Risk Aversion and Utility Theory

Utility Theory and Risk Aversion
John Y. Campbell
Ec2723
September 2011
John Y. Campbell (Ec2723)
September 2011
1 / 35
Outline
Outline
Utility Theory
I
I
Basics
Jensens Inequality
Risk Aversion
I
I
I
Measurement and Comparison

The Arrow-Pratt Approximation
Declining Absolute Risk Aversion
Tractable Utility Functions

Rabin Critique
Comparing Risks
September 2011
2 / 35
Utility Theory
Basics
Ordinal and Cardinal Utility

There is a basic contrast between:
Ordinal utility (x ) is invariant to monotonic transformations, so
(x ) is equivalent to ((x )) for any strictly increasing .
Cardinal utility (x ) is invariant to positive a ne (aka linear)
transformations, so (x ) is equivalent to a + b(x ) for any b > 0.
In nance we rely heavily on von Neumann-Morgenstern utility theory
which says that choice over lotteries, satisfying certain axioms, implies
maximization of the expectation of a cardinal utility function, dened over
outcomes.
September 2011
3 / 35
Utility Theory
Basics
Reducing the Dimension of Utility

Finance theory generally works with low-dimensional arguments of the
utility function. Proceeding from greater to lesser generality:
Multiple goods, dates, and states (ordinal utility)
Multiple goods and dates, taking expectation over states (cardinal
utility)
Multiple dates only (one-good simplication)
Time-separable utility, adding up over dates (U (C1 ) + U (C2 ) + ...)
Single-date utility U (C1 ), which is equivalent to utility dened over
wealth U (W1 ).
September 2011
4 / 35
Utility Theory
Jensens Inequality
Jensens Inequality
Consider a random variable e

z and a function f .
Denition: f is concave i for all 2 [0, 1] and values a, b,

f (a) + (1
)f (b )
f (a + (1
)b ).
If f is twice dierentiable, then concavity implies that f 00

Jensens Inequality : Ef (e
z)
0.
f (Ee
z ) for all possible e
z i f is concave.
September 2011
5 / 35
Utility Theory
Jensens Inequality
Jensens Inequality and Risk Aversion

Jensens Inequality : Ef (e
z)
f (Ee
z ) for all possible e
z i f is concave.
Denition: an agent is risk averse if he dislikes all zero-mean risk at all

levels of wealth. That is, for all w0 and risk e
x with Ee
x = 0,
This is equivalent to
where e
z = w0 + e
x.
Eu (w0 + e
x)
Eu (e
z)
u ( w0 ) .
u (Ee
z ),
Thus risk aversion is equivalent to concavity of the utility function.
September 2011
6 / 35
Risk Aversion
Measuring Risk Aversion
A natural measure of risk aversion is f 00 , scaled to avoid dependence on

the units of measurement for utility. Absolute risk aversion A is dened by
A=
f 00
.
f0
Note that in general this is a function of the initial level of wealth.
September 2011
7 / 35
Risk Aversion
Comparing Risk Aversion

Let u1 and u2 have the same initial wealth. u1 is more risk-averse than u2
if u1 dislikes all lotteries that u2 dislikes, regardless of the common initial
wealth level.
Dene (x ) = u1 (u2 1 (x )). What are the properties of this function?
1. u1 (z ) = (u2 (z )), so (.) turns u2 into u1 .
2. u10 (z ) = 0 (u2 (z ))u20 (z ), so 0 = u10 /u20 > 0.
3. u100 (z ) = 0 (u2 (z ))u200 (z ) + 00 (u2 (z ))u20 (z )2 , so
00 =
u100
0 u200
u20 2
u10
(A2
u20 2
A1 ).
September 2011
8 / 35
Risk Aversion
4. Consider a risk e
x that is disliked by u2 , that is a risk s.t.
Eu2 (w0 + e
x ) u2 (w0 ). We have
Eu1 (w0 + e
x ) = E(u2 (w0 + e
x ))
(Eu2 (w0 + e
x ))
for all e
x i is concave or equivalently 00
we must have A1 A2 .
(u2 (w0 )) = u1 (w0 )
0. But then from property 3
September 2011
9 / 35
Risk Aversion

