ECON0113 - Advanced Economics of Finance

Lecture Notes 0

Course Overview, Notation and Choice Under Uncertainty

Rodrigo Guimarães

September 25, 2020

1 Course overview

As one famous graduate textbook1 in empirical finance has put it:

”The starting point of every financial model is the uncertainty facing investors, and the sub-

stance of every financial model involves the impact of uncertainty on the behavior of investors

and, ultimately, on market prices”

Finance can be roughly split into two sub-disciplines: Asset Pricing and Corporate Finance. The focus

of this course is on asset pricing, which is the part of finance that looks at how an asset (essentially the

right to some future payoff: equity, bonds, derivatives, etc.) is valued given it’s cash flow properties. We

will only briefly discuss how the insights we have gained from asset pricing matter for Corporate Finance,

which looks at how choices (such as financing structure, timing of investment, project selection, project

execution, etc.) affect, and are affected by, valuation concerns. We will therefore not cover the wide range

of strategic issues that are central to Corporate Finance and arise from informational asymmetries, agency,

taxes and market structure among many others.

These notes cover the basic ingredient needed to study risk: how to deal with choice under uncertainty.

Our focus will be on the central role played by our preferences over (or distaste for) uncertainty, which

determines how much we value uncertain propositions. The very basic element of asset pricing, an ad-

justment to the value we attach to an uncertain outcome because of our preferences, will make it’s first

appearance. We will refer to it as risk-adjustment. We will then spend most of the course understanding
1 J. Campbell, A. Lo & A. MacKinlay, ‘The Econometrics of Financial Markets’, Princeton University Press.

1
how the risk adjustment is determined in equilibrium, how it appears in different problems, and how asset

pricing can be used in a few applications.

In the next lecture will first start from a very simple consumption choice problem familiar from microe-

conomics, focus on what it says about asset prices and slowly build up to the general result underlying all of

asset pricing: there exists a stochastic discount factor that allows us to price any asset. This is what could

be referred to as the fundamental theorem of asset pricing.2 This stochastic discount factor takes care of

the two aspects that determine the price of any asset, no matter how simple or fiendishly complicated: the

time to receive it’s payoffs (or delay) and the uncertainty attached to these payoffs.

We will then pause to think about what assumptions were needed to arrive at these general results and

what is the economic intuition behind them. The rest of the course can be thought of as simply expanding

or unpacking different aspects of the general result, sometimes adding more elements, sometimes just

narrowing in on one aspect of the problem to study it in more detail. We will also see how it shows up in

very distinct forms that are seemingly unrelated. Although there will be no explicit discussion of how to

estimate these models, we will also discuss real world applications and main empirical questions that have

motivated the development of new theories.

One way to see the importance (or not) of each assumption is to work out the full implications of some

particular assumptions and what happens when you remove some of them. In doing so we will cover some

well known models that historically were very useful in building our understanding of asset pricing, such

as CAPM, mean-variance portfolio selection and WACC.

The benefit of the approach adopted here is that it should allow you to clearly see the forest from the

trees: what is the same underlying economics in every asset pricing problem. You should then be able to (i)

see the relation to other areas of economics to allow you to have intelligent conversations where finance might

provide insight but is not the only dimension (if your interests are more in macro/labour/development/trade

or any other field of economics.); (ii) identify what are general results and what depends on particular

assumptions, which is also useful to distinguish the fundamental questions that need answers (because we

do have many open questions!) and what are ”red herrings”; (iii) use it as a good foundation to specialize

in any particular aspect of asset pricing or finance. In addition, at the end of the course you should be able

to read a general article on asset pricing, such as survey papers, advanced textbooks or handbook chapters3

and understand broadly the content as the core should be familiar even if the level of mathematics is much
2 There is a fundamental theorem that states the conditions under which a stochastic discount factor exists, which is

essentially the absence of arbitrage opportunities (you cannot get something for nothing). Though the general proof requires
incredibly powerful mathematical tools that are very much beyond the scope of this course, it boils down to assuming there
are no free lunches. So you can imagine it is indeed very general!
3 See for example John Campbell’s chapter in the Handbook of the Economics of Finance, vol 1B, or Sydney Ludvigson’s

chapter (available at https://www.sydneyludvigson.com/s/hbe.pdf) in volume 2B of the same Handbook. John Cochrane

also has several overview and synthesis papers. You should be able to read these after completing this course.

2
more advanced.

