4 Measuring Risk and Risk Aversion

A Measuring Risk Aversion

B Interpreting the Measures of Risk Aversion
C Risk Premium and Certainty Equivalent
D Assessing the Level of Risk Aversion
E The Concept of Stochastic Dominance
F Mean Preserving Spreads
Measuring Risk Aversion

We’ve already seen that within the von Neumann-Morgenstern

expected utility framework, risk aversion enters through the
concavity of the Bernoulli utility function.
Expected Utility Functions

When u is concave, a payoff of 5 for sure is preferred to a

payoff of 8 with probability 1/2 and 2 with probability 1/2.
Measuring Risk Aversion

We’ve also seen previously that concavity of the utility

function is related to convexity of indifference curves.

In standard microeconomic theory, this feature of preferences

represents a “taste for diversity.”

Under uncertainty, it represents a desire to smooth

consumption across future states of the world.
Expected Utility Functions

A risk averse consumer prefers cA = (cG + cB )/2 in both

states to cG in one state and cB in the other.
Measuring Risk Aversion

Mathematically, u 0 (p) > 0 means that an investor prefers

higher payoffs to lower payoffs, and u 00 (p) < 0 means that the
investor is risk averse.

But is there a way of quantifying an investor’s degree of risk

aversion?

And is there a criterion according to which we might judge one

investor to be more risk averse than another?
Measuring Risk Aversion

Since u 00 (p) < 0 makes an investor risk averse, one conjecture

would be to say that an investor with Bernoulli utility function
v (p) is more risk averse than another investor with Bernoulli
utility function u(p) if v 00 (p) < u 00 (p) for all payoffs p.
Measuring Risk Aversion
Recall, however, that the preference ordering of an investor
with vN-M utility function

U(z) = U(x, y , π) = πu(x) + (1 − π)u(y )

is also represented by the vN-M utility function

V (z) = αU(z) + β = πv (x) + (1 − π)v (y ),

where

v (x) = αu(x) + β and v (y ) = αu(y ) + β.

Measuring Risk Aversion

And with
v (p) = αu(p) + β,
for any payoff p,
v 0 (p) = αu 0 (p)
v 00 (p) = αu 00 (p),

By making α larger or smaller, the Bernoulli utility function

can be made “more” or “less” concave without changing the
underlying preference ordering.
Measuring Risk Aversion

Two alternative measures of risk aversion are

u 00 (Y )
RA (Y ) = − = coefficient of absolute risk aversion
u 0 (Y )

Yu 00 (Y )
RR (Y ) = − 0 = coefficient of relative risk aversion
u (Y )
where Y measures the investor’s income level.

Since v (p) = αu(p) + β implies v 0 (p) = αu 0 (p) and

v 00 (p) = αu 00 (p), these measures are invariant to affine
transformations of the Bernoulli utility function.
Measuring Risk Aversion

Two alternative measures of risk aversion are

u 00 (Y )
RA (Y ) = − = coefficient of absolute risk aversion
u 0 (Y )

Yu 00 (Y )
RR (Y ) = − = coefficient of relative risk aversion
u 0 (Y )
where Y measures the investor’s income level.

And since both measures have a minus sign out in front, both
are positive and increase when risk aversion rises.
Interpreting the Measures of Risk Aversion
To interpret the two measures of risk aversion, it is helpful to
recall from calculus the theorem stated by Brook Taylor
(England, 1685-1731), regarding the approximation of a
function f using its derivatives: the “first-order” approximation

f (x + a) ≈ f (x) + f 0 (x)a

and the “second-order” approximation

1
f (x + a) ≈ f (x) + f 0 (x)a + f 00 (x)a2 .
2
The second-order approximation is more accurate than the
first, and both become more accurate as a becomes smaller.
Interpreting the Measures of Risk Aversion

Focusing first on the measure of absolute risk aversion,

consider an investor with initial income Y who is offered a bet:
win h with probability π and lose h with probability 1 − π.

A risk-averse investor with vN-M expected utility would never

accept this bet if π = 1/2.

The question is: how much higher than 1/2 does π have to be
to get the investor to accept the bet?
Interpreting the Measures of Risk Aversion

Let π ∗ be the probability that is just high enough to get the

investor to accept the bet.

Then π ∗ must satisfy

u(Y ) = π ∗ u(Y + h) + (1 − π ∗ )u(Y − h).

