Chapter 1: Risk aversion

(Based on ch. 1 of EGS)

Zaruhi Sahakyan
Economics of Risk
ECON 469
University of Illinois
Fall 2015

August 24, 2015

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

1 / 89

Risk vs uncertainty

Frank Knight (1885 - 1972, University of Chicago) in his 1921 book Risk,
Uncertainty, and Profit discusses risk vs uncertainty.
Risk applies to situations where we do not know the outcome of a
given situation, but can accurately measure the odds (a decision leads
to consequences that are not precisely predictable, but follow a known
probability distribution).
Uncertainty, on the other hand, applies to situations where we cannot
know all the information we need in order to set accurate odds
(uncertainty or ambiguity means that the probability distribution of
outcomes is at least partially unknown to the decision maker).

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

2 / 89

Introduction

Risk exists whenever the consequences (final outcomes) of a decision

are uncertain: Economics, Finance, Insurance, Law, Medicine,
Weather, etc.
An understanding of risk and how to deal with it is an essential part
of modern economies as well as our everyday life.
It is not assumed people actually solve the mathematical problems
that we will study in this course.
Each of us has a relative who cannot solve an optimization problem,
yet decide every year to purchase an automobile insurance policy.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

3 / 89

Introduction
The focus of this course is:
To recognize, quantify, analyze, treat and incorporate risks into our
decision-making process.
To attempt to model human behavior towards risk using
1

Expected-utility theory developed by von Neumann and Morgenstern

(1948). It is based on the hypothesis that if certain axioms are satisfied,
the subjective value associated with a gamble by an individual is the
statistical expectation of that individuals valuations of the outcomes of
that gamble (not the expected value criterion, which takes into account
only the sizes of the payouts and the probabilities of occurrence).
Prospect Theory (non-expected-utility theory) which is a behavioral
economic theory, where people make decisions based on the potential
value of losses and gains rather than the final outcome. Prospect
Theory was developed by Daniel Kahneman and Amos Tversky in 1992.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

4 / 89

Historic introduction to behavior under risk

Early probability theory was developed in order to better understand

gambling problems.
Starting from the late middle ages, lotteries became an important
source of revenue for European princes:
I
I

Low administrative cost (relative to taxes)

High visibility/ difficult to avoid

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

5 / 89

Lotteries
To describe uncertainty we assume that there is a set of outcomes
(consequences) C = {c1 , c2 , ..., cs }. For example, C = { nothing, trip to
Italy, $5000}. Agent knows C and knows the probability of each outcome,
but he does not know which outcome will occur.

Definition
Lottery (or a gamble) is a probability distribution over outcomes C . That
is L = (c1 , p1 ; c2 , p2 ; ...; cs , ps ), where ci is the outcome in state i, pi is the
corresponding state probability, i pi > 0 and pi = 1.
For example,

or

1
1
1
L = (nothing , ; Italy , ; $5000, )
2
4
4
 1

1
1
2
4
4
L=
nothing Italy $5000

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

6 / 89

Lotteries
Definition
Let L denote the set of all lotteries. We will refer to these lotteries as
simple lotteries because they directly assign probability to each outcome.

Definition
Compound lottery is a lottery over lotteries, i.e. with some probability you
win one lottery and with another probability you win another lottery.
A compound lottery L could be expressed as:


p 1p
L1
L2

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

7 / 89

Lotteries
Example
1
1
Let L1 = (nothing , ; Italy , ; $5000, 0) and
2
2
1
1
L2 = (nothing , ; Italy , 0; $5000, ) are two simple lotteries. Then lottery
2
2
1
1
L = L1 + L2 is a compound lottery. Given L, the prob. of nothing
2
2
1 1 1 1
1
1
1
= + = ; prob. of Italy= ; prob. of $5000= .
2 2 2 2
2
4
4
Thus compound lottery L is equivalent to a simple lottery
 1

1
1
2

nothing

Italy

$5000

Consumer only cares about the final outcome, so we can look at simple
lotteries. Any compound lottery is equivalent to a simple lottery.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

8 / 89

Probability

Probability is a measure of the likelihood that a given state will occur

A number between 0 and 1, or a percentage
Probability of 0 means a state is impossible, probability of 1 means
its certain
Add the probabilities of two states of nature to obtain the probability
that one of those two states will occur
The probabilities of all states always add to 1 (its certain that
something will happen).

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

9 / 89

Probability
Different notions of randomness and associated probabilities:
Classical probability:
Deterministic but very complex phenomena may look random and be
modeled as such (e.g. radioactive decay; precipitation probability in
weather forecast)
Independent repeated trials, frequentist:
Example: Throw a coin 100 times
Beliefs, subjective probability:
Example: According to [Link] in January of 2013 the probability
for Lincoln to win Academy Award for Best Picture was .796
The same mathematical apparatus can handle all types of probability.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

10 / 89

Probability concepts
Sample space or probability space S
Examples: all possible outcomes of coin tosses, card deals etc.
Singletons are called elementary events cannot be decomposed any
further (Example: die roll results in a 6)
Others are composite events (Example: die roll results in a 5 or a 6)
Elementary events may depend on context (e.g. when two
indistinguishable coins are tossed, one head, one tail is an
elementary event; if the two coins are distinguishable, 1 head, 1 tail is
a composite of two elementary events)
Elementary events called states of the world: once you know which of
these has occurred, all uncertainty is resolved
Partial resolution: knowledge of composite event that contains the actual
state.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

11 / 89

Partial resolution: Conditional probability

If you know that an event E has occurred, you can update probabilities
of other events. For another event A, only the part A E remains relevant.
Conditional probability of event A, given event E is denoted by:
Pr (A|E ) =

Pr (A E )
Pr (E )

Other notations of conditional probability: Pr (A|E ) = P(A|E ) = PE (A)

Also it is useful to remember:
Pr (A E ) = Pr (A|E )Pr (E )
i.e. probability that event A and event E both occur, which is also known
as the probability of the intersection of A and E .

