Choice under Uncertainty

Lorenzo Bretscher

1 / 23
Uncertainty

We will try to understand the pricing of financial assets from an equilibrium perspective.
In an equilibrium, prices are such that they equate investors’ demand for assets with their
supply. To understand investors demand, we first need to understand how they make
choices in uncertain environments. We will consider a simple model of choice under
uncertainty.
For now, there is one period (i.e. two dates). This will be straightforward to
re-interpret in multi-period settings later on.
Uncertainty:
I The state of the world at the beginning of the period is known with certainty.
At the end of the period there are S discrete states, and we denote the state
space as ⌦ = {!1 , !2 , ..., !S }.
I The associated objective probabilities are ⇧ = {⇡ 1 , ⇡ 2 , ..., ⇡ S }, and these are
also the subjective probabilities of all agents. Clearly, ⇡ 1 + ... + ⇡ S = 1.

2 / 23
Consumption Plans

Individuals care about their consumption in each state of the world.

A consumption plan C is a random variable on {⌦, ⇧}. In our setup, we can
equivalently represent C as a vector (c1 , c2 , ...., cS ) where cj is the consumption
level when state !j is realized.
Example with S = 3:

state !j 1 2 3
probability ⇡j 0.2 0.3 0.5
cons plan C (1) 3 2 4
cons plan C (2) 3 1 5
cons plan C (3) 4 4 1
cons plan C (4) 1 1 4

We want to develop a notion of how individuals choose among the set C of all
consumption plans consistently. Preferences that lead to consistent choices are
called a preference relation.
3 / 23
Consumption Lotteries

Let us assume that individuals only care about the probability distributions over
consumption and thus that preferences are state-independent. Then, we can represent
consumption plans as consumption lotteries over outcomes rather than states.
Let Z denote the set of all consumption levels z generates by any C 2 C. Each
consumption plan C is associated with a probability distribution pC , which is the
function pC : Z ! [0, 1] given by
X
pC (z) = ⇡! (1)
!2⌦:c! =z

Example (continued): Z = {1, 2, 3, 4, 5} and

cons level z 1 2 3 4 5
cons plan C (1) , pC 1 0 0.3 0.2 0.5 0
cons plan C (2) , pC 2 0.3 0 0.2 0 0.5
cons plan C (3) , pC 3 0.5 0 0 0.5 0
cons plan C (4) , pC 4 0.5 0 0 0.5 0

Two different plans, C (3) and C (4) give rise to the same probability distribution over
Z.
4 / 23
Preferences over Consumption Lotteries

A probability distribution p on the set of all possible consumption levels Z is

called a lottery. We denote by P(Z ) the set of all lotteries on Z that are
generated by consumption plans in C.
For two lotteries, p and p 0 , and ↵ 2 [0, 1], any convex combination
↵p + (1 ↵)p 0 is also a lottery, namely the compound lottery. In other
words, P is convex.

5 / 23
A Safe and a Risky Lottery

USD 10 million with certainty ! LS = ($10m, 0, 100%)

Coin toss: USD 30 million with 50% probability ! LR = ($30m, 0, 50%)

More general:
Suppose there are two possible states of the world: x and y . Further assume that
x materializes with probability ⇡ and, hence, y with 1 ⇡.

Notation: L = (x, y , ⇡)

6 / 23
Preferences over Consumption Lotteries

We assume that individuals have preferences over outcomes and represent them
by a preference relation ⌫, where for two lotteries p and p 0 we denote

p ⌫ p0 , p is at least as good as p’ (2)

We say that p is strictly preferred to p 0 , or p p 0 , if p ⌫ p 0 but not p 0 ⌫ p. We

say that an agent is indifferent between p and p 0 , or p ⇠ p 0 , if p ⌫ p 0 and p 0 ⌫ p.
A preference relation is called complete, if for all p and p 0 2 P, either p ⌫ p 0
or p 0 ⌫ p, or both.
A preference relation is called transitive, if whenever p ⌫ p 0 and p 0 ⌫ p 00 ,
then p ⌫ p 00 .

7 / 23
An Example

L1 = (1, 5, 10%), L2 = (1, 5, 30%), L3 = (3, 4, 50%)

Suppose L3 ⌫ L2
and L2 ⌫ L1
! L1 ⌫ L3 violates transitivity

8 / 23
Choice Rules

We will assume that preference relations are complete and transitive in the following.

Given preferences, how will an agent behave? We assume that given a set of choices
(lotteries) B 2 P, the agent will choose the element of B she prefers most. To formalize
this, we define an agent’s choice rule given a preference relation ⌫

(B; ⌫) = {p 2 B|p ⌫ p 0 for all p 0 2 B}

Note that
(B; ⌫) may contain more than one element
If B is finite, (B; ⌫) is non-empty.
While ultimately, what we observe about a preference relation ⌫ are choices, i.e (B; ⌫),
it is convenient to try to represent them by a numerical ranking, namely a utility. We
now examine under what conditions this is possible.

9 / 23
Utility Indices

A utility index for a given preference relation ⌫ is a function U : P(Z ) ! R

which assigns to each lottery a real-valued number such that

p ⌫ p0 , U(p) U (p 0 )

Note that a utility index is only unique up to a strictly increasing transformation. If

U is a utility index for a preference relation ⌫ and f : R ! R is any strictly
increasing function, then the composite function V = f (U) is also a utility index for
⌫.
Armed with a utility index finding (B; ⌫) ’just’ amounts to finding the choice with the
greatest utility. We thus turn to the question under which conditions a preference
relation can be represented by a utility index.

10 / 23
Utility Indices: Axioms - Monotonicity

Suppose that p, p 0 2 P(Z ) with p p 0 and let a, b 2 [0, 1]. The preference
relation ⌫ has the monotonicity property if

a>b , ap + (1 a)p 0 bp + (1 b)p 0

In words: If you consider a weighted average of lotteries, you prefer a higher

weight on the better of the two lotteries.

11 / 23
Utility Indices: Axioms - Archimedean Property

Suppose that p, p 0 , p 00 2 P(Z ) with p p 0 p 00 . The preference relation ⌫

has the Archimedean property if there exist numbers a, b 2 (0, 1) such that

ap + (1 a)p 00 p0 bp + (1 b)p 00

In words: No matter how good a lottery is, if we put sufficiently low weight
on it, we find a compound lottery to which we prefer p 0 . Similarly, no matter
how bad a lottery is, if we put sufficiently low weight on it, we find a
compound lottery which we prefer to p 0 .

12 / 23
Utility Indices: Axioms - Continuity Property

Suppose that p, p 0 , p 00 2 P(Z ) with p p 0 p 00 . The preference relation ⌫

has the continuity property if there exists a unique number a 2 (0, 1) such
that

p 0 ⇠ ap + (1 a)p 00

In words: For any three lotteries, there is a unique combination between the
best and the worst lottery such that the agent is indifferent between the
middle lottery and that combination.

13 / 23
Continuity Property

Proposition: Let ⌫ be a preference relation satisfying the Monotonicity

Axiom and the Archimedean Axiom. Then it has the continuity property.

14 / 23
Existence of Utility Indices

Theorem: Let ⌫ be a preference relation satisfying the Monotonicity Axiom

and the continuity property. Then it can be represented by a utility index U .

15 / 23
Expected Utility

Under relatively uncontroversial conditions a preference relation can be represented by a

utility index. These are functions of an entire set of probability distributions (lotteries)
and can be difficult to work with. We now introduce another representation of preference
relations, which is both very common and easy to work with, but only obtains under
stronger and, arguably, more controversial conditions.
A preference relation ⌫ on P(Z ) has an expected utility representation if there
exists a function u : Z ! R such that
X X 0
p ⌫ p0 , p(z)u(z) p (z)u(z)
z2Z z2Z
P
Here z2Z p(z)u(z) is the expected end of period consumption given the
consumption lottery p.
P
In other words, z2Z p(z)u(z) = E [u(c)], where c is the random variable
representing end of period consumption given the lottery.
Note that the function u is defined on Z only. It is usually called the von
Neumann-Morgenstern utility function.
Given a PvN-M utility function u we can define a utility index by
U(p) = z2Z p(z)u(z).
16 / 23
Expected Utility

We now turn to conditions under which preference relations have an expected

utility representation.
Proposition A preference relation ⌫ has an expected utility representation if
and only if it can be represented by a linear utility index U in the sense that

U (↵p + (1 ↵)p 0 ) = ↵U (p) + (1 ↵)U (p 0 )

for any p, p 0 2 P(Z ) and any ↵ 2 (0, 1).