5. The risk premium is the amount one is willing to pay to avoid a pure
(zero-mean) risk. It solves
Dening z = w0
Eu (w0 + e
x ) = u ( w0
).
and ye = + e
x , this can be rewritten as
Eu (z + ye) = u (z ).
Now dene 2 as the risk premium for agent 2, and dene z2 and ye2
accordingly. We have
Eu2 (z2 + ye2 ) = u2 (z2 ).
If u1 is more risk-averse than u2 , then
which implies 1
2 .
Eu1 (z2 + ye2 )
u1 (z2 ),
September 2011
10 / 35
Risk Aversion
6. Consider a risk that may have a nonzero mean . It pays + e

x where
e
x has zero mean. The certainty equivalent C e satises
Eu (wo + + e
x ) = u ( w0 + C e ) .
This implies that
C e (w0 , u, + e
x) =
(w0 + , u, e
x ).
Thus if u1 is more risk-averse than u2 , then C1e
C2e .
September 2011
11 / 35
Risk Aversion
In summary, the following statements are equivalent:

u1 is more risk-averse than u2 .
u1 is a concave transformation of u2 .
A1
A2 .
2 .
C1e
C2e .
September 2011
12 / 35
Risk Aversion
Consider a pure risk ye = ke

x , where k is a scale factor. Write the risk
premium as a function g (k ): g (k ) = (w0, u, ke
x ) satises
Eu (w0 + ke
x ) = u ( w0
g (k )).
Note that g (0) = 0.

Now dierentiate w.r.t. k:
Ee
x u 0 (w0 + ke
x) =
g 0 ( k ) u 0 ( w0
g (k )).
Since Ee
x = 0, this implies that g 0 (0) = 0.
September 2011
13 / 35
Risk Aversion
Dierentiate w.r.t. k a second time:

Ee
x 2 u 00 (wo + ke
x ) = g 0 (k )2 u 00 (w0
g (k ))
g 00 (k )u 0 (w0
g (k )),
which implies that

g 00 (0) =
u 00 (w0 ) 2
Ee
x = A(w0 )Ee
x 2.
u 0 ( w0 )
September 2011
14 / 35
Risk Aversion

Now take a Taylor approximation of g (k ):
g (k )
1
g (0) + kg 0 (0) + k 2 g 00 (0).
2
Substituting in, we get
1
1
A(w0 )k 2 Ee
x 2 = A(w0 )Ee
y 2.
2
2
The risk premium is proportional to the square of the risk. This property
of dierentiable utility is known as second-order risk aversion. It implies
that people are approximately risk-neutral with respect to small risks.
We also nd that
1
A(w0 )k 2 Ee
x 2,
Ce
k
2
so a positive mean has a dominant eect for small risks.
September 2011
15 / 35
Risk Aversion
Relative Risks
b as
Dene a multiplicative risk by w
e = w0 (1 + ke
x ) = w0 (1 + ye). Dene
the share of wealth one would pay to avoid this risk:
b=
Then
(w0 , u, w0 ke
x)
.
w0
1
1
w0 A(w0 )k 2 Ee
x 2 = R (w0 )Ee
y 2,
2
2
where R (w0 ) = w0 A(w0 ) is the coe cient of relative risk aversion.
b
September 2011
16 / 35
Risk Aversion
Decreasing Absolute Risk Aversion
Decreasing Absolute Risk Aversion

The following conditions are equivalent:
is decreasing in w0 .
A(w0 ) is decreasing in w0 .
u 0 is a concave transformation of u, so u 000 /u 00
u 00 /u 0
000
00
everywhere. The ratio u /u = P has been called absolute
prudence by Kimball, who relates it to the theory of precautionary
saving.
Decreasing absolute risk aversion (DARA) seems intuitively appealing.
Certainly we should be uncomfortable with increasing absolute risk
aversion.
September 2011
17 / 35

Almost all applied theory and empirical work in nance uses some member
of the class of linear risk tolerance (LRT) or hyperbolic absolute risk
aversion (HARA) utility functions. These are dened by
u (z ) = ( +
z 1
)
dened over z s.t. + z/ > 0.

