7 Bayesian analysis

 ZHAN, Wenjie (Professor) 詹文杰

 School of Management, Huazhong Unive
rsity of Science & Technology
 Tel: 027-87556472
 Email: [email protected]
 7 Bayesian analysis
 Objectives:
 How to revise the prior probability in the ligh
t of new information.
 Why? – To know the Bayes’ theorem.
 Reliable of new information.
 Assessing the value of new information
 7 Bayesian analysis
 7.1 Introduction
 7.2 Bayes’ Theorem
 7.3 Application of Bayesian analysis
 7.4 Reliable of New Information
 7.5 Assessing the Value of New Information
 7.6 Examples for Bayesian analysis
 7.1 Introduction
 In this chapter we will look at the process of r
evising initial probability estimates in the light
of new information.
 The focus of our discussion will be Bayes’ the
orem, which is named after an English clergy
man,Thomas Bayes, whose ideas were publis
hed posthumously in 1763.
 Bayes’ theorem will be used as a normative t
ool, telling us how we should revise our prob
ability assessments when new information be
comes available.
 7.1 Introduction
 In Bayes’ theorem an initial probability
estimate is known as a prior probability.

 When Bayes’ theorem is used to modif

y a prior probability in the light of
new information the result is known
as a posterior probability.
 Example 1: assistant’s
forecast is high (1/6)
 Prior probability:
 A company’s sales manager estimates that there is a 0.2 probability
that sales in the coming year will be high, a 0.7 probability that they
will be medium and a 0.1 probability that they will be low:
 p (High)=0.2, p (Medium)=0.7, p (Low)=0.1
 New information 1:
 She then receives a sales forecast from her assistant and the
forecast suggests that sales will be high.
 Posterior probability:
 What should be the sales manager’s revised estimates of the
probability of (a) high sales, (b) medium sales and (c) low sales?
 p (High | high sales forecast)=?
 p (Medium | high sales forecast)=?
 p (Low | high sales forecast)=?
 Example 1: assistant’s
forecast is high (2/6)
 The accuracy of the assistant’s forecast
high:
 By examining the track record of the assistant’s
forecasts she is able to obtain the following
probabilities:

 P (high sales forecast | the market generated high sales) = 0.9

 P (high sales forecast | the market generated medium sales) = 0.6
 P (high sales forecast | the market generated low sales) = 0.3
Example 1: assistant’s
forecast is high (3/6)
 Example 1: assistant’s
forecast is high (4/6)
 The joint probabilities are:
 P (high sales occur and high sales forecast) = 0.
2 × 0.9 = 0.18
 P (medium sales occur and high sales forecast)
= 0.7 × 0.6 = 0.42
 p(low sales occur and high sales forecast) = 0.1
× 0.3 = 0.03
 so the sum of the joint probabilities is 0.63
 Example 1: assistant’s
forecast is high (5/6)
 Posterior probabilities:
 P (high | high sales forecast )
= 0.18/0.63 = 0.2857
 P (medium | high sales forecast )

= 0.42/0.63 = 0.6667
 P (low | high sales forecast )

= 0.03/0.63 = 0.0476
 Steps to get posterior probability
(6/6)
 (1) Construct a tree with branches representing all the possible
events which can occur and write the prior probabilities for these
events on the branches.
 (2) Extend the tree by attaching to each branch a new branch
which represents the new information which you have obtained. On
each branch write the conditional probability of obtaining this
information given the circumstance represented by the preceding
branch.
 (3) Obtain the joint probabilities by multiplying each prior
probability by the conditional probability which follows it on the
tree.
 (4) Sum the joint probabilities.
 (5) Divide the ‘appropriate’ joint probability by the sum of the joint
probabilities to obtain the required posterior probability.
 Example 2: assistant’s
forecast is medium (1/3)
 Prior probability:
 A company’s sales manager estimates that there is a 0.2 probability
that sales in the coming year will be high, a 0.7 probability that they
will be medium and a 0.1 probability that they will be low:
 p (High)=0.2, p (Medium)=0.7, p (Low)=0.1
 New information 1:
 She then receives a sales forecast from her assistant and the
forecast suggests that sales will be medium.
 Posterior probability:
 What should be the sales manager’s revised estimates of the
probability of (a) high sales, (b) medium sales and (c) low sales?
 p (high | medium sales forecast)=?
 p (medium| medium sales forecast)=?
 p (low | medium sales forecast)=?
 Example 2: assistant’s
forecast is medium (2/3)
 The accuracy of the assistant’s forecast
medium:
 By examining the track record of the assistant’s
forecasts she is able to obtain the following
probabilities:

 P (medium sales forecast | market generated high sales) = 0.05

 P (medium sales forecast | market generated medium sales) =0.2
 P (medium sales forecast | market generated low sales) = 0.1
 Example 2: assistant’s
forecast is medium (3/3)
Prior Prob. Con. Prob. Joint Prob. Post Prob.