17 / 23
Independence Axiom

Example:
z 1 2 3 4
p1 0 0.2 0.6 0.2
p2 0 0.4 0.2 0.4
p3 1 0 0 0
p4 0.5 0.1 0.3 0.1
p5 0.5 0.2 0.1 0.2
Suppose we want to compare p4 and p5 . They only differ in the probabilities of
consumption levels 2, 3 and 4, so it should suffice to compare those. More formally, we
have
p4 ⇠ 0.5p1 + 0.5p3 and p5 ⇠ 0.5p2 + 0.5p3
p1 is the conditional distribution of p4 given that z is different from 1, and p2 is the
conditional distribution of p5 given that z is different from 1. The independence axiom
then says
p4 p5 , p1 p2
18 / 23
Independence Axiom

Proposition: If a preference relation ⌫ satisfies the Independence Axiom, it will also

satisfy the Monotonicity Axiom.

19 / 23
von Neumann-Morgenstern Theorem

Theorem (von Neumann-Morgenstern): Assume that Z is finite and that

⌫ is a preference relation on P(Z ). Then ⌫ can be represented by a linear
utility index if and only if ⌫ satisfies the Archimedean and the Independence
Axiom.

20 / 23
von Neumann-Morgenstern Theorem

Proposition: A vN-M utility function is only determined up to a strictly

increasing affine transformation, ie, if u is a utility function for ⌫, then v will
be so iff there exist constants a > 0 and b such that v (z) = au(z) + b for all
z 2 Z.

21 / 23
The Allais Experiment

The Independence Axiom is key to establishing an expected utility representation of a

preference relation. One common concern with the Independence Axiom is that there are
multiple experiments showing that people’s behavior often violates it. One of the most
famous experiments is due to Allais (1955).
Problem 1: Choose between two lotteries. The first pays $55’000 with probability
0.33, $48’000 with probability 0.66, and zero with probability 0.01. The second
pays $48’000 for sure.
Problem 2: Choose between two lotteries. The first pays $55’000 with probability
0.33, and nothing with probability 0.67. The second pays $48’000 with probability
0.34 and nothing with probability 0.66

22 / 23
The Allais Experiment

The Independence Axiom is key to establishing an expected utility representation of a

preference relation. One common concern with the Independence Axiom is that there are
multiple experiments showing that people’s behavior often violates it. One of the most
famous experiments is due to Allais (1955).
Problem 1: Choose between two lotteries. The first pays $55’000 with probability
0.33, $48’000 with probability 0.66, and zero with probability 0.01. The second
pays $48’000 for sure.
Problem 2: Choose between two lotteries. The first pays $55’000 with probability
0.33, and nothing with probability 0.67. The second pays $48’000 with probability
0.34 and nothing with probability 0.66
Often, people choose the certain payoff in the first case and take the first lottery in the
second case. This, however, violates the the independence axiom. Normalizing u(0) = 0,
we have
u(48) > 0.33u(55) + 0.66u(48) , 0.34u(48) > 0.33u(55)
while the second choice implies the reverse inequality.
One explanation (due to Kahnemann and Tversky) is to posit a certainty effect: people
tend to overweigh a sure thing.
22 / 23
The Ellsberg Experiment

The expected utility framework assumes that agents assign probabilities to all outcomes.
In the terminology of Knight-Savage, they deal with risk (ie situations in which it is
possible to assign probabilities) as opposed to uncertainty (ie situations where one is just
clueless). In a famous experiment, Ellsberg (1961) casts doubt on the probability
assignments.
An urn contains 300 balls; 100 are red and 200 are some mix of white and blue. We
are going to draw a ball at random.
I You will receive $100 if you correctly guess the ball’s color. Would you rather

guess white or red?

I You will receive $100 if you correctly guess a color different than that of the

ball. Would you rather guess white or red?

23 / 23
The Ellsberg Experiment

The expected utility framework assumes that agents assign probabilities to all outcomes.
In the terminology of Knight-Savage, they deal with risk (ie situations in which it is
possible to assign probabilities) as opposed to uncertainty (ie situations where one is just
clueless). In a famous experiment, Ellsberg (1961) casts doubt on the probability
assignments.
An urn contains 300 balls; 100 are red and 200 are some mix of white and blue. We
are going to draw a ball at random.
I You will receive $100 if you correctly guess the ball’s color. Would you rather

guess white or red?

I You will receive $100 if you correctly guess a color different than that of the

ball. Would you rather guess white or red?

It turns out that people tend to choose red in both cases. This, however, is inconsistent
with people assigning probabilities to each event.
Suppose an agent has assigned a probability to the ball that is drawn being white.
If that probability is greater than 1/3, she must pick (white, red). If that probability
is less than 1/3, she must pick (red, white). Picking (red, red) implies a probability
on white which is both strictly less and strictly greater than 1/3.
23 / 23
Risk and Risk Attitudes

Lorenzo Bretscher

1 / 27
Risk and Risk Attitudes

Let us now assume that we are in an expected utility framework. Agents face
uncertainty, but can assign probabilities to all outcomes. In other words, they are
exposed to risk. We now want to examine what attitudes to risk agents’
preferences exhibit.

In an expected utility framework, this translates into properties of vN-M utility

functions, defined as u : Z ! R.

2 / 27
Risk Attitudes

Fix a consumption level W 2 Z . Consider a random variable "˜, with

"] = 0. Such a random variable is called a fair gamble.
E [˜
We can think of W + "˜ as a stochastically perturbed consumption plan.

3 / 27
Risk Attitudes

Fix a consumption level W 2 Z . Consider a random variable "˜, with

"] = 0. Such a random variable is called a fair gamble.
E [˜
We can think of W + "˜ as a stochastically perturbed consumption plan.
An individual is said to be (strictly) risk averse if for all W 2 Z and all fair
gambles "˜ she (strictly) prefers the certain consumption level W to W + "˜.
In other words, a risk averse individual rejects all fair gambles.

3 / 27
Risk Attitudes

Fix a consumption level W 2 Z . Consider a random variable "˜, with

"] = 0. Such a random variable is called a fair gamble.
E [˜
We can think of W + "˜ as a stochastically perturbed consumption plan.
An individual is said to be (strictly) risk averse if for all W 2 Z and all fair
gambles "˜ she (strictly) prefers the certain consumption level W to W + "˜.
In other words, a risk averse individual rejects all fair gambles.
An individual is said to be (strictly) risk loving if for all W 2 Z and all fair
gambles "˜, she (strictly) prefers W + "˜ to W .

3 / 27
Risk Attitudes

Fix a consumption level W 2 Z . Consider a random variable "˜, with

"] = 0. Such a random variable is called a fair gamble.
E [˜
We can think of W + "˜ as a stochastically perturbed consumption plan.
An individual is said to be (strictly) risk averse if for all W 2 Z and all fair
gambles "˜ she (strictly) prefers the certain consumption level W to W + "˜.
In other words, a risk averse individual rejects all fair gambles.
An individual is said to be (strictly) risk loving if for all W 2 Z and all fair
gambles "˜, she (strictly) prefers W + "˜ to W .
An individual is said to be risk neutral if for all W 2 Z and all fair gambles
"˜, she is indifferent between W + "˜ and W .

3 / 27
Risk Attitudes

Fix a consumption level W 2 Z . Consider a random variable "˜, with

"] = 0. Such a random variable is called a fair gamble.
E [˜
We can think of W + "˜ as a stochastically perturbed consumption plan.
An individual is said to be (strictly) risk averse if for all W 2 Z and all fair
gambles "˜ she (strictly) prefers the certain consumption level W to W + "˜.
In other words, a risk averse individual rejects all fair gambles.
An individual is said to be (strictly) risk loving if for all W 2 Z and all fair
gambles "˜, she (strictly) prefers W + "˜ to W .
An individual is said to be risk neutral if for all W 2 Z and all fair gambles
"˜, she is indifferent between W + "˜ and W .
Note that agents may be locally risk averse, locally risk loving and locally risk
neutral.