Interpreting the Measures of Risk Aversion

Take second-order Taylor approximations to u(Y + h) and

u(Y − h):

1
u(Y + h) ≈ u(Y ) + u 0 (Y )h + u 00 (Y )h2
2

1
u(Y − h) ≈ u(Y ) − u 0 (Y )h + u 00 (Y )h2
2
Interpreting the Measures of Risk Aversion

u(Y ) = π ∗ u(Y + h) + (1 − π ∗ )u(Y − h)

1
u(Y + h) ≈ u(Y ) + u 0 (Y )h + u 00 (Y )h2
2
1
u(Y − h) ≈ u(Y ) − u 0 (Y )h + u 00 (Y )h2
2
imply
 
∗ 0 1 00 2
u(Y ) ≈ π u(Y ) + u (Y )h + u (Y )h
2
 
∗ 0 1 00 2
+ (1 − π ) u(Y ) − u (Y )h + u (Y )h
2
Interpreting the Measures of Risk Aversion

 
∗ 1 00
0 2
u(Y ) ≈ π u(Y ) + u (Y )h + u (Y )h
2
 
∗ 0 1 00 2
+ (1 − π ) u(Y ) − u (Y )h + u (Y )h
2

implies
1
u(Y ) ≈ u(Y ) + (2π ∗ − 1)u 0 (Y )h + u 00 (Y )h2
2
Interpreting the Measures of Risk Aversion

1
u(Y ) ≈ u(Y ) + (2π ∗ − 1)u 0 (Y )h + u 00 (Y )h2
2
1
0 ≈ (2π ∗ − 1)u 0 (Y )h + u 00 (Y )h2
2
1
0 ≈ (2π ∗ − 1)u 0 (Y ) + u 00 (Y )h
2
1
2π ∗ u 0 (Y ) ≈ u 0 (Y ) − u 00 (Y )h
2
 00 
∗ 1 1 u (Y )
π ≈ + − 0 h
2 4 u (Y )
Interpreting the Measures of Risk Aversion

Since
u 00 (Y )
RA (Y ) = − = coefficient of absolute risk aversion,
u 0 (Y )

it follows from these calculations that

 00 
∗ 1 1 u (Y ) 1 1 1
π ≈ + − 0 h = + hRA (Y ) > .
2 4 u (Y ) 2 4 2

The boost in π above 1/2 required for an investor with income

Y to accept a bet of plus or minus h relates directly to the
coefficient of absolute risk aversion.
Interpreting the Measures of Risk Aversion

As an example, suppose that we ask an investor: What value

of π ∗ would you need to accept a bet of plus-or-minus
h = $1000?

And the investor says: I’ll take it if π ∗ = 0.75.

Interpreting the Measures of Risk Aversion

With h = $1000 and π ∗ = 0.75,

1 1
π∗ ≈ + hRA (Y )
2 4
implies
1000
0.75 ≈ 0.50 + RA (Y )
4
0.25 ≈ 250RA (Y )
0.25
RA (Y ) ≈ = 0.001.
250
Interpreting the Measures of Risk Aversion

Realistically, a bet over $1000 is probably going to seem more

risky to someone who starts out with less income.

In general, we are allowing for that. Since

u 00 (Y )
RA (Y ) = − = coefficient of absolute risk aversion,
u 0 (Y )

it also follows that investors with different income levels

generally display different levels of absolute risk aversion.
Interpreting the Measures of Risk Aversion

Suppose, however, that the Bernoulli utility function takes the

form
1
u(Y ) = − e −νY ,
ν
x
where ν > 0 and e is the exponential function (e ≈ 2.718).

Recall that exponential function has the special property that

f (x) = e x ⇒ f 0 (x) = e x

and by the chain rule

g (x) = e αx ⇒ g 0 (x) = αe αx
Interpreting the Measures of Risk Aversion
With
1
u(Y ) = − e −νY ,
ν
it follows that
1
u 0 (Y ) = − e −νY (−ν) = e −νY
ν
u 00 (Y ) = −νe −νY
u 00 (Y ) νe −νY
RA (Y ) = − = = ν,
u 0 (Y ) e −νY
so that this utility function displays constant absolute risk
aversion, which does not depend on income.
Interpreting the Measures of Risk Aversion

So if we were willing to make the assumption of constant

absolute risk aversion, we could use the results from our
example, where an investor requires π ∗ = 0.75 to accept a bet
with h = $1000 to set ν = 0.001 in
1
u(Y ) = − e −νY ,
ν
and thereby tailor portfolio decisions specifically for this
investor.
Interpreting the Measures of Risk Aversion

Absolute risk aversion describes an investor’s attitude towards

absolute bets of plus or minus h.

A similar analysis shows that relative risk aversion describes an

investor’s attitude towards relative bets of plus or minus kY ,
so that now, k is a fraction of total income.
Interpreting the Measures of Risk Aversion

Consider an investor with initial income Y who is offered a bet:

win kY with probability π and lose kY with probability 1 − π.

A risk-averse investor with vN-M expected utility would never

accept this bet if π = 1/2.

The question is: how much higher than 1/2 does π have to be
to get the investor to accept the bet?
Interpreting the Measures of Risk Aversion

Let π ∗ be the probability that is just high enough to get the

investor to accept the bet.

