11. Use data tables to perform sensitivity analysis when dealing with a sequence of decisions.

12. Use utilities to better reflect the values of payoffs.

13. Describe some common features in the practical application of decision analysis.
Chapter Nine The previous chapters have focused mainly on managerial decision making
when the consequences of alternative decisions are known with a reasonable
degree of certainty. This decision-making environment enabled formulating
Decision Analysis helpful mathematical models (linear programming, integer programming,
nonlinear programming, etc.) with objective functions that specify the estimated
consequences of any combination of decisions. Although these consequences
usually cannot be predicted with complete certainty, they could at least be
estimated with enough accuracy to justify using such models (along with what-if
analysis, etc.).
However, managers (or any other individuals) often must make decisions in
environments that are fraught with much more uncertainty. Here are a few
examples.

. A manufacturer introducing a new product into the marketplace. What will be

the reaction of potential customers? How much should be produced? Should
the product be test marketed in a small region before deciding upon full
distribution? How much advertising is needed to launch the product
successfully?
. A financial firm investing in securities. Which are the market sectors and
Learning Objectives individual securities with the best prospects? Where is the economy headed?
After completing this chapter, you should be able to
How about interest rates? How should these factors affect the investment
decisions?
1. Identify the kind of decision-making environment for which decision analysis is needed.
. A government contractor bidding on a new contract. What will be the actual
2. Describe the logical way in which decision analysis organizes a problem.
costs of the project? Which other companies might be bidding? What are their
3. Formulate a payoff table from a description of the problem.
likely bids?
4. Describe and evaluate several alternative criteria for making a decision based on a payoff
table. . An agricultural firm selecting the mix of crops and livestock for the upcoming
5. Apply Bayes’ decision rule to solve a decision analysis problem. season. What will be the weather conditions? Where are prices headed? What
6. Formulate and solve a decision tree for dealing with a sequence of decisions. will costs be?
7. Use Analytic Solver to construct and solve a decision tree. . An oil company deciding whether to drill for oil in a particular location. How
8. Perform sensitivity analysis with Bayes’ decision rule. likely is there to be oil in that location? How much? How deep will they need
9. Determine whether it is worthwhile to obtain more information before making a decision. to drill? Should geologists investigate the site further before drilling?
10. Use new information to update the probabilities of the states of nature.
 This list of examples could go on and on because so many decisions, or report about one of these tracts. A consulting geologist has just informed Max
sequences of decisions, need to be made in the face of great uncertainty. Such that he believes there is one chance in four of oil there.
decision making in all of business and all of life can be greatly improved by Max has learned from bitter experience to be skeptical about the chances of
using a rational art and science of decision making. This is the reason why the oil reported by consulting geologists. Drilling for oil on this tract would
field of decision analysis has been carefully developed over recent centuries. require an investment of about $100,000. If the land turns out to be dry (no oil),
Decision analysis provides a framework and methodology for rational decision the entire investment would be lost. Since his company does not have much
making when the outcomes are uncertain. capital left, this loss would be quite serious.
On the other hand, if the tract does contain oil, the consulting geologist
page 335 estimates that there would be enough there to generate a net revenue of
approximately $800,000, leaving an approximate profit of
The first section introduces a case study that will be carried throughout the
chapter to illustrate the various phases involved in applying decision analysis.
Section 9.2 focuses on choosing an appropriate decision criterion and then the
next section describes how decision trees can be used to structure and analyze a
decision analysis problem. Section 9.4 discusses how sensitivity analysis can Although this wouldn’t be quite the big strike for which Max has been waiting,
be performed efficiently with the help of decision trees. The subsequent four it would provide a very welcome infusion of capital into the company to keep it
sections deal with whether it would be worthwhile to obtain more information going until he hopefully can hit the really big gusher.
and, if so, how to use this information for making a sequence of decisions.
Section 9.9 then describes how to analyze the problem while calibrating the Should Max sell the land instead of drilling for oil there?
possible outcomes to reflect their true value to the decision maker. Finally,
Section 9.10 discusses the practical application of decision analysis. There is another option. Another oil company has gotten wind of the
In addition, a supplement to this chapter at www.mhhe.com/Hillier6e consulting geologist’s report and so has offered to purchase the tract of land
presents a detailed description and evaluation of various decision criteria. from Max for $90,000. This is very tempting. This too would provide a
welcome infusion of capital into the company, but without incurring the large
risk of a very substantial loss of $100,000.
9.1 A CASE STUDY: THE GOFERBROKE Table 9.1 summarizes the decision alternatives and prospective payoffs that
face Max.
COMPANY PROBLEM So Max is in a quandary about what to do. Fortunately, help is at hand.
Max’s daughter Jennifer has recently earned her degree from a fine business
Max Flyer is the founder and sole owner of the Goferbroke Company, which school and now has come to work for her proud dad. He asks her to apply her
develops oil wells in unproven territory. Max’s friends refer to him business training to help him analyze the problem. Having studied management
affectionately as a wildcatter. However, he prefers to think of himself as an science in college, she recommends applying decision analysis. Having paid
entrepreneur. He has poured his life’s savings into the company in the hope of for her fine education, he agrees to give it a try.
making it big with a large strike of oil.
Now his chance possibly has come. His company has purchased various TABLE 9.1
tracts of land that larger oil companies have spurned as unpromising even Prospective Profits for the Goferbroke Company
though they are near some large oil fields. Now Max has received an exciting
 Profit Jennifer: Now this is what I suggest we do. We’ll start out simple, without
Alternative \ Status of Land Oil Dry considering the option of the seismic survey and without getting into utilities.
Drill for oil $700,000 −$100,000 I’ll introduce you to how decision analysis organizes our problem and to the
Sell the land 90,000 90,000 options it provides for the criterion to use for making your decision. You’ll be
Chance of status 1 in 4 3 in 4 able to choose the criterion that feels right to you. Then we’ll look at whether it
might be worthwhile to do the seismic survey and, if so, how to best use its
information. After that, we’ll get into the nitty gritty of carefully analyzing the
page 336 problem, including incorporating utilities. I think when we finish the process
Jennifer begins by interviewing her dad about the problem. and you make your decision, you’ll feel quite comfortable that you are making
the best one.
Jennifer: How much faith do you put in the consulting geologist’s assessment Max: Good. Let’s get started.
that there is one chance in four of oil on this tract?
Here is the tutorial that Jennifer provided her dad about the logical way in
Max: Not too much. These guys sometimes seem to pull numbers out of the air.
which decision analysis organizes a problem.
He has convinced me that there is some chance of oil there. But it could just as
well be one chance in three, or one chance in five. They don’t really know.
Decision Analysis Terminology
Jennifer: Is there a way of getting more information to pin these odds down
better? This is an important option with the decision analysis approach. Decision analysis has a few special terms.
Max: Yes. We could arrange for a detailed seismic survey of the land. That The decision maker is the individual or group responsible for making the
would pin down the odds somewhat better. But you don’t really find out until decision (or sequence of decisions) under consideration. For the Goferbroke
you drill. Furthermore, these seismic surveys cost you an arm and a leg. I got a Co. problem, the decision maker is Max. Jennifer (the management scientist)
quote for this tract. 30,000 bucks! Then it might say oil is likely, so we drill and can help perform the analyses, but the objective is to assist the decision maker
we might not find anything. Then I’m out another 100,000 bucks! Losing in identifying the best possible decision from the decision maker’s perspective.
$130,000 would almost put us out of business. The alternatives are the options for the decision to be made by the
Jennifer: OK. Let’s put the seismic survey on the back burner for now. Here is decision maker. Max’s alternatives at this point are to drill for oil or to sell the
another key consideration. It sounds like we need to go beyond dollars and tract of land.
cents to look at the consequences of the possible outcomes. Losing $130,000 The outcome of the decision to be made will be affected by random factors
would hurt a lot more than gaining $130,000 would help. that are outside the control of the decision maker. These random factors
Max: That’s for sure! determine the situation that will be found when the decision is executed. Each
Jennifer: Well, decision analysis has a way of taking this into account by using of these possible situations is referred to as a possible state of nature. For
what are called utilities. The utility of an outcome measures the true value to the Goferbroke Co. problem, the possible states of nature are that the tract
you of that outcome rather than just the monetary value. contains oil or that it is dry (no oil).
The decision maker generally will have some information about the relative
The utility of an outcome measures the true value to the decision maker of that outcome. likelihood of the possible states of nature. This information may be in the form
of just subjective estimates based on the experience or intuition of an
Max: Sounds good. individual, or there may be some degree of hard evidence involved (such as is
contained in the consulting geologist’s report). When these estimates are
expressed in the form of probabilities, they are referred to as the prior When formulating the problem, it is important to identify all the relevant
probabilities of the respective states of nature. For the Goferbroke Co. decision alternatives and the possible states of nature. After identifying the
problem, the consulting geologist has provided the prior probabilities given in appropriate measure for the payoff from the perspective of the decision maker,
Table 9.2. Although these are unlikely to be the true probabilities based on the next step is to estimate the payoff for each combination of a decision
more information (such as through a seismic survey), they are the best available alternative and a state of nature. These payoffs then are displayed in a payoff
estimates of the probabilities prior to obtaining more information. (Later in the table.
chapter, we will analyze whether it would be worthwhile to conduct a seismic Table 9.3 shows the payoff table for the first Goferbroke Co. problem. The
survey, so the current problem of what to do without a seismic survey will be payoffs are given in units of thousands of dollars of profit. Note that the bottom
referred to hereafter as the first Goferbroke Co. problem.) row also shows the prior probabilities of the states of nature, as given earlier
in Table 9.2.
page 337 For clarity, this example has been kept as small as possible, with just two
decision alternatives and two states of nature, because subsequent analysis in
TABLE 9.2 later sections will enlarge the size of the problem. Most applications of
Prior Probabilities for the First Goferbroke Co. Problem
decision analysis have far larger numbers of decision alternatives and states of
State of Nature Prior Probability nature. Even for this particular problem, it might be of interest to consider more
The tract of land contains oil 0.25 than two decision alternatives (different types of drilling) and more than two
The tract of land is dry (no oil) 0.75 states of nature (different amounts of oil).

TABLE 9.3 Review Questions

Payoff Table (Profit in $1,000s) for the First Goferbroke Co. Problem
1. What are the decision alternatives being considered by Max?
State of Nature 2. What is the consulting geologist’s assessment of the chances of oil on the tract of land?
Alternative Oil Dry 3. How much faith does Max put in the consulting geologist’s assessment of the chances of oil?
Drill for oil 700 −100 4. What option is available for obtaining more information about the chances of oil?
Sell the land 90 90 5. What is meant by the possible states of nature?
Prior probability 0.25 0.75 6. What is meant by prior probabilities?
7. What do the payoffs represent in a payoff table?
Each combination of a decision alternative and a state of nature results in
some outcome. The payoff is a quantitative measure of the value to the
decision maker of the consequences of the outcome. In most cases, the payoff is 9.2 DECISION CRITERIA
expressed as a monetary value, such as the profit. As indicated in Table 9.1, the
payoff for the Goferbroke Co. at this stage is profit. (In Section 9.9, the Given the payoff table for the first Goferbroke Co. problem shown in Table
company’s payoffs will be reexpressed in terms of utilities.) 9.3, what criterion should be used in deciding whether to drill for oil or sell the
land? There is no single correct answer for this question that is appropriate for
The Payoff Table every decision maker. The choice of a decision criterion depends considerably
on the decision maker’s own temperament and attitude toward decision making,
as well as the circumstances of the decision to be made. Ultimately, Max Flyer,
as the owner of the Goferbroke Co., needs to decide which decision criterion is afterward if the decision does not turn out well, and the realism criterion that
most appropriate for this situation from his personal viewpoint. uses the decision maker’s pessimism–optimism index.)

There is no single decision criterion that is best for every situation.

Decision Making without Probabilities: The Maximax
Criterion
page 338
The maximax criterion is the decision criterion for the eternal optimist. It
Over a period of many decades (and even centuries), the development of says to focus only on the best that can happen to us. Here is how this criterion
the field of decision analysis has suggested a considerable number of criteria works:
for how to make a decision when given the kind of information provided by a
. Identify the maximum payoff from any state of nature for each decision
payoff table. All these criteria consider the payoffs in some way and some also
alternative.
take into account the prior probabilities of the states of nature, but other criteria
do not use probabilities in any way. Each criterion has some rationale as well . Find the maximum of these maximum payoffs and choose the corresponding
as some drawbacks. However, in recent decades, a substantial majority of decision alternative.
management scientists has concluded that one of these criteria (Bayes’ decision The rationale for this criterion is that it gives an opportunity for the best
rule) is a particularly appropriate criterion for most decision makers in most possible outcome (the largest payoff in the entire payoff table) to occur. All that
situations. Therefore, after describing and discussing Bayes’ decision rule later is needed is for the right state of nature to occur, which the eternal optimist
in this section, the rest of the chapter will focus on how to apply this particular believes is likely.
criterion in a variety of contexts.
However, before turning to Bayes’ decision rule, we will begin this section The maximax criterion always chooses the decision alternative that can give the largest
by briefly introducing three of the most important alternative decision criteria. possible payoff.
All three of these alternative criteria are particularly simple and intuitive. At
the same time, each criterion is quite superficial in the sense that it focuses on Table 9.4 shows the application of this criterion to the first Goferbroke
only one piece of information provided by the payoff table and ignores the rest problem. It begins with the payoff table (Table 9.3) without the prior
(including the information considered by the other two criteria). Nevertheless, probabilities (since these probabilities are ignored by this criterion). An extra
many individuals informally apply one or more of these criteria at various column on the right then shows the maximum payoff for each decision
times in their lives. The first two make no use of prior probabilities, which can alternative. Since the maximum of these maxima (700) must be the largest
be quite reasonable when it is difficult or impossible to obtain relatively payoff in the entire payoff table, the corresponding decision alternative (drill
reliable values for these probabilities. Bayes’ decision rule is quite different for oil) is selected by this criterion.
from these alternative criteria in that it makes full use of all the information in
This criterion ignores the prior probabilities.
the payoff table by applying a more structured approach to decision making.
www.mhhe.com/Hillier6e includes a supplement to this chapter entitled The biggest drawback of this criterion is that it completely ignores the prior
Decision Criteria that provides a much more detailed discussion and critique probabilities. For example, it always would say that Goferbroke should drill
of these three alternative decision criteria as well as three others that are for oil even if the chance of finding oil were minuscule. Another drawback is
somewhat more complicated. (These three others included in the supplement that it ignores all the payoffs except the largest one. For example, it again
are the equally likely criterion that assigns equal probabilities to all the states
of nature, the minimax regret criterion that minimizes the regret that can be felt
 would say that Goferbroke should drill for oil even if the payoff from selling This criterion also ignores the prior probabilities.
the land were 699 ($699,000).
The drawbacks of this criterion are the reverse of those for the maximax
TABLE 9.4 criterion. Because it completely ignores prior probabilities, it always would
Application of the Maximax Criterion to the First Goferbroke Co. Problem say that Goferbroke should sell the land even if it were almost certain to find
oil if it drilled. Because it ignores all the payoffs except the maximin payoff, it
again would say that Goferbroke should sell the land even if the payoff from
drilling successfully for oil were 10,000 ($10 million).

Decision Making with Probabilities: The Maximum

Likelihood Criterion
page 339 The maximum likelihood criterion says to focus on the most likely state of
nature as follows.
Decision Making without Probabilities: The Maximin
Criterion . Identify the state of nature with the largest prior probability.
The maximin criterion is the criterion for the total pessimist. In contrast to . Choose the decision alternative that has the largest payoff for this state of
the maximax criterion, it says to focus only on the worst that can happen to us. nature.
Here is how this criterion works:
The maximum likelihood criterion assumes that the most likely state of nature will occur
. Identify the minimum payoff from any state of nature for each decision and chooses accordingly.
alternative.
The rationale for this criterion is that by basing our decision on the
. Find the maximum of these minimum payoffs and choose the corresponding
assumption that the most likely state of nature will occur, we are giving
decision alternative.
ourselves a better chance of a favorable outcome than by assuming any other
The maximin criterion always chooses the decision alternative that provides the best state of nature. Furthermore, it is only necessary to identify the most likely state
guarantee for its minimum possible payoff. of nature without needing to develop close estimates of the prior probabilities
for the less likely states of nature.
The rationale for this criterion is that it provides the best possible protection Table 9.6 shows the application of this criterion to the first Goferbroke Co.
against being unlucky. Even if each possible choice of a decision alternative problem. This table is identical to the payoff table given in Table 9.3 except for
were to lead to its worst state of nature occurring, which the total pessimist also showing step 1 (select the dry state of nature) and step 2 (select the sell
thinks is likely, the choice indicated by this criterion gives the best possible the land alternative) of the criterion. Since dry is the state of nature with the
payoff under these circumstances. larger prior probability, we only consider the payoffs in its column (−100 and
The application of this criterion to the first Goferbroke problem is shown in 90). The larger of these two payoffs is 90, so we choose the corresponding
Table 9.5. The basic difference from Table 9.4 is that the numbers in the right- alternative, sell the land.
hand column now are the minimum rather than the maximum in each row. Since
90 is the maximum of these two numbers, the alternative to be chosen is to sell TABLE 9.5
the land. Application of the Maximin Criterion to the First Goferbroke Co. Problem
 nature and then summing these products. Using statistical terminology, refer to
this weighted average as the expected payoff (EP) for this decision
alternative.
. Bayes’ decision rule says to choose the alternative with the largest expected
payoff.
TABLE 9.6
Application of the Maximum Likelihood Criterion to the First Goferbroke Co. Problem
The spreadsheet in Figure 9.1 shows the application of this criterion to the
first Goferbroke Co. problem. Columns B, C, and D display the payoff table
first given in Table 9.3. Cells F5 and F6 then execute step 1 of the procedure by
using the equations entered into these cells, namely,
F5 = SUMPRODUCT ( PriorProbability, DrillPayoff ) = 0.25 ( 700 ) +
0.75 ( −100 ) = 100
F6 = SUMPRODUCT ( PriorProbability, SellPayoff ) = 0.25 ( 90 ) + 0.75 (
90 ) = 90

page 340
This criterion ignores all the payoffs except for the most likely state of nature.

This criterion has a number of drawbacks. One is that with a considerable

number of states of nature, the most likely state can have a fairly low prior
probability, in which case it would make little sense to base the decision solely
on this one state. Another more serious drawback is that it completely ignores
all the payoffs (including any extremely large payoffs and any disastrous
payoffs) throughout the payoff table except those for the single most likely state
of nature. For example, no matter how large the payoff for finding oil, it
automatically would say that Goferbroke should sell the land instead of drilling
for oil even if the dry state has only a slightly larger prior probability than the
oil state.

Decision Making with Probabilities: Bayes’ Decision

Rule
FIGURE 9.1
Bayes’ decision rule directly uses the prior probabilities of the possible This spreadsheet shows the application of Bayes’ decision rule to the first Goferbroke Co.
states of nature as summarized below. problem, where a comparison of the expected payoffs in cells F5:F6 indicates that the Drill
alternative should be chosen because it has the largest expected payoff.
. For each decision alternative, calculate the weighted average of its payoffs by
multiplying each payoff by the prior probability of the corresponding state of page 341
 Since expected payoff = 100 for the drill alternative (cell F5), (Section 9.6 describes how new information sometimes can be obtained to
versus a smaller value of expected payoff = 90 for the sell the land alternative improve prior probabilities and make them more objective.)
(cell F6), this criterion says to drill for oil. . By focusing on average outcomes, expected (monetary) payoffs ignore the
Like all the others, this criterion cannot guarantee that the selected effect that the amount of variability in the possible outcomes should have on
alternative will turn out to have been the best one after learning the true state of the decision making. For example, since Goferbroke does not have the
nature. However, it does provide another guarantee described below. financing to sustain a large loss, selling the land to assure a payoff of 90
($90,000) may be preferable to an expected payoff of 100 ($100,000) from
On the average, Bayes’ decision rule provides larger payoffs in the long run than any
other criterion.
drilling. Selling would avoid the risk of a large loss from drilling when the
land is dry. (Section 9.9 will discuss how utilities can be used to better reflect
The expected payoff for a particular decision alternative can be interpreted as what the average the value of payoffs.)
payoff would become if the same situation were to be repeated numerous times. Therefore, on the
average, repeatedly using Bayes’ decision rule to make decisions will lead to larger payoffs in the By considering only expected payoffs, Bayes’ decision rule fails to give special
long run than any other criterion (assuming the prior probabilities are valid). consideration to the possibility of disastrously large losses.
Thus, if the Goferbroke Co. owned many tracts of land with this same payoff
table, drilling for oil on all of them would provide an average payoff of about So why is this criterion commonly referred to as Bayes’ decision rule? The
100 ($100,000), versus only 90 ($90,000) for selling. As the following reason is that it is often credited to the Reverend Thomas Bayes, a
calculations indicate, this is the average payoff from drilling that results from nonconforming 18th century English minister who won renown as a philosopher
having oil in an average of one tract out of every four (as indicated by the prior and mathematician, although the same basic idea has even longer roots in the
probabilities). field of economics. Bayes’ philosophy of decision making still is very
influential today, and some management scientists even refer to themselves as
Bayesians because of their devotion to this philosophy.
More recently, it has become somewhat popular to also call this criterion
the expected monetary value (EMV) criterion. The reason for this
alternative name is that the payoffs in the payoff table often represent monetary
values (such as the number of dollars of profit), in which case the expected
However, achieving this average payoff might require going through a long payoff for each decision alternative is its expected monetary value. However,
stretch of dry tracts until the “law of averages” can prevail to reach 25 percent the name is a misnomer for those cases where the measure of the payoff is
of the tracts having oil. Surviving a long stretch of bad luck may not be feasible something other than monetary value (as in Section 9.9). Therefore, we will
if the company does not have adequate financing. consistently use the single name, Bayes’ decision rule, to refer to this criterion
This criterion also has its share of critics. Here are the main criticisms. in all situations.
Because of its popularity, the rest of the chapter focuses on procedures that
. There usually is considerable uncertainty involved in assigning values to prior
are based on this criterion.
probabilities, so treating these values as true probabilities will not reveal the
true range of possible outcomes. (Section 9.4 discusses how sensitivity
analysis can address this concern.) page 342
. Prior probabilities inherently are at least largely subjective in nature, whereas Max’s Reaction
sound decision making should be based on objective data and procedures.
Max: So where does this leave us? 9.3 DECISION TREES
Jennifer: Well, now you need to decide which criterion seems most
appropriate to you in this situation. The spreadsheet in Figure 9.1 illustrates one useful way of performing decision
Max: Well, I can’t say that I am very excited about any of the criteria. But it analysis with Bayes’ decision rule. Another enlightening way to apply this
sounded like this last one is a popular one. decision rule is to use a decision tree to display and analyze the problem
Jennifer: Yes, it is. graphically. The decision tree for the first Goferbroke Co. problem is shown in
Max: Why? Figure 9.2. Starting on the left side and moving to the right side shows the
progression of events. First, a decision is made as to whether to drill for oil or
Bayes’ decision rule uses all the information provided by the payoff table. sell the land. If the decision is to drill, the next event is to learn whether the
state of nature is that the land contains oil or is dry. Finally, the payoff is
Jennifer: Really, two reasons. First, this is the criterion that uses all the
obtained that results from these events.
available information. The prior probabilities may not be as accurate as we
In the terminology of decision trees, the junction points are called nodes
would like, but they do give us valuable information about roughly how likely
each of the possible states of nature is. Many management scientists feel that (or forks) and the lines emanating from the nodes are referred to as branches.
using this key information should lead to better decisions. A distinction is then made between the following two types of nodes.
Max: I’m not ready to accept that yet. But what is the second reason? A decision node, represented by a square, indicates that a decision needs to be made at that
point in the process. An event node (or chance node), represented by a circle, indicates that a
Jennifer: Remember that this is the criterion that focuses on what the average random event occurs at that point.
payoff would be if the same situation were repeated numerous times. We called
this the expected payoff. Consistently selecting the decision alternative that Thus, node A in Figure 9.2 is a decision node since the decision on whether to
provides the best expected payoff would provide the most payoff to the drill or sell occurs there. Node B is an event node since a random event, the
company in the long run. Doing what is best in the long run seems like rational occurrence of one of the possible states of nature, takes place there. Each of the
decision making for a manager. two branches emanating from this node corresponds to one of the possible
random events, where the number in parentheses along the branch gives the
Review Questions probability that this event will occur.
A decision tree can be very helpful for visualizing and analyzing a problem.
1. How does the maximax criterion select a decision alternative? What kind of person might find When the problem is as small as the one in Figure 9.2, using the decision tree in
this criterion appealing?
the analysis process is optional. However, one nice feature of decision trees is
2. What are some criticisms of the maximax criterion?
that they also can be used for more complicated problems where a sequence of
3. How does the maximin criterion select a decision alternative? What kind of person might find
this criterion appealing? decisions needs to be made. You will see this illustrated for the full Goferbroke
4. What are some criticisms of the maximin criterion? Co. problem in Sections 9.7 and 9.9 when a decision on whether to conduct a
5. Which state of nature does the maximum likelihood criterion focus on? seismic survey is made before deciding whether to drill or sell.
6. What are some criticisms of the maximum likelihood criterion?
7. How does Bayes’ decision rule select a decision alternative?
8. How is the expected payoff for a decision alternative calculated?
9. What are some criticisms of Bayes’ decision rule?
 the payoff associated with selling is 90 (the $90,000 selling price). After
making all of these entries as shown in Figure 9.3, clicking OK then yields the
decision tree shown in Figure 9.4.

Analytic Solver Tip: To create a new node at the end of a tree, select the cell
containing the terminal node and choose Add Node from the Decision Tree/Node menu
on the Analytic Solver ribbon. Select the type of node and enter names, values, and (if an
event node) probabilities for each branch.

FIGURE 9.2
The decision tree for the first Goferbroke Co. problem as presented in Table 9.3.
If the decision is to drill, the next event is to learn whether or not the land
contains oil. To create an event node, click on the cell containing the triangle
terminal node at the end of the drill branch (cell F3 in Figure 9.4) and choose
page 343 Add Node from the Decision Tree > Node menu on the Analytic Solver ribbon
to bring up the dialog box shown in Figure 9.5. The node is an event node with
Spreadsheet Software for Decision Trees
two branches, Oil and Dry, with probabilities 0.25 and 0.75, respectively, and
We will describe and illustrate how to use Analytic Solver to construct and values (partial payoffs) of 800 and 0, respectively, as entered into the dialog
analyze decision trees in Excel. Instructions for installing this software are box in Figure 9.5. After clicking OK, the final decision tree is shown in Figure
available at www.mhhe.com /Hillier6e. If you are a Mac user (Analytic Solver 9.6. (Note that Analytic Solver, by default, shows all probabilities as a
is not compatible with Mac versions of Excel) or you or your instructor simply percentage on the spreadsheet, with 25% and 75% in H1 and H6, rather than
prefer to use different software, a supplement to this chapter at 0.25 and 0.75.)
www.mhhe.com/Hillier6e contains instructions for TreePlan, another popular
Excel add-in for constructing and analyzing decision trees in Excel.

Analytic Solver Tip: The type of node, the branch names, and the values (or partial
payoffs) of each branch can be entered in the Analytic Solver Decision Tree dialog box.
Alternatively, the names and values can be entered (or changed) after the fact by typing
them directly into the spreadsheet.

To begin creating a decision tree using Analytic Solver, select Add Node
from the Decision Tree > Node menu on the Analytic Solver ribbon. This
brings up the dialog box shown in Figure 9.3. Here you can choose the type of
node (Decision or Event), give names to each of the branches, and specify a
value for each branch (the partial payoff associated with that branch). The
default names for the branches of a decision node in Analytic Solver are
Decision 1 and Decision 2. These can be changed (or more branches added) by
double-clicking on the branch name (or in the next blank row to add a branch)
and typing in a new name. The initial node in the first Goferbroke problem
(node A in Figure 9.2) is a decision node with two branches: Drill and Sell. FIGURE 9.3
The payoff associated with drilling is –100 (the $100,000 cost of drilling) and
 The Decision Tree dialog box used to specify that the initial node of the first Goferbroke claim days paid that would trigger claim-management intervention so as to minimize the
problem is a decision node with two branches, Drill and Sell, with values (partial payoffs) of expected cost of claim payments and intervention.
−100 and 90, respectively. This application of decision analysis with decision trees is saving WCB
approximately US$4 million per year while also enabling some injured workers to return
to work sooner.

Source: E. Urbanovich, E. E. Young, M. L. Puterman, and S. O. Fattedad, “Early

Detection of High-Risk Claims at the Workers’ Compensation Board of British
Columbia,” Interfaces 33, no. 4 (July–August 2003), pp. 15–26. (A link to this article is
provided at www.mhhe.com/Hillier6e.)

Analytic Solver Tip: To make changes to a node, select the node and choose an
FIGURE 9.4 appropriate option under the Decision Tree menu on the Analytic Solver ribbon.
The initial, partial decision tree created by Analytic Solver by selecting Add Node from the
Decision Tree/Node menu on the Analytic Solver ribbon and specifying a Decision node with
two branches named Drill and Sell, with partial payoffs of −100 and 90, respectively. At any time, you also can click on any existing node and make changes using
various choices under the Decision Tree menu on the Analytic Solver ribbon.
For example, under the Node submenu, you can choose Add Node, Change
page 344
Node, Delete Node, Copy Node, or Paste Node. Under the Branch submenu,
you can Add Branch, Change Branch, or Delete Branch.

An Application Vignette

The Workers’ Compensation Board (WCB) of British Columbia, Canada, is

responsible for the occupational health and safety, rehabilitation, and compensation
interests of the province’s workers and employers. The WCB serves more than 200,000
employers who employ about 2.3 million workers in British Columbia. It spends well over
US$1 billion annually on compensation and rehabilitation.
A key factor in controlling WCB costs is to identify those short-term disability claims
that pose a potentially high financial risk of converting into a far more expensive long-
term disability claim unless there is intensive early claim-management intervention to
provide the needed medical treatment and rehabilitation. The question was how to
accurately identify these high-risk claims so as to minimize the expected total cost of
claim compensation and claim-management intervention.
A management science team was formed to study this problem by applying decision
analysis. For each of numerous categories of injury claims based on the nature of the
injury, the gender and age of the worker, and so on, a decision tree was used to evaluate FIGURE 9.5
whether that category should be classified as low risk (not requiring intervention) or high The Decision Tree dialog box used to specify that the second node of the first Goferbroke
risk (requiring intervention), depending on the severity of the injury. For each category, a problem is an event node with two branches, Oil and Dry, with values (partial payoffs) of 800
calculation was made of the cutoff point on the critical number of short-term disability and 0, and with probabilities of 0.25 and 0.75, respectively.
 Max: I like this decision tree thing. It puts everything into perspective.
Jennifer: Good.
Max: But one thing still really bothers me.
Jennifer: I think I can guess.
Max: Yes. I’ve made it pretty plain that I don’t want to make my decision based
on believing the consulting geologist’s numbers. One chance in four of oil. Hah!
It’s just an educated guess.
Jennifer: Well, let me ask this. What is the key factor in deciding whether to
drill for oil or sell the land?
Max: How likely it is that there is oil there.
Jennifer: Doesn’t the consulting geologist help in determining this?
FIGURE 9.6
The decision tree constructed and solved by Analytic Solver for the first Goferbroke Co.
Max: Definitely. I hardly ever drill without his input.
problem as presented in Table 9.3, where the 1 in cell B9 indicates that the top branch (the Jennifer: So shouldn’t your criterion for deciding whether to drill be based
Drill alternative) should be chosen. directly on this input?
Max: Yes, it should.
page 345 Jennifer: But then I don’t understand why you keep objecting to using the
Analytic Solver identifies the optimal policy for the current decision tree according to
consulting geologist’s numbers.
Bayes’ decision rule. Max: I’m not objecting to using his input. This input is vital to my decision.
What I object to is using his numbers, one chance in four of oil, as being the
At each stage in constructing a decision tree, Analytic Solver automatically gospel truth. That is what this Bayes’ decision rule seems to do. We both saw
solves for the optimal policy with the current tree when using Bayes’ decision what a close decision this was, 100 versus 90. What happens if his numbers are
rule. The number inside each decision node indicates which branch should be off some, as they probably are? This is too important a decision to be based on
chosen (assuming the branches emanating from that node are numbered some numbers that are largely pulled out of the air.
consecutively from top to bottom). Thus, for the final decision tree in Figure Jennifer: OK, I see. Now he says that there is one chance in four of oil, a 25
9.6, the number 1 in cell B9 specifies that the first branch (the Drill alternative) percent chance. Do you think that is the right ballpark at least? If not 25 percent,
should be chosen. The number on both sides of each terminal node is the payoff how much lower might it be? Or how much higher?
if that node is reached. The number 100 in cells A10 and E6 is the expected Max: I usually add and subtract 10 percent from whatever the consulting
payoff (the measure of performance for Bayes’ decision rule) at those stages in geologist says. So I suppose the chance of oil is likely to be somewhere
the process. between 15 percent and 35 percent.
This description of the decision analysis tools of Analytic Solver may seem Jennifer: Good. Now we’re getting somewhere. I think I know exactly what
somewhat complicated. However, we think that you will find the procedure we should do next.
quite intuitive when you execute it on a computer. If you spend considerable Max: What’s that?
time with Analytic Solver, you also will find that it has many helpful features
Jennifer: There is a management science technique that is designed for just this
that haven’t been described in this brief introduction.
kind of situation. It is called sensitivity analysis. It will allow us to investigate
Max’s Reaction what happens if the consulting geologist’s numbers are off.
Max: Great! Let’s do it. This greatly simplifies sensitivity analysis. To change a piece of data, it needs
to be changed in only one place rather than searching through the entire tree to
Review Questions find and change all occurrences of that piece of data.1 A second advantage of
consolidating the data and results is that it makes it easy for anyone to interpret
1. What is a decision tree?
the model. It is not necessary to understand Analytic Solver or how to read a
2. What is a decision node in a decision tree? An event node?
decision tree in order to see what data were used in the model or what the
3. What symbols are used to represent decision nodes and event nodes?
suggested plan of action and expected payoff are.
The sum of the two prior probabilities must equal one, so increasing one of
page 346 these probabilities automatically decreases the other one by the same amount,
and vice versa. This is enforced on the decision tree in Figure 9.7 by the
equation in cell H6—the probability of a dry site = H6 = 1 − ProbabilityOfOil
9.4 SENSITIVITY ANALYSIS WITH DECISION (E22). Max has concluded that the true chances of having oil on the tract of land
TREES are likely to lie somewhere between 15 and 35 percent. In other words, the true
prior probability of having oil is likely to be in the range from 0.15 to 0.35, so
Sensitivity analysis (an important type of what-if analysis introduced in the corresponding prior probability of the land being dry would range from
Section 5.1) commonly is used with various applications of management 0.85 to 0.65.
science to study the effect if some of the numbers included in the mathematical We can begin sensitivity analysis by simply trying different trial values for
model are not correct. In this case, the mathematical model is represented by the prior probability of oil. This is done in Figure 9.8, first with this
the decision tree shown in Figure 9.6. The numbers in this tree that are most probability at the lower end of the range (0.15) and next with this probability at
questionable are the prior probabilities in cells H1 and H6, so we will initially the upper end (0.35). When the prior probability of oil is only 0.15, the
focus the sensitivity analysis on these numbers. decision swings over to selling the land by a wide margin (an expected payoff
It is helpful to start this process by consolidating the data and results on the of 90 versus only 20 for drilling). However, when this probability is 0.35, the
spreadsheet below the decision tree, as in Figure 9.7. As indicated by the decision is to drill by a wide margin (expected payoff = 180 versus only 90 for
formulas at the bottom of the figure, the cells giving the results (E24 and E26) selling). Thus, the decision is very sensitive to the prior probability of oil. This
make reference to the corresponding output cells on the decision tree (B9 and sensitivity analysis has revealed that it is important to do more, if possible, to
A10). Similarly, the data cells on the decision tree (D6, D14, H1, H4, and H9) pin down just what the true value of the probability of oil is.
now reference the corresponding data cells below the tree (E18:E22).
Consequently, the user can experiment with various alternative values in the Using a Data Table to Do Sensitivity Analysis
data cells below and the results will simultaneously change in both the decision Systematically
tree and the results section below the tree to reflect the new data. To pin down just where the suggested course of action changes, we could
Excel Tip: Consolidating the data and results on the spreadsheet makes it easier to do
continue selecting new trial values of the prior probability of oil at random.
sensitivity analysis and also makes the model and results easier to interpret. However, a better approach is to systematically consider a range of values. A
feature built into Excel, called a data table, is designed to perform just this sort
Consolidating the data and results offers a couple of advantages. First, it of analysis. Data tables are used to show the results of certain output cells for
assures that each piece of data is in only one place. Each time that piece of data various trial values of a data cell.
is needed in the decision tree, a reference is made to the single data cell below.
 A data table displays the results of selected output cells for various trial values of a data
cell.