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

12 / 89

Partial resolution: Bayes rule

Probability of event E after learning event A.
Pr (E |A) =

Pr (A|E )Pr (E )
Pr (A)

Can use conditional probability to derive it.

Pr (E |A) =

Pr (A E )
Pr (A)

Pr (A|E ) =

Pr (A E )
Pr (E )

Substitute Pr (A E ) = Pr (A|E )Pr (E ) into Pr (E |A) =

to derive Bayes rule.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

Pr (AE )
Pr (A)

expression

August 24, 2015

13 / 89

Partial resolution: General Bayes rule

Suppose E1 , E2 , . . . Em is a partition of a sample space S,
S that is,
collection of mutually exclusive and exhaustive events: ni=1 Ei = S.
Suppose there is another family of mutually exclusive and exhaustive
events A1 , A2 , . . . An .
Bayes rule
Pr (Aj |Ei )Pr (Ei )
Pr (Ei |Aj ) = Pm
k=1 Pr (Aj |Ek )Pr (Ek )
We know the conditional probabilities Pr (Aj |Ei ), and we do get to observe
which of the Aj actually occurs. Then Bayes formula gives us the reverse
probabilities.
The probabilities Pr (Ei ) are the ones initially held, which are then revised
in the light of the information as to which of the Aj occurs. Therefore the
Pr (Ei ) are called the prior probabilities and the relevant Pr (Ei |Aj ) the
posterior probabilities.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

14 / 89

Example: Dishonest and honest types

Example (Application of Bayes rule)

Suppose the world consists of honest guys and dishonest guys, and your
prior is that the probability of a random person being honest is 70%.
When you deal with these types in a business transaction, they have
different probabilities to make an error (all errors are to your
disadvantage if you dont catch them).
Honest types make errors only inadvertently, and the probability of that
event is 10%. A dishonest type adds some intentional mistakes and thus
errs with probability 80%.
Suppose you catch a business partner in making an error. What is your
posterior probability that this person is dishonest?

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

15 / 89

Notation

Let E1 = the person is dishonest, E2 = the person is honest. So your prior

probabilities are Pr (E1 ) = 0.3, Pr (E2 ) = 0.7 . Let A1 = an error is made,
A2 = no error is made. The conditional probabilities are:
Pr (A1 |E1 ) = 0.8, Pr (A2 |E1 ) = 0.2, Pr (A1 |E2 ) = 0.1, Pr (A2 |E2 ) = 0.9
What is your posterior probability that this person is dishonest (find
Pr (E1 |A1 ))?

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

16 / 89

Solution

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

17 / 89

Partial resolution: Bayes rule

Example (Early probability theory, France)

Peter and Paul each agree to pay $15 into a jackpot. The rules that are
initially agreed to are as follows: The game consists of 7 rounds; both
players have the same skill and are equally likely to win each round.
Whoever wins a majority of rounds (i.e., at least 4 rounds) wins the
jackpot.
After 3 rounds are played (Peter won twice, Paul won once), the game has
to be interrupted before all rounds can be played out. What is a fair way
to divide the jackpot?

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

18 / 89

Partial resolution: Bayes rule

Example (Early probability theory, France)
Peter and Paul each agree to pay $15 into a jackpot. The rules that are
initially agreed to are as follows: The game consists of 7 rounds; both
players have the same skill and are equally likely to win each round.
Whoever wins a majority of rounds (i.e., at least 4 rounds) wins the
jackpot.
After 3 rounds are played (Peter won twice, Paul won once), the game has
to be interrupted before all rounds can be played out. What is a fair way
to divide the jackpot?
Peter suggests: according to the number of victories so far ($20 for
Peter, $10 for Paul)
Paul suggests: according to the number of victories still required
(and then reversed). Peter needs 2 more victories, Paul needs 3
money should be split $18 for Peter, $12 for Paul
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

19 / 89

Partial resolution: Bayes rule: Solution

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

20 / 89

Review: Combinatorics
For any set containing n elements, the number of distinct k-element
subsets of it that can be formed (the

k-combinations

of its elements) is
n
n
given by the binomial coefficient k . Therefore k is often read as n
choose k.
 
n
n!
=
k!(n k)!
k
For example
 
4
4!
4321
=
=
=4
1
1!(4 1)!
1 (3 2 1)
 
4
4321
4!
=
=6
=
2
2!(4 2)!
2 1 (2 1)
 
 
  

n
0
n
n
= 1,
= 0,
=
0
k
k
nk
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

