEK CALLS WEEK CALLS
1 218 225 234 250
1 50 9 35 17 55
2 247 254 265 283
2 35 10 20 18 40
3 243 255 264 289
3 25 11 15 19 35
4 292 299 327 356
4 40 12 40 20 60
5 45 13 55 21 75
(a) Compute seasonal indices for each quarter based
6 35 14 35 22 50
on a CMA.
7 20 15 25 23 40 (b) Deseasonalize the data and develop a trend line
8 30 16 55 24 65 on the deseasonalized data.
(c) Use the trend line to forecast the sales for each
quarter of year 5.
(a) Compute the exponentially smoothed forecast of (d) Use the seasonal indices to adjust the forecasts
calls for each week. Assume an initial forecast of found in part (c) to obtain the final forecasts.
50 calls in the first week and use  = 0.1. What 5-32 Using the data in Problem 5-31, develop a multiple
is the forecast for the 25th week? regression model to predict sales (both trend and
(b) Reforecast each period using  = 0.6. seasonal components), using dummy variables to in-
(c) Actual calls during the 25th week were 85. corporate the seasonal factor into the model. Use
Which smoothing constant provides a superior this model to predict sales for each quarter of the
forecast? next year. Comment on the accuracy of this model.
5-28 Using the 911 call data in Problem 5-27, forecast 5-33 Trevor Harty, an avid mountain biker, always wanted
calls for weeks 2 through 25 using  = 0.9. Which to start a business selling top-of-the-line mountain
is best? (Again, assume that actual calls in week 25 bikes and other outdoor supplies. A little over
were 85 and use an initial forecast of 50 calls.) 6 years ago, he and a silent partner opened a store
5-29 Consulting income at Kate Walsh Associates for the called Hale and Harty Trail Bikes and Supplies.
period February–July has been as follows: Growth was rapid in the first 2 years, but since that
188 CHAPTER 5 • FORECASTING

time, growth in sales has slowed a bit, as expected.

YEAR DJIA YEAR 2 DJIA
The quarterly sales (in $1,000s) for the past 4 years
are shown in the table below: 2010 10,431 2000 11,502
2009 8,772 1999 9,213
YEAR 1 YEAR 2 YEAR 3 YEAR 4 2008 13,262 1998 7,908
QUARTER 1 274 282 282 296 2007 12,460 1997 6,448
QUARTER 2 172 178 182 210 2006 10,718 1996 5,117
QUARTER 3 130 136 134 158 2005 10,784 1995 3,834
QUARTER 4 162 168 170 182 2004 10,453 1994 3,754
2003 8,342 1993 3,301
(a) Develop a trend line using the data in the table. 2002 10,022 1992 3,169
Use this to forecast sales for each quarter of 2001 10,791 1991 2,634
year 5. What does the slope of this line indicate?
(b) Use the multiplicative decomposition model to
incorporate both trend and seasonal components 5-38 Using the DJIA data in Problem 5-37, use exponen-
into the forecast. What does the slope of this line tial smooth with trend adjustment to forecast the
indicate? opening DJIA value for 2011. Use  = 0.8 and
 = 0.2. Compare the MSE for this technique with
(c) Compare the slope of the trend line in part a to
the slope in the trend line for the decomposition the MSE for the trend line.
model that was based on the deseasonalized 5-39 Refer to the DJIA data in Problem 5-37.
sales figures. Discuss why these are so different (a) Use an exponential smoothing model with a
and explain which one is best to use. smoothing constant of 0.4 to predict the opening
5-34 The unemployment rates in the United States during DJIA index value for 2011. Find the MSE for this.
a 10-year period are given in the following table. (b) Use QM for Windows or Excel and find the
Use exponential smoothing to find the best forecast smoothing constant that would provide the low-
for next year. Use smoothing constants of 0.2, 0.4, est MSE.
0.6, and 0.8. Which one had the lowest MAD? 5-40 The following table gives the average monthly ex-
change rate between the U.S. dollar and the euro for
2009. It shows that 1 euro was equivalent to 1.324
YEAR 1 2 3 4 5 6 7 8 9 10 U.S. dollars in January 2009. Develop a trend line
Unemployment 7.2 7.0 6.2 5.5 5.3 5.5 6.7 7.4 6.8 6.1 that could be used to predict the exchange rate for
rate (%) 2010. Use this model to predict the exchange rate for
January 2010 and February 2010.
5-35 Management of Davis’s Department Store has used
time-series extrapolation to forecast retail sales for MONTH EXCHANGE RATE
the next four quarters. The sales estimates are January 1.324
$100,000, $120,000, $140,000, and $160,000 for
February 1.278
the respective quarters before adjusting for season-
ality. Seasonal indices for the four quarters have March 1.305
been found to be 1.30, 0.90, 0.70, and 1.10, respec- April 1.320
tively. Compute a seasonalized or adjusted sales May 1.363
forecast.
June 1.402
5-36 In the past, Judy Holmes’s tire dealership sold an av-
July 1.409
erage of 1,000 radials each year. In the past two
years, 200 and 250, respectively, were sold in fall, August 1.427
350 and 300 in winter, 150 and 165 in spring, and September 1.456
300 and 285 in summer. With a major expansion October 1.482
planned, Judy projects sales next year to increase to
November 1.491
1,200 radials. What will the demand be each season?
5-37 The following table provides the Dow Jones Indus- December 1.461
trial Average (DJIA) opening index value on the
first working day of 1991–2010: 5-41 For the data in Problem 5-40, develop an exponen-
Develop a trend line and use it to predict the open- tial smoothing model with a smoothing constant of
ing DJIA index value for years 2011, 2012, and 0.3. Using the MSE, compare this with the model in
2013. Find the MSE for this model. Problem 5-40.
 CASE STUDY 189

