System 1 and 2 Decision-Making

Under Uncertainty Part I:

Basic Probability Concepts
Suri Gurumurthi, Ph.D.
October 2020
Basic Probability Concepts
• An experiment is a process that results in some
outcome.
• The outcome of an experiment is a result that we
observe
• E.g.: the number of defective parts in the sample or the
length of time until the bulb fails.
• The collection of all possible outcomes of an
experiment is called the sample space.
• Probability is the likelihood that an outcome occurs.
2
Events B
A
A_c
• An event is a collection of one or more outcomes from a
sample space
• E.g.: finding 2 or fewer errors in the sample of 10 pages
edited, or having a bulb burn for more than 1000 hours.
• If A is any event, the complement of A, denoted as Ac,
consists of all outcomes in the sample space not in A.
• Two events are mutually exclusive if they have no
outcomes in common.

3
Probability Properties
• Label the n outcomes in a sample space as O1, O2, … On,
where Oi represents the ith outcome in the sample
space.
• The probability associated with any outcome must be
between 0 and 1
• 0 ≤ P(Oi) ≤ 1 for each outcome Oi
• The sum of the probabilities over all possible outcomes
must be 1.0
• P(O1) + P(O2) + … + P(On) = 1

4
Calculating Probabilities
A B

• Rule 1: The probability of any event is the sum of the

probabilities of the outcomes that compose that event.
• Rule 2: The probability of the complement of any event
A is P(Ac) = 1 – P(A).
• Rule 3: If events A and B are mutually exclusive, then
P(A or B) = P(A) + P(B)
• Rule 4: If two events A and B are not mutually exclusive,
then P(A or B) = P(A) + P(B) – P(A and B)

5
6
 How many ways are there to get a sequence of 4 cards from a deck ?
Basic counting rules = 52!/48! = 49 x 50 x 51 x 52.
How many ways are there to get a set of 4 cards from a deck ?
= 52!/(48! X 4!) = (49 x 50 x 51 x 52)/(4 x 3 x 2 x 1)
• Permutation:
• The number of ways (with order
specified) of r items selected from n!
n available items is: nPr =
• Combination: HHTT (n – r)!
HTHT
• The number of ways (without order
specified) of r items selected from HTTH
n available items: TTHH
• Example: Number of ways of getting THTH n!
exactly 2 heads in four coin tosses = THHT nCr =
4 (n – r)! r!
4C 2 = 2
= 4! / (4-2)! 2!
= 24 / 4
=6
• Use these permutations and combinations to determine probability of any one
permutation or combination:
• What is the probability of getting 2 heads in 4 coin tosses= 6/16
• What is the probability that we will draw four aces in a particular sequence from a deck of
cards: 7
• =1/6497400 = 1.5 in 10 million chance
Conditional Probability

• Conditional probability is the probability of occurrence of one event

A, given that another event B is known to be true or have already
occurred.

P(B|A)= P(A and B)/P(A)

• Multiplication rule of probability:

8
Independent Events

• Two events A and B are independent if P(A | B) = P(A).

• Example: Laptop fails independently of cell phone P(A and B)/= P(B) x P(A)

Example: A student brings a laptop to class as well as a smartphone

phone in case the laptop is not operational. The probability the laptop
stops working is 10% (battery issues), and independently the
probability that the smartphone does not work is 15% (again battery).
What is the probability that the student will have some electronic aid?
• Probability(No electronic aid) = 0.1 x 0.15 = 0.015 or 1.5%.
• Probability (Some electronic aid)= 1-0.015 = 0.985 or 98.5%

9
Conditional Probability Is Extremely Useful!
• A diagnostic test for a disease is 95% accurate in detecting a disease
in an infected person.
• The probability of infection in the general population is 1%.
• The probability of the test falsely stating infection in a healthy person
is 4%; this is the probability of a false positive

Question: What is probability that a person being tested, and for whom
the test says ”infected”, is actually infected with the disease?

10
 Conditional Probability Is Extremely Useful!
P(Test says infected and
P(Test says Infected)=0.95 x 0.01=
Infected|Infected)=0.95 0.0095
P(A|B) x P(B)
P(Infected)=0.01

P(Test says Not P(Test says Not Infected

Infected|Infected)=0.05 and Infected)=0.0005

Test administered
P(Test Says P(Test says Infected and
Infected|Not Not Infected)=0.0396
Infected)=0.04 P(A and B_c)
P(Not Infected)=0.99
P(Test Says Not P(Test says Not Infected
Infected|Not and Not
A: Event that the test is positive Infected)=0.96 Infected)=0.9504
B: Individual is infected
P(A|B)= 95%
P(B|A) = P(A and B)/(P(A)= P(A|B) x P(B)/P(A)
P(A)= P(A and B) + P(A and B_c)= 0.0491 11
P(A|B) x P(B)= 0.0095
Conditional Probability Is Extremely Useful!
• P(Test says Infected) = 0.0095+0.0396 = 0.0491
• P(Person is Infected|Test says Infected) = P(Person is Infected
and Test Says Infected)/P(Test Says Infected)
• = 0.0095/0.0491= 0.1934= 19.34%!
• In other words, there is only less than 20% chance that a person
who tests positive for the infection, is actually infected!
• Because of high false positive rate
• Because of lower “sensitivity”
• Because of low incidence of disease
• Beware of initial test results…
• Confirm your test results!

12
Conditional Probability Is Extremely Useful!

Infected
Test Positive= 4.9%

Test Positive and Infected=0.95%

Population

13
Confusion Matrix

CONDITION POSITIVE CONDITION NEGATIVE

True Positive False Positive (Type I Error)

∑ 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 ∑ 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃
Sensitivity= ∑ FPR= ∑
𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁

Test

False Negative (Type II Error) True Negative

∑ 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹 𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁 ∑ 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁
Miss Rate= ∑ Specificity = ∑
𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁

14
Fair Bet?
• What is the house advantage?
• All raffles, lotteries and casino games are designed to make a
profit
• House advantage = Amount bet – expected amount returned
• Example: ~35% for lotteries and ~10% for slot machines

• Payoff for $1 bet:

• $36 for “straight-up”
• $2 for red (or black)
• $2 for odd (or even)
• $3 for a dozen, etc.

15
Is this a fair payoff?
• This is how most people think when they bet $1 straight up on any
specific number:
• Expected payoff = 35/36 (-$1) + 1/36($35) = $0
• Even so, what happens when you play for a long time?

• A fair game?
• The roulette wheels actually has two
extra wedges, 0 and 00 that change
the probability
• Correct version of expected payoff:
37/38 (-$1) + 1/38 ($35) = -$0.0526
for each dollar you bet
• The house advantage is about 5.26%

16
Roulette: Is this a fair payoff?

17
 Are coincidences improbable?
• Coincidence
• A coincidence is a surprising concurrence of events, perceived as meaningfully
related, with no apparent causal connection (Diaconis and Mosteller, 1989)
• Winning the lottery twice
• A New York Times story of February 14, 1986, about Evelyn Marie Adams, who
won the NJ lottery twice in short time period
• The article claimed that the odds of a person winning the top prize twice were
about 1 in 17 trillion
• What do you think?