21 / 89

Partial resolution: Independence

Events A and B are said to be independent if the occurrence of one
provides no additional information about the probability of the other, that
is
Pr (B|A) = Pr (B), or equivalentlyPr (A B) = Pr (A)Pr (B)
Examples:
A = Ace or King; Pr (A) = 8/52 = 2/13. B = Hearts,
Pr (B) = 13/52 = 1/4.
A B = Ace or King of Hearts; Pr (A B) = 2/52 = 1/26.
Independent.
A = Ace; Pr(A) = 4 / 52 = 1 / 13. B = Ace or King of Hearts,
Pr(B) = 2 / 52 = 1 / 26.
A B = Ace of Hearts; Pr (A B) = 1/52 > Pr (A)Pr (B). Not
independent.
A = Spade; Pr (A) = 13/52 = 1/4. B = Hearts;
Pr (B) = 13/52 = 1/4.
A B = ; Pr (A B) = 0 < Pr (A)Pr (B). Not independent.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

22 / 89

Probability Distributions and Populations

Mean:

PN

i=1 Xi

N
Population mean: Expected Value of X for all x: the weighted average of
all of its values. The wieghts are the probabilities P(x)
E (X ) = =

xP(x) =

N
X

xi pi

i=1

Variance:
2

PN

i=1 (Xi

)2

N
Population Variance for all x: the weighted average of the squared
deviations from the mean.
V (X ) = 2 =

N
X
X
(x )2 P(x) =
(xi )2 pi

Population standard deviation: =

Zaruhi Sahakyan (University of Illinois)

i=1

Chapter 1: Risk Aversion

August 24, 2015

23 / 89

Laws of Expected Value and Variance

E (c) = c

E (X + c) = E (X ) + c

E (cX ) = cE (X )

V (c) = 0

V (X + c) = V (X )

V (cX ) = c 2 V (X )

where c is a constant number.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

24 / 89

Is the value of a game its expected payoff?

Previous example is based on the assumption that what matters for players
is expected payoff. Is this always true?
People choose to gamble, play lotteries etc. even though the expected
payoff is less than their cost ( expected loss)
Explanation: Excitement when buying a ticket; enjoyment of thinking
what one would do if lucky
Sempronius example from book

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

25 / 89

Is the value of a game its expected payoff? (example

continued)
Example (Sempronius (by Bernoulli); EGS, p.4)
Sempronius has a cargo of spices overseas. If the cargo arrives at home, it
is worth 8000 ducats. Ships sink with probability 1/2. (In addition,
Sempronius has 4000 ducats at home).
Sempronius faces a risk on his wealth.
Lottery x takes on a value of 4000 ducats with probability
ship is sunk), or 12000 ducats with probability 12 .

1
2

(if his

Lottery x is (4000, 12 ; 12000, 21 ).

Intuitively, Sempronius is better off if he ships the cargo in two
different ships (decreases the risk that all is lost; but also decreases
the chance that all comes through).
Lottery y is (4000, 14 ; 8000, 21 ; 12000, 14 ).
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

26 / 89

Is the value of a game its expected payoff? (example

continued)

However, expected value is independent of whether cargo is shipped in one

ship or in two ships.
One ship:
1
1
E x = 4000 + 12000 = 8000ducats
2
2
Two ships:
Do it!
Hence, the mathematical expectation of a lottery is not an adequate
measure of its value.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

27 / 89

Is the value of a game its Variance? (example continued)

Variance of x and y
One ship:
1
1
Var x = (4000 8000)2 + (12000 8000)2 = 16000000
2
2
Two ships:
Do it:

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

28 / 89

Variation on the example

Suppose Sempronius is actually 5000 ducats in debt at home. Does that

change the argument (i.e., should he use two ships or just one)?

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

29 / 89

Variation on the example: Solution

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

30 / 89

St. Petersburg paradox

Example (St. Petersburg paradox (Bernoulli))

Throw a coin until Heads appears for the first time (this could obviously
happen in the first throw, but also, with a lower probability, in throw 20).
If Heads appears for the first time in throw n, I pay $ 2n to you.
How much would you pay for the opportunity to play such a game?

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

31 / 89

St. Petersburg paradox

Expected payoff?
Prob(first H on throw 1)= pi = 1/2 payoff xi = 2
Prob(first H on throw 2)= pi = (1/2)2 payoff xi = 22
etc.


1/2 (1/2)2 (1/2)n
21
22

2n
The expected value of payoffs is

X
i=1

pi x i =

X
(1/2)i 2i =
i=1

However, most people would not even be willing to pay 100$ for the right
to play this game, even if you explain that the expected value is infinite.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

32 / 89

St. Petersburg paradox

What is the explanation for this paradox behavior?
Bernoulli: People do not care about the expected value of the amount of
money they get, but rather about expected utility; the higher ones
wealth, the lower the marginal utility (the utility gain by getting an extra
dollar).
Bernoulli suggested u(x) = ln(x), i.e. u(xi ) = ln(xi ) With this, expected
utility is
Eu(
x) =

pi u(xi ) =

i=1

X
X
1
1
( )i i = ln(4)
( )i ln(2i ) = ln 2
2
2
i=1

i=1

Hence the expected utility of the game is ln(4), so the payoffs would be as
good as an amount of $ 4 which the individuals gets with certainty, and so
people should be willing to pay only $ 4.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

33 / 89

Proof:

Denote

1 i
i=1 i( 2 )

1 i
i=1 i( 2 )

=2

by S. S is the expected waiting time till we have heads.