3 / 27
Risk Aversion

Any consumption plan can be thought of as a sum of its expected value, E [W̃ ]
and a fair gamble "˜ = W̃ E [W̃ ]. Under expected utility, an agent is thus risk
averse if and only if she prefers the certain consumption E [W̃ ] to the random
consumption W̃ , or
u(E [W̃ ]) E [u(W̃ )]
By Jensen’s inequality, this holds exactly whenever u is concave.
A function u : Z ! R is concave if for all z, z 0 2 Z and all ↵ 2 (0, 1) we
have
u(↵z + (1 ↵)z 0 ) ↵u(z) + (1 ↵)u(z 0 )
When the inequality is strict, u is strictly concave.
Whenever u is twice differentiable in the interior of Z , concavity is equivalent
to u 00  0.

4 / 27
Risk Aversion

Theorem: An individual with vN-M utility function u is (strictly) risk-averse

if and only if u is (strictly) concave.
An individual with vN-M utility function u is (strictly) risk-loving if and only
if u is (strictly) convex.
An individual with vN-M utility function u is risk-neutral if and only if u is
affine.

5 / 27
Risk Premium

The certainty equivalent of a random consumption plan W̃ is defined as

the Wc 2 Z such that
u(Wc ) = E [u(W̃ )]
If u is continuous and strictly increasing, Wc is uniquely defined.
For a risk-averse individual, the certainty equivalent Wc of a consumption
plan is smaller than the expected consumption E [W̃ ]. The risk premium
associated with the consumption plan W̃ is defined as ⇡ = E [W̃ ] Wc , so
that
E [u(W̃ )] = u(Wc ) = u(E [W̃ ] ⇡)
The risk premium is the consumption the individual is willing to give up in
order to eliminate risk.
We next want to come up with quantitative measures of risk aversion. These
will measure how much consumption an individual is willing to give up to
eliminate risk.
6 / 27
7 / 27
Local Measures of Risk Aversion

Notation:
wealth or consumption level W
2
fair gamble "˜ with Var(˜
") =
Wc certainty equivalent of W̃ = W + "˜
⇡=W Wc risk premium

8 / 27
Absolute Risk Aversion

The case of additive uncertainty: W̃ = W + "˜

" + 12 U 00 (W )˜
U(W + "˜) ⇡ U(W ) + U 0 (W )˜ "2
2
) E[U(W̃ )] ⇡ U(W ) + 2 U 00 (W )

and
U (Wc ) = U(W ⇡) ⇡ U(W ) U 0 (W )⇡
So the risk premium can be approximated by
2 2
U 00 (W )
⇡⇡ = A(W )
2 U 0 (W ) 2
U 00 (W )
where A(W ) = U 0 (W ) is the coefficient of absolute risk aversion

9 / 27
Absolute Risk Aversion

The risk premium for a small fair gamble around W is approximately equal to
the absolute risk aversion at W
Absolute risk aversion is constant if and only if ⇡ is independent of W
Absolute risk aversion measures the aversion to a fair gamble of a given
(dollar) amount, say winning or losing $1000 dollars with equal probability.
Intuitively, we would expect a wealthier investor to be less averse to such
gambles, ie we would expect absolute risk aversion to be a decreasing
function of wealth (or end of period consumption). We have
✓ ◆2
0 U 000 (W )u 0 (W ) U 00 (W )2 U 00 (W ) U 000 (W )
A (W ) = = <0
U 0 (W )2 U 0 (W ) U 0 (W )
, U 000 (W ) > 0

10 / 27
Comparing Risk Aversion

Let ⇢ be the fraction of W the agent would be prepared to give up in order to

avoid the risk of W̃ : Wc = (1 ⇢)W
2
E[U(W̃ )] ⇡ U(W ) + W 2 U 00 (W )
2
and
U (Wc ) = U(W ⇢W ) ⇡ U(W ) U 0 (W )⇢W
Therefore,
2 2
WU 00 (W )
⇢⇡ = R(W )
2 U 0 (W ) 2
WU 00 (W )
where R(W ) = U 0 (W ) = W A(W ) is coefficient of relative risk aversion.
Note that utility functions with constant relative risk aversion will exhibit
decreasing absolute risk aversion.
11 / 27
Comparing Risk Aversion

An individual with vN-M utility function U is said to be more risk averse than
an individual with vN-M utility function V if for any consumption plan W̃
and any fixed W with E [U(W̃ )] > U(W ), we have E [V (W̃ )] > V (W ).
In words: The agent with utility function v will accept all gambles that the
agent with utility function u accepts, and possibly more.
Theorem: Suppose U and V are twice continuously differentiable and
strictly increasing. Then the following conditions are equivalent:
I (a) U is more risk averse than V
I (b) AU (W ) AV (W ) for all W
I (c) a strictly increasing and concave function f exists such that U = f (V )

12 / 27
Common Utility Functions

CRRA utility: constant relative risk aversion utility or power utility. Utility
functions in this class are defined for W 0 as
W1
U(W ) =
1
for > 0 and 6= 1. Differentiating we find that the absolute and relative
risk aversion coefficients are given by

A(W ) = and R(W ) = W A(W ) =

W
CARA utility: constant absolute risk aversion utility or (negative)
exponential utility. Utility functions in this class are defined as
↵W
U(W ) = e
where ↵ is the coefficient of absolute risk aversion. Recall that CARA implies
a risk premium which is independent of wealth.
13 / 27
CRRA: Example

Consider two consumption plans, namely one with certain consumption level W and a
second one delivering (1 ↵)W with probability one half and (1 + ↵)W with probability
one half.
With CRRA utility, the certainty equivalent of the second plan is given by
1 1 1 1 1
(Wc )1 = ((1 ↵)W )1 + ((1 + ↵)W )1
1 21 21
which implies that

1 1/(1 )
Wc = [(1 ↵)1 + (1 + ↵)1 ]1/(1 )
W
2
The risk premium is
✓ ◆1/(1 !
)
1
⇡=W Wc = 1 [(1 ↵)1 + (1 + ↵)1 ]1/(1 )
W
2

14 / 27
CRRA: Example

Risk premia for different levels of relative risk aversion and different gambles:

= RRA ↵ = 1% ↵ = 10% ↵ = 50%

0.5 0.00% 0.25% 6.70%
1 0.01% 0.50% 13.40%
2 0.01% 1.00% 25.00%
5 0.02% 2.43% 40.72%
10 0.05% 4.42% 46.00%
20 0.10% 6.76% 48.14%
50 0.24% 8.72% 49.29%
100 0.43% 9.37% 49.65%

15 / 27
Risk Aversion and Portfolio Choice

Two assets:
risk-free asset, gross return R0
risky asset, return R̃ with mean R
Amount invested in risky asset: a
No restrictions (unlimited borrowing and short-selling)
Next period wealth:
⇣ ⌘
W̃ = R̃a + R0 (W0 a) = R̃ R0 a + R0 W0

Portfolio selection problem:

h ⇣⇣ ⌘ ⌘i
max E U R̃ R 0 a + R 0 W0
a2R

16 / 27
Risk Aversion and Portfolio Choice

Write objective function as f (a) and differentiate:

h ⇣⇣ ⌘ ⌘⇣ ⌘i
0 0
f (a) = E U R̃ R0 a + R0 W0 R̃ R0
 ⇣⇣ ⌘ ⌘⇣ ⌘2
00 00
f (a) = E U R̃ R0 a + R0 W0 R̃ R0

U 00 < 0 ) f 00 < 0 ) f strictly concave

First order condition:

h ⇣ ⌘⇣ ⌘i
0 ⇤ 0 ⇤
f (a ) = E U W̃ R̃ R0 =0
⇣ ⌘
⇤
with W̃ = R̃ R0 a⇤ + R0 W0 FOC gives us a unique a⇤

17 / 27
Portfolio Choice with a Single Risky Asset

18 / 27
The Principle of Participation

Proposition a⇤ > 0 if and only if R > R0

PROOF:
a⇤ > 0 iff f 0h(a)|a=0 > 0 ⇣ ⌘i
iff E U (R0 W0 ) R̃ R0 > 0
0

iff U 0 (R0 W0 ) (R R0 ) > 0
iff R > R0
So nonparticipation in asset markets cannot be explained by appealing to high risk
aversion; we need something like fixed costs of participation, or a kink in the
utility function that generates "first-order risk aversion"