For these utility functions, we get
T (z ) =
z
1
=+ ,
A(z )
which is linear in z, and

A(z ) =
z
+
which is hyperbolic in z.
September 2011
18 / 35
Important Special Cases

Quadratic utility has = 1. This implies increasing absolute risk
aversion and the existence of a bliss point at which u 0 = 0. These
are important disadvantages, although quadratic utility is tractable in
models with additive risk.
Exponential or constant absolute risk averse (CARA) utility is the
limit as ! . To obtain constant absolute risk aversion A, we
need
u 00 (z ) = Au 0 (z ).
Solving this dierential equation, we get
u (z ) =
exp( Az )
,
A
where A = 1/. This form of utility is tractable with normally

distributed risks because then utility is lognormally distributed.
September 2011
19 / 35
Important Special Cases

Power or constant relative risk averse (CRRA) utility has = 0 and
> 0. Relative risk aversion R = . For 6= 1,
u (z ) =
z1
.
1
For = 1,
u (z ) = log(z ).
Power utility is appealing because it implies stationary risk premia and
interest rates even in the presence of long-run economic growth. Also
it is tractable in the presence of multiplicative lognormally distributed
risks.
A negative represents a subsistence level. Rubinstein has argued
for this model, but economic growth renders any xed subsistence
level irrelevant in the long run. Models of habit formation have
time-varying subsistence levels which can grow with the economy.
September 2011
20 / 35
Rabin Critique
Rabin Critique
Matthew Rabin (2000) has argued that standard utility theory cannot
explain observed aversion to small gambles without implying ridiculous
aversion to large gambles. This follows from the fact that dierentiable
utility has second-order risk aversion.
To understand Rabins critique, consider a gamble that wins $11 with
probability 1/2, and loses $10 with probability 1/2. With diminishing
marginal utility, the utility of the win is at least 11u 0 (w0 + 11). The utility
cost of the loss is at most 10u 0 (w0 10). Thus if a person turns down
this gamble, we must have 10u 0 (w0 10) > 11u 0 (w0 + 11) which implies
10
u 0 (w0 + 11)
< .
0
u (w0 10)
11
September 2011
21 / 35
Rabin Critique
Rabin Critique
u 0 (w0 + 11)
10
< .
0
u (w0 10)
11
Now suppose the person turns down the same gamble at an initial wealth
level of w0 + 21. Then
u 0 (w0 + 21 + 11)
u 0 (w0 + 32)
10
=
< .
0
0
u (w0 + 21 10)
u (w0 + 11)
11
Combining these two inequalities,
u 0 (w0 + 32)
<
u 0 (w0 10)
10
11
100
.
121
If this iteration can be repeated, it implies extremely small marginal utility

at high wealth levels, which would induce people to turn down apparently
extremely attractive gambles.
September 2011
22 / 35
Rabin Critique
Responses to Rabin Critique
1. As we increase wealth, a person will continue to turn down a given

absolute gamble indenitely only if absolute risk aversion is constant or
increasing. Rabins most extreme results assume this, and can be
understood as a critique of constant or increasing absolute risk aversion.
2. Observed aversion to small risks probably results from some other
aspect of human psychology besides declining marginal utility of wealth.
But this does not necessarily mean that we should abandon utility theory
for studying large risks in nancial markets.
September 2011
23 / 35
Rabin Critique
3. What does explain the observed aversion to small risks? One simple
story would be that people suer xed disutility from any loss, regardless
of size. More attractive are nonstandard preference models with
rst-order risk aversion, in which the premium paid to avoid a small loss
is proportional to the loss, not the squared loss. Such models have a kink
in the utility function at the initial level of wealth. Kahneman and
Tverskys (1979) prospect theory has this feature.
September 2011
24 / 35
Rabin Critique

