Professional Documents
Culture Documents
5 Mean-Variance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7 Representative Investors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Part IV Topics
Single-Period Models
1
1.1. Calculate the risk tolerance of each of the five special utility functions in Section 1.7 to verify
the formulas given in the text.
u0 (w) 1
u(w) = e ↵w
) u0 (w) = ↵e ↵w
, u00 (w) = ↵2 e ↵w
, 00
= .
u (w) ↵
1 1 u0 (w)
u(w) = log w ) u0 (w) = , u00 (w) = , = w.
w w2u00 (w)
1 u0 (w) w
u(w) = w1 ⇢
) u0 (w) = w ⇢ , u00 (w) = ⇢w ⇢ 1 , 00
= .
1 ⇢ u (w) ⇢
1 1 0
u (w)
u(w) = log(w ⇣) ) u0 (w) = , u00 (w) = , =w ⇣.
w ⇣ (w ⇣)2 u00 (w)
✓ ◆1 ⇢ ✓ ◆ ⇢ ✓ ◆ ⇢ 1
⇢ w ⇣ 0 w ⇣ 00 w ⇣
u(w) = ) u (w) = , u (w) = ,
1 ⇢ ⇢ ⇢ ⇢
u0 (w) w ⇣
00
= .
u (w) ⇢
1.2. Let "˜ be a random variable with zero mean and variance equal to 1. Let ⇡( ) be the risk
premium for the gamble "˜ at wealth w, meaning
⇥ ⇤
u0 (w ⇡( ))⇡ 0 ( ) = E u0 (w + "˜)˜
" .
At = 0, this implies
u0 (w)⇡ 0 (0) = u0 (w)E[˜
"] = 0 .
⇥ ⇤
u00 (w ⇡( ))[⇡ 0 ( )]2 u0 (w ⇡( ))⇡ 00 ( ) = E u00 (w + "˜)˜
"2 .
At = 0, we obtain
u0 (w)⇡ 00 (0) = u00 (w)E[˜
"2 ] = u00 (w) .
1.3. Consider the five special utility functions in Section 1.7 (the utility functions with linear risk
tolerance). Which of these utility functions, for some parameter values, have decreasing absolute
risk aversion and increasing relative risk aversion? Which of these utility functions are monotone
increasing and bounded on the domain w 0?
Shifted log and shifted power utility functions have absolute risk aversion ⇢/(w ⇣) and relative
risk aversion ⇢w/(w ⇣), with ⇢ = 1 being shifted log. Absolute risk aversion is decreasing in w.
Relative risk aversion is increasing in w if ⇣ < 0. The shifted power utility function is monotone
increasing and bounded on the domain w 0 if ⇣ < 0 and ⇢ > 1.
(a) Verify that the fraction of wealth he will pay to avoid a gamble that is proportional to wealth
is independent of initial wealth (i.e., show that ⇡ defined in (1.10) is independent of w for
logarithmic and power utility).
(b) Consider a gamble "˜. Assume 1 + "˜ is lognormally distributed; specifically, assume 1 + "˜ = ez̃ ,
where z̃ is normally distributed with variance 2 and mean 2 /2. Note that by the rule for
means of exponentials of normals, E[˜
"] = 0. Show that ⇡ defined in (1.10) equals
⇢ 2 /2
1 e .
1 Utility Functions and Risk Aversion Coefficients 5
Note: This is consistent with the approximation (1.4), because a first-order Taylor series expan-
sion of the exponential function ex around x = 0 shows that ex ⇡ 1 + x when |x| is small.
so (1.10) is equivalent to
Hence, ⇡ does not depend on w. For power utility, the left-hand side of (1.10) is
1 1
((1 ⇡)w)1 ⇢
= w1 ⇢
(1 ⇡)1 ⇢
,
1 ⇢ 1 ⇢
so (1.10) is equivalent to
⇥ ⇤ ⇥ ⇤ 1
(1 ⇡)1 ⇢
= E (1 + "˜)1 ⇢
, ⇡=1 E (1 + "˜)1 ⇢ 1 ⇢
.
2 /2
⇡=1 e .
For ⇢ 6= 1,
⇥ ⇤ h i 2 /2+(1 ⇢)2 2 /2 2 /2
E (1 + "˜)1 ⇢
= E e(1 ⇢)z̃
=e (1 ⇢)
=e ⇢(1 ⇢)
.
⇢ 2 /2
⇡=1 e .
6 1 Utility Functions and Risk Aversion Coefficients
(a) Suppose the person has wealth of $100,000 and faces a gamble in which he wins or loses x with
equal probabilities. Calculate the amount he would pay to avoid the gamble, for various values
of ⇢ (say, between 0.5 and 40), and for x = $100, x = $1,000, x = $10,000, and x = $25,000.
For large gambles, do large values of ⇢ seem reasonable? What about small gambles?
(b) Suppose ⇢ > 1 and the person has wealth w. Suppose he is o↵ered a gamble in which he loses
x or wins y with equal probabilities. Show that he will reject the gamble no matter how large
y is if
x log(0.5) + log(1 x/w)
1 0.51/(⇢ 1)
, ⇢ .
w log(1 x/w)
For example, if w is $100,000, then the person would reject a gamble in which he loses $10,000
or wins 1 trillion dollars with equal probabilities when ⇢ satisfies this inequality for x/w = 0.1.
What values of ⇢ (if any) seem reasonable?
(a) Set " = x/100000. From the solution to the previous exercise, the amount the person would pay
is 100000⇡, where, for log utility,
For the largest gamble, ⇢ > 5 (or, perhaps ⇢ > 2) would seem unreasonable. But, for ⇢ 5, the
premium for the $100 gamble is $0.25 or less, which may be too small.
(b) Given 1 ⇢ < 0, the person rejects the gamble if
w1 ⇢
< 0.5(w x)1 ⇢
+ 0.5(w + y)1 ⇢
.
1.6. This exercise is a very simple version of a model of the bid-ask spread presented by Stoll
(1978).
Consider a person with constant absolute risk aversion ↵. Starting from a random wealth w̃,
(a) Compute the maximum amount the person would pay to obtain a random payo↵ x̃; i.e., compute
BID satisfying
E[u(w̃)] = E[u(w̃ + x̃ BID)] .
(b) Compute the minimum amount the person would require to accept the payo↵ x̃; i.e., compute
ASK satisfying
E[u(w̃)] = E[u(w̃ x̃ + ASK)] .
8 1 Utility Functions and Risk Aversion Coefficients
Note that ✓ ◆
1
E[u(w̃)] = exp ↵E[w̃] + ↵2 var(w̃) .
2
(a) We have
✓ ◆
1
E[u(w̃ + x̃ BID)] = exp ↵E[w̃] ↵E[x̃] + ↵BID + ↵2 [var(w̃) + 2 cov(x̃, w̃) + var(x̃)] .
2
This implies
1
BID = E[x̃] ↵ cov(x̃, w̃) ↵ var(x̃) .
2
(b) We have
✓ ◆
1
E[u(w̃ x̃ + ASK)] = exp ↵E[w̃] + ↵E[x̃] ↵ASK + ↵2 [var(w̃) 2 cov(x̃, w̃) + var(x̃)] .
2
This implies
1
ASK = E[x̃] ↵ cov(x̃, w̃) + ↵ var(x̃) .
2
1.7. Show that condition (ii) in the discussion of second-order stochastic dominance in the end-
of-chapter notes implies condition (i); i.e., assume ỹ = x̃ + z̃ + "˜ where z̃ is a nonpositive random
"|x̃ + z̃] = 0 and show that E[u(x̃)]
variable and E[˜ E[u(ỹ)] for every monotone concave function u.
Note: The statement of (ii) is that ỹ has the same distribution as x̃ + z̃ + "˜, which is a weaker
condition than ỹ = x̃ + z̃ + "˜, but if ỹ has the same distribution as x̃ + z̃ + "˜ and ỹ 0 = x̃ + z̃ + "˜,
then E[u(ỹ)] = E[u(ỹ 0 )] so one can without loss of generality take ỹ = x̃ + z̃ + "˜ (though this is not
true for the reverse implication (i) ) (ii)).
ỹ equals x̃ + z̃ plus mean-independent noise, so by concavity and Jensen’s inequality, E[u(x̃ + z̃)]
E[u(ỹ)], as shown in Section 1.8. Because z̃ is nonpositive and u is monotone, E[u(x̃)] E[u(x̃ + z̃)].
Therefore, E[u(x̃)] E[u(ỹ)].
1 Utility Functions and Risk Aversion Coefficients 9
"|ỹ]] = 0 .
E[ỹ "˜] = E[ỹE[˜
Furthermore,
E[˜ "|ỹ]] = 0 .
"] = E[E[˜
Therefore,
cov(ỹ, "˜) = E[ỹ "˜] E[ỹ]E[˜
"] = 0 .
1.9. Show that any monotone utility function with linear risk tolerance is a monotone affine
transform of one of the five utility functions: negative exponential, log, power, shifted log, or shifted
power. Hint: Consider first the special cases (i) risk tolerance = A and (ii) risk tolerance = Bw. In
case (i) use the fact that
u00 (w) d log u0 (w)
=
u0 (w) dw
and in case (ii) use the fact that
wu00 (w) d log u0 (w)
=
u0 (w) d log w
to derive formulas for log u0 (w) and hence u0 (w) and hence u(w). For the case A 6= 0 and B 6= 0,
define ✓ ◆
w A
v(w) = u ,
B
show that the risk tolerance of v is Bw, apply the results from case (ii) to v, and then derive the
form of u.
Hence,
10 1 Utility Functions and Risk Aversion Coefficients
This implies
Z w
u(w) = u(y) + u0 (x) dx
y
Z w
= u(y) + u0 (y)e↵y e ↵x
dx
y
1⇥ ⇤
= u(y) + u0 (y)e↵y e ↵y
e ↵w
.
↵
This is an affine transform of e ↵w . For u to be monotone, it must be a monotone affine transform
of e ↵w .
In case (ii), set ⇢ = 1/B. For any constant y > 0 and any w > 0,
Z w
0 0 d log u0 (x)
log u (w) = log u (y) + d log x
y d log x
Z w
= log u0 (y) ⇢ d log x
y
Hence,
u0 (w) = u0 (y)e ⇢(log w log y)
= u0 (y)y ⇢ w ⇢
.
This implies
Z w
u(w) = u(y) + u0 (x) dx
y
Z w
0 ⇢ ⇢
= u(y) + u (y)y x dx .
y
If ⇢ = 1, then
u(w) = u(y) + u0 (y)y(log w log y) ,
1
u(w) = u(y) + u0 (y)y ⇢ w1 ⇢
y1 ⇢
.
1 ⇢
which is an affine transform of w1 ⇢ /(1 ⇢). For u to be monotone, it must be a monotone affine
transform of w1 ⇢ /(1 ⇢).
For the case A 6= 0 and B 6= 0, set
✓ ◆
x A
v(x) = u
B
1 Utility Functions and Risk Aversion Coefficients 11
Therefore, from case (ii), on the region x > 0, either v(x) = log x if B = 1, or v(x) = x1 ⇢ /(1 ⇢)
for ⇢ = 1/B, up to an affine transform. Moreover,
Hence, for w such that A + Bw > 0, either u(w) = log(A + Bw) if B = 1, or u(w) = (A +
Bw)1 ⇢ /(1 ⇢) for ⇢ = 1/B, up to an affine transform. Setting ⇣ = A/B, we have, up to an affine
transform, u(w) = log(w ⇣) on the region w > ⇣ if B = 1, or
✓ ◆
1 w ⇣ 1 ⇢
u(w) = ,
1 ⇢ ⇢
on the region (w ⇣)/⇢ > 0. Monotonicity of u in the case B 6= 1 requires that u be a monotone
affine transform of ✓ ◆1 ⇢
⇢ w ⇣
.
1 ⇢ ⇢
1.10. Suppose a person has log utility: u(w) = log w for each w > 0.
n
(a) Consider flipping a sequence of coins and having wealth e2 if the first heads appears on the
n–th toss. The probability of the first heads appearing on the n–th toss is 2 n, so the expected
utility is
1
X 1
X
n n
2 log e2 = 1 = 1.
n=1 n=1
(b) Consider flipping coins and having wealth e 2n if the first heads appears on the n–th toss. The
expected utility is
12 1 Utility Functions and Risk Aversion Coefficients
1
X 1
X
n 2n
2 log e = 1= 1.
n=1 n=1
(c) Obviously, there are many ways to do this. Here is one. Let 0 < < 1 be such that
1
X
n 2n
w> 2 e .
1+
n=1
For n = 1, 2, . . ., let
2n
"˜ = e w
with probability p2 n.
Then w + "˜ > 0 in each state of the world, and
1
! "1 #
X X
n 2n n
E[˜
"] = (1 p) w 2 e +p 2 n e 2 w
n=1 n=1
1
!
X
n 2n
= [(1 p) p] w 2 e
n=1
= 0.
Moreover,
1
! " 1
#
X X
n 2n n 2n
E[u(w + "˜)] = (1 p) log (1 + )w 2 e +p 2 log e = 1.
n=1 n=1
1.11. Show that risk neutrality [u(w) = w for all w] can be regarded as a limiting case of negative
exponential utility as ↵ ! 0 by showing that there are monotone affine transforms of negative
exponential utility that converges to w as ↵ ! 0. Hint: Take an exact first-order Taylor series
expansion of negative exponential utility, expanding in ↵ around ↵ = 0. Writing the expansion as
c0 + c1 ↵, show that
e ↵w c0
!w
↵
as ↵ ! 0.
1 Utility Functions and Risk Aversion Coefficients 13
↵w ↵w
ˆ
e = 1 + we ↵.
This implies
e ↵w +1 ↵w
ˆ
= we ! w,
↵
as ↵ ! 0.
1.12. The notation and concepts in this exercise are from Appendix A. Suppose there are three
possible states of the world which are equally likely, so ⌦ = {!1 , !2 , !3 } with P({!1 }) = P({!2 }) =
P({!3 }) = 1/3. Let G be the collection of all subsets of ⌦:
Let x̃ and ỹ be random variables, and set ai = x̃(!i ) for i = 1, 2, 3. Suppose ỹ(!1 ) = b1 and
ỹ(!2 ) = ỹ(!3 ) = b2 6= b1 .
(a)
prob(x̃ = a1 | ỹ = b1 ) = 1 ,
prob(x̃ = a2 | ỹ = b1 ) = 0 ,
prob(x̃ = a3 | ỹ = b1 ) = 0 ,
prob(x̃ = a1 | ỹ = b2 ) = 0 ,
prob(x̃ = a2 | ỹ = b2 ) = 1/2 ,
prob(x̃ = a3 | ỹ = b2 ) = 1/2 .
(b)
E[x̃ | ỹ = b1 ] = a1 ,
1.13. Let ỹ = ex̃ , where x̃ is normally distributed with mean µ and variance 2. Show that
stdev(ỹ) p 2
= e 1.
E[ỹ]
2 /2
We have E[ỹ] = eµ+ , and
so
p
2
stdev(ỹ) = E[ỹ] e 1.
2
2.1. Consider the portfolio choice problem of a CARA investor with n risky assets having normally
distributed returns studied in Section 2.5, but suppose there is no risk-free asset, so the budget
constraint is 10 = w0 . Show that the optimal portfolio is
✓ ◆
1 1 ↵w0 10 ⌃ 1 µ 1
= ⌃ µ+ ⌃ 1.
↵ ↵10 ⌃ 1 1
Note: As will be seen in Section 5.1, the two vectors ⌃ 1µ and ⌃ 11 play an important role in
mean-variance analysis even without the CARA/normal assumption.
0 1 0
µ ↵ ⌃ .
2
Let denote the Lagrange multiplier for the constraint 10 = w0 . The Lagrangean is
0 1 0
(µ 1) ↵ ⌃ ,
2
which is solved by
1 1
= ⌃ (µ 1) .
↵
Imposing the constraint 10 = w0 yields
16 2 Portfolio Choice and Stochastic Discount Factors
1 0
1⌃ 1
µ 10 ⌃ 1
1 = w0 .
↵ ↵
Therefore,
10 ⌃ 1 µ ↵w0
= ,
10 ⌃ 1 1
and ✓ ◆
1 1 ↵w0 10 ⌃ 1 µ 1
= ⌃ µ+ ⌃ 1.
↵ ↵10 ⌃ 1 1
2.2. Suppose there is a risk-free asset and n risky assets. Adopt the notation of Section 2.5, but do
not assume the risky asset returns are normally distributed. Consider an investor with quadratic
utility who seeks to maximize
1 1
⇣E[w̃] E[w̃]2 var(w̃) .
2 2
Show that the optimal portfolio for the investor is
1 1
= (⇣ w0 Rf )⌃ (µ Rf 1) ,
1 + 2
where
2 = (µ Rf 1)0 ⌃ 1
(µ Rf 1) .
It is shown in Chapter 5 that is the maximum Sharpe ratio of any portfolio. Hint: In the first-order
conditions, define = (µ Rf 1)0 , solve for in terms of , and then compute .
The expected payo↵ of a portfolio of risky assets is w0 Rf + 0 (µ Rf 1), and the variance is 0⌃ .
The expected utility is
0 1 1 0
⇣[w0 Rf + (µ Rf 1)] [w0 Rf + 0 (µ Rf 1)]2 ⌃
2 2
0 1 2 2 1 0 1 0
= ⇣[w0 Rf + (µ Rf 1)] w0 Rf w0 Rf 0 (µ Rf 1) (µ Rf 1)(µ Rf 1)0 ⌃ .
2 2 2
⇣(µ Rf 1) w0 Rf (µ Rf 1) (µ Rf 1) ⌃ = 0,
with solution
2 Portfolio Choice and Stochastic Discount Factors 17
1
= (⇣ w 0 Rf )⌃ (µ Rf 1) .