Now π ∗ must satisfy

u(Y ) = π ∗ u(Y + Yk) + (1 − π ∗ )u(Y − Yk).

Interpreting the Measures of Risk Aversion

Take second-order Taylor approximations to u(Y + Yk) and

u(Y − Yk):

1
u(Y + Yk) ≈ u(Y ) + u 0 (Y )Yk + u 00 (Y )(Yk)2
2

1
u(Y − Yk) ≈ u(Y ) − u 0 (Y )Yk + u 00 (Y )(Yk)2
2
Interpreting the Measures of Risk Aversion

u(Y ) = π ∗ u(Y + Yk) + (1 − π ∗ )u(Y − Yk)

1
u(Y + Yk) ≈ u(Y ) + u 0 (Y )Yk + u 00 (Y )(Yk)2
2
1
u(Y − Yk) ≈ u(Y ) − u 0 (Y )Yk + u 00 (Y )(Yk)2
2
imply
 
∗ 0 1 00 2
u(Y ) ≈ π u(Y ) + u (Y )Yk + u (Y )(Yk)
2
 
∗ 0 1 00 2
+ (1 − π ) u(Y ) − u (Y )Yk + u (Y )(Yk)
2
Interpreting the Measures of Risk Aversion

 
∗ 0 1 00 2
u(Y ) ≈ π u(Y ) + u (Y )Yk + u (Y )(Yk)
2
 
∗ 0 1 00 2
+ (1 − π ) u(Y ) − u (Y )Yk + u (Y )(Yk)
2

implies
1
u(Y ) ≈ u(Y ) + (2π ∗ − 1)u 0 (Y )Yk + u 00 (Y )(Yk)2
2
Interpreting the Measures of Risk Aversion

1
u(Y ) ≈ u(Y ) + (2π ∗ − 1)u 0 (Y )Yk + u 00 (Y )(Yk)2
2
1
0 ≈ (2π ∗ − 1)u 0 (Y )Yk + u 00 (Y )(Yk)2
2
1
0 ≈ (2π ∗ − 1)u 0 (Y ) + u 00 (Y )Yk
2
1
2π ∗ u 0 (Y ) ≈ u 0 (Y ) − u 00 (Y )Yk
2
00
 
∗ 1 1 Yu (Y )
π ≈ + − 0 k
2 4 u (Y )
Interpreting the Measures of Risk Aversion

Since
Yu 00 (Y )
RR (Y ) = − = coefficient of relative risk aversion,
u 0 (Y )

it follows from these calculations that

Yu 00 (Y )
 
∗ 1 1 1 1 1
π ≈ + − 0 k = + kRR (Y ) > .
2 4 u (Y ) 2 4 2

The boost in π above 1/2 required for an investor with income

Y to accept a bet of plus or minus kY relates directly to the
coefficient of relative risk aversion.
Interpreting the Measures of Risk Aversion

Suppose that we ask an investor: What value of π ∗ would you

need to accept a bet of plus-or-minus one percent (k = 0.01)
of your income?

And the investor says: I’ll take it if π ∗ = 0.75.

Interpreting the Measures of Risk Aversion

With k = 0.01 and π ∗ = 0.75,

1 1
π∗ ≈ + kRR (Y )
2 4
implies
0.01
0.75 ≈ 0.50 + RR (Y )
4
0.25 ≈ 0.0025RR (Y )
0.25
RA (Y ) ≈ = 100.
0.0025
Interpreting the Measures of Risk Aversion

Since
Yu 00 (Y )
RR (Y ) = − = coefficient of relative risk aversion,
u 0 (Y )

it also follows that investors with different income levels

generally display different levels of relative risk aversion.

On the other hand, since the coefficient of relative risk

aversion describes aversion to risk over bets that are expressed
relative to income, it is more plausible to assume that
investors have constant relative risk aversion.
Interpreting the Measures of Risk Aversion

Suppose, therefore, that the Bernoulli utility function takes the

form
Y 1−γ − 1
u(Y ) =
1−γ
where γ > 0. For this function, Guillaume de l’Hôpital’s
(France, 1661-1704) rule implies that when γ = 1

Y 1−γ − 1
= ln(Y ),
1−γ
where ln denotes the natural logarithm. This was the form
that Daniel Bernoulli used to describe preferences over payoffs.
Interpreting the Measures of Risk Aversion

To see this, consider Y 1−γ , not as a function of Y but as a

function of γ,
f (γ) = Y 1−γ ,
and take the natural logarithm of both sides to obtain

ln(f (γ)) = (1 − γ) ln(Y ).