4. Barberis, Huang, and Thaler (2006) point out that even rst-order risk
aversion cannot generate substantial aversion to small delayed gambles.
During the time between the decision to take a gamble and the resolution
of uncertainty, the agent will be exposed to other risks and will merge
these with the gamble under consideration. If the gamble is uncorrelated
with the other risks, it is diversifying. In eect the agent will have
second-order risk aversion with respect to delayed gambles. To deal with
this problem, Barberis et al. argue that people treat gambles in isolation,
that is, they use narrow framing.
5. Chetty and Szeidl (2007) show that consumption commitments (xed
costs to adjust a portion of consumption) raise risk aversion over small
gambles, relative to risk aversion over large gambles where extreme
outcomes would justify paying the cost to adjust all consumption.
September 2011
25 / 35
QJE May 2007
Comparing Risks
Comparing Risks
We would like to compare the riskiness of dierent distributions. Three

possible notions of increasing risk:
1) Something that all concave utility functions dislike.
2) More weight in the tails of the distribution.
3) Added noise.
The classic analysis of Rothschild and Stiglitz shows that these are all
equivalent. They are not equivalent to higher variance.
September 2011
26 / 35
Comparing Risks
Comparing Risks
e and Y
e which have the same expectation.
Consider random variables X
e is weakly less risky than Y

e if no individual with a concave utility
1) X
e to X
e:
function prefers Y
h
i
h
i
e)
e)
E u (X
E u (Y
e is less risky (without qualication) if there is
for all concave u (.). X
e.
some concave u (.) which strictly prefers X
e and
Note that this is a partial ordering. It is not the case that for any X
e , either X
e is weakly less risky than Y
e or Y
e is weakly less risky than X
e.
Y
We can get a complete ordering if we restrict attention to a smaller class
of utility functions than the concave, such as the quadratic.
September 2011
27 / 35
Comparing Risks
Comparing Risks
e is less risky than Y
e if the density function of Y
e can be obtained from
2) X
e by applying a mean-preserving spread (MPS). An MPS s (x ) is
that of X
dened by
1
0
for c < x < c + t
B for c 0 < x < c 0 + t C
C
B
for d < x < d + t C
s (x ) = B
C
B
@ for d 0 < x < d 0 + t A
0 elsewhere
where , , t > 0 ; c + t < c 0 < d
(c 0 c ) = (d 0
can move it.
t < d + t < d 0 ; and
d ), that is, the more mass you move, the less far you
September 2011
28 / 35
Comparing Risks
Comparing Risks
An MPS is something you add to a density function f (x ). If
g (x ) = f (x ) + s (x ), then (i) g (x ) is also a density function, and (ii) it
has the same mean as f (x ).
(i) is obvious because
s (x ) dx = area under s (x ) = 0.
(ii) follows from the fact that the mean of s (x ),

follows from (c 0 c ) = (d 0 d ).
xs (x ) = 0, which
In what sense is an MPS a spread? Its obvious that if the mean of f (x ) is

between c 0 + t and d, then g (x ) has more weight in the tails. Its not so
obvious when the mean of f (x )is o to the left or the right. Nevertheless,
e with density g is riskier than X
e with density f in the
we can show that Y
sense of 1) above. In this sense the term spread is appropriate.
September 2011
29 / 35
Comparing Risks
Comparing Risks
e is less risky than Y

e if Y
e
3) A formal denition of added noise is that X
e + e, where E [ejX ] = 0 for all values of X .
has the same distribution as X
We say that e is a fair game with respect to X .
e
This condition is stronger than zero covariance, Cov e, X
e
weaker than independence, Cov f (e) , g X
e
and g . It is equivalent to Cov e, g X
= 0. It is
= 0 for all functions f
= 0 for all functions g .
September 2011
30 / 35
Comparing Risks
Comparing Risks
3) is su cient for 1):
h
i
e + e)jX
E U (X
h
i
e + ejX ) = U (X )
U (E X
h
i
h
i
e + e)
e)
) E U (X
E U (X
h
i
h
i
e)
e)
) E U (Y
E U (X
e and X
e + e have the same distribution.
because Y
In fact, Rothschild-Stiglitz show that 1), 2) and 3) are all equivalent. This
is a powerful result because one or the other condition may be most useful
in a particular application.
September 2011
31 / 35
Comparing Risks
Comparing Risks
e having greater variance than X