p(fore mH)=0.20.05 p(H|fore m)=0.01/0.16

p(fore m|H)=0.05 =0.01 =0.0625
p(H)=0.2
p(fore mM)=0.70.2 p(M|fore m)=0.14/0.16
p(M)=0.7 p(fore m|M)=0.2 =0.14 =0.875

p(fore mL)=0.10.1 p(L|fore m)=0.01/0.16

p(fore m|L)=0.1 =0.01 =0.0625
p(L)=0.1

p(fore m)=0.01+0.14+0.01=0.16
 Example 3: assistant’s
forecast is low (1/3)
 Prior probability:
 A company’s sales manager estimates that there is a 0.2 probability
that sales in the coming year will be high, a 0.7 probability that they
will be medium and a 0.1 probability that they will be low:
 p (High)=0.2, p (Medium)=0.7, p (Low)=0.1
 New information 1:
 She then receives a sales forecast from her assistant and the
forecast suggests that sales will be low.
 Posterior probability:
 What should be the sales manager’s revised estimates of the
probability of (a) high sales, (b) medium sales and (c) low sales?
 p (high | low sales forecast)=?
 p (medium| low sales forecast)=?
 p (low | low sales forecast)=?
 Example 3: assistant’s
forecast is low (2/3)
 The accuracy of the assistant’s forecast
low:
 By examining the track record of the assistant’s
forecasts she is able to obtain the following
probabilities:

 P (low sales forecast | market generated high sales) = 0.05

 P (low sales forecast | market generated medium sales) =0.2
 P (low sales forecast | market generated low sales) = 0.6
 Example 3: assistant’s
forecast is low (3/3)
Prior Prob. Con. Prob. Joint Prob. Post Prob.

p(fore lH)=0.20.05 p(H|fore l)=0.01/0.21

p(fore l |H)=0.05 =0.01 =0.0476
p(H)=0.2
p(fore lM)=0.70.2 p(M|fore l)=0.14/0.21
p(M)=0.7 p(fore l |M)=0.2 =0.14 =0.6667

p(fore lL)=0.10.6 p(L|fore l)=0.06/0.21

p(fore l |L)=0.6 =0.06 =0.2857
p(L)=0.1

p(fore m)=0.01+0.14+0.01=0.21
 Summary of Example 1-3:
Accuracy of assistant’s forecast:
 (1) Conditional Probabilities if real sales level is high:
 p (forecast h| H) = 0.9
 p (forecast m| H) = 0.05
 p (forecast l |H) = 0.05

 (2) Conditional Probabilities if real sales level is Medium:

 p (forecast h| M) = 0.6
 p (forecast m| M) =0.2
 p (forecast l | M) =0.2

 (3) Conditional Probabilities if real sales level is Low:

 p (forecast h| L) = 0.3
 p (forecast m| L) = 0.1
 p (forecast l | L) = 0.6
 Marg. Prob. Prior Prob. Cond. Prob. Joint Prob. Post. Prob.

p(H)=0.2 p(fore h|H)=0.9 p(fore hH)=0.18 p(H|fore h)=0.2857

p(M)=0.7
p(fore h|M)=0.6 p(fore hM)=0.42 p(M|fore h)=0.6667
p(fore h)=0.63
p(L)=0.1
p(fore h|L)=0.3 p(fore hL)=0.03 p(L|fore h)=0.0476

p(H)=0.2 p(fore m|H)=0.05 p(fore mH)=0.01 p(H|fore m)=0.0625

p(fore m)=0.16 p(M)=0.7

p(fore m|M)=0.2 p(fore mM)=0.14 p(M|fore m)=0.875

p(L)=0.1
p(fore m|L)=0.1 p(fore mL)=0.01 p(L|fore m)=0.0625

p(H)=0.2 p(fore l|H)=0.05 p(fore lH)=0.01 p(H|fore l)=0.0476

p(fore l)=0.21
p(M)=0.7
p(fore l|M)=0.2 p(fore lM)=0.14 p(M|fore l)=0.6667

p(L)=0.1
p(fore l|L)=0.6 p(fore lL)=0.06 p(L|fore l)=0.2857
 7.2 Bayes’ theorem
 If the events A and B are not
independent, the multiplication rule is:
 p (A and B) = p (A) × p (B|A), then
p (B|A)= p (A and B) / p (A)

 p (A and B) = p (B) × p (A|B), then

p (A|B)= p (A and B) / p (B)
 7.2 Bayes’ theorem
 Let’s suppose:
 A1=high;
 A2=medium;
 A3=low;
 B=high forecast
 We have:
 A1, A2, and A3 are mutually exclusive events, and
P(A1)+ P(A2)+ P(A3)=1, which are named as the
prior probability.
 Events Ai (i=1,2,3) and B are dependent.
 7.2 Bayes’ theorem
 By examining the track record of the
assistant’s forecasts, we have the following
conditional probabilities:
 P(B|A1)=0.9
 P(B|A2)=0.6
 P(B|A3)=0.3
 The problems here are to get the posterior
probability:
 P(A1|B)=?
 P(A2|B)=?
 P(A3|B)=?
 7.2 Bayes’ theorem
 According to the multiplication rule:
 P(A1|B)= P (A1 and B) / P (B)
 P(A2|B)= P (A2 and B) / P (B)
 P(A3|B)= P (A3 and B) / P (B)