The general structure, and much of the material in the lecture notes, draws very heavily from John

Cochrane’s graduate textbook ”Asset Pricing”. However it is not required reading because it is written for

a more advanced level than this course (but if you are comfortable with more mathematics and have a good

grasp of general equilibrium you might find it useful). Unfortunately most undergraduate level textbooks

still follow a ‘historical discovery’ approach, presenting models as they evolved over time, which can make

it hard to see the underlying common thread and what are the relevant predictions that are robust to

particular assumptions. It also means that an undergraduate level they never get past the models that

were the frontier in the 1970s, such as the CAPM. The lecture notes are a translation of Cochrane’s book

for this audience. Since there doesn’t exist a textbook that does this, it is extremely important you attend

lectures. You are responsible for covering all the material in the lecture notes, unless stated otherwise,

even if they are not covered in lectures.

What is assumed?4

• Statistics: You are comfortable with basic statistical concepts such as expectations, conditional

expectations and covariances;

• Mathematics: You are comfortable setting up and solving a constrained maximization problem (La-

grangian multipliers), though most of the time we will not deal with Lagrange multipliers;

• Microeconomics: You are familiar with utility functions, marginal utility, budget constraints, general

equilibrium and welfare concepts.

1.1 Notation convention

Throughout I will do my best to be explicit about names of variables, functions and parameters, and will

try to follow the following general convention (you don’t need to memorize these, but in doubt refer back

to this list):

• Greek letters will be reserved for parameters

• in particular µx , σx , σxy , ρxy will denote the mean of x, standard deviation of x, covariance of x, y,

and correlation between x, y, respectively

 
∂f (x)
• When no ambiguity is possible I will use f 0 to denote the partial derivative of f f0 = ∂x and
 2

f 00 the second partial derivative f 00 = ∂ ∂x
f (x)
2

4 See the Math and Statistics refresher note in the course’s Moodle page.

3
 • u will be reserved for utility functions; c for consumption, π for probabilities, w for wealth, and p for

prices.

• Et [] is the expectation relative to information at time t, and random variables will have subscripts

denoting in which period they become known, so xt+k is a known quantity at time t + k (or later,

there is no memory loss in these models, i.e. if xt+k is known at time t + k then it remains known at

any time after t + k) and hence



x

t+k for k ≤ 0
Et [xt+k ] =

Et [xt+k ]
 for k > 0

• capital R is a gross return, that is R = 1 + r or er if (using continuous compounded rates) where r

is the net return (like the interest rate in your savings account or mortgage)

• when referring to same variable for different agents/assets I will add subscripts (e.g. Ri,t is the Rt

for asset i, ui the utility function of agent i)

2 Choice Under Uncertainty: Expected Utility

Asset pricing can be succinctly summarized as the attempt to understand the effect of time and risk

(arising from uncertainty5 ) of cash flows on the equilibrium price of the right to those cash flows. The

really interesting economics happens on the risk dimension, and we will spend a considerable amount of

time trying to understand what risk is, how we deal with it and how it impacts choices and prices.

The objective of this lecture is not to explore deeply the economics of choice under uncertainty, which

could be a challenging advanced Ph.D. course on it’s own. We are only interested in building enough

intuition and familiarity with the basics to understand what are some of the determinants of the risk

adjustment term - the part of asset pricing that reflects the effect of risk on the price - that we will explore

throughout the course.

In your microeconomics courses you will have studied how preferences are characterized by utility

functions u such that x < y x is preferred to y ⇔ u (x) ≥ u (y). For simplicity we will assume utility
∂ 2 u(x)
functions are twice continuous differentiable (i.e. ∂x2 ≡ u00 exists).6 When we introduce uncertainty it

is convenient to think of preferences over money or wealth instead of goods.

5 We will later see that there is a subtle distinction that can be made between uncertainty and risk in asset pricing. For

now you can think of them as the same.

6 This excludes the possibility of kinks in the utility function. We will later briefly discuss what allowing for kinks implies.

4
Definition 1 The expected utility representation of preferences, also referred to as von Neumann-

Morgensten utility functions, expresses the utility over an uncertain income x defined by possible payoffs

xi and corresponding probabilities πi by the expectation over the utility of each payoff:

 P πi u (xi ) if discrete number of outcomes

i
U (x) = E [u (x)] = R∞
π u (xi ) di if outcomes are continuous

−∞ i


so for two random variables x, y we say x < y ⇔ E [u (x)] ≥ E [u (y)]

For example, the payoffs could be the roll of a dice, in which case x = {1, 2, 3, 4, 5, 6} and (if the dice
1
is fair) πi = for i = 1 : 6. Or it could be to win £100 if heads, or loose £50 if tails in a (fair) coin toss,
6

in which case x = {100, −50} , πi = 12 , 12 .



You will probably be familiar with the widespread assumption that

• u0 ≥ 0 which just means we prefer having more (positive marginal utility);

• u00 ≤ 0 to reflect the fact we get less additional satisfaction the more we have (decreasing marginal

utility).

We will now see that these quite general assumptions in economics have strong implications for attitudes

towards risk.