To use data tables, first make a table on the spreadsheet with headings as
shown in columns I, J, and K in Figure 9.9. In the first column of the table
(I19:I29), list the trial values for the data cell (the prior probability of oil),
except leave the first row blank. The headings of the next columns specify
which output will be evaluated. For each of these columns, use the first row of
the table (cells J18:K18) to write an equation that refers to the relevant output
cell. In this case, the cells of interest are Action (E24) and ExpectedPayoff
(E26), so the equations for J18:K18 are those shown below the spreadsheet in
Figure 9.9.

page 347

FIGURE 9.7
In preparation for performing sensitivity analysis on the first Goferbroke Co. problem, the
data and results have been consolidated on the spreadsheet below the decision tree.
 Next, select the entire table (I18:K29) and then choose Data Table from the
What-If Analysis menu of the Data tab. In the Data Table dialog box (as shown
at the bottom left of Figure 9.9), indicate the column input cell (E22), which
refers to the data cell that is being changed in the first column of the table.
Nothing is entered for the row input cell because no row is being used to list
the trial values of a data cell in this case.

page 348
 FIGURE 9.8 error (or algebra) can be used to pin this number down more precisely. It turns
Performing sensitivity analysis for the first Goferbroke Co. problem by trying alternative out to be 0.2375.
values (0.15 and 0.35) of the prior probability of oil.
For a problem with more than two possible states of nature, the most
straightforward approach is to focus the sensitivity analysis on only two states
page 349 at a time as described above. This again would involve investigating what
happens when the prior probability of one state increases as the prior
probability of the other state decreases by the same amount, holding fixed the
prior probabilities of the remaining states. This procedure then can be repeated
for as many other pairs of states as desired.

FIGURE 9.10
After the preparation displayed in Figure 9.9, clicking OK generates this data table that shows
the optimal action and expected payoff for various trial values of the prior probability of oil.

FIGURE 9.9
Expansion of the spreadsheet in Figure 9.7 to prepare for generating a data table, where the page 350
choice of E22 for the column input cell in the Data Table dialog box indicates that this is the
data cell that is being changed in the first column of the data table. Max’s Reaction
Clicking OK then generates the data table shown in Figure 9.10. For each Max: That data table paints a pretty clear picture. I think I’m getting a much
trial value for the data cell listed in the first column of the table, the better handle on the problem.
corresponding output cell values are calculated and displayed in the other Jennifer: Good.
columns of the table. (The entries in the first row of the table come from the Max: Less than a 23¾ percent chance of oil, I should sell. If it’s more, I should
original solution in the spreadsheet.) drill. It confirms what I suspected all along. This is a close decision, and it all
Figure 9.10 reveals that the best course of action switches from Sell to boils down to picking the right number for the chances of oil. I sure wish I had
Drill for a prior probability of oil somewhere between 0.23 and 0.25. Trial and more to go on than the consulting geologist’s numbers.
Jennifer: You talked earlier about the possibility of paying $30,000 to get a 4. What conclusion was drawn for the first Goferbroke Co. problem regarding how the decision
should depend on the prior probability of oil?
detailed seismic survey of the land.
Max: Yes, I might have to do that. But 30,000 bucks! I’m still not sure that it’s
worth that much dough.
9.5 CHECKING WHETHER TO OBTAIN MORE
The next section describes how to find and use the expected value of perfect information.
INFORMATION
Jennifer: I have a quick way of checking that. It’s another technique I learned
Prior probabilities may provide somewhat inaccurate estimates of the true
in my management science course. It’s called finding the expected value of
probabilities of the states of nature. Might it be worthwhile for Max to spend
perfect information (EVPI). The expected value of perfect information is some money for a seismic survey to obtain better estimates? The quickest way
the increase in the expected payoff you would get if the seismic survey could to check this is to pretend that it is possible for the same amount of money to
tell you for sure if there is oil there. actually determine which state is the true state of nature (“perfect information”)
Max: But it can’t tell you for sure. and then determine whether obtaining this information would make this
Jennifer: Yes, I know. But finding out for sure if oil is there is what we refer to expenditure worthwhile. If having perfect information would not be
as perfect information. So the increase in the expected payoff if you find out for worthwhile, then it definitely would not be worthwhile to spend this money just
sure is the expected value of perfect information. We know that’s better than to learn more about the probabilities of the states of nature.
you actually can do with a seismic survey.
Max: Right. Definitely identifying the true state of nature is referred to as perfect information. This
represents the best outcome of seeking more information.
Jennifer: OK, suppose we find that the expected value of perfect information is
less than $30,000. Since that is better than we can do with a seismic survey,
that tells us right off the bat that it wouldn’t pay to do the seismic survey. page 351
Max: OK, I get it. But what if this expected value of perfect information is The key quantities for performing this analysis are
more than $30,000?
Jennifer: Then you don’t know for sure whether the seismic survey is worth it
until you do some more analysis. This analysis takes some time, whereas it is
very quick to calculate the expected value of perfect information. So it is well
worth simply checking whether the expected value of perfect information is less
than $30,000 and, if so, saving a lot of additional work.
Max: OK. Let’s do it.

Review Questions The expected value of perfect information is calculated as

1. Why might it be helpful to use sensitivity analysis with Bayes’ decision rule?
EVPI = EP ( with perfect info ) − EP ( without more info )
2. When preparing to perform sensitivity analysis, what are a couple of advantages of
consolidating the data and results on the spreadsheet that contains the decision tree?
After calculating EP (with perfect info) and then EVPI, the last step is to
3. What is shown by a data table when it is used to perform sensitivity analysis?
compare EVPI with C.
 If C > EVPI, then it is not worthwhile to obtain more information. Max: So you’re telling me that if the seismic survey could really be definitive
If C ≤ EVPI, then it might be worthwhile to obtain more information. in determining whether oil is there, doing the survey would increase my
To calculate EP (with perfect info), we pretend that the decision can be average payoff by about $142,500?
made after learning the true state of nature. Given the true state of nature, we Jennifer: That’s right.
then would automatically choose the alternative with the maximum payoff for Max: So after subtracting the $30,000 cost of the survey, I would be ahead
that state. Thus, we drill if we know there is oil, whereas we sell if we know $112,500. Well, too bad the surveys aren’t that good. In fact, they’re not all that
the site is dry. The prior probabilities still give our best estimate of the reliable.
probability that each state of nature will turn out to be the true one. EP (with Jennifer: Tell me more. How reliable are they?
perfect info) is therefore the weighted average of the maximum payoff for each
state, multiplying each maximum payoff by the prior probability of the page 352
corresponding state of nature. Thus,

Analytic Solver also can be used to calculate EP (with perfect info) by

constructing and solving the decision tree shown in Figure 9.11. The clever
idea here is to start the decision tree with an event node whose branches are
the various states of nature (oil and dry in this case). Since a decision node
follows each of these branches, the decision is being made with perfect
information about the true state of nature. Therefore, the expected payoff of
242.5 obtained by Analytic Solver in cell A11 is the expected payoff with
perfect information.

Starting the decision tree with an event node whose branches are the various states of
nature corresponds to starting with perfect information about the true state of nature.

Since EP (with perfect info) = 242½, we now can calculate the expected
value of perfect information as FIGURE 9.11
By starting with an event node involving the states of nature, Analytic Solver uses this
decision tree to obtain the expected payoff with perfect information for the first Goferbroke
Co. problem.

Max: Well, they come back with seismic soundings. If the seismic soundings
Conclusion:EVPI > C, since 142.5 > 30. Therefore, it might be worthwhile
are favorable, then oil is fairly likely. If they are unfavorable, then oil is pretty
to do the seismic survey.
unlikely. But you can’t tell for sure.
Max’s Reaction Jennifer: OK. Suppose oil is there. How often would you get favorable
seismic soundings?
Max: I can’t give you an exact number. Maybe 60 percent. 2. How can the expected payoff with perfect information be calculated from the payoff table?

Jennifer: OK, good. Now suppose that the land is dry. How often would you 3. How should a decision tree be constructed to obtain the expected payoff with perfect
information by solving the tree?
still get favorable seismic soundings?
4. What is the formula for calculating the expected value of perfect information?
Max: Too often! I’ve lost a lot of money drilling when the seismic survey said 5. What is the conclusion if the cost of obtaining more information is more than the expected
to and then nothing was there. That’s why I don’t like to spend the 30,000 value of perfect information?
bucks. 6. What is the conclusion if the cost of obtaining more information is less than the expected
Jennifer: Sure. So it tells you to drill when you shouldn’t close to half the value of perfect information?
time? 7. Which of these two cases occurs in the Goferbroke Co. problem?

Max: No. It’s not that bad. But fairly often.

Jennifer: Can you give me a percentage?
Max: OK. Maybe 20 percent. 9.6 USING NEW INFORMATION TO UPDATE
Jennifer: Good. Thanks. Now I think we can do some analysis to determine THE PROBABILITIES
whether it is really worthwhile to do the seismic survey.
Max: How do you do the analysis? The prior probabilities of the possible states of nature often are quite
Jennifer: Well, I’ll describe the process in detail pretty soon. But here is the subjective in nature, so they may be only very rough estimates of the true
general idea. We’ll do some calculations to determine what the chances of oil probabilities. Fortunately, it frequently is possible to do some additional testing
would be if the seismic soundings turn out to be favorable. Then we’ll calculate or surveying (at some expense) to improve these estimates. These improved
the chances if the soundings are unfavorable. We called the consulting estimates are called posterior probabilities.
geologist’s numbers prior probabilities because they were prior to obtaining In the case of the Goferbroke Co., these improved estimates can be obtained
more information. The improved numbers are referred to as posterior at a cost of $30,000 by conducting a detailed seismic survey of the land. The
probabilities. possible findings from such a survey are summarized below.

Posterior probabilities are the revised probabilities of the states of nature after doing a test Possible Findings from a Seismic Survey
or survey to improve the prior probabilities. FSS: Favorable seismic soundings; oil is fairly likely.
USS: Unfavorable seismic soundings; oil is quite unlikely.
Max: OK.
Jennifer: Then we’ll use these posterior probabilities to determine the average To use either finding to calculate the posterior probability of oil (or of being
payoff, after subtracting the $30,000 cost, if we do the seismic survey. If this dry), it is necessary to estimate the probability of obtaining this finding for each
payoff is better than we would do without the seismic survey, then we should state of nature. During the conversation at the end of the preceding section,
do it. Otherwise, not. Jennifer elicited these estimates from Max, as summarized in Table 9.7. (Max
Max: That makes sense. actually only estimated the probability of favorable seismic soundings, but
subtracting this number from one gives the probability of unfavorable seismic
page 353 soundings.) The notation used in the table for each of these estimated
probabilities is
Review Questions
1. What is meant by perfect information regarding the states of nature?
This kind of probability is referred to as a conditional probability, because it calculations in this probability tree diagram are discussed in the next few
is conditioned on being given the state of nature. paragraphs.)
Recall that the prior probabilities are
Each joint probability in the third column of the probability tree diagram is the product of
P ( Oil ) = 0.25 the probabilities in the first two columns.

P ( Dry ) = 0.75 Having found each joint probability of both a particular state of nature and a
particular finding from the seismic survey, the next step is to use these
The next step is to use these probabilities and the probabilities in Table 9.7 to
probabilities to find each probability of just a particular finding without
obtain a combined probability called a joint probability. Each combination of a
specifying the state of nature. Since any finding can be obtained with any state
state of nature and a finding from the seismic survey will have a joint
of nature, the formula for calculating the probability of just a particular finding
probability that is determined by the following formula.
is
P ( state and finding ) = P ( state ) P ( finding | state )
P(finding) = P(Oil and finding) + P(Dry and finding)
TABLE 9.7 The probability of a finding is the sum of the corresponding joint probabilities in the third
Probabilities of the Possible Findings from the Seismic Survey, Given the State of Nature, for the column of the probability tree diagram.
Goferbroke Co. Problem

P(finding | state) For example, the probability of a favorable finding (FSS) is

State of Nature Favorable (FSS) Unfavorable (USS)
Oil P(FSS | Oil) = 0.6 P(USS | Oil) = 0.4
Dry P(FSS | Dry) = 0.2 P(USS | Dry) = 0.8

where the two joint probabilities on the right-hand side of this equation are
page 354 found on the first and third branches of the third column of the probability tree
diagram. The calculation of both P (FSS) and P (USS) is shown underneath the
For example, the joint probability that the state of nature is Oil and the finding diagram. (These are referred to as unconditional probabilities to differentiate
from the seismic survey is favorable (FSS) is them from the conditional probabilities of a finding given the state of nature,
shown in the second column.)

The calculation of all these joint probabilities is shown in the third column
of the probability tree diagram given in Figure 9.12. The case involved is
identified underneath each branch of the tree and the probability is given over
the branch. The first column gives the prior probabilities and then the
probabilities from Table 9.7 are shown in the second column. Multiplying each
probability in the first column by a probability in the second column gives the
corresponding joint probability in the third column. (The subsequent
 The arrows in the probability tree diagram show where the numbers come from for
calculating the posterior probabilities.

The fourth column of the probability tree diagram shows the calculation of
all the posterior probabilities. The arrows indicate how each numerator comes
from the corresponding joint probability in the third column and the
denominator comes from the corresponding unconditional probability below the
diagram.
By using the formulas given earlier for the joint probabilities and
unconditional probabilities, each posterior probability also can be calculated
directly from the prior probabilities (first column) and the conditional
probabilities (second column) as follows.

For example, the posterior probability of oil, given a favorable finding (FSS),
is
FIGURE 9.12
Probability tree diagram for the Goferbroke Co. problem showing all the probabilities leading
to the calculation of each posterior probability of the state of nature given the finding of the
seismic survey.

page 355
Finally, we now are ready to calculate each posterior probability of a This formula for a posterior probability is known as Bayes’ theorem, in
particular state of nature given a particular finding from the seismic survey. The honor of its discovery by the Reverend Bayes. The clever Reverend Bayes
formula involves combining the joint probabilities in the third column with the found that any posterior probability can be found in this way for any decision
unconditional probabilities underneath the diagram as follows. analysis problem, regardless of how many states of nature it has. The
denominator in the formula would contain one such term for each of the states
of nature. Note that the probability tree diagram also is applying Bayes’
theorem, but in smaller steps rather than a single long formula.
Table 9.8 summarizes all the posterior probabilities calculated in Figure
For example, the posterior probability that the true state of nature is oil, given a 9.12.
favorable finding (FSS) from the seismic survey, is
 After you learn the logic of calculating posterior probabilities, we suggest
that you use the computer to perform these rather lengthy calculations. We have
provided an Excel template (labeled Posterior Probabilities) for this purpose
in this chapter’s Excel files at www.mhhe.com/ Hillier6e. Figure 9.13
illustrates the use of this template for the Goferbroke Co. problem. All you do
is enter the prior probabilities and the conditional probabilities from the first
two columns of Figure 9.12 into the top half of the template (rows 6 and 7 in
this case). The posterior probabilities then immediately appear in the bottom
half, as shown in D15:E16 in this case. (The equations entered into the cells in
columns E through H are similar to those for column D shown at the bottom of
the figure.)

TABLE 9.8
Posterior Probabilities of the States of Nature, Given the Finding from the Seismic Survey, for
the Goferbroke Co. Problem

P(finding | state)
Finding Oil Dry
Favorable (FSS) P(Oil | FSS) = 1/2 P(Dry | FSS) = 1/2
Unfavorable (USS) P(Oil | USS) = 1/7 P(Dry | USS) = 6/7

page 356

FIGURE 9.13
The Posterior Probabilities template available at www.mhhe.com/ Hillier6e enables efficient
calculation of posterior probabilities, as illustrated here for the Goferbroke Co. problem.

Max’s Reaction
Max: So this is saying that even with favorable seismic soundings, I still only
have one chance in two of finding oil. No wonder I’ve been disappointed so
often in the past when I’ve drilled after receiving a favorable seismic survey. I
thought those surveys were supposed to be more reliable than that. So now I’m
even more unenthusiastic about paying 30,000 bucks to get a survey done.
Jennifer: But one chance in two of oil. Those are good odds.
Max: Yes, they are. But I’m likely to lay out 30,000 bucks and then just get an
unfavorable survey back.
Jennifer: My calculations indicate that you have about a 70 percent chance of DECISIONS
that happening.
Max: See what I mean? We now turn our attention to analyzing the full Goferbroke Co. problem with
Jennifer: But even an unfavorable survey tells you a lot. Just one chance in the help of a decision tree. For the full problem, there is a sequence of two
seven of oil then. That might rule out drilling. So a seismic survey really does decisions to be made. First, should a seismic survey be conducted? Second,
pin down the odds of oil a lot better. Either one chance in two or one chance in after obtaining the results of the seismic survey (if it is conducted), should the
seven instead of the ballpark estimate of one chance in four from the consulting company drill for oil or sell the land?
geologist. Applications of decision analysis commonly are similar to the full
Goferbroke Co. problem in the sense that they require making a sequence of
page 357 decisions. In some cases, a sequence of even more than two decisions need to
be made. Decision trees provide a convenient way of organizing and analyzing
Max: Yes, I suppose that’s right. I really would like to improve the consulting such problems. (Other powerful techniques for sequential decision making also
geologist’s numbers. It sounds like you’re recommending that we do the seismic are available, but are beyond the scope of this book.)
survey. As described in Section 9.3, a decision tree provides a graphical display
Jennifer: Well, actually, I’m not quite sure yet. What we’ll do is sketch out a of the progression of decisions and random events for the problem. Figure 9.2
decision tree, showing the decision on whether to do the seismic survey and in that section shows the decision tree for the first Goferbroke problem where
then the decision on whether to drill or sell. Then we’ll work out the average the only decision under consideration is whether to drill for oil or sell the land.
payoffs for these decisions on the decision tree. Figure 9.6 then shows the same decision tree as it would be constructed and
Max: OK, let’s do it. I want to make a decision soon. solved with Analytic Solver. We now will focus on expanding this decision
tree to consider the sequence of decisions needed for the full Goferbroke Co.
Review Questions problem.
1. What are posterior probabilities of the states of nature?
2. What are the possible findings from a seismic survey for the Goferbroke Co.? Constructing the Decision Tree
3. What probabilities need to be estimated in addition to prior probabilities in order to begin Now that a prior decision needs to be made on whether to conduct a seismic
calculating posterior probabilities? survey, the original decision tree needs to be expanded as shown in Figure 9.14
4. What five kinds of probabilities are considered in a probability tree diagram? (before including any numbers). Recall that each square in the tree represents a
5. What is the formula for calculating P(state and finding)? decision node, where a decision needs to be made, and each circle represents
6. What is the formula for calculating P(finding)? an event node, where a random event will occur.
7. What is the formula for calculating a posterior probability, P(state | finding), from P(state and Thus, the first decision (should we have a seismic survey done?) is
finding) and P(finding)?
represented by decision node a in Figure 9.14. The two branches leading out of
8. What is the name of the famous theorem for how to calculate posterior probabilities?
this node correspond to the two alternatives for this decision. Node b is an
event node representing the random event of the outcome of the seismic survey.
The two branches emanating from node b represent the two possible outcomes
9.7 USING A DECISION TREE TO ANALYZE of the survey. Next comes the second decision (nodes c, d, and e) with its two
THE PROBLEM WITH A SEQUENCE OF possible choices. If the decision is to drill for oil, then we come to another
event node (nodes f, g, and h), where its two branches correspond to the two occurring from this node has been inserted in parentheses along this branch.
possible states of nature. From event node h, the probabilities are the prior probabilities of these states
The next step is to insert numbers into the decision tree as shown in Figure of nature, since no seismic survey has been conducted to obtain more
9.15. The numbers under or over the branches that are not in parentheses are information in this case. However, event nodes f and g lead out of a decision to
the cash flows (in thousands of dollars) that occur at those branches. For each do the seismic survey (and then to drill). Therefore, the probabilities from these
path through the tree from node a to a final branch, these same numbers then are event nodes are the posterior probabilities of the states of nature, given the
added to obtain the resulting total payoff shown in boldface to the right of that outcome of the seismic survey, where these numbers are obtained from Table
branch. To differentiate payoffs from costs and probabilities, note that these 9.8 or from cells D15:E16 in Figure 9.13.
payoffs on the right side of Figure 9.15 are the only numbers in the decision
tree that are shown in boldface.

page 358

FIGURE 9.15
The decision tree in Figure 9.14 after adding both the probabilities of random events and the
payoffs.
FIGURE 9.14
The decision tree for the full Goferbroke Co. problem (before including any numbers) when
first deciding whether to conduct a seismic survey.
page 359

The numbers in parentheses are probabilities.

The last set of numbers to be inserted into the decision tree is the
probabilities of random events. In particular, since each branch emanating from
an event node represents a possible random event, the probability of this event
An Application Vignette
 Finally, we have the two branches emanating from event node b. The
Polio is a serious infectious disease that can lead to a permanent muscle weakness numbers here are the probabilities of these findings from the seismic survey,
(especially in the legs) or even to death. Children are especially vulnerable, but it also Favorable (FSS) or Unfavorable (USS), as given underneath the probability
can strike adults as well. There is no cure, so this was a particularly dreaded disease
tree diagram in Figure 9.12 or in cells C15:C16 of Figure 9.13.
throughout the first half of the 20th century. Fortunately, relatively effective polio vaccines
finally were developed in the 1950s, although they still failed to develop immunity for a
small percentage of recipients. Performing the Analysis
However, even with extensive vaccination and revaccination programs, some polio
outbreaks continued to occur. The polio virus is transmitted through person-to-person Having constructed the decision tree, including its numbers, we now are ready
contact from an infected person to a susceptible person. Such contact is impossible to to analyze the problem by using the following procedure.
avoid since some infected people show no symptoms and some susceptible people
may have been vaccinated in the past. Consequently, approximately 350,000 paralytic
The expected payoff needs to be calculated for each event node.
polio cases were reported worldwide as late as the year 1988.
Therefore, the Global Polio Eradication Initiative was begun in 1988 to eventually
eradicate polio completely from the face of the earth. One of the spearheading partners . Start at the right side of the decision tree and move left one column at a time.
of this campaign continues to be the U.S. Centers for Disease Control and For each column, perform either step 2 or step 3 depending on whether the
Prevention (CDC).
Considerable progress was made during the early years of this campaign. However,
nodes in that column are event nodes or decision nodes.
it then became clear that intensified efforts would be needed to complete the job in . For each event node, calculate its expected payoff by multiplying the expected
difficult parts of the world where civil war, internal strife, and hostility to outsiders greatly
limited what health workers could do there. In light of these challenges, it had become
payoff of each branch (shown in boldface to the right of the branch) by the
more difficult than ever to make the best possible sequence of interrelated decisions probability of that branch and then summing these products. Record this
(when and where to perform routine immunizations, supplemental immunizations, expected payoff for each event node in boldface next to the node, and designate
responses to outbreaks, stockpiling of vaccines, surveillance of possible infected areas,
etc.) as needed. Therefore, throughout the early part of the 21st century, CDC
this quantity as also being the expected payoff for the branch leading to this
management has relied on a variety of management science tools to identify how to node. (Note that payoffs and these expected payoffs are the only numbers being
make the best possible use of their limited resources. The most important of these recorded in boldface in the decision tree.)
management science techniques has been decision analysis. Very large decision trees
have been repeatedly formulated and solved to identify optimal sequences of interrelated . For each decision node, compare the expected payoffs of its branches and
decisions. choose the alternative whose branch has the largest expected payoff. In each
This approach has been a very major factor in the continuing success of the polio case, record the choice on the decision tree.
eradication campaign. The number of annual diagnosed cases worldwide went from the
hundreds of thousands at the beginning to less than 100 in 2015. The campaign has
realized an estimated $40–$50 billion in net benefits for the countries still working on page 360
complete eradication while protecting the significantly larger net benefits enjoyed by the
countries that have stopped any transmission of polio. Because of the vital contribution
of management science to this enormously important campaign, the U.S. Centers for To begin the procedure, consider the rightmost column of nodes, namely,
Disease Control and Prevention was awarded the prestigious first prize in the 2014 event nodes f, g, and h. Applying step 2, their expected payoffs (EP) are
Franz Edelman Award for Achievement in Operations Research and the Management calculated as
Sciences.

Source: K. M. Thompson, R. J. Duintjer Tebbens, M.A. Pallansch, S. G. F. Wassilak,

and S.L. Cochi, “Polio Eradicators Use Integrated Analytical Models to Make Better
Decisions,” Interfaces 45, no. 1 (January-February 2015), pp. 5–25. (A link to this article
is provided at www.mhhe.com/Hillier6e.)
These expected payoffs then are placed above these nodes, as shown in Figure
9.16. (The next few paragraphs will describe the analysis leading to additional
results displayed in this figure.)
Next, we move one column to the left, which consists of decision nodes c,
d, and e. The expected payoff for a branch that leads to an event node now is
recorded in boldface over that event node. Therefore, step 3 can be applied as
follows:

A double dash indicates a rejected decision.

The expected payoff for each chosen alternative now would be recorded in FIGURE 9.16
boldface over its decision node, as shown in Figure 9.16. The chosen The final decision tree that records the analysis for the full Goferbroke Co. problem when
using monetary payoffs.
alternative also is indicated by inserting a double dash as a barrier through
each rejected branch.
page 361
Next, moving one more column to the left brings us to node b. Since this is
an event node, step 2 of the procedure needs to be applied. The expected payoff
for each of its branches is recorded over the following decision node.
Therefore, the expected payoff is
EP = 0.7(60) + 0.3(270) = 123 for node b
as recorded over this node in Figure 9.16.
Finally, we move left to node a, a decision node. Applying step 3 yields
Node Do seismic survey has EP = 123
a: No seismic survey has EP = 100
123 > 100, so choose Do seismic survey.
This expected payoff of 123 now would be recorded over the node, and a
double dash inserted to indicate the rejected branch, as already shown in Figure
9.16. done
EP (without more = Expected payoff when the survey is
The open paths (no double dashes) provide the optimal decision at each decision node.
info) not done
This procedure has moved from right to left for analysis purposes. where Bayes’ decision rule is applied to find both quantities. EP (with more
However, having completed the decision tree in this way, the decision maker info) can be obtained from the decision tree in Figure 9.16 by referring to the
now can read the tree from left to right to see the actual progression of events. expected payoff of 123 for the branch leading from node a to node b (Do
The double dashes have closed off the undesirable paths. Therefore, given the seismic survey) except that the (unknown) cost of the seismic survey should not
payoffs for the final outcomes shown on the right side, Bayes’ decision rule be included, so 30 (the originally assumed cost of the survey) should be added
says to follow only the open paths from left to right to achieve the largest to this expected payoff. Therefore,
possible expected payoff.
Following the open paths from left to right in Figure 9.16 yields the EP (with more info) = 123 + 30 = 153
following optimal policy, according to Bayes’ decision rule. EP (without more info) is described at the beginning of Section 9.5 as the
Optimal Policy expected payoff from applying Bayes’ decision rule with the original prior
Do the seismic survey. probabilities (so not using the seismic survey), where this expected payoff was
If the result is unfavorable, sell the land. found to be 100 in Figure 9.1 (and in cell A10 of Figure 9.7). So
If the result is favorable, drill for oil. EP (without more info) = 100
The expected payoff (including the cost of the seismic survey) is 123
($123,000). page 362
Expected Value of Sample Information Now we can calculate the expected value of sample information (EVSI)
We have assumed so far that the cost of the seismic survey for the full as
Goferbroke Co. problem is known in advance to be $30,000. However,
suppose that there is uncertainty about this cost. How would this change the
analysis described above?
This uncertainty would raise the new key question of how much value
would be added by performing the seismic survey. In other words, how large Let
can the cost of the seismic survey be before it would no longer be worthwhile C = Best available estimate of the cost of the seismic survey (in thousands
to perform this survey. This quantity is referred to as the expected value of of dollars)
sample information (or EVSI for short), where “sample information” in this The final step in the analysis is to compare C and EVSI.
case refers to the information from the seismic survey. If C < EVSI, then perform the seismic survey.
The analysis needed to determine EVSI would begin by identifying two key If C ≥ EVSI, then do not perform the seismic survey.
quantities,
Using Analytic Solver
EP (with more info) = Expected payoff (excluding the cost
of the survey) when the survey is
Using the procedures described in Section 9.3, the Decision Tree tools of 5. When performing the analysis, where do you begin on the decision tree and in which direction
do you move for dealing with the nodes?
Analytic Solver can be used to construct and solve the decision tree in Figure
6. What calculation needs to be performed at each event node?
9.16 on a spreadsheet. Figure 9.17 shows the decision tree obtained with
7. What comparison needs to be made at each decision node?
Analytic Solver. Although the form is somewhat different, note that this
8. What is meant by the expected value of sample information and how might it be used?
decision tree is completely equivalent to the one in Figure 9.16. Besides the
convenience of constructing the tree directly on a spreadsheet, Analytic Solver
also provides the key advantage of automatically solving the decision tree. page 363
Rather than relying on hand calculations as in Figure 9.16, Analytic Solver
instantaneously calculates all the expected payoffs at each stage of the tree, as
shown below and to the left of each node, as soon as the decision tree is
constructed. Instead of using double dashes, Analytic Solver puts a number
inside each decision node indicating which branch should be chosen (assuming
the branches emanating from that node are numbered consecutively from top to
bottom).

Max’s Reaction
Max: I see that this decision tree gives me some numbers to compare
alternatives. But how reliable are those numbers?
Jennifer: Well, you have to remember that these average payoffs for the
alternatives at the decision nodes are based on both the payoffs on the right and
the probabilities at the event nodes. These probabilities are based in turn on the
consulting geologist’s numbers and the numbers you gave me on how frequently
you get favorable seismic soundings when you have oil or when the land is dry.
Max: That doesn’t sound so good. You know what I think about the consulting
geologist’s numbers. And the numbers I gave you were pretty rough estimates.
Jennifer: True. So the average payoffs shown in the decision tree are only
approximations. This is when some sensitivity analysis can be helpful, like we
did earlier before we considered doing the seismic survey.
Max: OK. So let’s do it.

Review Questions
1. What does a decision tree display?
2. What is happening at a decision node?
3. What is happening at an event node?
4. What kinds of numbers need to be inserted into a decision tree before beginning the
analysis?
 9.8 PERFORMING SENSITIVITY ANALYSIS
ON THE PROBLEM WITH A SEQUENCE
OF DECISIONS
Section 9.4 describes how the decision tree created with Analytic Solver
(Figures 9.6 and Figures 9.7) was used to perform sensitivity analysis for the
first Goferbroke problem where the only decision being made was whether to
drill for oil or sell the land (without conducting the seismic survey). The focus
was on one particularly critical piece of data, the prior probability of oil, so
the analysis involved checking whether the decision would change if the
original value of this prior probability (0.25) were changed to various other
trial values. New trial values first were considered in a trial-and-error manner
(Figure 9.8) and then were investigated more systematically by constructing a
data table (Figure 9.10).
Since Max Flyer wants to consider whether to have a seismic survey
conducted before deciding whether to drill or sell, the relevant decision tree
now is the one in Figure 9.17 instead of the one in Figure 9.6. With this
sequence of decisions and the resulting need to obtain and apply posterior
probabilities, conducting sensitivity analysis becomes somewhat more
involved. Let’s see how it is done.

Organizing the Spreadsheet

As was done in Section 9.4, it is helpful to begin by consolidating the data and
results into one section of the spreadsheet, as shown in Figure 9.18. The data
cells in the decision tree now make reference to the consolidated data cells to
the right of the decision tree (cells V4:V11). Similarly, the summarized results
to the right of the decision tree make reference to the output cells within the
decision tree (the decision nodes in cells B29, F41, J11, and J26, as well as the
expected payoff in cell A30).
FIGURE 9.17
Consolidating the data and results on the spreadsheet is important for sensitivity analysis.
The decision tree constructed and solved by Analytic Solver for the full Goferbroke Co.
problem that also considers whether to do a seismic survey.
The probability data in the decision tree are complicated by the fact that the
page 364 posterior probabilities will need to be updated any time a change is made in
any of the prior probability data. Fortunately, the template for calculating
posterior probabilities (as shown in Figure 9.13) can be used to do these To systematically determine how the decisions and expected payoffs change as
calculations. The relevant portion of this template (B3:H19) has been copied the prior probability of oil (or any other data) changes, a data table can be
(using the Copy and Paste commands in the Edit menu) to the spreadsheet in generated with Excel by using the same procedure described in Section 9.4.
Figure 9.18 (now appearing in U30:AA46). The data for the template refer to First make a table on the spreadsheet with headings as shown in columns Y
the probability data in the data cells PriorProbabilityOfOil (V9), through AD in Figure 9.19. In the first column of the table (Y5:Y15), list the
ProbFSSGivenOil (V10), and ProbUSSGivenDry (V11), as shown in the trial values for the data cell (the prior probability of oil), except leave the first
formulas for cells V33:X34 at the bottom of Figure 9.18. The template row blank. The headings of the next columns specify which output will be
automatically calculates the probability of each finding and the posterior evaluated. For each of these columns, use the first row of the table (cells
probabilities (in cells V42:X43) based on these data. The decision tree then Y4:AD4) to write an equation that refers to the relevant output cell. In this case,
refers to these calculated probabilities when they are needed, as shown in the the cells of interest are (1) the decision of whether to do the survey (V15), (2)
formulas for cells P3:P11 in Figure 9.18. if so, whether to drill if the survey is favorable or unfavorable (W19 and W20),
While it takes some time and effort to consolidate the data and results, (3) if not, whether to drill (U19), and (4) the value of ExpectedPayoff (V26).
including all the necessary cross-referencing, this step is truly essential for The equations for Y4:AD4 referring to these output cells are shown below the
performing sensitivity analysis. Many pieces of data are used in several places spreadsheet in Figure 9.19.
on the decision tree. For example, the revenue if Goferbroke finds oil appears
in cells P6, P21, and L36. Performing sensitivity analysis on this piece of data page 365
now requires changing its value in only one place (cell V6) rather than three
(cells P6, P21, and L36). The benefits of consolidation are even more
important for the probability data. Changing any prior probability may cause all
the posterior probabilities to change. By including the posterior probability
template, the prior probability can be changed in one place and then all the
other probabilities are calculated and updated appropriately.

Organize the spreadsheet so that any piece of data needs to be changed in only one place.

After making any change in the cost data, revenue data, or probability data
in Figure 9.18, the spreadsheet nicely summarizes the new results after the
actual work to obtain these results is instantly done by the posterior probability
template and the decision tree. Therefore, experimenting with alternative data
values in a trial-and-error manner is one useful way of performing sensitivity
analysis.
Now let’s see how this sensitivity analysis can be done more systematically
by using a data table.

Using a Data Table to Do Sensitivity Analysis

Systematically
 Next, select the entire table (Y4:AD15) and then choose Data Table from
the What-If Analysis menu of the Data tab. In the Data Table dialog box (as
shown at the bottom left of Figure 9.19), indicate the column input cell (V9),
which refers to the data cell that is being changed in the first column of the
table.
Clicking OK then generates the table shown in Figure 9.19. For each trial
value for the data cell listed in the first column of the table, the corresponding
output cell values are calculated and displayed in the other columns of the
table. Some of the output in the data table is not relevant. For example, when
the decision is to not do the survey in column Z, the results in columns AA and
AB (what to do given favorable or unfavorable survey results) are not relevant.
Similarly, when the decision is to do the survey in column Z, the results in
column AC (what to do if you don’t do the survey) are not relevant. The
relevant output has been formatted in boldface to make it stand out compared to
the irrelevant output.
Figure 9.19 reveals that the optimal initial decision switches from Sell
without a survey to doing the survey somewhere between 0.1 and 0.2 for the
prior probability of oil and then switches again to Drill without a survey
somewhere between 0.3 and 0.4. Using the spreadsheet in Figure 9.18, trial-
and-error analysis soon leads to the following conclusions about how the
optimal policy depends on this probability.