18
Someone, somewhere, someday…
• What is not improbable is that someone,
somewhere, someday will experience those
events or something similar
• Why?
• The 1 in 17 trillion is the probability that a specific
individual who plays the NJ state lottery exactly
twice will win both times
• But don’t forget…
• Millions of people play lottery every day, so not
surprising that someone, somewhere, someday
would win twice
• Stephen Samuels and George McCabe calculated at
least a 1 in 30 chance of a double winner in a 4-
month period and better than 50% probability that
there would be a double winner in a 7-year period
somewhere in the U.S.
19
 Lessons learned
• Probability of a specific event vs. Probability of the
general class of events
• Follow up questions
• What is the probability of someone winning the lottery (as
in previous example) twice in ten years given the following
additional assumptions?
- About 2 million lottery players
- They play twice weekly for 10 years (about 1,000 games)
• Answer:
1/17 trillion * 2 million * 1,000 = 1/8,500 = .000117
20
Basic Probability Distributions
Suri Gurumurthi, Ph.D.
Probability Distributions

• A random variable, X, is a numerical description of the

outcome of an experiment. Formally, a random variable is a
function that assigns a numerical value to every possible
outcome in a sample space.
• A probability distribution, f(x), is a characterization of the
possible values that a random variable may assume along
with the probability of assuming these values.
• The cumulative distribution function, F(x), specifies the
probability that the random variable X will assume a value
less than or equal to a specified value, x, denoted as P(X ≤
x).

2
Important Probability Distributions

• Discrete
• Binomial
• Poisson
• Uniform
• Continuous
• Normal
• Exponential
• Uniform

3
Binomial Distribution

• The binomial distribution describes the probability of obtaining

exactly x “successes” in a sequence of n identical experiments,
called trials.

4
Poisson Distribution

l = expected value or average number of occurrences,

x = 0, 1, 2, 3, … , and e = 2.71828…

5
 Continuous Probability Distributions

• A curve that characterizes outcomes of a continuous random variable

is called a probability density function, and is described by a
mathematical function f(x).
• Probabilities are only defined over intervals.
• The cumulative distribution function, F(x), represents the probability
P(X ≤ x).

6
 • Familiar bell-shaped curve.
• If a normal random variable
has a mean μ = 0 and a
standard deviation σ = 1, it is
called a standard normal
distribution, represented by
z.

Normal Distribution
7
Normal Distribution (PDF vs CDF)
• Normal (μ, σ2):
#"$ !
1 "
𝑓 𝑥 = 𝑒 !%!
2𝜋𝜎 !

8
Calculating Normal Probabilities

• If x is any value from a normal distribution with mean μ and

standard deviation σ, we may easily convert it to an equivalent
value from a standard normal distribution using:

• Excel function NORM.DIST(x, mean, standard deviation, true)

calculates the cumulative probability F(x) for a specified mean
and standard deviation.
• The Excel function NORM.S.DIST(z) calculates the cumulative
probability for any value of z for the standard normal
distribution.

9
Exponential Distribution

• The exponential distribution models the time between

randomly occurring events, such as the time to or between
failures of mechanical or electrical components.

• The Excel function EXPON.DIST(x, lambda, true) can be used to

compute cumulative exponential probabilities.

10
Uniform Distributions

• Continuous Example: Uniform Distribution

𝑓 𝑥 = 1/(𝑏 − 𝑎); 𝑖𝑓 𝑎 ≤ 𝑥 ≤ 𝑏;
𝑓 𝑥 = 0 otherwise
• Discrete Example: Tossing Dice

ì1 / 6 x = 1,2,3,4,5,6
f X ( x) = í
î 0 o.w.
Moments of a Random Variable
• Expected Values E(X)
• 𝐸 𝑋 = 𝜇 = ∫ 𝑥𝑓 𝑥 𝑑𝑥; if X is continuous
• 𝐸 𝑋 = 𝜇= ∑ 𝑥𝑃𝑟(𝑥) ; if X is discrete

• Var(X) = 𝜎 ! = 𝐸 𝑋 ! − 𝐸 𝑥 !

• Standard Deviation = σ
Moments of a Uniform Distribution

• When the distribution is continuous:

"#$
•𝐸 𝑋 =𝜇= !
%
• Var(X) = 𝜎 ! = %!
𝑏−𝑎 !

• When the distribution is discrete:

"#$
•𝐸 𝑋 =𝜇=
!
& ! '%
• Var(X) = 𝜎 ! =
%!
 Simple Decision Models for
Managing Uncertain Outcomes
Suri Gurumurthi, Ph.D.
November 2020
Model 1: Clinical Testing of a Drug in Phases
o There are 3 phases of clinical trials for a given therapeutic
o Phase 1 success rate is 50%
o Phase 2 success rate is 40%
o Phase 3 success rate is 20%
o What is the probability that the drug wins approval for release?
Phase 1 Phase 2 Phase 3

.5 .4 .2

Approval Probability = P(Phase 1 success AND Phase 2 success AND

Phase 3 success) = .5 X .4 X .2 = .04 or 4%
2
Model 1: Multiple drugs in development
o There are 3 phases of clinical trials
o Phase 1 success rate is 50% Drug 1 .04
o Phase 2 success rate is 40%
o Phase 3 success rate is 20%
o With 2 drugs in development Drug 2 .04
simultaneously,
o What is the probability of a
successful release of a drug? 4% + 4% - (4% x 4%)= 8- 0.16% = 7.84%

Probability any one Drug Fails to Win Approval = 1- 0.04 = 0.96

Probability BOTH Drugs Fail to Win Approval = 0.96 x 0.96 = 0.9216
Probability At least one Drug Wins Approval = 1- 0.9216 = 0.0784 = 7.84%

3
Model 1: Multiple drugs in development
o There are 3 phases of clinical trials
o Phase 1 success rate is 50%
o Phase 2 success rate is 40%
o Phase 3 success rate is 20%
o With 10 drugs in development simultaneously,
o What is the probability of at least one successful launch?
Probability a Drug Fails to Win Approval = 1- 0.04 = 0.96
Probability ALL Drugs Fail to Win Approval = 0.96 ^ 10 = 0.6648
Probability At least one Drug Wins Approval = 1- 0.6648 = 0.3351 = 33.5%
In other words, approximately 1 in 3 chance
1- BINOMDIST(S=0, N=10, Prob of Success=0.04, True)

4
Model 1: Binomial Probabilities of Success
o There are 3 phases of clinical trials
o Phase 1 success rate is 50%
o Phase 2 success rate is 40%
o Phase 3 success rate is 20%
o With 10 drugs in development simultaneously,
o What is the probability of 1 successful launch
From the binomial distribution,
Probability = 10!/((10-1)!1!) x (0.04)^1 x (1-0.04)^(10-1)= 0.277
= BINOMIAL(1,10,0.04,false)

5
 Model 2: Damned if you do, Damned if you don’t!
• Tour operator has to book tickets
for group travel
• Profitable venture
• But uncertain attendance
• Cost if you don’t reserve enough
tickets (lose profit margin and
goodwill)
• Underage Cost: u
• Cost if you reserve too many
tickets (sell at a discount)
• Overage Cost: o

Objective: Service Level = Critical Ratio= Set Cumulative Probability = u/(u+0)

Determine Quantity u/(u+o) [0,1]
If u=o; CR= ½
That Maximizes Expected If u>o; CR>1/2
Profits! If u<o; CR<1/2 6
Model 2: Uniform Distribution?
Uniform Distribution:
Probability and Cumulative Probabilities
1

0.9
Set Cumulative Probability = u/(u+0)
0.8

0.7

0.6

0.5

0.4

0.3

0.2
Corresponding Optimal Reservations

0.1

0
0 10 20 30 40 50 60 70 80 90 100
Probability Cumulative Probability 7
 Model 2: Marginal Analysis to Optimize Decision
Marginal Analysis
• Note that quantity
• Let Pr(D>Q) denote the probability that Pr(D>Q) x u
the attendance is greater than the ordered is non-increasing in Q. Why?
quantity Q.
• The expected marginal profit due to sale
• Note that quantity
of the (Q+1)st ticket is therefore:
Pr(D<Q) x o
Pr(D>Q) x u
is non-decreasing in Q. Why?
• The expected marginal loss from
discounting (Q+1)st ticket is
Pr(D≤Q) x o

8
Model 2: Marginal Analysis to Optimize Decision

Marginal Analysis
Expected marginal value of reserving
the (Q+1)st ticket Determine Q* so that expected
= Pr(D>Q) x u - Pr(D<Q) x o marginal value is zero!
Or, (1 - F(Q)) x u = F(Q) x o
= (1 - F(Q)) x u - F(Q) x o Or,
F(Q) = u/(u+o)
• Note that expected marginal value
is also non-increasing in Q. Why? Q= F^inv(u/(u+o))
• Hence at some level of Q, this expected
marginal value will cross, and then
decrease below 0.