Thus,
= (⇣ w 0 Rf )(µ Rf 1)0 ⌃ 1
(µ Rf 1)
= (⇣ w 0 Rf )2 ,
implying
2
= (⇣ w 0 Rf ) ,
1 + 2
and
1 1
= (⇣ w0 Rf )⌃ (µ Rf 1) .
1 + 2
2.3. This exercise provides another illustration of the absence of wealth e↵ects for CARA utility.
The investor chooses how much to consume at date 0 and how much to invest, but the investment
amount does not a↵ect the optimal portfolio of risky assets.
Consider the portfolio choice problem in which there is consumption at date 0 and date 1. Suppose
there is a risk-free asset with return Rf and n risky assets the returns of which are joint normally
distributed with mean vector µ and nonsingular covariance matrix ⌃. Consider an investor who
has time-additive utility and CARA utility for date–1 consumption:
↵c
u1 (c) = e .
The investor chooses date–0 consumption c0 and a portfolio of risky assets to maximize
✓ ◆
0 1 2 0
u(c0 ) exp ↵(w0 c0 )Rf ↵ (µ Rf 1) + ↵ ⌃ .
2
The optimal portfolio is the portfolio that maximizes
0 1 0
(µ Rf 1) ↵ ⌃ ,
2
just as in the problem with only date–1 consumption. Hence, (2.27) is the optimal portfolio.
2.4. This exercise repeats the previous one, but using asset payo↵s and prices instead of returns
and solving for the optimal number of shares to hold of each asset instead of the optimal amount
to invest.
18 2 Portfolio Choice and Stochastic Discount Factors
Suppose there is a risk-free asset with return Rf and n risky assets with payo↵s x̃i and prices pi .
Assume the vector x̃ = (x̃1 · · · x̃n )0 is normally distributed with mean µx and nonsingular covariance
matrix ⌃x . Let p = (p1 · · · pn )0 . Suppose there is consumption at date 0 and consider an investor
with initial wealth w0 and CARA utility at date 1:
↵c
u1 (c) = e .
Let ✓i denote the number of shares the investor considers holding of asset i and set ✓ = (✓1 · · · ✓n )0 .
The investor chooses consumption c0 at date 0 and a portfolio ✓, producing wealth (w0 c0
✓0 p)Rf + ✓0 x̃ at date 1. Show that the optimal vector of share holdings is
1
✓= ⌃ 1 (µx Rf p) .
↵ x
The investor chooses date–0 consumption c0 and a portfolio ✓ of risky assets to maximize
✓ ◆
0 0 1 2 0
u(c0 ) exp ↵(w0 c0 ✓ p)Rf ↵✓ µx + ↵ ✓ ⌃x ✓ .
2
1 0
Rf p0 ✓ + µ0x ✓ ↵✓ ⌃x ✓ .
2
with solution
1
✓= ⌃ 1 (µx Rf p) .
↵ x
2.5. Consider a utility function v(c0 , c1 ). The marginal rate of substitution is defined to be the
negative of the slope of an indi↵erence curve and is equal to
@v(c0 , c1 )/@c0
MRS(c0 , c1 ) = .
@v(c0 , c1 )/@c1
d log(c1 /c0 )
,
d log MRS(c0 , c1 )
where the marginal rate of substitution is varied holding utility constant. Show that, if
2 Portfolio Choice and Stochastic Discount Factors 19
1
v(c0 , c1 ) = c10 ⇢
+ c11 ⇢
,
1 ⇢ 1 ⇢
so ✓ ◆ ⇢
dc1 1 c0
= .
dc0 c1
This is the marginal rate of substitution. Setting x = c1 /c0 , we have
Hence,
d log MRS
= ⇢.
d log x
The elasticity of intertemporal substitution is the reciprocal 1/⇢.
2.6. This exercise shows that an improvement in the investment opportunity set leads to higher
saving (the substitution e↵ect dominates) when the elasticity of intertemporal substitution is high
and higher consumption (the wealth e↵ect dominates) when the elasticity of intertemporal substi-
tution is low.
Consider the portfolio choice problem with only a risk-free asset and with consumption at both the
beginning and end of the period. Assume the investor has time-additive power utility, so he solves
1 1 1
max c10 ⇢
+ c11 ⇢
subject to c0 + c1 = w0 .
1 ⇢ 1 ⇢ Rf
Show that the optimal consumption-to-wealth ratio c0 /w0 is a decreasing function of Rf if ⇢ < 1
and an increasing function of Rf if ⇢ > 1.
1 1
c10 ⇢
+ Rf1 ⇢
(w0 c0 )1 ⇢
,
1 ⇢ 1 ⇢
c0 ⇢ Rf1 ⇢
(w0 c0 ) ⇢
= 0.
20 2 Portfolio Choice and Stochastic Discount Factors
This implies
1/⇢ 1 1/⇢
c0 = Rf (w0 c0 ) ,
so
1/⇢ R1 1/⇢
f
c0 = w0 .
1+ 1/⇢ R1 1/⇢
f
The factor
1/⇢ R1 1/⇢
f
1+ 1/⇢ R1 1/⇢
f
2.7. Each part of this exercise illustrates the absence of wealth e↵ects for CARA utility and is
not true for general utility functions. The assumption in Part (c) that ỹ = aRf + bR̃ + "˜ where "˜
has zero mean and is uncorrelated with R̃ is without loss of generality (even without the normality
assumption): define b = cov(ỹ, R̃)/ var(R̃), a = (E[ỹ] bE[R̃])/Rf and "˜ = ỹ aRf bR̃. This is a
special case of an orthogonal projection (linear regression), which is discussed in more generality
in Section 4.5. The optimal portfolio in Part (c) can be interpreted as the optimal portfolio in the
absence of an endowment plus a hedge ( b) for ỹ.
Suppose there is a risk-free asset with return Rf and a risky asset with return R̃. Consider an
investor who maximizes expected end-of-period utility of wealth and who has CARA utility and
invests w0 . Suppose the investor has a random endowment ỹ at the end of the period, so his end-
of-period wealth is f Rf + R̃ + ỹ, where f denotes the investment in the risk-free asset and
the investment in the risky asset.
(a) Suppose ỹ and R̃ have a joint normal distribution. Derive the optimal portfolio. Show that
if ỹ and R̃ are uncorrelated, then the optimal is the same as if there were no end-of-period
endowment.
(b) Show that if ỹ and R̃ are independent, then the optimal is the same as if there were no
end-of-period endowment, regardless of whether ỹ and R̃ are normally distributed. Hint: Use
the law of iterated expectations as in Section 1.8 and the fact that if ṽ and x̃ are independent
random variables then E[ṽ x̃] = E[ṽ]E[x̃].
2 Portfolio Choice and Stochastic Discount Factors 21
(c) Suppose ỹ and R̃ have a joint normal distribution and ỹ = aRf +bR̃+ "˜ for constants a and b and
some "˜ that has zero mean and is uncorrelated with R̃. Show that the optimal is ⇤ b, where
⇤ denotes the optimal investment in the risky asset when there is no end-of-period endowment.
(a) The investment in the risk-free asset is f = w0 . The expected end-of-period wealth is
⇣ ⌘
w 0 Rf + E[R̃] Rf + E[ỹ] and the variance of end-of-period wealth is
2 2
var(R̃) + 2 cov(R̃, ỹ) + var(ỹ) = var(R̃) + var(ỹ) .
2 2
var(R̃) + 2 cov(R̃, ỹ) + var(ỹ) = ( + 2 b + b2 ) var(R̃) + var(˜
") .
22 2 Portfolio Choice and Stochastic Discount Factors
1 2
(E[R̃] Rf ) ↵( + 2 b) var(R̃) ,
2
2.8. This exercise illustrates the concept of “precautionary savings”—the risk imposed by ỹ results
in higher savings w0 c0 .
Consider the portfolio choice problem with only a risk-free asset and with consumption at both
the beginning and end of the period. Suppose the investor has time-additive utility with u0 = u
and u1 = u for a common function u and discount factor . Suppose the investor has a random
endowment ỹ at the end of the period, so he chooses c0 to maximize
Suppose the investor has convex marginal utility (u000 > 0) and suppose that E[ỹ] = 0. Show that
the optimal c0 is smaller than if ỹ = 0.
E[u0 ((w0 c0 )Rf + ỹ)] > u0 (E[(w0 c0 )Rf + ỹ]) = u0 ((w0 c0 )Rf ) .
Thus,
u0 (c0 ) > u0 ((w0 c0 )Rf ) .
The first-order condition if ỹ = 0 is for these to be equal. Because the left-hand side is decreasing
in c0 and the right-hand side increasing in c0 , equality requires that c0 be increased. Thus, the
optimal c0 would be larger if ỹ = 0.
2 Portfolio Choice and Stochastic Discount Factors 23
2.9. Letting c⇤0 denote optimal consumption in the previous problem, define the “precautionary
premium” ⇡ by
u0 ((w0 ⇡ c⇤0 )Rf ) = E[u0 ((w0 c⇤0 )Rf + ỹ)] .
(a) Show that c⇤0 would be the optimal consumption of the investor if he had no end-of-period
endowment and had initial wealth w0 ⇡.
(b) Assume the investor has CARA utility. Show that the precautionary premium is independent
of initial wealth (again, no wealth e↵ects with CARA utility).
with solution
1 ⇥ ↵ỹ
⇤
⇡= log E e .
↵Rf
2.10. The assumption in this exercise is a weak form of market completeness. The conclusion
follows in a complete market from the formulation (2.35) of the portfolio choice problem.
Suppose there is a stochastic discount factor m̃ with the property that for every function g there
exists a portfolio ✓ (depending on g) such that
n
X
✓i x̃i = g(m̃) .
i=1
Consider an investor with no labor income ỹ. Show that his optimal wealth is a function of m̃.
Hint: For any feasible w̃, define w̃⇤ = E[w̃|m̃], and use the result of Section 1.8.
24 2 Portfolio Choice and Stochastic Discount Factors
Setting w̃⇤ = E[w̃|m̃] and "˜ = w̃ w̃⇤ , we have that w̃ is w̃⇤ + "˜. Moreover,
Thus, "˜ is mean-independent noise. Hence w̃⇤ is preferred to w̃. Because w̃⇤ is a function of m̃,
there exists by assumption a portfolio ✓˜ with payo↵ equal to w̃⇤ . The cost of the portfolio is
3.1. Suppose each investor h has a concave utility function, and suppose a feasible allocation
(w̃1 , . . . , w̃m ) of market wealth w̃m satisfies the first-order condition
u0h (w̃h ) = h m̃
for each investor h, where m̃ is a stochastic discount factor and is the same for each investor. Show
that the allocation solves the social planner’s problem (3.2) with weights h = 1/ h. Note: The
first-order condition holds with the stochastic discount factor being the same for each investor in a
competitive equilibrium of a complete market, because there is a unique stochastic discount factor
in a complete market. Recall that h in the first-order condition is the Lagrange multiplier for the
investor’s budget constraint (see Section 2.1) and hence is the marginal value of beginning-of-period
wealth. Thus, the weights in the social planner’s problem can be taken to be the reciprocals of the
marginal values of wealth. Other things equal, investors with high wealth have low marginal values
of wealth and hence have high weights in the social planner’s problem.
in each state of the world, over all allocations (w̃1 , . . . , w̃H ), feasible or not. If (w̃10 , . . . , w̃H
0 ) is any
H
X H
X H
X H
X
1 1
uh (w̃h ) m̃w̃h uh (w̃h0 ) m̃w̃h0
h h
h=1 h=1 h=1 h=1
XH H
X
1
= uh (w̃h0 ) m̃w̃h .
h
h=1 h=1
Hence,
H
X H
X
1 1
uh (w̃h ) uh (w̃h0 ) .
h h
h=1 h=1
3.2. Suppose there are two investors, the first having constant relative risk aversion ⇢ > 0 and the
second having constant relative risk aversion 2⇢.
for ⌘ > 0. Hint: Use the first-order condition and the quadratic formula. Because ⌘ is arbitrary
in (0, 1), there are many equivalent ways to write the sharing rules.
(b) Suppose the market is complete and satisfies the law of one price. Show that the stochastic
discount factor in a competitive equilibrium is
⇣p ⌘ 2⇢
m̃ = ⌘ 2 + 2⌘ w̃m ⌘
(a) The marginal utility of the first investor is w ⇢, and the marginal utility of the second investor
is w 2⇢ . The first-order condition is
⇢ 2⇢ 2⇢
1 w̃1 = 2 w̃2 = 2 (w̃m w̃1 ) .
This implies
w̃1 = (w̃m w̃1 )2 ,
where =( 1/⇢ .
1/ 2) Thus
w̃12 2
(2w̃m + )w̃1 + w̃m = 0.
p
2w̃m + (2w̃m + )2
± 2
4w̃m
w̃1 =
2
p
2
= w̃m + ⌘ ± ⌘ + 2⌘ w̃m ,
for some ⌘ > 0. Thus, the second investor’s marginal utility at a competitive equilibrium equals
⇣p ⌘ 2⇢
⌘ 2 + 2⌘ w̃m ⌘ .
is a stochastic discount factor for some > 0. This is the unique stochastic discount factor in
a complete market.
3.3. Suppose there are n risky assets with payo↵s x̃i and no risk-free asset. Assume there is
consumption only at date 1. Let µ denote the mean and ⌃ the covariance matrix of the vector
X̃ = (x̃1 · · · x̃n )0 of asset payo↵s. Assume X̃ has a normal distribution, and assume ⌃ is nonsingular.
Let ✓¯ = (✓¯1 · · · ✓¯n )0 denote the vector of asset supplies. Assume all investors have CARA utility and
no endowments ỹh . Define ↵ to be the aggregate absolute risk aversion as in Section 1.3. Show that
the vector
p= µ ↵⌃ ✓¯
would be better o↵ if they had less wealth. Including consumption at date 0 or changing the budget
constraint to p0 ✓ p0 ✓¯h instead of p0 ✓ = p0 ✓¯h (i.e., allowing free disposal of wealth) eliminates the
equilibria with < 0.
1 0
✓0 µ ↵✓ ⌃✓
2
1 1
✓h = ⌃ (µ h p)
↵h
To check if markets clear, we need to compute h, which is determined by the budget equation of
investor h:
1 0 h 0
wh0 = p0 ✓h = p⌃ 1
µ p⌃ 1
p.
↵h ↵h
Instead of computing h,
¯
it suffices to sum this over h, noting that aggregate initial wealth is p0 ✓.
This yields !
H
X
1 h
p0 ✓¯ = p0 ⌃ 1
µ p0 ⌃ 1
p,
↵ ↵h
h=1
implying
H
X ✓ ◆
h 1 p0 ⌃ 1µ ↵p0 ✓¯
= .
↵h ↵ p0 ⌃ 1 p
h=1
For p = µ ↵⌃ ✓¯ , the expression in parentheses in the last displayed equation is 1/ . Thus, for
such prices, aggregate demand is
1 1 1 1 ⇥ ⇤
⌃ 1
µ ⌃ 1
p= ⌃ 1
µ ⌃ 1
µ ↵⌃ ✓¯ = ✓¯ .
↵ ↵ ↵ ↵
3.4. Reconsider the previous problem assuming there is a risk-free asset in zero net supply (meaning
investors can borrow from and lend to each other) and assuming there is consumption at date 0.
Both the price vector p of the risky assets and the risk-free return Rf are determined endogenously
in equilibrium. Suppose the utility functions of investor h are
3 Equilibrium and Efficiency 29
↵h c ↵h c
u0 (c) = e and u1 (c) = he .
PH
Let c̄0 denote the aggregate endowment h=1 yh0 at date 0 and define by
H
X ⌧h
log = log h,
⌧
h=1
PH
where ⌧h = 1/↵h and ⌧ = h=1 ⌧h . Using the result of Exercise 2.4 on the optimal demands for
the risky assets, show that the equilibrium risk-free return and price vector p are given by
✓ ◆
1 1 2 ¯0 ¯
Rf = exp ↵ ✓¯0 µ c̄0 ↵ ✓ ⌃✓ ,
2
1 ¯ .
p= (µ ↵⌃ ✓)
Rf
Explain in economic terms why the risk-free return is higher when ✓¯0 µ is higher and lower when ,
c̄0 , or ✓¯0 ⌃ ✓¯ is higher.
1 1
✓h = ⌃ (µ Rf p) .
↵h
1 1
✓¯ = ⌃ 1
(µ Rf p) ) p= (µ ¯ .
↵⌃ ✓)
↵ Rf
To derive Rf we need to clear the market for date–0 consumption (or the market for the risk-free
asset). Investor h chooses c0 and ✓ to maximize
exp( ↵h c0 ) hE exp ↵h [(wh0 c0 p0 ✓)Rf + ✓0 x̃]
0
= exp( ↵h c0 ) h exp ↵h (wh0 c0 p ✓)Rf E exp ↵h ✓0 x̃ .
✓ ◆
↵2 ¯0
E exp ↵h ✓0 x̃ = exp ↵✓¯0 µ + ✓ ⌃✓ ,
2
log h log Rf ↵ ¯0 ↵2 ¯0 ¯
ch0 = + (wh0 ch0 p0 ✓h )Rf + ✓µ ✓ ⌃✓ .
↵h ↵h ↵h 2↵h
By the definition of ,
H
X
log log h
= .
↵ ↵h
h=1
log log Rf ↵ ¯0 ¯
c̄0 = + ✓¯0 µ ✓ ⌃✓ .
↵ ↵ 2
3.5. Suppose the payo↵ of the market portfolio w̃m has k possible values. Denote these possible
values by a1 < · · · < ak . For convenience, suppose ai ai 1 is the same number for each i.
3 Equilibrium and Efficiency 31
Suppose there is a risk-free asset with payo↵ equal to 1. Suppose there are k 1 call options on the
market portfolio, with the exercise price of the i–th option being ai . The payo↵ of the i–th option
is max(0, w̃m ai ).