Internet Homework Problems

See our Internet home page, at www.pearsonhighered.com/render, for additional homework
problems, Problems 5-42 to 5-50.

Case Study
Forecasting Attendance at SWU Football Games
Southwestern University (SWU), a large state college in built in 1953, has seating for 54,000 fans. The following table
Stephenville, Texas, 30 miles southwest of the Dallas/Fort indicates attendance at each game for the past six years.
Worth metroplex, enrolls close to 20,000 students. In a typical One of Pitterno’s demands upon joining SWU had been a
town–gown relationship, the school is a dominant force in the stadium expansion, or possibly even a new stadium. With atten-
small city, with more students during fall and spring than per- dance increasing, SWU administrators began to face the issue
manent residents. head-on. Pitterno had wanted dormitories solely for his athletes
A longtime football powerhouse, SWU is a member of the in the stadium as an additional feature of any expansion.
Big Eleven conference and is usually in the top 20 in college SWU’s president, Dr. Marty Starr, decided it was time for
football rankings. To bolster its chances of reaching the elusive his vice president of development to forecast when the existing
and long-desired number-one ranking, in 2005 SWU hired the stadium would “max out.” He also sought a revenue projection,
legendary Bo Pitterno as its head coach. Although the number- assuming an average ticket price of $20 in 2011 and a 5% in-
one ranking remained out of reach, attendance at the five Satur- crease each year in future prices.
day home games each year increased. Prior to Pitterno’s arrival,
attendance generally averaged 25,000 to 29,000 per game. Sea- Discussion Questions
son ticket sales bumped up by 10,000 just with the announce-
1. Develop a forecasting model, justify its selection over
ment of the new coach’s arrival. Stephenville and SWU were
other techniques, and project attendance through 2012.
ready to move to the big time!
2. What revenues are to be expected in 2011 and 2012?
The immediate issue facing SWU, however, was not
3. Discuss the school’s options.
NCAA ranking. It was capacity. The existing SWU stadium,

Southwestern University Football Game Attendance, 2005–2010

2005 2006 2007

GAME ATTENDEES OPPONENT ATTENDEES OPPONENT ATTENDEES OPPONENT
1 34,200 Baylor 36,100 Oklahoma 35,900 TCU
2* 39,800 Texas 40,200 Nebraska 46,500 Texas Tech
3 38,200 LSU 39,100 UCLA 43,100 Alaska
4** 26,900 Arkansas 25,300 Nevada 27,900 Arizona
5 35,100 USC 36,200 Ohio State 39,200 Rice

2008 2009 2010

GAME ATTENDEES OPPONENT ATTENDEES OPPONENT ATTENDEES OPPONENT
1 41,900 Arkansas 42,500 Indiana 46,900 LSU
2* 46,100 Missouri 48,200 North Texas 50,100 Texas
3 43,900 Florida 44,200 Texas A&M 45,900 Prairie View A&M
4** 30,100 Miami 33,900 Southern 36,300 Montana
5 40,500 Duke 47,800 Oklahoma 49,900 Arizona State

*Homecoming games
**During the fourth week of each season, Stephenville hosted a hugely popular southwestern crafts festival. This event brought tens of thousands of
tourists to the town, especially on weekends, and had an obvious negative impact on game attendance.
Source: J. Heizer and B. Render. Operations Management, 6th ed. Upper Saddle River, NJ: Prentice Hall, 2001, p. 126.
190 CHAPTER 5 • FORECASTING