9
 Model 2 Insights:
Service Level and Quantity Decision

Service Level =
Probability of
Service = u/(u+o)
Tail Area=
Probability of No-Service
Expected Reservation Decision
Demand

Quantity Scale

Safety
Stock

0 z z-scale
Hence optimal Quantity = Expected Demand + z x Stdev;
Where z= Normsinv(u/(u+o))
10 The Safety Stock portion = z x Stdev
 Model 3: Asset Portfolio Optimization
• The reason for portfolio • Harry Markowitz’s efficient
theory: portfolios:
• To show why diversification is • Those portfolios providing the
a good idea maximum return for their level of
risk
• To show why diversification
makes sense logically • Those portfolios providing the
minimum risk for a certain level of
return

11
 Model 3: Performance is Expected Return
• A portfolio’s performance is • The expected return of a
the result of the performance portfolio is a weighted
of its components average of the expected
• The return realized on a returns of the components:
portfolio is a linear combination n
of the returns on the individual E ( R p ) = ∑  xi E ( Ri ) 
investments i =1

where xi = proportion of portfolio

• The variance of the portfolio is invested in security i and
not a linear combination of n
component variances ∑x
i =1
i =1
12
Model 3 Risk Measure: Variance of Portfolio Return
• Introduction • Portfolio variance is the essence
• Two-security case of the idea of diversification
• The variance of a linear
• Minimum variance portfolio combination of random variables
• Correlation and risk is not a weighted average of the
reduction component variances

13
 Model 3: Computing Portfolio Variance
• For an n-security portfolio, the portfolio variance is:
n n
σ = ∑∑ xi x j ρijσ iσ j
2
p
=i 1 =j 1

where xi = proportion of total investment in Security i

ρij = correlation coefficient between
Security i and Security j

• For a two-security portfolio containing Stock A and Stock B,

the variance is:
σ = x σ + x σ + 2 x A xB ρ ABσ Aσ B
2
p
2
A
2
A
2
B
2
B

14
 Model 3: Two Asset Case Solution: The expected return of this two-
security portfolio is:
n
E ( R p ) = ∑  xi E ( Ri ) 
Assume the following statistics for i =1

Stock A and Stock B: =  x A E ( R A )  +  xB E ( R B ) 

Stock A Stock B = [ 0.4(0.015) ] + [ 0.6(0.020) ]
Expected .015 .020 = 0.018
= 1.80%
return
Variance .050 .060 σ 2p = x A2σ A2 + xB2σ B2 + 2 x A xB ρ ABσ Aσ B
Standard .224 .245 = (.4) 2 (.05) + (.6) 2 (.06) + 2(.4)(.6)(.5)(.224)(.245)
deviation
= .0080 + .0216 + .0132
Weight 40% 60%
= .0428
Correlation .50
coefficient
15
 Model 3: Minimum Variance Portfolio

• The minimum variance • For a two-security minimum

portfolio is the particular variance portfolio, the proportions
combination of securities that invested in stocks A and B are:
will result in the least
possible variance σ B2 − σ Aσ B ρ AB
xA = 2
σ A + σ B − 2σ Aσ B ρ AB
2

• Solving for the minimum

variance portfolio requires
basic calculus xB = 1 − x A

16
 Model 3: Minimum Variance Portfolio
1.2

0.8

Weight A
0.6
Solution: The weights of the minimum
variance portfolios in the previous 0.4
portfolio problem are:
0.2

0
0 0.01 0.02 0.03 0.04 0.05 0.06

σ B2 − σ Aσ B ρ AB .06 − (.224)(.245)(.5) Portfolio Variance

xA = = 59.07%
σ A + σ B − 2σ Aσ B ρ AB .05 + .06 − 2(.224)(.245)(.5)
2 2

1 x A =−
xB =− 1 .5907 =40.93% 17
Model 3:
Minimum
Variance Portfolio
• The efficient frontier can also be
simulated using Monte Carlo
Simulation methods

Source: Mathworks.com 18
 Hypothesis Testing and
System 2 Decision Making
Suri Gurumurthi, Ph.D.
Hypothesis Testing
} Construct two contrasting propositions (competing hypotheses)
} relating to the value of one or more population parameters.
} H0 Null hypothesis: describes an existing theory
} Usually what we are trying to reject!
} H1 Alternative hypothesis: the complement of H0
} Usually what we are trying to prove!
} Using sample data, we either:
- reject H0 and conclude the sample data provides
sufficient evidence to support H1, or
- fail to reject H0 and conclude the sample data
does not support H1.

2
One Sample Hypothesis Tests
Three forms:
1. H0: parameter = constant
H1: parameter ≠ constant
2. H0: parameter ≤ constant
H1: parameter > constant
3. H0: parameter ≥ constant
H1: parameter < constant
The equality part of the hypotheses sign is always in the Null hypothesis.

3
One Sample: Type I vs Type II Error
We risk drawing an incorrect conclusion. Four outcomes are possible:
1. H0 is true and the test correctly fails to reject H0
2. H0 is false and the test correctly rejects H0
3. H0 is true and the test incorrectly rejects H0
4. H0 is false and the test incorrectly fails to reject H0
Outcome 3 is called a Type I error.
Outcome 4 is called a Type II error.

4
 Confusion Matrix

CONDITION POSITIVE CONDITION NEGATIVE

True Positive False Positive (Type I Error)

∑ "#$% &'()*)+% ∑ /01(% &'()*)+%
Sensitivity= ∑ FPR= ∑
,'-.)*)'- &'()*)+% producer's risk
,'-.)*)'- 2%30*)+%
H0 is that person is negative, and
the test incorrectly rejects H0
H0 is that person is negative, and
Test
is false here, but the test fails to
reject H0
False Negative (Type II Error) True Negative
∑ /01(% 2%30*)+% ∑ "#$% 2%30*)+%
Miss Rate= ∑ Specificity = ∑
,'-.)*)'- &'()*)+% ,'-.)*)'- 2%30*)+%

consumer's risk
5
One Sample: Inference Risks
Ø The probability of making a Type I error = α
Ø The probability of making a Type II error = β
Ø Type I Error: α = P( rejecting H0 | H0 is true)
Ø Type II Error: β = P(not rejecting H0 | H0 is false)
Ø The level of significance (α ) defines the risk that we are willing to take in
making the incorrect conclusion that the alternative hypothesis is true when
in fact the null hypothesis is true.
Ø The value of α can be controlled.
Ø Commonly used levels for α are 0.10, 0.05, and 0.01
Ø α is typically set to 0.01, 0.05, or 0.10.
Ø The value of β cannot be specified in advance and depends on the value of
the (unknown) population parameter.