FIGURE 9.18
In preparation for performing sensitivity analysis on the full Goferbroke Co. problem, the data
and results have been consolidated on the spreadsheet to the right of the decision tree.

page 366
 survey and then drill only if the survey is favorable. Otherwise, sell. So the
consulting geologist’s numbers can be off by quite a bit and we still would be
doing the right thing.
Jennifer: Yes, that was a key finding, wasn’t it? OK. Does this mean that you
are comfortable now with a decision to proceed with the seismic survey and
then either drill or sell depending on the outcome of the survey?
Max: Not quite. There is still one thing that bothers me.
Jennifer: What’s that?
Max: Suppose the seismic survey gives us a favorable seismic sounding, so we
drill. If the land turns out to be dry, then I’m out 130,000 bucks! As I said at the
beginning, that would nearly put us out of business. That scares me. I currently
am shorter of working capital than I normally am. Therefore, losing $130,000
now would hurt more than it normally does. It doesn’t look like this approach is
really taking that into account.
Jennifer: No, you’re right. It really doesn’t. This approach just looks at
average monetary values. That isn’t good enough when you’re dealing with
FIGURE 9.19 such large amounts. You wouldn’t be willing to flip a coin to determine whether
The data table that shows the optimal policy and expected payoff for various trial values of you win or lose $130,000, right?
the prior probability of oil.
Considering average monetary values isn’t good enough when uncomfortably large losses
page 367 can occur.

Optimal Policy Max: No, I sure wouldn’t.

Let p = Prior probability of oil. Jennifer: OK, that’s the tipoff. As I mentioned the first time we talked about
If p ≤ 0.168, then sell the land (no seismic survey). this problem, I think the circumstances here indicate that we need to go beyond
If 0.169 ≤ p ≤ 0.308, then do the survey: drill if favorable and sell if not. dollars and cents to look at the consequences of the possible outcomes.
If p ≥ 0.309, then drill for oil (no seismic survey). Fortunately, decision analysis has a way of doing this by introducing utilities.
The basic idea is that the utility of an outcome measures the true value to you of
Max’s Reaction that outcome rather than just the monetary value. So by expressing payoffs in
Max: Very interesting. I especially liked the way we were able to use that terms of utilities, the decision tree analysis would find the average utility at
sensitivity analysis spreadsheet to see immediately what would happen when each node instead of the average monetary value. So now the decisions would
we change some of the numbers. And there was one thing that I found be based on giving you the highest possible average utility.
particularly encouraging.
Jennifer: What was that? Review Questions
Max: When we changed that prior probability of oil to the most plausible 1. When preparing to perform sensitivity analysis, how should one begin organizing the
nearby values, it kept coming back with the same answer. Do the seismic spreadsheet that contains the decision tree?
2. Performing sensitivity analysis on a certain piece of data should require changing its value in as having a decreasing marginal utility for money. Such an individual is
how many places on the spreadsheet?
referred to as being risk averse.
3. What conclusion was drawn for the full Goferbroke problem regarding how the decision
should depend on the prior probability of oil?
Two different individuals can have very different utility functions for money.

However, not all individuals have a decreasing marginal utility for money.
9.9 USING UTILITIES TO BETTER REFLECT Some people are risk seekers instead of risk averse, and they go through life
THE VALUES OF PAYOFFS looking for the “big score.” The slope of their utility function increases as the
amount of money increases, so they have an increasing marginal utility for
Thus far, when applying Bayes’ decision rule, we have assumed that the money.
expected payoff in monetary terms is the appropriate measure of the Figure 9.21 compares the shape of the utility function for money for risk-
consequences of taking an action. However, in many situations where very averse and risk-seeking individuals. Also shown is the intermediate case of a
large amounts of money are involved, this assumption is inappropriate. risk-neutral individual, who prizes money at its face value. Such an
For example, suppose that an individual is offered the choice of (1) individual’s utility for money is simply proportional to the amount of money
accepting a 50–50 chance of winning $100,000 or (2) receiving $40,000 with involved. Although some people appear to be risk neutral when only small
certainty. Many people would prefer the $40,000 even though the expected amounts of money are involved, it is unusual to be truly risk neutral with very
payoff on the 50–50 chance of winning $100,000 is $50,000. A company may large amounts.
be unwilling to invest a large sum of money in a new product, even when the
expected profit is substantial, if there is a risk of losing its investment and
thereby becoming bankrupt. People buy insurance even though it is a poor
investment from the viewpoint of the expected payoff.

page 368
Do these examples invalidate Bayes’ decision rule? Fortunately, the answer
is no, because there is a way of transforming monetary values to an appropriate
scale that reflects the decision maker’s preferences. This scale is called the
utility function for money.

Utility Functions for Money

Figure 9.20 shows a typical utility function U(M) for money M. The intuitive
interpretation is that it indicates that an individual having this utility function
would value obtaining $30,000 twice as much as $10,000 and would value
obtaining $100,000 twice as much as $30,000. This reflects the fact that the FIGURE 9.20
A typical utility function for money, where U(M) is the utility of obtaining an amount of
person’s highest-priority needs would be met by the first $10,000. Having this money M.
decreasing slope of the function as the amount of money increases is referred to
 Thus, by weighting the two possible utilities (1 and 0) by their probabilities,
the expected utility is

In all three of these cases, a decision maker with the utility function in Figure 9.20 would
be indifferent between the two alternatives because they have the same expected utility.

FIGURE 9.21 Therefore, for each of the following three pairs of alternatives, the above
The shape of the utility function for money for (a) risk-averse, (b) risk-seeking, and (c) risk-
neutral individuals. fundamental property indicates that the decision maker is indifferent between
the first and second alternatives.
page 369 . First alternative: The offer with p = 0.25, so E(utility) = 0.25.
Second alternative: Definitely obtain $10,000, so utility = 0.25.
It also is possible to exhibit a mixture of these kinds of behavior. For
example, an individual might be essentially risk neutral with small amounts of . First alternative: The offer with p = 0.5, so E(utility) = 0.5.
money, then become a risk seeker with moderate amounts, and then turn risk Second alternative: Definitely obtain $30,000, so utility = 0.5.
averse with large amounts. In addition, one’s attitude toward risk can shift over . First alternative: The offer with p = 0.75, so E(utility) = 0.75.
time depending on circumstances. Second alternative: Definitely obtain $60,000, so utility = 0.75.
Managers of a business firm need to consider the company’s circumstances
This example also illustrates one way in which the decision maker’s utility
and the collective philosophy of top management in determining the appropriate
function for money can be constructed in the first place. The decision maker
attitude toward risk when making managerial decisions.
would be made the same hypothetical offer to obtain a large amount of money
The fact that different people have different utility functions for money has
(e.g., $100,000) with probability p, or nothing (utility = 0) otherwise. Then, for
an important implication for decision making in the face of uncertainty.
each of a few smaller amounts of money (e.g., $10,000, $30,000, and $60,000),
When a utility function for money is incorporated into a decision analysis approach to a problem, the decision maker would be asked to choose a value of p that would make him
this utility function must be constructed to fit the current preferences and values of the decision
maker involved. (The decision maker can be either a single individual or a group of people.) or her indifferent between the offer and definitely obtaining that amount of
money. The utility of the smaller amount of money then is p times the utility of
The key to constructing the utility function for money to fit the decision the large amount. When the utility of the large amount has been set equal to 1, as
maker is the following fundamental property of utility functions. in Figure 9.20, this conveniently makes the utility of the smaller amount simply
Fundamental Property: Under the assumptions of utility theory, the decision maker’s utility equal to p. The utility values in Figure 9.20 imply that the decision maker has
function for money has the property that the decision maker is indifferent between two chosen p = 0.25 when M = $10,000, p = 0.5 when M = $30,000, and p = 0.75
alternative courses of action if the two alternatives have the same expected utility.
when M = $60,000. (Constructing the utility function in this way is an example
To illustrate, suppose that the decision maker has the utility function shown of the equivalent lottery method described later in this section.)
in Figure 9.20. Further suppose that the decision maker is offered the following The scale of the utility function is irrelevant. In other words, it doesn’t
opportunity. matter whether the values of U(M) at the dashed lines in Figure 9.20 are 0.25,
Offer: An opportunity to obtain either $100,000 (utility = 1) with probability p or nothing (utility = 0) 0.5, 0.75, 1 (as shown) or 10,000, 20,000, 30,000, 40,000, or whatever. All the
with probability (1 − p). utilities can be multiplied by any positive constant without affecting which
decision alternative will have the largest expected utility. It also is possible to has gone heavily into debt to keep going. The worst-case scenario would be to
add the same constant (positive or negative) to all the utilities without affecting come up with $30,000 for a seismic survey and then still lose $100,000 by
which decision alternative will have the largest expected utility. drilling when there is no oil. This scenario would not bankrupt the company at
this point but definitely would leave it in a precarious financial position.
page 370 On the other hand, striking oil is an exciting prospect, since earning
$700,000 finally would put the company on fairly solid financial footing.
For these reasons, we have the liberty to set the value of U(M) arbitrarily Max is the decision maker for this problem. Therefore, to prepare for using
for two values of M, so long as the higher monetary value has the higher utility. utilities to analyze the problem, it is necessary to construct Max’s utility
It is particularly convenient to set U(M) = 0 for the smallest value of M under function for money, U(M), where we will express the amount of money M in
consideration and to set U(M) = 1 for the largest M, as was done in Figure units of thousands of dollars.
9.20. By assigning a utility of 0 to the worst outcome and a utility of 1 to the We start by assigning utilities of 0 and 1, respectively, to the smallest and
best outcome, and then determining the utilities of the other outcomes largest possible payoffs. Since the smallest possible payoff is M = −130 (a loss
accordingly, it becomes easy to see the relative utility of each outcome along of $130,000) and the largest is M = 700 (a gain of $700,000), this gives
the scale from worst to best. U(−130) = 0 and U(700) = 1.
The objective now is to maximize the expected utility rather than the expected payoff in
To determine the utilities for other possible monetary payoffs, it is
monetary terms. necessary to probe Max’s attitude toward risk. Especially important are his
feelings about the consequences of the worst possible loss ($130,000) and the
Now we are ready to summarize the basic role of utility functions in best possible gain ($700,000), as well as how he compares these
decision analysis. consequences. Let us eavesdrop as Jennifer probes these feelings with Max.
When the decision maker’s utility function for money is used to measure the relative worth of the
various possible monetary outcomes, Bayes’ decision rule replaces monetary payoffs by the
Interviewing Max
corresponding utilities. Therefore, the optimal decision (or series of decisions) is the one that Jennifer: Well now, these utilities are intended to reflect your feelings about
maximizes the expected utility.
the true value to you of these various possible payoffs. Therefore, to pin down
Only utility functions for money have been discussed here. However, we what your utilities are, we need to talk some about how you feel about these
should mention that utility functions can sometimes still be constructed when payoffs and their consequences for the company.
some or all of the important consequences of the decision alternatives are not Max: Fine.
monetary in nature. (For example, the consequences of a doctor’s decision Jennifer: A good place to begin would be the best and worst possible cases.
alternatives in treating a patient involve the future health of the patient.) This is The possibility of gaining $700,000 or losing $130,000.
not necessarily easy, since it may require making value judgments about the Max: Those are the big ones all right.
relative desirability of rather intangible consequences. Nevertheless, under Jennifer: OK, suppose you drill without paying for a seismic survey and then
these circumstances, it is important to incorporate such value judgments into the you find oil, so your profit is about $700,000. What would that do for the
decision process. company?
Max: A lot. That would finally give me the capital I need to become more of a
Dealing with the Goferbroke Co. Problem major player in this business. I then could take a shot at finding a big oil field.
Recall that the Goferbroke Co. is operating without much capital, so a loss of That big strike I’ve talked about.
$100,000 would be quite serious. As the owner of the company, Max already
Jennifer: OK, good. Now let’s talk about the consequences if you were to get than $100,000. The higher loss would be quite a bit more painful. I wouldn’t be
that biggest possible loss instead. Suppose you pay for a seismic survey, then willing to take this risk with just a one-in-four chance of gaining $700,000.
you drill and the land is dry. So you’re out about $130,000. How bad would Jennifer: OK, so now we know that the point at which you would be
that be? What kind of future would the company have? indifferent between going ahead or not is somewhere between having a one-in-
four chance and a 50–50 chance of gaining $700,000 rather than losing
page 371 $130,000. Let’s see if we can pin down just where your point of
indifference is within this range from one-in-four and 50–50. Let’s try a one-
Max: Well, let me put it this way. It would put the company in a pretty in-three chance. Would you go ahead and drill with a one-in-three chance of
uncomfortable financial position. I would need to work hard on getting some gaining $700,000 versus a two-in-three chance of losing $130,000, or would
more financing. Then we would need to cautiously work our way out of the hole you choose to sell the land for $90,000?
by forming some partnerships for some low-risk, low-gain drilling. But I think
we could do it. I’ve been in that position a couple times before and come out of The point of indifference is the point where the decision maker is indifferent between
two hypothetical alternatives.
it. We’d be OK.
Jennifer: It sounds like you wouldn’t be overly worried about such a loss as
Max: Hmm. That’s not so clear. What would be the average payoff in this case?
long as you have reasonable odds for a big payoff to justify this risk.
Jennifer: Almost $147,000.
Max: That’s right.
Max: Not bad. Hmm, one chance in three of gaining $700,000. That’s tempting.
Jennifer: OK, now let’s talk about those odds. What I’m going to do is set up a
But two chances in three of losing $130,000 with all the problems involved
simpler hypothetical situation. Suppose these are your alternatives. One is to
with that. I don’t know. $90,000 would be a sure thing. That’s a tough one.
drill. If you find oil, you clear $700,000. If the land is dry, you’re out $130,000.
Jennifer: OK, let’s try this. Suppose your chances of gaining $700,000 were a
The only other alternative is to sell the land for $90,000. I know this isn’t your
little better than one-in-three. Would you do it?
actual situation since $700,000 does not include the cost of a survey whereas
the loss of $130,000 does, but let’s pretend that these are your alternatives. Max: Yes, I think so.
Max: I don’t understand why you want to talk about a situation that is different Jennifer: And if your chances were a little under one-in-three?
from what we are facing. Max: Then I don’t think I would do it.
Jennifer: Trust me. Considering these kinds of hypothetical situations is going Jennifer: OK. You’ve convinced me that your point of indifference is one-in-
to enable us to determine your utilities. three. That’s exactly what I needed to know.
Max: OK.
Jennifer: Now presumably if you had a 50–50 chance of either clearing Finding U(90)
$700,000 or losing $130,000, you would drill. Max has indeed given Jennifer just the information she needs to determine
Max: Sure. U(90), Max’s utility for a payoff of 90 (a gain of $90,000). Recall that U(−130)
Jennifer: If you had a smaller chance, say one-in-four of gaining $700,000, already has been set at U(−130) = 0 and that U(700) already has been set at
versus a three-in-four chance of losing $130,000, would you choose to drill or U(700) = 1. Here is the procedure that Jennifer is using to find U(90).
sell the land for $90,000?
Max: Well, that’s almost the original decision we were trying to make, before page 372
we considered the seismic survey. However, there is one big difference. Now
you’re asking me to suppose that the loss if there is no oil is $130,000 rather The decision maker (Max) is offered two alternatives, A1 and A2.
 A1: Obtain a payoff of 700 with probability p. U(minimum) = 0
Obtain a payoff of −130 with probability (1 − p). . To determine the utility of another potential payoff M, the decision maker is
A2: Definitely obtain a payoff of 90. offered the following two hypothetical alternatives:

Question to the decision maker: What value of p makes you indifferent between A1: Obtain a payoff of maximum with probability p.
these two alternatives? Recall that Max has chosen p = 1/3. Obtain a payoff of minimum with probability 1 − p.
For a given choice of p, the expected utility for A1 is A2: Definitely obtain a payoff of M.

Question to the decision maker: What value of p makes you indifferent between
these two alternatives? Then U(M) = p.

Constructing Max’s Utility Function for Money

If the decision maker is indifferent between the two alternatives, the
fundamental property of utility functions says that the two alternatives must We now have found utilities for three possible payoffs (–130, 90, and 700) for
have the same expected utility. Therefore, the utility for A2 must also be p. Goferbroke. Plotting these values on a graph of the utility function U(M) versus
the monetary payoff M and then drawing a smooth curve through these points
Since Max chose a point of indifference of p = 1/3, the utility for A2 must be
gives the curve shown in Figure 9.22.
1/3, so U(90) = 1/3.
This curve is an estimate of Max’s utility function for money. To find the
utility values for the other possible payoffs (−100, 60, and 670), Max could
The Equivalent Lottery Method for Determining Utilities repeat step 3 of the equivalent lottery method for M = −100, M = 60, and M =
The above procedure for finding U(90) illustrates that the key to finding the 670. However, since −100 is so close to −130, 60 is so close to 90, and 670 is
utility for any payoff M is having the decision maker select a point of so close to 700, an alternative is to estimate these utilities as the values on the
indifference between two alternatives, where one of them (A1) involves a curve in Figure 9.22 at M = −100, M = 60, and M = 670. Following the
lottery between the largest payoff and the smallest payoff and the other corresponding dotted lines in the figure leads to U(−100) = 0.05, U(60) = 0.30,
alternative (A2) is to receive a sure payoff of M. At the point of indifference, and U(670) = 0.97. Table 9.9 gives the complete list of possible payoffs and
the lottery is equivalent to the sure payoff in the sense that they have the same their utilities.
expected utility, so the procedure is referred to as the equivalent lottery For comparative purposes, the dashed line in Figure 9.22 shows the utility
method. Here is an outline of the procedure. function that would result if Max were completely risk neutral. By nature, Max
is inclined to be a risk seeker. However, the difficult financial circumstances of
Equivalent Lottery Method his company that he badly wants to keep solvent has forced him to adopt a
moderately risk-averse stance in addressing his current decisions.
. Determine the largest potential payoff (call it maximum), and assign it a utility
of 1:
page 373
U(maximum) = 1
. Determine the smallest potential payoff (call it minimum), and assign it a
utility of 0:
 Thus, using Analytic Solver once again, our final decision tree with utilities
shown in Figure 9.23 closely resembles the one in Figure 9.17 given in Section
9.7. The nodes and branches are exactly the same, as are the probabilities for
the branches emanating from the event nodes. However, the key difference from
Figure 9.17 is that the monetary payoff at each terminal node now has been
replaced by the corresponding utility from Table 9.9. (This was accomplished
with Analytic Solver by entering this same utility as the “cash flow” at the
terminal branch and then entering “cash flows” of 0 at all the preceding
branches.) It is these utilities that have been used by Analytic Solver to
compute the expected utilities given next to all the nodes.

At each terminal branch, enter the utility of that outcome as the “cash flow” there and
then do not change the default value of 0 for the “cash flow” at the preceding branches.

These expected utilities lead to the same decisions as in Figure 9.17 at all
FIGURE 9.22 decision nodes except the bottom one in cell F41. The decision at this node
Max’s utility function for money as the owner of Goferbroke Co.
now switches to sell instead of drill. However, the solution procedure still
TABLE 9.9 leaves this node on a closed path, as indicated by the 1 in cell B29. Therefore,
Utilities for the Goferbroke Co. Problem the overall optimal policy remains the same as that obtained in Figure 9.17 (do
the seismic survey; sell if the result is unfavorable; drill if the result is
Monetary Payoff, M Utility, U(M)
favorable).
−130 0
−100 0.05
60 0.30 page 374
90 0.333
670 0.97
700 1

Using a Decision Tree to Analyze the Problem with

Utilities
Now that Max’s utility function for money has been identified in Table 9.9 (and
Figure 9.22), this information can be used with a decision tree as summarized
next.
The procedure for using a decision tree to analyze the problem now is identical to that described
in Section 9.7 except for substituting utilities for monetary payoffs. Therefore, the value obtained to
evaluate each node of the tree now is the expected utility there rather than the expected
(monetary) payoff. Consequently, the optimal decision selected by Bayes’ decision rule maximizes
the expected utility for the overall problem.
 The approach of maximizing the expected monetary payoff used in the
preceding sections was equivalent to assuming that the decision maker is
neutral toward risk. By using utility theory with an appropriate utility function,
the optimal solution now reflects the decision maker’s attitude about risk.
Because Max adopted only a moderately risk-averse stance, the optimal policy
did not change from before. For a somewhat more risk-averse decision maker,
the optimal solution would switch to the more conservative approach of
immediately selling the land (no seismic survey).

The previous approach of maximizing the expected monetary payoff assumes a risk-
neutral decision maker.

Jennifer and Max are to be commended for incorporating utilities into a

decision analysis approach to his problem. Utilities help to provide a rational
approach to decision making in the face of uncertainty. However, many
managers are not sufficiently comfortable with the relatively abstract notion of
utilities, or with working with probabilities to construct a utility function, to be
willing to use this approach. Consequently, utilities are not used nearly as
widely in practice as some of the other techniques of decision analysis
described in this chapter, including Bayes’ decision rule (with monetary
payoffs) and decision trees.

Another Approach for Estimating U(M)

The procedure described earlier for constructing U(M) asks the decision maker
to repeatedly apply the equivalent lottery method, which requires him (or her)
each time to make a difficult decision about which probability would make him
indifferent between two alternatives. Many managers would be uncomfortable
with making this kind of decision. Therefore, an alternative approach is
sometimes used instead to estimate the utility function for money.
This approach assumes that the utility function has a certain mathematical
form and then adjusts this form to fit the decision maker’s attitude toward risk
FIGURE 9.23 as closely as possible. For example, one particularly popular form to assume
The final decision tree constructed and solved by Analytic Solver for the full Goferbroke Co. (because of its relative simplicity) is the exponential utility function,
problem when using Max’s utility function for money to maximize expected utility.

page 375
where R is the decision maker’s risk tolerance. This utility function has the
kind of shape shown in Figure 9.21(a), so it is designed to fit a risk-averse
individual. A great aversion to risk corresponds to a small value of R (which
would cause the curve in this figure to bend sharply), whereas a small aversion
to risk corresponds to a large value of R (which gives a much more gradual
bend in the curve).

Since R measures the decision maker’s risk tolerance, the aversion to risk decreases as
R increases.

A decision maker’s risk tolerance tends to vary over time as his (or her) FIGURE 9.24
wealth changes. He tends to have a higher risk tolerance when he is relatively The Tree tab of the Analytic Solver Options dialog box allows you to set several options for
how the decision tree is solved. Here the options are set to use the exponential utility function,
wealthy than when he is not. However, given his current degree of wealth, the
to maximize profit, and to use an R value of 1,000.
exponential utility function assumes that he has a constant risk tolerance over
the entire range of potential outcomes that may be realized after the decision is
Using Analytic Solver with an Exponential Utility
made. This often is a reasonable assumption. Unfortunately, it is a questionable
assumption in Max’s situation because he is unusually risk averse about the Function
worst possible outcome (a loss of $130,000) but has a very high risk tolerance Analytic Solver includes the option of using the exponential utility function.
when comparing large potential gains. This is why Jennifer never raised the Clicking on the Options button on the Analytic Solver ribbon and choosing Tree
possibility of using an exponential utility function. reveals the options shown in Figure 9.24. The Certainty Equivalents section
In other situations where the consequences of the potential losses are not as gives two choices—Expected Values or Exponential Utility Function. Choosing
severe, assuming an exponential utility function may provide a reasonable the latter revises the decision tree to incorporate the exponential utility
approximation. In such a case, here is an easy way of estimating the appropriate function. The Decision Node EV/CE section has two choices, Maximize or
value of R. The decision maker would be asked to choose the number R that Minimize, which enable you to specify whether the objective is to maximize the
would make him indifferent between the following two alternatives. measure of performance (expected utility in this case) or to minimize that
measure. (Maximize is the default choice, as has been used throughout the
A1: A 50–50 gamble where he would gain R dollars with chapter.) The Risk Tolerance box is used to enter a value to be used for R when
probability 0.5 and lose R/2 dollars with probability calculating the exponential utility function. (These same options are also
0.5. available on the Platform tab of the Model pane in Analytic Solver.)
A2: Neither gain nor lose anything.
Analytic Solver Tip: The Decision Tree section of the Platform tab on the Model pane
lets you specify whether to use expected monetary values or the exponential utility
For example, if the decision maker were indifferent between doing nothing or function for applying Bayes’ decision rule. The entry in Decision Node EV/CE (either
taking a 50–50 gamble where he would gain $1,000 with probability 0.5 and Maximize or Minimize) also lets you specify whether the objective is to maximize the
lose $500 with probability 0.5, then R = 1,000. measure of performance (expected payoff or expected utility in this case) or to minimize
that measure.

page 376 To illustrate, suppose that the exponential utility function with a risk
tolerance of R = 1,000 were to be used as a rough approximation for analyzing
the full Goferbroke Co. problem. (Since this problem expresses payoffs in units
of thousands of dollars, R = 1,000 here is equivalent to using R = 1,000,000
when payoffs are in units of dollars.) The resulting decision tree is shown in
Figure 9.25. There are now two values calculated below and to the left of each
node. The lower number represents the expected utility value at that stage in the
decision tree. The upper number represents the certain payoff that is equivalent
to this expected utility value. For example, cell A31 indicates that the expected
value of the exponential utility function for this decision would be 0.0932. This
expected utility is equivalent to a certain payoff of $98,000, as indicated in cell
A30.
The exponential utility function leads to the same decisions as in Figure
9.23. The overall optimal policy remains to do the seismic survey; sell if the
result is unfavorable; drill if the result is favorable. However, the optimal
policy changes when the value of R is decreased far enough. For values of R
less than 728, the optimal policy switches to not doing the survey and selling
the land. Thus, a more risk-averse decision maker would make the safer
decision for Goferbroke—sell the land and receive $90,000 for sure.

Review Questions
1. What are utilities intended to reflect?
2. What is the shape of the utility function for money for a risk-averse individual? A risk-seeking
individual? A risk-neutral individual?
3. What is the fundamental property of utility functions?
4. What is the lottery when using the equivalent lottery method?
5. Given two hypothetical alternatives where one of them involves a probability p, what is meant
by the point of indifference between these two alternatives?
6. When using utilities with a decision tree, what kind of value is obtained to evaluate each node
of the tree?
7. What decisions did Max make regarding the full Goferbroke Co. problem?

page 377

FIGURE 9.25
The final decision tree constructed and solved by Analytic Solver for the full Goferbroke Co.
problem when using an exponential utility function with R = 1,000.
 page 378 has become quite popular is called the influence diagram, which provides
another helpful way of showing the interrelationships among a decision maker’s
alternatives, uncertainties, and values.
9.10 THE PRACTICAL APPLICATION OF Although the Goferbroke problem only involved a single decision maker
(Max) assisted by a single analyst (Jennifer), many strategic business decisions
DECISION ANALYSIS are made collectively by management. One technique for group decision making
is called decision conferencing. This is a process where the group comes
In one sense, the Goferbroke Co. problem is a very typical application of together for discussions in a decision conference with the help of an analyst and
decision analysis. Like other applications, Max needed to make his decisions a group facilitator. The facilitator works directly with the group to help it
(Do a seismic survey? Drill for oil or sell the land?) in the face of great structure and focus discussions, think creatively about the problem, bring
uncertainty. The decisions were difficult because their payoffs were so assumptions to the surface, and address the full range of issues involved. The
unpredictable. The outcome depended on factors that were outside Max’s analyst uses decision analysis to assist the group in exploring the implications
control (does the land contain oil or is it dry?). He needed a framework and of the various decision alternatives. With the assistance of a computerized
methodology for rational decision making in this uncertain environment. These group decision support system, the analyst builds and solves models on the
are the usual characteristics of applications of decision analysis. spot, and then performs sensitivity analysis to respond to what-if questions
However, in other ways, the Goferbroke problem is not such a typical from the group.
application. It was oversimplified to include only two possible states of nature Applications of decision analysis commonly involve a partnership between
(oil and dry), whereas there actually would be a considerable number of the managerial decision maker (whether an individual or a group) and an
distinct possibilities. For example, the actual state might be dry, a small amount analyst (whether an individual or a team) with training in management science.
of oil, a moderate amount, a large amount, and a huge amount, plus different Some managers are not as fortunate as Max in having a staff member (let alone
possibilities concerning the depth of the oil and soil conditions that impact the a daughter) like Jennifer who is qualified to serve as the analyst. Therefore, a
cost of drilling to reach the oil. Max also was considering only two alternatives considerable number of management consulting firms specializing in decision
for each of two decisions. Real applications commonly involve more analysis have been formed to fill this role.
decisions, more alternatives to be considered for each one, and many possible
If you would like to do more reading about the practical application of
states of nature.
decision analysis, a good place to begin would be the leadoff article2 in the
The Goferbroke Co. problem could have included many more states of nature. first issue of the journal Decision Analysis that was founded in 2004 to focus
on applied research in decision analysis. This leadoff article provides a
Problems as tiny as the Goferbroke problem can be readily analyzed and detailed discussion of various publications that present applications of decision
solved by hand. However, real applications typically involve large decision analysis. The many subsequent issues of this journal have continued to further
trees, whose construction and analysis require the use of a software package develop the theory of decision analysis and to describe various innovative
(such as Analytic Solver). In some cases, the decision tree can explode in size applications.
with many thousand terminal branches. Special algebraic techniques are being
developed and incorporated into the solvers for dealing with such large Review Questions
problems.
1. How does the Goferbroke Co. problem compare with typical applications of decision
Other kinds of graphical techniques also are available to complement the analysis?
decision tree in representing and solving decision analysis problems. One that 2. What is the purpose of an influence diagram?
3. Who are the typical participants in a decision-conferencing process? Decision analysis is widely used. Versatile software packages for personal computers have
4. Where can a manager go for expert help in applying decision analysis if a qualified analyst is become an integral part of the practical application of decision analysis.
not available on staff?

page 379 Glossary

alternatives The options available to the decision maker for the decision under consideration.
(Section 9.1), 336
Bayes’ decision rule A popular criterion for decision making that uses probabilities to
9.11 Summary calculate the expected payoff for each decision alternative and then chooses the one with the
largest expected payoff. (Section 9.2), 340

Decision analysis is a valuable technique for decision making in the face of great uncertainty. It Bayes’ theorem A formula for calculating a posterior probability of a state of nature.
provides a framework and methodology for rational decision making when the outcomes are (Section 9.6), 355
uncertain. branch A line emanating from a node in a decision tree. (Section 9.3), 342
In a typical application, a decision maker needs to make either a single decision or a
decision conferencing A process used for group decision making. (Section 9.10), 378
sequence of decisions (with additional information perhaps becoming available between
decisions). A number of alternatives are available for each decision. Uncontrollable random decision maker The individual or group responsible for making the decision under
factors affect the payoff that would be obtained from a decision alternative. The possible consideration. (Section 9.1), 336
outcomes of the random factors are referred to as the possible states of nature. decision node A point in a decision tree where a decision needs to be made. (Section 9.3),
Which state of nature actually occurs will be learned only after making the decisions. 342
However, prior to the decisions, it often is possible to estimate prior probabilities of the
decision tree A graphical display of the progression of decisions and random events to be
respective states of nature.
considered. (Sections 9.3 and 9.7), 342, 357
Various alternative decision criteria are available for making the decisions. A particularly
popular one is Bayes’ decision rule, which uses the prior probabilities to determine the equivalent lottery method The procedure for finding the decision maker’s utility for a
expected payoff for each decision alternative and then chooses the one with the largest specific amount of money by comparing two hypothetical alternatives where one involves a
expected payoff. This is the criterion (accompanied by sensitivity analysis) that is mostly used in gamble. (Section 9.9), 372
practice, so it is the focus of much of the chapter. event node A point in a decision tree where a random event will occur. (Section 9.3), 342
Sensitivity analysis is very helpful for evaluating the effect of having inaccurate estimates of
expected monetary value (EMV) criterion An alternative name for Bayes’ decision rule
the data for the problem, including the probabilities, revenues, and costs. Data tables can be
when the payoffs have monetary values. (Section 9.2), 341
used to systematically vary the data and see how it affects the optimal decisions or expected
payoffs. expected payoff (EP) For a decision alternative, it is the weighted average of the payoffs,
It sometimes is possible to pay for a test or survey to obtain additional information about the using the probabilities of the states of nature as the weights. (Section 9.2), 340
probabilities of the various states of nature. Calculating the expected value of perfect expected value of perfect information (EVPI) The increase in the expected payoff that
information provides a quick way of checking whether doing this might be worthwhile. could be obtained if it were possible to learn the true state of nature before making the decision.
When more information is obtained, the updated probabilities are referred to as posterior (Sections 9.4 and 9.5), 350, 351
probabilities. A probability tree diagram is helpful for calculating these new probabilities.
For problems involving a sequence of decisions (including perhaps a decision on whether to
obtain more information), a decision tree commonly is used to graphically display the page 380
progression of decisions and random events. The calculations for applying Bayes’ decision rule
then can be performed directly on the decision tree one event node or decision node at a time. expected value of sample information (EVSI) The increase in the expected payoff that
Spreadsheet packages, such as Analytic Solver, are very helpful for constructing and solving could be obtained by performing a test to obtain more information, excluding the cost of the test.
decision trees. (Section 9.7), 362
When the problem involves the possibility of uncomfortably large losses, utilities provide a exponential utility function A utility function designed to fit some risk-averse individuals.
way of incorporating the decision maker’s attitude toward risk into the analysis. Bayes’ decision (Section 9.9), 375
rule then is applied by expressing payoffs in terms of utilities rather than monetary values.
influence diagram A diagram that complements the decision tree for representing and
analyzing decision analysis problems. (Section 9.10), 378
maximax criterion A very optimistic decision criterion that does not use prior probabilities
and simply chooses the decision alternative that could give the largest possible payoff. (Section
9.2), 338
maximin criterion A very pessimistic decision criterion that does not use prior probabilities
Learning Aids for This Chapter
and simply chooses the decision alternative that provides the best guarantee for its minimum
possible payoff. (Section 9.2), 339 All learning aids are available at www.mhhe.com/Hillier6e.
maximum likelihood criterion A criterion for decision making with probabilities that focuses Excel Files:
on the most likely state of nature. (Section 9.2), 339
Bayes’ Decision Rule for First Goferbroke Problem
node A junction point in a decision tree. (Section 9.3), 342
Decision Tree for First Goferbroke Problem
payoff A quantitative measure of the outcome from a decision alternative and a state of Data Table for First Goferbroke Problem
nature. (Section 9.1), 337
EP with Perfect Info for First Goferbroke Problem
payoff table A table giving the payoff for each combination of a decision alternative and a
state of nature. (Section 9.1), 337 Decision Tree for EVPI for First Goferbroke Problem

point of indifference The point where the decision maker is indifferent between the two Template for Posterior Probabilities
hypothetical alternatives in the equivalent lottery method. (Section 9.9), 371 Decision Tree for Full Goferbroke Problem (with Data Table)
posterior probabilities Revised probabilities of the states of nature after doing a test or Decision Tree for Full Goferbroke Problem with Max’s Utility Function
survey to improve the prior probabilities. (Sections 9.5 and 9.6), 352, 353 Decision Tree for Full Goferbroke Problem with Exponential Utility Function
prior probabilities The estimated probabilities of the states of nature prior to obtaining
Excel Add-in:
additional information through a test or survey. (Section 9.1), 337
Analytic Solver
probability tree diagram A diagram that is helpful for calculating the posterior probabilities
of the states of nature. (Section 9.6), 354 Supplements to This Chapter:
risk-averse individual An individual whose utility function for money has a decreasing slope Decision Criteria
as the amount of money increases.(Section 9.9), 368 Using TreePlan Software for Decision Trees
risk-neutral individual An individual whose utility for money is proportional to the amount of
Excel Files for Supplement 1 to This Chapter:
money involved. (Section 9.9), 368
Template for Maximax Criterion
risk seeker An individual whose utility function for money has an increasing slope as the
amount of money increases. (Section 9.9), 368 Template for Maximin Criterion
sensitivity analysis The study of how other plausible values for the probabilities of the states Template for Realism Criterion
of nature (or for the payoffs) would affect the recommended decision alternative. (Sections 9.4 Template for Minimax Regret Criterion
and 9.8), 346, 364
Template for Maximum Likelihood Criterion
states of nature The possible outcomes of the random factors that affect the payoff that Template for Equally Likely Criterion
would be obtained from a decision alternative. (Section 9.1), 336
utility The utility of an outcome measures the instrinsic value to the decision maker of that
outcome. (Sections 9.1 and 9.9), 336 page 381
utility function for money, U(M) A plot of utility versus the amount of money M being
received. (Section 9.9), 368
Solved Problems
The solutions are available at www.mhhe.com/Hillier6e.
 9.S1. New Vehicle Introduction Meredith Delgado owns a small firm that has developed software for organizing and playing
music on a computer. Her software contains a number of unique features that she has patented,
The General Ford Motors Corporation (GFMC) is planning the introduction of a brand new
so her company’s future has looked bright.
SUV—the Vector. There are two options for production. One is to build the Vector at the
However, there now has been an ominous development. It appears that a number of her
company’s existing plant in Indiana, sharing production time with its line of minivans that are
patented features were copied in similar software developed by MusicMan Software, a huge
currently being produced there. If sales of the Vector are just moderate, this will work out well
software company with annual sales revenue in excess of $1 billion. Meredith is distressed.
as there is sufficient capacity to produce both types of vehicles at the same plant. However, if
MusicMan Software has stolen her ideas and that company’s marketing power is likely to
sales of the Vector are strong, this option would require the operation of a third shift, which
enable it to capture the market and drive Meredith out of business.
would lead to significantly higher costs.
In response, Meredith has sued MusicMan Software for patent infringement. With attorney
A second option is to open a new plant in Georgia. This plant would have sufficient capacity
fees and other expenses, the cost of going to trial (win or lose) is expected to be $1 million. She
to meet even the largest projections for sales of the Vector. However, if sales are only
feels that she has a 60 percent chance of winning the case, in which case she would receive $5
moderate, the plant would be underutilized and therefore less efficient.
million in damages. If she loses the case, she gets nothing. Moreover, if she loses the case,
This is a new design, so sales are hard to predict. However, GFMC predicts that there
there is a 50 percent chance that the judge would also order Meredith to pay for court expenses
would be about a 60 percent chance of strong sales (annual sales of 100,000), and a 40 percent
and lawyer fees for MusicMan (an additional $1 million cost). MusicMan Software has offered
chance of moderate sales (annual sales of 50,000). The average revenue per Vector sold is
Meredith $1.5 million to settle this case out of court.
$30,000. Production costs per vehicle for the two production options depend upon sales, as
indicated in the table below. a. Construct and use a decision tree to determine whether Meredith should go to court or
Expected Production Cost per Vehicle for the Vector ($thousands) accept the settlement offer, assuming she wants to maximize her expected payoff.
b. To implement the equivalent lottery method to determine appropriate utility values for all the
Moderate Sales Strong Sales possible payoffs in this problem, what questions would need to be asked of Meredith?
c. Suppose that Meredith’s attitude toward risk is such that she would be indifferent between
Shared plant in Indiana 16 24
doing nothing and a gamble where she would win $1 million with 50 percent probability and
Dedicated plant in lose $500 thousand with 50 percent probability. Use the exponential utility function to re-solve
Georgia 22 20 the decision tree from part a.