(a) Show for each i = 1, . . . , k 2 that a portfolio that is long one unit of option i and short one
unit of option i + 1 pays if w̃m ai+1 and 0 otherwise. This portfolio of options is a bull
spread.
(b) Consider the following k portfolios. Show that the payo↵ of portfolio i is 1 when w̃m = ai and
0 otherwise. Thus, these are Arrow securities for the events on which w̃m is constant.
(i) i = 1: long one unit of the risk-free asset, short 1/ units of option 1, and long 1/ units
of option 2. This portfolio of options is a short bull spread.
(ii) 1 < i < k: long 1/ units of option i 1, short 2/ units of option i, and long 1/ units
of option i + 1. These portfolios are butterfly spreads.
(iii) i = k 1: long 1/ units of option k 2 and short 2/ units of option k 1.
(iv) i = k: long 1/ units of option k 1.
(c) Given any function f , define z̃ = f (w̃m ). Show that there is a portfolio of the risk-free asset
and the call options with payo↵ equal to z̃.
(a) When w̃m ai+1 , a long position in option i pays w̃m ai , and a short position in option i + 1
has cash flow ai+1 w̃m . The sum of these is ai+1 ai = .
(b) (i) Being long 1/ units of option 2 and short 1/ units of option 1 pays 1 when w̃m a2
and 0 otherwise. Combining this with a payo↵ of 1 yields 1 when w̃m = a1 and 0 otherwise.
(ii) Being long 1/ units of option i 1 and short 1/ units of option i pays 1 when w̃m ai
and 0 otherwise. Being short 1/ units of option i and long 1/ units of option i+1 pays 1
when w̃m ai+1 and 0 otherwise. The sum of these pays 1 when w̃m = ai and 0 otherwise.
(iii) Being long 1/ units of option k 2 produces a payo↵ of 1 when w̃m = ak 1 and 2 when
w̃m = ak . Being short 2/ units of option k 1 produces a payo↵ of 2 when w̃m = ak .
The sum of the two pays 1 when w̃m = ak 1 and 0 otherwise.
(iv) Being long 1/ units of option k 1 produces a payo↵ of 1 when w̃m = ak and 0 otherwise.
(c) Let zi denote the value of z̃ when w̃m = ai . The portfolio consisting of z1 units of the risk-free
asset, (z2 z1 )/ units of option 1, and (zi+1 2zi +zi 1 )/ units of option i for i = 2, . . . , k 1
has payo↵ equal to z̃.
32 3 Equilibrium and Efficiency
w̃h = ah + bh w̃m
PH
where bh = ⌧h /⌧ and h=1 ah = 0. Show that the allocation is Pareto optimal. Hint: Show that it
solves the social planner’s problem with weights h defined as h = ⌧h eah /⌧h .
↵h w̃h
(8 h) ↵h he = ⌘˜ .
Setting w̃h = ah + bh w̃m and h = ⌧h eah /⌧h , the left-hand side is e ⌧ w̃m . Thus, the first-order
condition holds for ⌘˜ = e ⌧ w̃m . By concavity, the first-order condition is sufficient for optimality.
3.7. Suppose all investors have shifted CRRA utility with the same coefficient ⇢ > 0. Suppose
w̃m > ⇣. Consider an allocation
w̃h = ⇣h + bh (w̃m ⇣)
PH
where h=1 bh = 1. Show that the allocation is Pareto optimal. Hint: Show that it solves the social
planner’s problem with weights h defined as h = b⇢h .
⇢
(8 h) h w̃h = ⌘˜ .
Setting w̃h = ⇣h + bh (w̃m ⇣) and h = b⇢h , the left-hand side is (w̃m ⇣) ⇢. Thus, the first-
order condition holds for ⌘˜ = (w̃m ⇣) ⇢. By concavity, the first-order condition is sufficient for
optimality.
3.8. Show that if each investor has shifted CRRA utility with the same coefficient ⇢ > 0 and shift
⇣h , then, as asserted in Section 3.6, any Pareto optimal allocation involves an affine sharing rule.
⇢
(8 h) h (w̃h ⇣h ) = ⌘˜
1/⇢
for some weights h and Lagrange multiplier ⌘˜. This implies w̃h ⇣h = h ⌘ ˜ 1/⇢ and adding over
investors yields
3 Equilibrium and Efficiency 33
H
!
X 1/⇢ 1/⇢
w̃m ⇣= h ⌘˜ ,
h=1
so !⇢
H
X 1/⇢ ⇢
⌘˜ = h (w̃m ⇣) .
h=1
which implies
1/⇢
h
w̃h ⇣h = P 1/⇢
(w̃m ⇣) .
H
j=1 j
3.9. Consider an economy with date–0 consumption as in Section 3.7. Assume the investors have
time-additive utility and the date–1 allocation solves the social planning problem (3.1). Using the
first-order condition (2.390 ), show that the equilibrium allocation is Pareto optimal. Hint: Using
the first-order condition (3.4) with ⌘˜ = ⌘˜1 , show that
0
(8 h) h uh0 (c̃h0 ) = Rf E[˜
⌘1 ] .
By (2.390 ),
u0h1 (c̃h1 ) ⌘˜1
=
u0h0 (ch0 ) 0
h uh0 (ch0 )
This implies
0
h uh0 (c̃h0 ) = ⌘0 ,
defining ⌘0 = Rf E[˜
⌘1 ]. Thus, the first-order conditions for the social planner’s problem are satisfied.
By concavity, this implies Pareto optimality.
4
4.1. Assume there are two possible states of the world: !1 and !2 . There are two assets, a risk-free
asset returning Rf in each state, and a risky asset with initial price equal to 1 and date–1 payo↵
x̃. Let Rd = x̃(!1 ) and Ru = x̃(!2 ). Assume without loss of generality that Ru > Rd .
(a) What conditions on Rf , Rd and Ru are equivalent to the absence of arbitrage opportunities?
(b) Assuming the conditions from the previous part hold, compute the unique vector of state prices,
and compute the unique risk neutral probabilities of states !1 and !2 .
(c) Suppose another asset is introduced into the market that pays max(x̃ K, 0) for some constant
K. Compute the price at which this asset should trade, assuming the conditions from part (a)
hold.
(a) The payo↵ of a zero-cost portfolio is (R̃ Rf ) for some . For this to be nonnegative in both
states and positive in one state, we must have either (i) > 0 and Ru > Rd Rf or (ii) <0
and Rf Ru > Rd . Thus, a necessary and sufficient condition for the absence of arbitrage
opportunities is that Ru > Rf > Rd .
(b) Let qd denote the state price of state !1 and qu the state price of state !2 . The state prices
satisfy
q d Rf + q u Rf = 1 ,
q d Rd + q u Ru = 1 .
Ru Rf Rf Rd
qd = , and qu = .
Rf (Ru Rd ) Rf (Ru Rd )
36 4 Arbitrage and Stochastic Discount Factors
4.2. Show that, if there is a strictly positive stochastic discount factor, then there are no arbitrage
opportunities.
If x̃ is a nonnegative marketed payo↵, then its price is E[m̃x̃] 0, and E[m̃x̃] = 0 if and only if
x̃ = 0 with probability one. Therefore, there are no arbitrage opportunities.
4.3. Show by example that the law of one price can hold but there can still be arbitrage opportu-
nities.
Suppose there are two possible states of the world, and the market consists of the two Arrow
securities having prices pi . Then the market is complete, and each payo↵ x̃ = (x1 , x2 ) has a unique
cost p1 x1 + p2 x2 . If p1 < 0, then buying the first asset is an arbitrage opportunity.
4.4. Suppose there is no risk-free asset. For what value of ⌫ = E[m̃] does the projection m̃⌫p equal
the projection m̃p ?
In the absence of a risk-free asset, m̃⌫p belongs to the span of the assets if and only if the constant
in (4.9) is zero, i.e., if and only if
4.5. Assume there are three possible states of the world: !1 , !2 , and !3 . Assume there are two
assets: a risk-free asset returning Rf in each state, and a risky asset with return R1 in state !1 , R2
in state !2 , and R3 in state !3 . Assume the probabilities are 1/4 for state !1 , 1/2 for state !2 , and
1/4 for state !3 . Assume Rf = 1.0, and R1 = 1.1, R2 = 1.0, and R3 = 0.9.
4 Arbitrage and Stochastic Discount Factors 37
m̃ = (m1 , m2 , m3 ) ,
where mi denotes the value of the stochastic discount factor in state !i . Verify that m1 = 4,
m2 = 2, m3 = 4 is a stochastic discount factor.
(d) Consider the formula
ỹp = E[ỹ] + Cov(X̃, ỹ)0 ⌃x 1 (X̃ E[X̃])
for the projection of a random variable ỹ onto the linear span of a constant and a random vector
X̃. When the vector X̃ has only one component x̃ (is a scalar), the formula simplifies to
where
cov(x̃, ỹ)
= .
var(x̃)
Apply this formula with ỹ being the stochastic discount factor m1 = 4, m2 = 2, m3 = 4 and
x̃ being the risky asset return R1 = 1.1, R2 = 1.0, R3 = 0.9 to compute the projection of the
stochastic discount factor onto the span of the risk-free and risky assets.
(e) The projection in part (d) is by definition the payo↵ of some portfolio. What is the portfolio?
(a) Let R̃ denote the risky asset return. A zero-cost portfolio has payo↵ (R̃ Rf ) for some . This
equals 0.1 in state 1, 0 in state 2, and 0.1 in state 3. Obviously, there is no such that
(R̃ Rf ) is nonnegative in all states and positive in some state.
(b) State prices must satisfy
q1 + q2 + q3 = 1
1.1q1 + q2 + 0.9q3 = 1 .
Subtracting the top from the bottom shows that q3 = q1 and substituting this into the first
shows that q2 = 1 2q1 . q1 is arbitrary.
(c) Stochastic discount factors are given by
4.6. Suppose there is a risk-free asset and the risky asset returns have a joint normal distribution.
Use the reasoning in Exercise 2.10 and the formula (5.31) for m̃p to show that the optimal portfolio
of risky assets for an investor with zero labor income is ⇡ = ⌃ 1 (µ Rf 1) for some real number
, where ⌃ denotes the covariance matrix of the risky asset returns. Hint: if w̃ and m̃ are joint
normal, then E[w̃|m̃] is the orthogonal projection of w̃ on a constant and m̃—i.e.,
cov(w̃, m̃)
E[w̃|m̃] = E[w̃] + (m̃ E[m̃]) .
var(m̃)
For any feasible w̃, let w̃⇤ = E[w̃ | m̃p ]. Then w̃ equals w̃⇤ plus mean-independent noise, so w̃⇤ is
preferred to w̃. From (5.31),
1
m̃p E[m̃p ] = (µ Rf 1)0 ⌃ 1
(R̃vec µ) .
Rf
Hence, ✓ ◆
⇤ 1 cov(w̃, m̃p )
w̃ = E[w̃] (µ Rf 1)0 ⌃ 1
(R̃vec µ) .
Rf var(m̃p )
This is in the span of the assets and budget feasible, because E[m̃p w̃⇤ ] = E[m̃p w] by iterated
expectations. Thus, w̃⇤ is feasible. The portfolio of risky assets producing w̃⇤ is ⌃ 1 (µ Rf 1) for
✓ ◆ ✓ ◆
1 cov(w̃, m̃p ) 1 cov(w̃⇤ , m̃p )
= = ,
Rf var(m̃p ) Rf var(m̃p )
the second equality following from iterated expectations.
4.7. Assume there is a finite number of assets, and the payo↵ of each asset has a finite variance.
Assume the law of one price holds. Apply facts stated in Section 4.8 to show that there is a unique
stochastic discount factor m̃p in the span of the asset payo↵s. Show that the orthogonal projection
of any other stochastic discount factor onto the span of the asset payo↵s equals m̃p .
4 Arbitrage and Stochastic Discount Factors 39
The span of the assets is a finite-dimensional subspace of L2 . The law of one price states that there
is a unique price C[x̃] for each x̃ in the span of the payo↵s. The function C[·] is linear. Therefore,
it has a Riesz representation C[x̃] = E[x̃m̃p ] for a unique m̃p in the span of the assets. Given any
stochastic discount factor m̃, we have m̃ = m̃⇤ + "˜, where the orthogonal projection m̃⇤ is in the
span of the assets and "˜ is orthogonal to the span of the assets. Hence, C[x̃] = E[m̃x̃] = E[x̃m̃⇤ ] for
all x̃ in the span of the assets. Thus, m̃⇤ is also in the span of the assets and represents the price
function. By the uniqueness of the Riesz representation, it must be that m̃⇤ = m̃p .
5
Mean-Variance Analysis
5.1. Derive the minimum global variance portfolio directly by solving the problem: minimize ⇡ 0 ⌃⇡
subject to 10 ⇡ = 1.
5.2. Assume there is a risk-free asset. Consider an investor with quadratic utility, who seeks to
maximize
1 1
⇣E[w̃] E[w̃]2 var(w̃) .
2 2
(a) Show that the investor will choose a portfolio on the mean-variance frontier.
(b) Assume (5.14) holds, so the tangency portfolio is efficient. Under what circumstances will the
investor choose a mean-variance efficient portfolio? Explain the economics of the condition you
derive. Hint: Compare Exercise 2.2.
utility function. When ⇣ < w0 Rf , the investor can exceed the bliss level by simply holding the
risk-free asset. To avoid this, he holds an inefficient portfolio of risky assets.
5.3. Suppose that the risk-free return is equal to the expected return of the global minimum
variance portfolio (Rf = B/C). Show that there is no tangency portfolio. Hint: Show there is no
and satisfying
1
⌃ (µ Rf 1) = ⇡µ + (1 )⇡1 .
The mean-variance frontier considering only the risky assets is the set ⇡µ + (1 )⇡1 for some
, and the mean-variance frontier including the risk-free asset is the set ⌃ 1 (µ Rf 1) for some
. For the frontiers to intersect, we must have
1
⌃ (µ Rf 1) = ⇡µ + (1 )⇡1 .
This is equivalent to
✓ ◆ ✓ ◆
1 1 1
⌃ µ= Rf + 0 ⌃ 1,
10 ⌃ 1µ 1⌃ 11
1
= Rf + = 0.
10 ⌃ 1 µ 10 ⌃ 11
This implies
1
Rf + = 0,
10 ⌃ 1 µ 10 ⌃ 11
1
= 0,
10 ⌃ 1 1
which is impossible.
5.4. Consider the problem of choosing a portfolio ⇡ of risky assets, a proportion b 0 to borrow
and a proportion ` 0 to lend to maximize the expected return ⇡ 0 µ + ` R` b Rb subject to the
5 Mean-Variance Analysis 43
constraints (1/2)⇡ 0 ⌃⇡ k and 10 ⇡ + ` b = 1. Assume B/C > Rb > R` , where B and C are
defined in (5.6). Define
1 1
⇡b = ⌃ (µ Rb 1) ,
10 ⌃ 1 (µ Rb 1)
1 1
⇡` = ⌃ (µ R` 1) .
10 ⌃ 1 (µ R` 1)
Using the Kuhn-Tucker conditions, show that the solution is either (i) ⇡ = (1 ` )⇡` for 0 ` 1,
(ii) ⇡ = ⇡` + (1 )⇡b for 0 1, or (iii) ⇡ = (1 + b )⇡b for b 0.
µ ⌃⇡ 1 = 0,
R` + ⌘` = 0 ,
Rb + + ⌘b = 0 ,
`, b , ⌘` , ⌘b , 0,
1 0
⇡ ⌃⇡ k ,
2
10 ⇡ + ` b = 1,
✓ ◆
1 0
⌘` ` = ⌘b b = ⇡ ⌃⇡ k = 0 .
2
If ` > 0, then ⌘` = 0, = R` , and
1 1
⇡= ⌃ (µ R` 1) .
If ` = b = 0, then
1 1
⇡= ⌃ (µ 1) ,
5.5. Establish the properties claimed for the risk-free return proxies:
44 5 Mean-Variance Analysis
(a) Show that var(R̃) var(R̃p + bm ẽp ) for every return R̃.
(b) Show that cov(R̃p , R̃p + bz ẽp ) = 0.
(c) Prove (5.18), showing that R̃p + bc ẽp represents the constant bc times the expectation operator
on the space of returns.
(a) By Fact 15, the minimum variance return is R̃p + bẽp for some b. Using Fact 8, we have
By Fact 16, E[ẽp ] > 0, so the minimum variance return is found by minimizing (b2 (1 E[ẽp ])
2bE[R̃p ] in b, with solution b = bm .
(b) Using Fact 8, we have cov(R̃p , R̃p + bz ẽp ) = var(R̃p ) bz E[R̃p ]E[ẽp ] = 0.
(c) Using Fact 11 and the definition of bc , we have
Thus,
bc E[R̃] = E[R̃(R̃p + bc ẽp )] .
We have
1 1
E[x̃R̃] = E[(1 ẽp )R̃p ] + E[(1 ẽp )(R̃ R̃p )]
E[R̃p ] E[R̃p ]
1
= E[(1 ẽp )R̃p ]
E[R̃p ]
= 1,
using E[R̃ R̃p ] = E[ẽp (R̃ R̃p ] for the second equality and Fact 8 for the third. Thus, x̃ is a
stochastic discount factor. This implies
1 1 E[ẽp ]
= E[x̃] = .
Rf E[R̃p ]
Moreover, x̃ = m̃p implies
x̃
R̃p = ,
E[x̃2 ]
and
1 1 E[ẽp ] 1
E[x̃2 ] = (1 2E[ẽp ] + E[ẽ2p ]) = = ,
E[R̃p ]2 E[R̃p ] 2 Rf E[R̃p ]
using Fact 16 for the second equality. Thus,
!
1
R̃p = Rf E[R̃p ] (1 ẽp ) = Rf (1 ẽp ) .