Definition 2 Jensen’s inequality: If the function f is concave (f 00 ≤ 0) then for any random variable

x7

f (E [x]) ≥ E [f (x)]

with strict inequality if the function is strictly concave (f 00 < 0). The inequalities are the opposite if the

function is (strictly) convex. Clearly (can you see why?) if the function is linear equality holds.

If u00 ≤ 0 (concave) then Jensen’s inequality implies an expected utility agent prefers to have the average

of the outcomes with certainty (get the expected value) than the expected utility of getting each of those

outcomes, that is they dislike uncertainty:

u (E [x]) ≥ E [u (x)]

We will say that this agent (with concave utility function) is risk averse: they do not like the uncertainty

and would rather just get the expected value than take their chances.
7 Note that if x is not a random variable, then f (x) = E [f (x)] for any function.

5
Example 3 Lets consider a coin toss example where you either win 100 or get nothing ( x = {100, 16} , πi =
1 1 1/2
2 , 2 ) and preferences are described by Expected utility with utility function u (x) = 2x , which is risk

averse (u0 = x−1/2 , u00 = − 21 x−3/2 < 0 for x > 0 so u is strictly concave) then

E [x] E [u (x)] u (E [x])

58 14 15.23

Jensen’s inequality has two elements: uncertainty and the curvature of the function. We will now see

how each contributes to the amount we are willing to pay to avoid uncertainty, which are the underlying

determinants of the risk adjustment in asset pricing. We will later see that there will be equilibrium effects

in addition to pure preference and uncertainty effects we focus on for now.

2.1 The role of curvature.

Definition 4 The preference risk premium % (u, w, x) attached to a gamble x for an expected utility

agent with utility function u and wealth w is the fixed amount (not random) that satisfies

E [u (w + x)] = u (w + E [x] − % (u, w, x))

The risk premium defined this way measures the size of the discount relative to the expected value of

the gamble that the agent applies when valuing the gamble because of the uncertainty. Note that if x is

not random, i.e. x = E [x], then % (u, w, x) = 0 for any u, w. It can be thought of as the amount of money

the agent would be willing to pay (receive if negative) to eliminate the uncertainty of the gamble. Note

that in addition to the gamble itself and the agent’s preferences, this may depend on the agent’s wealth.

From Jensen’s inequality we can easily see that

% (u, w, x) > 0 if u00 < 0 (concave: risk averse)

% (u, w, x) < 0 if u00 > 0 (convex: risk loving)

% (u, w, x) = 0 if u00 = 0 (linear: risk neutral)

Note that we refer to u00 < 0 as risk averse because it implies that, because u0 > 0, E [u (w + x)] <

u (w + E [x]) so the agent would prefer to get the expected value insted of the gamble. The opposite is

true when u00 > 0, they prefer the gamble than the expected value (E [u (w + x)] > u (w + E [x])), so we

refer to them as risk loving. If u00 = 0 then E [u (w + x)] = u (w + E [x]) so the agent doesn’t care about

6
uncertainty, so re refer to them as risk neutral.

Therefore the curvature is what determines whether we are willing to pay to eliminate the uncertainty.

In this course (and in economics in general, as decreasing marginal utility is a standard assumption) we

will always assume agents are risk averse.

Some typical choices for risk averse utility functions are

x1−γ
• (power) u (x) = 1−γ , γ > 0, γ 6= 1 ⇒ u0 (x) = x−γ > 0 and u00 (x) = −γx−(1+γ) < 0 for x > 0

• (log) u (x) = ln (x) ⇒ u0 (x) = x−1 > 0 and u00 (x) = −x−2 < 0 for x > 0 (special case of power

utility as γ → 1)
h i
• (quadratic) u (x) = βx + ζ2 x2 for x ∈ 0, − βζ ⇒ u0 (x) = β + ζx > 0 and u00 (x) = ζ < 0 for
 
ζ < 0, x ∈ 0, − βζ (necessary to ensure u0 (x) > 0)

The preference risk premium is closely related to the concept of certainty equivalence widely used in

macroeconomics.

Definition 5 The certainty equivalent CE (u, w, x) of random income x is the fixed amount (i.e. not

random) an individual with preferences described by expected utility with utility function u and wealth w

would be willing to exchange for the gamble

u (w + CE (u, w, x)) = E [u (w + x)]

Since CE (u, w, x) is not random E [u (w + CE (u, w, x))] ≡ u (w + CE (u, w, x)). The certainty equiv-

alent is another way of thinking of the uncertainty adjustment, or risk premia:

u (w + CE (u, w, x)) = E [u (w + x)] = u (w + E [x] − % (u, w, x))

therefore

CE (u, w, x) = E [x] − % (u, w, x)

Hence, given u0 > 0, then for u00 < 0 we have % (u, w, x) > 0 ⇐⇒ CE (u, w, x) < E [x] . We will see that

the more uncertain is x and/or more risk averse the agent is (more concave u) then the higher is % (u, w, x),

so lower is CE (u, w, x).