The amortized annual cost of plant construction and other associated fixed costs for the
Georgia plant would total $400 million per year (regardless of sales volume). The fixed costs for
adding Vector production to the plant in Indiana would total $200 million per year (regardless of Problems
sales volume).
To the left of the following problems (or their parts), we have inserted the symbol A whenever
a. Construct a decision tree to determine which production option maximizes the expected
Analytic Solver can be used. The symbol T indicates that the Excel template for posterior
annual profit, considering fixed costs, production costs, and sales revenues.
probabilities can be helpful. Nearly all the problems can be conveniently formulated in a
b. Because of the uncertainty in expected sales for the Vector, GFMC is considering conducting spreadsheet format, so no special symbol is used to designate this. An asterisk on the problem
a marketing survey to determine customer attitudes toward the Vector to better predict the number indicates that at least a partial answer is given in the back of the book.
likelihood of strong sales. The marketing survey would give one of two results—a positive
attitude or a negative attitude toward the design. GFMC has used this marketing survey for
other vehicles. For vehicles that eventually had strong sales, the marketing survey indicated page 382
positive attitudes toward the design 70 percent of the time and negative attitudes 30 percent
of the time. For vehicles that eventually had moderate sales, the marketing survey indicated 9.1. You are given the following payoff table (in units of thousands of dollars) for a decision
positive attitudes toward the design 20 percent of the time and negative attitudes 80 percent analysis problem without probabilities.
of the time. Assuming GFMC conducts such a survey, construct a decision tree to determine
how the company should proceed and what the expected annual profit would be (ignoring the
cost of the survey).
c. What is the expected value of sample information in part b? What does this say about how
large the cost of the marketing survey can be before it would no longer be worthwhile to
conduct the survey?

9.S2. Settle or Go to Trial

 a. Which alternative should be chosen under the maxi-max criterion? where the economy is headed, and so has assigned prior probabilities
b. Which alternative should be chosen under the maxi-min criterion? of 0.1, 0.5, and 0.4, respectively, to these three scenarios. He also
9.2. Follow the instructions of Problem 9.1 with the following payoff table. estimates that his profits under these respective scenarios are those
given by the following table.

9.3. Jean Clark is the manager of the Midtown Saveway Grocery Store. She now needs to
Which investment should Warren make under each of the following criteria?
replenish her supply of strawberries. Her regular supplier can provide as many cases as she
wants. However, because these strawberries already are very ripe, she will need to sell them a. Maximax criterion.
tomorrow and then discard any that remain unsold. Jean estimates that she will be able to sell b. Maximin criterion.
10, 11, 12, or 13 cases tomorrow. She can purchase the strawberries for $3 per case and sell c. Maximum likelihood criterion.
them for $8 per case. Jean now needs to decide how many cases to purchase.
d. Bayes’ decision rule.
Jean has checked the store’s records on daily sales of strawberries. On this basis, she
estimates that the prior probabilities are 0.2, 0.4, 0.3, and 0.1 for being able to sell 10, 11, 12, and 9.5. Reconsider Problem 9.4. Warren Buffy decides that Bayes’ decision rule is his most
13 cases of strawberries tomorrow. reliable decision criterion. He believes that 0.1 is just about right as the prior probability of an
improving economy, but is quite uncertain about how to split the remaining probabilities between
a. Develop a decision analysis formulation of this problem by identifying the decision
a stable economy and a worsening economy. Therefore, he now wishes to do sensitivity
alternatives, the states of nature, and the payoff table.
analysis with respect to these latter two prior probabilities.
b. If Jean is dubious about the accuracy of these prior probabilities and so chooses to
ignore them and use the maximax criterion, how many cases of strawberries should
she purchase? page 383
c. How many cases should be purchased if she uses the maximin criterion?
d. How many cases should be purchased if she uses the maximum likelihood criterion? a. Reapply Bayes’ decision rule when the prior probability of a stable economy is 0.3
and the prior probability of a worsening economy is 0.6.
e. How many cases should be purchased according to Bayes’ decision rule?
b. Reapply Bayes’ decision rule when the prior probability of a stable economy is 0.7
f. Jean thinks she has the prior probabilities just about right for selling 10 cases and
and the prior probability of a worsening economy is 0.2.
selling 13 cases, but is uncertain about how to split the prior probabilities for 11 cases
and 12 cases. Reapply Bayes’ decision rule when the prior probabilities of 11 and 12 c. Construct a decision tree by hand for this problem with the original prior probabilities.
cases are (i) 0.2 and 0.5, (ii) 0.3 and 0.4, and (iii) 0.5 and 0.2. A d. Use Analytic Solver to construct and solve a decision tree for this problem with the
9.4.* Warren Buffy is an enormously wealthy investor who has built his fortune through his original prior probabilities.
legendary investing acumen. He currently has been offered three major investments and he A e. In preparation for performing sensitivity analysis, consolidate the data and results on
would like to choose one. The first one is a conservative investment that would perform very the same spreadsheet as the decision tree constructed in part d (as was done in
well in an improving economy and only suffer a small loss in a worsening economy. The second Figure 9.7 for the case study).
is a speculative investment that would perform extremely well in an improving economy but
would do very badly in a worsening economy. The third is a countercyclical investment that
A f. Use the spreadsheet (including the decision tree) obtained in parts d and e to do
would lose some money in an improving economy but would perform well in a worsening parts a and b.
economy. A g. Expanding the spreadsheet as needed, generate a data table that shows which
Warren believes that there are three possible scenarios over the investment Warren should make and the resulting expected profit for the following
prior probabilities of a stable economy: 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9.
lives of these potential investments: (1) an improving economy, (2) a
stable economy, and (3) a worsening economy. He is pessimistic about
 h. For each of the three investments, find the expected profit when the prior probability
of a stable economy is 0 and then when it is 0.9 (with the prior probability of an
improving economy fixed at 0.1). Plot these expected profits on a single graph that
has expected profit as the vertical axis and the prior probability of a stable economy
as the horizontal axis. For each of the three investments, draw a line segment
connecting its two points on this graph to show how its expected profit would vary
with the prior probability of a stable economy. Use this graph to describe how the
choice of the investment depends on the prior probability of a stable economy.
9.6. Read the referenced article that fully describes the management science study a. Which alternative should be chosen under the maxi-max criterion?
summarized in the application vignette presented in Section 9.3. Briefly describe how decision b. Which alternative should be chosen under the maxi-min criterion?
analysis was applied in this study. Then list the various financial and non-financial benefits that
resulted from this study. c. Which alternative should be chosen under the maximum likelihood criterion?
d. Which alternative should be chosen under Bayes’ decision rule?
9.7.* Consider a decision analysis problem whose payoffs (in units of thousands of dollars)
are given by the following payoff table. e. Construct a decision tree by hand for this problem. A f. Use Analytic Solver to
construct and solve a decision tree for this problem.
A g. Perform sensitivity analysis with this decision tree by generating a data table that
shows what happens when the prior probability of S 1 increases in increments of 0.05
from 0.3 to 0.7 while the prior probability of S 3 remains fixed at its original value.
Then use trial and error to estimate the value of the prior probability of S 1 at which
the best alternative changes as this prior probability increases.
A h. Repeat part g when it is the prior probability of S 2 that remains fixed at its original
value.
A i. Repeat part g when it is the prior probability of S 1 that remains fixed at its original
a. Which alternative should be chosen under the maximax criterion? value while the prior probability of S 2 increases in increments of 0.05 from 0 to 0.4.
b. Which alternative should be chosen under the maximin criterion?
j. If you feel that the true probabilities of the states of nature should be within 10
c. Which alternative should be chosen under the maximum likelihood criterion? percent of the given prior probabilities, which alternative would you choose?
d. Which alternative should be chosen under Bayes’ decision rule?
9.9. Dwight Moody is the manager of a large farm with 1,000 acres of arable land. For
e. Construct a decision tree by hand for this problem. A f. Use Analytic Solver to greater efficiency, Dwight always devotes the farm to growing one crop at a time. He now
construct and solve a decision tree for this problem. needs to make a decision on which one of four crops to grow during the upcoming growing
f. Which alternative should be chosen under Bayes’ decision rule? season. For each of these crops, Dwight has obtained the following estimates of crop yields and
net incomes per bushel under various weather conditions.
A e. Use Analytic Solver to construct and solve a decision tree for this problem.
A f. Expanding the spreadsheet containing this decision tree as needed, perform
sensitivity analysis with the decision tree by re-solving when the prior probability of page 384
S 1 is 0.2 and again when it is 0.6.
A g. Now perform this sensitivity analysis systematically by generating a data table that
shows the best alternative (according to Bayes’ decision rule) and the resulting
expected payoff as the prior probability of S 1 increases in increments of 0.04 from
0.2 to 0.6.
9.8. You are given the following payoff table (in units of thousands of dollars) for a decision
analysis problem.

After referring to historical meteorological records, Dwight also has estimated the following
prior probabilities for the weather during the growing season:
 Dry 0.3 d. You are given the opportunity to spend $1,000 to obtain more information about
Moderate 0.5 which state of nature is likely to occur. Given your answer to part b, might it be
Damp 0.2 worthwhile to spend this money?
9.12.* Betsy Pitzer makes decisions according to Bayes’ decision rule. For her current
a. Develop a decision analysis formulation of this problem by identifying the decision problem, Betsy has constructed the following payoff table (in units of dollars).
alternatives, the states of nature, and the payoff table.
A b. Construct a decision tree for this problem and use Bayes’ decision rule to determine
which crop to grow.
A c. Using Bayes’ decision rule, do sensitivity analysis with respect to the prior
probabilities of moderate weather and damp weather (without changing the prior
probability of dry weather) by re-solving when the prior probability of moderate
weather is 0.2, 0.3, 0.4, and 0.6.
9.10. Barbara Miller makes decisions according to Bayes’ decision rule. For her current
problem, Barbara has constructed the following payoff table (in units of hundreds of dollars) and
she now wishes to maximize the expected payoff. a. Which alternative should Betsy choose?
b. Find the expected value of perfect information.
A c. Check your answer in part b by recalculating it with the help of a decision tree.
d. What is the most that Betsy should consider paying to obtain more information about
which state of nature will occur?
9.13. Using Bayes’ decision rule, consider the decision analysis problem having the following
payoff table (in units of thousands of dollars).

The value of x currently is 50, but there is an opportunity to increase x by spending some money
now.
What is the maximum amount Barbara should spend to increase x to 75?
9.11. You are given the following payoff table (in units of thousands of dollars) for a decision
analysis problem.

a. Which alternative should be chosen? What is the resulting expected payoff?

b. You are offered the opportunity to obtain information that will tell you with certainty
whether the first state of nature S 1 will occur. What is the maximum amount you
should pay for the information? Assuming you will obtain the information, how should
this information be used to choose an alternative? What is the resulting expected
payoff (excluding the payment)?
page 385
c. Now repeat part b if the information offered concerns S 2 instead of S 1.
a. According to Bayes’ decision rule, which alternative should be chosen? d. Now repeat part b if the information offered concerns S 3 instead of S 1.
b. Find the expected value of perfect information.
A e. Now suppose that the opportunity is offered to provide information that will tell you
A c. Check your answer in part b by recalculating it with the help of a decision tree. with certainty which state of nature will occur (perfect information). What is the
maximum amount you should pay for the information? Assuming you will obtain the
 information, how should this information be used to choose an alternative? What is b. Assuming the credit-rating organization is not used, use Bayes’ decision rule to
the resulting expected payoff (excluding the payment)? determine which decision alternative should be chosen.
f. If you have the opportunity to do some testing that will give you partial additional c. Find the expected value of perfect information. Does this answer indicate that
information (not perfect information) about the state of nature, what is the maximum consideration should be given to using the credit-rating organization?
amount you should consider paying for this information? d. Assume now that the credit-rating organization is used. Develop a probability tree
9.14. Reconsider the Goferbroke Co. case study, including its analysis in Sections 9.6 and diagram to find the posterior probabilities of the respective states of nature for each
9.7. With the help of the consulting geologist, Jennifer Flyer now has obtained some historical of the three possible credit evaluations of this potential customer.
data that provides more precise information than Max could supply on the likelihood of obtaining T e. Use the corresponding Excel template to obtain the answers for part d.
favorable seismic soundings on similar tracts of land. Specifically, when the land contains oil,
f. Draw the decision tree for this entire problem by hand. Use this decision tree to
favorable seismic soundings are obtained 80 percent of the time. This percentage changes to 40
determine Vincent’s optimal policy.
percent when the land is dry.
A g. Use Analytic Solver to construct and solve this decision tree.
a. Revise Figure 9.12 to find the new posterior probabilities.
A h. Find the expected value of sample information. If the fee for using the credit-rating
T b. Use the corresponding Excel template to check your answers in part a.
organization is open to negotiation, how large can the fee be before the use of this
c. Revise Figure 9.16 to find the new decision tree. What is the resulting optimal organization would no longer be worthwhile?
policy?
9.17. You are given the following payoff table (in units of dollars).
A d. Use Analytic Solver to construct and solve this new decision tree.
9.15. Read the referenced article that fully describes the management science study
summarized in the application vignette presented in Section 9.7. Briefly describe how decision
analysis was applied in this study. Then list the various financial and non-financial benefits that
resulted from this study.
9.16.* Vincent Cuomo is the credit manager for the Fine Fabrics Mill. He is currently faced
with the question of whether to extend $100,000 of credit to a potential new customer, a dress
manufacturer. Vincent has three categories for the creditworthiness of a company—poor risk,
average risk, and good risk—but he does not know which category fits this potential customer.
Experience indicates that 20 percent of companies similar to this dress manufacturer are poor
risks, 50 percent are average risks, and 30 percent are good risks. If credit is extended, the You have the option of paying $100 to have research done to better predict which state of
expected profit is –$15,000 for poor risks, $10,000 for average risks, and $20,000 for good risks. nature will occur. When the true state of nature is S 1, the research will accurately predict S 1 60
If credit is not extended, the dress manufacturer will turn to another mill. Vincent is able to percent of the time (but will inaccurately predict S 2 40 percent of the time). When the true state
consult a credit-rating organization for a fee of $5,000 per company evaluated. For companies of nature is S 2, the research will accurately predict S 2 80 percent of the time (but will
whose actual credit records with the mill turn out to fall into each of the three categories, the inaccurately predict S 1 20 percent of the time).
following table shows the percentages that were given each of the three possible credit
evaluations by the credit-rating organization. a. Given that the research is not done, use Bayes’ decision rule to determine which
decision alternative should be chosen.
A b. Use a decision tree to help find the expected value of perfect information. Does this
answer indicate that it might be worthwhile to do the research?
page 386
c. Given that the research is done, find the joint probability of each of the
following pairs of outcomes: (i) the state of nature is S 1 and the research predicts
S 1, (ii) the state of nature is S 1 and the research predicts S 2, (iii) the state of nature
is S 2 and the research predicts S 1, and (iv) the state of nature is S 2 and the research
a. Develop a decision analysis formulation of this problem by identifying the decision predicts S 2.
alternatives, the states of nature, and the payoff table when the credit-rating
d. Find the unconditional probability that the research predicts S 1. Also find the
organization is not used.
unconditional probability that the research predicts S 2.
 e. Given that the research is done, use your answers in parts c and d to determine the e. Use Analytic Solver to construct and solve the decision tree for this entire problem.
posterior probabilities of the states of nature for each of the two possible predictions
9.20. The Hit-and-Miss Manufacturing Company produces items that have a probability p of
of the research.
being defective. These items are produced in lots of 150. Past experience indicates that p for
T f. Use the corresponding Excel template to obtain the answers for part e. each item in a lot is either 0.05 or 0.25. Furthermore, in 80 percent of the lots produced, p
g. Given that the research predicts S 1, use Bayes’ decision rule to determine which equals 0.05 (so p equals 0.25 in 20 percent of the lots). These items are then used in an
assembly, and ultimately their quality is determined before the final assembly leaves the plant.
decision alternative should be chosen and the resulting expected payoff.
Initially the company can either screen each item in a lot at a cost of $10 per item and replace
h. Repeat part g when the research predicts S 2. defective items or use the items directly without screening. If the latter action is chosen, the
i. Given that research is done, what is the expected payoff when using Bayes’ decision cost of rework is ultimately $100 per defective item. Because screening requires scheduling of
rule? inspectors and equipment, the decision to screen or not screen must be made two days before
the screening is to take place. However, one item can be taken from the lot and sent to a
j. Use the preceding results to determine the optimal policy regarding whether to do the
laboratory for inspection, after which its quality (defective or nondefective) can be reported
research and which decision alternative should be chosen.
before the screen/no-screen decision must be made. The cost of this initial inspection is $125.
A k. Construct and solve the decision tree to show the analysis for the entire problem.
a. Develop a decision analysis formulation of this problem by identifying the decision
(Using Analytic Solver is optional.)
alternatives, the states of nature, and the payoff table if the single item is not
9.18. An athletic league does drug testing of its athletes, 10 percent of whom use drugs. The inspected in advance.
test, however, is only 95 percent reliable. That is, a drug user will test positive with probability b. Assuming the single item is not inspected in advance, use Bayes’ decision rule to
0.95 and negative with probability 0.05, whereas a non-user will test negative with probability determine which decision alternative should be chosen.
0.95 and positive with probability 0.05.
c. Find the expected value of perfect information. Does this answer indicate that
Develop a probability tree diagram to determine the posterior probability of each of the
consideration should be given to inspecting the single item in advance?
following outcomes of testing an athlete.
T d. Assume now that the single item is inspected in advance. Find the posterior
a. The athlete is a drug user, given that the test is positive.
probabilities of the respective states of nature for each of the two possible outcomes
b. The athlete is not a drug user, given that the test is positive. of this inspection.
c. The athlete is a drug user, given that the test is negative. A e. Construct and solve the decision tree for this entire problem.
d. The athlete is not a drug user, given that the test is negative.
A f. Find the expected value of sample information. If the cost of using the laboratory to
T e. Use the corresponding Excel template to check your answers in the preceding parts. inspect the single item in advance is open to negotiation, how large can the cost of
9.19. Management of the Telemore Company is considering developing and marketing a new using the laboratory be and still be worthwhile?
product. It is estimated to be twice as likely that the product would prove to be successful as 9.21.* Silicon Dynamics has developed a new computer chip that will enable it to begin
unsuccessful. If it were successful, the expected profit would be $1,500,000. If unsuccessful, producing and marketing a personal computer if it so desires. Alternatively, it can sell the rights
the expected loss would be $1,800,000. A marketing survey can be conducted at a cost of to the computer chip for $15 million. If the company chooses to build computers, the profitability
$100,000 to predict whether the product would be successful. Past experience with such of the venture depends on the company’s ability to market the computer during the first year. It
surveys indicates that successful products have been predicted to be successful 80 percent of has sufficient access to retail outlets that it can guarantee sales of 10,000 computers. On the
the time, whereas unsuccessful products have been predicted to be unsuccessful 70 percent of other hand, if this computer catches on, the company can sell 100,000 machines. For analysis
the time. purposes, these two levels of sales are taken to be the two possible outcomes of marketing the
a. Develop a decision analysis formulation of this problem by identifying the decision computer, but it is unclear what their prior probabilities are. The cost of setting up the assembly
alternatives, the states of nature, and the payoff table when the market survey is not line is $6 million. The difference between the selling price and the variable cost of each
conducted. computer is $600.
b. Assuming the market survey is not conducted, use Bayes’ decision rule to determine
which decision alternative should be chosen.
page 387
c. Find the expected value of perfect information. Does this answer indicate that
consideration should be given to conducting the market survey? a. Develop a decision analysis formulation of this problem by identifying the decision
T d. Assume now that the market survey is conducted. Find the posterior probabilities of alternatives, the states of nature, and the payoff table.
the respective states of nature for each of the two possible predictions from the b. Construct a decision tree for this problem by hand.
market survey.
A
A
 c. Assuming the prior probabilities of the two levels of sales are both 0.5, use Analytic
Solver to construct and solve this decision tree. According to this analysis, which
decision alternative should be chosen?
9.22.* Reconsider Problem 9.21. Management of Silicon Dynamics now is considering doing
full-fledged market research at an estimated cost of $1 million to predict which of the two levels
of demand is likely to occur. Previous experience indicates that such market research is correct
two-thirds of the time.
a. Find the expected value of perfect information for this problem.
b. Does the answer in part a indicate that it might be worthwhile to perform this market
research?
c. Develop a probability tree diagram to obtain the posterior probabilities of the two
levels of demand for each of the two possible outcomes of the market research.
T d. Use the corresponding Excel template to check your answers in part c.
R 9.23.* Reconsider Problem 9.22. The management of Silicon Dynamics now wants to see a
decision tree displaying the entire problem.
A
a. Use Analytic Solver to construct and solve this decision tree.
a. Analyze this decision tree to obtain the optimal policy.
b. Find the expected value of sample information. How large can the cost of doing full-
fledged market research be and still be worthwhile? A b. Use Analytic Solver to construct and solve the same decision tree.
c. Assume now that the estimate of $1 million for the cost of doing full-fledged market 9.25. You are given the following decision tree, with the probabilities at event nodes shown in
research is correct but that there is some uncertainty in the financial data ($15 parentheses and with the payoffs at terminal points shown on the right. Analyze this decision
million, $6 million, and $600) stated in Problem 9.21. Each could vary from its base tree to obtain the optimal policy.
value by as much as 10 percent. For each one, perform sensitivity analysis to find
what would happen if its value were at either end of this range of variability (without
any change in the other two pieces of data). Then do the same for the eight cases
where all these pieces of data are at one end or the other of their ranges of
variability.
9.24. You are given the following decision tree, where the numbers in parentheses are
probabilities and the numbers on the right are payoffs at these terminal points.

9.26.* The Athletic Department of Leland University is considering whether to hold an

extensive campaign next year to raise funds for a new athletic field. The response to the
campaign depends heavily on the success of the football team this fall. In the past, the football
team has had winning seasons 60 percent of the time. If the football team has a winning season
(W) this fall, then many of the alumni will contribute and the campaign will raise $3 million. If
the team has a losing season (L), few will contribute and the campaign will lose $2 million. If no
campaign is undertaken, no costs are incurred. On September 1, just before the football season
begins, the Athletic Department needs to make its decision about whether to hold the campaign The probabilities of growth, recession, and depression for the first year are 0.7, 0.3, and 0,
next year. respectively. If growth occurs in the first year, these probabilities remain the same for the
second year. However, if a recession occurs in the first year, these probabilities change to 0.2,
0.7, and 0.1, respectively, for the second year.
page 388 a. Construct by hand the decision tree for this problem and then analyze the decision
tree to identify the optimal policy.
a. Develop a decision analysis formulation of this problem by identifying the decision
alternatives, the states of nature, and the payoff table. A b. Use Analytic Solver to construct and solve the decision tree.
b. According to Bayes’ decision rule, should the campaign be undertaken? 9.28. On Monday, a certain stock closed at $10 per share. Before the stock market opens on
c. What is the expected value of perfect information? Tuesday, you expect the stock to close at $9, $10, or $11 per share, with respective probabilities
0.3, 0.3, and 0.4. Looking ahead to Wednesday, you expect the stock to close 10 percent lower,
d. A famous football guru, William Walsh, has offered his services to help evaluate
unchanged, or 10 percent higher than Tuesday’s close, with the following probabilities.
whether the team will have a winning season. For $100,000, he will carefully
evaluate the team throughout spring practice and then throughout preseason
workouts. William then will provide his prediction on September 1 regarding what
kind of season, W or L, the team will have. In similar situations in the past when
evaluating teams that have winning seasons 50 percent of the time, his predictions
have been correct 75 percent of the time. Considering that this team has more of a
winning tradition, if William predicts a winning season, what is the posterior
probability that the team actually will have a winning season? What is the posterior
probability of a losing season? If William predicts a losing season instead, what is the
posterior probability of a winning season? Of a losing season? Show how these Early on Tuesday, you are directed to buy 100 shares of the stock before Thursday. All
answers are obtained from a probability tree diagram. purchases are made at the end of the day, at the known closing price for that day, so your only
T e. Use the corresponding Excel template to obtain the answers requested in part d. options are to buy at the end of Tuesday or at the end of Wednesday. You wish to determine the
optimal strategy for whether to buy on Tuesday or defer the purchase until Wednesday, given
f. Draw the decision tree for this entire problem by hand. Analyze this decision tree to
the Tuesday closing price, to minimize the expected purchase price.
determine the optimal policy regarding whether to hire William and whether to
undertake the campaign. a. Develop and evaluate a decision tree by hand for determining the optimal strategy.
A g. Use Analytic Solver to construct and solve this decision tree. A b. Use Analytic Solver to construct and solve the decision tree.
A h. Find the expected value of sample information. If the fee for hiring William Walsh is A 9.29. Jose Morales manages a large outdoor fruit stand in one of the less affluent
open to negotiation, how large can William’s fee be and still be worthwhile? neighborhoods of San Jose, California. To replenish his supply, Jose buys boxes of fruit early
each morning from a grower south of San Jose. About 90 percent of the boxes of fruit turn out
9.27. The comptroller of the Macrosoft Corporation has $100 million of excess funds to
to be of satisfactory quality, but the other 10 percent are unsatisfactory. A satisfactory box
invest. She has been instructed to invest the entire amount for one year in either stocks or bonds
contains 80 percent excellent fruit and will earn $200 profit for Jose. An unsatisfactory box
(but not both) and then to reinvest the entire fund in either stocks or bonds (but not both) for one
contains 30 percent excellent fruit and will produce a loss of $1,000. Before Jose decides to
year more. The objective is to maximize the expected monetary value of the fund at the end of
accept a box, he is given the opportunity to sample one piece of fruit to test whether it is
the second year.
excellent. Based on that sample, he then has the option of rejecting the box without paying for
The annual rates of return on these investments depend on the economic environment, as
it. Jose wonders (1) whether he should continue buying from this grower, (2) if so, whether it is
shown in the following table.
worthwhile sampling just one piece of fruit from a box, and (3) if so, whether he should be
accepting or rejecting the box based on the outcome of this sampling.
Use Analytic Solver (and the Excel template for posterior probabilities) to construct and
solve the decision tree for this problem.
9.30.* The Morton Ward Company is considering the introduction of a new product that is
believed to have a 50–50 chance of being successful. One option is to try out the product in a
test market, at an estimated cost of $2 million, before making the introduction decision. Past
experience shows that ultimately successful products are approved in the test market 80
percent of the time, whereas ultimately unsuccessful products are approved in the test market
only 25 percent of the time. If the product is successful, the net profit to the company will be
$40 million; if unsuccessful, the net loss will be $15 million.
 Chelsea feels that her chance of winning the nomination depends largely on having
page 389 substantial funds available after the Super Tuesday primaries to carry on a vigorous campaign
the rest of the way. Therefore, she wants to choose the strategy (whether to run in the New
a. Discarding the test market option, develop a decision analysis formulation of the Hampshire primary and then whether to run in the Super Tuesday primaries) that will maximize
problem by identifying the decision alternatives, states of nature, and payoff table. her expected funds after these primaries.
Then apply Bayes’ decision rule to determine the optimal decision alternative.
a. Construct and solve the decision tree for this problem.
b. Find the expected value of perfect information.
b. There is some uncertainty in the estimates of a gain of $16 million or a loss of $10
A c. Now including the option of trying out the product in a test market, use Analytic million depending on the showing on Super Tuesday. Either amount could differ from
Solver (and the Excel template for posterior probabilities) to construct and solve the this estimate by as much as 25 percent in either direction. For each of these two
decision tree for this problem. financial figures, perform sensitivity analysis to check how the results in part a would
A d. Find the expected value of sample information. How large can the cost of trying out change if the value of the financial figure were at either end of this range of
the product in a test market be and still be worthwhile to do? variability (without any change in the value of the other financial figure). Then do the
same for the four cases where both financial figures are at one end or the other of
A e. Assume now that the estimate of $2 million for the cost of trying out the product in a their ranges of variability.
test market is correct. However, there is some uncertainty in the stated profit and
loss figures ($40 million and $15 million). Either could vary from its base by as much A 9.32. The executive search being conducted for Western Bank by Headhunters Inc. may
as 25 percent in either direction. For each of these two financial figures, perform finally be bearing fruit. The position to be filled is a key one—Vice President for Information
sensitivity analysis to check how the results in part c would change if the value of the Processing— because this person will have responsibility for developing a state-of-the-art
financial figure were at either end of this range of variability (without any change in management information system that will link together Western’s many branch banks. However,
the value of the other financial figure). Then do the same for the four cases where Headhunters feels it has found just the right person, Matthew Fenton, who has an excellent
both financial figures are at one end or the other of their ranges of variability. record in a similar position for a mid-sized bank in New York.
After a round of interviews, Western’s president believes that Matthew has a probability of
A 9.31. Chelsea Bush is an emerging candidate for her party’s nomination for president of the
0.7 of designing the management information system successfully. If Matthew is successful, the
United States. She now is considering whether to run in the high-stakes Super Tuesday
company will realize a profit of $2 million (net of Matthew’s salary, training, recruiting costs, and
primaries. If she enters the Super Tuesday (S.T.) primaries, she and her advisers believe that
expenses). If he is not successful, the company will realize a net loss of $600,000.
she will either do well (finish first or second) or do poorly (finish third or worse) with
probabilities 0.4 and 0.6, respectively. Doing well on Super Tuesday will net the candidate’s For an additional fee of $40,000, Headhunters will provide a detailed investigative process
campaign approximately $16 million in new contributions, whereas a poor showing will mean a (including an extensive background check, a battery of academic and psychological tests, etc.)
loss of $10 million after numerous TV ads are paid for. Alternatively, she may choose not to run that will further pinpoint Matthew’s potential for success. This process has been found to be 90
at all on Super Tuesday and incur no costs. percent reliable, that is, a candidate who would successfully design the management information
system will pass the test with probability 0.9, and a candidate who would not successfully design
Chelsea’s advisors realize that her chances of success on Super the system will fail the test with probability 0.9.
Tuesday may be affected by the outcome of the smaller New Western’s top management needs to decide whether to hire Matthew and whether to have
Hampshire (N.H.) primary occurring three weeks before Super Headhunters conduct the detailed investigative process before making this decision.