Recall, also, that if g (x) = ln(x),

d 1
g 0 (x) = ln(x) = .
dx x
Interpreting the Measures of Risk Aversion
Hence, the chain rule implies that if

ln(f (γ)) = (1 − γ) ln(Y ),

then
d f 0 (γ) − ln(Y ) 1
ln(f (γ)) = = =−
dγ f (γ) (1 − γ) ln(Y ) 1−γ

and hence
1 1
f 0 (γ) = − f (γ) = − (1 − γ) ln(Y ) = − ln(Y )
1−γ 1−γ
Interpreting the Measures of Risk Aversion

Hence, by l’Hôpital’s rule

d
Y 1−γ − 1 dγ
(Y 1−γ − 1)
lim = lim d
γ→1 1−γ γ→1
dγ
(1 − γ)
− ln(Y )
= lim
γ→1 −1
= ln(Y ).
Interpreting the Measures of Risk Aversion

With
Y 1−γ − 1
u(Y ) =
1−γ
it follows that
u 0 (Y ) = Y −γ
u 00 (Y ) = −γY −γ−1
Yu 00 (Y ) Y γY −γ−1
RR (Y ) = − 0 = = γ,
u (Y ) Y −γ
so that this utility function displays constant relative risk
aversion, which does not depend on income.
Interpreting the Measures of Risk Aversion

So if we were willing to make the assumption of constant

relative risk aversion, we could use the results from our
example, where an investor requires π ∗ = 0.75 to accept a bet
with k = 0.01 to set γ = 100 in

Y 1−γ − 1
u(Y ) =
1−γ
and thereby tailor portfolio decisions specifically for this
investor.
Interpreting the Measures of Risk Aversion
Finally, suppose that we do away with the concavity of the
Bernoulli utility function and simply assume that

u(p) = αp + β,

where α > 0, so that higher payoffs are preferred to lower

payoffs. For this utility function,

u 0 (Y ) = α and u 00 (Y ) = 0

u 00 (Y ) Yu 00 (Y )
RA (Y ) = − = 0 and RR (Y ) = − = 0.
u 0 (Y ) u 0 (Y )
This investor is risk-neutral and cares only about expected
payoffs.
Risk Premium and Certainty Equivalent

Our thought experiments so far have asked about how

probabilities need to be boosted in order to induce a
risk-averse investor to accept an absolute or relative bet.

Let’s take step away from gambling and towards investing by

asking: suppose that an investor with income Y has the
opportunity to buy an asset with a payoff Z̃ that is random
and has expected value E (Z̃ ).
Risk Premium and Certainty Equivalent

If this investor is risk-averse and has vN-M expected utility, he

or she will always prefer an alternative asset that pays off
E (Z̃ ) for sure. Mathematically,

u[Y + E (Z̃ )] ≥ E [u(Y + Z̃ )],

“the utility of the expectation is greater than the expectation

of utility.”
Risk Premium and Certainty Equivalent

This follows from a result proven by Johan Jensen (Denmark,

1859-1925).

Theorem (Jensen’s Inequality) Let g be a concave function

and x̃ be a random variable. Then

g [E (x̃)] ≥ E [g (x̃)].

Furthermore, if g is strictly concave and the probability that

x̃ 6= E (x̃) is greater than zero, the inequality is strict.
Risk Premium and Certainty Equivalent

This graph illustrates a special case of Jensen’s inequality. The

result holds much more generally.
Risk Premium and Certainty Equivalent

An implication of Jensen’s inequality is that the maximum

riskless payoff that a risk-averse investor is willing to exchange
for the asset with random payoff Z̃ , called the certainty
equivalent for that asset, will always be less than E (Z̃ ).

Since
u[Y + E (Z̃ )] ≥ E [u(Y + Z̃ )],
the certainty equivalent CE (Z̃ ) defined by

u[Y + CE (Z̃ )] = E [u(Y + Z̃ )]

also satisfies CE (Z̃ ) ≤ E (Z̃ ).

Risk Premium and Certainty Equivalent

Since
u[Y + E (Z̃ )] ≥ E [u(Y + Z̃ )],
the certainty equivalent CE (Z̃ ) defined by

u[Y + CE (Z̃ )] = E [u(Y + Z̃ )]

also satisfies CE (Z̃ ) ≤ E (Z̃ ).

The difference between the higher expected value E (Z̃ ) and

the smaller certainty equivalent CE (Z̃ ) can then be used to
define the positive risk premium Ψ(Z̃ ) for the asset:

Ψ(Z̃ ) = E (Z̃ ) − CE (Z̃ ) ≥ 0.