Example 6 Continuing with the example above, what is the certainty equivalent and the risk premium of

the coin toss? In our previous example we did not consider any starting wealth, so equivalent to setting it

7
to zero. With w = 0 to find the CE we need the amount x0 such that u (x0 ) = E [u (x)]. In this simple
1/2
example we can solve for x0 : 2x0 = 14 → x0 = ( 14 2
2 ) = 49. So the CE of this gamble is 11 lower than

the expected value of the gamble, which is the size of the preference risk premia.

2.2 The role of uncertainty

Uncertainty is clearly key for Jensen’s inequality (it is defined for random variables only!). If x is known

(no uncertainty, E [x] = x) then irrespective of the curvature of the utility function u (E [x]) = u (x) =

E [u (x)]. To see that more uncertainty over x implies a bigger risk premia % (u, w, x), or equivalently a

smaller CE (u, w, x), for a given curvature consider the second order Taylor expansion8 around x̄ = E [x] :

1 2
u (x) = u (x̄) + u0 (x̄) (x − x̄) + u00 (x̄) (x − x̄) + higher order terms
2
∼ 1 2
= u (x̄) + u0 (x̄) (x − x̄) + u00 (x̄) (x − x̄)
2

Taking expectations over both sides9

1 h i
E [u (x)] ∼
2
= u (x̄) + u00 (x̄) E (x − x̄)
2
1
= u (E [x]) + u00 (x̄) V ar (x)
2

so
1 00
E [u (x)] − u (E [x]) = u (x̄) V ar (x)
2

where we have used the definition of x̄ = E [x]. Therefore, for a given curvature (u00 ) the higher is the

uncertainty (here proxied by variance) the higher is the loss in utility (remember that u00 < 0) due to

uncertainty (E [u (x)] − u (E [x])). Also, for given level of uncertainty, the more risk averse (the more

concave or more negative the second derivative) the higher is the loss in utility.

Note that if we just linearize the utility function around x̄ there would be no risk premium, which is

why macro models that are linearized around the steady state feature no risk premia even if the utility

functions are (and they always are) assumed to be concave/risk averse. I will return to this point a few

times during this course, as it is crucial in understanding some seemingly contradictory results from finance

and macroeconomics.

We have used a 2nd order Taylor approximation, so this is only a local argument. There are two cases
8 For
an n-differentiable (first n derivatives exist) function f , the n-th order Taylor expansion of f around x̄ is given by
k
1 ∂ f (x̄)
f (x) − f (x̄) ∼
Pn
= k=1 k! ∂xk
(x − x̄)k
9 Remember that if x is a random variable, then f (x) is a random variable for any function with f 0 6= 0. If x is not random,

then neither is f (x).

8
when variance is a sufficient statistic for uncertainty (so not just a local approximation). First, when x
h i
k
has a normal distribution (symmetric, so E (x − x̄) = 0 for all odd k) and the even derivatives of u are

non-positive, then % (u, w, x) is strictly increasing in the variance. Second, irrespective of the distribution

of x, if derivatives of u higher than 2 are all zero (as in the quadratic utility function above), then % (u, w, x)

is a function of 21 u00 (x̄) V ar (x) only (no approximation). More generally it will depend both on the type of

uncertainty and the utility function. In more realistic cases uncertainty is not described only by variance:

skewness, kurtosis and higher moments will matter.

The next definition formalizes the role of curvature in capturing risk aversion. As we will see shortly, it

will allow us to compare the levels of risk aversion across utility functions and make (relative) predictions

about the preference risk premia.

00
Definition 7 The Arrow-Pratt coefficient of absolute risk aversion : A (u, w) = − uu0 (w)
(w)
and the
00
coefficient of relative risk aversion given by R (u, w) = −w uu0 (w)
(w)
= wA (u, w)

These measures of risk aversion are useful in summarizing the degree of risk aversion and allow us to

obtain explicit formulas for the risk premium.

Consider first an additive gamble x with E [x] = 0, V ar (x) = σx2 and the amount k exposed in the

gamble, so E [u (w + kx)] = u (w − % (u, w, kx)). For small k, the risk premium is a function of A (u, w)

and the variance of the gamble y = kx given by σy2 = k 2 σx2 :

1
% (u, w, y) ∼
= σy2 A (u, w)
2

If instead we have a multiplicative gamble (so a proportion of wealth is at risk), so E [u (w (1 + kx))] =

u (w − % (u, w, kx)) , then the risk premia as a proportion of wealth is

% (u, w, y) ∼ 1 2
= σy R (u, w)
w 2

We can show these approximations hold by using Taylor expansions around k = 0.