Tuesday. Political analysts feel that the results of New Hampshire’s a. Construct and solve the decision tree for this problem to identify the optimal policy.
primary are correct two-thirds of the time in predicting the results of b. Now suppose that Headhunters’s fee for administering its detailed investigative
process is negotiable. What is the maximum amount that Western Bank should pay?
the Super Tuesday primaries. Among Chelsea’s advisers is a decision
9.33. Reconsider the Goferbroke Co. case study, including the application of utilities in
analysis expert who uses this information to calculate the following Section 9.9. Max Flyer now has decided that, given the company’s precarious financial situation,
probabilities: he needs to take a much more risk-averse approach to the problem. Therefore, he has revised
the utilities given in Table 9.9 as follows: U(−130) = 0, U(−100) = 0.07, U(60) = 0.40, U(90) =
P(Chelsea does well in S.T. primaries, given she does well in N.H.) = 0.45, U(670) = 0.99, and U(700) = 1.
P(Chelsea does well in S.T. primaries, given she does poorly in N.H.) = ¼

P(Chelsea does well in N.H. primary) =

page 390
The cost of entering and campaigning in the New Hampshire primary is estimated to be $1.6 a. Analyze the revised decision tree corresponding to Figure 9.23 by hand to obtain the
million. new optimal policy.
A b. Use Analytic Solver to construct and solve this revised decision tree. 9.38. You are given the following payoff table.
9.34.* You live in an area that has a possibility of incurring a massive earthquake, so you are
considering buying earthquake insurance on your home at an annual cost of $180. The
probability of an earthquake damaging your home during one year is 0.001. If this happens, you
estimate that the cost of the damage (fully covered by earthquake insurance) will be $160,000.
Your total assets (including your home) are worth $250,000.
a. Apply Bayes’ decision rule to determine which alternative (take the insurance or not)
maximizes your expected assets after one year.
b. You now have constructed a utility function that measures how much you value
having total assets worth x dollars (x ≥ 0). This utility function is U ( x ) .
Compare the utility of reducing your total assets next year by the cost of the a. Assume that your utility function for the payoffs is . Plot the expected
earthquake insurance with the expected utility next year of not taking the earthquake utility of each decision alternative versus the value of p on the same graph. For each
insurance. Should you take the insurance? decision alternative, find the range of values of p over which this alternative
9.35. For your graduation present from college, your parents are offering you your choice of maximizes the expected utility.
two alternatives. The first alternative is to give you a money gift of $19,000. The second A b. Now assume that your utility function is the exponential utility function with a risk
alternative is to make an investment in your name. This investment will quickly have the tolerance of R = 50. Use Analytic Solver to construct and solve the resulting decision
following two possible outcomes. tree in turn for p = 0.25, p = 0.5, and p = 0.75.
Outcome Probability A 9.39. Dr. Switzer has a seriously ill patient but has had trouble diagnosing the specific
Receive $10,000 0.3 cause of the illness. The doctor now has narrowed the cause down to two alternatives: disease
A or disease B. Based on the evidence so far, she feels that the two alternatives are equally
Receive $30,000 0.7 likely.
Beyond the testing already done, there is no test available to determine if the cause is
disease B. One test is available for disease A, but it has two major problems. First, it is very
Your utility for receiving M thousand dollars is given by the utility function expensive. Second, it is somewhat unreliable, giving an accurate result only 80 percent of the
. Which choice should you make to maximize expected utility? time. Thus, it will give a positive result (indicating disease A) for only 80 percent of patients
9.36. Reconsider Problem 9.35. You now are uncertain about what your true utility function who have disease A, whereas it will give a positive result for 20 percent of patients who
for receiving money is, so you are in the process of constructing this utility function by using the actually have disease B instead.
equivalent lottery method and units of thousands of dollars. You have concluded that you are Disease B is a very serious disease with no known treatment. It is sometimes fatal, and
indifferent between the two alternatives offered to you by your parents. Use this information to those who survive remain in poor health with a poor quality of life thereafter. The prognosis is
find U(19) after setting U(10) = 0 and U(30) = 1. similar for victims of disease A if it is left untreated. However, there is a fairly expensive
9.37. You wish to construct your personal utility function U(M) for receiving M thousand treatment available that eliminates the danger for those with disease A, and it may return them
dollars. After setting U(0) = 0, you next set U(10) = 1 as your utility for receiving $10,000. You to good health. Unfortunately, it is a relatively radical treatment that always leads to death if the
next want to find U(1) and then U(5). patient actually has disease B instead.
a. You offer yourself the following two hypothetical alternatives: The probability distribution for the prognosis for this patient is given for each case in the
following table, where the column headings (after the first one) indicate the disease for the
patient.

page 391
You then ask yourself the question: What value of p makes you indifferent between
these two alternatives? Your answer is p = 0.125. Find U(1) by using the equivalent
lottery method.
b. You next repeat part a except for changing the second alternative to definitely
receiving $5,000. The value of p that makes you indifferent between these two
alternatives now is p = 0.5625. Find U(5).
c. Repeat parts a and b, but now use your personal choices for p.
 a terminal branch. The resulting utilities for the various terminal branches are shown to the right
of the decision tree.
Use these utilities to analyze the decision tree. Then determine the value of x for which the
decision maker is indifferent between decision alternatives A1 and A2.
A 9.41. Reconsider the Goferbroke Co. case study when using utilities, as presented in
Section 9.9.
a. Beginning with the decision tree shown in Figure 9.23 (available in one of this
chapter’s Excel files), prepare to perform sensitivity analysis by expanding and
organizing the spreadsheet to (1) consolidate the data and results in one section and
The patient has assigned the following utilities to the possible outcomes. (2) incorporate the Excel template for posterior probabilities in another section
In addition, these utilities should be incremented by −2 if the patient incurs the cost of the (similar to what was done in Figure 9.18).
test for disease A and by −1 if the patient (or the patient’s estate) incurs the cost of the b. Perform sensitivity analysis by re-solving the decision tree (after using the Excel
treatment for disease A. template for posterior probabilities to revise these probabilities) when the prior
Use decision analysis with a complete decision tree to determine if probability of oil is changed in turn to 0.15, 0.2, 0.3, and 0.35.

the patient should undergo the test for disease A and then how to
proceed (receive the treatment for disease A?) to maximize the
patient’s expected utility.
9.40. Consider the following decision tree, where the probabilities for each event node are
shown in parentheses. Case 9-1
Who Wants to Be a Millionaire?
You are a contestant on “Who Wants to Be a Millionaire?” You already have answered the $250,000
question correctly and now must decide if you would like to answer the $500,000 question. You can choose
to walk away at this point with $250,000 in winnings or you may decide to answer the $500,000 question. If
you answer the $500,000 question correctly, you can then choose to walk away with $500,000 in winnings
or go on and try to answer the $1,000,000 question. If you answer the $1,000,000 question correctly, the
game is over and you win $1,000,000. If you answer either the $500,000 or the $1,000,000 question
incorrectly, the game is over immediately and you take home “only” $32,000.

page 392
A feature of the game “Who Wants to Be a Millionaire?” is that you have three “lifelines”—namely
“50–50,” “ask the audience,” and “phone a friend.” At this point (after answering the $250,000 question),
Outcome Utility
you already have used two of these lifelines, but you have the “phone a friend” lifeline remaining. With this
Die 0 option, you may phone a friend to obtain advice on the correct answer to a question before giving your
Survive with poor health 10 answer. You may use this option only once (i.e., you can use it on either the $500,000 question or the
$1,000,000 question, but not both). Since some of your friends are smarter than you are, “phone a friend”
Return to good health 30
significantly improves your odds for answering a question correctly. Without “phone a friend,” if you choose
to answer the $500,000 question you have a 65 percent chance of answering correctly, and if you choose to
The dollar amount given next to each branch is the cash flow generated along that branch, answer the $1,000,000 question you have a 50 percent chance of answering correctly (the questions get
where these intermediate cash flows add up to the total net cash flow shown to the right of progressively more difficult). With “phone a friend,” you have an 80 percent chance of answering the
each terminal branch. (The unknown amount for the top branch is represented by the variable $500,000 question correctly and a 65 percent chance of answering the $1,000,000 question correctly.
x.) The decision maker has a utility function U(y) = y1/3 where y is the total net cash flow after a. Use Analytic Solver to construct and solve a decision tree to decide what to do. What is the best course
of action, assuming your goal is to maximize your expected winnings?
b. Use the equivalent lottery method to determine your personal utility function (in particular, your utility b. Use Analytic Solver to develop and solve a decision tree to help University Toys decide the best course
values for all of the possible payoffs in the game). of action and the expected payoff.
c. Re-solve the decision tree, replacing the payoffs with your utility values, to maximize your expected c. Now suppose that University Toys is uncertain of the probability that the LSPAFs will enter the market
utility. Does the best course of action change? before the test marketing would be completed (if it were done). How would you expect the expected
payoff to vary as the probability that the LSPAFs will enter the market changes?
d. Generate a data table that shows how the expected payoff and the test marketing decision changes as
the probability that the LSPAFs will enter the market varies from 0 percent to 100 percent (at 10 percent
Case 9-2 increments).
e. At what probability does the test marketing decision change?
University Toys and the Business Professor Action
Figures page 393
University Toys has developed a brand new product line—a series of Business Professor Action Figures
(BPAFs) featuring likenesses of popular professors at the local business school. Management needs to
decide how to market the dolls. Case 9-3
One option is to immediately ramp up production and simultaneously launch an ad campaign in the
university newspaper. This option would cost $1,000. Based on past experience, new action figures either
take off and do well or fail miserably. Hence, the prediction is for one of two possible outcomes—total
Brainy Business
sales of 2,500 units or total sales of only 250 units. University Toys receives revenue of $2 per unit sold.
While El Niño is pouring its rain on northern California, Charlotte Rothstein, CEO, major shareholder,
Management currently thinks that there is about a 50 percent chance that the product will do well (sell
and founder of Cerebrosoft, sits in her office, contemplating the decision she faces regarding her
2,500 units) and a 50 percent chance that it will do poorly (sell 250 units).
company’s newest proposed product, Brainet. This has been a particularly difficult decision. Brainet might
Another option is to test-market the product. The company could build some units, put up a display in
catch on and sell very well. However, Charlotte is concerned about the risk involved. In this competitive
the campus bookstore, and see how they sell without any further advertising. This would require less capital
market, marketing Brainet also could lead to substantial losses. Should she go ahead anyway and start the
for the production run and no money for advertising. Again, the prediction is for one of two possible
marketing campaign? Or just abandon the product? Or perhaps buy additional marketing research
outcomes for this test market, namely, the product will either do well (sell 200 units) or do poorly (sell 20
information from a local market research company before deciding whether to launch the product? She has
units). The cost for this option is estimated to be $100. University Toys receives revenue of $2 per unit sold
to make a decision very soon and so, as she slowly drinks from her glass of high-protein-power multivitamin
for the test market as well. The company has often test marketed toys in this manner. Products that sell
juice, she reflects on the events of the past few years.
well when fully marketed have also sold well in the test market 80 percent of the time. Products that sell
poorly when fully marketed also sell poorly in the test market 60 percent of the time. Cerebrosoft was founded by Charlotte and two friends after they had graduated from business school.
The company is located in the heart of Silicon Valley. Charlotte and her friends managed to make money in
There is a complication with the test market option, however. A rival toy manufacturer is rumored to be
their second year in business and have continued to do so every year since. Cerebrosoft was one of the
considering the development of Law School Professor Action Figures (LSPAF). After doing the test
first companies to sell software online and to develop PC-based software tools for the multimedia sector.
marketing, if University Toys decides to go ahead and ramp up production and fully market the BPAF, the
Two of the products generate 80 percent of the company’s revenues: Audiatur and Videatur. Each product
cost of doing so would still be $1,000. However, the sales prospects depend upon whether LSPAF has been
has sold more than 100,000 units during the past year. Customers can download a trial version of the
introduced into the market or not. If LSPAF has not entered the market, then the sales prospects will be the
software, test it, and if they are satisfied with what they see, they can purchase the product. Both products
same as described above (i.e., 2,500 units if BPAF does well, or 250 units if BPAF does poorly, on top of
are priced at $75.95 and are sold exclusively online.
any units sold in the test market). However, if LSPAF has been introduced, the increased competition will
Charlotte is interrupted in her thoughts by the arrival of Jeannie Korn. Jeannie is in charge of marketing
diminish sales of BPAF. In particular, management expects in this case to sell 1,000 if BPAF does well or
for online products and Brainet has had her particular attention from the beginning. She is more than ready
100 units if it does poorly (on top of any units sold in the test market). Note that the probability of BPAF
to provide the advice that Charlotte has requested. “Charlotte, I think we should really go ahead with
doing well or doing poorly is not affected by LSPAF, just the final sales totals of each possibility. The
Brainet. The software engineers have convinced me that the current version is robust and we want to be
probability that LSPAF will enter the market before the end of the test market is 20 percent. On the other
on the market with this as soon as possible! From the data for our product launches during the past two
hand, if University Toys markets BPAF immediately, they are guaranteed to beat the LSPAF to market
years, we can get a rather reliable estimate of how the market will respond to the new product, don’t you
(thus making LSPAF a nonfactor).
think? And look!” She pulls out some presentation slides. “During that time period we launched 12 new
a. Suppose that the test marketing is done. Use the Posterior Probabilities template to determine the products altogether and 4 of them sold more than 30,000 units during the first six months alone! Even
likelihood that the BPAF would sell well if fully marketed, given that it sells well in the test market and better: The last two we launched even sold more than 40,000 copies during the first two quarters!”
then given that it sells poorly in the test market. Charlotte knows these numbers as well as Jeannie does. After all, two of these launches have been
products she herself helped to develop. But she feels uneasy about this particular product launch. The
company has grown rapidly during the past three years and its financial capabilities are already rather
stretched. A poor product launch for Brainet would cost the company a lot of money, something that isn’t
available right now due to the investments Cerebrosoft has recently made.
Later in the afternoon, Charlotte meets with Reggie Ruffin, a jack of all trades and the production
manager. Reggie has a solid track record in his field and Charlotte wants his opinion on the Brainet project.
“Well, Charlotte, quite frankly, I think that there are three main factors that are relevant to the success
of this project: competition, units sold, and cost—ah, and, of course, our pricing. Have you decided on the
price yet?”
“I am still considering which of the three strategies would be most beneficial to us. Selling for $50.00 TABLE 3
and trying to maximize revenues—or selling for $30.00 and trying to maximize market share. Of course, Probability Distribution of Unit Sales, Given a Low Price ($30)
there is still your third alternative; we could sell for $40.00 and try to do both.”
At this point, Reggie focuses on the sheet of paper in front of him. “And I still believe that the $40.00
alternative is the best one. Concerning the costs, I checked the records; basically we have to amortize the
development costs we incurred for Brainet. So far we have spent $800,000 and we expect to spend
another $50,000 per year for support.” Reggie next hands a report to Charlotte. “Here we have some data
on the industry. I just received that yesterday, hot off the press. Let’s see what we can learn about the
industry here.” He shows Charlotte some of the highlights. Reggie then agrees to compile the most relevant
information contained in the report and have it ready for Charlotte the following morning. It takes him long
into the night to gather the data from the pages of the report, but in the end he produces three tables, one
for each of the three alternative pricing strategies. Each table shows the corresponding probability of “How much do they want for the study?”
various amounts of sales given the level of competition (severe, moderate, or weak) that develops from “I knew you’d ask that, Reggie. They want $10,000, and I think it’s a fair deal.”
other companies. At this point, Charlotte steps into the conversation. “Do we have any data on the quality of the work of
The next morning, Charlotte is sipping from another power drink. Jeannie and Reggie will be in her this marketing research company?”
office any moment now and, with their help, she will have to decide what to do with Brainet. Should they “Yes, I do have some reports here. After analyzing them, I have come to the conclusion that the
launch the product? If so, at what price? predictions of the marketing research company are pretty good: Given that the competition turned out to be
When Jeannie and Reggie enter the office, Jeannie immediately bursts out: “Guys, I just spoke to our severe, they predicted it correctly 80 percent of the time, while 15 percent of the time they predicted
marketing research company. They say that they could do a study for us about the competitive situation for moderate competition in that setting, with a prediction of weak competition only 5 percent of the time.
the introduction of Brainet and deliver the results within a week.” Given that the competition turned out to be moderate, they predicted severe competition 15 percent of the
time and moderate competition 80 percent of the time, with a prediction of weak competition just 5 percent
of the time. Finally, for the case of weak competition, the numbers were 90 percent of the time a correct
TABLE 1 prediction, 7 percent of the time a ‘moderate’ prediction and 3 percent of the time a ‘severe’ prediction.”
Probability Distribution of Unit Sales, Given a High Price ($50) Charlotte feels that all these numbers are too much for her. “Don’t we have a simple estimate of how
the market will react?”
“Some prior probabilities, you mean? Sure, from our past experience, the likelihood of facing severe
competition is 20 percent, whereas it is 70 percent for moderate competition and 10 percent for weak
competition,” says Jeannie, her numbers always ready when needed.
All that is left to do now is to sit down and make sense of all this . . .

a. For the initial analysis, ignore the opportunity of obtaining more information by hiring the marketing
research company. Identify the decision alternatives and the states of nature. Construct the payoff table.
Then formulate the decision problem in a decision tree. Clearly distinguish between decision and event
page 394 nodes and include all the relevant data.
b. What is Charlotte’s decision if she uses the maximum likelihood criterion?
TABLE 2 c. What is Charlotte’s decision if she uses Bayes’ decision rule?
Probability Distribution of Unit Sales, Given a Medium Price ($40) d. Now consider the possibility of doing the market research. Develop the corresponding decision tree.
Calculate the relevant probabilities and analyze the decision tree. Should Cerebrosoft pay the $10,000 for
the marketing research? What is the overall optimal policy?
 Max, Julie’s assistant, just shakes his head and murmurs, “those techno-nerds. . . .”
Case 9-4 Brian starts to explain: “Sorry for the confusion. Let’s just go through it again, step by step.”
“Good idea—and perhaps make smaller steps!” Julie obviously dislikes Brian’s style of presentation.
Smart Steering Support “OK, the first decision we are facing is whether to invest in research for the road scanning device.”
“How much would that cost us?” asks Marc.
On a sunny May morning, Marc Binton, CEO of Bay Area Automobile Gadgets (BAAG), enters the “Our estimated budget for this is $300,000. Once we invest that money, the outcome of the research
conference room on the 40th floor of the Gates building in San Francisco, where BAAG’s offices are effort is somewhat uncertain. Our engineers assess the probability of successful research at 80 percent.”
located. The other executive officers of the company have already gathered. The meeting has only one “That’s a pretty optimistic success rate, don’t you think?” Julie remarks sarcastically. She still
item on its agenda: planning a research and development project to develop a new driver support system remembers the disaster with Brian’s last project, the fingerprint-based car-security-system. After spending
(DSS). Brian Huang, manager of Research and Development, is walking around nervously. He has to half a million dollars, the development engineers concluded that it would be impossible to produce the
inform the group about the R&D strategy he has developed for the DSS. Marc has identified DSS as the security system at an attractive price.
strategic new product for the company. Julie Aker, vice president of Marketing, will speak after Brian. She Brian senses Julie’s hostility and shoots back: “In engineering, we are quite accustomed to these
will give detailed information about the target segment, expected sales, and marketing costs associated with success rates—something we can’t say about marketing. . . .”
the introduction of the DSS. “What would be the next step?” intervenes Marc.
BAAG builds electronic nonaudio equipment for luxury cars. Founded by a group of Stanford “Hm, sorry. If the research is not successful, then we can only sell the DSS in its current form.”
graduates, the company sold its first product—a car routing system relying on a technology called Global “The profit estimate for that scenario is $2 million,” Julie throws in.
Positioning Satellites (GPS)—a few years ago. Such routing systems help drivers to find directions to their “If, however, the research effort is successful, then we will have to make another decision, namely,
desired destinations using satellites to determine the exact position of the car. To keep up with technology whether to go on to the development stage.”
and to meet the wishes of its customers, the company has added a number of new features to its router
“If we wouldn’t want to develop a product at that point, would that mean that we would have to sell the
during the last few years. The DSS will be a completely new product, incorporating recent developments in
DSS as it is now?” asks Max.
GPS as well as voice recognition and display technologies. Marc strongly supports this product, as it will
“Yes, Max. Except that additionally we would earn some $200,000 from selling our research results to
give BAAG a competitive advantage over its Asian and European competitors.
GM. Their research division is very interested in our work and they have offered that money for our
findings.”
page 395 “Ah, now that’s good news,” remarks Julie.
Brian continues, “If, however, after successfully completing the research stage, we decide to develop a
Driver support systems have been a field of intense research for more than two decades. These new product, then we’ll have to spend another $800,000 for that task, at a 35 percent chance of not being
systems provide the driver with a wide range of information, such as directions, road conditions, traffic successful.”
updates, and so forth. The information exchange can take place verbally or via projection of text onto the “So you are telling us we’ll have to spend $800,000 for a ticket in a lottery where we have a 35 percent
windscreen. Other features help the driver avoid obstacles that have been identified by cars ahead on the chance of not winning anything?” asks Julie.
road (these cars transmit the information to the following vehicles). Marc wants to incorporate all these “Julie, don’t focus on the losses but on the potential gains! The chance of winning in this lottery, as you
features and other technologies into one support system that would then be sold to BAAG’s customers in call it, is 65 percent. I believe that that’s much more than with a normal lottery ticket,” says Marc.
the automobile industry. “Thanks, Marc,” says Brian. “Once we invest that money in development, we have two possible
After all the attendees have taken their seats, Brian starts his presentation: “Marc asked me to inform outcomes: either we will be successful in developing the road scanning device or we won’t. If we fail, then
you about our efforts with the driver support system, particularly the road scanning device. We have once again we’ll sell the DSS in its current form and cash in the $200,000 from GM for the research
reached a stage where we basically have to make a go or no-go decision concerning the research for this results. If the development process is successful, then we have to decide whether to market the new
device, which, as you all know by now, is a key feature in the DSS. We have already integrated the other product.”
devices, such as the GPS-based positioning and direction system. The question with which we have to deal “Why wouldn’t we want to market it after successfully developing it?” asks Marc.
is whether to fund basic research into the road scanning device. If this research is successful, we then will “That’s a good question. Basically what I mean is that we could decide not to sell the product ourselves
have to decide if we want to develop a product based on these results— or if we just want to sell the but instead give the right to sell it to somebody else, to GM for example. They would pay us $1 million for
technology without developing a product. If we do decide to develop the product ourselves, there is a it.”
chance that the product development process might not be successful. In that case, we could still sell the “I like those numbers!” remarks Julie.
technology. In the case of successful product development, we would have to decide whether to market the “Once we decide to build the product and market it, we will face the market uncertainties and I’m sure
product. If we decide not to market the developed product, we could at least sell the product concept that that Julie has those numbers ready for us. Thanks.”
was the result of our successful research and development efforts. Doing so would earn more than just At this point, Brian sits down and Julie comes forward to give her presentation. Immediately some
selling the technology prematurely. If, on the other hand, we decide to market the driver support system, colorful slides are projected on the wall behind her as Max operates the computer.
then we are faced with the uncertainty of how the product will be received by our customers.” “Thanks, Brian. Well, here’s the data we have been able to gather from some marketing research. The
“You completely lost me,” snipes Marc. acceptance of our new product in the market can be high, medium, or low.” Julie is pointing to some figures
projected on the wall behind her. “Our estimates indicate that high acceptance would result in profits of 2 D. L. Keefer, C. W. Kirkwood, and J. L. Corner, “Perspective on Decision Analysis Applications,”
$8.0 million and that medium acceptance would give us $4.0 million. In the unfortunate case of a poor Decision Analysis 1 (2004), pp. 4–22.
reception by our customers, we still expect $2.2 million in profit. I should mention that these profits do not
include the additional costs of marketing or R&D expenses.”
“So, you are saying that in the worst case we’ll make barely more money than with the current
product?” asks Brian.
“Yes, that’s what I am saying.”

page 396
“What budget would you need for the marketing of our DSS with the road scanner?” asks Marc.
“For that we would need an additional $200,000 on top of what has already been included in the profit
estimates,” Julie replies.
“What are the chances of ending up with a high, medium, or low acceptance of the new DSS?” asks
Brian.
“We can see those numbers at the bottom of the slide,” says Julie, while she is turning toward the
projection behind her. There is a 30 percent chance of high market acceptance and a 20 percent chance of
low market acceptance, while the chance of a medium market acceptance is 50 percent.
At this point, Marc moves in his seat and asks: “Given all these numbers and bits of information, what
are you suggesting that we do?”

a. Organize the available data on cost and profit estimates in a table.

b. Formulate the problem in a decision tree. Clearly distinguish between decision and event nodes.
c. Calculate the expected payoffs for each node in the decision tree.
d. What is BAAG’s optimal policy according to Bayes’ decision rule?
e. What would be the expected value of perfect information on the outcome of the research effort?
f. What would be the expected value of perfect information on the outcome of the development effort?
g. Marc is a risk-averse decision maker. In a number of interviews, it has been determined that he would
just barely be willing to consider taking a 50-50 gamble where he either makes $1.2 million or loses $600
thousand. Based on Marc’s level of risk aversion, use the exponential utility function to determine
BAAG’s optimal policy.

Additional Cases
Additional cases for this chapter also are available at the University of Western Ontario Ivey School of
Business website, cases.ivey.uwo.ca/cases, in the segment of the CaseMate area designated for this
book.

1In this very simple decision tree, this advantage does not become evident since each piece of
data is only used once in the tree anyway. However, in later sections, when the possibility of
seismic testing is considered, some data will be repeated many times in the tree and this
advantage will become more clear.
page 397
Chapter Ten

Forecasting

Learning Objectives
After completing this chapter, you should be able to

1. Describe some important types of forecasting applications.

2. Identify two common measures of the accuracy of forecasting methods.
3. Adjust forecasting data to consider seasonal patterns.
4. Describe several forecasting methods that use the pattern of historical data to forecast a
future value.
5. Apply these methods either by hand or with the software provided.
6. Compare these methods to identify the conditions when each is particularly suitable.
7. Describe and apply an approach to forecasting that relates the quantity of interest to one or
more other quantities.
8. Describe several forecasting methods that use expert judgment.
How much will the economy grow over the next year? Where is the stock templates for various forecasting methods are included at
market headed? What about interest rates? How will consumer tastes be www.mhhe.com/Hillier6e.
changing? What will be the hot new products?
Forecasters have answers to all these questions. Unfortunately, these page 398
answers will more than likely be wrong. Nobody can accurately predict the
future every time.
Nevertheless, the future success of any business depends heavily on how 10.1 AN OVERVIEW OF FORECASTING
savvy its management is in spotting trends and developing appropriate
strategies. The leaders of the best companies often seem to have a sixth sense TECHNIQUES
for when to change direction to stay a step ahead of the competition, but this
sixth sense actually is guided by frequent use of the best forecasting To illustrate various forecasting techniques, consider the following problem.
techniques. These companies seldom get into trouble by badly misestimating
what the demand will be for their products. Many other companies do. The A Forecasting Problem
ability to forecast well makes the difference. Fastchips is a leading producer of microprocessors. Six months ago, it
The ability to forecast future sales of their products is particularly launched the sales of its latest microprocessor. The month-by-month sales (in
important for many companies. When historical sales data are available, thousands) of the microprocessor over the initial six months have been
some proven statistical forecasting methods have been developed for
17 25 24 26 30 28
using these data to forecast future demand. Such methods assume that
historical trends will continue, so management then needs to make In this highly competitive market, sales can shift rather quickly, depending
adjustments to reflect current changes in the marketplace. largely on when competitors launch the latest version of their
Several judgmental forecasting methods that solely use expert microprocessors. Therefore, it always is important to have a forecast of the
judgment also are available. These methods are especially valuable when next month’s sales to guide what the production level should be.
little or no historical sales data are available or when major changes in the Let us look at some alternative ways of obtaining this forecast.
marketplace make these data unreliable for forecasting purposes.
Forecasting product demand is just one important application of these Some Forecasting Techniques
forecasting methods. For other applications, forecasts might be needed for The most straightforward technique is the last-value forecasting
such diverse quantities as the need for spare parts, production yields, and method (sometimes also called the naive method), which says simply to
staffing needs. Forecasting techniques also are heavily used for forecasting use the last month’s sales as the forecast for the next month. For Fastchips,
economic trends on a regional, national, or even international level. this yields
We begin the chapter with an overview of forecasting techniques in their
simplest form. Section 10.2 then introduces a typical case study that will be Forecast = 28
carried through much of the chapter to illustrate how these techniques This is a reasonable forecasting method when conditions tend to change so
commonly need to be adapted to incorporate such practical considerations as quickly that sales before the last month’s are not a reliable indicator of future
seasonal demand patterns. Sections 10.3–10.5 focus on statistical forecasting sales.
methods and Section 10.6 on judgmental forecasting methods. Excel
 The averaging forecasting method says to use the average of all the uses a two-dimensional graph with sales measured along the vertical axis
monthly sales to date as the forecast for the next month. This gives and time measured along the horizontal axis. After plotting the page 399
sales data month by month, this method finds a line passing
through the midst of the data as closely as possible. The extension of the line
into future months provides the forecast of sales in these future months.
for Fastchips. This is a reasonable forecasting method when conditions tend Section 10.5 fully presents the linear regression method. The other
to remain so stable that even the earliest sales are a reliable indicator of forecasting methods mentioned above are described in detail in Section 10.3
future sales (a dubious assumption for Fastchips). and then evaluated from a broader perspective in Section 10.4. The
The moving-average forecasting method provides a middle discussion in all three sections is in the context of the case study introduced
ground between the last-value and averaging method by using the average of in Section 10.2.
the monthly sales for only the most recent months as the forecast for the next Which of these forecasting techniques should Fastchips use? On the basis
month. The number of months being used must be specified. For example, a of the sales data so far, it appears that either the moving-average forecasting
three-month moving average forecast for Fastchips is method or the exponential smoothing forecasting method would be a
reasonable choice. However, as time goes on, further analysis should be
done to see which of the forecasting methods provides the smallest
forecasting errors (the absolute value of the difference between the actual
sales and fore-casted sales).
This is a reasonable forecasting method when conditions tend to change
After determining the forecasting error in each of a number of months for
occasionally but not extremely rapidly.
any forecasting method, one common measure of the accuracy of that method
The exponential smoothing forecasting method provides a more is the average of these forecasting errors. (This average is more precisely
sophisticated version of the moving-average method in the way that it gives
defined as the mean absolute deviation, which is abbreviated as MAD,
primary consideration to sales in only the most recent months. In particular,
but we commonly will refer to it more simply as the average forecasting
rather than giving equal weight to the sales in the most recent months, the
error.) This is a reasonable measure of accuracy when the seriousness of a
exponential smoothing method gives the greatest weight to the last month and
forecasting error is considered to be roughly proportional to the size of the
then progressively smaller weights to the older months. (The formula for this
forecasting error.
method will be given in Section 10.3.) This forecasting method is a
However, instead of assuming proportionality, there are some situations
reasonable one under the same conditions as described for the moving-
where the consequences of incurring forecasting errors somewhat larger than
average method.
hoped are very serious. For such situations, another popular measure of the
Additional sophistication is added to the exponential smoothing
accuracy of a forecasting method is the average of the square of its
forecasting method by using exponential smoothing with trend. This
forecasting errors. (This measure is referred to as the mean square error,
latter method adjusts exponential smoothing by also directly considering any
which is abbreviated as MSE.) Throughout this chapter, MAD and MSE
current upward or downward trend in the sales. (Formulas are given in
values are used to help analyze which forecasting method should be used in
Section 10.3.)
the case study.
If the sales data show a relatively constant trend in some direction, then
For some types of products, the sales to be anticipated in a particular
linear regression provides a reasonable forecasting method. This method
month are influenced by the time of the year. For example, a product which is
popular for Christmas presents could well have December sales that are page 400
twice as large as January sales. For any product influenced by seasonal
factors, it is important to incorporate these seasonal factors into the
forecasts. This plays a fundamental role in the analysis of the case study
throughout the chapter.
10.2 A CASE STUDY: THE COMPUTER CLUB
Although we have described the various forecasting techniques in terms WAREHOUSE (CCW) PROBLEM
of forecasts of sales month by month for the Fastchips problem, other
forecasting applications can be somewhat different. The quantity being The Computer Club Warehouse (commonly referred to as CCW) sells
forecasted might be something other than sales and the periods involved various computer products at bargain prices by taking telephone orders (as
might be something like quarters or years instead of months. For example, well as website and fax orders) directly from customers. Its products include
this chapter’s case study involves forecasts of the number of calls to a call desktop and laptop computers, peripherals, hardware accessories, supplies,
center on a quarterly basis. software (including games), and computer-related furniture. The company
When using any of these forecasting techniques, it is also important to mails catalogs to its customers and numerous prospective customers several
look behind the numbers to try to understand what is driving the quantity times per year, as well as publishing minicatalogs in computer magazines.
being forecasted in order to adjust the forecast provided by the forecasting These catalogs prominently display the 800 toll-free telephone number to
technique in an appropriate way. This is a key lesson provided by the call to place an order. These calls come into the company’s call center.
analysis of the case study. When there are factors that are driving changes in
the quantity being forecasted, the judgmental forecasting methods described The CCW Call Center
in Section 10.6 also can play a useful role. The call center is never closed. During busy hours, it is staffed by dozens of
A summary of the formulas for all the forecasting techniques described agents. Their sole job is to take and process customer orders over the
above (and elaborated upon throughout the chapter) is provided at the end of telephone. (A second, much smaller call center uses another 800 number for
the chapter. callers making inquiries or reporting problems. This case study focuses on
just the main call center.)
Review Questions New agents receive a week’s training before beginning work. This
training emphasizes how to efficiently and courteously process an order. An
1. What is the last-value forecasting method and when might it be a reasonable method to
use?
agent is expected not to average more than five minutes per call. Records are
2. What is the averaging forecasting method and when might it be a reasonable method to
kept and an agent who does not meet this target by the end of the probationary
use? period will not be retained. Although the agents are well paid, the tedium and
3. What is the moving-average forecasting method and when might it be a reasonable time pressure associated with the job leads to a fairly high turnover rate.
method to use? A large number of telephone trunks are provided for incoming calls. If an
4. How does the exponential smoothing forecasting method differ from the moving-average agent is not free when the call arrives, it is placed on hold with a recorded
forecasting method?
message and background music. If all the trunks are in use (referred to as
5. How does exponential smoothing with trend differ from the exponential smoothing
forecasting method? saturation), an incoming call receives a busy signal instead.
6. How does the linear regression forecasting method obtain forecasts? Although some customers who receive a busy signal, or who hang up
7. What are the two main measures of the accuracy of a forecasting method? after being on hold too long, will try again later until they get through, many
do not. Therefore, it is very important to have enough agents on duty to upcoming quarter, based on her forecast of the call volume, the forecast
minimize these problems. On the other hand, because of the high labor costs usually has turned out to be considerably off. Therefore, she still isn’t getting
for the agents, CCW tries to avoid having so many agents on duty that they the right staffing levels.
have significant idle time.
Consequently, obtaining forecasts of the demand for the agents is crucial
to the company.

The Call Center Manager, Lydia Weigelt An Application Vignette

The current manager of the call center is Lydia Weigelt. As the top student in
her graduating class from business school, she was wooed by several top
companies before choosing CCW. Extremely bright and hard driving, she is L.L. Bean, Inc., is a widely known retailer of high-quality outdoor goods and apparel,
being groomed to enter top management at CCW in the coming years. with an annual sales volume fairly close to $2 billion. The company markets its
products primarily through the mailing of millions of copies of various catalogs each
When Lydia was hired a little over three years ago, she was assigned to year. Thus, in addition to online and mail orders, much of its sales volume is
her current position in order to learn the business from the ground up. The generated through orders taken at the company’s call center, which is open seven
call center is considered to be the nerve center of the entire CCW operation. days per week. The sales volume is seasonal, with an especially large spike during
the Christmas season. The sales within each week tend to slowly decrease from
Before Lydia’s arrival, the company had suffered from serious Monday through Sunday, except for a sharp drop on the day of a holiday and a sharp
management problems with the call center. Orders were not being processed rise immediately after the arrival of a catalog.
efficiently. A few were even misdirected. Staffing levels never seemed to be Staffing the call center at an appropriate level day by day is critical for the
company. Understaffing results in lost sales from customers who don’t get through to
right. Management directives to adjust the levels kept overcompensating in the call center and then give up. Overstaffing results in excessive labor costs.
the opposite direction. Data needed to get a handle on the staffing level Therefore, accurate forecasts of daily call volumes are needed.
problem hadn’t been kept. Morale was low. Because previous subjective forecasting methods had proven to be
unsatisfactory, L.L. Bean management hired a team of management science
All that changed when Lydia arrived. One of her first moves was to consultants to improve the forecasting procedures for orders coming into the call
install procedures for gathering the data needed to make decisions on staffing center. After L.L. Bean call-center managers compiled an exhaustive list of 35
levels. The key data included a detailed record of call volume and how much possible factors that could logically affect call volumes, this team developed and
helped implement a highly sophisticated time-series forecasting method (the Box and
of the volume was being handled by each agent. Efficiency improved Jenkins autoregressive/ integrated/moving-average method). This methodology
substantially. Despite running a tight ship, Lydia took great pains to praise incorporates all the important factors, including seasonal patterns, the effect of
and reward good work. Morale increased dramatically. holidays, and the effect of the arrival of catalogs. Each week, forecasts are obtained
for the daily call volumes for the next three weeks. The forecasts for the last of the
Although gratified by the great improvement in the operation of the call three weeks are then used to determine the Monday-to-Sunday work schedule at the
center, Lydia still has one major frustration. At the end of each quarter, when call center two weeks in advance.
she knows how many agents are not being retained at the end of their The improved precision of this forecasting methodology is estimated to have
saved L.L. Bean $300,000 annually through enhanced scheduling efficiency. The
probationary period, she makes a decision on how many new agents to hire computerization of the methodology also greatly reduced the labor costs required to
to go through the next training session (held at the beginning of each quarter). prepare the forecast every week.
She has developed an excellent procedure for estimating the staffing level Although this study was conducted back in the 1990s, L.L. Bean has continued
ever since to make especially effective use of forecasting to improve its bottom line.
needed to cover any particular call volume. However, each page 401
In fact, this impressive actual application to a call center inspired the writing of the
time she has used this procedure to set the staffing level for the CCW case study being used in this chapter.
 Source: B. H. Andrews and S. M. Cunningham, “L.L. Bean Improves Call-Center
Forecasting,” Interfaces 25, no. 6 (November–December 1995), pp. 1–13. (A link to
this article is provided at www.mhhe.com/ Hillier6e.)
This is the forecasting method that Lydia has been using. (Referring back to
the forecasting methods introduced in Section 10.1, this 25 percent rule is a
Lydia has concluded that her next project should be to develop a better last-value forecasting method along with a seasonal adjustment for
forecasting method to replace the current one. Christmas sales.)
Better forecasts of call volume are needed.
page 402
Lydia’s Current Forecasting Method
Thanks to Lydia’s data-gathering procedures installed shortly after her
arrival, reliable data on call volume now are available for the past three
years. Figure 10.1 shows the average number of calls received per day in
each of the four quarters of these years. The right side also displays these
same data to show the pattern graphically. This graph was generated by
selecting the data in D4:D15 and then selecting a Line chart from the Insert
tab of the ribbon.