19 / 27
Comparative Statics of a⇤

a⇤ solves h ⇣ ⌘⇣ ⌘i
0 ⇤
E U W̃ R̃ R0 =0

where ⇣ ⌘
⇤
W̃ = R̃ R 0 a ⇤ + R 0 W0

Differentiate with respect to W0 :

 ⇣ ⌘ ✓⇣ ⌘ da⇤ ◆⇣ ⌘
E U 00 W̃ ⇤ R̃ R0 + R0 R̃ R0 =0
dW0

so that h ⌘⇣ ⇣
⌘i
da⇤ E U W̃ 00
R̃ R0 R0 ⇤

=  ⇣ ⌘⇣ ⌘2
dW0
E U 00 W̃ ⇤ R̃ R0

20 / 27
Comparative Statics of a⇤

Denominator is always negative, so

✓ ⇤◆ ⇣ h ⇣ ⌘⇣ ⌘i⌘
da 00 ⇤
sign = sign E U W̃ R̃ R0
dW0
⇣ h ⇣ ⌘ ⇣ ⌘⇣ ⌘i⌘
⇤ 0 ⇤
= sign E A W̃ U W̃ R̃ R0
⇣ ⌘
Under CARA, A W̃ ⇤
= constant, so
h ⇣ ⌘ ⇣ ⌘⇣ ⌘i
⇤ 0 ⇤
E A W̃ U W̃ R̃ R0 =0

da⇤
by first order condition, hence dW0 = 0, so a⇤ is invariant to wealth

21 / 27
Comparative Statics of a⇤

DARA, with R > R0 (so a⇤ > 0)

⇣ ⌘
⇤ ⇤
If R̃ > R0 , then W̃ > R0 W0 , A W̃ < A (R0 W0 )
⇣ ⌘
⇤ ⇤
If R̃ < R0 , then W̃ < R0 W0 , A W̃ > A (R0 W0 )

In either case, we have

⇣ ⌘ ⇣ ⌘⇣ ⌘ ⇣ ⌘⇣ ⌘
⇤ 0 ⇤ 0 ⇤
A W̃ U W̃ R̃ R0 < A (R0 W0 ) U W̃ R̃ R0
so
h ⇣ ⌘ ⇣ ⌘⇣ ⌘i h ⌘⇣ ⇣ ⌘i
⇤ 0 ⇤ 0 ⇤
E A W̃ U W̃ R̃ R0 < E A (R0 W0 ) U W̃ R̃ R0
h ⇣ ⌘⇣ ⌘i
0 ⇤
= A (R0 W0 ) E U W̃ R̃ R0
=0
da⇤
and dW0 > 0. A similar argument (with signs reversed) works for IARA
22 / 27
Risk Aversion and Portfolio Choice

Proposition If R > R0 , then a change in the initial level of wealth has the
following effect on the optimal amount invested in the risky asset:
8
> 0 for DARA;
da⇤ <
= 0 for CARA;
dW0 :
< 0 for IARA

Only DARA seems plausible: Bill Gates probably has more money invested in risky
assets than you do

23 / 27
Comparing Risks

While (under the stated assumptions) we can use a utility index or a utility
function to compare risky prospects (consumption plans or lotteries), in some
cases it is possible to rank different risks directly according to their distribution,
with minimal knowledge of an individual’s utility function.

We will now describe lotteries by means of their distribution functions. Let two
lotteries be represented by their respective cumulative distribution functions, say F
and G , defined on an interval [a, b] 2 Z . Assume that an individual’s utility
function u is weakly increasing. We will show that under these conditions
when a lottery F dominates G in the sense of first-order stochastic
dominance, individuals always prefer F to G as longs as U is weakly
increasing.
when a lottery F dominates G in the sense of second-order stochastic
dominance, individuals always prefer F to G as longs as she is risk-averse
and u is weakly increasing.
24 / 27
First-order Stochastic Dominance

For any lotteries F and G , F first order stochastically dominates G if and

only if
F (W )  G (W )
for all W .

In words: For any wealth or consumption level W , the probability of ending

up with an outcome less than W is lower under F than under G .

25 / 27
Second-order Stochastic Dominance

Let us now assume that F and G have the same mean, so neither dominates in
the sense of first-order stochastic dominance. However, intuitively, for a
risk-averse individual one lottery will still dominate the other if it ’involves more
risk’. Second-order stochastic dominance formalizes this intuition.
For any lotteries F and G with the same mean, F second-order
stochastically dominates G if and only if an individual weakly prefers F to
G under every weakly increasing and concave utility function U.
For any lotteries F and G , G is a mean-preserving spread of F if and only if

Y = Z + "˜

for some Z F and Y G and "˜ such that E [˜

"|Z ] = 0.

In words: For any realization Z drawn from F , realizations drawn from G

can be obtained by adding noise to Z , thus adding risk without improving its
expectation.
26 / 27
Second-order Stochastic Dominance

For convenience, we will now state the main theorem concerning second-order
stochastic dominance using continuous distributions.
R R
Theorem: Assume that zdG = zdF . Then the following are equivalent
R R
I (1) u(z)dF (z) u(z)dG (z) for every weakly increasing concave utility
function.
I (2) G is a mean preserving
Rt spread Rof F .
t
I (3) For every t 0, a G (z)dz a
F (z)dz.

27 / 27
Modern Portfolio Theory

Lorenzo Bretscher

1 / 36
Motivation

Take the perspective of an individual investor: How should I invest my

money? Which assets should I buy, and in what proportions?
Markowitz’s Modern Portfolio Theory - published 1952, Nobel prize 1990

2 / 36
Setup

We assume that an investor only cares about the mean and variance of
overall portfolio return, Rp : for any given mean return the investor prefers to
minimize variance
We assume that the investor knows the mean return of each individual asset,
R̄i , i = 1, . . . , N, and its variance and covariance with all other assets
0
Notation: we stack the mean returns into a vector R = R̄1 , . . . , R̄N , and
form the variance covariance matrix ⌃
The entry in row i, column j of ⌃ is the covariance of Ri with Rj

3 / 36
MPT - investor preferences

Investors are risk averse:

for the same expected return, prefer less risk
for the same risk, prefer more return

4 / 36
Mean-variance analysis with two risky assets

Two risky assets with returns R1 and R2

Choose portfolio share wi in asset i, where w1 + w2 = 1
The portfolio return is
R p = w1 R 1 + w 2 R 2
The mean portfolio return is

R̄p = w1 R̄1 + w2 R̄2

The variance of the portfolio return is

2
p = var [w1 R1 + w2 R2 ]
= w12 var R1 + 2w1 w2 cov (R1 , R2 ) + w22 var R2
= w12 2
1 + 2w1 w2 ⇢ 1 2 + w22 2
2

5 / 36
Mean-variance analysis with two risky assets

With only two assets, we have w2 = 1 w1 , so

R̄p = R̄2 + w1 R̄1 R̄2

Thus specifying a target mean portfolio return pins down

R̄p R̄2
w1 =
R̄1 R̄2
The variance of the portfolio return is
2 2
p = w12 2
1 + 2w1 (1 w1 ) ⇢ 1 2 + (1 w1 ) 2
2

This is a quadratic in w1 , and hence in the mean portfolio return R̄p

Plot mean against variance: parabola
Plot mean against standard deviation: hyperbola
6 / 36
Mean-variance analysis with two risky assets

Mean-s.d. frontiers with R̄1 = 10%, 1 = 15%, R̄2 = 5% and 2 = 6%.

Dashed: ⇢ = 1. Solid: ⇢ = 0. Dotted: ⇢ = 1.
7 / 36
Mean-variance analysis with two risky assets

Consider the case ⇢ = 0, and suppose you start with a portfolio that is fully
invested in the less risky asset: w2 = 1
What happens to the portfolio mean if you shift a very small amount of
wealth into the riskier asset? What happens to portfolio variance?