For the power utility function the coefficient of relative risk aversion is given by R(u, w) = γ, which

controls the curvature of the function. Because it does not depend on the level of wealth (just the constant

γ) the power utility function is referred to as constant relative risk aversion (CRRA). This is a very

desirable property. Over the last century the level of real per capita consumption and wealth has grown

substantially. If our utility functions did not feature CRRA, then any asset price moments that depend on

risk aversion would have trends, which (we will see) is counterfactual. This is why power utility function

9
is a standard choice of utility functions, and one of the main drawbacks of the quadratic utility function

(which is convenient in deriving closed form solutions, as we will later see).

To see numerically what varying γ in the power utility case entails, consider the following simple

experiment. Suppose there is a 50% chance your salary will be £100 and a 50% chance it will be £50, and

you have no other income/wealth. For different values of γ, the first line of the table below gives the CE,

that is the amount you would be willing to take now for sure instead of having this salary uncertainty:10

γ 0 2 4 8 20 40 80

CE100,50 75 66.7 60.6 55.1 51.9 50.9 50.44

CE125,25 75 41.7 31.4 27.6 25.9 25.4 25.2

This table is important to understand some important macro-finance puzzles. At γ ≥ 20 the individual

is willing to take just £1.9 above the worst outcome (£51.9 instead of the guaranteed minimum of £50)

instead of having the opportunity of winning another £50 with 50% chance. This is quite extreme, and

considered highly implausible. Another way to think about his is that you have £50 guaranteed and then

you are asked how much is the possibility of getting another £50 by flipping a fair coin worth. The answer

is CE in the table above - 50. If you are risk neutral, you pay the expected value (which would be £25,

hence the CE for γ = 0 is £50+£25) because uncertainty doesn’t bother you. For γ = 80 you would value

this lottery at £0.4. Ask yourself what would you be willing to pay.

The second line gives the CE for a gamble with same expected value but higher uncertainty: 50/50

chance to get £125 or £25. This is a more risky gamble: you can win more, but are guaranteed less (or

loose more relative to the initial gamble). Note how the CE is much smaller for any curvature parameter

except when risk neutral. This is because for any level of risk aversion, the more risky the proposition, the

less you like it, so you would be willing to exchange it for less. Or put the other way around, you are willing

to pay more to get rid of uncertainty. The next table shows the risk premia (% (u, 0, x) = E [x]−CE (u, 0, x))

for different levels of relative risk averion (γ) for these two gambles, which have same expected value but

differ in uncertainty:
10 To x1−γ
calculate it we use u (CE (x)) = E [u (x)] , u (x) = 1−γ
⇒
  1−γ  1
x 1−γ
CE (x) = (1 − γ) E
1−γ
 1
1001−γ 501−γ
 
1−γ
= (1 − γ) 0.5 + 0.5
1−γ 1−γ

1
0.5 ∗ 1001−γ + 501−γ 1−γ

=

10
 γ 0 2 4 8 20 40 80

%100,50 0 8.3 14.4 19.9 23.1 24.1 24.6

%125,25 0 33.3 43.6 47.4 49.1 49.6 49.8

In the previous tables we have taken wealth to be zero. What if we vary wealth? The Table below

shows the risk premia for the same gamble, with γ = 4 and different wealth levels, from 0 to 1 billion.

w 0 100 1000 10k 100k 1mn 1bn

%100,50 14.4 6.9 1.2 0.1 0.01 0.001 0

%125,25 43.6 25.6 4.6 0.5 0.05 0.005 0

We still see a higher risk premia for the more risky proposition, but both are much smaller for larger

levels of wealth. What explains the patterns in the previous table? The power utility has absolute risk
γ


aversion that is decreasing in wealth A (u, w) = w . As you become really wealthy, the risk premium for

a same (absolute size of) gamble becomes smaller, and for 1 billion risking 50 or 100 is really tiny. The

exercises at the end will take you through some more variations similar to the examples above to help

clarify these concepts.

There is a whole field looking at all the intricacies of choice over uncertainty. We have only covered

enough to go back to our asset pricing objective with some understanding of the determinants of the

magnitude of corrections for risk. Before putting it together and linking back to asset pricing, let’s see a

final way in which the effect of uncertainty can be summarized, which will be useful in building intuition

for one of the ways to express the main asset pricing equation, which is by twisting probabilities instead

of the amount we are willing to pay to get rid of the uncertainty.

