Chapter 8 Terminology

Decision
Analysis States of
nature

Payoffs /
payoff table

MGS3100 Probability
Julie Liggett De Jong

The payoff table is a fundamental

component in decision analysis Terminology
models
Expected Return
State of Nature
Decision 1 2 … m Regret
d1 r11 r12 … r1m

d2 r21 r22 … r2m EVPI

… … … … …
dn rn1 rn2 … rnm EVSI
Table 1, p81

Three Classes of Decisions

Decision Models under
certainty
Decisions under:

certainty

risk

uncertainty

1
If I know for sure that it will be raining If I know for sure that it will be raining
when I leave work this afternoon, when I leave work this afternoon,
should I take my umbrella to work should I take my umbrella to work
today? today?

Rain

Take Umbrella 0
Do Not -7.00

Table 2, p82

Decisions under risk Multiple states of nature

We size up the likelihood of each Historical frequencies

state of nature happening

2
 Historical frequencies We calculate Expected Returns

Subjective estimates

We choose the alternative that yields the maximum

E(X) = Σpixi expected return. In other words, i* is the optimal
decision where

ERi* = maximum overall i of ERi

All-Ways-Open Market The Newsvendor Model

c) Calculate expected values: expected shortage
(S) & expected excess (E) inventory Selling Price: $ .75
Purchase Price: $ .40
Week RN Demand Prob. S E Exp(s) Exp(E) Goodwill cost: $ .50
1 .97 45 0.14 3 0.14*3=0.42
2 .02 40 0.08 2 0.08*2=0.16
3 .80 44 0.16 2 0.16*2=0.32
4 .66 43 0.18 1 0.18*1=0.18
5 .96 45 0.14 3 0.14*3=0.42
6 .55 43 0.18 1 0.18*1=0.18
7 .50 42 0.24
8 .29 42 0.24
9 .58 43 0.18 1 0.18*1=0.18
10 .51 42 0.24
Expected 1.70 0.16

3
 The Newsvendor Model The Newsvendor Model

A B C D E Selling Price: $ .75

1 Selling Price 75 Purchase Price: $ .40
2 Purchase Cost 40 Goodwill cost: $ .50
3 Goodwill Cost 50
4
5 States of Nature
6 Decision 0 1 2 3
7 0 0 -50 -100 -150 Demand distribution:
8 1 -40 35 -15 -65
9 2 -80 -5 70 20
10 3 -120 -45 30 105 P0 = Prob(demand = 0) = 0.1
P1 = Prob(demand = 1) = 0.3
P2 = Prob(demand = 2) = 0.4
P3 = Prob(demand = 3) = 0.2
Table 4, p84

The Newsvendor Model

Decisions
1
A
Selling Price
B
75
C D E F under
2
3
Purchase Cost
Goodwill Cost
40
50
uncertainty
4
5 States of Nature
6
7
Decision
0
0
0
1
-50
2
-100
3 Expected Return
-150 -85
Multiple states
8
9
1
2
-40
-80
35
-5
-15
70
-65
20
-12.5
22.5
of nature
10 3 -120 -45 30 105 7.5
11
12 Probabilities 0.1 0.3 0.4 0.2 Don’t know what
state of nature
What is the Expected Return? will occur

Decisions Laplace
under
uncertainty

Laplace
Maximin
Maximax
Minimax regret

4
 The Newsvendor Model
Laplace
A B C D E F
1 Selling Price 75

Assume all 2
3
Purchase Cost
Goodwill Cost
40
50
states of nature 4
5 States of Nature
are equally likely 6
7
Decision
0
0
0
1
-50
2
-100
3 Expected Return
-150 -85
to occur 8
9
1
2
-40
-80
35
-5
-15
70
-65
20
-12.5
22.5
10 3 -120 -45 30 105 7.5
11
12 Probabilities 0.1 0.3 0.4 0.2

What is the Expected Return?

Maximin Maximin

extremely Evaluate
conservative or minimum
pessimistic possible return
approach to associated with
making each decision.
decisions

Maximin
Maximin

Select decision
yielding
maximum
max
value of
minimum
min
returns.

Table 1, p81

5
 Different criterion yields different decisions.
Consider the decision table below: Maximax

optimistic approach to
making decisions

• Under the Maximin criterion, you would

choose decision 1.