8 / 36
Mean-variance analysis with two risky assets

If ⇢ = ±1, so the two assets are perfectly positively or negatively correlated,

then we can "complete the square" in the expression
2 2 2 2 2
p = w1 1 + 2w1 w2 ⇢ 1 2 + w2 2
Eg with ⇢ = 1
2 2
p = w12 2
1 2w1 w2 1 2 + w22 2
2 = (w1 1 w2 2)

and we can choose w1 and w2 so that this variance is zero (though an

investor would not necessarily want to do so)

9 / 36
Mean-variance analysis with two risky assets

If 1 < ⇢ < 1, portfolio variance is strictly positive

We also have
d p2 2 2 2
= 2w1 1 2⇢ 1 2 + 2 2 2 ⇢ 1 2
dw1 | {z }
0

We can find the minimum variance portfolio by setting this to zero:

d p2 2
2 ⇢ 1 2
= 0 =) wMV ,1 = 2 2
dw1 1 2⇢ 1 2 + 2

If 1 = 2, then wMV ,1 = 1/2

2 2 2
If ⇢ = 0, then wMV ,1 = 2/ 1 + 2

10 / 36
Mean-variance analysis with two risky assets

If asset 2 is riskless, 2 = 0, then we write R2 = Rf , and we have

R̄p = Rf + w1 R̄1 Rf and p2 = w12 12
If w1 > 0, we have w1 = p / 1, and
✓ ◆
R̄1 Rf
R̄p = Rf + p
1

This defines a line in the mean-standard deviation diagram that is known as

the capital allocation line
Its slope,
R̄1 Rf
S1 =
1

is the Sharpe ratio of the risky asset

11 / 36
Mean-variance analysis with N risky assets

Starting from pairwise combinations of 3 stocks

12 / 36
Mean-variance analysis with N risky assets

All possible combinations of 3 stocks and the efficient frontier

13 / 36
Mean-variance analysis with N risky assets

If you continue to add risky securities:

Efficient frontier improves but at a decreasing rate: decreasing marginal
benefits of diversification
Efficient frontier dominates individual securities

14 / 36
Mean-variance analysis with N risky assets

N risky assets to choose from with returns R1 , R2 , . . . , RN

0
Write R for the vector of mean returns, R̄1 , . . . , R̄N , and ⌃ for the
variance-covariance matrix of returns, which is assumed to be non-singular
0
w = (w1 , . . . , wN ) is the vector of portfolio weights on the N assets
The overall mean portfolio return is

R̄p = w1 R 1 + · · · + wN R N = w0 R

and the variance of the portfolio return is

p = w12 11 + 2w1 w2 12 + · · · + wN2 NN = w0 ⌃w

15 / 36
Mean-variance analysis with N risky assets

When N = 2, the mean portfolio return uniquely identifies the portfolio

weights and hence the portfolio variance
When N > 2, the portfolio choice problem is to solve

minw 12 w0 ⌃w
s.t. w0 R = R p
w0 e = 1

The notation e means the vector of ones, e = (1, . . . , 1)0

Portfolios that solve this problem for some R p are said to lie on the
minimum-variance frontier. Portfolios on the "upper half" of the
minimum-variance frontier are said to be mean-variance efficient

16 / 36
Mean-variance analysis with N risky assets

This is a standard constrained optimization problem

Form the Lagrangian
1 0
L (w , 1, 2) = w ⌃w + 1 R̄p w 0R + 2 (1 w 0 e)
2
The first-order conditions are

⌃w 1R 2e =0

So
1 1
w= 1⌃ R+ 2⌃ e (1)
Don’t be misled by the notation: there are N first-order conditions

17 / 36
Mean-variance analysis with N risky assets

To solve for the Lagrange multipliers 1 and 2, we use the two constraints:
0 1 0 1
Rp = 1R ⌃ R+ 2e ⌃ R= 1A + 2B
0 1 0 1
1= 1e ⌃ R+ 2e ⌃ e= 1B + 2C
0 1 1 1
where A ⌘ R ⌃ R,B ⌘ e 0 ⌃ R and C ⌘ e 0 ⌃ e
These imply that
C Rp B
1 =
D
A BR p
2 =
D
where D ⌘ AC B2
While A, C and D are all mathematically guaranteed to be positive, we
cannot say the same about B
18 / 36
Mean-variance analysis with N risky assets

The optimized variance is

2
C Rp 2BR p + A
w 0 ⌃w = w 0 1R + 2e = 1R p + 2 =
D
which, as before, is quadratic in R p
A, B, C , and D are quantities that can in principle be estimated empirically.
They represent all relevant information about the investment opportunity set
It is up to an individual investor to decide what R p she wants, and hence
what 1 and 2 are

19 / 36
Mean-variance analysis with N risky assets

The portfolio choice problem is a convex optimization problem (because we

have a quadratic objective subject to linear constraints) so equation (1) is
necessary and sufficient for a portfolio to be mean-variance efficient
Another way to see it: Suppose that the portfolio w satisfies

⌃w = 1R + 2e

We aim to show that any other portfolio v with the same mean has a
variance that is at least as large. So, suppose v 0 R = w 0 R
Because w and v are both portfolios, we know that v0 e = w0 e = 1

20 / 36
Mean-variance analysis with N risky assets

Let ⌘ = v w . So, ⌘ 0 R = 0 and ⌘ 0 e = 0

The variance of the candidate portfolio v is

v 0 ⌃v = (w + ⌘)0 ⌃(w + ⌘)
= w 0 ⌃w + 2⌘ 0 ⌃w + ⌘ 0 ⌃⌘
= w 0 ⌃w + 2⌘ 0 1R + 2e + ⌘ 0 ⌃⌘
= w 0 ⌃w + ⌘ 0 ⌃⌘
w 0 ⌃w

So w was indeed mean-variance efficient

21 / 36
Mean-variance analysis with N risky assets

Because we have both necessity and sufficiency, we can say that the
mean-variance efficient portfolios are precisely those portfolios w such that
w = 1 ⌃ 1 R + 2 ⌃ 1 e for some 1 and 2
Moreover, given any two distinct mean-variance efficient portfolios, we can
construct any other mean-variance efficient portfolio by combining these two

22 / 36
The minimum-variance portfolio

An investor who wants to minimize variance at all costs will choose the global
minimum variance portfolio
To find this portfolio, we drop the constraint that w 0 R = R p , and find
⌃ 1e
1
w MV = ⌃ e =
e 0⌃ 1e
The expected return on this portfolio is
e 0⌃ 1R B
R MV = w 0MV R = =
e 0⌃ 1e C
Empirically, the minimum-variance portfolio has a positive expected return,
so this suggests that B > 0.
The variance that this portfolio achieves is
1 1
e 0⌃ ⌃⌃ e 1
w 0MV ⌃w MV = 2
=
e 0⌃ 1
e e 0⌃ 1e
23 / 36
The global minimum-variance portfolio

If all assets have equal variance 2 and correlation ⇢ between any two assets,
then the minimum variance portfolio puts equal weight on each asset:
w MV = e/N
2 2 2
Then, w 0MV ⌃w MV = ⇢ + (1 ⇢) /N ⇡ ⇢ for large N
The GMV portfolio has the same covariance with every portfolio:

0 w 0e 1
w ⌃w MV = 1
= 1
>0
e⌃ e
0 e⌃ e
0

Intuition: Suppose you had a portfolio w 1 and w 2 that had different

covariances with w MV . Can you find a portfolio with lower variance than
w MV ?
In particular, the global minimum-variance portfolio is positively correlated
with every asset

24 / 36
A (two) mutual fund theorem

Another natural portfolio to look at is w

e,

⌃ 1R
w
e ⌘
e 0⌃ 1R
We can use this portfolio together with the minimum variance portfolio to
rewrite (1) as
0 1 0 1
w= 1e ⌃ Rw
e+ 2e ⌃ ew MV = 1B w
e + 2 C w MV

Mutual fund theorem: any portfolio on the minimum-variance frontier can be

expressed as a linear combination of w
e and w MV
A mean-variance investor would be equally happy whether
I choosing between the menu of N assets, or
I choosing between just two mutual funds, one of which holds portfolio w
e and
the other of which holds portfolio w MV
25 / 36
Adding a riskless asset