6
One-Sample Hypothesis Tests

Computing the test statistic

One-sample test on a mean, s unknown

<= Standard error

7-7
One-Sample Hypothesis Tests

Drawing Conclusions in a Two-tailed Test

H0: parameter = constant

H1: parameter ≠ constant

7-8
One-Sample Hypothesis Tests

Drawing Conclusions in One-tailed tests

H0: parameter ≥ constant H0: parameter ≤ constant
H1: parameter < constant H1: parameter > constant

7-9
Two Tailed Test Example Number Mean mg/dl Std. Deviation
20 311.15 64.3929

Ø Doctors may recommend having a total cholesterol level below 200 mg/dl.
Ø Given the sample of data above, determine whether the sample gathered is statistically
different, on average, from this recommended level.
Ø Null hypothesis: H0: μ = 200 vs. H1: μ ≠ 200
Ø Our sample of n=20 has sample average= 311.15 and s = 64.3929.
Ø Degrees of Freedom df = 19, so reject H0 if |t| > 2.093
Ø t=(311.15-200)/(64.3929/SQRT(20)) = 7.72
Ø A bigger t-stat implies a greater confidence in rejecting the Null Hypothsis
Ø Therefore proving our desired statement or hypothesis
Ø |t| > 2.093 so we reject H0; Critical t-value = T.INV.2T(α,df)
Ø df=n - 1
Ø The 95% confidence limits around the mean are
Ø 311.15 ± (2.093)(64.3929/√20), OR
Ø 311.15 ± 30.14, OR
Ø (281.01, 341.29)
10
Two-Sample Hypothesis Tests

Excel’s Two-sample Hypothesis Tests on Means

} If the population variances are unknown then, in most situations, use:
t-Test: Two-Sample Assuming Unequal Variances

7-11
One Sample Hypothesis Test Example: Chips-
A-Hoy! Mean = 1238.19, SD = 94.282, SE = 23.5705
Critical value = T.INV(0.05,16-1) = 1.75305, t-stat = 10.1053

ØIn a recent advertising campaign, the Nabisco # Chips

1219
Company announced a “1000 Chips Challenge,” 1214
1087
Øclaiming that every 18-ounce bag of their Chips Ahoy! 1200
1419
cookies contained more than 1000 chocolate chips. 1121
1325
ØThe analysts randomly selected bags of cookies and 1345
counted the chocolate chips. 1244
1258
ØThe data to the right report their counts. 1356
1132

ØIs it safe to say that the Company will keep its 1191
1270
promise of > 1000 chips in a bag of Chips-Ahoy? 1295
1135
H0: x <= 1000 H1: x > 1000, t-stat > C then reject H0 (safe) 16 items
12
Upper Tail Test Example
Step Mean Time Std. Deviation
Answer 0.1023 0.0183
Service 0.5290 0.0902

➢ A bank call center has a mandate to answer telephone customers calls in 0.1
minute or less.
➢ A sample of 25 answer times was drawn to evaluate the performance.
➢ Operators must determine customer needs and respond or refer the customer
within 0.50 minute.
➢ A sample of 25 service times was again taken to evaluate performance.
➢ If these are independent variables, is the average time taken to perform each step
statistically greater than the required standard?

3
Upper Tail Test Example

Specification for answer time for the steps at the call center is :

H0: Mean response time:  1 < 0.10

H1: Mean response time:  1 > 0.10
1 = 0.1023, s1 = 0.0183
x the t-test is:
and

, t24,.05 = 1.711
x − 0.10 0.1023 − 0.10 0.0023
t1 = = = = 0.6284
s/ n 0.0183 / 25 0.0037

4
Analysis

Specification for service time is:

H0: Mean service time:  2 < 0.50

H1: Mean service time:  2 > 0.50
x 2 = 0.5290, s2 = 0.0902
and the t-test is:

x − 0.50 0.529 − 0.50 0.029

t2 = = = = 1.608 , t24,.05 = 1.711
s/ n 0.0902 / 25 0.0180

5
 Conclusion of Upper Tail Test Example
➢ Because t24, .05 = 1.711 for both measures, we cannot reject the null hypothesis for t1 =
0.6284, and we also cannot reject the hypothesis for t2 = 1.608.
➢ Therefore, there is no statistical evidence that the mean response time exceeds 0.10 for
the answer component. The statistical evidence also doesn’t support the fact that the
service component time exceeds the standard of 0.50 minutes.

7-6
One-Sample Hypothesis Tests
p-Values
 One decision rule for any hypothesis test uses the p-value
 The p-value is the observed significance level.
 The p-value decision rule is to:

Reject H0 if the p-value < α

7-7
Two-Sample Hypothesis Tests
Interpreting Excel’s Two-sample Hypothesis Test Output

➢ If the test statistic is negative, Excel’s one-tail p-value is correct for

a lower tail test. For an upper-tail test, subtract the reported p-value
from 1
➢ The reverse is true when the test statistic is non-negative
➢ The critical value reported in Excel should be multiplied by -1 if

7-28
Simulation Modeling
Suri Gurumurthi, Ph.D.
November 2020
Optimization vs. Simulation
Optimization
Determines the best
levels of variables
under our control

Simulation

Helps assess the

impact of specific
choices we make
for variables under
our control

2
 Simulation Application
• Model = Representation of Reality • Manufacturing:
• scheduling, inventory
• Model IS NOT EQUAL Reality!
• Staffing: personal-service
• We use simulation when other methods are operations
too context poor • Banks, fast food, theme parks, Post
Office, ...
• We look at “what-ifs”, or consequences of our
decisions in the model • Distribution and logistics
• Air traffic operations
• Principle: Approximate answer to exact • Warehouse operations
problem is sometimes better than an exact • Ports and Shipping
• Urban Planning
answer to approximate problem
• Health care:
• Staffing, Emergency and hospital
operations
• Telecommunications
• Switching and Call Routing
• Military
• Logistics and war-games
4
Limitations of Simulation Modeling

Don’t get exact

Get random output
answers, only Simpler models, yield Complex models yield
(RIRO) from stochastic
approximations, more robust results less robust results
simulations
estimates

Statistical design,
Also true of many
analysis of Fewer degrees of Higher degrees of
other modern
simulation freedom freedom
methods
experiments

Results are
Higher statistical Lower statistical
conditioned on
confidence in confidence in
statistical
results results
confidence levels

7
 Simulation by Hand:
The Buffon Needle Problem
https://www.youtube.com/watch?v=szUH1rzwbAw Silly Experiment?
• Estimate p (George Louis Leclerc, c. • Experiment to estimate something
1733) hard to compute exactly (in 1733)
• Toss needle of length l onto table with • Randomness, so estimate will not be
stripes d (>l) apart exact; estimate the error in the
estimate
• P (needle crosses a line) = 2l • Replication (the more the better) to
pd
reduce error
• Repeat; tally p! = proportion of times a • Sequential sampling to control error
line is crossed -- keep tossing until probable error
in estimate is “small enough”
• Variance reduction
2l
• Estimate p by
!
pd
9
One Time Quantity Decisions Under Uncertainty
Underage Cost u per unit
Prob(Demand D > Q)

Uncertain Demand D Set F(Q) = u/(u+o)

Prob(Demand<=Q)=u/(u+o)
To maximize Expected Profits
Order Amount Q

Prob(Demand D < Q) Overage Cost o per unit

6
 Normally Distributed Demand
• Uncertain Demand

• Underage (Shortage) Cost: u per unit

• Overage (Excess) Cost: o per unit

Objective:
Determine Quantity
That Maximizes Expected
Profits! 9
Monte Carlo Simulation: Repeated Trials