S=

Solve for S:

Zaruhi Sahakyan (University of Illinois)

1
1
1 + (S + 1)
2
2

1
S =1S =2
2

Chapter 1: Risk Aversion

August 24, 2015

34 / 89

St. Petersburg paradox

Bernoullis theory can explain why people are not willing to pay large
amounts of money for this lottery. However, there are several problems
with Bernoullis suggestion:
The choice of the utility function is arbitrary (mostly influenced by
the desire to make computation easy).
More generally, under which basic assumptions (axioms) can a
behavior which maximizes expected utility be derived?
Are there other reasons why individuals may be reluctant to pay large
amounts of money in this game?

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

35 / 89

Risk aversion
Definition
An agent is risk averse if, at any wealth level w , he or she dislikes every
lottery with an expected payoff of zero:
For all w , z with E z = 0, u(w ) Eu(w + z).
(Strictly risk averse if the inequality is strict.) In other words u(w ) is
concave.
Intuition:
Your wealth is $10. I toss a coin and offer you $1 if it is heads and take $1
from you if it is tails
E z = 0.5 1 + 0.5 (1) = 0
Expected payoff is 0.5 11 + 0.5 9 = 10, but you reject it.
Reason: your gain in utility from another $1 is less than your loss in
utility from losing $1
Your Marginal Utility diminishes, you are risk averse (u 00 (w ) < 0).
Conversely, if you are risk averse, your Marginal Utility diminishes.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

36 / 89

Risk aversion

Equivalent Concepts
A person is risk averse
A persons marginal utility of wealth (money) diminishes
A persons utility function, u(x), is concave
A persons indifference curves are convex.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

37 / 89

Risk aversion: Sempronius example

Recall,
Initial income w = 4000, z = (0, 12 ; 8000, 21 )
Final wealth with one ship w + z x = (4000, 21 ; 12000, 12 ),
Final wealth with two ships w + z y = (4000, 14 ; 8000, 12 ; 12000, 14 ),
Final wealth with fair insurance 4000 + 8000 4000 = 8000 (pay
expected loss of 4000 as a premium and be fully covered in event of a
loss).

u(x) = x.
Should he go ahead with the shipment, i.e. should he accept the lottery?
The optimal decision is to accept the lottery z with mean 6= 0 and
variance 2 , if
Eu(w + z) u(w )

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

38 / 89

Risk aversion: Fair Premium

An insurance policy is actuarially fair if its expected net payoff is zero.

Let p denote probability of avoiding the accident. Actuarial fairness
requires that:
p(M) + (1 p)(B M) = 0
M = B(1 p)
where M is premium and B is the benefit paid by insurance company in
case of an accident.
Insurance policies are less than actuarially fair M > B(1 p)

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

39 / 89

Risk aversion: Sempronius example

The expected utilities are:

if he decides not to ship goods
at all, then he has a sure income of
4000 ducats and u(4000) = 4000 = 63.25

if using one ship: Eu(

x ) = 12 12000 + 12 4000 86.4

if using two ships: Eu(

y ) = 14 12000 + 12 8000 + 14 4000 87.9
if Sempronius
could insure his cargo for a fair premium:

u = 8000 = 89.4
Sempronius is best off with full insurance (if available), then with
diversification, and worst off without diversification.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

40 / 89

Risk aversion Figure 1.1 from EGS

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

41 / 89

Concavity and risk aversion

EGS, Proposition 1.2: A decision maker with utility function u is risk

averse if and only if u is concave.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

42 / 89

Risk premium
Definition
Risk premium associated with a zero-mean lottery is the maximum
amount of money that an individual is willing to pay in order to get rid of
a zero-mean lottery z.
Eu(w + z)
| {z }

utility in risky situation

u(w )
| {z }

utility with certain wealth, reduced by risk premium

An individual ends up with the same welfare either by accepting the risk or
by paying the risk premium .
Risk premium is a money measure rather than a utility measure.
Risk premium can be compared between different individuals
(Sempronius is more/less risk averse than Alexander).
Risk premium is willingness to pay for insurance on top of the fair
premium.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

43 / 89

Risk premium Sempronius example

Suppose Sempronius can only ship his cargo in one ship.

Eu(w + z)
| {z }

utility in risky situation

u(w )
| {z }

utility with certain wealth, reduced by risk premium

w := expected wealth (= 4000 + 12 8000 + 21 0 = 8000)

(
+4000 with probability 1/2
z =
4000 with probability 1/2

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

44 / 89

Risk premium Sempronius example

Eu(w + z) =12 4000 + 12 12000 = 86.395

u(w )
= 8000
86.395 = 8000 8000 = 86.3952 = 7464.10
= 535.90
7464.10 is called Semproniuss certainty equivalent.

Definition
Certainty equivalent is the certain amount of money that makes an
individual as well off as with the lottery.
Risk premium = Expected wealth - Certainty equivalent

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

45 / 89

Risk premium and certainty equivalent Figure 1.2

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

46 / 89

Risk premium and certainty equivalent

Risk aversion = disliking mean-zero risk

Risk aversion 6= disliking any risk (sufficiently favorable risks are
attractive even for risk averse individuals).
Risk-aversion = the certainty equivalent is smaller than the expected
prize/payoff.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

47 / 89

Risk premium Exercise

What is Semproniuss certainty equivalent and risk premium if he can use

two ships to transport his cargo?