Forecasts need to take into account the seasonal pattern of increased sales in Quarter
4 due to Christmas purchases.

Note that the sales in Quarter 4 jump up each year due to Christmas
purchases. When Lydia first joined CCW, the president told her about the “25
percent rule” that the company had traditionally used to forecast call volume FIGURE 10.1
(and sales). The average number of calls received per day at the CCW call center in each of the four
quarters of the past three years.
The 25 Percent Rule: Since sales are relatively stable through the year except for a
substantial increase during the Christmas season, assume that each quarter’s call volume will be
the same as for the preceding quarter, except for adding 25 percent for Quarter 4. Thus, Figure 10.2 shows the forecasts that Lydia obtained with this method.
Column F gives the forecasting error (the deviation of the forecast from
Forecast for Quarter 2 = Call volume for Quarter 1 what then turned out to be the true value of the call volume) in each case.
Forecast for Quarter 3 = Call volume for Quarter 2 Since the total of the 11 forecasting errors is 4,662, the average is
Forecast for Quarter 4 = 1.25(Call volume for Quarter 3)
The forecast for the next year’s Quarter 1 then would be obtained from
the current year’s Quarter 4 by
MAD is simply the average of the forecasting errors.
 As mentioned in Section 10.1, the average forecasting error is commonly decreasing their weight is desirable. It is the large forecasting errors that
called MAD, which stands for mean absolute deviation. Its formula is have serious consequences. Therefore, it is good to give a relatively large
penalty to a forecasting method that occasionally allows large page 403
forecasting errors while rewarding a forecasting method that
consistently keeps the errors reasonably small. When comparing two such
Thus, in this case, cell I5 gives methods, this can result in the former kind of method receiving the larger
value of MSE even though it has the smaller MAD value. Thus, MSE
MAD = 424 provides a useful complement to MAD by providing additional information
about how consistently a forecasting method avoids seriously large errors.
To put this value of MAD = 424 into perspective, note that 424 is over 5
However, the disadvantage of MSE compared to MAD is a greater difficulty
percent of the average daily call volume in most quarters. With forecasting
in interpreting the significance of its value for an individual forecasting
errors ranging as high as 1,127, two of the errors are well over 10 percent.
method. (For example, it is difficult to attach an intuitive meaning to the
Although errors of this size are common in typical applications of
value of MSE = 317,815 obtained above.) Consequently, Lydia (who is
forecasting, greater accuracy is needed for this particular application. Errors
familiar with both measures) will focus most of her attention on MAD values
of 5 and 10 percent make it impossible to properly set the staffing level for a
while keeping her eye on MSE values as well.
quarter. No wonder Lydia is mad about the poor job that the 25 percent rule
is doing. A better forecasting method is needed.

MSE is the average of the square of the forecasting errors.

Another popular measure of the accuracy of forecasting methods is called

the mean square error and is abbreviated as MSE. Its formula is

Thus, in Figure 10.2,

A bad forecasting error greatly increases the value of the MSE.

as given in cell I8. The advantage of squaring the forecasting errors is that it
increases the weight of the large errors relative to the weight given to the
small errors. Small errors are only to be expected with even the best
forecasting methods and, since such errors have no serious consequences,
 A time series is a series of observations over time of some quantity of interest. For example,
the series of observations of average daily call volume for the most recent 12 quarters, as given
in Figure 10.1, constitute a time series.

She also is reminded that a variety of statistical methods are available for
using the historical data from a time series to forecast a future observation in
the series. These types of statistical forecasting methods are referred to as
time-series forecasting methods. Her task now is to review these
methods, along with other statistical forecasting methods, and assess which
one is best suited for her particular forecasting problem.
To assist her in this task for a few weeks, Lydia gains approval from the
CCW president to contract for the services of a consultant (a former
classmate) from a management science consulting firm that specializes
largely in forecasting.
The next section begins to describe their approach to this problem.

Review Questions
1. How does the Computer Club Warehouse (CCW) operate?
2. What are the consequences of not having enough agents on duty in the CCW call center?
Of having too many?
3. What is Lydia’s current major frustration?
4. What is CCW’s 25 percent rule?
FIGURE 10.2 5. What is MAD?
This spreadsheet records the results of applying the 25 percent rule over the past three
6. What is MSE?
years to forecast the average daily call volume for the upcoming quarter.
7. What is a time series?

The Plan to Find a Better Forecasting Method

Lydia remembers taking a management science course in college. She recalls
that one of the topics in the course was forecasting, so she decides to review 10.3 APPLYING TIME-SERIES FORECASTING
her textbook and class notes on this topic. METHODS TO THE CASE STUDY
page 404 As mentioned at the end of the preceding section, time-series forecasting
methods are statistical forecasting methods that use a time series (a series of
This review reminds her that she is dealing with what is called a time observations over time of some quantity of interest) to forecast a future
series. observation in the series based on its previous values. Several of these
methods were introduced in Section 10.1. We focus in this section on further
describing these methods and applying them to the case study, where the
quantity of interest to be forecasted is the average daily call volume for the
next quarter.
Figure 10.1 in the preceding section highlights the seasonal pattern of
CCW’s call volumes, with a large jump up each fourth quarter due to
Christmas shopping. Therefore, before considering specific forecasting
methods, Lydia and the consultant begin by addressing how to deal with this
seasonal pattern.

Considering Seasonal Effects

For many years, the folklore at CCW has been that the call volume (and
sales) will be pretty stable over the first three quarters of a year and then In general, the seasonal factor for any period of a year (a quarter, a month, etc.) measures
how that period compares to the overall average for an entire year. Specifically, using historical
will jump up by about 25 percent in Quarter 4. This has been the basis for the data, the seasonal factor is calculated to be
25 percent rule.
To check how close this folklore still is to reality, the consultant uses the
data previously given in Figure 10.1 to calculate the average daily call
volume for each quarter over the past three years. For example, the average
An Excel template is available at www.mhhe.com/Hillier6e for
for Quarter 1 is
calculating these seasonal factors. Figure 10.3 shows this template applied to
the CCW problem.

This Excel template calculates seasonal factors on either a monthly or quarterly basis.

These averages for all four quarters are shown in the second column of Table Note the significant differences in the seasonal factors for the first three
10.1. Underneath this column, the overall average over all four quarters is quarters, with Quarter 3 considerably above the other two. This makes sense
calculated to be 7,529. Dividing the average for each quarter by this overall to Lydia, who has long suspected that back-to-school buying should give a
average gives the seasonal factor shown in the third column. small boost to sales in Quarter 3.

page 405
TABLE 10.1
Calculation of the Seasonal Factors for the CCW Problem
 the 25 percent rule, to indicate seasonal patterns until such time as future data indicate a shift in
these patterns.

The Seasonally Adjusted Time Series

It is much easier to analyze sales data and detect new trends if the data are
first adjusted to remove the effect of seasonal patterns. To remove the
seasonal effects from the time series shown in Figure 10.1, each of these
average daily call volumes needs to be divided by the corresponding
seasonal factor given in Table 10.1 and Figure 10.3. Thus, the formula is

Applying this formula to all 12 call volumes in Figure 10.1 gives the
seasonally adjusted call volumes shown in column F of the Excel template in
Figure 10.4.

FIGURE 10.3
The Excel template available at www.mhhe.com/ Hillier6e for calculating seasonal
factors is applied here to the CCW problem.

page 406
In contrast to the 25 percent rule, the seasonal factor for Quarter 4 is only
19 percent higher than that for Quarter 3. (However, the Quarter 4 factor is
about 25 percent above 0.94, which is the average of the seasonal factors for
the first three quarters.)
Although data on call volumes are not available prior to the most recent
three years, reliable sales data have been kept. Upon checking these data
several years back, Lydia finds the same seasonal patterns occurring.
Conclusion: The seasonal factors given in Table 10.1 appear to accurately reflect subtle but FIGURE 10.4
important differences in all the seasons. Therefore, these factors now will be used, instead of The seasonally adjusted time series for the CCW problem obtained by dividing each actual
average daily call volume in Figure 10.1 by the corresponding seasonal factor obtained in
 Figure 10.3. Outline for Forecasting Call Volume

page 407 . Select a time-series forecasting method.

. Apply this method to the seasonally adjusted time series to obtain a forecast
In effect, these seasonally adjusted call volumes show what the call of the seasonally adjusted call volume for the next quarter.1
volumes would have been if the calls that occur because of the time of the
. Multiply this forecast by the corresponding seasonal factor in Table 10.1 to
year (Christmas shopping, back-toschool shopping, etc.) had been spread
obtain a forecast of the actual call volume (without seasonal adjustment).
evenly throughout the year instead. Compare the plots in Figures 10.4 and
Figures 10.1. After considering the smaller vertical scale in Figure 10.4, note If seasonal adjustments are not needed, you can obtain your forecast directly from the
how much less fluctuation this figure has than Figure 10.1 because of original time series and then skip step 3.
removing seasonal effects. However, this figure still is far from completely
flat because fluctuations in call volume occur for other reasons besides just The following descriptions of forecasting methods focus on how to perform
seasonal effects. For example, hot new products attract a flurry of calls. A step 2, that is, how to forecast the next data point for a given time series. We
jump also occurs just after the mailing of a catalog. Some random also include a spreadsheet in each case that applies steps 2 and 3 throughout
fluctuations occur without any apparent explanation. Figure 10.4 enables the past three years and then calculates both MAD (the average forecasting
seeing and analyzing these fluctuations in sales volumes that are not caused error) and MSE (the average of the square of the forecasting errors). Lydia
by seasonal effects. and the consultant are paying particular attention to the MAD values to assess
which method seems best suited for forecasting CCW call volumes.
Removing seasonal effects provides a much clearer picture of trends.

The Last-Value Forecasting Method

The pattern in these remaining fluctuations in the seasonally adjusted
The last-value forecasting method ignores all the data points in a time
time series (especially the pattern for the most recent data points) is
series except the last one. It then uses this last value as the forecast of what
particularly helpful for forecasting where the next data point will fall. Thus,
the next data point will turn out to be, so the formula is simply
in Figure 10.4, the data points fall in the range between 6,635 and 8,178,
with an average of 7,529. However, the last few data points are trending Forecast = Last value
upward above this average, and the last point is the highest in the entire time
series. This suggests that the next data point for the upcoming quarter Figure 10.5 shows what would have happened if this method had been
probably will be above the 7,529 average and may well be near or even applied to the CCW problem over the past three years. (We are supposing
above the last data point of 8,178. that the seasonal factors given in Table 10.1 already were being used then.)
The various time-series forecasting methods use different Column E gives the values of the seasonally adjusted call volumes from
approaches to projecting forward the pattern in the seasonally adjusted time column F of Figure 10.4. Each of these values then becomes the seasonally
series to forecast the next data point. The main methods will be presented in adjusted forecast for the next quarter, as shown in column F.
this section. Rows 22–33 show separate plots of these values in columns E and F.
After obtaining a forecast for the seasonally adjusted time series, all Note how the plot of the seasonally adjusted forecasts follows exactly the
these methods then convert this forecast to a forecast of the actual call same path as the plot of the values of the seasonally adjusted call volumes
volume (without seasonal adjustments), as outlined below. but shifted to the right by one quarter. Therefore, each time there is a large
shift up or down in the call volume, the forecasts are one quarter late in The Excel template available at www.mhhe.com/Hillier6e for the last-value method with
seasonal adjustments is applied here to the CCW problem.
catching up with the shift.

page 408 page 409

Multiplying each seasonally adjusted forecast in column F by the
corresponding seasonal factor in column K gives the forecast of the actual
call volume (without seasonal adjustment) presented in column G. The
difference between this forecast and the actual call volume in column D
gives the forecasting error in column H.
Thus, column G is using the following formula:
Actual forecast = Seasonal factor × Seasonally adjusted forecast
as indicated by the equations at the bottom of the figure. For example, since
cell K9 gives 0.93 as the seasonal factor for Quarter 1, the forecast of the
actual call volume for Year 2, Quarter 1 given in cell G10 is
Actual forecast = (0.93)(7,005) = 6,515
Since the true value of this call volume turned out to be 7,257, the forecasting
error calculated in cell H10 for this quarter is
Forecasting error = 7,257 − 6,515 = 742
Summing these forecasting errors over all 11 quarters of forecasts gives a
total of 3,246, so the average forecasting error given in cell K23 is

This compares with MAD = 424 for the 25 percent rule that Lydia has been
using (as described in the preceding section).
Similarly, the average of the square of these forecasting errors is
calculated in cell K26 as

FIGURE 10.5
This value also is considerably less than the corresponding value, MSE = managers who are anything but naive do occasionally use this method under
317,815, shown in Figure 10.2 for the 25 percent rule. such circumstances.
Except for its graph, Figure 10.5 displays one of the templates in this
This method is a good one to use when conditions are changing rapidly.
chapter’s Excel files. In fact, two Excel templates are available at
www.mhhe.com/Hillier6e for each of the forecasting methods presented in
this section. One template performs all the calculations for you for the case page 410
where no seasonal adjustments are needed. The second template does the
same when seasonal adjustments are included, as illustrated by this figure. The Averaging Forecasting Method
With all templates of the second type, you have complete flexibility for what The averaging forecasting method goes to the other extreme. Rather
to enter as the seasonal factors. One option is to calculate these factors than using just a sample size of one, this method uses all the data points in the
based on historical data (as was done with another Excel template in Figure time series and simply averages these points. Thus, the forecast of what the
10.3). Another is to estimate them based on historical experience, as with the next data point will turn out to be is
25 percent rule.
Forecast = Average of all data to date
The 25 percent rule actually is a last-value forecasting method as well,
but with different seasonal factors. Since this rule holds that the call volume Using the corresponding Excel template to apply this method to the CCW
in the fourth quarter will average 25 percent more than each of the first three problem over the past three years gives the seasonally adjusted forecasts
quarters, its seasonal factors are essentially 0.94 for Quarters 1, 2, and 3 and shown in column F of Figure 10.6. At the bottom of the figure, the equation
1.18 (25 percent more than 0.94) for Quarter 4. Thus, the lower value of entered into each of the column F cells is just the average of the column E
MAD in Figure 10.5 is entirely due to refining the seasonal factors in Table cells in the preceding rows. The middle of the figure shows a plot of these
10.1. seasonally adjusted forecasts for all three years next to the values of the
Lydia is enthusiastic to see the substantial improvement obtained by seasonally adjusted call volumes. Note how each forecast lies at the average
simply refining the seasonal factors. However, the consultant quickly adds a of all the preceding call volumes. Therefore, each time there is a large shift
note of caution. The forecasts obtained in Figure 10.5 are using the same data in the call volume, the subsequent forecasts are very slow in catching up with
that were used to calculate these refined seasonal factors, which creates the shift.
some bias for these factors to tend to perform better than on new data (future Multiplying all the seasonally adjusted forecasts in column F by the
call volumes). Fortunately, Lydia also has checked older sales data to corresponding seasonal factors in column K then gives the forecasts of the
confirm that these seasonal factors seem quite accurate. The consultant actual call volumes shown in column G. Based on the resulting forecasting
agrees that it appears that these factors should provide a significant errors given in column H, the average forecasting error in this case (cell
improvement over the 25 percent rule. K23) is
The last-value forecasting method sometimes is called the naive
MAD = 400
method, because statisticians consider it naive to use just a sample size of
one when additional relevant data are available. However, when conditions which is considerably larger than the 295 obtained for the last-value
are changing rapidly, it may be that the last value is the only relevant data forecasting method. Similarly, the average of the square of the forecasting
point for forecasting the next value under current conditions. Therefore, errors given in cell K26 is
 MSE = 242,876
which also is considerably larger than the corresponding value of 145,909
for the last-value forecasting method.
Lydia is quite surprised, since she expected an average to do much better
than a sample size of one. The consultant agrees that averaging should
perform considerably better if conditions remain the same throughout the time page 411
series. However, it appears that the conditions affecting the CCW call
volume have been changing significantly over the past three years. The call
volume was quite a bit higher in year 2 than in year 1, and then jumped up
again late in year 3, apparently as popular new products became available.
Therefore, the Year 1 values were not very relevant for forecasting under the
changed conditions of years 2 and 3. Including the year 1 call volumes in the
overall average caused every forecast for years 2 and 3 to be too low,
sometimes by large amounts.

The averaging forecasting method is a good one to use when conditions are very
stable, which is not the case for CCW.

The Moving-Average Forecasting Method

Rather than using old data that may no longer be relevant, the moving-
average forecasting method averages the data for only the most recent
time periods. Let
n = Number of most recent periods considered particularly
relevant for forecasting the next period
Then the forecast for the next period is
Forecast = Average of last n values
Lydia and the consultant decide to use n = 4, since conditions appear to
be relatively stable for only about four quarters (one year) at a time.
With n = 4, the first forecast becomes available after four quarters of call
volumes have been observed. Thus, the initial seasonally adjusted forecasts
in cells F10:F12 of Figure 10.7 are
 An Application Vignette

Taco Bell Corporation has over 6,400 quick-service restaurants in the United
States and a growing international market. It serves approximately 2 billion meals per
year, generating well over $5 billion in annual sales income. Taco Bell also co-brands
with KFC and Pizza Hut under the corporate umbrella of Yum! Brands, a Fortune 500
corporation.
At each Taco Bell restaurant, the amount of business is highly variable throughout
the day (and from day to day), with a heavy concentration during the normal meal
times. Therefore, determining how many employees should be scheduled to perform
what functions in the restaurant at any given time is a complex and vexing problem.
To attack this problem, Taco Bell management instructed a team of management
scientists (including several consultants) to develop a new labor–management
system. The team concluded that the system needed three major components: (1) a
forecasting model for predicting customer transactions at any time, (2) a computer
simulation model (such as those described in Chapters 12 and 13) to translate
customer transactions to labor requirements, and (3) an integer programming model
to schedule employees to satisfy labor requirements and minimize payroll.
To enable application of a forecasting model in each restaurant, a procedure is
needed to continually gather data on the number of customer transactions during
each 15-minute interval throughout the day, each day of the week. Therefore, the
management science team developed and implemented a rolling database
containing six weeks of in-store and drive-through transaction data to be stored on
each restaurant’s computer. After some testing of alternative forecasting methods,
the team concluded that a six-week moving average is best. In other words, the
forecast of the number of transactions in a particular 15-minute period on a particular
day of the week should be the average of the number of transactions during the
corresponding period in the preceding six weeks. However, the restaurant manager
has the authority to modify the forecast if unusual events distorted the data being
used.
The implementation of this forecasting procedure along with the other
components of the labor–management system have provided Taco Bell with
documented savings of $13 million per year in labor costs. The impact of this labor–
management system has been particularly impressive because, although it was
FIGURE 10.6 developed back in the 1990s, it has provided the foundation for continued efforts to
manage labor costs over the many subsequent years at Taco Bell, as well as at KFC
The Excel template available at www.mhhe.com/Hillier6e for the averaging method with
and Pizza Hut.
seasonal adjustments is applied here to the CCW problem.

Source: J. Hueter and W. Swart, “An Integrated Labor–Management System for Taco
page 412 Bell,” Interfaces 28, no. 1 (January–February 1998), pp. 75–91. (A link to this article is
provided at www.mhhe.com/Hillier6e.)
 sharply, as with the big jump up in call volumes at the beginning of year 2,
Note how each forecast is updated from the preceding one by lopping off one and then again in the middle of year 3, the next few forecasting errors tend to
observation (the oldest one) and adding one new one (the most recent be very large.
observation). Thus, the moving-average method is somewhat slow to respond to
Column F of Figure 10.7 shows all the seasonally adjusted forecasts changing conditions. One reason is that it places the same weight on each of
obtained in this way with the equations at the bottom. For each of these the last n values in the time series even though the older values may be less
forecasts, note in the plot how it lies at the average of the four preceding representative of current conditions than the last value observed.
(seasonally adjusted) call volumes. Consequently, each time there is a large The next method corrects this weighting defect.
shift in the call volume, it takes four quarters for the forecasts to fully catch
up with this shift (by which time another shift may already have occurred). page 413
Consequently, the average of the eight forecasting errors in column H is
MAD = 437
the highest of any of the methods so far, including even the 25 percent rule.
The average of the square of the forecasting errors is somewhat better at
MSE = 238,816
since this is slightly lower than for the averaging method and considerably
below the value for the 25 percent rule, but it is still substantially higher than
for the last-value method.
Lydia is very puzzled about this surprisingly high MAD value. The
moving-average method seemed like a very sensible approach to forecasting,
with more rationale behind it than any of the previous methods. (It uses only
recent history and it uses multiple observations.) So why should it do so
poorly?

The moving-averaging forecasting method is a good one to use when conditions don’t
change much over the number of time periods included in the average.

The consultant explains that this is indeed a very good forecasting method
when conditions remain pretty much the same over n time periods (or four
quarters in this case). For example, the seasonally adjusted call volumes
remained reasonably stable throughout year 2 and the first half of year 3.
Consequently, the forecasting error dropped all the way down to 68 (cell
H15) for the last of these six quarters. However, when conditions shift
 page 414

The Exponential Smoothing Forecasting Method

The exponential smoothing forecasting method modifies the
moving-average method by placing the greatest weight on the last value in the
time series and then progressively smaller weights on the older values.
However, rather than needing to calculate a weighted average each time, it
uses a simpler formula to obtain the same result.
This formula for forecasting the next value in the time series combines
the last value and the last forecast (the one used one time period ago to
forecast this last value) as follows:
Forecast = α(Last value) + (1 − α)(Last forecast)
where α (the Greek letter alpha) is a constant between 0 and 1 called the
smoothing constant. For example, if the last value in a particular time
series is 24, the last forecast is 20, and α = 0.25, then the forecast of the next
value in the time series is
Forecast = 0.25(24) + 0.75(20)
= 21
Two Excel templates (one without seasonal adjustments and one with)
are available at www.mhhe.com/Hillier6e for applying this formula to
generate a series of forecasts (period by period) for a time series when you
specify the value of α.
The choice of the value for the smoothing constant α has a substantial
effect on the forecast, so the choice should be made with care. A small value
(say, α = 0.1) is appropriate if conditions are remaining relatively stable.
However, a larger value (say, α = 0.3) is needed if significant changes in the
conditions are occurring relatively frequently. Because of the frequent shifts
in the CCW seasonally adjusted time series, Lydia and the consultant
FIGURE 10.7 conclude that α = 0.5 would be an appropriate value. (The values selected
The Excel template available at www.mhhe.com/Hillier6e for the moving-average for most applications are between 0.1 and 0.3, but a larger value can be used
method with seasonal adjustments is applied here to the CCW problem. in this kind of situation.)
 The more unstable the conditions, the larger the smoothing constant α needs to be (but
never bigger than 1).

When making the first forecast, there is no last forecast available to plug
into the right-hand side of the above formula. Therefore, to get started, a
reasonable approach is to make an initial estimate of the average value page 415
anticipated for the time series. This initial estimate is then labeled as the last
forecast when using the formula for forecasting the second value in the time Similarly, the forecast for Quarter 4 is
series after observing its first value (now labeled as the last value).
CCW call volumes have averaged just over 7,500 for the past three
years, and the level of business just prior to Year 1 was comparable.
Consequently, Lydia and the consultant decide to use
Initial estimate = 7,500 The exponential smoothing forecasting method places the greatest weight on the last
value and then decreases the weights as the values get older.
to begin retrospectively generating the forecasts over the past three years.
Recall that the first few seasonally adjusted call volumes are 7,322, 7,183, Thus, this latter forecast places a weight of 0.5 on the last value, 0.25 on the
and 6,635. Thus, using the above formula with α = 0.5 for the second quarter next-to-last value, and 0.125 on the next prior value (the first one), with the
onward, the first few seasonally adjusted forecasts are remaining weight on the initial estimate. In subsequent quarters as more and
more values are considered by the forecast, the same pattern continues of
Y1, Q1: Seas. adj. forecast = 7,500
placing less and less weight on the values as they get older. Thus, with any
Y1, Q2: Seas. adj. forecast = 0.5(7,322) + choice of α, the weights prior to the initial estimate would be α, α(1 − α),
0.5(7,500) = 7,411 α(1 − α)2, and so forth.
Y1, Q3: Seas. adj. forecast = 0.5(7,183) + Therefore, choosing the value of α amounts to using this pattern to choose
0.5(7,411) = 7,297 the desired progression of weights on the time series values. With frequent
Y1, Q4: Seas. adj. forecast = 0.5(6,635) + shifts in the time series, a large weight needs to be placed on the most recent
0.5(7,297) = 6,966 value, with rapidly decreasing weights on older values. However, with a
relatively stable time series, it is desirable to place a significant weight on
To see why these forecasts are weighted averages of the time series many values in order to have a large sample size.
values to date, look at the calculations for Quarters 2 and 3. Since Further insight into the choice of α is provided by an alternative form of
the forecasting formula.
0.5(7,322) + 0.5(7,500) = 7,411
the forecast for Quarter 3 can be written as
where the absolute value of (Last value − Last forecast) is just the last Although he isn’t ready to mention it to Lydia yet, the consultant is
forecasting error. Therefore, the bottom form of this formula indicates that beginning to develop an idea for a whole new approach that might give her
each new forecast is adjusting the last forecast by adding or subtracting the the forecasting precision she needs. But first he has one more time-series
quantity α times the last forecasting error. If the forecasting error usually is forecasting method to present.
mainly due to random fluctuations in the time-series values, then only a small To lay the groundwork for this method, the consultant explains a major
value of α should be used for this adjustment. However, if the forecasting reason why exponential smoothing did not fare well in this case. Look at the
error often is largely due to a shift in the time series, then a large value of α plot of seasonally adjusted call volumes in Figure 10.8. Note the distinct
is needed to make a substantial adjustment quickly. trend downward in the first three quarters, then a sharp trend upward for the
Using α = 0.5, the Excel template in Figure 10.8 provides all the results next two, and finally a major trend upward for the last five quarters. Also
for CCW with this forecasting method. Rows 22–32 show a plot of all the note the large gap between the two plots (meaning large forecasting errors)
seasonally adjusted forecasts next to the true values of the seasonally by the end of each of these trends. The reason for these large errors is that
adjusted call volumes. Note how each forecast lies midway between the exponential smoothing forecasts lag well behind such a trend because they
preceding call volume and the preceding forecast. Therefore, each time there place significant weight on values near the beginning of the trend. Although a
is a large shift in the call volume, the forecasts largely catch up with the shift large value of α = 0.5 helps, exponential smoothing forecasts tend to lag
rather quickly. The resulting average of the forecasting errors in column H is further behind such a trend than last-value forecasts.
given in cell K28 as
MAD = 324 page 416

This is significantly smaller than for the previous forecasting methods, except
for the value of MAD = 295 for the last-value forecasting method. The same
comparison also holds for the average of the square of the forecasting errors,
which is calculated in cell K31 as
MSE = 157,836
Lydia is somewhat frustrated at this point. She feels that she needs a
method with average forecasting errors well below 295. Realizing that the
last-value forecasting method is considered the naive method, she had
expected that such a popular and sophisticated method as exponential
smoothing would beat it easily.
The consultant is somewhat surprised also. However, he points out that
the difference between MAD = 324 for exponential smoothing and MAD =
295 for last-value forecasting is really too small to be statistically
significant. If the same two methods were to be applied the next three years,
exponential smoothing might come out ahead. Lydia is not impressed.
 The next method adjusts exponential smoothing by also estimating the
current trend and then projects this trend forward to help forecast the next
value in the time series.

Exponential Smoothing with Trend

Exponential smoothing with trend uses the recent values in the time
series to estimate any current upward or downward trend in these values. It
is especially designed for the kind of time series depicted in Figure 10.9
where an upward (or downward) trend tends to continue for a considerable
number of periods (but not necessarily indefinitely). This particular figure
shows the estimated population of a certain state of the United States at
midyear over a series of years. The line in the figure (commonly referred to
as a trend line) shows the basic trend that the time series is following, but
with fluctuations on both sides of the line. Because the basic trend is upward
in this case, forecasts based on any of the preceding forecasting methods
would tend to be considerably too low. However, by developing an estimate
of the current slope of this trend line, and then adjusting the forecast to
consider this slope, considerably more accurate forecasts should be
obtained. This is the basic idea behind exponential smoothing with trend.
Trend is defined as
Trend = Average change from one time-series value to the
next if the current pattern continues
The formula for forecasting the next value in the time series, then is modified
from the preceding method by adding the estimated trend. Thus, the new
formula is
Forecast = α(Last value) + (1 − α)(Last forecast) + Estimated trend

Adding the estimated trend enables the forecast to keep up with the current trend in
FIGURE 10.8 the data.
The Excel template available at www.mhhe.com/Hillier6e for the exponential smoothing
method with seasonal adjustments is applied here to the CCW problem. (A separate box describes how this formula can be easily modified to
forecast beyond the next value in the time series as well.)
page 417
 with trend provides a way of doing this. In particular, after determining
the estimated trend, this method’s forecast for n time periods into the
future is

Exponential smoothing also is used to obtain and update the estimated

trend each time. The formula is
Estimate of trend = β(Latest trend) + (1 − β)(Last estimate of trend)

The trend smoothing constant β is used to apply exponential smoothing to estimating

the trend.

where β (the Greek letter beta) is the trend smoothing constant, which,
like α, must be between 0 and 1. Latest trend refers to the trend based on just
the last two values in the time series and the last two forecasts. Its formula is
FIGURE 10.9
A time series that gives the estimated population of a certain state of the United States
over a series of years. The trend line shows the basic upward trend of the population.

page 418 Getting started with this forecasting method requires making two initial
estimates about the status of the time series just prior to beginning
forecasting. These initial estimates are
. Initial estimate of the average value of the time series if the conditions just
FORECASTING MORE THAN ONE TIME prior to beginning forecasting were to remain unchanged without any trend.
PERIOD AHEAD . Initial estimate of the trend of the time series just prior to beginning
We have focused thus far on forecasting what will happen in the next
time period (the next quarter in the case of CCW). However,
forecasting.
managers sometimes need to forecast further into the future. How
can the various time-series forecasting methods be adapted to do The forecast for the first period being forecasted then is
this?
In the case of the last-value, averaging, moving-average, and First forecast = Initial estimate of average value + Initial estimate of trend
exponential smoothing methods, the forecast for the next period also
is the best available forecast for subsequent periods as well. The second forecast is obtained from the above formulas, where the initial
However, when there is a trend in the data, it is important to take this estimate of trend is used as the last estimate of trend in the formula for
trend into account for long-range forecasts. Exponential smoothing
estimated trend and the initial estimate of average value is used as both the Y1, Q1: Seas. adj. forecast = 7,500 + 0 = 7,500
next-to-last value and the next-to-last forecast in the formula for latest trend. Y1, Q2: Latest trend = 0.3(7,322 − 7,500) +
The above formulas then are used directly to obtain subsequent forecasts. 0.7(7,500 − 7,500) = −53.4
Since the calculations involved with this method are relatively involved,
a computer commonly is used to implement the method. Two Excel templates Estimated trend = 0.3(−53.4) + 0.7(0) = −16
(one without seasonal adjustments and one with) are available at Seas. adj. forecast = 0.3(7,322) +
www.mhhe.com/Hillier6e for this method. 0.7(7,500) − 16 = 7,431
Y1, Q3: Latest trend = 0.3(7,183 − 7,322) +
The Excel templates for this method perform the calculations for you.
0.7(7,431 − 7,500) = −90
The considerations involved in choosing the trend smoothing constant β Estimated trend = 0.3(−90) + 0.7(−16) =
are similar to those for α. A large value of β (say, β = 0.3) is more −38.2
responsive to recent changes in the trend, whereas a relatively small value Seas. adj. forecast = 0.3(7,183) +
(say, β = 0.1) uses more data in a significant way to estimate trend. 0.7(7,431) − 38.2 = 7,318
After trying various combinations of α and β on the CCW problem, the
consultant concludes that α = 0.3 and β = 0.3 perform about as well as any. The Excel template in Figure 10.10 shows the results from these calculations
Both values are on the high end of the typically used range (0.1 to 0.3), but for all 12 quarters over the past three years, as well as for the upcoming
the frequent changes in the CCW time series call for large values. However, quarter. The middle of the figure shows the plots of all the seasonally
lowering α from the 0.5 value used with the preceding method seems adjusted call volumes and seasonally adjusted forecasts. Note how each
justified since incorporating trend into the analysis would help respond more trend up or down in the call volumes causes the forecasts to gradually trend
quickly to changes. in the same direction, but then the trend in the forecasts takes a couple
When applying exponential smoothing without trend earlier, Lydia and quarters to turn around when the trend in call volumes suddenly reverses
the consultant chose 7,500 as the initial estimate of the average value of the direction. The resulting forecasting errors in column J then give an average
seasonally adjusted call volumes. They now note that there was no forecasting error (cell M30) of
noticeable trend in these call volumes just prior to the retrospective
When the trend in the data suddenly reverses direction, it takes a little while for the
generation of forecasts three years ago. Therefore, to apply exponential estimated trend to turn around.
smoothing with trend, they decide to use
MAD = 345
a little above the 324 value for regular exponential smoothing and 295 for
last-value forecasting. A similar result is obtained when using the square of
page 419 the forecasting errors since the mean square error given in cell M33,

Working with the seasonally adjusted call volumes given in several MSE = 180,796
recent figures, these initial estimates lead to the following seasonally
also is a little above the MSE values for these other two forecasting methods.
adjusted forecasts.
 Table 10.2 summarizes the values of MAD and MSE for all the Consultant: Cheer up. I have an idea for how we can use one of these
forecasting methods so far. Here is Lydia’s reaction to the large MAD value time-series forecasting methods in a different way that may do the job you
for exponential smoothing with trend. want.
Lydia: Really? Tell me more.
Lydia: I’m very discouraged. These time-series forecasting methods just Consultant: Well, let me hold off on the details until we can check out
aren’t doing the job I need. I thought this one would. It sounded like an whether this is going to work. What I would like you to do is contact CCW’s
excellent method that also would deal with the trends we keep encountering. marketing manager and set up a meeting between the three of us. Also, send
Consultant: Yes, it is a very good method under the right circumstances. him your data on call volumes for the past three years. Ask him to dig out his
When you have trends that may occasionally shift some over time, it should sales data for the same period and compare it to your data.
do a great job.
Lydia: So what went wrong here?
page 420
Consultant: Well, look at the trends you have here in the seasonally
adjusted time series. You have a fairly sharp downward trend the first three
quarters and then suddenly a very sharp upward trend for a couple quarters.
Then it flattens out before a big drop in the eighth quarter. Then suddenly it is
going up again. It is really tough to keep up with such abrupt big shifts in the
trends. This method is better suited for much more gradual shifts in the
trends.
Lydia: OK. But aren’t there any other methods? None of these will do.
Consultant: There is one other main time-series forecasting method. It is
called the ARIMA method, which is an acronym for AutoRegressive
Integrated Moving Average. It also is sometimes called the Box-Jenkins
method, in honor of its founders. It is a very sophisticated method, but some
excellent software is available for implementing it. Another nice feature is
that it is well-suited for dealing with strong seasonal patterns.
Lydia: Sounds good. So shouldn’t we be trying this ARIMA method?
Consultant: Not at this point. It is such a sophisticated method that it
requires a great amount of past data, say, a minimum of 50 time periods. We
don’t have nearly enough data.