• Increasing the number of trials increases the accuracy of your

estimate of the expected outcome
• Remember formula for standard deviation of average:

s
sX =
X

n
• Excel and related tools such as Oracle Crystal Ball make it much
easier to conduct and analyze a large number of trials

11
Monte Carlo Simulation with Excel

• Suppose a retailer has a one time • There could be several criteria for
opportunity to sell an apparel product
• Product has expected demand = 4000 units the decision:
• Standard deviation of demand = 800 units
• What is the number the retailer
should produce to maximize
• In order to buy the intellectual property
rights, the retailer has to pay $20,000 to the expected profits?
designer • What is the number the retailer
• The selling price of the product is estimated should produce so that he suffers no
at $18 per unit
more than say 5% chance of a loss?
• The royalty to be paid to the designer is $3
per unit • What is the standard deviation of
• The manufacturing cost of the item is $4.5 profits?
• If the product remains unsold, we have to
destroy the product, and lose the
manufacturing cost

12
Monte Carlo Simulation with Excel No.
Sampled
Demand
1 4666
2 3465
3 4983
• Excel has simple functions that allow us 4 5168
to model this problem 5 3964
6 3013
7 3459
• For e.g., the norminv (probability, 8 3247
Mean, Stdev) function allows us to 9 4554
determine the quantity that can provide 10 3926
11 4094
a target probability of serving demand 12 3283
• Combined with the rand() function, this 13 3626
can allow us to generate a sample of 14 4238
15 4213
demand from an underlying distribution 16 4669
• For e.g. NORMINV(RAND(),4000,800) 17 4231
generates a single sample (random) 18 3334
19 4136
demand point from a normal 20 2296
distribution with mean 4000, and Stdev 21 4415
800 22 4910
23 4191
24 4923 13
25 5254
Monte Carlo Simulation with Excel

• Suppose we wish to simulate the Sampled Overage Net

outcomes from producing 4400 units No. Demand Sales Profits Overage Cost Income
• We can then evaluate the sales, profits, $- $26,200
1 4666 4400 $46,200 0
and the excess (overage) costs for a
sample (trial) 2 3465 3465 $36,384 935 $4,207 $12,177
• By examining relatively large number 3 4983 4400 $46,200 0 $- $26,200
of trials or sampled demand points, 4 5168 4400 $46,200 0 $- $26,200
we can estimate the average profits 5 3964 3964 $41,618 436 $1,964 $19,654
from producing 4400 units
6 3013 3013 $31,637 1387 $6,241 $5,396
• We can then change the decision,
7 3459 3459 $36,314 941 $4,237 $12,078
and re-evaluate the expected profits.
8 3247 3247 $34,093 1153 $5,189 $8,904
• We can similarly capture other $- $26,200
9 4554 4400 $46,200 0
measures such as standard deviation
of profits, or probability of loss 10 3926 3926 $41,224 474 $2,133 $19,091

14
Monte Carlo Simulation with Excel

• We can then change the decision, and Simulation Results

re-evaluate the expected profits. Production Decision 4400
• We can similarly capture other Expected Net Income $18,206
measures such as standard deviation of Stdev of Net Income $8,889
profits, or probability of loss Minimum Profits $-23,293
Maximum Profits $26,200

Simulation Results and Tables

Production
Decision 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Expected Net
Income -$9,500 $974 $10,760 $17,335 $16,434 $13,211 $8,204 $4,367 -$657 -$4,845
Stdev of Net
Income $- $341 $2,930 $7,332 $11,281 $11,581 $12,095 $11,778 $11,775 $12,059
Minimum
Profits -$9,500 -$5,379 -$19,743 -$32,020 -$46,208 -$16,722 -$30,297 -$30,959 -$35,454 -$40,330
Maximum 15
Profits -$9,500 $1,000 $11,500 $22,000 $32,500 $43,000 $47,079 $46,994 $40,591 $30,630
Cash Flow NPV and Portfolio
Simulation Models
Suri Gurumurthi, Ph.D.
Fall 2020
A Portfolio Management System
• Design of a project portfolio  Selection Criteria
system:  Financial models: payback, net present
value (NPV)
• Classification of  Non-financial models: assets/projects of
investments/projects strategic importance to the firm
• Selection criteria depending
upon classification  Multi-Criteria Selection Models
 Use several weighted selection criteria to
• Sources of proposals evaluate project proposals.
• Evaluating proposals Say you have 3 choices for a solar panel installation:
1. Big system (24 panels): 20K to install: $1200 savings/year
• Managing the portfolio of - $20K/$1200 = 16.66 years
- With tax savings $15K investment; hence 15K/1200=12.5 years
assets/projects payback
2. Small (12 panels): 10K to install: $800 savings/year
- $10K/$800= 12.5 years to payback
- With tax savings 7.5K/800 = 9.375 years
3. Small system + battery : 21K to install: $1200 savings/year
- $21K/1500=14 years
2–2 -With tax savings $15K investment; hence 10 years to payback.
NPV Calculation PV=$1/(1+0.03)= 0.97
PV= $1/(1+0.03)^2 𝑛𝑛
𝐹𝐹𝑡𝑡
𝑁𝑁𝑁𝑁𝑁𝑁 = 𝐼𝐼0 + �
 The Net Present Value (NPV) Model 𝑡𝑡=1
(1 + 𝑘𝑘)𝑡𝑡

 Uses management’s minimum desired rate-of-return (discount rate) to

compute the present value of all net cash inflows.
 Positive NPV: project meets minimum desired rate
of return and is eligible for further consideration.
 Negative NPV: project is rejected for initial investment I0; Interest
or discount rate k;
 Net cash inflow Ft for period t. This represents project or
investment performance

2–3
NPV with Returns
 The Net Present Value (NPV) Model
 Negative NPV: project is rejected for initial investment I0; Interest
or discount rate k;
 Net cash inflow Ft for period t. This represents company or
project performance
 For investments, the return on the investment compounds over
time; say the return rate is r each period
𝐼𝐼0 (1 + 𝑟𝑟)𝑛𝑛
𝑁𝑁𝑁𝑁𝑁𝑁 =
(1 + 𝑘𝑘)𝑛𝑛

2–4
NPV Calculation: Potential Issues with the model

 The Net Present Value (NPV) Model

 Cash flows could be subject to uncertainty
 Discount rate is assumed fixed, but even that could change over time
 E.g. in tight cash flow situation, we are more impatient; wealth level influences discount
rate
 When comparing different streams of investment growth, they could be
correlated, hence the returns may not be independent

2–5
 Asset Allocation Decisions
• Portfolio Theory:
• Shows why diversification is a good idea
• Trade-off between higher expected returns and lower levels of
risk
• Say two assets S and B
• Say S is stock; B is a bond
• S has mean return rS; Stdev σS
• B has mean return rB; Stdev σB
• Typically stocks have higher expected returns and higher risk
• Bonds have lower expected returns and lower risk too

6
 Asset Allocation Decisions
• For any investment level
• Assume x is the proportion of investment allocated to Asset S
• Then (1-x) is the proportion of investment allocated to Asset B
• What is the best value of x?
• In other words, what proportion of the investment should be
invested in safe versus risky asset?

7
 Performance: Expected Return
• A portfolio’s performance is the • The expected return of a
result of the performance of its portfolio is a weighted average
components of the expected returns of the
• The return realized on a portfolio is a components
linear combination of the returns on
the individual investments
• Hence Expected return is:
• ER = xrS + (1-x)rB
• The variance of the portfolio is not a
linear combination of component
variances

8
 Performance: Variance of Portfolio Return
• Understanding portfolio variance is the essence of understanding the mathematics
of diversification
• The variance of a linear combination of random variables is not a weighted average of the
component variances
• Further, there are 3 cases: If Independent: Var_P= x^2Var_S + (1-x)^2Var_B
• The assets are positively corrected in performance
• The assets are negatively corrected in performance
• The assets are independent in performance
• Suppose the correlation co-efficient between assets is ⍴
• If Rho= +1, then perfectly positively correlated
• If Rho= -1, then perfectly negatively correlated
• If Rho=0, then independent asset classes
• ⍴ is covariance of the variables divided by the product of their standard deviations.