E[R̃p ]
5.7. Show that E[R̃2 ] E[R̃p2 ] for every return R̃ (thus, R̃p is the “minimum second-moment
return”). The returns having a given second moment a are the returns satisfying E[R̃2 ] = a, which
is equivalent to
var(R̃) + E[R̃]2 = a ;
thus, they plot on the circle x2 + y 2 = a in (standard deviation, mean) space. Use the fact that R̃p
is the minimum second-moment return to illustrate graphically that R̃p must be on the inefficient
part of the frontier, with and without a risk-free asset (assuming E[R̃p ] > 0 in the absence of a
risk-free asset).
With a risk-free asset, the cone intersects the vertical axis at Rf > 0, and the point on the cone
closest to the origin is on the lower part. In the absence of a risk-free asset, the assumption E[R̃p ] > 0
46 5 Mean-Variance Analysis
implies that global minimum variance portfolio has a positive expected return (use the definition
of bm and Facts 16 and 17 — which imply 1 E[ẽp ] > 0 — to deduce this). Thus, the point on the
hyperbola closest to the origin must be on the lower part of the hyperbola.
5.8. If all returns are joint normally distributed, then R̃p , ẽp and "˜ are joint normally distributed
in the orthogonal decomposition R̃ = R̃p + bẽp + "˜ of any return R̃ (because R̃p is a return and ẽp
and "˜ are excess returns). Assuming all returns are joint normally distributed, use the orthogonal
decomposition to compute the optimal return for a CARA investor.
When returns are normally distributed, a CARA investor chooses the return R̃ that maximizes
1
E[R̃] ↵w0 var(R̃) .
2
Given R̃ = R̃p + bẽp + "˜ and Facts 11 and 15, the objective function is
1
E[R̃ + bẽp ] ↵w0 [var(R̃p + bẽp ) + var(˜
")] ,
2
1
bE[ẽp ] ↵w0 [2b cov(R̃p , ẽp ) + b2 var(ẽp )] ,
2
implying
E[ẽp ] cov(R̃p , ẽp )
b= .
↵w0 var(ẽp ) var(ẽp )
Using Facts 8 and 17, we can simplify this further to
1 + ↵w0 E[R̃p ]
b= .
↵w0 (1 E[ẽp ])
5.9. The return R̃p is on the mean-variance frontier. Hence, when there is a risk-free asset, it must
be the return of a weighted average of the tangency portfolio and the risk-free asset. The purpose
of this exercise is to compute the weighted average. When there is a risk-free asset, the projection
m̃p defined in Section 4.4 is the same as the projection m̃⌫p defined in Section 4.5. Furthermore, the
vector of asset payo↵s in the formula (4.10) for m̃⌫p can always be replaced by the vector R̃vec of
5 Mean-Variance Analysis 47
returns whenever the asset prices are positive, because the linear span of the returns is the same as
the linear span of the payo↵s. Substituting the returns for the payo↵s and substituting E[m̃] = 1/Rf
and E[m̃R̃vec ] = 1 in (4.10), we have
✓ ◆0
1 1 1
m̃p = + 1 µ ⌃ (R̃vec µ) .
Rf Rf
Using this formula, show that R̃p = ⇡⇤0 R̃vec + (1 )Rf for some (i.e., calculate ).
We have
1 ⇥ ⇤ 1
m̃p = 1 + (µ Rf 1)0 ⌃ 1
µ (µ Rf 1)0 ⌃ 1
R̃vec .
Rf Rf
Hence
2
var(m̃p ) = ,
Rf2
where 2 = (µ Rf 1)0 ⌃ 1 (µ Rf 1) is the squared maximum Sharpe ratio. Because E[m̃p ] = 1/Rf ,
this implies
1 + 2
E[m̃2p ] = .
Rf2
Therefore
Rf ⇥ ⇤ Rf
R̃p = 2
1 + (µ Rf 1)0 ⌃ 1 µ (µ Rf 1)0 ⌃ 1
R̃vec
1+ 1 + 2
Rf (1 + A Rf B) Rf (B Rf C) 0 vec
= ⇡⇤ R̃ ,
1 + 2 1 + 2
Rf (B Rf C)
= ,
1 + 2
we have
1 + 2 + R f B Rf2 C 1 + A Rf B
1 = = ,
1+ 2 1 + 2
because 2 = A 2Rf B + Rf2 C. Thus,
5.10. Assuming there is a risk-free asset, show that in Exercise 5.9 is negative when Rf <
B/C and positive when Rf > B/C. (This verifies that the portfolio generating R̃p is short the
tangency portfolio when the tangency portfolio is efficient and long the tangency portfolio when it
is inefficient.)
48 5 Mean-Variance Analysis
Rf (B Rf C)
= <0
1 + 2
when B > Rf C and positive when B < Rf C.
6
6.1. Suppose there is a risk-free asset. Use the formula (5.9) for frontier portfolios to show that
the beta-pricing model (6.5) implies the return R̃⇤ is on the mean-variance frontier.
Let R̃⇤ = Rf + ⇡ 0 (Rvec Rf 1). Stacking the formula (6.5) for assets 1, . . . , n gives
E[R̃⇤ ] Rf
µ = Rf 1 + ⌃⇡ ,
var(R̃⇤ )
so
1
⇡= ⌃ (µ Rf 1) ,
where
var(R̃⇤ ) ⇡ 0 ⌃⇡
= = 0 .
E[R̃⇤ ] Rf ⇡ (µ Rf 1)
6.2. Suppose there is no risk-free asset. Use the formula (5.1) for frontier portfolios to show that
a beta-pricing model (6.5) is equivalent to the return R̃⇤ being on the mean-variance frontier and
not equal to the global minimum variance return.
Let R̃⇤ = ⇡ 0 Rvec . Stacking the formula (6.5) for assets 1, . . . , n gives
E[R̃⇤ ] ↵
µ = ↵1 + ⌃⇡ ,
var(R̃⇤ )
so
1 1
⇡= ⌃ µ+ ⌃ 1,
where
var(R̃⇤ ) var(R̃⇤ )
= and = ↵ .
E[R̃⇤ ] Rf E[R̃⇤ ] Rf
To have ⇡ = ⇡1 , we must have = 0, which implies = 0 and ⇡ = 0, contradicting 10 ⇡ = 1.
50 6 Beta Pricing Models
6.3. Suppose investors can borrow and lend at di↵erent rates. Let Rb denote the return on borrow-
ing and R` the return on lending. Suppose B/C > Rb > R` , where B and C are defined in (5.6).
Suppose each investor chooses a mean-variance efficient portfolio, as described in Exercise 5.4. Show
that the CAPM holds with R` Rz Rb .
From Exercise 5.4, the optimal portfolio of risky assets for each investor h is
h [ h ⇡` + (1 h )⇡b ] ,
where 0 h 1 and h 0 is the fraction of investor h’s wealth that is invested in risky assets. It
follows that the market portfolio of risky assets is ⇡` + (1 )⇡b for some 0 1. The vector
of covariances of the risky asset returns with the market return is therefore
1
Cov(R̃vec , R̃m ) = ⌃[ ⇡` + (1 )⇡b ] = (µ Rb 1) + (µ R` 1) .
B CRb B CR`
1 ✓ b Rb + ✓ ` R`
µ= Cov(R̃vec , R̃m ) + 1.
✓b + ✓` ✓b + ✓`
Thus,
✓ b Rb + ✓ ` R `
Rz = ,
✓b + ✓`
which is a convex combination of Rb and R` .
6.4. Assuming normally distributed returns, no end-of-period endowments, and investors with
CARA utility, derive the CAPM from the portfolio formula (2.27), i.e., from
1 1
h = ⌃ (µ Rf 1) ,
↵h
where ↵h denotes the absolute risk aversion of investor h. Show that the factor risk premium is
↵w0 var(R̃m ), where ↵ is the aggregate absolute risk aversion defined in Section 1.3 and w0 =
P
10 Hh=1 h is the market value of risky assets at date 0.
PH
Setting = h=1 h and adding the optimal portfolios over investors gives
1 1
= ⌃ (µ Rf 1) ,
↵
which in equilibrium is the vector of market values of risky assets at date 0. Thus,
6 Beta Pricing Models 51
µ = Rf 1 + ↵⌃ .
The market portfolio is ⇡ = (1/10 ) . The return on the market portfolio is R̃m = 0 R̃vec / 0 1, and
the vector of betas of the asset returns with respect to the market return is
1
⌃⇡ .
⇡ 0 ⌃⇡
We have
1
µ = Rf 1 + ↵(10 )⌃⇡ = ↵(10 )(⇡ 0 ⌃⇡) ⌃⇡ ,
⇡ 0 ⌃⇡
so the factor risk premium is
↵(10 )(⇡ 0 ⌃⇡) = ↵w0 var(R̃m ) .
6.5. Assume there exists a return R̃⇤ that is on the mean-variance frontier and is an affine function
of a vector F̃ ; i.e., R̃ = a + b0 F̃ . Assume either (i) there is a risk-free asset and R̃⇤ 6= Rf , or (ii)
there is no risk-free asset and R̃⇤ is di↵erent from the global minimum variance return. Show that
there is a beta pricing model with factors F̃ . Note: In the context of a factor model with factors
F̃ , a return R̃ = a + b0 F̃ is called well diversified, because it has no idiosyncratic risk. If there is
a finite number of assets satisfying a factor model, then there is no risky well diversified return,
P P 2
because var( ⇡ "˜i ) = ⇡ var(˜"i ) > 0 if ⇡ 6= 0. However, if there is an infinite number of assets,
then one can take ⇡i = 1/n for i = 1, . . . , n and n ! 1 to obtain a well diversified limit return.
Under either assumption (i) or (ii), there is a beta pricing model with R̃⇤ as the factor. Thus, for
every return R̃,
= ↵ + b0 Cov(F̃ , R̃)
= ↵ + ˆ 0 ⌃F 1 Cov(F̃ , R̃) ,
where we define
ˆ = ⌃F b .
Show that there is a beta pricing model with factors F̃ . Hint: Pareto optimality implies sharing
rules w̃h = fh (w̃m ).
Complete markets and strictly monotone preferences implies Pareto optimality. Given the sharing
rules, the marginal utility of any investor depends only on w̃m and hence depends only on F̃ . Thus,
there is a stochastic discount factor that is proportional to an investor’s marginal utility and equal
to m̃ = g(F̃ ) for some function g. By the strict monotonicity of preferences, m̃ is strictly positive.
Hence, E[m̃] 6= 0. This implies a beta pricing model with m̃ as the factor. This further implies an
APT pricing formula with pricing errors E[m̃˜
"i ]/E[m̃]. The pricing errors are zero, because,
Thus, there is exact APT pricing, i.e., a beta pricing model with F̃ as the factors.
(a) Show that R̃1⇤ and R̃2⇤ do not satisfy a one-factor model with factor f˜.
(b) Show that R̃1⇤ and R̃2⇤ satisfy a zero-factor model, i.e.,
(a) R̃2⇤ = E[R̃2⇤ ] + (2⇡ 1)f˜ + "˜⇤2 , where "˜⇤2 = ⇡ "˜1 + (1 "2 is uncorrelated with f˜. However,
⇡)˜
"1 , "˜⇤2 ) = ⇡
cov(˜ 2 6= 0.
(b) Define "˜⇤1 = f + "˜1 and "˜⇤2 = (2⇡ 1)f + ⇡ "˜1 + (1 ⇡)˜
"2 . Then
6.8. Assume there are H investors with CARA utility and the same absolute risk aversion ↵.
Assume there is a risk-free asset. Assume there are two risky assets with payo↵s x̃i that are joint
normally distributed with mean vector µ and nonsingular covariance matrix ⌃. Assume HU < H
investors are unaware of the second asset and invest only in the risk-free asset and the first risky
asset. If all investors invested in both assets (HU = 0), then the equilibrium price vector would be
1 ↵
p⇤ = µ ⌃ ✓¯ ,
Rf HRf
where ✓¯ is the vector of supplies of the risky assets (see Exercise 3.4). Assume HU > 0, and set
HI = H HU .
(a) Show that the equilibrium price of the first asset is p1 = p⇤1 , and the equilibrium price of the
second asset is ✓ ◆✓ ◆
↵ HU cov(x̃1 , x̃2 )2
p2 = p⇤2 var(x̃2 ) < p⇤2 .
HRf HI var(x̃1 )
(b) Show that there exist A > 0 and such that
cov(R̃1 , R̃m )
E[R̃1 ] = Rf + ,
var(R̃m )
cov(R̃2 , R̃m )
E[R̃2 ] = A + Rf + ,
var(R̃m )
= E[R̃m ] Rf A⇡2 ,
where ⇡2 = p2 ✓¯2 /(p1 ✓¯1 + p2 ✓¯2 ) is the relative date–0 market capitalization of the second risky
asset. Note that is less than in the CAPM, and the second risky asset has a positive “alpha,”
relative to .
54 6 Beta Pricing Models
(a) The optimal portfolio of investors who do not invest in the second risky asset is ✓U = (✓U 1 0)0 ,
where
E[x̃1 ] Rf p1
✓U 1 = .
↵ var(x̃1 )
The optimal portfolio of investors who invest in both risky assets satisfies
↵⌃✓I = µ Rf p .
Let ij denote the (i, j)–th element of ⌃. The market clearing condition ✓¯ = HI ✓I + HU ✓U
implies
1 0
HU @ 1 0
↵⌃ ✓¯ = HI ↵⌃✓I + HU ↵⌃✓U = HI (µ Rf p) + ⌃ A (µ Rf p)
11 0 0
0 1
H 0
=@ A (µ Rf p) .
HU 12 / 11 HI
Thus,
0 1 1
H 0
µ Rf p = ↵ @ A ⌃ ✓¯
HU 12 / 11 HI
0 1
1/H 0
= ↵@ A ⌃ ✓¯
HU
HHI 1/HI 12
11
2 0 1 0 13
1 1 0 H U @ 0 0
= ↵4 @ A+ A5 ⌃ ✓¯
H 0 1 HHI 12
1
11
0 1
↵ H U ↵ @0 0
= ⌃ ✓¯ + 2
A ✓¯
H HHI 0 12
22 11
Hence, 0 1
✓ ◆✓ 2
◆
1 ↵ ↵ HU 0
p= µ ⌃ ✓¯ 22
12 @ A.
Rf HRf HRf HI 11 ✓¯2
Note that p2 < p⇤2 because 11 12 > 2
12 .
Hence,
= E[R̃m ] Rf ⇡2 A < E[R̃m ] Rf .
6.9. Suppose there is no risk-free asset and the minimum-variance return is di↵erent from the
constant-mimicking return, i.e., bm 6= bc . From Section 6.2, we know there is a beta pricing model
with the constant-mimicking return as the factor:
for every return R̃. From Section 6.6, we can conclude there is a stochastic discount factor that is an
affine function of the constant-mimicking return unless ↵ = 0. However, the existence of a stochastic
discount factor that is an affine function of the constant-mimicking return would contradict the
result of Section 5.11. So, it must be that ↵ = 0 in (6.25). Calculate ↵ to demonstrate this.
56 6 Beta Pricing Models
We have
Thus,
E[R̃] = ↵ + cov(R̃, R̃p + bc ẽp )
implies
E[R̃]{1 bc + E[R̃p + bc ẽp ]} = ↵ .
6.10. Suppose there is no risk-free asset and the minimum-variance return is di↵erent from the
constant-mimicking return, i.e., bm 6= bc . From Section 5.11, we know that there is a stochastic
discount factor that is an affine function of the minimum-variance return:
m̃ = + (R̃p + bm ẽp )
for some and . From Section 6.6, we can conclude there is a beta pricing model with the
minimum-variance return as the factor unless E[m̃] = 0. However, this would contradict the result
of Section 6.2. So it must be that E[m̃] = 0 for the stochastic discount factor m̃ in (6.26). Calculate
E[m̃] to demonstrate this.
We have
0 = E[m̃ẽp ]
= E[ẽp ] + bm E[ẽp ]
E[R̃p ]E[ẽp ]
= E[ẽp ] +
1 E[ẽp ]
(1 E[ẽp ]) + E[R̃p ]
= E[ẽp ]
1 E[ẽp ]
= E[ẽp ]E[m̃] ,
using Facts 8 and 16 for the third equality. Because E[ẽp ] 6= 0 (by Fact 17), this implies E[m̃] = 0.
7
Representative Investors
7.1. Show that if uh0 and uh1 are concave for each h, then the social planner’s utility functions û0
and û1 are concave.
Consider û0 . The argument is identical for û1 . For convenience, drop the subscript 0 from û0 and
uh0 . Consider two aggregate consumption levels c1 and c2 . Let {ch1 | h = 1, . . . , H} and {ch2 | h =
P PH
1, . . . , H} satisfy Hh=1 ch1 = c1 and h=1 ch2 = c2 . Let 0 < < 1. We have
H
X
û( c1 + (1 )c2 ) h uh ( ch1 + (1 )ch2 )
h=1
XH H
X
h uh (ch1 ) + (1 ) h uh (ch1 ) ,
h=1 h=1
PH
using the definition of û — and the fact that h=1 ch1 + (1 )ch2 = c1 + (1 )c2 — for the first
inequality and concavity of the uh for the second. Because this is true for every {ch1 | h = 1, . . . , H}
P PH
satisfying H h=1 ch1 = c1 and for every and {ch2 | h = 1, . . . , H} satisfying h=1 ch2 = c2 , we have
7.2. Use the results on affine sharing rules in Section 3.6 to establish (b0 ) in Section 7.2.
ch = ⇣h + bh (c ⇣) ,
where
1/⇢
bh = P h 1/⇢
.
H
j=1 j
60 7 Representative Investors
Thus, the representative investor’s utility for consumption at dates 0 and 1 is, up to an affine
transform,
1
(c0 ⇣)1 ⇢
+ (c1 ⇣)1 ⇢
.
1 ⇢ 1 ⇢
7.3. Suppose there is a representative investor with quadratic utility u(w) = (⇣ w)2 . Assume
E[w̃m ] 6= ⇣. Show that in the CAPM (6.12) equals
var(w̃m )
,
E[⌧ (w̃m )]
where ⌧ (w) denotes the coefficient of risk tolerance of the representative investor at wealth level w.