The ARIMA method is another good forecasting method, but it requires much more
data than CCW currently has available.

Lydia: A pity. So what are we going to do? I haven’t seen anything that
will do the job.
 TABLE 10.2 page 421
The Average Forecasting Error (MAD) and Mean Square Error (MSE) for
the Various Time-Series Forecasting Methods When Forecasting CCW Call Volumes

Forecasting Method MAD MSE

CCW’s 25 percent rule 424 317,815
Last-value method 295 145,909
Averaging method 400 242,876
Moving-average method 437 238,816
Exponential smoothing 324 157,836
Exponential smoothing with trend 345 180,796

Lydia: OK. What should I tell him is the purpose of the meeting?
Consultant: Explain what we’re trying to accomplish here about
forecasting call volumes. Then tell him that we’re trying to understand better
what has been causing these sudden shifts in call volumes. He knows more
about what has been driving sales up or down than anybody. We just want to
pick his brain about this.
Lydia: OK. Will do.

The Meeting with the Marketing Manager

This meeting takes place a few days later. As you eavesdrop (after the
preliminaries), you will find it helpful to refer to the call volume data in one
of the recent spreadsheets, such as Figure 10.10.

Lydia: Did you receive the call volume data I mailed to you?
Marketing manager: Yes, I did.
Consultant: How does it compare with your own sales data for these three
years?
Marketing manager: Your data track mine pretty closely. I see the same ups
and downs in both sets of data.
FIGURE 10.10 Lydia: That makes sense, since it’s the calls to my call center that generate
The Excel template available at www.mhhe.com/Hillier6e for the exponential smoothing those sales.
with trend method with seasonal adjustments is applied here to the CCW problem. Marketing manager: Right.
Consultant: Now, let me check on what caused the ups and downs. Three The one big factor that drives total sales up or down is whether the company has just
released any hot new products.
years ago, what we labeled as Year 1 in our data, there was a definite trend
down for most of the year. What caused that?
Marketing manager: There really is just one big factor: Do we have any hot
Marketing manager: Yes, I remember that year all too well. It wasn’t a very new products out there? We have well over a hundred products. But most of
good year. The new Klugman operating system had been scheduled to come them just fill a small niche in the market. Many of them are old standbys that,
out early that year. Then they kept pushing the release date back. People kept with updates, just keep going indefinitely. All these small-niche products
waiting. They didn’t manage to get it out until the beginning of the next year, together provide most of our total sales. A nice stable market base. Then, in
so we even missed the Christmas sales. addition, we should have three or four major new products out there. Maybe
Lydia: But our call volume did jump up a little more than usual during that a couple that have been out for a few months but still have some life left in
holiday season. them. Then one or two just coming out that we hope will do very well.
Marketing manager: Yes. So did sales. I remember that we came out with a Consultant: I see. A large market base and then three or four major new
new networking tool, one with faster data transfer, in time for the holiday products.
season. It turned out to be very popular for a few months. It really bailed us Marketing manager: That’s what we shoot for.
out during that slow period.
Consultant: Are you able to predict how well a major new product will
Consultant: Then the Klugman operating system was released and sales do?
jumped the next year.
Marketing manager: I try. I’ve gotten better at it. I’m usually fairly close on
Marketing manager: Right. what the initial response will be, but it is difficult to predict how long the
Lydia: What happened late in the year? We weren’t as busy as we expected product will hold up. I would like to have a better handle on it.
to be. Consultant: Thanks very much for all your information. It has verified
Marketing manager: I assume that most people already had updated to the what I’ve been suspecting for awhile now.
new operating system by then. There wasn’t any major change in our product Lydia: What’s that?
mix during that period.
Consultant: Then sales moved back up the next year. Last year. Forecasting call volumes better requires coordinating directly with what is driving
sales.
Marketing manager: Yes, last year was a rather good year. We had a couple
new products that did very well. One was a new data storage device that
Consultant: That we really need to coordinate directly with what is
came out early in the year. Very inexpensive. The other was a color plain-
driving sales in order to do a better job of forecasting call volumes.
paper printer that was released in July. We were able to offer it at a very
competitive price and our customers gobbled it up. Lydia: Good thought!
Consultant: How would the two of you feel about coordinating in
developing better procedures for forecasting both sales and call volumes?
page 422
Lydia: You bet.
Consultant: Thanks. That really clarifies what lies behind those call Marketing manager: Sounds good.
volume numbers we’ve been working with. Now I have another key question
We will return to this story in the next two sections. However, as you
for you. When you look at your sales data and do your own forecasting, what
have seen above, the calculations required for the more sophisticated
do you see as the key factors that drive total sales up or down?
forecasting methods are fairly cumbersome. Consequently, software 4. Why is the last-value forecasting method sometimes called the naive method?
commonly is used for applications of forecasting like the CCW study. 5. Why did the averaging forecasting method not perform very well on the case study?
Therefore, let us pause briefly to describe the availability of forecasting 6. What is the rationale for replacing the averaging forecasting method by the moving-
average forecasting method?
software.
7. How does the exponential smoothing forecasting method modify the moving-average
forecasting method?
Forecasting Software 8. With exponential smoothing, when is a small value of the smoothing constant appropriate?
Because of its widespread usage in practice, an article in the June 2016 issue A larger value?
of ORMS Today (a publication of INFORMS) presents a detailed survey of 9. What is the formula for obtaining the next forecast with exponential smoothing? What is
added to this formula when using exponential smoothing with trend?
forecasting software. This article points out that at that time, 19 different
10. What does the marketing manager say is the one big factor that drives CCW’s total sales
vendors were marketing a total of 26 forecasting software packages. up or down?
Detailed information was provided about each of these packages.
Four of these packages were being marketed by Frontline Systems, which
also is the developer of Analytic Solver. The main features of these packages
now have been incorporated into Analytic Solver. The various forecasting
10.4 THE TIME-SERIES FORECASTING
methods are included on the “Data Mining” tab (along with some data mining METHODS IN PERSPECTIVE
tools not included in this book). In addition to the more advanced forecasting
methods described in this chapter, Analytic Solver includes a powerful Section 10.3 presented several methods for forecasting the next value of a
method called ARIMA. (Recall that the consultant told Lydia about ARIMA time series in the context of the CCW case study. We now will take a step
during their conversation just before meeting with the marketing manager.) back to place into perspective just what these methods are trying to
ARIMA is widely used for detailed forecasting studies based on accomplish. After providing this perspective, CCW’s consultant then will
extensive data. However, it also is a sophisticated method that requires a give his recommendation for setting up a forecasting system.
strong background in statistics, so we will not describe it further.
Furthermore, due to the advanced nature of this and other forecasting The Goal of the Forecasting Methods
software included within Analytic Solver, we will not go into the details of It actually is something of a misnomer to talk about forecasting the value of
how to use this software. We instead have provided convenient Excel the next observation in a time series (such as CCW’s call volume in the next
templates that are available at www.mhhe.com/Hillier6e for each of the quarter). It is impossible to predict the value precisely, because this next
forecasting methods covered in this chapter. value can turn out to be anything over some range. What it will be depends
upon future circumstances that are beyond our control.
Review Questions
The next value in a time series cannot be forecasted with certainty because it has a
1. What does a seasonal factor measure? probability distribution rather than a fixed value that definitely will occur.
2. What is the formula for calculating the seasonally adjusted call volume from the actual call
volume and the seasonal factor? In other words, the next value that will occur in a time series is a random
page 423 variable. It has some probability distribution. For example, Figure 10.11
3. What is the formula for calculating the forecast of the actual call volume from shows a typical probability distribution for the CCW call volume in a future
the seasonal factor and the seasonally adjusted forecast?
quarter in which the mean of this distribution happens to be 7,500. This
distribution indicates the relative likelihood of the various possible values of for each and every time period, then the averaging forecasting method
the call volume. Nobody can say in advance which value actually will occur. provides the best estimate of the mean.
So what is the meaning of the single number that is selected as the However, other forecasting methods commonly are used instead because
“forecast” of the next value in the time series? If it were possible, we would the distribution may be changing over time.
like this number to be the mean of the distribution. The reason is that random
observations from the distribution tend to cluster around the mean of the Problems Caused by Shifting Distributions
distribution. Therefore, using the mean as the forecast would tend to Section 10.3 began by considering seasonal effects. This then led to
minimize the average forecasting error. estimating CCW’s seasonal factors as 0.93, 0.90, 0.99, and 1.18 for Quarters
1, 2, 3, and 4, respectively.
If the overall average daily call volume for a year is 7,500, these
seasonal factors imply that the probability distributions for the four quarters
of that year fall roughly as shown in Figure 10.12. Since these distributions
have different means, we should no longer simply average the random
observations (observed call volumes) from all four quarters to estimate the
mean for any one of these distributions.
This complication is why the preceding section seasonally adjusted the
time series. Dividing each quarter’s call volume by its seasonal factor shifts
the distribution of this seasonally adjusted call volume over to basically the
distribution shown in Figure 10.11 with a mean of 7,500. This allows
averaging the seasonally adjusted values to estimate this mean.
FIGURE 10.11 Unfortunately, even after seasonally adjusting the time series, the
A typical probability distribution of what the average daily call volume will be for CCW in
a quarter when the mean is 7,500.
probability distribution may not remain the same from one year to the next (or
even from one quarter to the next). For example, as CCW’s marketing
manager explained, total sales jumped substantially at the beginning of Year
page 424 2 when the new Klugman operating system became available. This also
Unfortunately, we don’t actually know what this probability distribution caused the average daily call volume to increase by about 10 percent, from
is, let alone its mean. The best we can do is use all the available data (past just over 7,000 in Year 1 to over 7,700 in Year 2. Figure 10.13 compares the
values from the time series) to estimate the mean as closely as possible. resulting distributions for typical quarters (seasonally adjusted) in the two
years.
The goal of time-series forecasting methods is to estimate the mean of the underlying
probability distribution of the next value of the time series as closely as possible.

Given some random observations from a single probability distribution,

the best estimate of its mean is the sample average (the average of all these
observations). Therefore, if a time series has exactly the same distribution
 Judging from the marketing manager’s information, it appears that some
shift in the distribution also occurred several times from just one quarter to
the next rather than from only one year to the next. This further added to the
forecasting errors.

Comparison of the Forecasting Methods

Section 10.3 presented five methods for forecasting the next value in a time
series. Which of these methods is particularly suitable for a given
FIGURE 10.12
Typical probability distributions of CCW’s average daily call volumes in the four quarters application depends greatly on how stable the time series is.
of a year in which the overall average is 7,500. A time series is said to be stable if its underlying probability distribution usually remains the
same from one time period to the next. (Any shifts that do occur in the distribution are both
infrequent and small.) A time series is unstable if both frequent and sizable shifts in the
distribution tend to occur.

CCW’s seasonally adjusted time series shown in Figure 10.4 (and many
subsequent figures) appears to have had several shifts in the distribution,
including the sizable one depicted in Figure 10.13. Therefore, this time
series is an example of a relatively unstable one.
Here is a summary of which type of time series fits each of the
forecasting methods.
FIGURE 10.13
Comparison of typical probability distributions of CCW’s average daily call volumes
The key factor in choosing a forecasting method is how stable the time series is.
(seasonally adjusted) in years 1 and 2.

Last-value method: Suitable for a time series that is so unstable that even the next-to-last value
page 425 is not considered relevant for forecasting the next value.
Averaging method: Suitable for a very stable time series where even its first few values are
Random observations from the Year 1 distribution in this figure provide a considered relevant later for forecasting the next value.
poor basis for estimating the mean of the Year 2 distribution. Yet, except for Moving-average method: Suitable for a moderately stable time series where the last
the last-value method, each of the forecasting methods presented in the
few values are considered relevant for forecasting the next value. The number of values
preceding section placed at least some weight on these observations from included in the moving average reflects the anticipated degree of stability in the time series.
Year 1 to estimate the mean for each quarter in Year 2. This was a major part
Exponential smoothing method: Suitable for a time series in the range from somewhat
of the reason why both the average forecasting errors (MAD) and the mean unstable to rather stable, where the value of the smoothing constant needs to be adjusted
square errors (MSE) were higher for these methods than for the last-value
to fit the anticipated degree of stability. Refines the moving-average method by placing the
method. greatest weight on the most recent values, but is not as readily understood by managers as the
moving-average method.
When the probability distribution for a time series shifts frequently, recent data quickly
become outdated for forecasting purposes. Exponential smoothing with trend: Suitable for a time series where the mean of the
distribution tends to follow a trend either up or down, provided that changes in the trend occur
 only occasionally and gradually. . Exponential smoothing with trend, with relatively large smoothing constants,
Unfortunately for CCW, its seasonally adjusted time series proved to be a is suggested for forecasting sales of each of the major new products. Once
little too unstable for any of these methods except the last-value method, again, retrospective testing should be conducted to check this decision and
which is considered to be the least powerful of these forecasting methods. to guide choosing values for the smoothing constants. The marketing
Even when using exponential smoothing with trend, the changes in the trend manager should be asked to provide the initial estimate of anticipated sales
occurred too frequently and sharply. in the first month for a new product. He also should be asked to check the
In light of these considerations, the consultant now is ready to present his subsequent exponential smoothing forecasts and make any adjustments he
recommendations to Lydia for a new forecasting procedure. feels are appropriate based on his knowledge of what is happening in the
marketplace.
The Consultant’s Recommendations . Because of the strong seasonal sales pattern, seasonally adjusted time series
should be used for each application of these forecasting methods.
. Forecasting should be done monthly rather than quarterly in order to respond . After separately obtaining forecasts of actual sales for each of the major
more quickly to changing conditions. components of total sales identified in recommendation 5, these forecasts
. Hiring and training of new agents also should be done monthly instead of should be summed to obtain a forecast of total sales.
quarterly in order to fine-tune staffing levels to meet changing needs. . Causal forecasting with linear regression (as described in the next section)
. Recently retired agents should be offered the opportunity to work part-time should be used to obtain a forecast of call volume from this forecast of total
on an on-call basis to help meet current staffing needs more closely. sales.
. Since sales drive call volume, the forecasting process should begin by
The next section describes how to obtain a forecast of call volume from a forecast of
forecasting sales. total sales.
page 426

. For forecasting purposes, total sales should be broken down into the major Lydia accepts these recommendations with considerable enthusiasm. She
components described by the marketing manager, namely, (1) the relatively also agrees to work with the marketing manager to gain his cooperation.
stable market base of numerous small-niche products and (2) each of the Read on to see how the last recommendation is implemented.
few (perhaps three or four) major new products whose success or failure
can significantly drive total sales up or down. These major new products Review Questions
would be identified by the marketing manager on an ongoing basis. 1. What kind of variable is the next value that will occur in a time series?
. Exponential smoothing with a relatively small smoothing constant is 2. What is the goal of time-series forecasting methods?
suggested for forecasting sales of the marketing base of numerous small- 3. Is the probability distribution of CCW’s average daily call volume the same for every
niche products. However, before making a final decision on the forecasting quarter?
method, retrospective testing should be conducted to check how well this 4. What is the explanation for why the average forecasting errors were higher for the other
time-series forecasting methods than for the supposedly less powerful last-value method?
particular method would have performed over the past three years. This
5. What is the distinction between a stable time series and an unstable time series?
testing also should guide selection of the value of the smoothing constant.
6. What is the consultant’s recommendation regarding what should be forecasted instead of
Forecasting call volume should begin by separately forecasting the major components of total sales.
call volumes to begin the forecasting process?
7. What are the major components of CCW’s total sales?
10.5 CAUSAL FORECASTING WITH LINEAR Table 10.3 shows some examples of the kinds of situations where causal
forecasting sometimes is used. In each of the first four cases, the indicated
REGRESSION dependent variable can be expected to go up or down rather directly with the
independent variable(s) listed in the rightmost column. The last case also
We have focused so far on time-series forecasting methods, that is, methods applies when some quantity of interest (e.g., sales of a product) tends to
that forecast the next value in a time series based on its previous values. follow a steady trend upward (or downward) with the passage of time (the
These methods have been used retrospectively in Section 10.3 to forecast independent variable that drives the quantity of interest).
CCW’s call volume in the next quarter based on its previous call volumes.
Linear Regression
Causal Forecasting
At Lydia’s request, the marketing manager brought sales data for the past
However, the consultant’s last recommendation suggests another approach to three years to their recent meeting. These data are summarized in Figure
forecasting. It is really sales that drive call volume, and sales can be 10.14. In particular, column D gives the average daily sales (in units of
forecasted considerably more precisely than call volume. Therefore, it thousands of dollars) for each of the 12 past quarters. Column E repeats the
should be possible to obtain a better forecast of call volume by relating it data given previously on average daily call volumes. None of the data have
directly to forecasted sales. This kind of approach is called causal been seasonally adjusted.
forecasting (or correlational forecasting). The right side of the figure was generated by selecting the data in D5:E16
For the CCW problem, call volume is the dependent variable and total sales is the
and then selecting an XY Scatter Chart from the Chart tab of the ribbon. This
independent variable. graph shows a plot of the data in columns D and E on a two-dimensional
graph. Thus, each of the 12 points in the graph shows the combination of
Causal forecasting obtains a forecast of the quantity of interest (the dependent sales and call volume for one of the 12 quarters (without identifying which
variable) by relating it directly to one or more other quantities (the independent
variables) that drive the quantity of interest.
quarter).
This graph shows a close relationship between call volume and sales.
Each increase or decrease in sales is accompanied by a roughly proportional
page 427 increase or decrease in call volume. This is not surprising since the sales are
TABLE 10.3 being made through the calls to the call center.
Possible Examples of Causal Forecasting It appears from this graph that the relationship between call volume and
Type of Forecasting Possible Dependent Possible Independent sales can be approximated by a straight line. Figure 10.15 shows such a line.
Variable Variables (This line was generated by right-clicking on one of the data points in the
Sales Sales of a product Amount of advertising graph in Figure 10.14, choosing Add Trendline and then specifying Linear
Spare parts Demand for spare parts Usage of equipment Trend. The equation above the line is added by choosing Display Equation
Economic trends Gross domestic product Various economic factors on Chart under Trendline Options.) This line is referred to as a linear
CCW call volume Call volume Total sales regression line.
Any quantity This same quantity Time
 When doing causal forecasting with a single independent variable, linear regression involves
approximating the relationship between the dependent variable (call volume for CCW) and the
independent variable (sales for CCW) by a straight line. This linear regression line is
drawn on a graph with the independent variable on the horizontal axis and the dependent
variable on the vertical axis. The line is constructed after plotting a number of points showing
each observed value of the independent variable and the corresponding value of the dependent
variable.

A linear regression line estimates what the value of the dependent variable should be
for any particular value of the independent variable.

Thus, the linear regression line in Figure 10.15 can be used to estimate
what the call volume should be for a particular value of sales. In general, the
equation for the linear regression line has the form
y = a + bx
FIGURE 10.14 where
The data needed to do causal forecasting for the CCW problem by relating call volume to y = Estimated value of the dependent variable, as given by the linear
sales.
regression line
a = Intercept of the linear regression line with the y-axis
page 428 b = Slope of the linear regression line
x = Value of the independent variable

The equation for a linear regression line is called the regression equation.

(If there is more than one independent variable, then this regression
equation has a term, a constant times the variable, added on the right-hand
side for each of these variables.) For the linear regression line in this figure,
the exact values of a and b happen to be
a = −1,223.9 b = 1.6324
Figure 10.16 shows the Excel template available at
www.mhhe.com/Hillier6e that can also be used to find these values of a and
b. You need to input all the observed values of the independent variable
(sales) and of the dependent variable (call volume) in columns C and D, and
FIGURE 10.15 then the template performs all the calculations. On the right, note that you
Figure 10.14 has been modified here by adding a trend line to the graph. have the option of inserting a value for x (sales) in cell J10 and then the
template calculates the corresponding value of y (call volume) that lies on The method of least squares chooses the values of a and b that make the sum of the
resulting numbers in column G as small as possible.
the linear regression line. This calculation can be repeated for as many
values of x as desired. In addition, column E already shows these
The average of the square of the estimation errors in column G (1,838)
calculations for each value of x in column C, so each cell in column E gives
has an interesting interpretation. Suppose that the sales for a quarter could be
the estimate of call volume provided by the linear regression line for the
known in advance (either because of advance orders or an exact prediction).
corresponding sales level in column C. The difference between page 429
In this case, using the linear regression line to forecast call volume would
this estimate and the actual call volume in column D gives the
give a mean square error (MSE) of 1,838. What the method of least squares
estimation error in column F. The square of this error is shown in column G.
has done is place the linear regression line exactly where it would minimize
MSE in this situation. Note that this minimum value of MSE = 1,838 is
roughly 1 percent of the MSE values given earlier in Table 10.2 for the time-
series forecasting methods presented in Section 10.3.
The numbers in column F also are interesting. Averaging these numbers
reveals that the average estimation error for the 12 quarters is only 35. This
indicates that if the sales for a quarter were known in advance (or could be
predicted exactly), then using the linear regression line to forecast call
volume would give an average forecasting error (MAD) of only 35. This is
only about 10 percent of the values of MAD obtained for the various time-
series forecasting methods in Section 10.3.
For some applications of causal forecasting, the value of the independent
variable will be known in advance. This is not the case here, where the
independent variable is the total sales for the upcoming time period.
However, the consultant is confident that a very good forecast of total sales
can be obtained by following his recommendations. This page 430
forecast can then be used as the value of the independent
variable for obtaining a good forecast of call volume from the linear
FIGURE 10.16 regression line.
The Excel template available in Connect for doing causal forecasting with linear
regression, as illustrated here for the CCW problem.
CCW’s New Forecasting Procedure
The procedure used to obtain a and b is called the method of least . Obtain a forecast of total (average daily) sales for the upcoming month by
squares. This method chooses the values of a and b that minimize the sum implementing the consultant’s recommendations presented at the end of
of the square of the estimation errors given in column G of Figure 10.16. Section 10.4.
Thus, the sum of the numbers in column G (22,051) is the minimum possible.
. Use this forecast as the value of total sales in forecasting the average daily
Any significantly different values of a and b would give different estimation
call volume for the upcoming month from the linear regression line
errors that would cause this sum to be larger.
identified in Figures 10.15 and Figures 10.16.
 Seasonal adjustments play a role in step 1 of this procedure, but not in where x1 and x2 are the independent variables with b1 and b2 as their
step 2. Based on the consultant’s recommendations (see recommendation no. respective coefficients. (Each additional independent variable would add a
8), seasonal adjustments should be incorporated into whatever forecasting similar term in the regression equation.) With two independent variables, the
method is used in step 1. However, a forecast of seasonally adjusted sales corresponding linear regression line now would lie in a three-dimensional
would then be converted back into a forecast of actual sales for use in step 2. graph with y (the dependent variable) as the vertical axis and both x1 and x2
By using a forecast of actual sales, the forecast of call volume obtained from as horizontal axes in the other two dimensions. Regardless of the number of
the linear regression line in step 2 will be the desired forecast of actual call independent variables, the method of least squares still can be used to
volume rather than seasonally adjusted call volume. choose the value of a, b1, b2, etc., that minimizes the sum of the square of the
estimation errors. As before, these estimation errors are calculated by
Multiple Linear Regression comparing the values of the dependent variable at the various data points
Before completing the CCW story at the end of this section, let us pause here with the corresponding estimates given by the regression equation.
to look at an important extension of the linear regression method described Textbooks in statistics often provide many more details about multiple
above. linear regression. However, we will not delve further into this more
advanced topic.
Many applications of linear regression have multiple independent variables.

The application of linear regression to the CCW problem involves only The CCW Case Study a Year Later
one independent variable (total sales) that drives the dependent variable A year after implementing the consultant’s recommendations, Lydia gives him
(call volume). However, some applications of causal forecasting with linear a call.
regression involve multiple independent variables that together drive the
dependent variable. This type of linear regression is called multiple linear Lydia: I just wanted to let you know how things are going. And to
regression. congratulate you on the great job you did for us. Remember that the 25
Multiple linear regression is a widely used forecasting technique because percent rule was giving us MAD values over 400? And then the various
it is common to have multiple independent variables. We live in a time-series forecasting methods were doing almost as badly?
complicated business world where such applications frequently arise. Consultant: Yes, I do remember. You were pretty discouraged there for a
Economic forecasting also often uses multiple linear regression. For while.
example, when using linear regression to forecast the nation’s gross domestic
product for the next quarter (the dependent variable), the independent page 431
variables might include such leading indicators of future economic
performance as the current level of the stock market, the current index of Lydia: I sure was! But I’m feeling a lot better now. I just calculated MAD
consumer confidence, the current index of business activity (measuring for the first year under your new procedure. 120. Only 120!
orders placed), and so forth. Consultant: Great. That’s the kind of improvement we like to see. What do
If there are, say, two independent variables, the regression equation you think has made the biggest difference?
would have the form Lydia: I think the biggest factor was tying our forecasting into forecasting
sales. We never had much feeling for where call volumes were headed. But
y = a + b1x1 + b2x2
we have a much better handle on what sales will be because, with the 6. How does the MAD value for CCW’s new forecasting procedure compare with that for the
old procedure that used the 25 percent rule?
marketing manager’s help, we can see what is causing the shifts.
Consultant: Yes. I think that is a real key to successful forecasting. You
saw that we can get a lot of garbage by simply applying a time-series
forecasting method to historical data without understanding what is causing 10.6 JUDGMENTAL FORECASTING
the shifts. You have to get behind the numbers and see what is really going METHODS
on, then design the forecasting procedure to catch the shifts as they occur, like
we did by having the marketing manager identify the major new products that We have focused so far on statistical forecasting methods that base the
impact total sales and then separately forecasting sales for each of them. forecast on historical data. However, such methods cannot be used if no data
are available, or if the data are not representative of current conditions. In
A key to successful forecasting is to understand what is causing shifts so that they can
be caught as soon as they occur.
such cases, judgmental forecasting methods can be used instead.

Judgmental forecasting methods use expert opinion to make forecasts.

Lydia: Right. Bringing the marketing manager in on this was a great move.
He’s a real supporter of the new procedure now, by the way. He says it is
Even when good data are available, some managers prefer a judgmental
giving him valuable information as well.
method instead of a formal statistical method. In many other cases, a
Consultant: Good. Is he making adjustments in the statistical forecasts, combination of the two may be used. For example, in the CCW case study,
based on his knowledge of what is going on in the marketplace, like I the marketing manager uses his judgment, based on his long experience and
recommended? his knowledge of what is happening in the marketplace, to adjust the sales
Lydia: Yes, he is. You have a couple fans here. We really appreciate the forecasts obtained from time-series forecasting methods.
great job you did for us. Here is a brief overview of the main judgmental forecasting methods.
A good forecasting procedure combines a well-constructed statistical forecasting . Manager’s opinion: This is the most informal of the methods, because it
procedure and a savvy manager who understands what is driving the numbers and so
is able to make appropriate adjustments in the forecasts. simply involves a single manager using his or her best judgment to make the
forecast. In some cases, some data may be available to help make this
Review Questions judgment. In others, the manager may be drawing solely on experience and
an intimate knowledge of the current conditions that drive the forecasted
1. What is causal forecasting? quantity.
2. When applying causal forecasting to the CCW problem, what is the dependent variable
and what is the independent variable? . Jury of executive opinion: This method is similar to the first one, except
3. When doing causal forecasting with a single independent variable, what does linear now it involves a small group of high-level managers who pool their best
regression involve? judgments to collectively make the forecast. This method may be used for
4. What is the form of the equation for a linear regression line with a single independent more critical forecasts for which several executives share responsibility and
variable? With more than one independent variable? can provide different types of expertise.
5. What is the name of the method for obtaining the value of the constants in the regression
equation for a linear regression line?
page 432
 corporate chain of command, with managerial review at each level, to be
aggregated into a corporate sales forecast.
An Application Vignette The salesforce composite method uses a bottom-up approach.
. Consumer market survey: This method goes even further than the
preceding one in adopting a grass-roots approach to sales forecasting. It
Compañia Sud Americana de Vapores (CSAV) is one of the world’s largest involves surveying customers and potential customers regarding their future
shipping companies. It is the most global firm in Chile, with regional offices on four purchasing plans and how they would respond to various new features in
continents. It has operations in over 100 countries while using more than 2,000
terminals and depots worldwide. The company’s main business is shipping cargo
products. This input is particularly helpful for designing new products and
using containers on large container ships. The container fleet transports about then in developing the initial forecasts of their sales. It also is helpful for
700,000 20-foot containers. planning a marketing campaign.
Once loaded containers reach their destination and are unloaded, it is important
to quickly send the empty containers to where they will be most needed for . Delphi method: This method employs a panel of experts in different
subsequent shipments. Therefore, CSAV and researchers from the University of locations who independently fill out a series of questionnaires. However,
Chile conducted a project to apply management science to more effectively make the results from each questionnaire are provided with the next one, so each
these decisions quickly on a global basis. The result was the development of an
Empty Container Logistics Optimization System (ECO). expert then can evaluate this group information in adjusting his or her
ECO uses both an advanced network optimization model and a sophisticated responses next time. The goal is to reach a relatively narrow spread of
inventory model. However, both of these models need to be driven by relatively conclusions from most of the experts. The decision makers then assess this
precise forecasts of the demand for empty containers. The forecasting methods
used to generate these forecasts are a combination of certain time-series input from the panel of experts to develop the forecast. This involved
forecasting methods (moving average and seasonally adjusted moving average with process normally is used only at the highest levels of a corporation or
trend) and two types of judgmental forecasting methods. One of the latter types is government to develop long-range forecasts of broad trends.
sales-force composite—a consensus forecast model in which the sales agents
worldwide register their demand expectations. The other is manager’s opinion—
whereby logistics planners provide information that must be considered in the Review Questions
forecasts.
This extensive application of management science, including a combination of 1. Statistical forecasting methods cannot be used under what circumstances?
forecasting methods, resulted in savings of $81 million in its first full year. These 2. Is judgmental forecasting used only when statistical forecasting methods cannot be used?
cost reductions were mainly the result of reducing empty container inventories by 50
percent and increasing container turnover by 60 percent. 3. How does the jury of executive opinion method differ from the manager’s opinion method?
4. How does the salesforce composite method begin?
Source: R. Epstein and 14 co-authors, “A Strategic Empty Container Logistics 5. When is a market survey particularly helpful?
Optimization in a Major Shipping Company,” Interfaces 42, no. 1 (January–February 6. When might the Delphi method be used?
2012), pp. 5–16. (A link to this article is provided at www.mhhe.com/Hillier6e.)

page 433
. Salesforce composite: This method is often used for sales forecasting when
a company employs a salesforce to help generate sales. It is a bottom-up
approach whereby each sales person provides an estimate of what sales
will be in his or her region. These estimates then are sent up through the 10.7 Summary
The future success of any business depends heavily on the ability of its management to exponential smoothing forecasting method A method that uses a weighted average of
forecast well. Forecasting may be needed in several areas, including sales, the need for the last value from a time series and the last forecast to obtain the forecast of the next
spare parts, production yields, economic trends, and staffing needs. value. (Sections 10.1 and 10.3), 398, 414
The Computer Club Warehouse (CCW) case study illustrates a variety of approaches to exponential smoothing with trend An adjustment of the exponential smoothing
forecasting, some of which prove unsatisfactory for this particular application. Ultimately, it forecasting method that projects the current trend forward to help forecast the next value of
becomes necessary to get behind the CCW data to understand just what is driving the call a time series (and perhaps subsequent values as well). (Sections 10.1 and 10.3), 398, 417
volumes at its call center in order to develop a good forecasting system.
A time series is a series of observations over time of some quantity of interest. Several forecasting error The deviation of the forecast from the realized quantity. (Sections 10.1
statistical forecasting methods use these observations in some way to forecast what the next and 10.2), 399, 402
value will be. These methods include the last-value method, the averaging method, the independent variable A quantity that drives the value of the dependent variable in causal
moving-average method, the exponential smoothing method, and exponential smoothing with forecasting. (Section 10.5), 426
trend.
judgmental forecasting methods Methods that use expert judgment to make forecasts.
The goal of all these methods is to estimate the mean of the underlying probability (Introduction and Section 10.6), 397, 431
distribution of the next value of the time series as closely as possible. This may require using
seasonal factors to seasonally adjust the time series, as well as identifying other factors that jury of executive opinion A judgmental forecasting method that involves a small group of
may cause this underlying probability distribution to shift from one time period to the next. high-level managers pooling their best judgment to collectively make the forecast. (Section
Another statistical forecasting approach is called causal forecasting. This approach 10.6), 431
obtains a forecast of the quantity of interest (the dependent variable) by relating it directly to last-value forecasting method A method that uses the last value of a time series as the
one or more other quantities (the independent variables) that drive the quantity of interest. forecast of the next value. (Sections 10.1 and 10.3), 398, 407
Frequently, this involves using linear regression to approximate the relationship between the
linear regression Approximating the relationship between the dependent variable and the
dependent variable and the independent variable(s) by a straight line.
independent variable(s) by a straight line. (Sections 10.1 and 10.5), 398, 428
The software accompanying this book includes Excel templates for the various statistical
forecasting methods. linear regression line The line that approximates the relationship between the dependent
Still another key category of forecasting methods is judgmental methods. This category variable and the independent variable(s) when using causal forecasting. (Section 10.5), 428
involves basing the forecast on a manager’s opinion, a jury of executive opinion, a salesforce MAD An acronym for mean absolute deviation, which refers to the average forecasting
composite, a consumer market survey, or the Delphi method. error. (Sections 10.1 and 10.2), 399, 402
manager’s opinion A judgmental forecasting method that involves using a single
manager’s best judgment to make the forecast. (Section 10.6), 431
Glossary
page 434
ARIMA An acronym for the AutoRegressive Integrated Moving Average method, a
sophisticated time-series forecasting method commonly referred to as the Box-Jenkins mean absolute deviation (MAD) The average forecasting error. (Sections 10.1 and
method. (Section 10.3), 419 10.2), 399, 402
averaging forecasting method A method that uses the average of all the past mean square error (MSE) The average of the square of the forecasting errors. (Sections
observations from a time series as a forecast of the next value. (Sections 10.1 and 10.3), 10.1 and 10.2), 399, 402
398, 410 method of least squares The procedure used to obtain the constants in the equation for a
causal forecasting Obtaining a forecast of the dependent variable by relating it directly to linear regression line. (Section 10.5), 429
one or more independent variables. (Section 10.5), 426 moving-average forecasting method A method that uses the average of the last n
consumer market survey A judgmental forecasting method that uses surveys of observations from a time series as a forecast of the next value. (Sections 10.1 and 10.3),
customers and potential customers. (Section 10.6), 432 398, 410
Delphi method A judgmental forecasting method that uses input from a panel of experts MSE An acronym for mean square error, the average of the square of the forecasting
in different locations. (Section 10.6), 432 errors. (Sections 10.1 and 10.2), 399, 402
dependent variable The quantity of interest when doing causal forecasting. (Section multiple linear regression Linear regression with multiple independent variables.
10.5), 426 (Section 10.5), 430
naive method Another name for the last-value forecasting method. (Section 10.3), 409
regression equation The equation for a linear regression line. (Section 10.5), 428
salesforce composite A judgmental forecasting method that aggregates the sales
forecasts of the salesforce from their various regions. (Section 10.6), 432
seasonal factor A factor for any period of a year that measures how that period compares
to the overall average for an entire year. (Section 10.3), 405
seasonally adjusted time series An adjustment of the original time series that removes
seasonal effects. (Section 10.3), 407
smoothing constant A parameter of the exponential smoothing forecasting method that
gives the weight to be placed on the last value in the time series. (Section 10.3), 414
stable time series A time series whose underlying probability distribution usually remains
the same from one time period to the next. (Section 10.4), 425
statistical forecasting methods Methods that use historical data to forecast future
quantities. (Introduction and Sections 10.1–10.5), 397–431
time series A series of observations over time of some quantity of interest. (Section 10.2),
404
time-series forecasting methods Methods that use the past observations in a time
series to forecast what the next value will be. (Sections 10.2 and 10.3), 404, 407
trend The average change from one time series value to the next if the current pattern
continues. (Section 10.3), 417
trend smoothing constant A smoothing constant for estimating the trend when using
exponential smoothing with trend. (Section 10.3), 418
unstable time series A time series that has frequent and sizable shifts in its underlying
probability distribution. (Section 10.4), 425