Case Study
Forecasting Monthly Sales
For years The Glass Slipper restaurant has operated in a resort shore. They also knew that hiring the right manager would al-
community near a popular ski area of New Mexico. The restau- low James and Deena the time to begin a semi-retirement in a
rant is busiest during the first 3 months of the year, when the corner of paradise.
ski slopes are crowded and tourists flock to the area. To make this happen, James and Deena would have to sell
When James and Deena Weltee built The Glass Slipper, The Glass Slipper for the right price. The price of the business
they had a vision of the ultimate dining experience. As the view would be based on the value of the property and equipment, as
of surrounding mountains was breathtaking, a high priority was well as projections of future income. A forecast of sales for the
placed on having large windows and providing a spectacular next year is needed to help in the determination of the value of
view from anywhere inside the restaurant. Special attention was the restaurant. Monthly sales for each of the past 3 years are
also given to the lighting, colors, and overall ambiance, result- provided in Table 5.14.
ing in a truly magnificent experience for all who came to enjoy
gourmet dining. Since its opening, The Glass Slipper has devel- Discussion Questions
oped and maintained a reputation as one of the “must visit”
1. Prepare a graph of the data. On this same graph, plot a
places in that region of New Mexico.
12-month moving average forecast. Discuss any apparent
While James loves to ski and truly appreciates the moun-
trend and seasonal patterns.
tains and all that they have to offer, he also shares Deena’s
2. Use regression to develop a trend line that could be used
dream of retiring to a tropical paradise and enjoying a more
to forecast monthly sales for the next year. Is the slope of
relaxed lifestyle on the beach. After some careful analysis of
this line consistent with what you observed in question 1?
their financial condition, they knew that retirement was many
If not, discuss a possible explanation.
years away. Nevertheless, they were hatching a plan to bring
3. Use the multiplicative decomposition model on these
them closer to their dream. They decided to sell The Glass
data. Use this model to forecast sales for each month of
Slipper and open a bed and breakfast on a beautiful beach in
the next year. Discuss why the slope of the trend equation
Mexico. While this would mean that work was still in their
with this model is so different from that of the trend equa-
future, they could wake up in the morning to the sight of the
tion in question 2.
palm trees blowing in the wind and the waves lapping at the

TABLE 5.14 MONTH 2008 2009 2010

Monthly Revenue
(in $1,000s) January 438 444 450
February 420 425 438
March 414 423 434
April 318 331 338
May 306 318 331
June 240 245 254
July 240 255 264
August 216 223 231
September 198 210 224
October 225 233 243
November 270 278 289
December 315 322 335

Internet Case Study

See our Internet home page, at www.pearsonhighered.com/render, for the additional case study
on Akron Zoological Park. This case involves forecasting attendance at Akron’s zoo.
 APPENDIX 5.1 FORECASTING WITH QM FOR WINDOWS 191

Bibliography
Berenson, Mark L., David M. Levine, and Timothy C. Kriehbiel. Business Heizer, J., and B. Render. Operations Management, 9th ed. Upper Saddle
Statistics: Concepts and Applications, 10th ed. Upper Saddle River, NJ: River, NJ: Prentice Hall, 2008.
Prentice Hall, 2006. Hyndman, Rob J. “The Interaction Between Trend and Seasonality,”
Billah, Baki, Maxwell L. King Ralph D. Snyder, and Anne B. Koehler. “Expo- International Journal of Forecasting 20, 4 (October–December 2004):
nential Smoothing Model Selection for Forecasting,” International Jour- 561–563.
nal of Forecasting 22, 2, (April–June 2006): 239–247. Hyndman, Rob J., and Anne B. Koehler. “Another Look at Measures of Fore-
Black, Ken. Business Statistics: For Contemporary Decision Making, 6th ed. cast Accuracy,” International Journal of Forecasting 22, 4 (October
John Wiley & Sons, Inc., 2009. 2006): 679–688.
Diebold, F. X. Elements of Forecasting, 2nd ed. Cincinnati: South-Western Li, X. “An Intelligent Business Forecaster for Strategic Business Planning,”
College Publishing, 2001. Journal of Forecasting 18, 3 (May 1999): 181–205.
Gardner, Everette Jr. “Exponential Smoothing: The State of the Art—Part II,” Meade, Nigel. “Evidence for the Selection of Forecasting Methods,” Journal
International Journal of Forecasting 22, 4 (October 2006): 637–666. of Forecasting 19, 6 (November 2000): 515–535.
Granger, Clive W., and J. M. Hashem Pesaran. “Economic and Statistical Snyder, Ralph D., and Roland G. Shami. “Exponential Smoothing of Seasonal
Measures of Forecast Accuracy,” Journal of Forecasting, 19, 7 (Decem- Data: A Comparison,” Journal of Forecasting 20, 3 (April 2001):
ber 2000): 537–560. 197–202.
Hanke, J. E., and D. W. Wichern. Business Forecasting, 9th ed. Upper Saddle Yurkiewicz, J. “Forecasting Software Survey,” OR/MS Today 35, 3 (August
River, NJ: Prentice Hall, 2009. 2008): 54–63.