9
 Problem: Simulate correlated variables?
• Simulating one evolution of one asset is relatively easy:
• Say, NORMINV(rand(),rS, σS)
• Similarly, NORMINV(rand(), rB, σB)
• But how to simulate asset performance B in a way that is
correlated with S (or vice versa)
• We will use approximate simulation method…

10
 Problem: Simulate correlated variables?
• Suppose Z1= NORMINV(rand(),0, 1); Z2 = NORMINV(rand(), 0, 1)
• Notice, that Z1 and Z2 will almost surely sample different values
• Then we will simulate S= rS+ σS x Z1
• Simulate B= rB+ σB x [⍴Z1 + SQRT(1- ⍴2) Z2]
• Doing so will result in CORREL(S,B) approximately equal to ⍴

• Then we can simulate the evolution of S and B and compare the actual or
discounted value of portfolio over a time-span

11
What is a Decision Tree?
• A Visual Representation of Choices, Consequences,
Probabilities, and Opportunities.
• A Way of Breaking Down Complicated Situations Down to
Easier-to-Understand Scenarios.
 •Graduate School Costs
Simple Decision Tree Example • Tuition and Fees
•Rent / Food / etc.
•Opportunity Cost
•Graduate School Benefits
Graduate
Job
School – 2 years •Learning
Decision Graduate •Future Growth in Earnings
School- 2 years
Job – 2 years
Decision in 2
...
years!

Present Value of Income in Year i= Net Income in Year i / (1+ r)^i

Probabilistic Decision Trees

• For sure you can get $1M

• But if you want to gamble: flip a coin
• If it’s heads: you win $2M
• If its tails: you get nothing
• Expected value of the gamble = ½ x $2M + ½ x 0 = $1M
• IN many cases, people would be willing to accept a reduced sure payoff
• They would be willing to accept less than $1M in order to not take
the gamble.
• Risk premium
https://www.youtube.com/watch?v=GfDaplU9zuw

4
Monty Hall Problem:
Pick a door to win a car!

In search of a new car, the player picks a door, say 1. The game
host then opens one of the other doors, say 3, to reveal a goat
and offers to let the player switch from door 1 to door 2.

PLAYER PROBLEM:
SWITCH OR NOT TO SWITCH THEIR ORIGINAL CHOICE!
 SWITCH OUTCOME
Monty Hall Problem 1/6 Host Opens Door 2 Lose

Pick Door 1 1/6 Host Opens Door 3 Lose

1/3 Host Opens Door 3 Win

Pick Door 2
Prize behind 1/3
Door 1 Host Opens Door 2 Win
Pick Door 3
Host Opens Door 3
1/3 Win
1/3 Pick Door 1
Prize behind Host Opens Door 1 Lose
1/6
Door 2
1/3 Pick Door 2
Host Opens Door 3 Lose
1/6
Pick Door 3 Win
1/3 1/3 Host Opens Door 1

Pick Door 1 1/3 Host Opens Door 2 Win

Prize behind
Door 3 1/3 Host Opens Door 1 Win
Pick Door 2
1/6 Host Opens Door 1 Lose
Pick Door 3
Probability[Win Prize after Switch ] = 2/3 1/6 Host Opens Door 2 Lose 6
 Breakeven Analysis
• Profit Margin = Revenue Per Unit – Variable Cost Per Unit
• Breakeven Point = Fixed Costs/(Profit Margin Per Unit)
• Increasing revenue per unit reduces breakeven
• Increasing variable cost per unit increases breakeven
• Increasing fixed costs increase breakeven
• Remember, lower break-evens are better!
• Lower risk from capacity investment
• Quicker return-on-investment

14
Breakeven Analysis Example – Bakery Co.
Tsim Sha Tsui Hang Hau
Lease Cost (per month) $300,000.00 $100,000.00
Employee Salaries (per month) $120,000 $100,000
Maximum Sales Volume (customers/month) 15000 9000
Average Order Revenue (per customer) $75 $55
Average Cost per Order (per customer) $40 $35
Average Profit Margin (per unit) $35 $20
Total Profits (per month) $525,000 $180,000
Net Income (per month)
$105,000.00 ($20,000.00)
Max Sales x Profit Margin – Fixed Cost
Breakeven Point (customers per month) 12000 10000
15
What if: We Hired an Extra Worker?
Tsim Sha Tsui Hang Hau
Lease Cost (per month) $300,000.00 $100,000.00
Employee Salaries (per month) $150,000 $120,000
Maximum Sales Volume (customers/month) 17000 11000
Average Order Revenue (per customer) $75 $55
Average Cost per Order (per customer) $40 $35
Average Profit Margin (per unit) $35 $20
Total Profits (per month) $595,000 $220,000
Net Income (per month) $145,000.00 $0
Breakeven Point (customers per month) 12857 11000

18
Simple Decision Trees for Capacity Investment
Net income 105K Net income 105K

Open TST Revenue

$69/cust

Net income -70K Net income 105K

First
Decision HH Base Cost
Case $21/cust

Net income 0K Net income 55K Net income 110K

Add Capacity Revenue Cost

$20K/month $60/cust $30/cust

19
What if: Demand were Uncertain?

Net Income with

Net Income with
Net Income with Additional Worker
Net Additional Worker
Probability TTT Additional Worker in Low and Medium
Income in Low and Medium
in All Scenarios Scenarios with
Scenarios
Advertising

High 0.2 16K 140K 140K 180K 140K

Medium 0.4 12K 0 40K 40K 40K

Low 0.4 10K -70K -30K -30K 30K

Expected
12K 0 32K 40K 56K
Value
20
Decision Trees for Uncertain Scenarios- Recourse
Net income 140K

High
Sales
(16K)

20% Probability

Net income 0K Net income 40K

Initial 40% Probability Medium Hire Additional

Decision Sales Worker ($30K) Sales (+2K)
Open TST (12K)

40% Probability Net income -70K Net income -30K Net income 30K

Low Hire Additional Advertising

Sales Worker ($30K) Sales (+2K) ($10K) Sales (+2K)
(10K)

21
Financial and
Real Options
and Related
Decision
Trees
SURI GURUMURTHI, PH.D.

N O V E M B E R 2 0 21
What are Options?
Ø An option is the right, but not the obligation, to buy (or sell) an asset for a pre-determined
price within a pre-determined period of time

Ø Options are financial derivatives traded in financial markets

Ø The value of an option depends on:

Ø The time before exercising the option (+)
Ø The value of the underlying asset (+)
Ø The volatility of the asset (+)
Ø The strike price (-)
Ø The interest rate (+)
Call Option
on a Share
of Tesla
Gives the holder the right to
buy the underlying stock at a
predetermined (exercise) price

Source of value in an option:

The asymmetry from having
the right but not the obligation
to exercise the option.
Put Option
on a Tesla
Share
Gives the holder the right to
sell the underlying stock at a
predetermined (exercise) price
Real Options vs Financial Options
The real world is characterized by change, uncertainty and competitive interactions

As new information arrives and uncertainty about market conditions is resolved, there could be
value in the flexibility to alter the initial decision in order to:
◦ capitalize on favorable future opportunities
◦ react to mitigate losses.