Let the portfolio weights on the risky assets be w1 , . . . , wN as before, and let
the weight on the riskless asset be w0
We want to solve
minw0 ,...,wN 12 w 0 ⌃w
s.t. w0 Rf + w 0 R = R p
w0 + w 0 e = 1
We can use the second constraint to substitute out for w0 . Then the problem
becomes
minw 12 w 0 ⌃w
s.t. w 0 R Rf e = R̄p Rf

26 / 36
Adding a riskless asset

The first-order condition for this problem is

⌃w = R Rf e

So the optimal choice of risky assets is

1
w= ⌃ R Rf e

chosen so that the constraint w 0 R Rf e = R p Rf is satisfied:

Rp Rf
=
E
0 1
where E ⌘ R Rf e ⌃ R Rf e

27 / 36
Adding a riskless asset

With this choice of , the portfolio variance is

2
0 1 1 2 R̄p Rf
w 0 ⌃w = R Rf e ⌃ ⌃ ⌃ R Rf e = E=
E
The portfolio standard deviation satisfies
p
R̄p Rf = p E

This is a straight line on a plot of mean against standard deviation

The line is tangent to the minimum-variance frontier achievable by trading
only in risky assets
The slope of the line is the Sharpe ratio of the tangency portfolio, which has
the highest Sharpe ratio of any portfolio of risky assets

28 / 36
Efficient frontier

29 / 36
Adding a riskless asset

Similarly to the previous section, this says that mean-variance investors would
be indifferent between trading all N + 1 assets, and trading just the riskless
asset and tangency portfolio. This is sometimes called Tobin’s mutual fund
theorem
This does not imply that all investors should hold the same portfolio.
Extremely risk-averse investors may want almost all of their wealth in the
riskless asset. Conversely, investors with high risk tolerance may want to lever
up, shorting the riskless asset in order to buy even more of the tangency
portfolio

30 / 36
Covariance properties of efficient portfolios

Consider a mean-variance efficient portfolio p, and suppose we are

considering increasing slightly the weight wi on one of the assets, i, financed
by a decrease in the weight on the riskless asset
This will affect the mean and variance of the portfolio return:

d R̄p
= R̄i Rf
dwi
dvarRp
= 2 cov (Ri , Rp )
dwi
because the terms in var Rp that involve wi are

2w1 wi cov (R1 , Ri ) + · · · + wi2 var Ri + · · · + 2wN wi cov (RN , Ri )

31 / 36
Covariance properties of efficient portfolios

The ratio of the effects on mean and on variance is

d R̄p /dwi R̄i Rf
=
d var Rp /dwi 2 cov (Rp , Ri )

If portfolio p is efficient, this ratio should be the same for all assets.
Why? Consider adjusting two different portfolio weights wi and wj (financed
in each case by changes in holdings of the riskless asset). The effects on
mean and variance of Rp are

d R̄p = R̄i Rf dwi + R̄j Rf dwj

d var Rp = 2 cov (Rp , Ri ) dwi + 2 cov (Rp , Rj ) dwj

32 / 36
Covariance properties of efficient portfolios

We can choose dwj so that the mean portfolio return d R̄p = 0 :

R̄i Rf
dwj = dwi
R̄j Rf

If we do so, the portfolio variance must also be unchanged

Otherwise we could achieve a lower variance with the same mean,
contradicting the assumption that p is efficient. So we must have

R̄i Rf
d var Rp = 2 cov (Rp , Ri ) 2 cov (Rp , Rj ) =0
R̄j Rf

33 / 36
Covariance properties of efficient portfolios

This implies that

R i Rf R j Rf
=
cov (Rp , Ri ) cov (Rp , Rj )
for any assets i and j
This must also hold, therefore, for the portfolio p itself:

R i Rf R p Rf
=
cov (Rp , Ri ) var Rp

Rearranging,

cov (Ri , Rp )
Ri Rf = Rp Rf = ip Rp Rf
var Rp

34 / 36
Issues

Of course, all this assumes that we know the variance-covariance matrix ⌃

and the vector of mean returns R
In practice, we have to estimate them from the data, which is not a trivial
exercise ! data requirements grow very quickly, especially number of
pairwise correlations

35 / 36
Issues

In general, for N securities, need to estimate:

N expected returns, N variances
2
BUT: N 2 N pairwise correlations (equivalently covariances)
That is, 4,950 correlation values if N = 100

This motivates a search for ways of identifying optimal portfolios that are based
on further theoretical considerations
36 / 36
The Capital Asset Pricing Model

Lorenzo Bretscher

1 / 30
Motivation

What if everyone behaves like a mean-variance investor? What predictions

can we make for the bahavior of the returns on different assets? What
happens in equilibrium?
CAPM was developed by William Sharpe - published 1964, Nobel prize 1990
How do preferences relate to the mean-variance portfolio choice problem?
How general is mean-variance?
! When will an expected utility maximizer have mean-variance preferences?

2 / 30
The Capital Asset Pricing Model (CAPM)

Assumptions:
1 All investors are mean-variance optimizers
2 All investors perceive the same means, variances, and covariances
3 All investors can borrow or lend at a given riskfree interest rate
4 All risky assets are in positive supply (i.e., the total market value of every
asset is positive) and the riskless asset is in zero net supply (i.e., the total
amount of borrowing equals the total amount of lending)

3 / 30
The Capital Asset Pricing Model (CAPM)

Assumptions (i)-(iii) imply that all investors work with the same
mean-standard-deviation diagram; every investor is solving the same
mean-variance portfolio choice problem
So every investor holds a mean-variance efficient portfolio
Since all mean-variance efficient portfolios combine the riskless asset with a
fixed portfolio of risky assets, we can conclude that all investors hold risky
assets in the same proportions (namely, the proportions of the tangency
portfolio, w t )
Thus, investor i has risky asset portfolio shares w (i) = (i)
wt
I The entries of w t add up to one, but w (i) need not add up to one because of
the presence of the riskless asset. (If you allocate 80% of your wealth, in total,
to the risky assets, then you can put the remaining 20% in the riskless asset.)

4 / 30
The Capital Asset Pricing Model (CAPM)

Now suppose we add up everyone’s risky asset holdings, weighted by their

wealth (where W (i) is investor i’s wealth and N is the number of investors in
the economy)
This is a vector, and the top entry of the vector is the amount of cash
invested in asset 1, which must equal the total market value of asset 1 (and
so on):
W (1) w (1) + W (2) w (2) + · · · + W (N) w (N) = W w m
where W = W (1) + · · · + W (N) is aggregate wealth and wm is the market
portfolio, i.e. the value-weighted index that contains all risky assets in
proportion to their market values

5 / 30
The Capital Asset Pricing Model (CAPM)

Written in terms of the tangency portfolio, this equation becomes

⇣ ⌘
(1) (1) (2) (2) (N) (N)
W +W + ··· + W wt = W wm

Pre-multiplying by e 0 , and remembering that e 0 w t = e 0 w m = 1 (why?), we

find that W (1) (1) + W (2) (2) + · · · + W (N) (N) = W , and hence that

wt = wm

Thus the market portfolio is the tangency portfolio and, in particular,

is mean-variance efficient
This is the content of the CAPM

6 / 30
The Capital Asset Pricing Model (CAPM)

As a result, we can revisit the capital allocation line (CAL) equation

cov (Ri , Rp )
Ri Rf = Rp Rf = ip Rp Rf ,
var Rp

using the market portfolio as an example of a mean-variance efficient

portfolio (i.e., substituting m for p ):

cov (Ri , Rm )
Ri Rf = Rm Rf = im Rm Rf (1)
var Rm
The CAPM beta, im = cov (Ri , Rm ) / var Rm , is the coefficient on Rm in a
regression of Ri on Rm

7 / 30
The Capital Asset Pricing Model (CAPM)

R i = Rf + im Rm Rf
What it does/says:
gives the relationship between expected return and risk (as measured by )
that holds for all individual securities and portfolios of securities
the expected return on any security i is equal to:
1 the risk-free rate plus
2 a risk premium that is equal to the security’s beta multiplied by the market
risk premium (excess expected return)

8 / 30
The Capital Market Line (CML)

Capital market line is the capital allocation line for the market portfolio!