Definition 8 For a level of wealth w and utility function u, the probability premium π̃ (u, w, x) is the

excess in winning probability over fair odds required to make the individual indifferent between gambling x

or not:
   
1 1
u (w) = + π̃ (u, w, x) u (w + x) + − π̃ (u, w, x) u (w − x)
2 2

In other words, the probability premium is the minimum twist in probabilities in your favour relative

to fair odds required to make a pure bet (win or lose x) acceptable.

It can be shown that

∂ π̃ (u, w, 0)
A (u, w) = 4
∂x

which shows that the degree of probability twisting is just another way of capturing the effect of risk

11
aversion.
1 1
Example 9 Using a slight modification to our earlier example (w = 100 x = {50, −50} , πi = 2, 2 and

Expected utility with power utility and γ = 0.5, u (x) = 2x1/2 ) we can find the probability premium by

solving for π̃ : u (w) = 12 + π̃ u (w + 50) + 21 − π̃ u (w − 50) ⇒ 11 π̃ = 0.066. That is, instead of fair

odds, you need a 56.6% chance of winning (43.4% of losing) to accept this gamble. If instead of γ = 0.5

we have γ = 4, then π̃ = 0.4, i.e. a 90% chance of winning is required to take the gamble! The table below

shows some variations:

w, γ 100, 0.5 100, 4 100, 20 1000, 0.5 1000, 4 1000, 20 1mn, 20

π̃50,−50 0.066 0.409 0.4999 0.006 0.05 0.23 0.00025

Note that as you increase risk aversion from γ = 0.5 to γ = 4 and then γ = 20, the probability of

winning you would require increases dramatically to 90.9% (50%+π̃=50%+40.9%) and 99.99% (i.e. you

essentially require a sure thing at γ = 20 for initial wealth of 100). Note how the probability twist decreases

as the wealth increases, just as the preference risk premia decreases for a given gamble size as we increased

wealth for this class (CRRA) of utility functions.

An alternative, perhaps more intuitive, way to express probability twists is to define a risk-adjusted

probability as the probabilities π ∗ such that

E [u (w + x)] = u (w + CE (u, w, x))

= u (w + E [x] − % (u, w, x))

= u (w + E ∗ [x])

where the expectation E ∗ is calculated according to probabilities π ∗ . This would imply

E ∗ [x] = E [x] − % (u, w, x) = CE (u, w, x)

Since risk aversion implies % (u, w, x) > 0, it implies E ∗ [x] < E [x]. Can you see how the risk-

adjusted probabilities imply that risk averse agents value uncertain income as if they were risk-neutral

but pessimists? That is, they act as risk neutral by valuing gambles as they value it’s expected value

(E [u (w + x)] = u (w + E ∗ [x])), but the expected value is calculated by overweighting the bad outcomes

and under-weighting the good outcomes (E ∗ [x] < E [x]). It is an alternative way of expressing aversion to

risk. We can do it either as a risk premium or twist in probability. We will see this equivalence later in
11 π̃ 1

= u (w) − 2
[u (w + 50) + u (w − 50)] / [u (w + 50) − u (w − 50)]

12
equilibrium asset pricing. It will prove incredibly useful, particularly in pricing derivatives (at this point

it should not be clear why it miht be helpful).

2.3 Summarizing risk aversion

These concepts allow us to rank attitudes towards risk, and are the key take-away from this lecture.

The next two propositions summarize the relation between them (which we will state without proof here).

Proposition 10 For a expected utility agent with utility function u (), the following are equivalent (mean-

ing, any of them implies all the other) for (w, x):

1. agent is strictly risk averse

2. % (u, w, x) > 0

3. CE (u, w, x) < E [x]

4. u00 < 0 (strictly concave utility)

5. π̃ (u, w, x) > 0

Proposition 11 For two expected utility agents with utility functions uA , uB the following are equivalent

for every (w, x):

1. A is more risk averse than B

2. A (uA , w) > A (uB , w)

3. CE (uA , w, x) < CE (uB , w, x)

4. % (uA , w, x) > % (uB , w, x)

5. π̃ (uA , w, x) > π̃ (uB , w, x)

This ranking of attitude towards risk is not always possible. For example, some utility functions display

decreasing absolute risk aversion (ARA) for some pair (w, x) while having constant ARA for other values

of (w, x), so the second condition will be true only over some range of values of (w, x).

13
2.4 Time-separable utilities

This is a small aside, and not purely related to choice under uncertainty, but more about intertemporal

choice. But it interacts with uncertainty. To see why, consider a non-time-separable expected utility

function

U (ct , ct+1 ) = E [u (ct , ct+1 )]

In this case we would just define risk aversion as

Et [u (ct , ct+1 )] ≤ u (ct , Et [ct+1 ])

When u is not time separable, then the measures of t + 1 risk aversion depend on date t consumption. With

time-separable utility, measures of risk aversion for any period only depend on that period. This makes

life simpler for us now, but comes at a cost of flexibility in matching data, which we’ll discuss later.