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

48 / 89

Variation of the Example

Suppose Sempronius has to pay 2000 ducats to purchase his cargo

abroad (and then faces the transportation risk); is it worthwhile for
him?
Suppose an insurance company offers to cover Sempronius losses, but
the premium now is higher than 4000 ducats. Is it worthwhile? (I.e.,
what is the maximum willingness to pay for insurance, in order to get
rid of the risk)

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

49 / 89

Risk premium preliminaries: Taylor expansion

Approximation of a function through its derivatives

For some functions, the value at a certain point is very hard to calculate
directly.
In order to approximate these functions, we fit a polynomial to the
function.
Simplest form of polynomial: Straight line
f (x) f (x0 ) + f 0 (x0 )(x x0 )

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

50 / 89

Taylor
The line we fitted has the same function value and slope at x0 , but not the
same second derivative as the original function
Better fit:
g (x) = f (x0 ) + f 0 (x0 )(x x0 ) + c(x x0 )2
Can fit:
value
first derivative
second derivative
g 0 (x)

= f 0 (x0 ) + 2c(x x0 ) ( g 0 (x0 ) = f 0 (x0 ))

= 2c choose c = f 00 (x0 )/2.
This gives an even better approximation than the straight line.
g 00 (x)

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

51 / 89

Taylor series

In general:
f (x) f (x0 ) +

N
X
f (n) (x0 )
n=1

n!

(x x0 )n

where f (n) denotes the nth derivative of f .

This is called a (Nth order) Taylor expansion of f around x0 .

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

52 / 89

Risk premium
Definition of risk premium :
Eu(w + z)
| {z }

u(w )
| {z }

utility in risky situation

utility with certain wealth, reduced by risk premium

Second order expansion of the left-hand side (take x0 = w and x = w + zi

hence x x0 =

Pzi ):
P 
Eu(w + z) = pi u(w + zi ) pi u(w ) + u 0 (w )zi + 12 u 00 (w )zi2
= u(w )

X
X
1
pi +u 0 (w )
pi zi + u 00 (w )
pi zi2
2
| {z }
| {z }
| {z }
X

=1

Zaruhi Sahakyan (University of Illinois)

=0

Chapter 1: Risk Aversion

=Var (
z)

August 24, 2015

53 / 89

Risk premium: derivation of the formula

First-order expansion of the right-hand side (take x0 = w and x = w

hence x x0 = ):
u(w ) u(w ) u 0 (w )
u(w ) + 12 u 00 (w )Var (
z ) = u(w ) u 0 (w ) Solve for

Var (
z)
2

 00

u (w )
0
u (w )

This is known as the Arrow-Pratt approximation (Arrow (1963) and Pratt

(1964)).

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

54 / 89

Risk premium
Var (
z)

 00

u (w )
0
u (w )

Observation 1:
Risk premium is (for small risks) linear in the variance of the risk.
Risk premium is quadratic in the size of the risk
Suppose z = k (k is a factor that scales the risk, e.g., k = 2
twice as much risk, E (
z ) = 0)
2
Var (
z ) = Var (k ) = k Var (
)
for small risks (k goes to zero), the risk premium is not just small,
but very small (i.e., small even relative to the size of the risk)
Therefor at the margin, accepting a small zero-mean risk has no effect
on the welfare of risk-averse agents (risk-neutral towards small risks).
Application: Should you buy travel insurance or extended warranties for
small appliances because of risk aversion?
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

55 / 89

Risk premium
Var (
z)

 00

u (w )
0
u (w )
|
{z
}

Arrow-Pratt measure of risk aversion

Observation 2:
The Arrow-Pratt measure of risk aversion shows how risk averse a person
is: Higher Arrow-Pratt measure higher risk premium.
Observation 3:
Arrow-Pratt measure of risk aversion is invariant to (i.e., does not change
when we have) a linear transformation of the utility function:
Let v (x) = a + bu(x), where b > 0. Then

bu 00 (x)
u 00 (x)
v 00 (x)
=

.
v 0 (x)
bu 0 (x)
u 0 (x)

v () has the same measure of risk aversion as u()

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

56 / 89

Cardinal utility
In consumer theory under certainty, only the ordering of utility numbers
matters: an indifference curve higher up in the preferences should be
assigned a larger number, but how much larger is immaterial and has no
significance. Therefore that utility is said to be ordinal. Under uncertainty,
and specifically under the expected utility theory, the size of differences in
utility numbers matters, at least up to a choice of scale. For example, if
the consequences are magnitudes of wealth with w1 < w2 < w3 , then
whether [u(w2 ) u(w1 )] is bigger or less than [u(w3 ) u(w2 )] matters for
whether u(w2 ) > or < 12 u(w1 ) + 12 u(w3 ), and therefore for whether the
decision-maker would prefer to have w2 for sure, or a 50 : 50 gamble
between w1 and w3 . Therefore a non-linear transformation of the u(w )
numbers, which could change [u(w2 ) u(w1 )] and [u(w3 ) u(w2 )] in
quite different ways, can affect this comparison and therefore would not
represent the same underlying attitude toward risk.
Since the magnitudes of utility numbers matter (up to the choice of origin
and scale), the utility function that goes into expected utility calculations
is said to be cardinal.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