Risk Premium and Certainty Equivalent

The certainty equivalent and risk premium are “two sides of

the same coin”

Ψ(Z̃ ) = E (Z̃ ) − CE (Z̃ )

CE (Z̃ ) = lesser amount the investor is willing to accept to

remain invested in the risk-free asset

Ψ(Z̃ ) = extra amount the investor needs to take on additional

risk
Risk Premium and Certainty Equivalent

Combining the definitions of the certainty equivalent CE (Z̃ ),

E [u(Y + Z̃ )] = u[Y + CE (Z̃ )],

and the risk premium Ψ(Z̃ ),

CE (Z̃ ) = E (Z̃ ) − Ψ(Z̃ ),

yields
E [u(Y + Z̃ )] = u[Y + E (Z̃ ) − Ψ(Z̃ )],
which we can use to link the risk premium Ψ(Z̃ ) to our
measures of risk aversion.
Risk Premium and Certainty Equivalent

Recall that if the random variable X can take on n possible

values, X1 , X2 , . . . , Xn , with probabilities π1 , π2 , . . . , πn , then
the expected value of X is defined as

E (X ) = π1 X1 + π2 X2 + . . . + πn Xn .

It follows from this definition that the random variable defined

by αX , with α ∈ R, is such that

E (αX ) = π1 αX1 + π2 αX2 + . . . + πn αXn

= α(π1 X1 + π2 X2 + . . . + πn Xn ) = αE (X ).
Risk Premium and Certainty Equivalent
With this fact in mind, return to

E [u(Y + Z̃ )] = u[Y + E (Z̃ ) − Ψ(Z̃ )],

but let

Y ∗ = Y + E (Z̃ )
= income plus expected payout from the risky asset

so that

E {u[Y ∗ + Z̃ − E (Z̃ )]} = u[Y ∗ − Ψ(Z̃ )].

Risk Premium and Certainty Equivalent

Take a second-order Taylor approximation to

u[Y ∗ + Z̃ − E (Z̃ )], viewing Z̃ − E (Z̃ ) as the “size of the bet”

u[Y ∗ + Z̃ − E (Z̃ )] ≈ u(Y ∗ ) + u 0 (Y ∗ )[Z̃ − E (Z̃ )]

1
+ u 00 (Y ∗ )[Z̃ − E (Z̃ )]2 .
2
Risk Premium and Certainty Equivalent

Now take the expected value on both sides and simplify, using
the fact that Y ∗ is not random:

E {u[Y ∗ + Z̃ − E (Z̃ )]} ≈ E [u(Y ∗ )]

+ E {u 0 (Y ∗ )[Z̃ − E (Z̃ )]}
1
+ E { u 00 (Y ∗ )[Z̃ − E (Z̃ )]2 }
2
= u(Y ) + u 0 (Y ∗ )E [Z̃ − E (Z̃ )]
∗

1
+ u 00 (Y ∗ )E {[Z̃ − E (Z̃ )]2 }.
2
Risk Premium and Certainty Equivalent

Finally, use the fact that E [Z̃ − E (Z̃ )] = 0 and the definition
of the variance of Z̃ to simplify further:

E {u[Y ∗ + Z̃ − E (Z̃ )]} ≈ u(Y ∗ ) + u 0 (Y ∗ )E [Z̃ − E (Z̃ )]

1
+ u 00 (Y ∗ )E {[Z̃ − E (Z̃ )]2 }
2
1
= u(Y ∗ ) + σ 2 (Z̃ )u 00 (Y ∗ ).
2
Risk Premium and Certainty Equivalent

On the other side of our original equation, consider a

first-order Taylor approximation to u[Y ∗ − Ψ(Z̃ )]:

u[Y ∗ − Ψ(Z̃ )] ≈ u(Y ∗ ) − u 0 (Y ∗ )Ψ(Z̃ ).

Risk Premium and Certainty Equivalent
Hence, the equation defining the risk premium

E {u[Y ∗ + Z̃ − E (Z̃ )]} = u[Y ∗ − Ψ(Z̃ )],

and the approximations

1
E {u[Y ∗ + Z̃ − E (Z̃ )]} ≈ u(Y ∗ ) + σ 2 (Z̃ )u 00 (Y ∗ )
2

u[Y ∗ − Ψ(Z̃ )] ≈ u(Y ∗ ) − u 0 (Y ∗ )Ψ(Z̃ )

imply
1 2
σ (Z̃ )u 00 (Y ∗ ) ≈ −u 0 (Y ∗ )Ψ(Z̃ ).
2
Risk Premium and Certainty Equivalent

1 2
σ (Z̃ )u 00 (Y ∗ ) ≈ −u 0 (Y ∗ )Ψ(Z̃ )
2
 00 ∗ 
1 2 u (Y )
Ψ(Z̃ ) ≈ σ (Z̃ ) − 0 ∗
2 u (Y )
1 1
Ψ(Z̃ ) ≈ σ 2 (Z̃ )RA (Y ∗ ) = σ 2 (Z̃ )RA (Y + E (Z̃ )),
2 2
indicating that the risk premium depends directly on the
coefficient of absolute risk aversion and the absolute “size of
the bet” σ 2 (Z̃ ).
Risk Premium and Certainty Equivalent
As an example, consider an investor with income Y = 50000
and utility function of the constant relative risk aversion form

Y 1−γ − 1
u(Y ) =
1−γ
with γ = 5, who is considering buying an asset with random
payoff Z̃ that equals 2000 with probability 1/2 and 0 with
probability 1/2. For this asset

E (Z̃ ) = (1/2)2000 + (1/2)0 = 1000

σ 2 (Z̃ ) = (1/2)(2000 − 1000)2 + (1/2)(0 − 1000)2 = 10002 .