3 Link to Finance

We will now see how these measures of risk aversion have direct implications for demand for risky

assets and insurance. We will have a more detailed analysis of portfolio choice in a few lectures ahead.

3.1 Demand for risky asset

Suppose there are two assets, a safe asset that returns £1 for every pound invested (so Rf = 1) for

computational simplicity, and a risky asset the returns z per pound invested (so R = z). If the individual

has wealth w and denoting by α (u, w) the investment in in the risky asset, it can be shown that:

1. If u00 < 0, then α (u, w) > 0 ⇔ E [z] > 1

2. If A (uA , w) > A (uB , w) ⇒ α (uA , w) ≤ α (uB , w)

To show this notice that the problem can be written as

max E [u (w + α (z − 1))]
α

It is useful to define φ (α) = E [u (w + α (z − 1))], so we can express the FOC as

φ0 (α) = E [u0 (w + α (z − 1)) (z − 1)] = 0

14
 h i
2
Because u00 < 0, φ00 (α) = E u00 (w + α (z − 1)) (z − 1) < 0, so φ0 (α) is strictly decreasing for any α. It

follows that α (u, w) > 0 ⇔ φ0 (0) > 0 (why? if φ0 (0) > 0, φ00 < 0 then the only way to achieve φ0 (α) = 0,

that is satisfy the FOC, is by increasing α, so α > 0). φ0 (0) = E [u0 (w) (z − 1)] > 0 ⇔ E [(z − 1)] > 0.

The exact same steps show that α (u, w) = 0 ⇔ E [z] = 1 and α (u, w) < 0 ⇔ E [z] < 1. (note that risk

aversion u00 < 0 was crucial in establishing that the FOC was decreasing in α)

In words, this means that agents will only invest in risky assets if it offers returns that are higher than

the risk free rate (in this case R = E [z] > 1 = Rf ) , and that the more risk averse an agent is, the

smaller will be the portfolio allocation on the risky asset (if any at all). The first is a manifestation of the

underlying intuition of asset pricing: we require a positive expected excess return if we are going to hold a

risky asset, and the more risk averse we are the higher the excess return will have to be for us to be willing

to hold a given amount of the risky asset. When we combine this with equilibrium considerations in the

next few lectures we will determine the required excess return in equilibrium.

3.2 Demand for Insurance

Suppose you have wealth w, expected utility u () and are exposed to some risk x. If the price of

insurance is the expected value, then the individual will fully insure. If there is an insurance premium,

then insurance will be partial. This just says that we would all buy full insurance for all the risks we are

exposed to if there were no insurance premium.

This is just the flip side of the demand for risky asset. We will only accept a negative excess return

(pay more for an asset than we expect to get relative to investing in a risk-free bond) if the asset provides

a hedge for our risk. If it provides a hedge at no premium, then we’d choose full insurance. This is the first

glimpse of one of the fundamental insights of asset pricing: how much an asset adds (making your overall

outcome more uncertain) or subtracts (by hedging/insuring your other cources of uncertainty) your overall

risk is key to the pricing of that asset.

4 Key concepts and review questions

The usual assumption in all of economics about preferences, of decreasing marginal utility, implies

we dislike risk. Risk aversion is key to understanding asset pricing. Knowing the absolute and relative

risk aversion of an agent allows us to make a few basic predictions and comparative statics (see summary

of equivalent statements). The degree of risk aversion and quantity of risk will be important ingredients

(but not the whole story) for understanding the economics and intuition of the equilibrium asset pricing

relations we will derive. You should work through the exercises below to gain some intuition about these

15
concepts.

−2 1
Exercise 1 (Risk aversion ranking) Suppose that ua (w) = − w2 , ub (w) = ln (w) , uc (w) = w 2 and all

have initial wealth w0 (you will need Excel or a calculator to find the numerical values)

1. are they risk averse, risk neutral or risk lovers?

2. compute the coefficients of absolute and relative risk aversion as a function of wealth

3. can you rank their attitudes towards risk at w0 = 10?

4. can you rank the Certainty Equivalents/preference risk premia for a gamble that pays 1 with prob-

ability 0.5 and nothing with probability 0.5 at w0 = 10 (hint: to compute the CE remember that

eln(x) = ln (ex ) = x)

5. does your answer (about ranking) change for w0 = 2? why or why not?

Exercise 2 (Risk-adjusted expectations) Suppose an asset offers returns of either 10% with probability

75%, or 1% with probability 25% at the end of 1 year (that is, if you invest 100 you will either receive £110

with probability 75%, or £101 with probability 25% after 1 year)

1. What is the expected return of the investment?

−2
2. If an investor has utility uA (w) = − w2 and wealth 10, what risk-free return would make A indifferent

between investing all of her wealth in the risky investment or in the risk free asset?