57 / 89

Invariance to linear transformations

Why is it necessary that AP-measure is invariant to linear transformations
of the utility function?
Linear transformations of the utility function never change the choice
behavior of an individual:
Suppose an individual with utility function u prefers z to y
X
X
pi u(zi ) >
pi u(yi )
This will remain true if we multiply both sides by b > 0:
X
X
pi bu(zi ) >
pi bu(yi )
P
We can also add a, and because
pi = 1 on both sides, we can draw that
term under the sum:
X
X
X
X
pi [a + bu(zi )] >
pi [a + bu(yi )]
pi v (zi ) >
pi v (yi )
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

58 / 89

Invariance to linear transformations

Thus, whenever the individual with utility function u prefers z to y, then

the individual with utility function v will also prefer z to y
The two utility functions u and v are observationally indistinguishable or
equivalent

Application: If Sempronius utility function is v (x) = 5 x + 99, he has

exactly the same certainty equivalent and risk premium as above, with

utility function u(x) = x.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

59 / 89

Concave transformations
Suppose v is a concave transformation of u. For example, v (u) =

u = x. Then,
q
p

x = 4 x = x 0.25
v (u(x)) = u(x) =

u and

The function v is more risk averse than the function u:

u 0 (x) = 0.5x 0.5 ; u 00 (x) = (0.5)0.5x 1.5

u 00 (x)
(0.5)0.5x 1.5
1
=

=
0
0.5
u (x)
0.5x
2x

v 0 (x) = 0.25x 0.75 ; v 00 (x) = (0.75)0.25x 1.75

v 00 (x)
(0.75)0.25x 1.75
3
=

=
0
0.75
v (x)
0.25x
4x

v has a higher risk premium than u for all risks. One can show (EGS Prop.
1.5) that this is true whenever v is a concave transformation of u.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

60 / 89

Example

Calculate Semproniuss risk premium for v (x) =

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

61 / 89

Example: Solution

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

62 / 89

Do it Yourself

Suppos Sempronius can buy an insurance in order to hedge himself against

the risk. Given the preferences on the previous slide, find Simproniuss
maximum willingness to pay for insurance (maximum insurance premium
that insurance company can charge and Semronious will be willing to pay
for it). Assume he will buy full coverage.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

63 / 89

Decreasing absolute risk aversion

It is plausible that a person has a higher risk premium for the same risk
when he is poor than when he is rich
Example:
(
+100 with probability 1/2
z =
100 with probability 1/2
Case 1: w = $101
Case 2: w = $1, 000, 000
Show that rich person is almost risk-neutral against this particular risk.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

64 / 89

Decreasing absolute risk aversion

Solution:

(
+100
z =
100

with probability 1/2

with probability 1/2

Case 1: w = $101

1
1
1+
201 = 101 101 = 57.59 = $43.41
2
2
Case 2: w = $1, 000, 000

1p
1
999, 900+
1000100 = 1000000 1000000 = 999999.9975
2
2
= $0.0025
Rich person is almost risk-neutral (against this risk)

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

65 / 89

Decreasing absolute risk aversion (DARA)

00

(x)
Let A(x) = uu0 (x)
(degree of absolute risk aversion of the agent)

 00 2
u (x)
u 000 (x)u 0 (x) [u 00 (x)]2
u 000 (x) u 00 (x)
A (x) =
=
+ 0
= 00
0
2
0
[u (x)]
u (x) u (x)
u (x)
0

= A(x)[A(x) P(x)]
000

(x)
: the degree of absolute prudence of agent with utility u
P(x) = uu00 (x)
0
For A (x) < 0, it must be that P(x) > A(x). A necessary condition for
this is u 000 (x) > 0.

1
is decreasing in x.
Example: For u = x, A(x) = 2x
000
DARA requires that u is positive (marginal utility is convex):

Zaruhi Sahakyan (University of Illinois)

000

3 5
= x2
8

Chapter 1: Risk Aversion

August 24, 2015

66 / 89

Relative risk aversion

Absolute risk aversion is the rate of decay (negative rate of growth) of
marginal utility when wealth is increased by one dollar.
Rate of growth of function f (x) is given by:
f (x) = u 0 (x)

df (x) 1
dx f (x) ,

where

Problem: This measure depends on units (whether wealth is

measured in dollar, euro, yen). Economists prefer unit-free measures
(cf. elasticities)
Relative risk aversion: The rate at which marginal utility decreases
when wealth is increased by one percent ( measurement units are
irrelevant).
R(w ) =

du 0 (w )/u 0 (w )
wu 00 (w )
= 0
= wA(w )
dw /w
u (w )

R(w ) is the wealth-elasticity of marginal utility.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

67 / 89

Relative risk aversion

Relative risk aversion: Important for relative risk premium/portfolio risk

Eu(w (1 + z)) = u(w (1 ))

Second order Taylor expansion of the left-hand side:
1
u(w ) + u 0 (w )E (w (
z )) + u 00 (w )E (w 2 z2 )
2

First order Taylor expansion of the right hand side: u(w ) u 0 (w )w

Setting these terms equal and simplifying:
= 1 w 2 u 00 (w )E (
z 2)
u 0 (w )w
2
00

z ) u (w )w
= Var (

2
u 0 (w )
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

68 / 89

Relative risk aversion: Application

z ) u 00 (w )w
Var (
z)
= Var (

=
R(w )
0
2
u (w )
2
Suppose you face a relative risk of 20 percent of your expected wealth
level, with equal probability, i.e. z = (0.2, 21 ; 0.2, 12 )
shows the share of your initial wealth that you are ready to pay to

get rid of the proportional risk z.