e ? It is
Why are these not equivalent to Y
e
e
e
obvious from 3) that if Y is riskier than X then Y has greater variance
e . The problem is that the reverse is not true in general. Greater
than X
e could have
variance is necessary but not su cient for increased risk. Y
e but still be preferred by some concave utility
greater variance than X
functions if it has more desirable higher-moment properties. This
possibility can only be eliminated if we conne attention to a limited class
of distributions such as the normal distribution.
September 2011
32 / 35
Comparing Risks
Stochastic Dominance
The following are denitions.
e dominates Y
e if Y
e =X
e +e
X
, where e
0.
e rst-order stochastically dominates Y

e if Y
e has the distribution of
X
e
e
e
e and G (.) is the
X + , where 0. Equivalently, if F (.) is the cdf of X
e , then X
e rst-order stochastically dominates Y
e if F (z ) G (z )
cdf of Y
for every z. First-order stochastic dominance implies that every increasing
e.
utility function will prefer X
e second-order stochastically dominates Y

e if Y
e has the distribution of
X
e
e
e
X + + e, where 0 and E [ejX + ] = 0. Second-order stochastic
dominance implies that every increasing, concave utility function will prefer
e . Increased risk is the special case of second-order stochastic dominance
X
where e
= 0.
September 2011
33 / 35
Comparing Risks
Application: The Principle of Diversication
Consider n lotteries with payos e

x1 , e
x2 , ..., e
xn that are independent and
identically distributed (iid). You are asked to choose weights 1 , 2 , ....,
n subject to the constraint that i i = 1. It seems obvious that the
best choice is complete diversication, with weights i = 1/n for all i.
The payo is then
1 n
xi .
e
z = e
n i =1
September 2011
34 / 35
Comparing Risks
Application: The Principle of Diversication
To prove that this is optimal, note that the payo on any other strategy is
i exi = ez +
i
and
E [ejz ] =
1
n
1
n
e
xi = e
z + e,
E [e
xi jz ] = k i
i
1
n
= 0.
September 2011
35 / 35

Risk Aversion and Utility Theory

Caricato da

Informazioni sul documento

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Risk Aversion and Utility Theory

Caricato da

Copyright:

Formati disponibili

Utility Theory and Risk Aversion

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Measurement and Comparison

Tractable Utility Functions

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Ordinal and Cardinal Utility

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Reducing the Dimension of Utility

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Consider a random variable e

Denition: f is concave i for all 2 [0, 1] and values a, b,

If f is twice dierentiable, then concavity implies that f 00

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Jensens Inequality and Risk Aversion

Denition: an agent is risk averse if he dislikes all zero-mean risk at all

Thus risk aversion is equivalent to concavity of the utility function.

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Measurement and Comparison

Measuring Risk Aversion

A natural measure of risk aversion is f 00 , scaled to avoid dependence on

Note that in general this is a function of the initial level of wealth.

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Measurement and Comparison

Comparing Risk Aversion

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Measurement and Comparison

Comparing Risk Aversion

John Y. Campbell (Ec2723)

(u2 (w0 )) = u1 (w0 )

0. But then from property 3

Utility Theory and Risk Aversion

Measurement and Comparison

Comparing Risk Aversion

If u1 is more risk-averse than u2 , then

Eu1 (z2 + ye2 )

Utility Theory and Risk Aversion

Measurement and Comparison

Comparing Risk Aversion

6. Consider a risk that may have a nonzero mean . It pays + e

This implies that

Thus if u1 is more risk-averse than u2 , then C1e

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

Measurement and Comparison

Comparing Risk Aversion

In summary, the following statements are equivalent:

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

The Arrow-Pratt Approximation

The Arrow-Pratt Approximation

Consider a pure risk ye = ke

Note that g (0) = 0.

John Y. Campbell (Ec2723)

Utility Theory and Risk Aversion

The Arrow-Pratt Approximation

The Arrow-Pratt Approximation