Risk Premium and Certainty Equivalent

Our approximation formula

1
Ψ(Z̃ ) ≈ σ 2 (Z̃ )RA (Y + E (Z̃ ))
2
indicates that
 
1 5
Ψ(Z̃ ) ≈ (1000)2 = 49.02
2 51000

since RA (Y ) = RR (Y )/Y .
Risk Premium and Certainty Equivalent

Now go back to the original formula defining the risk premium,

E [u(Y + Z̃ )] = u[Y + E (Z̃ ) − Ψ(Z̃ )],

and plug in the numbers to see that in this case, the exact
value of Ψ(Z̃ ) must satisfy

52000−4 − 1 50000−4 − 1
   
(1/2) + (1/2)
−4 −4
(51000 − Ψ(Z̃ ))−4 − 1
= .
−4
Risk Premium and Certainty Equivalent

52000−4 − 1 50000−4 − 1
   
(1/2) + (1/2)
−4 −4
(51000 − Ψ(Z̃ ))−4 − 1
= .
−4

(1/2)52000−4 + (1/2)50000−4 = (51000 − Ψ(Z̃ ))−4

[(1/2)52000−4 + (1/2)50000−4 ]−1/4 = 51000 − Ψ(Z̃ )
Ψ(Z̃ ) = 51000 − [(1/2)52000−4 + (1/2)50000−4 ]−1/4
Ψ(Z̃ ) = 48.97.
Risk Premium and Certainty Equivalent

The approximation Ψ(Z̃ ) ≈ 49.02 or the exact solution

Ψ(Z̃ ) = 48.97 imply that an investor with Y = 50000 and
constant coefficient of relative risk aversion equal to 5 will give
up a riskless payoff of up to about

CE (Z̃ ) = E (Z̃ ) − Ψ(Z̃ ) ≈ 1000 − 49 = 951

for this risky asset with expected payoff equal to 1000.

Assessing the Level of Risk Aversion

We can use similar calculations to work through thought

experiments that shed light on our own levels of risk aversion.

Suppose your income is Y = 50000 and you have the chance

to buy an asset that pays 50000 with probability 1/2 and 0
with probability 1/2.

This asset has E (Z̃ ) = (1/2)50000 + (1/2)0 = 25000, but

what is the maximum riskless payoff you would exchange for
it?
Assessing the Level of Risk Aversion

Suppose your utility function is of the constant relative risk

aversion form
Y 1−γ − 1
u(Y ) =
1−γ
and recall that the most you should pay for the asset is given
by the certainty equivalent CE (Z̃ ) defined by

E [u(Y + Z̃ )] = u[Y + CE (Z̃ )].

Assessing the Level of Risk Aversion

E [u(Y + Z̃ )] = u[Y + CE (Z̃ )]

1000001−γ − 1 500001−γ − 1
   
(1/2) + (1/2)
1−γ 1−γ
(50000 + CE (Z̃ ))1−γ − 1
=
1−γ

CE (Z̃ ) = [(1/2)1000001−γ + (1/2)500001−γ ]1/(1−γ) − 50000

Assessing the Level of Risk Aversion
Certainty equivalent for an asset that pays 50000 with
probability 1/2 and 0 with probability 1/2 when income is
50000 and the coefficient of relative risk aversion is γ.

γ CE (Z̃ )
0 25000 (risk neutrality)
1 20711 (logarithmic utility, proposed by D Bernoulli)
2 16667
3 13246
4 10571
5 8566
10 3991
20 1858
50 712
The Concept of Stochastic Dominance

It is important to recognize that the coefficients of absolute

and relative risk aversion, RA (Y ) and RR (Y ), and the
certainty equivalent CE (Z̃ ) and the risk premium Ψ(Z̃ ), all
help describe or summarize investors’ preferences over risky
cash flows.

They do not directly represent differences in market or

equilibrium prices or rates of return across riskless and risky
assets.
The Concept of Stochastic Dominance

Since individuals will differ in their attitudes towards risk as in

their preferences over everything else (“chacun à son goût,” as
the French say), it is useful to ask whether there are properties
of payoff distributions that will allow “preference-free”
comparisons to be made across risky cash flows.