Thus, the risk premium is higher when market wealth is riskier or when the representative investor
is more risk averse.
We have u0 (w) = 2(⇣ w), u00 (w) = 2, and ⌧ (w) = u0 (w)/u00 (w) = ⇣ w. There is a stochastic
discount factor m̃ = u0 (w̃m ) = 2 (⇣ w̃m ) for some , and E[m̃] 6= 0 by virtue of the assumption
E[w̃m ] 6= ⇣. Thus,
7 Representative Investors 61
1 1
E[R̃] = cov(m̃, R̃)
E[m̃] E[m̃]
for each return R̃. We have E[m̃] = 2 E[⌧ (w̃m )] and cov(m̃, R̃) = 2 cov(w̃m , R̃). Therefore,
1 1
E[R̃] = + cov(m̃, R̃)
2 E[⌧ (w̃m )] E[⌧ (w̃m )]
1 var(w̃m ) cov(m̃, R̃)
= + .
2 E[⌧ (w̃m )] E[⌧ (w̃m )] var(w̃m )
Therefore, = var(w̃m )/E[⌧ (w̃m )].
7.4. Suppose returns and end-of-period endowments are joint normally distributed and there is a
representative investor with constant absolute risk aversion ↵. Show that in the CAPM (6.12)
equals ↵ var(w̃m ). Thus, the risk premium is higher when market wealth is riskier or when the
representative investor is more risk averse.
We have
1 1
E[R̃] = cov(m̃, R̃) ,
E[m̃] E[m̃]
where m̃ = ↵e ↵w̃m for some . By Stein’s Lemma,
↵w̃m
cov( ↵e , R̃) = ↵2 E[e ↵w̃m
] cov(w̃m , R̃) .
Thus,
1
E[R̃] = ↵w̃m ]
+ ↵ cov(w̃m , R̃)
↵E[e
1 cov(w̃m , R̃)
= ↵w̃m ]
+ ↵ var(w̃m ) .
↵E[e var(w̃m )
This implies = ↵ var(w̃m ).
7.5. Assume in (7.11) that log R̃ and log(c̃1 /c0 ) are joint normally distributed. Specifically, let
R̃ = eỹ and c̃1 /c0 = ez̃ with E[ỹ] = µ, var(ỹ) = 2, E[z̃] = µc , var(z̃) = 2
c, and corr(ỹ, z̃) = .
Note that c is the covariance of the continuously compounded return y with the continuously
compounded consumption growth rate z, so (7.21b) has the usual form
with ✓ = ⇢, except for the extra term 2 /2. The extra term, which involves the total and hence
idiosyncratic risk of the return, is usually called a Jensen’s inequality term, because it arises from
2 /2
the fact that E[eỹ ] = eµ+ > eµ .
2 /2 1 1 ⇣ ⌘
⇢2 2
i cov R̃, c̃1 ⇢ .
eµ+ = e⇢µc c /2 h
E c̃1 ⇢
⇣ ⌘ h i h i h i
Substituting cov R̃, c̃1 ⇢ = E R̃c̃1 ⇢ E R̃ E c̃1 ⇢ yields
h i
2 /2 1 E R̃c̃1 ⇢
⇢2 2 2 /2
eµ+ = e⇢µc c /2 h i + eµ+ .
E c̃1 ⇢
Also,
h i
⇥ ⇤
E R̃c̃1 ⇢ E eỹ ⇢z̃
h i = ⇢z̃ ]
E c̃1 ⇢ E [e
exp µ ⇢µc + 12 ( 2 2⇢ c + ⇢2 2
c)
=
exp ⇢µc + 12 ⇢2 2
c
2 /2
= eµ ⇢ c+
.
Thus,
2 /2 1 ⇢2 2
eµ ⇢ c+
= e⇢µc c /2 ,
which implies
1 2 2 1 2
µ= log + ⇢ c + ⇢µc ⇢ c .
2 2
7 Representative Investors 63
(b) From
1 c0 ⇢
Rf = = h i,
E[m̃] E c̃1 ⇢
h i
and E c̃1 ⇢ = c0 ⇢ e ⇢µc +⇢2 2 ⇢2 2
c /2 , we obtain Rf = e⇢µc c /2 / . Hence, r = log +⇢µc ⇢2 2
c /2.
7.6. Suppose there is a risk-free asset and a representative investor with power utility, so (7.10)
is a stochastic discount factor. Let z̃ = log(c˜1 /c0 ) and assume z̃ is normally distributed with mean
µc and variance 2
c. Let denote the maximum Sharpe ratio of all portfolios. Use the Hansen-
Jagannathan bound (4.13) to show that
p
log(1 + 2 )
⇢ .
c
Hint: Apply the result of Exercise 1.15. Note that (7.22) implies risk aversion must be larger if
consumption volatility is smaller or the maximum Sharpe ratio is larger. Also, using the approxi-
mation log(1 + x) ⇡ x (a first-order Taylor series approximation of log(1 + x) around x = 0), the
lower bound on ⇢ in (7.22) is approximately / c .
We have
stdev(m̃) stdev e ⇢z̃ p
= ⇢z̃
= e ⇢2 2
c 1,
E[m̃] E [e ]
using the result of Exercise 1.15. Thus, the Hansen-Jagannathan bound (4.13) implies
p
e ⇢2 2
c 1 ,
which implies
2 2
e⇢ c 1 + 2 ,
and p
log(1 + 2 )
⇢ .
c
7.7. Suppose there is a risk-free asset in zero net supply and the risky asset returns have a factor
structure
R̃i = ai + b0i F̃ + "˜i ,
where the "˜i have zero means and are independent of each other and of F̃ . Assume there are no
end-of-period endowments and there is a representative investor with CARA utility. Let ↵ denote
64 7 Representative Investors
the risk aversion of the representative investor. Let ⇡ ⇤ denote the vector of market weights. Denote
initial market wealth by w0 and end-of-period market wealth by w̃m = w0 R̃m . Let i denote the
APT pricing error defined in Section 6.7. Assume "˜i with probability one, for some constant .
Via the following steps, show that
Hint: Use an exact first-order Taylor series expansion of the exponential function.
(a) There is a stochastic discount factor m̃ = ⌘↵e ↵w̃m for some ⌘. Because E[m̃] = 1/Rf ,
1
⌘↵ = ↵w̃m ]
,
Rf E[e
so
e ↵w̃m
m̃ = .
Rf E[e ↵w̃m ]
This implies
E[e ↵w̃m "˜i ]
i = .
Rf E[e ↵w̃m ]
7 Representative Investors 65
and the independence of "˜i from F̃ and from R̃j for j 6= i, we obtain
P
↵w̃m ↵(w0 ⇡j⇤ R̃j +w0 ⇡i⇤ ai +w0 ⇡i⇤ b0i F̃ ) ↵w0 ⇡i⇤ "˜i
E[e "˜i ] = E[e j6=i ]E[e "˜i ] ,
P
↵w̃m ↵(w0 ⇡j⇤ R̃j +w0 ⇡i⇤ ai +w0 ⇡i⇤ b0i F̃ ) ↵w0 ⇡i⇤ "˜i
E[e ] = E[e j6=i ]E[e ].
Thus,
E [exp( ↵w0 ⇡i⇤ "˜i )˜
"i ]
i = ⇤ .
Rf E [exp( ↵w0 ⇡i "˜i )]
(c) The random variable ↵w0 ⇡i⇤ "˜i has mean zero and the exponential function is convex, so
Jensen’s inequality implies
↵w0 ⇡i⇤ "˜i
E[e ] e0 = 1 .
ex = 1 + e y x
for some y between x and 0. Applying this with x = ↵w0 ⇡i⇤ "˜i gives
↵w0 ⇡i⇤ ˜ 2
E [exp( ↵w0 ⇡i⇤ "˜i )˜
"i ] = ↵w0 ⇡i⇤ E[e "˜i ] .
↵w0 ⇡i⇤ ˜ ⇤
e e↵w0 ⇡i
Hence,
⇤
|E [exp( ↵w0 ⇡i⇤ "˜i )˜
"i ]| ↵w0 ⇡i⇤ e↵w0 ⇡i var(˜
"i ) .
Part II
Dynamic Models
8
8.1. Suppose there is a risk-free asset with constant return Rf each period. Suppose there is a
single risky asset with dividends given by
8
>
< h Dt with probability 1/2 ,
Dt+1 =
>
: D
` t with probability 1/2 ,
where h > ` are constants, and D0 > 0 is given. Suppose the price of the risky asset satisfies
Pt = kDt for a constant k. Suppose the information in the economy consists of the history of
dividends, so the information structure can be represented by a tree as in Figure ?? with two
branches emanating from each node (corresponding to the outcomes h and `). For each date t > 0
⌫t t ⌫t
and each path, let ⌫t denote the number of dates s t such that Ds = h Ds 1 , so Dt = D0 h ` .
Recall that, for 0 n t, the probability that ⌫t = n is the binomial probability
t t!
2 .
n!(t n)!
(a) State a condition implying that there are no arbitrage opportunities for each finite horizon T .
(b) Assuming the condition in part (a) holds, show that there is a unique one-period stochastic
discount factor from each date t to t + 1, given by
8
>
<zh if Dt+1 /Dt = h,
Zt+1 =
>
:z if D /D =
` t+1 t `,
(c) Assuming the condition in part (a) holds, show that there is a unique SDF process M , and
show that Mt depends on ⌫t and the parameters Rf , k, h and `.
70 8 Dynamic Securities Markets
(d) Assuming the condition in part (a) holds, show that there is a unique risk neutral probability
measure for any given horizon T < 1, and show that the risk neutral probability of any path
depends on ⌫t and the parameters Rf , k, h and `.
(a) This is a dynamic version of Exercise 4.1. The return on the risky asset when Dt+1 = h Dt is
Pt+1 + Dt+1 (k + 1)Dt+1 k+1
= = h.
Pt kDt k
Likewise, the return when Dt+1 = ` Dt is (k + 1) ` /k. Define
k+1 k+1
Rh = h, R` = `.
k k
The necessary and sufficient condition for absence of arbitrage opportunities is
R h > Rf > R` .
(b) As in Exercise 4.1, the unique one-period conditional state prices are
Rf R` Rh Rf
qh = , q` = .
Rf (Rh R` ) Rf (Rh R` )
The unique one-period stochastic discount factor is obtained by dividing the state prices by the
probabilities, so zh = 2qh , and z` = 2q` .
(c) The SDF process is obtained by compounding the one-period stochastic discount factors. This
produces
Mt = 2t qh⌫t q`t ⌫t
.
8 Dynamic Securities Markets 71
(d) There are 2T distinct events (paths) that can be distinguished at date T, each of which has
probability 2 T. The risk neutral probability of any path is
⌫t t ⌫t
Given that Dt = D0 h ` , this holds for all values of ⌫t if and only if
✓ ◆ 1/⇢ ✓ ◆ 1/⇢
2qh 2q`
h = , ` = ,
so
⇢ ⇢
qh = h , q` = ` .
2 2
This implies !
1 2 1
Rf = = ⇢ ⇢ .
qh + q` h + `
Also,
1 = q h Rh + q ` R`
✓ ◆ ✓ ◆
k+1 1 ⇢ k+1 1 ⇢
= h + ` ,
2 k 2 k
This implies ⇣ ⌘
1 ⇢ 1 ⇢
h + `
k= ⇣ ⌘.
1 ⇢ 1 ⇢
2 h + `
72 8 Dynamic Securities Markets
1 ⇢ 1 ⇢ 2
h + ` < ,
Note that ✓ ◆1 ✓ ◆1
⇢ ⇢
D1 Dt
t
Dt1 ⇢ = D01 ⇢ ··· ,
D0 Dt 1
and, by independence,
h i
E t
Dt1 ⇢
= D01 ⇢ t
.
Thus, " #
1
X 1
X
E t
Dt1 ⇢ = D01 ⇢ t
<1 , < 1.
t=1 t=1
8.2. This exercise verifies the assertion in Section 8.10 regarding marketed consumption processes.
Consider any T < 1, and suppose Ct is a marketed date–t payo↵, for t = 0, . . . , T . Show that there
exists a wealth process W and portfolio process ⇡ such that C, W and ⇡ satisfy
for t = 0, . . . , T 1, and CT = WT . Hint: Add up the wealth processes and take a weighted average
of the portfolio processes associated with the individual payo↵s.
For each t T , let Wt,s denote the self-financing wealth process for s 2 {0, . . . , t} such that
Wt,t = Ct . Let ⇡t,s for s 2 {0, . . . , t} denote the corresponding portfolio process, so
0
Wt,s+1 = Wt,s ⇡t,s Rs+1
! T
1 X
(Ws Cs )⇡s0 Rs+1 = (Ws Cs ) PT
0
Wt,s ⇡t,s Rs+1
t=s+1 Wt,s t=s+1
T
X
= Wt,s+1
t=s+1
= Ws+1 .
8.3. This exercise shows a sense in which the log-optimal portfolio is asymptotically superior to
any other portfolio.
Suppose the return vectors R1 , R2 , . . . are independent and identically distributed. Let w be a
positive constant. Assume max⇡ E[log(⇡ 0 Rt )] > 1 and let ⇡ ⇤ be a solution to
max E[log(⇡ 0 Rt )] .
⇡
Let W ⇤ be the wealth process defined by the intertemporal budget constraint (8.1) with ⇡t = ⇡ ⇤
and Yt = Ct = 0 for each t and W0⇤ = w. Consider any other portfolio ⇡ for which
Let W be the wealth process defined by the intertemporal budget constraint (8.1) with ⇡t = ⇡ and
Yt = Ct = 0 for each t and W0 = w. Show that, with probability one, there exists T (depending on
the state of the world) such that
Wt⇤ > Wt
for all t T . Hint: Apply the Strong Law of Large Numbers to (1/T ) log WT⇤ and to (1/T ) log WT .
⇤
log Wt+1 = log Wt⇤ + log (⇡ ⇤ )0 Rt+1 ,
for all t, so
T 1
1 ⇤ log w 1 X
log WT = + log (⇡ ⇤ )0 Rt+1 .
T T T
t=0
Likewise,
T 1
1 log w 1 X
log WT = + log ⇡ 0 Rt+1 .
T T T
t=0
By the Strong Law of Large Numbers,
74 8 Dynamic Securities Markets
T 1
1 X
log (⇡ ⇤ )0 Rt+1 ! E[log((⇡ ⇤ )0 Rt )]
T
t=1
with probability one as T ! 1. Therefore, with probability one, there exists T such that
1 1
log Wt⇤ > log Wt , Wt⇤ > Wt
t t
for all t T.
8.4. This exercise shows that a probability Q consistent with risk-neutral pricing on an infinite
horizon is generally not equivalent to P.
In the setting of Exercise 8.1, let P denote the probability measure under which the random variables
Dt /Dt 1 are independent and identically distributed and equal to h or ` with probability 1/2
each. Assume
Pt+1 + Dt+1
E 6= Rf .
Pt
Suppose there is an infinite horizon. Show that there is no probability measure Q on the space of
infinite paths that is (a) equivalent to P, and (b) satisfies
⇤ Pt+1 + Dt+1
Et = Rf
Pt
for each t. Hint: Apply the Strong Law of Large Numbers to show that any Q satisfying (b) cannot
be equivalent to P.
The condition
Pt+1 + Dt+1
E⇤t = Rf
Pt
implies that the date–t conditional probability of the event Dt+1 /Dt = h is Rf qh , and the date–t
conditional probability of the event Dt+1 /Dt = ` is Rf q` , where qh and q` are the conditional state
prices calculated in Exercise 8.1. This implies that the random variables Dt /Dt 1 are i.i.d. under
Q. By the Strong Law of Large Numbers,
T
1 X Dt Dt
!E
T Dt 1 Dt 1
t=1
8 Dynamic Securities Markets 75
9.1. Consider the problem of maximizing the expected utility of terminal wealth with i.i.d. returns
studied in Section 9.4. Suppose the investor has log utility. Let U = max⇡ E[log ⇡ 0 Rt+1 ]. Assume U
is finite.
(b) Show that the optimal portfolio at each date t is the one that maximizes
E[log ⇡ 0 Rt+1 ] .
= (T s)U + log w .
argmax⇡ E[Vs+1 (w⇡ 0 Rs+1 )] = argmax⇡ {(T s 1)U + log w + E[log(⇡ 0 Rs+1 )]} .
9.2. Consider the finite horizon problem with intermediate consumption and i.i.d. returns studied
in Section 9.5. Suppose the investor has log utility. Assume max⇡ E[log(⇡ 0 Rt+1 )] is finite. Define
1
⇠t = T +1 t
.
1
78 9 Portfolio Choice by Dynamic Programming
E[log ⇡ 0 Rt+1 ] .
C t = ⇠t Wt .
9.3. Consider a finite horizon problem with intermediate consumption, state variables Xt , log
utility, and no endowments Yt . Assume max⇡ Et [log(⇡ 0 Rt+1 )] is finite for each t with probability
one. The value function at date T is
JT (x, w) = log w .
Define
1
⇠t = T +1 t
.
1
(a) Show that
1
Jt (x, w) = log w + Jt (x, 1) .
⇠t
(b) Show that the optimal portfolio at each date t is the one that maximizes
Et [log ⇡ 0 Rt+1 ] .
C t = ⇠t Wt .
(a) The proof is by induction. We have JT (x, w) = log w, which implies JT (x, 1) = 0. Also, ⇠T = 1.
Therefore,
1
JT (x, w) = log w + JT (x, 1) .