 The two problems are:

 P (Ai and B), i=1,2,3, which is named as joint
probability.
 P (B).
 7.2 Bayes’ theorem
 How to get P(Ai and B):
 According to the multiplication rule, we have two
ways to get P(Ai and B):
 (1) P(Ai and B)=P(Ai)×P(B|Ai)
 (2) P(Ai and B)=P(B)×P(Ai|B)
 We choose the equation (1), for the reason w
e know P(Ai) and P(B|Ai) , but don’t know the
P(B) and P(Ai|B).
 P(A1 and B)=P(A1)×P(B|A1) =0.2×0.9=0.18
 P(A2 and B)=P(A2)×P(B|A2) = 0.7 × 0.6 = 0.42
 P(A3 and B)=P(A3)×P(B|A3) = 0.1 × 0.3 =0.03
 7.2 Bayes’ theorem
 Getting the P (B) according to Theorem of total probabil
ity: If A1, A2, …,An are mutually exclusive, and constitu
te the sample space S that means that P(A1)+ P(A2)+
… P(An)=1; then for any event B in S :

P( B)  P( B and A1)  P( B and A2)  ...  P( B and An)

 P( A1)  P( B | A1)  P( A2)  P( B | A2)  ...  P( An)  P( B | An)

P (B) = sum joint probability.

The equation must satisfy two conditions:
(1) A1, A2, …, An are mutually exclusive;
(2) P(A1)+ P(A2)+… P(An)=1.
 7.2 Bayes’ theorem
 Getting the posterior probability accordi
ng to the multiplication rule:
 P(Ai|B)= P(Ai and B) / P(B)
= P(Ai and B) / sum [P(Ai and B)]
 7.3 Application of Bayesian
analysis
 The application of Bayesian analysis inv
olves the use of the posterior probabiliti
es, rather than the prior probabilities, in
the decision model.
 Example 4: (1/9)
 A retailer has to decide whether to hold a lar
ge or a small stock of a product for the comi
ng summer season.
 A payoff table for the courses of action and o
utcomes is shown below:
 Example 4: (2/9)
 The following table shows the retailer’s utilities
for the above sums of money (it can be assume
d that money is the only attribute which he is c
oncerned about):
 Example 4: (3/9)
 The retailer estimates that there is a 0.4
probability that sales will be low and a 0
.6 probability that they will be high.
 P(low) =0.4
 P(high) =0.6
 What level of stocks should he hold?
Example 4: (5/9)
 Example 4: (6/9)

Decision

Future sales Prior Hold small stocks Hold large stocks

level probability

Low 0.4 0.5 0

High 0.6 0.8 1
 Example 4: Decision with
posterior probability (6/9)
 New information:
 Before implementing his decision the retailer receives a sales
forecast which suggests that sales will be high.

 Conditional Probability:
 In the past when sales turned out to be high the forecast had
correctly predicted high sales on 75% of occasions.
p (high forecast | sales came to be high) =0.75
 However, in seasons when sales turned out to be low the
forecast had wrongly predicted high sales on 20% of
occasions.
p (high forecast | sales came to be low) =0.2
Example 4: Posterior
probability (7/9)
Example 4: Decision tree with
posterior probability (8/9)
 Example 4: Decision matrix
with posterior probability (9/9)

Decision

Future sales Posterior Hold small stocks Hold large stocks

level probability

Low 0.15 0.5 0

High 0.85 0.8 1
 7.4 Reliable of New Information
 Example 5:
 Suppose that a geologist is involved in a search for
new sources of natural gas in southern England.
 In one particular location he is asked to estimate,
on the basis of a preliminary survey, the probability
that gas will be found in that location.
 Having made his estimate, he will receive new
information from a test drilling.
 Example 5:
 Let us first consider a situation where the geologist is not
very confident about his estimate (prior probabilities) and
where the test drilling is very reliable.
 The ‘vague’ prior probability distribution that the
geologist can put forward is to assign probabilities of 0.5
to the two events ‘gas exists at the location’ and ‘gas
does not exist at the location’.
 Suppose that having put forward the prior probabilities of
0.5 and 0.5, the result of the test drilling is received.
 This indicates that gas is present and the result can be
regarded as 95% reliable. By this we mean that there is
only a 0.05 probability that it will give a misleading
indication.
Probability tree
7.4 Reliable of new information
 By considering the distance of the curves
from the diagonal line, it can be seen that the
more reliable the new information, the
greater will be the modification of the prior
probabilities.