1
u(w ) = w , R(w ) = wA(w ) = w 2w
= 12

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

69 / 89

Relative risk aversion: Application

Private sector job (risky option): Lifetime wage earnings either

1,600,000 or 2,400,000 (i.e., 40 years at $40,000 per year, or at
$60,000 per year, in constant dollar terms).
State sector job (no risk option).
Question: How much certain income per year would you need to be
indifferent (monetarily) between state and private jobs?
as a percentage of
Remember to calculate relative risk premium
expected wealth (E w
= $50, 000 per year)

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

70 / 89

Relative risk aversion: Application: Solution

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

71 / 89

Quadratic utility

Early researchers in finance, such as Markowitz and Sharpe, used just the
mean and the variance of the return rate of an asset to describe it.
Mean-variance characterization is often easier than using an von NM
utility function.
Q: But is it compatible with vNM theory?
A: Yes, but under some conditions.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

72 / 89

Quadratic utility (IARA)

Quadratic utility is concave.
u(w ) = aw 12 w 2 for w < a.
We consider the domain of w where u is nondecreasing.
1
1
A(w ) = aw
= aw
1
Increasing absolute risk aversion: A0 (w ) = (aw
> 0 (bad feature
)2
for a utility function)
Preferences over lotteries depend only on average and variance. Lets show
that:
1 2
1
Eu(w
) = E (aw
w
) = aE (w
) E (w
2)
2
2
From statistics:
Var (w
) = E ([w
E (w
)]2 ) = E (w
2 ) 2E (w
)E (w
) + [E (w
)]2 , so that
E (w
2 ) = Var (w
) + [E (w
)]2 .

1
1
Eu(w
) = aE (w
) [E (w
)]2 Var (w
).
2
2
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

73 / 89

Exponential utility (CARA)

u(w ) = e rw . Concave and increasing. (Negative, but this does not

matter.)
2 rw

e
Constant absolute risk aversion (CARA): A(w ) = rre rw

=r

Increasing relative risk aversion (IRRA): R(w ) = rw

Individuals risk premium for a given fixed risk does not depend on the
level of wealth
Portfolio investment: Same amount of risky assets, independent of
weath level (e.g., invest $10,000 in shares, no matter whether initial
wealth is $12,000 or $1,000,000)

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

74 / 89

Power utility (CRRA)

1

u(w ) = w1 , > 0 and 6= 1.

Think what happens if < 1 / > 1.
1

Decreasing absolute risk aversion (DARA): A(w ) = w

w

Constant relative risk aversion (CRRA): R(w ) = wA(w ) = w w = .

Individuals risk premium for a given fixed risk decreases with the level
of wealth
Individuals risk premium for a given relative risk does not depend on
the level of wealth
Portfolio investment: Same percentage of risky assets, independent of
wealth level (e.g., invest $10,000 in shares when wealth level is
$20,000 invest $500,000 in shares if initial wealth is $1,000,000)
What if = 1? Power utility function is not defined, but A(w ) =
appears perfectly normal.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

1
w

75 / 89

Logarithmic utility (CRRA)

Let u(w ) = ln(w ).

1

Decreasing absolute risk aversion (DARA): A(w ) = w12 =

w

1
w

Constant relative risk aversion (CRRA): R(w ) = wA(w ) = 1

Logarithmic utility is the missing member of the power utility family.
All other properties of the CRRA family apply.
Lots of human senses satisfy logarithmic scales (e.g. sounds measured
in dB (decibel); magnitude of stars)

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

76 / 89

Expected utility theory

So far: assume that individual choices can be summarized by
expected utility maximization for some utility function.
Under which fundamental assumptions on preferences over risky
situations (lotteries L) is it actually possible to capture preferences
by a utility function?
An expected utility function is an extremely dense way to describe an
individuals preferences over lotteries. There are infinitely many
different lotteries.
Describing preferences in a list (e.g. LA  LB , LB  LC , etc) is not
practically feasible because there are infinitely many different lotteries.
If I know an individuals expected utility function (1 line!), I know
how this individual will behave in any choice situation between
lotteries (I just need to plug payoff and probabilities of the different
lotteries in the expected utility function)
But: When can we actually do this?? John von Neumann and
Oskar Morgenstern (1944, Theory of Games)
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

77 / 89

Axioms
Notation: A lottery is described by

p1 p2
x1 x2

pn
xn

Let L be the set of all lotteries.

A standard lottery is a lottery that has just two outcomes, a good one
(xmax ) and a bad one (xmin < xmax ).


1u
u
xmin xmax
There are 4 axioms of expected utility theory.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

78 / 89

Axioms: Ordering and non-satiation

Axiom (Ordering)
The agent has a complete and transitive ordering on L.