• Under the Maximax criterion, you would

choose decision 2.

Maximax Maximax

Evaluate maximum Select decision yielding

possible return maximum
max of these
associated with each maximum
max returns.
decision

Maximax Different criterion yields different decisions.

• Under the Maximin criterion, you would

choose decision 1.

• Under the Maximax criterion, you would

choose decision 2.

Which is the best choice?

6
Minimax regret Regret measures
the desirability of
an outcome.

a) Find the
Choose the maximum
decision that value in
minimizes the column 1
regret for making
that choice. b)Subtract every
value in
column 1 from
this value

c) Repeat for
each column

a) Find the a) Find the

maximum maximum
value in value in
column 1 column 1

b)Subtract every b)Subtract every

value in value in
column 1 from column 1 from
this value this value

c) Repeat for c) Repeat for

each column each column

7
After regret table After regret table
is built: is built:

d)Choose the d)Choose the

maximum maximum
value in each value in each
row row

e) choose the e) Choose the

smallest smallest
(minimum of
the maximum)

Minimax Regret Each method yields different

A B C D E F
decisions regarding the newsvendor
1 Selling Price 75
2 Purchase Cost 40 data:
3 Goodwill Cost 50
4
5 States of Nature
6 Decision 0 1 2 3
7 0 0 -50 -100 -150
8 1 -40 35 -15 -65
Criteria Decision
9 2 -80 -5 70 20
10 3 -120 -45 30 105
LaPlace Cash Flow Order 2 papers
11 0 35 70 105
12 Maximin Cash Flow Order 1 paper
13 Regret MinMax Regret
14 0 0 85 170 255 255
15 1 40 0 85 170 170 Maximax Cash Flow Order 3 papers
16 2 80 40 0 85 85
17 3 120 80 40 0 120 Minimax Regret Order 2 papers
18
19 85

A B C D E F
1 Selling Price 75
2 Purchase Cost 40
3 Goodwill Cost 50
4
5 States of Nature
6 Decision 0 1 2 3 Expected Return
7 0 0 -50 -100 -150 -85
8 1 -40 35 -15 -65 -12.5
9 2 -80 -5 70 20 22.5
10 3 -120 -45 30 105 7.5
11
12 Probabilities 0.1 0.3 0.4 0.2

What is the most money the newsvendor

should be willing to pay for perfect
How much would you be willing to
information?
pay for perfect information?

8
 A B C D E F
1
2
Selling Price
Purchase Cost
75
40
Decision Trees
3 Goodwill Cost 50
4
5
6 Decision 0
States of Nature
1 2 3 Expected Return
Graphical tool
7
8
0
1
0
-40
-50
35
-100
-15
-150
-65
-85
-12.5
used to analyze
9 2 -80 -5 70 20 22.5 decisions under
10 3 -120 -45 30 105 7.5
11 risk
12 Probabilities 0.1 0.3 0.4 0.2

expected return maximum possible Useful to analyze

EVPI = with perfect - expected return sequences of
information without sample decisions
information

TreePlan Sonoralo
An add-in used to Cellular
draw decision Phones
trees in Excel.
3 strategies
Bayes’ Theorem
Allows us to
incorporate new
information into
the process.

Aggressive Basic

Major commitment Move production

to existing facility
Major capital
expenditure Modify current line

Large inventory Maintain inventory

for popular items
Major global
marketing Local/regional
campaign advertising

9
 Cautious States of
Nature
Use excess
capacity
Strong Demand
Minimize retooling (S)

Produce enough to Weak Demand

satisfy demand (W)
Advertise at
discretion of local
dealer

Sonoralo Cellular Phones A square node represents a point

Payoff table at which a decision must be made.

Each line (branch) leading from a

square represents a possible
decision.

TREE PLAN
A square node represents a point
at which a decision must be made. • Insert the CD into the CD-ROM drive.
• Select Run... from the Windows Start
menu.
A circular node represents an • Type d:\html\TreePlan\Treeplan.xla &
event (a situation when the select "OK".
outcome is not certain). • TreePlan will launch in Microsoft Excel
as an add-in to the Tools menu.
Each line (branch) leading from a • In the Microsoft Excel dialog box,
circle represents a possible select Enable Macros.
outcome. • For additional assistance go to Help.