9 / 30
Security Market Line

The security market line (SML) is the plot of the linear relationship between
expected return and beta implied by the CAPM

10 / 30
The Capital Asset Pricing Model (CAPM)

Since the market portfolio is mean-variance efficient, a mean-variance

investor need not even perform the mean-variance analysis!
The investor just needs to decide how much wealth to put in the riskless asset
and how much to put in the market portfolio (in practice, a broad index fund)
Other implications of (1):
I Capital budgeting (what discount rate to use in evaluating investment projects)
I Performance evaluation (how large a return should one expect given the risk
that a fund manager is taking?)

11 / 30
The CAPM with no Riskless Asset

In the absence of a riskless asset, it is still possible to derive an equation

analogous to (1)
The same logic as before implies that the market portfolio is mean-variance
efficient. For investor k ’s portfolio is mean-variance efficient, and so satisfies
(k) (k) (k) (k)
the criterion ⌃w (k) = 1 R + 2 e for some 1 and 2
Summing over all investors k in proportion to their wealth, we get
⌃w m = R + e, so the market portfolio is mean-variance efficient

12 / 30
The CAPM with no Riskless Asset

Turning this expression round,

1
R+ e= ⌃w m (2)

This, like (1), is a statement about expected returns

Notice that
0 1
X X
(⌃w m )i = @
cov (Ri , Rj ) wm,j = cov Ri , wm,j Rj A = cov (Ri , Rm )
j j

To get rid of the Lagrange multipliers and , we need a portfolio that plays
the role of the riskless asset in the previous analysis
Natural guess: the minimum-variance portfolio? Wrong!
The right notion is that of the zero-beta portfolio
13 / 30
The CAPM with no Riskless Asset

The zero-beta portfolio is the minimum-variance portfolio that is uncorrelated

with the market:
minwz 12 w 0z ⌃w z
s.t. w 0z ⌃w m = 0
w 0z e = 1

14 / 30
The CAPM with no Riskless Asset

The important feature is that because w z is uncorrelated with w m we can

premultiply equation (2) by w 0z , to get

w 0z R + =0

So,
1
R Rz e = ⌃w m

Pinning down any expected excess return will get rid of the constant of
proportionality 1/ . It is natural to choose the expected excess return on the
market

15 / 30
The CAPM with no Riskless Asset

The above equation implies (after premultiplying by w m t) that

1
Rm Rz = var Rm

Using this to eliminate the 1/ , we have

⌃w m
R Rz e = Rm Rz
var Rm
or equivalently
Ri Rz = im Rm Rz (3)
where im = cov (Ri , Rm ) / var Rm

16 / 30
More on Zero Beta Portfolios

The property that w z is uncorrelated with w m is independent of the

existence or nonexistence of a riskless asset
If there were a riskless asset, both the standard CAPM and equation (3)
would hold. Comparing the two, we see that R z = Rf
This provides a way to find the location of the zero-beta portfolio on the
risky-asset efficient frontier: draw in the tangency line, and check where it
cuts the y -axis. This gives the mean return on the zero-beta portfolio, and
indicates where on the efficient frontier the portfolio lies

17 / 30
More on Zero Beta Portfolios

More generally, given an arbitrary portfolio w p on the risky-asset efficient

(p)
frontier, we can define w z to be the portfolio of risky assets that has
minimum variance, subject to being uncorrelated with w p
(p)
The above argument implies that the expected return on w z can be
calculated by seeing where the tangency line associated with w p cuts the
y -axis
(p)
The portfolio w z is the unique portfolio on the efficient frontier with this
expected return
With one exception: there is one risky portfolio that does not have an
associated zero-beta portfolio. Which is it, and why?

18 / 30
What Have We Done So Far?

The mean-variance analysis delivered two results: first, the mutual fund
theorem, which reduced the dimensionality from N assets to two mutual
funds (and if a riskless asset exists, we can identify one of these as the
riskless asset, and the other as the tangency portfolio); second, the
characterization of expected returns
Imposing equilibrium lets us identify a particular efficient risky-asset
portfolio: the market portfolio
Perhaps one of the most concrete successes of this framework is the fact that
index funds, which date from the 1970s and were inspired by these results,
are now so prevalent

19 / 30
The General One-Period Portfolio Problem

In general, an expected utility maximizing investor does not necessarily

choose a mean-variance efficient portfolio
N risky securities, gross returns R̃n (n = 1, 2, . . . , N)
wn fraction of wealth invested in security n
Portfolio selection problem:
" N
!# N
X X
max E U W0 wn R̃n s.t. wn = 1
(w1 ,...wN )2RN
n=1 n=1

20 / 30
The General One-Period Portfolio Problem

Lagrangian: !
N
X
L = E[U] + W0 1 wn
n=1

First order conditions for optimal portfolio weights:

h ⇣ ⌘ i
0 ⇤
E U W̃ R̃n = (n = 1, . . . , N)

where
N
X
W̃ ⇤ = W0 wn⇤ R̃n
n=1

Expected marginal utility of investing is the same for each asset

21 / 30
The General One-Period Portfolio Problem

Multiply each condition by corresponding wn⇤ , and add up:

h ⇣ ⌘ i N
X
E U 0 W̃ ⇤ R̃ ⇤ = where R̃ ⇤ = wn⇤ R̃n
n=1

measures marginal indirect expected utility from an additional unit of initial

wealth: h ⇣ ⌘i
d ⇤
= E U W0 R̃
dW0

22 / 30
The General One-Period Portfolio Problem

Now introduce security 0 with risk-free gross return R0

Allow unlimited borrowing and lending
Portfolio problem:
" ( N
)!# N
X X
max E U W0 w0 R 0 + wn R̃n s.t. wn = 1
(w0 ,...wN )2RN+1
n=1 n=0

Additional first order condition:

h ⇣ ⌘ i
0 ⇤
E U W̃ R0 =

where ( )
N
X
W̃ ⇤ = W0 R̃ ⇤ = W0 w0⇤ R0 + wn⇤ R̃n
n=1

23 / 30
The General One-Period Portfolio Problem

Eliminate :
h ⇣ ⌘ i h ⇣ ⌘ i
0 ⇤ 0 ⇤
E U W̃ R̃n = E U W̃ R0 (n = 1, . . . , N)

In particular, by the same argument as above (multiply by wn⇤ and sum

over n ): h ⇣ ⌘ i h ⇣ ⌘ i
0 ⇤ ⇤ 0 ⇤
E U W̃ R̃ = E U W̃ R0

24 / 30
The General One-Period Portfolio Problem

The key is that for any two assets i and j

h ⇣ ⌘ i h ⇣ ⌘ i
0 ⇤ 0 ⇤
E U W̃ R̃i = E U W̃ R̃j

Intuition: suppose I gave you an extra £ to invest ( very small)

Put it in asset i : earnh random extra
⇣ amount
⌘i £ R̃i ; change in your
expected utility is E R̃i ⇥ U 0 W̃ ⇤ , to first order in
Put it in asset j : earnh random extra
⇣ amount
⌘i £ R̃j ; change in your
expected utility is E R̃j ⇥ U 0 W̃ ⇤ , to first order in
If you have chosen your portfolio optimally, you should be indifferent about
which asset you put the marginal pounds in

25 / 30
The General One-Period Portfolio Problem

We’ve now done the analogue of the mean-variance portfolio choice problem
h ⇣ ⌘⇣ ⌘i
The first-order conditions (in any form, eg E U W̃
0 ⇤
R̃n R0 = 0 )
are often known as the Euler equation(s))
They are a restriction on the demand for assets that must hold in any
equilibrium model
Fully specified equilibrium models make further assumptions about asset
supply and about U(·)

26 / 30
Mean-Variance Analysis vs Expected Utility

Suppose vector of risky asset returns is normally distributed

Return R̃ on any portfolio (with or without the riskless asset) also has a
normal distribution
Consider a non-satiable risk-averse investor with utility function U [so
U 0 > 0, U 00 < 0] and initial wealth W0
2
Portfolio with return R̃ ⇠ N R, yields wealth

W̃ = W0 R̃ = W0 (R + Z̃ )

where Z̃ ⇠ N(0, 1)

27 / 30
Mean-Variance Analysis vs Expected Utility

Density function of Z̃ :
✓ ◆
1 z2
(z) = p exp
2⇡ 2
Expected utility:
Z 1
2
E[U(W̃ )] = U (W0 [R + z]) (z)dz = V R,
1