⇠T
Suppose the claim is true for t = s + 1. We have
⇢
⇥ ⇤
Js (x, w) = max log c + Es Js+1 (Xs+1 , (w c)⇡ 0 Rs+1 )
c,⇡
⇢
= max log c + Es log((w c)⇡ 0 Rs+1 ) + Es [Js+1 (Xs+1 , 1)]
c,⇡ ⇠s+1
⇢
⇥ ⇤
= max log c + log(w c) + Es log(⇡ 0 Rs+1 ) + Es [Js+1 (Xs+1 , 1)]
c,⇡ ⇠s+1 ⇠s+1
⇢
⇥ ⇤
= max log c + log(w c) + max Es log(⇡ 0 Rs+1 ) + Es [Js+1 (Xs+1 , 1)] .
c ⇠s+1 ⇠s+1 ⇡
80 9 Portfolio Choice by Dynamic Programming
⇠s+1
c= w = ⇠s w .
+ ⇠s+1
1
log c + log(w c) = log ⇠s + log w + log(1 ⇠s ) .
⇠s+1 ⇠s ⇠s+1
9.4. Consider the infinite horizon problem with i.i.d. returns studied in Section 9.6. Suppose the
investor has log utility. Assume max⇡ E[log ⇡ 0 Rt+1 ] is finite.
log w
J(w) = +A
1
ˆ log w
J(w) = +A
1
for a constant A. Thus, this solution of the Bellman equation is the value function J.
(c) Show that the portfolio attaining the maximum in the Bellman equation, when Jˆ = J, is the
one that maximizes
E[log ⇡ 0 Rt+1 ] .
(d) Show that the consumption attaining the maximum in the Bellman equation is
c = (1 )w .
9 Portfolio Choice by Dynamic Programming 81
lim T
E[J(WT⇤ )] = 0
T !1
holds.
Then
Thus,
log w
J(w) = + A,
1
where we define
" 1 t 1
!#
X Y
A= max E t
log ⇠t (1 ⇠s )⇡s0 Rs+1 .
(⇠t ,⇡t )t=0,1,...
t=0 s=0
Note: This optimization problem can easily be solved directly, without using dynamic program-
ming, to compute the optimal ⇠t and ⇡t .
(b) The Bellman equation is
⇢ h i
ˆ
J(w) = max ˆ
log c + E J((w c)⇡ 0 Rt+1 ) .
c,⇡
82 9 Portfolio Choice by Dynamic Programming
ˆ
Substituting J(w) = A + (log w)/(1 ) yields
⇢
log w log ((w c)⇡ 0 Rt+1 )
+ A = max log c + E +A
1 c,⇡ 1
⇢
log(w c) ⇥ ⇤
= A + max log c + + max E log ⇡ 0 Rt+1
c 1 1 ⇡
log log w U
= A + log(1 )+ + + ,
1 1 1
where we use the fact that the maximum in c occurs at c = (1 )w, and where we set
⇥ ⇤
U = max E log ⇡ 0 Rt+1 .
⇡
Thus,
log U
A = A + log(1 )+ + .
1 1
This has the unique solution
(1 ) log(1 ) + log + U
A= .
(1 )2
= WT⇤ 0
1 ⇡T 1 RT
= ...
TY1
= W0 T
⇡t0 Rt+1 .
t=0
This implies
T
X1
log WT⇤ = log W0 + T log + log ⇡t0 Rt+1 .
t=0
Now we have
9 Portfolio Choice by Dynamic Programming 83
as T ! 1.
9.5. This exercise computes the value function and optimal portfolio for an investor with finite
horizon and CARA utility for intermediate consumption, assuming i.i.d. returns.
Consider the finite horizon problem with intermediate consumption and i.i.d. returns studied in
Section 9.5. Suppose one of the assets is risk-free with return Rf . Let R denote the vector of risky
asset returns, let µ denote the expected value of R, and let ⌃ denote the covariance matrix of R.
Assume ⌃ is nonsingular.
where
0w
J0 (w) = 0 e .
Show that
1w
J1 (w) = 1 e
(b) Using dynamic programming, deduce from the result of part (a) that the value function Jt of
an investor with CARA utility and horizon T t has constant absolute risk aversion. How does
the risk aversion depend on the remaining time T t until the horizon?
(c) If the investor has an infinite horizon, then the value function is independent of t. A good guess
would therefore be J(w) = 0 e 0w where 0 and are such that 1 = 0 and = in
0 1 0
part (a); i.e., (0 , 0) is a fixed point of the map (0 , 0) 7! (1 , 1) calculated in part (a). Show
that this implies ✓ ◆
Rf 1
0 = ↵.
Rf
84 9 Portfolio Choice by Dynamic Programming
(a) We have
⇢ h i
↵c 0 (R
0 (w c)Rf Rf 1)
J1 (w) = max e 0 E e 0
c,
⇢ h i
↵c 0 (R
0 (w c)Rf Rf 1)
= max e 0 e min E e 0
c
⇢
↵c 0 (w c)Rf
= max e 0 U 0 e ,
c
where we define
h 0 (R
i
Rf 1)
U0 = min E e 0
.
Maximizing in c yields
✓ ◆
0 Rf 1 0 0 R f U 0
c= w log
↵+ 0 Rf ↵+ 0 Rf ↵
and
2 3
✓ ◆ ↵ ✓ ◆ 0 Rf
↵ 0 Rf
4 0 0 R f U 0 ↵+ 0 Rf 0 0 R f U 0 ↵+ 0 Rf
5e ↵+ 0 Rf
w
J1 (w) = + 0 U 0 .
↵ ↵
Thus
↵ 0 Rf
1 = ,
↵ + 0 Rf
✓ ◆ ↵ ✓ ◆ 0 Rf
0 0 Rf U0 ↵+ 0 Rf 0 0 R f U 0 ↵+ 0 Rf
1 = + 0 U 0
↵ ↵
✓ ◆✓ ◆ 0 Rf
↵ + 0 Rf 0 0 Rf U0 ↵+ 0 Rf
= 0 U0 .
↵ ↵
(b) We have
↵w
JT (w) = e
and ⇢
↵c
⇥ 0
⇤
Jt (w) = max E + E Jt+1 ((w c)Rf + (R Rf 1)) .
c,
Thus, applying the result of the previous part successively shows that Jt is CARA utility for
all t. Let ↵⌧ denote the absolute risk aversion of the value function JT ⌧, so ↵0 = ↵. We can
show that
↵
↵ ⌧ = P⌧ s ,
s=0 Rf
✓ ◆
Rf 1
↵
Rf
as ⌧ ! 1. The proof of this is by induction. It is true for ⌧ = 0. Suppose it is true for arbitrary
⌧ . Then we have
↵↵⌧ Rf
↵⌧ +1 =
↵ + ↵ ⌧ Rf
P
↵2 Rf / ⌧s=0 Rf s
= P
↵ + ↵Rf / ⌧s=0 Rf s
↵Rf
= P⌧ s
s=0 Rf + Rf
↵
= P⌧ +1 s
.
s=0 Rf
(c) 1 = 0 implies
↵ 0 Rf
0 = ,
↵ + 0 Rf
of which the solution is ✓ ◆
Rf 1
0 = ↵.
Rf
9.6. This exercise demonstrates that there can be multiple solutions Jˆ of the Bellman equation
in an infinite-horizon positive dynamic programming problem and that attaining the maximum in
the Bellman equation may not produce the optimum.
Suppose there is a single asset that is risk-free with return Rf > 1. Consider an investor with an
infinite horizon, utility function u(c) = c, and discount factor = 1/Rf . Suppose he is constrained
to consume 0 Ct Wt .
(a) Show that the value function for this problem is J(w) = w.
(b) Show that the value function solves the Bellman equation.
ˆ
(c) Show that J(w) = 2w also solves the Bellman equation.
ˆ
(d) Show that, using the true value function J(w) = w in the Bellman equation, the suboptimal
policy Ct = 0 for every t achieves the maximum for every value of w.
(a) This is a complete market, and the investor can choose at each date t any consumption sequence
(Ct , Ct+1 , . . .) satisfying the budget constraint
86 9 Portfolio Choice by Dynamic Programming
1
X
Rft u
C u = Wt
u=t
and the constraint 0 Cu Wu for u t. The last constraint implies finiteness of the sums in
the following.
1
X 1
X
u u t
Vt (w) = max Cu subject to Cu = w
u=t u=t
1
X 1
X
t u t u t
= max Cu subject to Cu = w
u=t u=t
t
= w.
(b) We have
=w
= J(w) .
(c) We have
ˆ
max {c + J((w c)Rf )} = max {c + 2 (w c)Rf )}
0cw 0cw
= 2w
ˆ
= J(w) .
(d) We have
= max w ,
0cw
9.7. This exercise illustrates the fact that the transversality condition (9.25) holds in bounded and
negative dynamic programming.
Consider the infinite horizon problem with i.i.d. returns studied in Section 9.6. Denote the investor’s
utility function by u(c).
9 Portfolio Choice by Dynamic Programming 87
(a) Case B: Assume there is a constant K such that K u(c) K for each c. Show that the
transversality condition (9.25) holds.
(b) Case N: Assume u(c) 0 for each c and J(w) > 1 for each w. Show that the transversality
condition (9.25) holds. Hint: Use (9.28) and the definition of a value function to deduce that
the limit in (9.25) is nonnegative.
as T ! 1.
(b) By the definition of a value function,
" 1
#
X
V0 (w) E u(Ct⇤ ) ,
t=0
so (9.28) implies
lim E[VT +1 (WT⇤ +1 )] 0.
T !1
9.8. This exercise illustrates the fact that (9.29) is a sufficient condition for any solution Jˆ of the
Bellman equation to be the true value function and a sufficient condition for the argmax in the
Bellman equation to be the optimum.
Consider the infinite horizon problem with i.i.d. returns studied in Section 9.6. Denote the investor’s
utility function by u(c). Let Jˆ be a function that solves the Bellman equation. Assume (9.29) holds.
ˆ t )] are finite for each t. Suppose (C ⇤ , ⇡ ⇤ )
For arbitrary decisions (Ct , ⇡t ), assume E[u(Ct )] and E[J(W t t
attain the maximum in the Bellman equation. Show that Jˆ is the value function and (Ct⇤ , ⇡t⇤ ) are
optimal.
88 9 Portfolio Choice by Dynamic Programming
ˆ t)
J(W ˆ t+1 )]
E[u(Ct ) + J(W
ˆ 0)
J(W ˆ 1 )]
E[u(C0 ) + J(W
E[u(c0 ) + u(C1 ) + 2 ˆ 2 )]
E[J(W
···
"T 1 #
X
E t
u(Ct ) + T ˆ T )]
E[J(W
t=0
For the decisions that attain the maximum in the Bellman equation, the first inequality is an
equality, and the transversality condition (9.29b) implies that the last inequality is an equality
also, so Jˆ is the value function, and the decisions are optimal.
10
10.1. Suppose all investors have constant absolute risk aversion, with possibly di↵erent risk aversion
coefficients Ah . Assume there are no endowments Yht , the return vectors Rt are i.i.d. over time, and
there is a risk-free asset with return Rf each period. Assume all investors have infinite horizons,
and assume the value functions are as described in Exercise 9.5.
H H
!
1 X Rf Rf X 1
⇡ = .
↵t (Rf 1)Ah Rf 1 Ah
h=1 h=1
Therefore, ↵t is approximately (Rf 1)/Rf times the aggregate absolute risk aversion of the
utility functions.
11
11.1. Consider a generalization of the model studied in Sects. 11.2–11.5 in which, at each date t,
with µ being a Markov process (observed by the representative investor at each date) and indepen-
dent of "1 , "2 , . . .. This implies that the distribution of µt+1 , µt+2 , . . . at any date t, conditional on
the information at that date, depends only on the value of µt . Adopting the other assumptions in
Section 11.2, show that the market price-dividend ratio at date t depends on µt as follows:
1
" ( s 1
) #
Pt X X
= s t E exp (1 ⇢) µ` µt ,
Dt
s=t+1 `=t
where ✓ ◆
1 2 2
= exp (1 ⇢) .
2
Note: the model analyzed by Mehra and Prescott (1985) is of this form, with µt following a “finite-
state Markov chain.”
Hence,
92 11 Some Dynamic Equilibrium Models
" 1 ✓ ◆1 #
Pt X Ds ⇢
s t
= Et
Dt Dt
s=t+1
1
" s 1
! s
!#
X X X
s t
= Et exp (1 ⇢) µ` exp (1 ⇢) "`
s=t+1 `=t `=t+1
1
" s 1
!#
X X
s t 2 2
= exp (1 ⇢) (s t) Et exp (1 ⇢) µ`
s=t+1 `=t
1
" s 1
!#
X X
= s t Et exp (1 ⇢) µ` .
s=t+1 `=t
11.2. In the model of Sects. 11.2–11.5, the return on the market portfolio is given by (11.5), which
implies log Rmt = log ⌫ + µ + "t , where the "t are i.i.d. standard normals. Using the parameter
values µ = 0.017, = 0.00125, = 0.99 and ⇢ = 10, calculate the standard deviation of the market
return. Note: you will find it is much smaller than the sample standard deviation of 16.54% in the
data studied by Mehra and Prescott (1985). This is a very simple version of the excess volatility
puzzle.
p
The parameter values imply ⌫ = 0.8937. Thus, log Rmt = 0.129 + 0.00125 "t . From the calculation
in the previous exercise, this implies
p
stdev(Rmt ) = e2⇥0.129+2⇥0.0125 e2⇥0.129+0.0125 = 0.040 = 4.0% .
11.3. The model studied in Sections 11.2–11.5 has the properties assumed in Exercise 7.6; hence,
(7.22), derived from the Hansen-Jagannathan bound, must hold. In our current notation, (7.22) is
p
log(1 + 2 )
⇢ ,
where denotes the maximum Sharpe ratio of all portfolios. In particular, the inequality (11.37)
must hold for equal to the market Sharpe ratio. Combining the estimate of the standard deviation
of the market return of 16.54% from Mehra and Prescott (1985) with the historical equity premium
cited in Section 11.4 and the estimate = 0.035, derive a numerical lower bound on ⇢ from (11.37).
Note: you will find this bound is more reasonable than the estimate of 47.6 presented in the text.
This is because of two o↵setting factors: both the risk premium of the market and the volatility
of the market are higher in the data than the model would predict, given reasonable values of
and ⇢.
11 Some Dynamic Equilibrium Models 93
An equity premium of 6.18% and a standard deviation of 16.54% imply a Sharpe ratio of 0.374.
Using = 0.035, we have p
log(1 + 0.3742 )
⇢ = 16.098 .
0.035
11.4. This exercise shows that the condition ⌫ < 1 is necessary for the Bellman equation of a
representative investor to have a solution, in the model of Sections 11.2–11.5.
Consider the infinite-horizon portfolio choice model studied in Section 9.6. Suppose there is a risk-
free asset and a single risky asset with return
1 µ+ "t
Rt = e
⌫
for a sequence of independent standard normals "t , where ⌫ is defined in (11.2). Suppose B 1 ⇢ <1
as in Section 9.6 and that it is optimal for the investor to always hold a zero position in the risk-free
asset. Show that ⌫ < 1.
Let ⇡ denote the weight on the risky asset. It is shown in Section 9.6 that the optimal portfolio is
given by ⇡ solving
1
max E ((1 ⇡)Rf + ⇡R)1 ⇢
.
⇡ 1 ⇢
If the optimum is ⇡ = 1, then
1 1 ⇢ 1
B ⌘ max E ((1 ⇡)Rf + ⇡R)1 ⇢
1 ⇢ ⇡ 1 ⇢
✓ ◆1 ⇢
1 1 ⇢)µ+(1 ⇢)2 2 /2
= e(1
1 ⇢ ⌫
✓ ◆1 ⇢⇣
1 1 ⌫⌘
= .
1 ⇢ ⌫
12.1. Simulate the path of a Brownian motion over a year (using your favorite programming
language or Excel) by simulating N standard normal random variables zi and calculating Bti =
p
B t i 1 + zi t for i = 1, . . . , N , where t = 1/N and B0 = 0. To simulate a standard normal
random variable in a cell of an Excel worksheet, use the formula = NORMSINV(RAND()).
✓ ◆2
dS 1 dS
dY = .
S 2 S
(d) Suppose Xi is a strictly positive process with dXi /Xi = ↵i dt+✓i dBi for i = 1, 2, and Brownian
motions Bi . Define Yt = X1t X2t . Show that
✓ ◆✓ ◆
dY dX1 dX2 dX1 dX2
= + + .
Y X1 X2 X1 X2
(e) Suppose Xi is a strictly positive process with dXi /Xi = ↵i dt+✓i dBi for i = 1, 2, and Brownian
motions Bi . Define Yt = X1t /X2t . Show that
✓ ◆ ✓ ◆✓ ◆
dY dX1 dX2 dX2 2 dX1 dX2
= + .
Y X1 X2 X2 X1 X2
1
dR = f 0 (X) dX + f 00 (X)(dX)2
2
1
= eX r dt + eX (r dt)2
2
= eX r dt .
1
dY = eX dX + eX (dX)2 .
2
Hence, dY /Y = dX + (1/2)(dX)2 .
(c) For f (x) = log x, we have f 0 (x) = 1/x and f 00 (x) = 1/x2 , so
1
dY = f 0 (S) dS + f 00 (S)(dS)2
2
dS 1 (dS)2
=
S 2 S2
✓ ◆
dS 1 dS 2
= .
S 2 S
@f @f @2f @2f
= x2 , = x1 , = 1, = 0.
@x1 @x2 @x1 @x2 @x2i
Hence,
dY = X2 dX1 + X1 dX2 + (dX1 )(dX2 ) .
12 Brownian Motion and Stochastic Calculus 97
This implies ✓ ◆✓ ◆
dY dX1 dX2 dX1 dX2
= + + .
Y X1 X2 X1 X2
(e) For f (x1 , x2 ) = x1 /x2 , we have
Hence,
1 X1 1 X1
dY = dX1 dX2 (dX1 )(dX2 ) + 3 (dX2 )2 .
X2 X22 X22 X2
This implies
✓ ◆✓ ◆ ✓ ◆2
dY X2 dX1 dX2 dX1 dX2 dX2
= dY = + .