 Question: True or false?

 At any given level of reliability, the modification is
relatively small where the prior probability is high,
while this modification is relatively big where the
prior probability is low.
 7.4 Reliable of new information
 Two extreme situation:
 (1) If your prior probability of an event occurring is zero,
then the posterior probability will also be zero.
 (2) If your prior probability of an event occurring is one,
then the posterior probability will also be one.
 At the two extreme situation, whatever new
information you receive, no matter how reliable it is,
you will not change your prior probability to posterior
probability.
 Question: Does the Bayesian analysis work at the
two extreme situation?
 7.5 Assessing the Value of
New Information
 New information can remove or reduce the
uncertainty involved in a decision and thereby
increase the expected payoff/utility.
 However, in many circumstances it may be expensive
to obtain information since it might involve, for
example, the use of scientific tests, the engagement
of the services of a consultant or the need to carry
out a market research survey.
 If this is the case, then the question arises as to
whether or not it is worth obtaining the new
information. That is we should assess the value of
new information.
 The expected value of
perfect information (EVPI)
 In many decision situations it is not possible to
obtain perfectly reliable information, but nevertheless
the concept of the expected value of perfect
information (EVPI) can still be useful.
 We will use the following problem to show how the
value of perfect information can be measured.
 For simplicity, we will assume that the decision
maker is neutral to risk so that the expected
monetary value criterion can be applied.
 Example 6:
 The Hewlett-Packard (HP) company specializes in
assembling and selling PC systems for use by family
doctor practices throughout the U.S.A. The company is
developing a new PC-based system. At present the
company is trying to decide on the manufacturing and
assembly process to be used. One aspect of this relates
to the keyboard that will be used in the system, which
will have specially labeled function keys.

 The company has decided that it faces three

alternatives:
 It can manufacture/assemble the keyboard itself.
 It can buy the keyboards from a domestic manufacturer.
 It can buy the keyboards from a manufacturer in the Far East
 Example 6:
 Uncertainty in the decision problem
 There is uncertainty as to which decision to take
because there is uncertainty over future sale.
 To help simplify the situation, the company is
planning for one three possible sales levels in the
future:
 Low
 Medium
 high
 Decision Tree
Low 0.2
-15
M Medium 0.5
10
High 0.3
55
Low 0.2 10
BD
HP Medium 0.5
30
High 0.3
25
Low 0.2
5
BA Medium 0.5
20
High 0.3
40
Fig.2-3 Completed decision tree (pay-off and
probability)
 Decision Matrix (Pay-off
Table)
Table 2-3 Pay-off table: probability and pay-off ($000s)
Decision

Sales Probability Manufacture Buy domestic Buy abroad

level (M) (BD) (BA)
Low 0.2 -15 10 5
Mediu 0.5 10 30 20
m
High 0.3 55 25 40
How to get the prior probability
 Prior probability: 先验概率
 History data of sale level for similar product i
n HP
θ High Medium Low Sum
(H) (M) (L)
times 20 50 30 100

According to relative frequency approach, we have prior

probabilities as following:
p (H)=20/100=0.2
p (M)=50/100=0.5
p (L)=30/100=0.3
 To get new information
 To get new information:
 HP plans to ask a professional market research company (AB
C) to forecast the sale level of the new product for it. And A
BC is famous for its forecast accuracy.

 Posterior probability: 后验概率

 How to adjust HP’s prior probabilities based on the accuracy
forecase of ABC.
 p (H|forecase result)=?
 p (M|forecase result)=?
 p (L |forecase result)=?

 Question: Does HP know the forecast result of sales l

evel before to pay ABC company?
 History data of forecast accuracy
for ABC
Table 1: History data of forecast result for ABC
forecas
t
h m l Sum
Real sale lev
el

H 18 1 1 20
M 5 40 5 50
L 3 3 24 30
Sum 26 44 30 100
 Question ： How to figure out conditional probabilities a
ccording to Table 1?
 Conditional Probabilities for ABC
 (1) Conditional Probabilities if real sales level is high:
 p (forecast h|H)=18/20=0.9
 p (forecast m|H)=1/20=0.05
 p (forecast l |H)=1/20=0.05

 (2) Conditional Probabilities if real sales level is Medium:

 p (forecast h|M)=5/50=0.1
 p (forecast m|M)=40/50=0.8
 p (forecast l |M)=5/50=0.1

 (3) Conditional Probabilities if real sales level is Low:

 p (forecast h|L)=3/30=0.1
 p (forecast m|L)=3/30=0.1
 p (forecast l |L)=24/30=0.8
（ 1 ） How to adjust prior pro
babilities if forecast is h
 We have following conditional probabilities:
 p (forecast h|H)=18/20=0.9
 p (forecast h|M)=5/50=0.1
 p (forecast h|L)=3/30=0.1