State-by-state dominance, as we’ve already seen, is one such

property. But are there any others, which might be more
widely applicable?
The Concept of Stochastic Dominance

Consider two assets, with random payoffs Z1 and Z2 :

Payoffs 10 100 1000

Probabilities for Z1 0.40 0.60 0.00
Probabilities for Z2 0.40 0.40 0.20

There may be no state-by-state dominance, if the payoffs

Z1 = 10 and Z2 = 100 can occur together and the payoffs
Z1 = 100 and Z2 = 10 can occur together.
The Concept of Stochastic Dominance

Payoffs 10 100 1000

Probabilities for Z1 0.40 0.60 0.00
Probabilities for Z2 0.40 0.40 0.20

Because E (Z1 ) = 64, σ(Z1 ) = 44, E (Z2 ) = 244, and

σ(Z2 ) = 380, there is no mean-variance dominance either.
The Concept of Stochastic Dominance

Payoffs 10 100 1000

Probabilities for Z1 0.40 0.60 0.00
Probabilities for Z2 0.40 0.40 0.20

But, intuitively, asset 2 “looks” better, because its distribution

takes some of the probability of a payoff of 100 and “moves”
that probability to the even higher payoff of 1000.

We can make this idea more concrete by looking at the

distributions of these random payoffs in a different way.
The Concept of Stochastic Dominance

In probability theory, the cumulative distribution function (cdf)

for a random variable X keeps track of the probability that the
realized value of X will be less than or equal to x:

F (x) = Prob(X ≤ x).

The Concept of Stochastic Dominance

Payoffs 10 100 1000

Probabilities for Z1 0.40 0.60 0.00
Probabilities for Z2 0.40 0.40 0.20

cdfs x < 10 10 ≤ x < 100 100 ≤ x < 1000 1000 ≤ x

F1 (x) 0.00 0.40 1.00 1.00
F2 (x) 0.00 0.40 0.80 1.00
The Concept of Stochastic Dominance

Cumulative distribution functions are always nondecreasing.

The Concept of Stochastic Dominance

Cumulative distribution functions always satisfy F (−∞) = 0,

F (∞) = 1 and 0 ≤ F (x) ≤ 1.
The Concept of Stochastic Dominance

Cumulative distribution functions are always càdlàg (“continue

à driote, limite à gauche”) or RCLL “right continuous with left
limits.”
The Concept of Stochastic Dominance

The fact that F2 (x) always lies below F1 (x) formalizes the
first-order stochastic dominance of Z2 over Z1 .
The Concept of Stochastic Dominance

cdfs x < 10 10 ≤ x < 100 100 ≤ x < 1000 1000 ≤ x

F1 (x) 0.00 0.40 1.00 1.00
F2 (x) 0.00 0.40 0.80 1.00

Asset 2 displays first-order stochastic dominance over asset 1

because F2 (x) ≤ F1 (x) for all possible values of x.
The Concept of Stochastic Dominance

Theorem Let F1 (x) and F2 (x) be the cumulative distribution

functions for two assets with random payoffs Z1 and Z2 . Then

F2 (x) ≤ F1 (x) for all x,

that is, asset 2 displays first-order stochastic dominance over

asset 1, if and only if

E [u(Z2 )] ≥ E [u(Z1 )]

for any nondecreasing Bernoulli utility function u.

The Concept of Stochastic Dominance

First-order stochastic dominance is a weaker condition than

state-by-state dominance, in that state-by-state dominance
implies first-order stochastic dominance but first-order
stochastic dominance does not necessarily imply state-by-state
dominance.

But first-order stochastic dominance remains quite a strong

condition. Since an asset that displays first-order stochastic
dominance over all others will be preferred by any investor
with vN-M utility who prefers higher payoffs to lower payoffs,
the price of such an asset is likely to be bid up until the
dominance goes away.
The Concept of Stochastic Dominance

Consider two more assets, with random payoffs Z3 and Z3 :

Payoffs 1 4 5 6 8 9
Probabilities for Z3 0.33 0.00 0.00 0.33 0.33 0.00
Probabilities for Z4 0.00 0.25 0.50 0.00 0.00 0.25

Asset 4 looks at least slightly better, since it always pays off at

least 4 and has a non-trivial probability of 9. On the other
hand, asset 3 has a higher probability of a payoff of 6 or more.
The Concept of Stochastic Dominance

Payoffs 1 4 5 6 8 9
Probabilities for Z3 0.33 0.00 0.00 0.33 0.33 0.00
Probabilities for Z4 0.00 0.25 0.50 0.00 0.00 0.25

cdfs x <1 1≤x <4 4≤x <5 5≤x <6

F3 (x) 0.00 0.33 0.33 0.33
F4 (x) 0.00 0.00 0.25 0.75

cdfs 6≤x <8 8≤x <9 9≤x

F3 (x) 0.66 1.00 1.00
F4 (x) 0.75 0.75 1.00
The Concept of Stochastic Dominance

There is no first-order stochastic dominance, since

F4 (x) ≤ F3 (x) for some values of x but F3 (x) ≤ F4 (x) for
other values of x.
The Concept of Stochastic Dominance

Still, as we move from left to right . . .