Summary of Key Formulas

y = a + bx
Forecasting error = Difference between a forecasted value and the true value then obtained
(Section 10.2)
page 435

Learning Aids for This Chapter

All learning aids are available at www.mhhe.com/Hillier6e.
 Excel Files:
Template for Seasonal Factors
Problems
Templates for Last-Value Method (with and without Seasonality) The first 18 problems should be done by hand without using the templates in this chapter’s
Templates for Averaging Method (with and without Seasonality) Excel files. To the left of the subsequent problems (or their parts), we have inserted the
Templates for Moving-Average Method (with and without Seasonality) symbol E (for Excel) to indicate that one of these templates can be helpful. An asterisk on
the problem number indicates that at least a partial answer is given in the back of the book.
Templates for Exponential Smoothing Method (with and without Seasonality)
Templates for Exponential Smoothing with Trend (with and without Seasonality) 10.1.* The Hammaker Company’s newest product has had the following sales during its
first five months: 5, 17, 29, 41, 39. The sales manager now wants a forecast of sales in the
Template for Linear Regression next month.
Excel Add-In: a. Use the last-value method.
Analytic Solver b. Use the averaging method.
c. Use the moving-average method with the three most recent months.
d. Given the sales pattern so far, do any of these methods seem inappropriate for
Solved Problem obtaining the forecast? Why?
10.2. Sales of stoves have been going well for the Good-Value Department Store. These
sales for the past five months have been 15, 18, 12, 17, 13. Use the following methods to
The solution is available at www.mhhe.com/Hillier6e. obtain a forecast of sales for the next month.
a. The last-value method.
10.S1. Forecasting Charitable Donations at the Union
b. The averaging method.
Mission c. The moving-average method with three months.
Cash donations (in thousands of dollars) at the Union Mission for 2015–2017 were as shown
d. If you feel that the conditions affecting sales next month will be the same as in
below.
the last five months, which of these methods do you prefer for obtaining the
forecast? Why?

page 436
10.3.* You have been forecasting sales the last four quarters. These forecasts and the true
values that subsequently were obtained are shown below.
a. Ignoring seasonal effects, compare both the MAD and MSE values for the last-value
Quarter Forecast True Value
method, the averaging method, the moving-average method (based on the most recent 4
quarters), the exponential smoothing method (with an initial estimate of 275 and a 1 327 345
smoothing constant of α = 0.2), and the exponential smoothing method with trend (with 2 332 317
initial estimates of 275 for the average value, 2 for the trend, along with smoothing
3 328 336
constants of α = 0.2 and β = 0.2) when they are applied retrospectively to the years
2015–2017. 4 330 311
b. Determine the seasonal factors for the four quarters.
c. Repeat part a, but now consider the seasonal effects. Calculate the forecasting error for each quarter. Then calculate MAD and MSE.
d. Using the forecasting method from part a or c with the lowest MAD value, make long- 10.4. Sharon Johnson, sales manager for the Alvarez-Baines Company, is trying to
range forecasts for charitable donations in each of the quarters of 2018. choose between two methods for forecasting sales that she has been using during the past
five months. During these months, the two methods obtained the forecasts shown next for
the company’s most important product, where the subsequent actual sales are shown on the a. Determine the seasonal factors for the four quarters.
right. b. Over the next year, the unemployment rates (not seasonally adjusted) for the four
quarters turn out to be 7.8 percent, 7.4 percent, 8.7 percent, and 6.1 percent.
Determine the seasonally adjusted unemployment rates for the four quarters.
What does this progression of rates suggest about whether the state’s economy is
improving?
10.8. Ralph Billett is the manager of a real estate agency. He now wishes to develop a
forecast of the number of houses that will be sold by the agency over the next year.
The agency’s quarter-by-quarter sales figures over the last three years are shown
below.
a. Calculate and compare MAD for these two forecasting methods. Then do the
same with MSE.
b. Sharon is uncomfortable with choosing between these two methods based on
such limited data, but she also does not want to delay further before making her
choice. She does have similar sales data for the three years prior to using these
forecasting methods the past five months. How can these older data be used to
further help her evaluate the two methods and choose one?
10.5. Read the referenced article that fully describes the management science study
summarized in the application vignette presented in Section 10.2. Briefly describe how a. Determine the seasonal factors for the four quarters.
forecasting was applied in this study. Then list the various financial and nonfinancial benefits b. After considering seasonal effects, use the last-value method to forecast sales in
that resulted from this study. Quarter 1 of next year.
10.6. Figure 10.1 shows CCW’s average daily call volume for each quarter of the past c. Assuming that each of the quarterly forecasts is correct, what would the last-
three years and Figure 10.4 gives the seasonally adjusted call volumes. Lydia Weigelt now value method forecast as the sales in each of the four quarters next year?
wonders what these seasonally adjusted call volumes would have been had she started using
d. Based on his assessment of the current state of the housing market, Ralph’s best
seasonal factors two years ago rather than applying them retrospectively now.
judgment is that the agency will sell 100 houses next year. Given this forecast for
a. Use only the call volumes in year 1 to determine the seasonal factors for year 2 the year, what is the quarter-by-quarter forecast according to the seasonal
(so that the “average” call volume for each quarter is just the actual call volume factors?
for that quarter in year 1).
10.9.* You are using the moving-average forecasting method based on the last four
b. Use these seasonal factors to determine the seasonally adjusted call volumes for observations. When making the forecast for the last period, the oldest of the four
year 2. observations was 1,945 and the forecast was 2,083. The true value for the last period then
c. Use the call volumes in years 1 and 2 to determine the seasonal factors for year turned out to be 1,977. What is your new forecast for the next period?
3.
d. Use the seasonal factors obtained in part c to determine the seasonally adjusted
call volumes for year 3. page 437
10.7. Even when the economy is holding steady, the unemployment rate tends to
10.10. You are using the moving-average forecasting method based on sales in the last
fluctuate because of seasonal effects. For example, unemployment generally goes up in
three months to forecast sales for the next month. When making the forecast for last month,
Quarter 3 (summer) as students (including new graduates) enter the labor market. The
sales for the third month prior to last month were 805. The forecast for last month was 782
unemployment rate then tends to go down in Quarter 4 (fall) as students return to school and
and then the actual sales turned out to be 793. What is your new forecast for next month?
temporary help is hired for the Christmas season. Therefore, using seasonal factors to obtain
10.11. After graduating from college with a degree in mathematical statistics, Ann
a seasonally adjusted unemployment rate is helpful for painting a truer picture of economic
Preston has been hired by the Monty Ward Company to apply statistical methods for
trends.
forecasting the company’s sales. For one of the company’s products, the moving-average
Over the past 10 years, one state’s average unemployment rates (not seasonally
method based on sales in the 10 most recent months already is being used. Ann’s first task is
adjusted) in Quarters 1, 2, 3, and 4 have been 6.2 percent, 6.0 percent, 7.5 percent, and 5.5
to update last month’s forecast to obtain the forecast for next month. She learns that the
percent, respectively. The overall average has been 6.3 percent.
forecast for last month was 1,551 and that the actual sales then turned out to be 1,532. She Calculate the forecast of sales for next month.
also learns that the sales for the 10th month before last month were 1,632. What is Ann’s 10.19.* Ben Swanson, owner and manager of Swanson’s Department Store, has decided
forecast for next month? to use statistical forecasting to get a better handle on the demand for his major products.
10.12. The J. J. Bone Company uses exponential smoothing to forecast the average daily However, Ben now needs to decide which forecasting method is most appropriate for each
call volume at its call center. The forecast for last month was 782, and then the actual value category of product. One category is major household appliances, such as washing
turned out to be 792. Obtain the forecast for next month for each of the following values of machines, which have a relatively stable sales level. Monthly sales of washing machines last
the smoothing constant: α = 0.1, 0.3, 0.5. year are shown below.
10.13.* You are using exponential smoothing to obtain monthly forecasts of the sales of a
certain product. The forecast for last month was 2,083, and then the actual sales turned out
to be 1,973. Obtain the forecast for next month for each of the following values of the
smoothing constant: α = 0.1, 0.3, 0.5.
10.14. Read the referenced article that fully describes the management science study
summarized in the application vignette presented in Section 10.3. Briefly describe how
forecasting was applied in this study. Then list the various financial and nonfinancial benefits
that resulted from this study. a. Considering that the sales level is relatively stable, which of the most basic
10.15. Three years ago, the Admissions Office for Ivy College began using exponential forecasting methods—the last-value method, the averaging method, or the
smoothing with a smoothing constant of 0.25 to forecast the number of applications for moving-average method—do you feel would be most appropriate for forecasting
admission each year. Based on previous experience, this process was begun with an initial future sales? Why?
estimate of 5,000 applications. The actual number of applications then turned out to be 4,600
E b. Use the last-value method retrospectively to determine what the forecasts would
in the first year. Thanks to new favorable ratings in national surveys, this number grew to
have been for the last 11 months of last year. What are MAD and MSE?
5,300 in the second year and 6,000 last year.
E c. Use the averaging method retrospectively to determine what the forecasts would
a. Determine the forecasts that were made for each of the past three years.
have been for the last 11 months of last year. What are MAD and MSE?
b. Calculate MAD and MSE for these three years.
E d. Use the moving-average method with n = 3 retrospectively to determine what the
c. Determine the forecast for next year. forecasts would have been for the last nine months of last year. What are MAD
10.16. Reconsider Problem 10.15. Notice the steady trend upward in the number of and MSE?
applications over the past three years— from 4,600 to 5,300 to 6,000. Suppose now that the e. Use their MAD values to compare the three methods.
Admissions Office of Ivy College had been able to foresee this kind of trend and so had
f. Use their MSE values to compare the three methods.
decided to use exponential smoothing with trend to do the forecasting. Suppose also that the
initial estimates just over three years ago had been average value = 3,900 and trend = g. Do you feel comfortable in drawing a definitive conclusion about which of the
700. Then, with any values of the smoothing constants, the forecasts obtained by this three forecasting methods should be the most accurate in the future based on
forecasting method would have been exactly correct for all three years. these 12 months of data?
Illustrate this fact by doing the calculations to obtain these forecasts when the smoothing
constant is α = 0.25 and the trend smoothing constant is β = 0.25.
10.17.* Exponential smoothing with trend, with a smoothing constant of α = 0.2 and a trend
page 438
smoothing constant of β = 0.3, is being used to forecast values in a time series. At this point,
E 10.20. Reconsider Problem 10.19. Ben Swanson now has decided to use the exponential
the last two values have been 535 and then 550. The last two forecasts have been 530 and
smoothing method to forecast future sales of washing machines, but he needs to decide on
then 540. The last estimate of trend has been 10. Use this information to forecast the next
which smoothing constant to use. Using an initial estimate of 24, apply this method
value in the time series.
retrospectively to the 12 months of last year with α = 0.1, 0.2, 0.3, 0.4, and 0.5. Compare
10.18. The Healthwise Company produces a variety of exercise equipment. Healthwise
MAD for these five values of the smoothing constant α. Then do the same with MSE.
management is very pleased with the increasing sales of its newest model of exercise
10.21. Management of the Jackson Manufacturing Corporation wishes to choose a
bicycle. The sales during the last two months have been 4,655 and then 4,935.
statistical forecasting method for forecasting total sales for the corporation. Total sales (in
Management has been using exponential smoothing with trend, with a smoothing
millions of dollars) for each month of last year are shown below.
constant of α = 0.1 and a trend smoothing constant of β = 0.2, to forecast sales for the next
month each time. The forecasts for the last two months were 4,720 and then 4,975. The last
estimate of trend was 240.
 E 10.24. The choice of the smoothing constants, α and β, has a considerable effect on the
accuracy of the forecasts obtained by using exponential smoothing with trend. For each of
the following time series, set α = 0.2 and then compare MAD obtained with β = 0.1, 0.2, 0.3,
0.4, and 0.5. Begin with initial estimates of 50 for the average value and 2 for the trend.
a. 52, 55, 55, 58, 59, 63, 64, 66, 67, 72, 73, 74
b. 52, 55, 59, 61, 66, 69, 71, 72, 73, 74, 73, 74
a. Note how the sales level is shifting significantly from month to month—first
c. 52, 53, 51, 50, 48, 47, 49, 52, 57, 62, 69, 74
trending upward and then dipping down before resuming an upward trend.
Assuming that similar patterns would continue in the future, evaluate how well 10.25. The Andes Mining Company mines and ships copper ore. The company’s sales
you feel each of the five forecasting methods introduced in Section 10.3 would manager, Juanita Valdes, has been using the moving-average method based on the last three
perform in forecasting future sales. years of sales to forecast the demand for the next year. However, she has become
dissatisfied with the inaccurate forecasts being provided by this method.
E b. Apply the last-value method, the averaging method, and the moving-average
The annual demands (in tons of copper ore) over the past 10 years are 382, 405, 398,
method (with n = 3) retrospectively to last year’s sales and compare their MAD
421, 426, 415, 443, 451, 446, 464.
values. Then compare their MSE values.
a. Explain why this pattern of demands inevitably led to significant inaccuracies in
E c. Using an initial estimate of 120, apply the exponential smoothing method
the moving-average forecasts.
retrospectively to last year’s sales with α = 0.1, 0.2, 0.3, 0.4, and 0.5. Compare
both MAD and MSE for these five values of the smoothing constant α. E b. Determine the moving-average forecasts for the past seven years. What are
MAD and MSE? What is the forecast for next year?
E d. Using initial estimates of 120 for the average value and 10 for the trend, apply
exponential smoothing with trend retrospectively to last year’s sales. Use all E c. Determine what the forecasts would have been for the past 10 years if the
combinations of the smoothing constants where α = 0.1, 0.3, or 0.5 and β = 0.1, exponential smoothing method had been used instead with an initial estimate of
0.3, or 0.5. Compare both MAD and MSE for these nine combinations. 380 and a smoothing constant of α = 0.5. What are MAD and MSE? What is the
e. Which one of the above forecasting methods would you recommend that forecast for next year?
management use? Using this method, what is the forecast of total sales for E d. Determine what the forecasts would have been for the past 10 years if
January of the new year? exponential smoothing with trend had been used instead. Use initial estimates of
10.22. Reconsider Problem 10.21. Use the lessons learned from the CCW case study to 370 for the average value and 10 for the trend, with smoothing constants α = 0.25
address the following questions. and β = 0.25.
e. Based on the MAD and MSE values, which of these three methods do you
a. What might be causing the significant shifts in total sales from month to month
recommend using hereafter?
that were observed last year?
b. Given your answer to part a, how might the basic statistical approach to E 10.26.* The Pentel Microchip Company has started production of its new microchip.
forecasting total sales be improved? The first phase in this production is the wafer fabrication process. Because of the great
difficulty in fabricating acceptable wafers, many of these tiny wafers must be rejected
c. Describe the role of managerial judgment in applying the statistical approach
because they are defective. Therefore, management places great emphasis on continually
developed in part b.
improving the wafer fabrication process to increase its production yield (the percentage of
E 10.23. Choosing an appropriate value of the smoothing constant α is a key decision when wafers fabricated in the current lot that are of acceptable quality for producing page 439
applying the exponential smoothing method. When relevant historical data exist, one microchips).
approach to making this decision is to apply the method retrospectively to these data with So far, the production yields of the respective lots have been 15 percent, 21 percent, 24
different values of α and then choose the value of α that gives the smallest MAD. Use this percent, 32 percent, 37 percent, 41 percent, 40 percent, 47 percent, 51 percent, and 53
approach for choosing α with each of the following time series representing monthly sales. percent. Use exponential smoothing with trend to forecast the production yield of the next
In each case, use an initial estimate of 50 and compare α = 0.1, 0.2, 0.3, 0.4, and 0.5. lot. Begin with initial estimates of 10 percent for the average value and 5 percent for the
a. 51, 48, 52, 49, 53, 49, 48, 51, 50, 49 trend. Use smoothing constants of α = 0.2 and β = 0.2.
10.27. The Centerville Water Department provides water for the entire town and outlying
b. 52, 50, 53, 51, 52, 48, 52, 53, 49, 52
areas. The number of acre-feet of water consumed in each of the four seasons of the three
c. 50, 52, 51, 55, 53, 56, 52, 55, 54, 53 preceding years is shown below.
 E 10.29. Transcontinental Airlines maintains a computerized forecasting system to forecast
the number of customers in each fare class who will fly on each flight in order to allocate the
available reservations to fare classes properly. For example, consider economy-class
customers flying in midweek on the noon flight from New York to Los Angeles. The
following table shows the average number of such passengers during each month of the year
just completed. The table also shows the seasonal factor that has been assigned to each
month based on historical data.
E a. Determine the seasonal factors for the four seasons.
E b. After considering seasonal effects, use the last-value method to forecast water
consumption next winter.
c. Assuming that each of the forecasts for the next three seasons is correct, what
would the last-value method forecast as the water consumption in each of the
four seasons next year?
E d. After considering seasonal effects, use the averaging method to forecast water
consumption next winter.
E e. After considering seasonal effects, use the moving-average method based on four
a. After considering seasonal effects, compare both the MAD and MSE values for
seasons to forecast water consumption next winter.
the last-value method, the averaging method, the moving-average method (based
f. After considering seasonal effects, use the exponential smoothing method with an on the most recent three months), and the exponential smoothing method (with an
initial estimate of 46 and a smoothing constant of α = 0.1 to forecast water initial estimate of 80 and a smoothing constant of α = 0.2) when they are applied
consumption next winter. retrospectively to the past year.
E g. Compare both the MAD and MSE values of these four forecasting methods b. Use the forecasting method with the smallest MAD value to forecast the average
when they are applied retrospectively to the last three years. number of these passengers flying in January of the new year.
10.28. Reconsider Problem 10.8. Ralph Billett realizes that the last-value method is 10.30. Reconsider Problem 10.29. The economy is beginning to boom so the management
considered to be the naive forecasting method, so he wonders whether he should be using of Transcontinental Airlines is predicting that the number of people flying will steadily
another method. Therefore, he has decided to use the available Excel templates that increase this year over the relatively flat (seasonally adjusted) level of last year. Since the
consider seasonal effects to apply various statistical forecasting methods retrospectively to forecasting methods considered in Problem 10.29 are relatively slow in adjusting to such a
the past three years of data and compare both their MAD and MSE values. trend, consideration is being given to switching to exponential smoothing with trend.
E a. Determine the seasonal factors for the four quarters. Subsequently, as the year goes on, management’s prediction proves to be true. The
following table shows the corresponding average number of passengers in each month of the
E b. Apply the last-value method. new year.
E c. Apply the averaging method.
d. Apply the moving-average method based on the four most recent quarters of
E
data.
page 440
E e. Apply the exponential smoothing method with an initial estimate of 25 and a
smoothing constant of α = 0.25.
E f. Apply exponential smoothing with trend with smoothing constants of α = 0.25 and
β = 0.25. Use initial estimates of 25 for the average value and 0 for the trend.
E g. Compare both the MAD and MSE values for these methods. Use the one with
the smallest MAD to forecast sales in Quarter 1 of next year.
h. Use the forecast in part g and the seasonal factors to make long-range forecasts E a. Repeat part a of Problem 10.29 for the two years of data.
now of the sales in the remaining quarters of next year.
E b. After considering seasonal effects, apply exponential smoothing with trend to just
the new year. Use initial estimates of 80 for the average value and 2 for the
 trend, along with smoothing constants of α = 0.2 and β = 0.2. Compare MAD for and 0 for the trend, along with smoothing constants of α = 0.2 and β = 0.2.
this method to the MAD values obtained in part a. Then do the same with MSE. e. Compare both the MAD and MSE values obtained in parts b, c, and d.
c. Repeat part b when exponential smoothing with trend is begun at the beginning of f. Calculate the combined forecast for each month by averaging the forecasts for
the first year and then applied to both years, just like the other forecasting that month obtained in parts b, c, and d. Then calculate MAD for these combined
methods in part a. Use the same initial estimates and smoothing constants except forecasts.
change the initial estimate of trend to 0.
g. Based on these results, what is your recommendation for how to do the forecasts
d. Based on these results, which forecasting method would you recommend that next year?
Transcontinental Airlines use hereafter?
10.31. Quality Bikes is a wholesale firm that specializes in the distribution of bicycles. In
the past, the company has maintained ample inventories of bicycles to enable filling orders page 441
immediately, so informal rough forecasts of demand were sufficient to make the decisions on
when to replenish inventory. However, the company’s new president, Marcia Salgo, intends 10.32.* Long a market leader in the production of heavy machinery, the Spellman
to run a tighter ship. Scientific inventory management is to be used to reduce inventory levels Corporation recently has been enjoying a steady increase in the sales of its new lathe. The
and minimize total variable inventory costs. At the same time, Marcia has ordered the sales over the past 10 months are shown below.
development of a computerized forecasting system based on statistical forecasting that
considers seasonal effects. The system is to generate three sets of forecasts—one based on
the moving-average method, a second based on the exponential smoothing method, and a
third based on exponential smoothing with trend. The average of these three forecasts for
each month is to be used for inventory management purposes.
The following table gives the available data on monthly sales of 10-speed bicycles over
the past three years. The last column also shows monthly sales this year, which is the first
year of operation of the new forecasting system.

Because of this steady increase, management has decided to use causal forecasting,
with the month as the independent variable and sales as the dependent variable, to forecast
sales in the coming months.
a. Plot these data on a two-dimensional graph with the month on the horizontal axis
and sales on the vertical axis.
E b. Find the formula for the linear regression line that fits these data.
c. Plot this line on the graph constructed in part a.
d. Use this line to forecast sales in month 11.
e. Use this line to forecast sales in month 20.
f. What does the formula for the linear regression line indicate is roughly the
average growth in sales per month?
E a. Determine the seasonal factors for the 12 months based on past sales. 10.33. Reconsider Problems 10.15 and 10.16. Since the number of applications for
E b. After considering seasonal effects, apply the moving-average method based on admission submitted to Ivy College has been increasing at a steady rate, causal forecasting
the most recent three months to forecast monthly sales for each month of this can be used to forecast the number of applications in future years by letting the year be the
year. independent variable and the number of applications be the dependent variable.
E c. After considering seasonal effects, apply the exponential smoothing method to a. Plot the data for years 1, 2, and 3 on a two-dimensional graph with the year on
forecast monthly sales this year. Use an initial estimate of 420 and a smoothing the horizontal axis and the number of applications on the vertical axis.
constant of α = 0.2. b. Since the three points in this graph line up in a straight line, this straight line is the
linear regression line. Draw this line.
E d. After considering seasonal effects, apply exponential smoothing with trend to
forecast monthly sales this year. Use initial estimates of 420 for the average value E c. Find the formula for this linear regression line.
 d. Use this line to forecast the number of applications for each of the next five years
(years 4 through 8).
e. As these next years go on, conditions change for the worse at Ivy College. The
favorable ratings in the national surveys that had propelled the growth in
applications turn unfavorable. Consequently, the number of applications turn out to
be 6,300 in year 4 and 6,200 in year 5, followed by sizable drops to 5,600 in year 6 page 442
and 5,200 in year 7. Does it still make sense to use the forecast for year 8
obtained in part d? Explain. a. To use causal forecasting to forecast sales for a given amount of advertising,
E f. Plot the data for all seven years. Find the formula for the linear regression line which need to be the dependent variable and the independent variable?
based on all these data and plot this line. Use this formula to forecast the number b. Plot the data on a graph.
of applications for year 8. Does the linear regression line provide a close fit to the
data? Given this answer, do you have much confidence in the forecast it provides E c. Find the formula for the linear regression line that fits these data. Then plot this
for year 8? Does it make sense to continue to use a linear regression line when line on the graph constructed in part b.
changing conditions cause a large shift in the underlying trend in the data? d. Forecast the sales that would be attained by expending $300,000 on advertising.
E g. Apply exponential smoothing with trend to all seven years of data to forecast the e. Estimate the amount of advertising that would need to be done to attain a booking
number of applications in year 8. Use initial estimates of 3,900 for the average of 22,000 passengers.
and 700 for the trend, along with smoothing constants of α = 0.5 and β = 0.5. f. According to the linear regression line, about how much increase in sales can be
When the underlying trend in the data stays the same, causal forecasting provides attained on the average per $1,000 increase in the amount of advertising?
the best possible linear regression line (according to the method of least squares) 10.36. To support its large fleet, North American Airlines maintains an extensive
for making forecasts. However, when changing conditions cause a shift in the inventory of spare parts, including wing flaps. The number of wing flaps needed in inventory
underlying trend, what advantage does exponential smoothing with trend have to replace damaged wing flaps each month depends partially on the number of flying hours
over causal forecasting? for the fleet that month, since increased usage increases the chances of damage.
10.34. Reconsider Problem 10.25. Despite some fluctuations from year to year, note that The following table shows both the number of replacement wing flaps needed and the
there has been a basic trend upward in the annual demand for copper ore over the past 10 number of thousands of flying hours for the entire fleet for each of several recent months.
years. Therefore, by projecting this trend forward, causal forecasting can be used to
forecast demands in future years by letting the year be the independent variable and the
demand be the dependent variable.
a. Plot the data for the past 10 years (years 1 through 10) on a two-dimensional
graph with the year on the horizontal axis and the demand on the vertical axis.
E b. Find the formula for the linear regression line that fits these data. a. Identify the dependent variable and the independent variable for doing causal
c. Plot this line on the graph constructed in part a. forecasting of the number of wing flaps needed for a given number of flying
d. Use this line to forecast demand next year (year 11). hours.
e. Use this line to forecast demand in year 15. b. Plot the data on a graph.
f. What does the formula for the linear regression line indicate is roughly the E c. Find the formula for the linear regression line. d. Plot this line on the graph
average growth in demand per year? constructed in part b.
10.35. Luxury Cruise Lines has a fleet of ships that travel to Alaska repeatedly every d. Forecast the average number of wing flaps needed in a month in which 150,000
summer (and elsewhere during other times of the year). A considerable amount of flying hours are planned.
advertising is done each winter to help generate enough passenger business for that summer. e. Repeat part e for 200,000 flying hours.
With the coming of a new winter, a decision needs to be made about how much advertising E 10.37. Joe Barnes is the owner of Standing Tall, one of the major roofing companies in
to do this year. town. Much of the company’s business comes from building roofs on new houses. Joe has
The following table shows the amount of advertising (in thousands of dollars) and the learned that general contractors constructing new houses typically will subcontract the
resulting sales (in thousands of passengers booked for a cruise) for each of the past five roofing work about two months after construction begins. Therefore, to help him develop
years. long-range schedules for his work crews, Joe has decided to use county records on the
 number of housing construction permits issued each month to forecast the number of roofing One year after proposing his vision to management, Mark received the go-ahead from Cutting Edge
jobs on new houses he will have two months later. corporate headquarters. He prepared his “to do” list—specifying computer and phone systems
Joe has now gathered the following data for each month over the past year, where the requirements, installing hardware and software, integrating data from the 35 separate administration
second column gives the number of housing construction permits issued in that month and centers, standardizing record-keeping and response procedures, and staffing the administration center.
the third column shows the number of roofing jobs on new houses that were subcontracted Mark delegated the systems requirements, installation, and integration jobs to a competent group of
out to Standing Tall in that month. Use a causal forecasting approach to develop a technology specialists. He took on the responsibility of standardizing procedures and staffing the
forecasting procedure for Joe to use hereafter. administration center.
Mark had spent many years in human resources and therefore had little problem with standardizing
record-keeping and response procedures. He encountered trouble in determining the number of
representatives needed to staff the center, however. He was particularly worried about staffing the call
center since the representatives answering phones interact directly with customers—the 60,000 Cutting
Edge employees. The customer service representatives would receive extensive training so that they
would know the records and benefits policies backwards and forwards—enabling them to answer
questions accurately and process changes efficiently. Overstaffing would cause Mark to suffer the high
costs of training unneeded representatives and paying the surplus representatives the high salaries that
go along with such an intense job. Understaffing would cause Mark to continue to suffer the headaches
from customer complaints— something he definitely wanted to avoid.
10.38. Read the referenced article that fully describes the management science study
summarized in the application vignette presented in Section 10.6. Briefly describe how The number of customer service representatives Mark needed to hire
forecasting methods, including judgmental forecasting methods, were applied as part of this depended on the number of calls that the records and benefits call center
study. Then list the various financial and nonfinancial benefits that resulted from this study.
would receive. Mark therefore needed to forecast the number of calls that the
new centralized center would receive. He approached the forecasting
problem by using judgmental forecasting. He studied data from one of the 35
decentralized administration centers and learned that the decentralized center
Case 10-1 had serviced 15,000 customers and had received 2,000 calls per month. He
concluded that since the new centralized center would service four times the
Finagling the Forecasts number of customers—60,000 customers—it would receive four times the
Mark Lawrence has been pursuing a vision for more than two years. This pursuit began when he number of calls—8,000 calls per month.
became frustrated in his role as director of Human Resources at Cutting Edge, a large company
manufacturing computers and computer peripherals. At that time, the Human Resources page 443
Mark slowly checked off the items on his “to do” list, and the centralized
Department under his direction provided records and benefits administration to the 60,000 records and benefits administration center opened one year after Mark had
Cutting Edge employees throughout the United States, and 35 separate records and benefits received the go-ahead from corporate headquarters.
administration centers existed across the country. Employees contacted these records and benefits
centers to obtain information about dental plans and stock options, change tax forms and personal Now, after operating the new center for 13 weeks, Mark’s call center
information, and process leaves of absence and retirements. The decentralization of these administration forecasts are proving to be terribly inaccurate. The number of calls the center
centers caused numerous headaches for Mark. He had to deal with employee complaints often since receives is roughly three times as large as the 8,000 calls per month that
each center interpreted company policies differently—communicating inconsistent and sometimes
inaccurate answers to employees. His department also suffered high operating costs since operating 35 Mark had forecasted. Because of demand overload, the call center is slowly
separate centers created inefficiency. going to hell in a handbasket. Customers calling the center must wait an
His vision? To centralize records and benefits administration by establishing one administration average of five minutes before speaking to a representative, and Mark is
center. This centralized records and benefits administration center would perform two distinct functions:
data management and customer service. The data management function would include updating
receiving numerous complaints. At the same time, the customer service
employee records after performance reviews and maintaining the human resource management system. representatives are unhappy and on the verge of quitting because of the stress
The customer service function would include establishing a call center to answer employee questions created by the demand overload. Even corporate headquarters has become
concerning records and benefits and to process records and benefits changes over the phone.
aware of the staff and service inadequacies, and executives have been What is the day-by-day seasonally adjusted call volume for the past 13
breathing down Mark’s neck demanding improvements. weeks?
Mark needs help, and he approaches you to forecast demand for the call 2. Using the seasonally adjusted call volume, forecast the daily demand for
center more accurately. the next week using the last-value forecasting method.
Luckily, when Mark first established the call center, he realized the 3. Using the seasonally adjusted call volume, forecast the daily demand for
importance of keeping operational data, and he provides you with the number the next week using the averaging forecasting method.
of calls received on each day of the week over the last 13 weeks. The data
(shown next) begins in week 44 of the last year and continues to week 5 of 4. Using the seasonally adjusted call volume, forecast the daily demand for
the current year. the next week using the moving- average forecasting method. You decide
to use the five most recent days in this analysis.
5. Using the seasonally adjusted call volume, forecast the daily demand for
the next week using the exponential smoothing forecasting method. You
decide to use a smoothing constant of 0.1 because you believe that
demand with seasonal effects remains relatively stable. Use the daily
call volume average over the past 13 weeks for the initial estimate.
b. After one week, the period you have forecasted passes. You realize that you are able to determine
the accuracy of your forecasts because you now have the actual call volumes from the week you
had forecasted. The actual call volumes are shown below.

page 444
c. At the end of the hour, Mark arrives at your desk with two data sets: weekly case volumes for the
Mark indicates that the days where no calls were received were decentralized center and weekly case volumes for the centralized center. You ask Mark if he has
holidays. data for daily case volumes, and he tells you that he does not. You therefore first have to forecast the
weekly demand for the next week and then break this weekly demand into daily demand.
a. Mark first asks you to forecast daily demand for the next week using the data from the past 13
weeks. You should make the forecasts for all the days of the next week now (at the end of week 5), The decentralized center was shut down last year when the new
but you should provide a different forecast for each day of the week by treating the forecast for a centralized center opened, so you have the decentralized case data spanning
single day as being the actual call volume on that day.
from week 44 of two years ago to week 5 of last year. You compare this
1. From working at the records and benefits administration center, you decentralized data to the centralized data spanning from week 44 of last year
know that demand follows “seasonal” patterns within the week. For to week 5 of this year. The weekly case volumes are shown in the following
example, more employees call at the beginning of the week when they table.
are fresh and productive than at the end of the week when they are For each of the forecasting methods, calculate the mean absolute
planning for the weekend. You therefore realize that you must account for deviation for the method and evaluate the performance of the method. When
the seasonal patterns and adjust the data that Mark gave you accordingly. calculating the mean absolute deviation, you should use the actual forecasts
you found in part a above. You should not recalculate the forecasts based on
 the actual values. In your evaluation, provide an explanation for the weekly case volume; you need the daily call volume. To calculate call
effectiveness or ineffectiveness of the method. volume from case volume, you perform further analysis and determine that
You realize that the forecasting methods that you have investigated do not each case generates an average of 1.5 calls. To calculate daily call volume
provide a great degree of accuracy, and you decide to use a creative from weekly call volume, you decide to use the seasonal factors as page 445
approach to forecasting that combines the statistical and judgmental conversion factors. Given the following case volume data from the
approaches. You know that Mark had used data from one of the 35 decentralized center for week 6 of last year, forecast the daily call volume
decentralized records and benefits administration centers to perform his for the new center for week 6 of this year.
original forecasting. You therefore suspect that call volume data exists for
Decentralized Case Centralized Case
this decentralized center. Because the decentralized centers performed the Volume Volume
same functions as the new centralized center currently performs, you decide Week 44 612 2,052
that the call volumes from the decentralized center will help you forecast the Week 45 721 2,170
call volumes for the new centralized center. You simply need to understand Week 46 693 2,779
how the decentralized volumes relate to the new centralized volumes. Once Week 47 540 2,334
you understand this relationship, you can use the call volumes from the Week 48 1,386 2,514
decentralized center to forecast the call volumes for the centralized center. Week 49 577 1,713
You approach Mark and ask him whether call center data exist for the Week 50 405 1,927
decentralized center. He tells you that data exist, but data do not exist in the Week 51 441 1,167
format that you need. Case volume data—not call volume data—exist. You Week 52/1 655 1,549
do not understand the distinction, so Mark continues his explanation. There Week 2 572 2,126
are two types of demand data—case volume data and call volume data. Case Week 3 475 2,337
volume data count the actions taken by the representatives at the call center. Week 4 530 1,916
Call volume data count the number of calls answered by the representatives Week 5 595 2,098
at the call center. A case may require one call or multiple calls to resolve it.
Thus, the number of cases is always less than or equal to the number of calls.
Week 6
You know you only have case volume data for the decentralized center,
Decentralized case volume 613
and you certainly do not want to compare apples and oranges. You therefore
ask if case volume data exist for the new centralized center. Mark gives you a
wicked grin and nods his head. He sees where you are going with your . Using the actual call volumes given in part b, calculate the mean absolute
forecasts, and he tells you that he will have the data for you within the hour. deviation and evaluate the effectiveness of this forecasting method.
. Which forecasting method would you recommend Mark use and why? As the
. Find a mathematical relationship between the decentralized case volume
call center continues its operation, how would you recommend improving
data and the centralized case volume data.
the forecasting procedure?
. Now that you have a relationship between the weekly decentralized case
volume and the weekly centralized case volume, you are able to forecast the
weekly case volume for the new center. Unfortunately, you do not need the Additional Cases
Additional cases for this chapter also are available at the University of Western Ontario Ivey School of
Business website, cases.ivey.uwo.ca/cases, in the segment of the CaseMate area designated for this
book.

1 This forecast also can be projected ahead to subsequent quarters, but we are focusing on
just the next quarter.