This flexibility is similar to the financial options decision, and is known as a Strategic Option, or
Real Option.
 Wait/Learn
• To wait before taking an action until more is known or timing is
expected to be more favorable
• E.g., When to introduce a new product, or replace an existing piece
of equipment
Expand or contract
Common • To increase or decrease the scale of an operation in response to
demand
Types of Real • E.g. adding or subtracting to a service offering, or adding memory to
a computer
Options Switch inputs or outputs
• To alter the mix of inputs or outputs of a production process in
response to market prices
• E.g., The output mix of telephone/internet/mobile services for a
bank
Abandon
• To discontinue an operation and liquidate the assets
• E.g., discontinuation of a research project, or product/service line
 Call Option Decision Tree
Price > Value of Contract
Strike
p= P(Price >Strike Price) Price • If Expiration Price> Strike Price
• = (Expiration Price-Strike Price)*100 – c x 100
Buy Option?
Cost =
• But If Expiration Price< Strike Price
c x 100 = – c x 100

• Expected Value of Contract

1- P(Price >Strike Price)
Price < = p* (Expiration Price-Strike Price)*100 - c x 100
Strike
Price

Contract Expiration
 Put Option Decision Tree
Price >
Strike Value of Contract
p= P(Price >Strike Price) Price • If Expiration Price< Strike Price
= (Strike Price - Expiration Price)*100 – c x 100
Buy Option?
Cost = • But If Expiration Price> Strike Price
c x 100 = – c x 100

1- P(Price >Strike Price)

• Expected Value of Contract
Price <
Strike
= (1-p)* (Strike Price - Expiration Price)*100 - c x 100
Price

Contract Expiration
 Call Option Decision Tree (Reality)
Keep Option?
Price > p* Price >
Gain =
Strike Strike
p= P(Price >Strike Price) Price (𝛿*Price – c) x Price
100
(1-p*)

Buy Option?
Cost =
c x 100
p**

1- P(Price >Strike Price) Keep Option?

Price < Loss = (c- (1-p**) Price <
Strike Strike
Price 𝛿*Price) x Price
100
Contract Review Expiration
Risk Premium
If we sell the option:
◦ Sure Gain = (𝛿*Price – c) x 100

If we keep the option:

◦ Expected Value of Contract
= p* (Expiration Price-Strike Price)*100 - c x 100
Risk Premium =
= p* (Expiration Price-Strike Price)*100 - c x 100 - (𝛿*Price – c) x 100
 Risk Premium (in General)
Best Case G
Probability p Risk Premium
= p*G + (1-p)*L – Certainty
Equivalent

Probability (1-p)
Take Risk ? Worst Case L

Sure Gain or
Certainty Equivalent
Estimating Value of Information:
Example
Management estimates the net profits of choosing from three decision alternatives (A, B, and C)
under the differing probable levels of demand. These net profit estimates, in millions of dollars,
are:

0.1 0.5 0.4

Low Medium High EV
Decisions:
A,B, or C A 10 50 90 62
B -120 25 200 80.5
C 20 40 60 46
EVA = 0.1(10) + 0.5(50) + 0.4(90) = 62
EVB = 0.1(-120) + 0.5(25) + 0.4(200) = 80.5
EVC = 0.1(20) + 0.5(40) + 0.4(60) = 46
With perfect information: EV = 0.1(20)+0.5(50)+0.4(200)=107

14
Expected value of perfect
information
EVPI
◦ How much you are willing to pay for information that
help you make the best decision?
◦ EVPI is the difference between the expected value with
perfect information (crystal ball) and the expected value
without perfect information (without crystal ball)

EVPI = EV with PI – EV without PI

◦ EVPI for previous example =107 – 80.5 = 26.5M !

15
 What would you prefer?
Option 1 VS Option 2
• Win the money in envelope A • One randomly drawn bank note is
removed from envelope B, and then
20 x $1000 notes + 10 x $500 notes placed back into the same envelope
=$25000
• After observing the note from B, then
choose between envelopes A and B
and win the money in the envelope
10 x $1000 notes + 20 x $500 notes
= $20000
• Option 1?
• Option 2?
• Make no difference?

5
 Option 1
What is the expected value for option 1?

Expected value of option 1:

1 1
$25,000 + $20,000 = $22,500
2 2

6
 Option 2
One randomly drawn bank note is removed from
envelope B

– Scenario 1: One $500 bank note is removed from, and put back into
envelope B

– Scenario 2: One $1,000 bank note is removed from, and put back
into envelope B

With this additional information, which envelope would

you choose?
Envelope A? or Envelope B?

7
 Option 2: Scenario 1
Scenario 1: One $500 bank note is removed from
envelope B

Envelope B contains $20,000 10×$1,000

10 B contains
Envelope 𝑛𝑜𝑡𝑒$25,000
20×$500
20 𝑛𝑜𝑡𝑒
Probability that Envelope B contains Probability
(Totally that
there areEnvelope
$20,000)B contains
$20,000 given that $500 is removed $25,000 given that $500 is removed

20 2 10 1
= = = =
10 + 20 3 10 + 20 3

! #
Given $500 is removed, with 𝑃𝑟𝑜𝑏 = Given $500 is removed, with 𝑃𝑟𝑜𝑏 =
" "
Envelope A contains $25,000 Envelope A contains $20,000
Envelope B contains $20,000 Envelope B contains $25,000

8
 Option 2: Scenario 1
Scenario 1: One $500 bank note is removed from
envelope B
Which envelope would you choose?
Envelope A? or EnvelopeProb
B?= 1/3 Prob = 2/3
$ in envelope A after draw $20,000 $25,000
$ in envelope B after draw $25,000 $20,000

Expected value for envelope A:

2 4
3
$20,000 + 3 $25,000 = $23,333.33 Choose Envelope A

Expected value for envelope B:

2 4
$25,000 + $20,000 = $21,666.67
3 3
9
 Option 2: Scenario 2
Scenario 2: One $1,000 bank note is removed from
envelope B
Which envelope would you choose?
Envelope A? or EnvelopeProb
B?= 1/3 Prob = 2/3
$ in envelope A after draw $25,000 $20,000
$ in envelope B after draw $20,000 $25,000

Expected value for envelope A:

2 4
$25,000 + $20,000 = $21,666.67
3 3

Expected value for envelope B:

2 4
3
$20,000 + 3
$25,000 = $23,333.33 Choose Envelope B
10
 Option 2: Summary
Comparing the two scenarios, what is the expected value
for option 2?
$500 bank note is drawn $1,000 bank note is drawn

Expected value
in envelope A $23,333.33 $21,666.67
Expected value
in envelope B $21,666.67 $23,333.33

Expected value of option 2:

1 1
$23,333.33 + $23,333.33 = $23,333.33
2 2
Expected Value of Information =
Expected Value of Option 2 – Expected Value of Option 1= 23.33K-22.5K
= 833 dollars 11
 Option 2
30 1 30 1
= =
60 2 60 2

Scenario 1: Scenario 2:
Remove $500 Remove $1,000

Probability 𝟏𝟎 𝟏 𝟐𝟎 𝟐 Probability 𝟏𝟎 𝟏 𝟐𝟎 𝟐
= = = =
𝟑𝟎 𝟑 𝟑𝟎 𝟑 𝟑𝟎 𝟑 𝟑𝟎 𝟑
Envelope B $25,000 $20,000 Envelope B $20,000 $25,000
Envelope A $20,000 $25,000 Envelope A $25,000 $20,000

12
 Option 1 VS Option 2
• Win the money in envelope A • One randomly drawn bank
note is removed from
envelope B
• Choose between envelopes A
and B and win the money in
the envelope