The Concept of Stochastic Dominance

. . . the areas over which F4 (x) ≤ F3 (x) . . .

The Concept of Stochastic Dominance

. . . seem to be consistently larger than the areas over which

F3 (x) ≤ F4 (x).
The Concept of Stochastic Dominance

Thus, asset 4 displays second-order stochastic dominance over

asset 3.
The Concept of Stochastic Dominance

cdfs x <1 1≤x <4 4≤x <5 5≤x <6

F3 (x) 0.00 0.33 0.33 0.33
F4 (x) 0.00 0.00 0.25 0.75

cdfs 6≤x <8 8≤x <9 9≤x

F3 (x) 0.66 1.00 1.00
F4 (x) 0.75 0.75 1.00

Asset 4 displays second-order stochastic dominance over asset

4 since Z x̄
[F4 (x) − F3 (x)]dx ≤ 0
−∞

for all possible values of x̄.

The Concept of Stochastic Dominance

Asset 4 displays second-order stochastic dominance over asset

4 since Z x̄
[F4 (x) − F3 (x)]dx ≤ 0
−∞

for all possible values of x̄.

Note the integral in the definition runs from −∞ up to x̄,

since we want to penalize assets with higher probabilities of
lower payoffs (we moved from left to right in the graphs).
Note, also, that the condition must hold for all values of x̄.
The Concept of Stochastic Dominance

Theorem Let F3 (x) and F4 (x) be the cumulative distribution

functions for two assets with random payoffs Z3 and Z4 . Then
Z x̄
[F4 (x) − F3 (x)]dx ≤ 0 for all x̄
−∞

that is, asset 4 displays second-order stochastic dominance

over asset 3, if and only if

E [u(Z4 )] ≥ E [u(Z3 )]

for any nondecreasing and concave Bernoulli utility function u.

The Concept of Stochastic Dominance

Second-order stochastic dominance is a weaker condition than

first-order stochastic dominance, in that first-order stochastic
dominance implies second-order stochastic dominance but
second-order stochastic dominance does not necessarily imply
first-order stochastic dominance.

But second-order stochastic dominance remains a strong

condition. Since an asset that displays second-order stochastic
dominance over all others will be preferred by any risk-averse
investor with vN-M utility, the price of such an asset is likely
to be bid up until the dominance goes away.
Mean Preserving Spreads

Comparisons based on state-by-state dominance, first-order

stochastic dominance, and second-order stochastic dominance
can reflect differences in the mean, or expected, payoff as well
as in the standard deviation or variance of the payoff.

It is also useful, therefore, to consider an alternative criterion

that focuses entirely on the standard deviation of a random
payoff, as a measure of the riskiness of the corresponding
asset, holding the mean or expected value fixed.
Mean Preserving Spreads

Graphically, a mean preserving spread takes probability from

the center of a distribution and shifts it to the tails.
Mean Preserving Spreads

Mathematically, one way of producing a mean preserving

spread is to take one random variable X1 and define a second,
X2 , by adding “noise” in the form of a third, zero-mean
random variable Z :
X2 = X1 + Z ,
where E (Z ) = 0.
Mean Preserving Spreads
As an example, suppose that

5 with probability 1/2
X1 =
2 with probability 1/2

+1 with probability 1/2
Z=
−1 with probability 1/2
then 

 6 with probability 1/4
4 with probability 1/4

X2 = X1 + Z =

 3 with probability 1/4
1 with probability 1/4


E (X1 ) = E (X2 ) = 3.5, but if these are random payoffs, asset 2

seems riskier.
Mean Preserving Spreads

The following theorem relates the concept of a mean

preserving spread to the previous concept of second-order
stochastic dominance.

Theorem Let X1 and X2 , with E (X1 ) = E (X2 ), be random

payoffs on two assets. Then the following two statements are
equivalent: (i) X2 = X1 + Z for some random variable Z with
E (Z ) = 0 and (ii) asset 1 displays second-order stochastic
dominance over asset 2.
Mean Preserving Spreads

Theorem Consider two assets with random payoffs Z1 and Z2 .

Asset 2 displays second-order stochastic dominance over asset
1 if and only if
E [u(Z2 )] ≥ E [u(Z1 )]
for any nondecreasing and concave Bernoulli utility function u.

This theorem imply that any risk-averse investor with vN-M

preferences will avoid “pure gambles,” in the form of assets
with payoffs that simply add more randomness to the payoff of
another asset.