 How to get following post probabilities ：

 p (H| forecast h)=?
 p (M| forecast h)=?
 p (L | forecast h)=?
 （ 1 ） How to adjust prior pro
babilities if forecast is h
Prior Prob. Con. Prob. Joint Prob. Post Prob.

p(fore hH)=0.90.2 p(H|fore h)=0.18/0.26

p(fore h|H)=0.9 =0.18 =0.692
p(H)=0.2
p(fore hM)=0.10.5 p(M|fore h)=0.05/0.26
p(M)=0.5 p(fore h|M)=0.1 =0.05 =0.192

p(fore hL)=0.10.3 p(L|fore h)=0.03/0.26

p(fore h|L)=0.1 =0.03 =0.115
p(L)=0.3

p(fore h)=0.18+0.05+0.03=0.26
（ 2 ） How to adjust prior proba
bilities if forecast is m
 We have following conditional probabilities:
 p (forecast m|H)=1/20=0.05
 p (forecast m|M)=40/50=0.8
 p (forecast m|L)=3/30=0.1

 How to get following post probabilities ：

 p (H|forecast m)=?
 p (M| forecast m)=?
 p (L | forecast m)=?
 （ 2 ） How to adjust prior proba
bilities if forecast is m
Prior Prob. Con. Prob. Joint Prob. Post Prob.

p(fore mH)=0.050.2 p(H|fore m)=0.01/0.44

p(fore m|H)=0.05 =0.01 =0.022
p(H)=0.2
p(fore mM)=0.80.5 p(M|fore m)=0.4/0.44
p(M)=0.5 p(fore m|M)=0.8 =0.4 =0.909

p(fore mL)=0.10.3 p(L|fore m)=0.03/0.44

p(fore m|L)=0.1 =0.03 =0.068
p(L)=0.3

p(fore m)=0.01+0.4+0.03=0.44
（ 3 ） How to adjust prior pro
babilities if forecast is l
 We have following conditional probabilities:
 p (forecast l|H)=1/20=0.05
 p (forecast l|M)=5/50=0.1
 p (forecast l|L)=24/30=0.8

 How to get following post probabilities ：

 p (H| forecast l)=?
 p (M| forecast l)=?
 p (L | forecast l)=?
 （ 3 ） How to adjust prior pro
babilities if forecast is l
Prior Prob. Con. Prob Joint Prob. Post Prob.

p(fore lH)=0.050.2 p(H|fore l)=0.01/0.3

p(fore l|H)=0.05 =0.01 =0.033
p(H)=0.2
p(fore lM)=0.10.5 p(M|fore l)=0.05/0.3
p(M)=0.5 p(fore l|M)=0.1 =0.05 =0.167

p(fore lL)=0.80.3 p(L|fore l)=0.24/0.3

p(fore l|L)=0.8 =0.24 =0.80
p(L)=0.3

p(fore l)=0.01+0.05+0.24=0.3
 How HP predict the forecast
result of ABC before paying
forecast
h m l Sum
Real sales leve
l

H 18 1 1 20
M 5 40 5 50
L 3 3 24 30
Sum 26 44 30 100

p(fore h) p(fore m) p(fore l)

Marginal
=26/100 =44/100 =30/100
Probability:
=0.26 =0.44 =0.3
 Marg. Prob. Prior Prob. Cond. Prob. Joint Prob. Post. Prob.

p(H)=0.2 p(fore h|H)=0.9 p(fore hH)=0.18 p(H|fore h)=0.692

p(M)=0.5
p(fore h|M)=0.1 p(fore hM)=0.05 p(M|fore h)=0.192
p(fore h)=0.26
p(L)=0.3
p(fore h|L)=0.1 p(fore hL)=0.03 p(L|fore h)=0.115

p(H)=0.2 p(fore m|H)=0.05 p(fore mH)=0.01 p(H|fore m)=0.022

p(fore m)=0.44 p(M)=0.5

p(fore m|M)=0.8 p(fore mM)=0.4 p(M|fore m)=0.909

p(L)=0.3
p(fore m|L)=0.1 p(fore mL)=0.03 p(L|fore m)=0.068

p(H)=0.2 p(fore l|H)=0.05 p(fore lH)=0.01 p(H|fore l)=0.033

p(fore l)=0.3
p(M)=0.5
p(fore l|M)=0.1 p(fore lM)=0.05 p(M|fore l)=0.167

p(L)=0.3
p(fore l|L)=0.8 p(fore lL)=0.24 p(L|fore l)=0.80
How much should HP pay for ABC
before knowing its forecast result
 The way:
 How different forecast results improve the
EMV of HP.
 HP’s EMV without forecast
High 0.2
55
The EMV for the decision of MU is:

MU Medium 0.5
10 550.2+ 100.5-150.3=11.5
Low 0.3
-15
High 0.2 25
BD The EMV for the decision of BD is:


H Medium 0.5 250.2+ 300.5+ 100.3=23

30
P Low 0.3
10
High 0.2
40
BA The EMV for the decision of BA is:

Medium 0.5
20 400.2+ 200.5+ 50.3=19.5
Low 0.3
5
Fig.2-3 Completed decision tree (pay-
off and probability)
 (1) HP’s EMV if forecast is h
p(H|fore h)=0.692
55
MU p(M|fore h)=0.192 The EMV for the decision of MU is:


10 550.692+ 100.192 -150.115

p(L|fore h)=0.115
-15 =38.255
p(H|fore h)=0.692
25
BD p(M|fore h)=0.192 The EMV for the decision of BD is:

H 30 250.692 + 300.192 +100.115
P p(L|fore h)=0.115
10 =34.56
p(H|fore h)=0.692
40 The EMV for the decision of BA is:

BA p(M|fore h)=0.192 400.692 + 200.192 +50.115+
20 =32.095
p(L|fore h)=0.115
5

To get EMV with posterior

probability
 (2) HP’s EMV if forecast is m
p(H|fore m)=0.022
55
MU p(M|fore m)=0.909 The EMV for the decision of MU is:


10 550.022 +10  0.909-150.068

p(L|fore m)=0.068
-15 =9.28
p(H|fore m)=0.022
25
BD p(M|fore m)=0.909 The EMV for the decision of BD is:

H 30 25 0.022 + 300.909+ 10 0.068
P p(L|fore m)=0.068
10
=28.5
p(H|fore m)=0.022
40 The EMV for the decision of BA is:


BA p(M|fore m)=0.909 40 0.022 + 200.909+ 5 0.068

20 =19.4
p(L|fore m)=0.068
5
To get EMV with posterior probability
 (2) HP’s EMV if forecast is l
p(H|fore l)=0.033
55
MU p(M|fore l)=0.167 The EMV for the decision of MU is:


10 550.033+100.167-150.80
p(L|fore l)=0.80
-15 =-8.515
p(H|fore l)=0.033 25
BD p(M|fore l)=0.167 The EMV for the decision of BD is:

H 30 250.033+ 300.167+ 100.80
P p(L|fore l)=0.80
10
=13.835

p(H|fore l)=0.033
40 The EMV for the decision of BA is:


BA p(M|fore l)=0.167 400.033+ 200.167+ 50.80

20 =8.66
p(L|fore l)=0.80
5
To get EMV with posterior
probability
 HP’s EMV with forecast before pa
ying
Marg. Prob. Action EMV with Post. Prob
38.255
MU
38.255 34.56
BD
p(fore h)=0.26
BA 32.095

9.28
MU
p(fore m)=0.44 28.5
BD 28.5
HP
BA 19.4

-8.515
p(fore l)=0.3
13.835 MU
BD 13.835

BA
8.66

The EMV with forecast: 38.2550.26+28.50.44+13.835 0.3=26.6368

 The expected value of
imperfect information (EVII)
HP’s highest EMV without forecast ：
　 The EMV for the decision of BD is:
250.2+ 300.5+ 100.3=23

HP’s EMV with forecast before paying

　 The EMV with forecast: 38.2550.26+28.50.44
+13.835 0.3=26.6368

Improvement of EMV ：
　　 26.6368 － 23 ＝ 3.6368 (EVII)
 The expected value of
perfect information (EVPI)
High 0.2
55
MU Medium 0.5
10  The EMV with perfect informatio
Low 0.3
-15 n is :
High 0.2 25 550.2+ 300.5+ 100.3=29
BD Medium 0.5
H 30
P Low 0.3  Improvement of EMV with perfec
10 t information:
High 0.2 29-23=6 (EVPI)
40
BA Medium 0.5
20
Low 0.3
5
Fig.2-3 Completed decision tree
(pay-off and probability)
 Accuracy of New Information
 Suppose: the forecast accuracy of ABC improves as follows.

forecast

Real sales leve h m l Sum

l

H 19 1 0 20
M 2 46 2 50
L 2 1 27 30
Sum 23 48 29 100
 7.6 Examples for Bayesian
analysis
 Example 7: North Holt Farm
 Example 8:
 Example 7: North Holt Farm
(1/13)
 A year ago a major potato producer suffered
serious losses when a virus affected the crop at
the company’s North Holt farm.
 Since then, steps have been taken to eradicate
the virus from the soil and the specialist who
directed these operations estimates, on the basis
of preliminary evidence, that there is a 70%
chance that the eradication program has been
successful.
 Example 7: North Holt Farm
(2/13)
 The manager of the farm now has to decide
on his policy for the coming season and he
has identified two options:
 (1) He could go ahead and plant a full crop of
potatoes. If the virus is still present an estimated
net loss of $20 000 will be incurred. However, if
the virus is absent, an estimated net return of $90
000 will be earned.
 (2) He could avoid planting potatoes at all and
turn the entire acreage over to the alternative
crop. This would almost certainly lead to net
returns of $30 000.
 Example 7: Decision matrix
without perfect information (3/13)
Decision
Situation of Virus Probability Plant potatoes Plant alternative