Y X1 X1 X2 X1 X2 X2
dS
= µ dt + dB
S
dC
= ↵ dt + ✓ dB
C
for constants ↵ and ✓ and a Brownian motion B. Assume the statistics for aggregate U.S.
consumption reported by Mehra and Prescott (1985):
✓ ◆
Ct+1 Ct+1
Et = 1.018 and stdevt = 0.036 .
Ct Ct
(a) We have
Z t+1 ✓ ◆ Z t+1
1 2
log St+1 = log St + µ ds + dBs
t 2 t
1 2
= log St + µ + (Bu Bt ) .
2
Hence, log St+1 log St is normally distributed with mean µ 2 /2 and variance 2. Thus,
St+1 /St is the exponential of a normally distributed variable with mean µ 2 /2 and variance
2. This is the same assumption as in Exercise 1.15 (with µ in that exercise being µ 2 /2
here). Thus,
✓ ◆ 2⇣ ⌘
St+1 St+1 2
vart = Et e 1
St St
⇣ 2 ⌘
= e2µ e 1 .
we obtain = 0.1537.
(c) From e↵ = 1.018, we obtain ↵ = 0.0178. From
p
e↵ e✓ 2 1 = 0.036 ,
we obtain ✓ = 0.035.
What technical conditions are needed to ensure that St is well defined for t 2 [0, T ]?
Define
Z t✓ ◆ Z t
1 2
Xt = log S0 + µs s ds + s dBs .
0 2 0
Then,
12 Brownian Motion and Stochastic Calculus 99
✓ ◆
1 2
dX = µ dt + dB .
2
Define St = eXt . From part (b) of the preceding exercise, we have
dS
= dX + (1/2)(dX)2
S ✓ ◆
1 2 1
= µ dt + dB + ( dB)2
2 2
= µ dt + dB .
is well defined and finite for each t, with probability one and that
Z t
2
s <1
0
12.5. Assume
Z t
t (t s)
Xt = ✓ e (✓ X0 ) + e dBs
0
dX = (✓ X) dt + dB .
Note: The process X is called an Ornstein-Uhlenbeck process. Assuming > 0, ✓ is called the long-
run or unconditional mean, and is the rate of mean reversion. This is the interest rate process in
the Vasicek model (Section 16.8).
Define
Z t
Zt = es dBs
0
and
t
f (t, z) = ✓ e (✓ X0 z) .
@f t @f t @2f
= e (✓ X0 z) , =e , = 0.
@t @z @z 2
100 12 Brownian Motion and Stochastic Calculus
Hence,
t t
dXt = e (✓ X0 Zt ) dt + e dZt
= (✓ Xt ) dt + dBt .
dX = X dt + dB
for constants ̂, ✓ˆ and ˆ . Note: The squared Ornstein-Uhlenbeck process Y is a special case of
the interest rate process in the Cox-Ingersoll-Ross model (Section 17.2) and a special case of the
variance process in the Heston model (Section 16.8) — special because ̂✓ˆ = ˆ 2 /4.
dY = 2X dX + (dX)2
= 2X 2 dt + 2 X dB + 2 dt
p
= ( 2 2Y ) dt + 2 Y dB
✓ 2 ◆
p
= 2 Y dt + 2 Y dB
2
p
= ̂(✓ˆ Y ) dt + ˆ Y dB ,
where we define
2
̂ = 2 , ✓ˆ = , ˆ=2 .
2
12.7. Suppose dS/S = µ dt + dB for constants µ and and a Brownian motion B. Let r be a
constant. Consider a wealth process W as defined in Section 12.6:
dW dS
= (1 ⇡)r dt + ⇡ ,
W S
where ⇡ is a constant.
(a) By observing that W is a geometric Brownian motion, derive an explicit formula for Wt .
12 Brownian Motion and Stochastic Calculus 101
(a) We have
dW
= [r + ⇡(µ r)] dt + ⇡ dB .
W
Thus, ✓✓ ◆ ◆
1 2 2
Wt = W0 exp r + ⇡(µ r) ⇡ t + ⇡ Bt .
2
(b) For any s < t, we have
✓ ◆
1 2 2
log Wt = log Ws + r + ⇡(µ r) ⇡ (t s) + ⇡ (Bt Bs ) .
2
Hence,
Wt1 ⇢
= e(1 ⇢) log Wt
✓ ✓ ◆ ◆
1 2
= Ws1 ⇢ exp (1 ⇢) r + ⇡(µ r) ⇡ 2
(t s) + (1 ⇢)⇡ (Bt Bs ) .
2
By the usual rule for means of exponentials of normals and the fact that Bt Bs is normally
distributed with mean zero and variance t s, this implies
h i ✓ ✓ ◆ ◆
1 ⇢ 1 ⇢ 1 2 2 1
Es W t = Ws exp (1 ⇢) r + ⇡(µ r) ⇡ (t s) + (1 ⇢)2 ⇡ 2 2 (t s)
2 2
✓ ✓ ◆ ◆
1 ⇢ 1 2 2 1 2 2
= Ws exp (1 ⇢) r + ⇡(µ r) ⇡ + (1 ⇢)⇡ (t s)
2 2
✓ ✓ ◆ ◆
1 ⇢ 1 2 2
= Ws exp (1 ⇢) r + ⇡(µ r) ⇢⇡ (t s) .
2
(c) The objective is to maximize
✓ ✓ ◆◆
1 1 2 2
exp T (1 ⇢) r + ⇡(µ r) ⇢⇡ .
1 ⇢ 2
The first-order condition is
✓ ✓ ◆◆
1 1 2 2 2
exp T (1 ⇢) r + ⇡(µ r) ⇢⇡ T (1 ⇢) µ r ⇢ ⇡ = 0.
1 ⇢ 2
Equivalently, µ r ⇢ 2⇡ = 0, which implies ⇡ = (µ r)/⇢ 2.
102 12 Brownian Motion and Stochastic Calculus
12.8. This exercise shows how to create correlated Brownian motions from independent Brownian
motions.
Let B1 and B2 be independent Brownian motions. Define a stochastic process B̂2 by B̂20 = 0 and
p
dB̂2 = ⇢ dB1 + 1 ⇢2 dB2 for any ⇢ 2 [ 1, 1].
12.9. This exercise shows how to project imperfectly correlated Brownian motions onto each other,
leaving a residual that is a stochastic integral with respect to an independent Brownian motion.
Equation (12.27) identifies ⇢ dB1 as the orthogonal projection of dB2 on dB1 in the sense that
p
dB2 = ⇢ dB1 + 1 ⇢2 dZ2
with Z2 being independent of B1 . The construction of Z1 and Z2 can be continued for n > 2. This
process is Gram-Schmidt orthogonalization.
Let B1 and B2 be imperfectly correlated Brownian motions with correlation process ⇢. Define
Z1 = B1 . Define Z20 = 0 and
1
dZ2 = p (dB2 ⇢ dZ1 ) .
1 ⇢2
Show that Z1 and Z2 are independent Brownian motions.
12.10. This exercise illustrates the implementation of Gram-Schmidt orthogonalization via the
Cholesky decomposition discussed in Sections 4.9 and 12.13.
for i = 1, 2.
(b) Let 0 1
a 0
L=@ A.
b c
Calculate a, b and c with a > 0 and c > 0 such that LL0 = A.
(c) Define Z = (Z1 Z2 )0 by Zi0 = 0 and
0 10 1
11 12 dB1
L dZ = @ A@ A,
21 22 dB2
(a) We have
(dY1 )2 = ( 2
11 + 2
12 ) dt , (dY1 )(dY2 ) = ( 11 21 + 12 22 ) dt , (dY2 )2 = ( 2
21 + 2
22 ) dt .
Hence, 0 1
2 + 2 +
11 12 11 21 12 22
A=@ A.
+ 2 + 2
11 21 12 22 21 22
Hence,
a2 = 2
11 + 2
12 ,
ab = 11 21 + 12 22 ,
b2 + c 2 = 2
21 + 2
22 .
This implies
q
a= 2 2 ,
+ 12
11
11 21 + 12 22
b= p ,
2 + 2
11 12
s
( 2
2 2 11 21 + 12 22 )
c= 21 + 22 2 2
11 + 12
| 11 22 12 21 |
= p .
2 + 2
11 12
Thus,
This implies
12 Brownian Motion and Stochastic Calculus 105
1
dZ1 = ( 11 dB1 + 12 dB2 ) ,
a
1
dZ2 = ( 21 dB1 + 22 dB2 b dZ1 ) .
c
One can calculate directly that (dZ1 )2 = (dZ2 )2 = dt and (dZ1 )(dZ2 ) = 0, but it also follows
from Exercise 12.9 and Part (d) of this exercise.
(d) Substituting for a from the previous part shows that Z1 = B̂1 . The correlation process of B̂1
and B̂2 is ⇢ given by
11 21 + 12 22
⇢ dt = (dB̂1 )(dB̂2 ) = p 2 2 2 2
dt
( 11 + 12 )( 21 + 22 )
Simple algebra shows that
1 1
( 21 dB1 + 22 dB2 ) =p dB̂2
c 1 ⇢2
and
b ⇢
dZ1 = p dZ1 .
c 1 ⇢2
Thus,
1
dZ2 = p (dB̂2 ⇢ dZ1 ) .
1 ⇢2
12.11. This exercise is to express the conditional variance formula (12.25b) and conditional co-
variance formula (12.25d) in terms of processes being martingales. A more general fact, which does
not require the finite variance assumption, and which can be used as the definition of (dMi )2 and
Rt
(dMi )(dMj ), is that 0 (dMis )2 is the finite-variation process such that (12.28a) is a local martingale,
Rt
and 0 (dM1s ) (dM2s ) is the finite-variation process such that (12.28b) is a local martingale.
Suppose dMi = ✓ dBi for i = 1, 2, where Bi is a Brownian motion and ✓i satisfies (12.5), so Mi is
a finite-variance martingale.
being a martingale.
(b) Show that (12.25d) is equivalent
Z t
M1t M2t (dM1s ) (dM2s )
0
being a martingale.
106 12 Brownian Motion and Stochastic Calculus
⇥ ⇤
vars (Mu Ms ) = Es (Mu Ms )2 Es [Mu Ms ] 2
⇥ ⇤
= Es (Mu Ms )2
⇥ ⇤
= Es Mu2 2Ms Mu + Ms2
= Es [Mu2 ] Ms2 ,
which is (12.25b).
(b) We have, for u > s,
covs (M1u M1s , M2u M2s ) = Es [(M1u M1s )(M2u M2s )] Es [M1u M1s ]Es [M2u M2s ]
which is (12.25d).
(a) Use the fact that a Brownian motion has independent zero-mean increments with variance equal
to the length of the time interval to show that Y is a martingale.
(b) Apply Itô’s formula to calculate dY and verify condition (12.5) to show that Y is a martingale.
Hint: To verify (12.5) use the fact that
Z T Z T
E Bt2 dt = E[Bt2 ] dt .
0 0
(c) Let dM = ✓ dB for a Brownian motion B. Use Itô’s formula to show that
Z t
2
Mt (dMs )2
0
is a local martingale.
(d) Let dMi = ✓i dBi for i = 1, 2, and Brownian motions B1 and B2 . Use Itô’s formula to show that
Z t
M1t M2t (dM1s ) (dM2s )
0
is a local martingale.
(a) We want to show that Es [Bt2 t] = Bs2 s. Equivalently, Es [Bt2 ] Bs2 = t s. By the calculation
in Part (a) of the preceding exercise,
@f @f @2f
= 1, = 2x , = 2.
@t @x @x2
Thus,
dYt = df (t, Bt ) = dt + 2Bt dBt + (dB)2 = 2Bt dBt .
108 12 Brownian Motion and Stochastic Calculus
We have
Z T Z T Z T
2
E (2Bt ) dt = 4 E[Bt2 ] dt =4 t dt = 2T 2 < 1 .
0 0 0
Then,
dYt = dZt (dMt )2 .
Thus, dYt = 2Mt dMt , which inherits the local martingale property of M .
(d) Set Zt = M1t M2t and
Z t
Yt = Zt (dM1s ) (dM2s ) .
0
Then
dYt = dZt (dM1t ) (dM2t ) .
Thus,
dYt = M1t dM2t + M2t dM1t ,
12.13. Let dMi = ✓i dBi for i = 1, 2 and Brownian motions B1 and B2 . Suppose ✓1 and ✓2
satisfy condition (12.5), so M1 and M2 are finite-variance martingales. Consider discrete dates
s = t0 < t1 < · · · < tN = u for some s < u. Show that
2 3
N
X
covs (M1u M1s , M2u M2s ) = Es 4 (M1tj M1tj 1 )(M2tj M2tj 1 )5 .
j=1
12 Brownian Motion and Stochastic Calculus 109
Hint: This is true of discrete-time finite-variance martingales, and the assumption that the Mi are
stochastic integrals is neither necessary nor helpful in this exercise. However, it is interesting to
compare this to (12.25d).
Etj 1 [(M1tj M1tj 1 )(M2tj M2tj 1 )] = Etj 1 [M1tj M2tj ] M1tj 1 M2tj 1 .
13.1. Let rd denote the instantaneously risk-free rate in the domestic currency, and let Rd denote
the domestic currency price of the domestic money market account:
✓Z t ◆
d d
Rt = exp rs ds .
0
As in Section 13.12, let X denote the price of a unit of a foreign currency in units of the domestic
currency. Let rf denote the instantaneously risk-free rate in the foreign currency, and let Rf denote
the foreign currency price of the foreign money market account:
✓Z t ◆
f f
Rt = exp rs ds .
0
Suppose M d is an SDF process for the domestic currency, so M f ⌘ M d X/X0 is an SDF process
for the foreign currency. Assume
dX
= µx dt + x dBx
X
for a Brownian motion Bx .
dX ⇤
= (rd rf ) dt + x dBx ,
X
where Bx⇤ is a Brownian motion under the risk neutral probability.
Note: The result of Part (c) is called uncovered interest parity under the risk neutral probability.
Suppose for example that r⇤ < r. Then it may appear profitable to borrow in the foreign currency
and invest in the domestic currency money market. The result states that, under the risk neutral
probability, the cost of the foreign currency is expected to increase so as to exactly o↵set the interest
rate di↵erential.
dY dM f
= + rf dt ,
Y Mf
so
dM f
= rf dt + dZ ,
Mf
where Z is a local martingale defined by dZ = dY /Y .
(b) From the previous part, the formula M f = M d X/X0 , and the formula (13.16a) for the SDF
process M d , we have
dM f
rf dt + dZ =
Mf
✓ ◆✓ ◆
dM d dX dX dM d
= + +
Md X X Md
✓ ◆✓ ◆
d 0 dX dM d
= r dt dB + µx dt + x dBx + .
X Md
For this to be true, the dt terms on each side must match, implying
✓ ◆✓ ◆
d f dX dM d
µx dt = (r r ) dt .
X Md
(c) If M d XRf is a martingale under the physical probability, then XRf /Rd is a martingale under
the risk neutral probability. Setting Y = XRf /Rd , we have
dY dX
= (rf rd ) dt + .
Y X
This process has zero drift under the risk neutral probability, so the drift of dX/X under the
risk neutral probability is (rd rf ) dt. Because volatilities do not change when we make an
equivalent change of measures, we know that
13 Continuous-Time Securities Markets and SDF Processes 113
dX ⇤
= (rd rf ) dt + x dBx
X
for some Brownian motion Bx⇤ under the risk neutral probability. If fact, defining Bx⇤ by
✓ ◆
⇤ 1 dX d f µx r d + r f
dBx = (r r ) dt = dt + dBx ,
x X x
the facts that Bx⇤ has no drift under the risk neutral probability and (dBx⇤ )2 = (dBx )2 = dt
implies Bx⇤ is a Brownian motion under the risk neutral probability, by Levy’s theorem.
13.2. Generalizing from the discrete-time model in Section 8.7, assume an investor’s optimal
consumption satisfies the first-order condition
e t u0 (C
t)
= Mt
u0 (C 0)
for a constant discount factor and stochastic discount factor process M . Assume
1
u(c) = c1 ⇢
.
1 ⇢
(a) Assume (or prove) that the optimal consumption process is an Itô process:
dC
= ↵ dt + ✓ dZ
C
for a Brownian motion Z. Use the fact that the drift of dM/M must be r to show that
1
r = + ⇢↵ ⇢(1 + ⇢)✓2 .
2
(b) Interpreting the investor as a representative investor, give an economic explanation for why
the equilibrium interest rate r should be higher when (i) the discount rate is higher, (ii) the
expected growth rate ↵ of consumption is higher, or (iii) the variance ✓2 of consumption growth
is smaller.
(c) Assume there is a constant risk-free rate and a single risky asset. Assume the dividend-reinvested
price of the risky asset is a geometric Brownian motion:
dS
= µ dt + dB .
S
Assume the Brownian motion B is the only source of uncertainty in the economy. Show that
µ r
✓= .
⇢
114 13 Continuous-Time Securities Markets and SDF Processes
(d) Assume the risk-free rate is consistent with the statistics reported by Mehra and Prescott (1985);
i.e., r = log 1.008. Use the numbers for ↵, ✓, µ and calculated in Exercise 12.3. Show that
they imply ⇢ > 10 and < 0.
and using Levy’s theorem to construct the Brownian motion Z. However, to derive the formula
for r, it is convenient to use
Mt = C0⇢ e t
Ct ⇢
.
Setting f (t, c) = C0⇢ e tc ⇢, we have @f /@t = f , @f /@c = ⇢f /c and @ 2 f /@c2 = ⇢(1+⇢)f /c2 .
Hence,
✓ ◆
dM dC 1 dC 2
= dt ⇢ + ⇢(1 + ⇢)
M C 2 C
✓ ◆
1
= + ⇢↵ ⇢(1 + ⇢)✓2 dt ⇢✓ dZ .
2
dM µ r
= r dt dB .
