14.

13 Economics and Psychology

(Lecture 5)

Xavier Gabaix

February 19, 2003

Second order risk aversion for EU

The agent takes the 50/50 gamble + , i:
1
1
B () = u (x + + ) + u (x + ) u (x)
2
2
i.e. where:
B () = u (x)

Assume that u is twice dierentiable and take a look at the Taylor

expansion of the above equality for small . .
h
i

1 0
1 00
2
2
2
2
B () = u (x)+ u (x) 2+ u (x) 2 + +o + = u (x)
2
4

then
i

h 2
2
2
2
+ +o +
=
2
00

where = uu0

h
i

2
2
To solve : = 2 + for small . Call 0 = /2.

 Barbarian way : Solve:

1
2
0 + 2 = 0

Exactly. Then take Taylor. One finds:

= 0 2 = 2
2

h
i
0
2
2
Elegant way: = + for small .

will be small. Take a guess. If the expansion is = k, then

we get:
h
i
0
2
2
2
k = + k
h
i
0
2
k = 1+k

contraction for 0, the RHS goes to 0 and the LHS is k. This

guess doesnt work.
Lets try instead = k 2. Then:
h
i
2
0
2
2
4
k = + k

0
2
= + o 2

k = 0 + o(1) after dividing both side by 2

that works, with k = 0. Conclusion:

= 2.
2
Note this method is really useful when the equation to solve doesnt
have a closed form solution. For example, solve for small
= 0( 2 + 2 + 7)
solution postulate = k 2, plug it back in the equation to solve, then
take 0 and it works for k = 0
The 2 indicates second order risk aversion.

First order risk aversion of PT

Consider same gamble as for EU. Take the gamble i where
(.5)u( + ) + (.5)u( ) = 0
We will show that in PT, as 0, the risk premium is of the order
of when reference wealth x = 0. This is called the first order risk
aversion.
Lets compute for u (x) = x for x 0 and u (x) = |x| for
x 0.

 The premium at x = 0 satisfies

1
1
0 = ( ) ( + ) + ( ) () | + |
2
2
cancel ( 12 ) and use the fact that + < 0 to get
0 = ( + ) ( )

( + ) = ( )
+ = 1/ [ ]

then
=
where k is defined appropriately.

1
+1

= k

 Empirically:
= 2, ' 1
21
1
k'
=
2+1
3
Note that when = 1, the agent is risk neutral and the risk premium
is 0.

2.1

Calibration 1

Consider an EU agent with a constant elasticity of substitution, CES,

1
utility, i.e. u (c) = c1 .
Gamble 1
$50,000 with probability 1/2
$100,000 with probability 1/2
Gamble 2. $x for sure.
Typical x that makes people indierent between the two gambles belongs to (60k, 75k) (though some people are risk loving and ask for
higher x).

 If x = 65k, what is
.5 u(W + 50) + .5 u(W + 100) = u(W + x)
.5 W 1 501 + .5 W 1 1001 = W 1 x1
5 501 + .5 1001 = x1

Note the relation between x and the elasticity of substitution :

x 75k 70k 63k 58k 54k 51.9k 51.2k
0
1
3
5
10
20
30
Right seems to be between 1 and 10.

Evidence on financial markets calls for bigger than 10. This is the
equity premium puzzle.

2.2

Calibration 2

Gamble 1
$11 with probability 1/2
$-10 with probability 1/2

Gamble 2. Get $0 for sure.

If someone prefers Gamble 2, she or he satisfies

1
1
u (W ) > u (W + ) + u (W + + ) .
2
2
Here, = $.5 and = $10.5. We know that in EU

< = 2
2

u00(W )
W 1

And thus with CES utility = u0(W ) = W = W

2
2W
< 2 =
<

2
2
2W

forces large as the wealth W is larger than 105 easily.

Here:

2W
2 105 .5
3
>
=

10
2
10.52

Conclusion: very hard to calibrate the same model to large and small
gambles using EU.

2.3

Calibration Conclusions

What would a PT agent do? If = 1, = 2, in calibration 2 he wont

take gamble 1 as
(.5)11 + (.5)( 10) = 9(.5) < 0
In PT we have = k. For W = 104, = 2, and = 0.5 the risk
2
$.00002
premium is = k = 13 .5 $.2 while in EU = 2W
If we want to fit an EU parameter to a PT agent we get
P T () = EU ()
2
k =

2W

then

=
and this explodes as 0.

2kW

 If someone is averse to 50-50 lose $100/gain g for all wealth levels

then he or she will turn down 50-50 lose L-gain G in the table

Guess:

L\g
$101 $105 $110 $125
$400
$400 $420 $550
$800
$800
$1000 $1, 010
$2000
$10, 000

L\g
$101
$105
$110
$125
$400
$400
$420
$550 $1, 250
$800
$800 $1, 050 $2, 090

$1000 $1, 010 $1, 570

$2000 $2, 320

$10, 000

cf paper by Matt Rabin

2.4

What does it mean?

EU is still good for modelling.

Even behavioral economists stick to it when they are not interested in
risk taking behavior, but in fairness for example.

The reason is that EU is nice, simple, and parsimonious.

Two extensions of PT
Both outcomes, x and y, are positive, 0 < y < x. Then,
V = v (y) + (p) (v (x) v (y)) .
Why not V = (p) v (x) + (1 p) v (y)? Because it becomes selfcontradictory when x = y and we stick to K-T calibration that puts
(.5) < .5.

 Continuous gambles, distribution f (x)

EU gives:
V =

Z +

u (x) f (x) dx

PT gives:
V =
+

Z +

0
Z 0

u (x) f (x) 0 (P (g x)) dx

u (x) f (x) 0 (P (g x)) dx