3. If another investor has utility uB (w) = a + bw, for positive a and b, what alternative probabilities for

the returns of the risky asset would make B be indifferent with the same risk-free return as investor

A was? How is this related to risk-adjusted expectations?

4. Can you predict what would have happened to your answers above if A instead had log utility?

Exercise 3 (Risk aversion and required returns) An investor has a log utility function. She has £10,000

and can buy an investment A that can earn her an extra £1,000 with probability p and zero otherwise, or

put her money in a (risk-free) saving account and get £500 in return.

1. For what value of p will she choose A?

2. Repeat previous question with utility function u (x) = 1 − e−ax , with a = 1

1000 .

3. Repeat previous question with utility function u (x) = x

16
 4. Compare your answers to 1 through 3.

Exercise 4 (Ranking assets based on returns and preferences) Assets X,Y,Z have the following returns
r
(the way to read the table is that for return r and investment of w they pay back w(1 + 100 )) that depend

on the particular state of the world:

return

Probability state X Y Z

0.2 s1 6 12 4.5

0.3 s2 5 6 4.5

0.4 s3 4 2 4.5

0.1 s4 3 -2 4.5

Investors must place all of their money, w = 1, in only one of these assets. For each of the following
x1−γ
investors, rank the assets: u (x) = ln (x) , u (x) = 1−γ for γ = {0.5, 2, 10, 40} (hint: use Excel or other

software to economize on computations). Interpret your results in light of the properties of these assets.

Exercise 5 (Risk aversion and insurance demand - harder)Suppose a strictly risk averse individual that

has wealth w but runs the risk of loosing D (D ≤ w), the probability of loss is π. The individual can buy

insurance. One unit of insurance costs p and pays 1 if the loss occurs. Thus, if α units of insurance are

bought, the wealth of the individual will be w − αp if no loss occurs and ω − αp − D + α if a loss occurs

1. what is the expected wealth (as a function of α)

2. show that if p = π, it is optimal to fully insure (α∗ = D)

3. what does p = π mean?

5 Reading

No further reading beyond these Lecture Notes and the exercises are required.

Danthine & Donaldson (3rd edition): provide a very detailed and extensive treatment of choice under

uncertainty, much more than needed for this course. Most of the relevant material is spread out in Chapters

3 and 4, with application to investment decisions in Chapter 5.

Cochrane: doesn’t really devote any Chapters or Sections to choice under uncertainty.

If you are interested in Behavioral Finance, Hersh Shefrin’ ”A Behavioral Approach to Asset Pricing”

(Associated Press, 2008, 2nd Edition) provides a comprehensive treatment with an emphasis on making

17
explicit the stochastic discount factor, which will be our focus from the next lecture until the end of the

course.

6 Appendix: constrained demand for risky asset is positive if

positive excess return (not required )

We now consider the problem of demand for risky asset with no selling or borrowing (so 0 ≤ α ≤ w) and

allow the risk free return to be Rf and the return on £1 invested in the risky security to be R:

max E u wRf + α R − Rf



α

st.

0≤α≤w

We can write the Lagrangian as

L = max E u wRf + α R − Rf



− λ1 (α − w) − λ2 (−α)
α

with FOC:

∂L
= 0
∂α
λ1 (w − α) = 0

λ2 α = 0

λ1 , λ2 , α, w − α ≥ 0

which results in

E u0 wRf + α R − Rf R − Rf − λ1 + λ2 = 0





λ1 (w − α) = 0

λ1 , λ2 , α, w − α ≥ 0

18
If α < w → λ1 = 0, α > 0 → λ2 = 0 Therefore to satisfy the FOC must be one of the following:

α∗ (0, w) , E u0 wRf + α∗ R − Rf R − Rf = 0




∈

α∗ = 0, E u0 wRf + α∗ R − Rf R − Rf ≤ 0





α∗ = w, E u0 wRf + α∗ R − Rf R − Rf ≥ 0





Let φ(α) = E u0 wRf + α∗ R − Rf R − Rf then φ(0) = E u0 wRf R − Rf = u0 wRf E [R] − Rf .




 



Therefore if u0 (w) > 0 




 > 0 if E [R] > Rf

φ(0)
 = 0 if E [R] = Rf


 < 0 if E [R] < Rf


h
2 i
And φ0 = E u00 wRf + α∗ R − Rf ≤ 0 if u00 < 0. Therefore



R − Rf

if E [R] ≤ Rf ⇒ α∗ = 0

if E [R] > Rf ⇒ α∗ > 0 otherwise the FOC cannot be satisfied.

19