Differentiate:
Z 1
@V
= W0 U 0 (W0 [R + z]) (z)dz > 0
@R 1
Z 1
@V 1 @
= U (W0 [R + z]) (z)dz
@ 2 2 @ 1
Z 1
W0
= U 0 (W0 [R + z]) z (z)dz
2 1
<0
28 / 30
Mean-Variance Analysis vs Expected Utility

Alternatively, suppose the investor’s utility function is quadratic

Without loss of generality,

aW 2
U(W ) = W
2
1
with a > 0, W < a (so that U 0 > 0 and U 00 < 0)
Consider initial wealth W0 and a portfolio with return R̃, mean R, and
variance 2

29 / 30
Mean-Variance Analysis vs Expected Utility

Expected utility:
h ⇣ ⌘i
E[U(W̃ )] = E U W0 R̃
aW02 h 2 i
= W0 E[R̃] E R̃
2
aW02 2
= W0 R + R2
2
Higher order moments are irrelevant
1 2
Expected utility increases in R (if W0 R = E(W̃ ) < a ) and decreases in
Problems: 1) all concave quadratic functions are decreasing after a certain
point (the ’bliss point’), which is incompatible with non-satiation and 2)
quadratic utility implies IARA

30 / 30
Arbitrage Pricing Theory

Lorenzo Bretscher

1 / 18
Reminder: Arbitrage

An asset with price equal to zero and (PV of) E[future CF] > 0!

An asset with a price different from zero and all future CF are equal to zero.

2 / 18
Factor Pricing and Arbitrage Pricing Theory

The central prediction of the CAPM is that

ERie = e
im Rm

where im is asset i’s market beta, cov (Ri , Rm ) / var Rm

We can also run the regression
e e
Ri,t = ↵i + im Rm,t + "it (1)

as a purely statistical exercise

Suppose we did so, and suppose that we found the errors "it were
uncorrelated across stocks: E ["it "jt ] = 0 for all i 6= j and t
2
Covariances would be easy to estimate because cov (Rit , Rjt ) = im so
jm m ,
we only have to estimate N betas with the market rather than N 2 N /2
covariances
If N is large, we should expect ↵i typically to be very small
3 / 18
Factor Pricing and Arbitrage Pricing Theory

A portfolio of N assets with weights wj will have the return

e e
Rpt = ↵p + pm Rmt + "pt
PN PN PN
↵p = j=1 wj ↵j , pm = j=1 wj jm , and "pt = j=1 wj "jt
The variance of the portfolio residual error is
N
X
var "pt = wj2 var "jt
j=1

4 / 18
Factor Pricing and Arbitrage Pricing Theory

This variance is very small for large N. Suppose that the maximum
idiosyncratic variance of any asset j is 2 , and that we choose to equally
weight the portfolio (wj = 1/N). Then
2
var "pt 
N
As N ! 1 this goes to zero: the portfolio is well diversified
So we can ignore "pt , and we have
e e
Rpt = ↵p + pm Rmt

But then we must have ↵p = 0 : If ↵p > 0, we could go long portfolio p and

short the market, earning ↵p without taking any risk or putting up any
money up front
5 / 18
Factor Pricing and Arbitrage Pricing Theory

This is the arbitrage pricing theory (APT) of Ross (1976): if excess returns
can be explained by a market factor Rm,t
e
as in (1), with uncorrelated
residuals, then
e e
ERi,t = im ERm,t
We have derived the beta pricing equation of the CAPM without using any of
the apparatus of mean-variance analysis
The key assumption is that the residual risks "it are uncorrelated (still!)

6 / 18
Anectodal Empirical Evidence

CAPM (or any one-factor model):

Ri Rf = ↵i + im [Rm R f ] + "i

Is it realistic to assume that "˜i ’s are independent?

7 / 18
Advance Micro Devices vs. Intel

8 / 18
AMD vs. INTC
Do CAPM regression for AMD/INTC
AMD: = 3.8%, SPY = 1.42
INTC: = 2.0%, SPY = 1.02
Correlation between "AM and "INTC is 0.2

Maybe, indeed a better model of risk and return is needed!

9 / 18
Factor Pricing and Arbitrage Pricing Theory

The "market" can be any broadly diversified portfolio that produces

uncorrelated residual risk
In practice, no single factor accounts for all the correlations between asset
returns
I Industry effects
I Macroeconomic variables
I Small firms seem to move together

10 / 18
Factor Pricing and Arbitrage Pricing Theory

If there are K portfolios capturing the influence of K underlying common

sources of risk, such that the regression
K
X
Rite = ↵i + e
ik Rkt + "it
k=1

delivers uncorrelated residuals, then the prediction of the model is that ↵i

should be very close to zero for almost all stocks
We want K ⌧ N why?

11 / 18
Factor Pricing and Arbitrage Pricing Theory

Ways to pick factors for multifactor models:

I The first factor is normally an excess return on a broad market index
I You can do principal component analysis to find the most important common
factors. Unfortunately, you tend to find more and more factors as you increase
the number of assets
I Macro factors (inflation, interest rates, industrial production)
I Portfolios of stocks with common characteristics (size, book-to-market,
momentum). This leads to a data-snooping problem if the characteristics are
chosen to produce high average returns in sample
I Factors suggested by equilibrium models, for example innovations in variables
that predict market returns

12 / 18
Factor Pricing and Arbitrage Pricing Theory

The strength of the APT is the minimality of its assumptions

But this is also a weakness
I Some portfolio is always ex post mean-variance efficient. So we can always
find some 1 -factor model that fits the data. What does this tell us about the
world?
I The theory does not determine the signs or magnitudes of risk prices. Much
recent work on general equilibrium asset pricing seeks to pin down risk prices
more precisely based on theoretical considerations

13 / 18
Factors are Useful for Hedging

Consider Barrick Gold (ABX), a gold mining firm

Assume a factor model such as the following, where the factors are the
market (SPY) and the price of gold (GLD)

r˜abx rf = ↵abx + SPY [˜

rSPY rf ] + GLD [˜
rGLD rf ] + "˜abx

If you form a portfolio with the same SPY and GLD , then you track
Barrick’s exposure to these common factors. It’s like replication with factors.
You can track exposure to both factors or exposure to just one (e.g., GLD)

14 / 18
Factors are Useful for Hedging

You want to buy Barrick Gold (ABX), but you want to hedge out the exposure to
the price of gold

15 / 18
Factors are Useful for Hedging

Regress returns of ABX on two factors: SPY and GLD (a gold ETF)

r˜abx rf = 0.00 + 0.45 ⇥ [˜

rSPY rf ] + 1.62 ⇥ [˜
rGLD rf ] + "˜abx

The portfolio that tracks the GLD exposure has GLD = 1.62:
I wGLD = 1.62 and wrF = 0.62
I r˜track = wGLD r˜GLD + wrf rf = 1.62 ⇥ r˜GLD + ( 0.62) ⇥ rf = rf + 1.62 ⇥ [˜
rGLD rf ]
The tracking portfolio has GLD = 1.62
So to hedge out the GLD exposure, short $ 1.62 of GLD for every $ 1
invested in ABX and invest $ 0.62 in the risk-free bond

16 / 18
Factors are Useful for Hedging

Now imagine a portfolio that tracks BOTH SPY and GLD exposure
I wGLD = 1.62, wSPY = 0.45and wrF = 1 (0.62 + 0.45) = 1.07
What is the return of a portfolio long Barrick Gold and short the tracking
portfolio?
r˜hedge = r˜abx r˜track = ↵abx + "˜abx
Let’s assume " = 0. In reality, they are not, but let’s assume this for a
moment ! r˜hedge = ↵abx + "˜abx = ↵abx
If epsilons are always zero, then we have a zero cost portfolio with certain
returns ! what should the return be?

↵abx = 0

17 / 18
Factors are Useful for Hedging

If we have a factor model, we can create a tracking portfolio

If there is no idiosyncratic risk, then we’ve replicated the security and we can
write down the return equation:

r˜abx = rf + 0.45 ⇥ [˜
rSPY rf ] + 1.62 ⇥ [˜
rGLD rf ]

This is the idea behind the Arbitrage Pricing Theory (APT)!

18 / 18