10
 The Completed Decision Tree

Decision Trees:
Incorporating New
Information

Before implementing the Terminology

Basic strategy, the
corporate marketing
research group performs a
marketing study and Prior Probabilities
reports on whether the
study is encouraging (E) or
Conditional Probabilities / Reliabilities
discouraging (D).

Joint & Marginal Probabilities

We will consider the new
information before we make Posterior Probabilities
a decision.

A MARKET RESEARCH STUDY FOR A MARKET RESEARCH STUDY FOR

CELLULAR PHONES CELLULAR PHONES

Prior Probabilities: Conditional Probabilities / Reliabilities:

Initial estimates, such as P(S) and P(W). For two events A and B, the conditional
probability [P(A|B)], is the probability of event
A occurs given that event B will occur.
Sonorola has estimated the prior probabilities
as P(S) = 0.45 and P(W) = 0.55.

11
A MARKET RESEARCH STUDY FOR A MARKET RESEARCH STUDY FOR
CELLULAR PHONES CELLULAR PHONES

Conditional Probabilities / Reliabilities: Conditional Probabilities / Reliabilities:

For example, P(E|S) is the conditional If marketing were perfectly reliable, P(E|S) = 1.
probability that marketing gives an
encouraging report given that the market is in
fact going to be strong.

A MARKET RESEARCH STUDY FOR

Marketing has the following “track record” in CELLULAR PHONES
predicting the market:
Posterior Probabilities:
P(E|S) = 0.6
P(D|S) = 1 - P(E|S) = 0.4 Conditional probabilities, such as P(S|E).

P(D|W) = 0.7 We’ll use Bayes’ Theorem to calculate the

posterior probabilities.
P(E|W) = 1 - P(D|W) = 0.3

Calculating Posterior Probabilities:

1. Enter given Reliabilities (conditional P(E|W)

P(D|W)
probabilities).
2. Calculate Joint Probabilities by multiplying
Reliabilities by Prior Probabilities. P(S) P(W)

3. Compute Marginal Probabilities by summing the

P(E&S)
entries in each row.
4. Generate Posterior Probabilities by dividing each
row entry of joint probability table by its row sum.

P(W|E)
P(W|D)

12
A new decision tree!

) 30
P(S|E
S P(W|E)
IV W -8
A P(S|E
) 20 How much should
S
B
II
C
V
W
P(W|E)
7 we be willing to
S P(S|E) 5
VI P(W |E) spend on sample
E)

E W 15
P(

I )
30 information?
S P(S|D
D
P(

P(W|D)
VII W -8
D
)

A P(S|D) 20
B S
III VIII P(W|D)
W 7
C P(S|D) 5
S
IX P(W |D
)
W
15

THE EXPECTED VALUE OF SAMPLE INFORMATION

Sonoralo Cellular Phones
maximum possible maximum possible
Payoff table (w/out sample info) EVSI = expected return - expected return
with sample without sample
information information

EVSI = 13.46 – 12.85 = $0.61 million.

EVSI is the upper bound of how much one would be

willing to pay for this particular sample information.

THE EXPECTED VALUE OF PERFECT INFORMATION THE EXPECTED RETURN WITH PERFECT
INFORMATION
expected return maximum possible
EVPI = with perfect - expected return
information without perfect
information

EVPI is the maximum possible increase in the

expected return that can be obtained from new
information. ERPI = 30(0.45) + 15(0.55) = 21.75

13
THE EXPECTED VALUE OF PERFECT INFORMATION

expected return maximum possible maximum possible maximum possible

EVPI = with perfect - expected return EVSI = expected return - expected return
information without perfect with sample without sample
information information information

EVPI = 21.75 – 12.85 = $8.90 million

expected return maximum possible
EVPI = with perfect - expected return
information without perfect
information
EVPI is the maximum possible increase in the
expected return that can be obtained from new
information.

Sequential Decisions:
To Test or Not to Test

The value in performing the market research test

depends on how Sonorola uses the information
generated by the test.

The value of an initial decision depends on a

sequence of decisions and uncertain events that will
follow the initial decision. This is called a sequential
decision model.
model

14