Virus is absent 0.7 $90 000 $30 000

Virus is present 0.3 -$20 000 $30 000
EMV $57 000 $30 000
For simplicity, we will assume that the decision maker is neutral to
risk so that the expected monetary value criterion can be applied.
 Example 7: North Holt Farm
New Information (1/13)
 The manager is now informed that Ceres Laboratories
could carry out a test on the farm which will indicate
whether or not the virus is still present in the soil.
 The manager has no idea as to how accurate the
indication will be or the fee which Ceres will charge.
 However, he decides initially to work on the
assumption that the test is perfectly accurate. If this is
the case, what is the maximum amount that it would
be worth paying Ceres to carry out the test?
 Example 7: Decision tree
without /with perfect information
(5/13)
Example 7: Assessing the value
of perfect information (6/13)
 Example 7: The expected value of
imperfect information (EVII) (7/13)
 Suppose that, after making further enquiries, the
farm manager discovers that the Ceres test is not
perfectly reliable.
 If the virus is still present in the soil the test has only
a 90% chance of detecting it, while if the virus has
been eliminated there is a 20% chance that the test
will incorrectly indicate its presence.
 How much would it now be worth paying for the
test? To answer this question it is necessary to
determine the expected value of imperfect
information (EVII).
 Example 7: To get the conditional
probability (8/13)
 If the virus is still present in the soil the test has
only a 90% chance of detecting it.
 P (Test indicates present | Virus is present) =0.9
 P (Test indicates absent | Virus is present) =0.1

 If the virus has been eliminated there is a 20%

chance that the test will incorrectly indicate its
presence.
 P (Test indicates present | Virus is absent) =0.2
 P (Test indicates absent | Virus is absent) =0.8
Example 7: Decision tree without/with
imperfect information (9/13)
Example 7: If test indicates virus
present (10/13)
Example 7: If test indicates virus
absent (11/13)
Example 7: Decision tree without/with
imperfect information (12/13)
Example 7: EVII (13/13)
 Example 8: Thompson Lumber
(1/10)

Table 1: Payoff Table for Thompson Lumber

STATE OF NATURE
FAVORABLE UNFAVORABLE
ALTERNATIVE MARKET ($) MARKET ($) EMV ($)
Construct a large
200,000 –180,000 10,000
plant
Construct a small
100,000 –20,000 40,000
plant
Do nothing 0 0 0
Probabilities 0.50 0.50
Largest EMV
 (2/10)
State of Nature
Alternative Favorable Unfavorable EMV
Market ($) Market ($)
Construct a large plant 200,000 -180,000 10,000
Construct a small plant 100,000 -20,000 40,000
Do nothing 0 0 0
EVwPI =
Perfect Information 200,000 0
100,000

Compute EVwPI
The best alternative with a favorable market is to build a large
plant with a payoff of $200,000. In an unfavorable market the
choice is to do nothing with a payoff of $0
EVwPI = ($200,000)*.5 + ($0)(.5) = $100,000
Compute EVPI = EVwPI – max EMV = $100,000 - $40,000 = $60,000
The most we should pay for any information is $60,000
 Thompson’s Decision Tree (3/10)
EMV for Node = (0.5)($200,000) + (0.5)(–$180,000)
1 = $10,000
Payoffs
Favorable Market (0.5)
$200,000
Alternative with best
EMV is selected 1
Unfavorable Market (0.5)
ct nt –$180,000
r u
n st Pla
e
Co a r g
L Favorable Market (0.5)
$100,000
Construct
2
Small Plant Unfavorable Market (0.5)
–$20,000
Do
No
th EMV for Node = (0.5)($100,000)
in
g 2 = $40,000 + (0.5)(–$20,000)

$0
Figure 3.3
 Thompson’s Complex Decision Tree:
Using Sample Information (4/10)
 Thompson Lumber has two decisions two make,
with the second decision dependent upon the
outcome of the first:
 First, whether or not to conduct their own marketing
survey, at a cost of $10,000, to help them decide
which alternative to pursue (large, small or no plant)
 The survey does not provide perfect information
 Then, to decide which type of plant to build ：
 Note that the $10,000 cost was subtracted from each
of the first 10 branches. The, $190,000 payoff was
originally $200,000 and the $-10,000 was originally $0.
 (5/10)

Thompson’s Complex Decision Tree

First Decision Second Decision Payoffs
Point Point

Favorable Market (0.78)

$190,000
nt 2 Unfavorable Market (0.22)
e P la –$190,000
g
Lar Small
Favorable Market (0.78)
$90,000
5) Plant 3 Unfavorable Market (0.22)
–$30,000
0.4
(
e y s e No Plant
u rv sult abl –$10,000
S r
Su Re avo
1 rv F Favorable Market (0.27)
e $190,000
Re y (0
.5 4 Unfavorable Market (0.73)
y