Appendix 5.1 Forecasting with QM for Windows

In this section, we look at our other forecasting software package, QM for Windows. QM for
Windows can project moving averages (both simple and weighted), do simple and trend-
adjusted exponential smoothing, handle least squares trend projection, solve regression prob-
lems, and use the decomposition method.
To develop forecasts in QM for Windows, select Module on the toolbar and select Forecasting.
Then either click the new document icon or click File—New—Time Series Analysis to enter a new
time series problem. Specify the number of past observations and enter a title, if desired.
To illustrate QM for Windows, we will use the Port of Baltimore data from Table 5.5. The
number of past observations was eight in that example. When you enter those data, the screen
shown in Program 5.8A opens and allows for data input. Once the data is entered, click the

PROGRAM 5.8A
QM for Windows Click the arrow in the Method window to select the desired methods.
Forecasting Methods

Enter the data.

You can change these names by typing over them.

192 CHAPTER 5 • FORECASTING

arrow on the message box to see all the options and select the one desired. In selecting exponen-
tial smoothing for this example, a box appears where  (alpha) may be entered and a column
where any previous forecasts (if available) may be entered, as shown in Program 5.8B. With
other forecasting methods, other types of input boxes may appear. Click the Solve button, and
the Forecasting Results screen appears, as shown in Program 5.8C. If you want to try a different
value for , click Edit to return to the input screen, where you can change . Note that you can
enter an initial forecast if desired, but the error analysis will begin with the first forecast gener-
ated by the computer. Any forecasts entered by the user are ignored in the error analysis.
Notice that additional output, including detailed results of the procedure and a graph, are
available from the Window option in the toolbar once the problem has been solved. With expo-
nential smoothing, one output is called Errors as a function of alpha. This will display the MAD
and MSE for all values of  from 0 to 1, in increments of 0.01. You can simply scroll down this
screen to find the value for  that minimizes the MAD or MSE.
For another example, we will use the decomposition method on the Turner Industries exam-
ple from Table 5.10. Enter a time-series problem with 12 past periods of data and select
Multiplicative Decomposition under Method. When this is done, additional input is needed, so
indicate that there are four seasons, select Centered Moving Average as the basis for smoothing,
and specify that the seasonal factors should not be rescaled, as shown in Program 5.9. This out-
put screen provides both the unadjusted forecasts found using the trend equation on the desea-
sonalized data and the final or adjusted forecasts, which are found by multiplying the unadjusted
forecast by the seasonal factor or index. Additional details can be seen by selecting Details and
Error Analysis under Window.

PROGRAM 5.8B
Exponential Smoothing Click Solve.
in the Port of Baltimore
Example with QM for
Windows

After exponential smoothing is selected,

the option of selecting alpha appears.
Input the desired value. Other options
appear for other forecasting methods.

You can input the initial forecast,

if there is one. Otherwise, it will
be assumed to be the same as the
actual value (180 in this example).
 APPENDIX 5.1 FORECASTING WITH QM FOR WINDOWS 193

PROGRAM 5.8C Other output options are available

Exponential Smoothing under Window on the toolbar.
in the Port of Baltimore
Output with QM for
Windows
Watch for notes such as this one,
which specifies that the error
analysis begins with the first period
with a forecast that was not
entered by the user.

A graph is available for all

time-series methods.

The forecast for the next

period is here.

PROGRAM 5.9 When data were input, 4, Centered Moving

QM for Windows Average, and Rescale were all specified. If
the ‘Do not rescale’ option is selected, there
Decomposition Output
will be slight differences in the output.
for the Turner Industries
Example

Additional output, including a graph

and detailed analysis, is available
from the Window drop-down menu.

Seasonal indices are found here. The

unadjusted forecast is multiplied by
the appropriate seasonal index to get
the adjusted (final) forecast.

The trend equation is here. The unadjusted forecasts come from this.
This page intentionally left blank