Expected value of option 1 Expected value of option 2

= 22,500 = 23,333.33

13
Decision-Making With
Big Data and AI
Suri Gurumurthi, Ph.D.
November 2019
Virtual & Augmented Reality
• Virtual Reality
• “The computer-generated simulation of a 3D
image or environment that a person can be
interact with in a seemingly real or physical
way by a person using special electronic
devices…”
• Augmented Reality
• “Superimposition of a computer-generated
image on a person's view of the real world,
providing a composite and altered
perspective...”
 Decision Making in VR and AR
• US Army trains its soldiers on high pressure
combat environments using VR
• De Beers allows potential customers to virtually
try on jewelry
• Entice and attract potential buyers
• Personalize the experience
• IKEA allows customers to virtually place the
catalog items in your home using mobile apps
• Visualize before purchase
• Reduce regret and returns!
 VR is a 3D Simulates space
and higher Simulates time
dimensional Simulates interactions!
simulation Can be used to model multi-agent decision-making

Differences
Between VR
and AR
Augmented Marries personal experiences with related imagery and
simulations
reality is
Adds value to the real world experience
more
Enables superior decision-making with additional
personal superimposed data
and Looks for and provides the missing elements that can
complex accelerate decision-making!
 Big Data are high-volume, high-velocity, and/or high-variety

“Big Data” Defined…

information assets that require new forms of processing to
enable enhanced decision making, insight discovery and
process optimization (Gartner, 2012)

2.5 quintillion bytes (10^18 bytes) of data

10,000
is created every day as of 2018)…

• More than 90% is unstructured data UNSTRUCTURED DATA

• Approx. 500 quadrillion files

• Quantity doubles every 2 years
• Most unstructured data is neither stored nor
analyzed!
GB of Data
(IN BILLIONS)

STRUCTURED DATA

2005 2010 2015

 Opportunities for Decision Makers…

European
As of 2017, administrators Poor data
53% of Forbes 40% projected
could save across Poor data can
surveyed Bad data or growth in
more than businesses cost
companies poor data global data
€100 billion and the businesses
adopted Big quality costs generated per
($149 billion) government 20%–35% of
Data US businesses year vs. 5%
in operational costs the U.S. their
Analytics. $600 billion growth in
efficiency economy $3.1 operating
annually. global IT
Up From 15% improvements trillion dollars revenue.
spending.
in 2015 alone by using a year.
big data
 Unstructured The key need in today’s digital
Data environment is to:
1. Reduce the dimensionality of big data
(aggregate)
Structured
Data 2. Transform and synthesize
3. Extract intelligence (usable
information)
Useful
Information 4. Analyze (understand relationships)
5. Value-added decision-making

Analytics

Decision-
Making
Big Data = Big Nothing, Unless…
Descriptive Analytics: Inference
Identify Draw from
goals sample
Start Population Sample

Draw Collect raw data

conclusions and summarize

Make inferences
about population
Population Sample Statistics
parameters

Source: Bennett, Briggs, and Triola (2014), Statistical Reasoning for Everyday Life.
9
Correlation Coefficients… Correlation = 0

Dependent variable
• Correlation 30
25
• If two events (features or 20
measurement variables) 15
tend to occur together, they 10
5
are said to be correlated 0
• Example: Lighting and 0 2 4 6 8 10 12
thunder, weight and height Independent variable
• Correlation coefficient:
Measuring the strength of • Outliers
linear relationship of • An outlier that is consistent with the trend of
measurement variables the rest of the data can inflate the correlation
• An outlier that is not consistent with the rest
• r è +1: Positive correlation of the data can deflate the correlation
• r è -1: Negative correlation
• Simpson’s paradox
• r è 0: No correlation (?) • Groups combined inappropriately may mask
relationships
• The missing link (effect of a third variable) 11
Example: Death rate vs. speed limit

• Correlation between death rate and speed limit is

r = 0.55 (Rivkin, 1986)
• If Italy is removed, correlation drops to 0.098
• If Britain is also removed, correlation jumps to 0.70
12
 Lurking vs. confounding variables
• Confounding occurs in a study
A
when you are unable to
B Causation
distinguish among the effects of
two or more explanatory
variables
• Potential problem
A B
• Example: Testing the effectiveness
Common response:
of a new drug
A B C = Lurking variable • Design of experiments: Give
C placebo to patients in clinic A and
real drug to patients in clinic B
Correlation • Even if we find difference between
exists patients in clinics A and B, what
kind of conclusion we can draw?
A ? B • The difference is due to the new
C= Confounding drug or clinics?
variable
C

14
Why should we care about correlation?
• How to confirm causation?
• Use of randomized experiments
(treatment-control study)
• Use of observational (case-control)
•
•
Easily observable
No proof needed
study if experiments are not possible or
ethical
• Evidence of a possible causal
connection:
• Not easily observable • There is a reasonable explanation of
• Require proof cause and effect
• The connection happens under varying
conditions
• Potential confounding variables are
ruled out
• Has research potential
• Most useful to us • Example: Developing a new medicine
15
 Defining AI
• For e.g. Science which
makes machines Systems mimicking Rational thinking
perform real-time tasks human reasoning systems
that require higher
intelligence,
independent of human
intervention and
decision-making
• AI is the study of formal
AI
structures and problem
solving methods, that
aims to equip
computing machines Systems mimicking Rational acting
with capabilities to
solve problems, that
are comparable to
human behavior systems
human beings
 Using formal rules to draw valid conclusions
Philosophy What does knowledge truly mean?
How can we convert knowledge into action?
Creating formal rules to draw valid conclusions
Measures and computation
Mathematics
Probability and inference
Fuzzy logic

Operations Utility theory and decision theory

Research and Game Theory
Economics Probabilistic Decision Making & Control

Fundamental Neuroscience Information processing in the brain

Structures in AI
Behavioral conditioning
Psychology Cognitive psychology
Cognitive science

Computer Efficient algorithms

Science Autonomous agents

Natural language processing

Linguistics
Knowledge representation
 Systems That Think Like Humans
• Cognitive Science
• Computational Models integrated with
psychology
• Understands Emotional States
• Robots that understand jokes (not just
tell them)
• Robots that understand human moods
• Conversation with Alexa
• Me: Alexa, tell me all about the Beatles Song
“Imagine”
• Alexa: John Lennon composed “Imagine”…
• Me: Play “it”
• Alexa: Playing “Imagine” by John Lennon
• Develop second order associations
Systems Mimicking Human
Behavior (Act like humans)
• I propose to consider the question, 'Can machines think?
• To pass the Turing Test a computer must have the
following skills:
• Natural language processing
• Knowledge representation
• Automated reasoning
• Machine Learning
• Code-Breaking Algorithms
• Decoding message & identifying who sent the message!
• British and US Navy destroyed the German Navy in
WW-II
System of logical rules
• All men are mortal;
Systems that
• à Socrates is a man; Socrates is mortal Think
• Developing first order associations… Rationally
Software Programs & Algorithms
• Optimization of Routes
• Medical diagnoses (process of
elimination)
• Decision-making under uncertainty
(maximizing likelihood)
Systems that Act Rationally
• Autonomous Agents
• Achieves the best outcome
• Under uncertainty, the best expected outcome
• May act differently from humans
 1 2 3
Formal (Structured) Tasks Unstructured (Mundane) Expert tasks (structured and
• Games Tasks mundane)
• Mathematical Problems • Perception (visual vs speech) • Design
• Natural language processing • Scientific Analysis
• Judgement • Medical decision-making
• Control • Financial analysis and stock-
picking

Common Tasks in AI