Ne su t
la n
ve

ga lts 5) P –$190,000
ge
ur

tiv Lar Favorable Market (0.27)

tS

e Small $90,000
ke

Plant 5 Unfavorable Market (0.73)

–$30,000
ar
M
ct

No Plant
du

–$10,000
n
Co

Do Favorable Market (0.50)

Not $200,000
Con
duc nt 6 Unfavorable Market (0.50)
tS urv e P la –$180,000
g
e y Lar Small
Favorable Market (0.50)
$100,000
Plant 7 Unfavorable Market (0.50)
–$20,000
No Plant
$0
 (6/10)

Thompson’s Complex Decision Tree

1. Given favorable survey results
(market favorable for sheds),
EMV(node 2) = EMV(large plant | positive survey)
= (0.78)($190,000) + (0.22)(–$190,000) =
$106,400
EMV(node 3) = EMV(small plant | positive
survey)
= (0.78)($90,000) + (0.22)(–$30,000) =
$63,600
EMV for no plant = –$10,000

2. Given negative survey results,

EMV(node 4) = EMV(large plant | negative
survey)
= (0.27)($190,000) + (0.73)(–$190,000) = –
$87,400
 Thompson’s Complex Decision Tree
(7/10)

Compute the expected value of the market survey,

EMV(node 1) = EMV(conduct survey)
= (0.45)($106,400) + (0.55)($2,400)
= $47,880 + $1,320 = $49,200
f the market survey is not conducted,
EMV(node 6) = EMV(large plant)
= (0.50)($200,000) + (0.50)(–$180,000) = $10,000
EMV(node 7) = EMV(small plant)
= (0.50)($100,000) + (0.50)(–$20,000) = $40,000
EMV for no plant = $0
Best choice is to seek marketing information
 (8/10)

Thompson’s Complex Decision Tree

First Decision Second Decision Payoffs
Point Point
$106,400 Favorable Market (0.78)
$190,000
nt Unfavorable Market (0.22)
Pla –$190,000

$106,400
ge $63,600
Lar Small
Favorable Market (0.78)
$90,000
) Plant Unfavorable Market (0.22)
. 45 –$30,000
(0
e y s e No Plant
u rv sult abl –$10,000
S e r
$49,200Su R avo
rv F –$87,400 Favorable Market (0.27)
e $190,000
Re y (0
.5 Unfavorable Market (0.73)
y

Ne su nt
ve

ga lts 5) Pla –$190,000

e
ur

g $2,400
tiv Lar Favorable Market (0.27)
$2,400
tS

e Small $90,000
ke

Plant Unfavorable Market (0.73)

–$30,000
ar
M
ct

No Plant
du

–$10,000
n
$49,200
Co

Do $10,000 Favorable Market (0.50)

Not $200,000
Con
duc nt Unfavorable Market (0.50)
tS Pla –$180,000
urv ge
$40,000

$40,000
e y Lar Small
Favorable Market (0.50)
$100,000
Plant Unfavorable Market (0.50)
–$20,000
No Plant
$0
 (9/10)
Complex Decision Tree

(1/10)
 Expected Value of Sample Information
(10/10)
 Thompson wants to know the actual value of
doing the survey
Expected value Expected value
with sample of best decision
EVSI = information, assuming – without sample
no cost to gather it information

= (EV with sample information + cost)

– (EV without sample information)

EVSI = ($49,200 + $10,000) – $40,000 = $19,200

Thompson could have paid up to $19,200 for a market
study and still come out ahead since the survey actually
costs $10,000
 Homework 6-1:
 (2) In January a sales manager estimates that there
is only a ‘30% chance’ that the sales of a new
product will exceed one million units in the coming
year. However, he is then handed the results of a
sales forecast. This suggests that the sales will
exceed one million units. The probability that this
indication will be given when sales will exceed a
million units is 0.8. However, the probability that the
forecast will give this indication when sales will not
exceed a million units is 0.4. Revise the sales
manager’s estimate in the light of the sales forecast.
 Homework 6-2:
 (7) A company has just received some ‘state of the art’ electronic
equipment from an overseas supplier. The packaging has been
damaged during delivery and the company must decide whether to
accept the equipment. If the equipment itself has not been damaged,
it could be sold for a profit of $10 000. However, if the batch is
accepted and it turns out to be damaged, a loss of −$5000 will be
made. Rejection of the equipment will lead to no change in the
company’s profit. After a cursory inspection, the company’s engineer
estimates that there is a 60% chance that the equipment has not been
damaged. The company has another option. The equipment could be
tested by a local specialist company.
 Q1: If their test is prefect reliable, how much would it be worth paying
for the perfect information?
 Q2: If their test is not perfectly reliable and has only an 80% chance
of giving a correct indication, how much would it be worth paying for
the imperfect information?
 Assume that the company’s objective is to maximize expected profit
and risk neutral.
The end