Axiom (Non-satiation)


 

1 u2 u2
1 u1 u1

u1 > u2
xmin xmax
xmin xmax

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

79 / 89

Axioms: Continuity

Axiom (Continuity)
For all x [xmin , xmax ], there exists a probability u(x) such that


1 u(x) u(x)
x
xmin
xmax
u(x) is called the equivalent winning probability.
Example: Getting $ 1000 for sure is as good as getting $ 5000 with
probability 1/4. (Note: This is just one individuals preference, and it is
certainly not true for everybody!)

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

80 / 89

Axioms: Independence
Axiom (Independence)
Let Li , Lj , Lk L

 and Li  Lj (Lk is arbitrary). Then, for all p, we have
1p p
1p p

Li
Lk
Lj
Lk
Reason: With prob. 1 p, the two big lotteries deliver different payoffs
(either Li or Lj ), and we know that Li  Lj . So in this case, the first big
lottery is better than the second big lottery.
With prob. p, both lotteries deliver the same payoff Lk and hence in this
case, both lotteries are equally good.
=First big lottery is more attractive
Intuitive, but apparently the critical axiom of the theory.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

81 / 89

Allais experiment: Violation of independence axiom




100%
Choice 1: Choose either A1 =
or
$ 1 million


1%
89%
10%
B1 =
0 $ 1 million $ 5 million

Choice 2: Choose either A2 =





89%
11%
90%
10%
or B2 =
0
$ 1 million
0
$ 5 million
How would you choose?
Number of people choosing:
A1 , A2 A1 , B2 B1 , A2 B1 , B2

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

82 / 89

Allais experiment
Consider
 choice 1; rewrite A1 and
 B1 (without any substantive change) as
89%
11%
A1 =
and
$ 1 million $ 1 million

89%
11%


1/11
10/11
B1 =
$ 1 million
0
$ 5 million
Independence axiom: Any preference for A1 or B1 cannot depend on the
respective first parts of the lotteries (i.e., 89% chance of getting 1 mio.;
this is the Lk in the independence axiom).


1/11
10/11
Thus, A1  B1 if and only if $ 1 million 
0
$ 5 million
Similarly, the choice between A2 and B2 can be broken down to the same
question. (Do it!)
A1  B1 and A2 B2 is inconsistent with the independence axiom.
A1 B1 and A2  B2 is inconsistent with the independence axiom.
Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

83 / 89

Expected-utility theorem
Proposition (Expected-utility theorem)
If an individual satisfies axioms 1-4, he behaves as if he were maximizing
expected utility for some function u
X
pi u(xi )
i

General idea of the proof:

For every lottery, find an equivalent standard lottery.
According to the non-satiation axiom, preference among standard
lotteries is determined by the winning probabilities
Winning probabilities for the standard lotteries are in expected
utility form.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

84 / 89

Expected-utility theorem
Proof: Let LA  LB . (Axiom 1 guarantees that we can always order two
lotteries). Consider x1 (first possible prize);
by the continuity
axiom, there


1 u1 u1
exists a probability u1 such that x1
.
xmin xmax
Independence axiom: we can substitute this lottery for x1 in

A
A
 A

p1A

 p2 pn
p1 p2A pnA

LA =
1 u1 u1
x1 x2 xn
x2 xn
xmin xmax
Substitute sequentially all xi in LA and LB equivalent standard lotteries

A
A
p
p

1
2

 


1 u2 u2
L0A = 1 u1 u1

xmin xmax
xmin xmax

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

85 / 89

Expected-utility theorem
Similar substitution for L0B .

B
B
p
p

1
2

 


1 u2 u2
L0B = 1 u1 u1

xmin xmax
xmin xmax
Now L0A and L0B can be compared using the non-satiation axiom. The logic
goes as follows:
LA  LB (by completeness)
L0A LA and L0B LB (by construction)
L0A LA  LB L0B L0A  P
L0B (by transitivity)
P
0
0
Hence LA  LB if and only if i piA ui i piB ui (by non-satiation)
This is the Expected utility form that we wanted to arrive at.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

86 / 89

Important properties of u

1. u is not unique! It depends on xmin and xmax which were arbitrary.

2. v (x) = a + bu(x), with b > 0, reflects the same preferences.
The ui s in the proof were probabilities (0 ui 1), but functions
with negative utility (or utility bigger than 1) are feasible utility
functions as well.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

87 / 89

Important properties of u

3. However, arbitrary (non-linear) transformations of u are not allowed

since they change the behavior of the
 individual.

3/4 1/4
Example: Compare $ 4 certain and
$0 $16

u(x) = x u(4) = 2;
(3/4)u(0) + (1/4)u(16) = 1 this individual prefers $ 4 certain over
the lottery.
v (x) = [u(x)]4 = x 2 . v (4) = 16, and
(3/4)v (0) + (1/4)v (16) = 64 this individual prefers the lottery
over $4 certain.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

88 / 89

Important properties of u

4. Concavity corresponds to risk aversion.

A function u is concave if for all x1 , x2 and 0 < < 1 the following
inequality holds:
u(x1 + (1 )x2 ) u(x1 ) + (1 )u(x2 ).
If can be replaced by > (for all x1 6= x2 ), then u is strictly
concave. A function v is convex if v is concave.

Zaruhi Sahakyan (University of Illinois)

Chapter 1: Risk Aversion

August 24, 2015

89 / 89