M
µ r µ r
= ⇢✓ , ✓= .
⇢
(d) Given µ = 0.0675, = 0.1537, ↵ = 0.0178, ✓ = 0.035, and r = log 1.008, from (µ r)/ = ⇢✓,
we obtain ⇢ = 11.07. Given this value of ⇢, the formula
1
r = + ⇢↵ ⇢(1 + ⇢)✓2
2
implies = 0.107.
13 Continuous-Time Securities Markets and SDF Processes 115
13.3. Assume there is a representative investor with utility u(c) = log c and suppose that his
optimal consumption satisfies the first-order condition (13.48). Assume WT = 0, so the wealth of
the investor at any date t is the value of receiving the market dividend from date t to T . Assume
(13.34) is a martingale. Show that the market price-dividend ratio Wt /Ct is given by
Z T
(s t)
e ds .
t
We have, for s t,
Ms e (s t) C
t
=
Mt Cs
and Z T Z T
Ms (s t)
W t = Et Cs ds = Ct e .
t Mt t
13.4. Adopt the same assumptions as in the previous problem, but suppose that (13.49) is the
representative investor’s utility function. Assume that the market dividend is a geometric Brownian
motion:
dC
= ↵ dt + ✓ dZ
C
for constants ↵ and ✓ and a Brownian motion Z. Define
1
⌫= (1 ⇢)↵ + ⇢(1 ⇢)✓2 .
2
Conclude that if the horizon is infinite and ⌫ > 0, then Wt /Ct = 1/⌫.
We have, for s t,
Ms e (s t) C ⇢
t
=
Mt Cs ⇢
and
116 13 Continuous-Time Securities Markets and SDF Processes
Z T ✓ ◆
Wt Ms Cs
= Et ds
Ct Mt Ct
"Zt ✓ ◆1 #
T ⇢
(s t) Cs
= Et e ds .
t Ct
From
dC
= ↵ dt + ✓ dZ ,
C
it follows that ✓✓ ◆ ◆
1 2
Cs = Ct exp ↵ ✓ (s t) + ✓(Zs Zt ) ,
2
so ✓ ◆1 ✓ ✓ ◆ ◆
⇢
Cs 1 2
= exp (1 ⇢) ↵ ✓ (s t) + (1 ⇢)✓(Zs Zt ) .
Ct 2
Given date–t information, this is the exponential of a normally distributed random variable with
mean ✓ ◆
1 2
(1 ⇢) ↵ ✓ (s t)
2
and variance
(1 ⇢)2 ✓2 (s t) .
Therefore,
"Z ✓ ◆1 # Z "✓ ◆1 #
T ⇢ T ⇢
(s t) Cs (s t) Cs
Et e ds = e Et ds
t Ct t Ct
Z T
(s t) ((1 ⇢)↵(s t) (1 ⇢)✓ 2 (s t)/2+(1 ⇢)2 ✓ 2 (s t)/2
= e e ds
t
Z T
⌫(s t)
= e ds
t
1⇣ ⌫(T t)
⌘
= 1 e .
⌫
13.5. Consider an investor with initial wealth W0 > 0 who seeks to maximize E[log WT ]. Assume
the investor must choose among portfolio processes ⇡ satisfying (13.13a) and the following stronger
version of (13.13b):
Z T
E ⇡t0 ⌃t ⇡t dt < 1 .
0
Z T
E ⇡t0 t dBt = 0.
0
Using the formula (13.12) for Wt show that the optimal portfolio process is
⇡t = ⌃t 1 (µt rt 1) .
Hint: the objective function obtained by substituting the formula (13.12) for Wt can be maximized
in ⇡ separately at each date and in each state of the world.
1 0
⇡t0 (µt rt 1) ⇡ ⌃t ⇡t
2 t
⇡t = ⌃t 1 (µt rt 1) .
13.6. This exercise establishes the market completeness result asserted in Section 13.10. It uses
martingale representation under the physical probability measure.
for t < T .
(a) Apply the result of Section 12.7 to deduce that there is a stochastic process such that, for all
t T,
Z t Z t
0
Ms Cs ds + Mt Wt = W0 + dB .
0 0
118 13 Continuous-Time Securities Markets and SDF Processes
(b) Take the di↵erential of the formula in Part (a) to show that
✓ ◆0 ✓ ◆
1 dM
dW = C dt + rW dt + W p + dB (dW ) .
M M
(c) Define
1 1 1 0
= W⌃ (µ r1) + ( ) .
M
Show that (W, C, ) satisfies the intertemporal budget constraint (13.33).
(a) We have
Z t Z t Z T
Ms Cs ds + Mt Wt = Ms Cs ds + Et Ms Cs ds + MT WT
0 0 t
Z T
= Et Ms Cs ds + MT WT ,
0
which is a martingale adapted to the vector B. Hence, the martingale representation theorem
implies
Z t Z t
0
Ms Cs ds + Mt Wt = W0 + dB
0 0
0
M C dt + M dW + W dM + (dW )(dM ) = dB .
so ✓ ◆0 ✓ ◆0
1 1
dW = C dt + rW dt + W p + p dt + W p + dB .
M M
(c) From the definition of and the definition of p,
0 1
(µ r1) = W (µ r1)0 ⌃ 1
(µ r1) + 0 1
(µ r1)
M
0 1 0
=W p p + p,
M
0 1
= W (µ r1)0 ⌃ 1
+ 0 1
M
0 1 0
=W p + .
M
13.7. This exercise establishes the market completeness result asserted in Section 13.10, using
martingale representation under the risk neutral probability.
Adopt the assumptions of Section 13.10. Let M denote the unique stochastic discount factor process.
Assume M R is a martingale. Consider T < 1, and define the probability QT in terms of ⇠T =
MT RT by (13.38). Define B ⇤ by (13.41). Let x be a random variable that depends only on the
path of the vector process B ⇤ up to time T and let C be a process adapted to B ⇤ . Assume
Z T
⇤ Cs x
E ds + < 1.
0 Rs RT
For t T , define Z T
Cs x
Wt⇤ = E⇤t ds + .
t Rs RT
Observe that Z
Z t T
Cs Cs x
ds + Wt⇤ = E⇤t ds + ,
0 Rs 0 Rs RT
which is a QT –martingale.
(a) Apply the result of Section 12.7 to deduce that there is a stochastic process ⌘ such that, for all
t T,
Z t Z t Z t
Cs
ds + Wt⇤ = W0⇤ + 0
⌘ (µ r1) ds + ⌘ 0 dB .
0 Rs 0 0
(b) Set W = RW ⇤ (so, in particular, WT = x) and = R⌘. Apply Itô’s formula and use the result
of the previous part to deduce that W , C and satisfy the intertemporal budget constraint
(13.33).
120 13 Continuous-Time Securities Markets and SDF Processes
By the martingale representation theorem, applied under the risk neutral probability, there
exists such that
Z t Z t
Cs
ds + Wt⇤ = W0⇤ + 0
dB ⇤
0 Rs 0
Z t Z t
= W0⇤ + 0 1
(µ r1) dt + 0
dB .
0 0
Setting ⌘ = 1 0 , we have
Z t Z t Z t
Cs
ds + Wt⇤ = W0⇤ + ⌘ (µ0
r1) dt + ⌘ 0 dB .
0 Rs 0 0
C
dt + dW ⇤ = ⌘ 0 (µ r1) dt + ⌘ 0 dB .
R
Setting W = RW ⇤ yields
1 W
dW ⇤ = dW dR ,
R R2
so
1 W C
dW dR = dt + ⌘ 0 (µ r1) dt + ⌘ 0 dB
R R2 R
C 1 0 1 0
= dt + (µ r1) dt + dB .
R R R
0 0
dW = W r dt C dt + (µ r1) dt + dB .
13.8. This exercise verifies that, as asserted in Section 13.9, condition (13.43) is sufficient for M W
to be a martingale.
Let M be an SDF process such that M R is a martingale. Define B ⇤ by (13.41). Let W be a positive
self-financing wealth process. Define W ⇤ = W/R.
(a) Use Itô’s formula, (13.8), (13.16b) and (13.41) to show that
1
dW ⇤ = 0
dB ⇤ .
R
13 Continuous-Time Securities Markets and SDF Processes 121
(a) We have
dW ⇤ dW dR
⇤
=
W W R
= ⇡ 0 (µ r1) dt + ⇡ 0 dB
= ⇡0 dt + ⇡ 0 dB
= ⇡ 0 dB ⇤ ,
using Itô’s formula, (13.8), (13.16b) and (13.41) successively. This implies
dW ⇤ = W ⇤ ⇡ 0 dB ⇤
W 0
= ⇡ dB ⇤
R
1 0
= dB ⇤ ,
R
where = W ⇡.
p 0
(b) Define ⌘ = ⌃ , Z0 = 0 and
1 0
dZ = dB ⇤ .
⌘
By Levy’s theorem, Z is a Brownian motion under the risk neutral probability. We have
⌘
dW ⇤ = dZ .
R
The sufficient condition (12.5) for W ⇤ to be a martingale under the risk neutral probability is
that Z T
⌘2
E⇤ dt < 1 .
0 R2
This is equivalent to Z T
⇤ 1 0
E ⌃ dt < 1 .
0 R2
122 13 Continuous-Time Securities Markets and SDF Processes
(c) We have
Z T Z T
⇤ 1 0 ⇤ 1 0
E ⌃ dt = E t ⌃t t dt
0 R2 0 Rt2
Z T
MT R T 0
= E t ⌃t t dt
0 Rt2
Z T
Mt 0
= E t ⌃t t dt ,
0 Rt
using iterated expectations and Et [MT RT ] = Mt Rt for the last equality. Thus, (13.43) is a
sufficient condition for W ⇤ = W/R to be a martingale on [0, T ] under the risk neutral probability,
hence a sufficient condition for M W to be a martingale under the physical probability.
13.9. This exercise verifies that Novikov’s condition can be expressed as (13.32), as asserted in
Section 13.7.
For a local martingale Y satisfying dY /Y = ✓0 dB for some stochastic process ✓, Novikov’s condition
is that ⇢ Z T
1
E exp ✓0 ✓ dt < 1.
2 0
Under this condition, Y is a martingale on [0, T ]. Consider Y = M W , where M is an SDF process
and W is a self-financing wealth process.
= ( 0⇡ )0 dB
= ( 0⇡ p ⇣)0 dB ,
using =µ r1 and = p + ⇣.
(b) Setting ✓ = 0⇡ ⇣, we have
p
✓0 ✓ = ⇡ 0 ⌃⇡ + 0
p p + ⇣ 0⇣ 2⇡ 0 p 2⇡ 0 ⇣ + 2 0
p⇣ .
13 Continuous-Time Securities Markets and SDF Processes 123
0
p⇣ = (µ r1)0 ⌃ 1
⇣ = 0,
we obtain
✓0 ✓ = ⇡ 0 ⌃⇡ + 0
p p + ⇣ 0⇣ 2⇡ 0 (µ r1) .
To apply (13.32) to W = Si , take the portfolio process ⇡ to be the i–th basis vector ei . Condition
(13.32) in this case is
⇢ Z T
1 0
E exp p p + ⇣ 0 ⇣ + e0i ⌃ei 2(µi r) dt < 1.
2 0
Note that e0i ⌃ei is the squared volatility of the asset return, i.e., the i–th diagonal element of ⌃.
Suppose W > 0, C and ⇡ satisfy the intertemporal budget constraint (13.33). Define the
consumption-reinvested wealth process W † by (13.36).
C
dY = dt .
W†
Conclude that
Z t
Wt Cs
=1 ds
Wt† 0 Ws†
for all t.
124 13 Continuous-Time Securities Markets and SDF Processes
(c) Let M be an SDF process and assume M W † is a martingale. Use this assumption and iterated
expectations to show that, for any t < T ,
Z T Z t Z T
† Cs † Cs
E t MT W T †
ds = Mt Wt †
ds + Et Ms Cs ds .
0 Ws 0 Ws t
(d) Let M be an SDF process and assume M W † is a martingale. Use the results of the previous
two parts to show that (13.34) is a martingale.
(c) For t < s < T , iterated expectations and the fact that Es [MT WT† ] = Ms Ws† imply
h i
† Cs Cs †
Et MT W T † = Et E s M T W T
Ws Ws†
= Et [Ms Cs ] .
13 Continuous-Time Securities Markets and SDF Processes 125
Therefore,
Z T Z t h i Z T
Cs Cs Cs
Et MT WT† ds = ds ⇥ Et MT WT† + Et MT WT† ds
0 Ws† †
0 Ws t Ws†
Z t Z T
Cs Cs
= Mt Wt† †
ds + Et MT WT† ds
0 Ws t Ws†
Z t Z T
Cs
= Mt Wt† †
ds + Et [Ms Cs ] ds
0 Ws t
Z t Z T
Cs
= Mt Wt† †
ds + Et Ms Cs ds .
0 Ws t
(d) Let Z T
Cs
X t = Etds . MT WT†
Ws† 0
This is a martingale. Using Parts (c) and (b) successively, we have
Z t Z T
† Cs
X t = Mt W t †
ds + Et Ms Cs ds
0 Ws t
Z T
†
= Mt (Wt Wt ) + Et Ms Cs ds .
t
This implies
Z t Z t Z T
Mt W t + Ms Cs ds = Ms Cs ds + Et Ms Cs ds + Et [MT WT ]
0 0 t
Z T
= E t MT W T + Ms Cs ds .
0
13.12. This exercise provides an alternate proof that requiring M W to be a martingale for self-
financing wealth processes W validates the valuation formula (13.350 ) for consumption processes.
It considers reinvesting consumption in the money market account rather than in the portfolio
generating the wealth process.
(a) We have
✓Z t ◆
† Cs
dW = dW + C dt + ds dR
0 Rs
dR
= dW + C dt + (W † W)
R
0 0
= rW dt + (µ r1) dt + dB + (W † W )r dt
= rW † dt + 0
(µ r1) dt + 0
dB .
Thus,
Z T Z T Z T Z t Z T
Cs Cs
Et Ms Cs ds MT R T ds = Et Ms Cs ds Mt R t ds Et Ms Cs ds
0 0 Rs 0 0 Rs t
Z t Z t
Cs
= Ms Cs ds Mt Rt ds
0 0 Rs
14
for each asset i, where Wm denotes the value of the market portfolio, ⇢ = ↵Wm , and ↵ denotes the
p
aggregate absolute risk aversion. Define i = e0i ⌃ei to be the volatility of asset i, as described in
Section 13.2, so we have
dSi
= µi dt + i dZi
Si
for a Brownian motion Zi . Likewise, the return on the market portfolio is
dWm
= µm dt + m dZm
Wm
for some µm , m and Brownian motion Zm . Let im denote the correlation process of the Brownian
motions Zi and Zm .
(a) Using the fact that the market return must also satisfy the continuous-time CAPM relation,
show that the continuous-time CAPM can be written as
i m im
µi r= 2
(µm r) .
m
(b) Suppose r, µi , µm , i, m and ⇢i are constant over a time interval t, so both Si and Wm
are geometric Brownian motions over the time interval. Define the annualized continuously
compounded rates of return over the time interval:
log Si log Wm
ri = and rm = .
t t
128 14 Continuous-Time Portfolio Choice and Beta Pricing
Let r̄i and r̄m denote the expected values of ri and rm . Show that the continuous-time CAPM
implies
cov(ri , rm ) 1
r̄i r= (r̄m r) + [cov(ri , rm ) var(ri )] t.
var(rm ) 2
(a) We have
✓ ◆2
dWm 2
(µm r) dt = ⇢ =⇢ m dt ,
Wm
so ⇢ = (µm r)/ 2
m. Therefore
✓ ◆✓ ◆
µm r dSi dWm
(µi r) dt = 2
m Si Wm
µm r
= 2
( i m im ) dt .
m
1 2 1 2
r̄i = µi i , and r̄m = µm m.
2 2
1 2
r̄i r = µi r i
2
i m im 1 2
= 2
(µm r) i
m 2
✓ ◆
i m im 1 2 1 2
= 2
r̄m + m r i
m 2 2
i m im 1 1 2
= 2
(r̄m r) + i m im i .
m 2 2
Also,
2
i
var(ri ) = ,
t
2
m
var(rm ) = ,
t
i m im
cov(ri , rm ) = .
t
14.2. This exercise derives the ICAPM from the portfolio choice formula (14.24).
For each investor h = 1, . . . , H, let ⇡h denote the optimal portfolio presented in (14.24). Using the
notation of Section 14.7, (14.24) implies
14 Continuous-Time Portfolio Choice and Beta Pricing 129
X̀
W h ⇡h = ⌧ h ⌃ 1
(µ r1) ⌧h ⌘hj ⌃ 1
⌫ 0 ej .
j=1
(a) We have
X̀
⌧h (µ r1) = Wh ⌃⇡h + ⌧h ⌘hj ⌫ 0 ej .
j=1
which implies !
H
X̀ X ⌧h ⌘hj
µ r1 = ↵W ⌃⇡ + ⌫ 0 ej .
⌧
j=1 h=1
X̀
= ↵W ( dB)(dB 0 ) 0 ⇡ + ⌘j ( dB)(dB 0 )⌫ 0 ej
j=1
X̀
= ↵W ⌃⇡ dt + ⌘j ⌫ 0 ej dt .
j=1
14.3. Consider an investor with log utility and an infinite horizon. Assume the capital market line
is constant, so we can write J(w) instead of J(w, x) for the stationary value function defined in
Section 14.10.
130 14 Continuous-Time Portfolio Choice and Beta Pricing
log r + 2 /2
K= + 2
.
holds, where W ⇤ denotes the wealth process generated by the consumption and portfolio pro-
cesses in part (a).
(a) Substituting J = K + log w/ , Jw = 1/( w) and Jww = 1/( w2 ), the HJB equation (14.31) is
⇢
1h ci 1 0
0 = max log c K log w + r + ⇡ 0 (µ r1) ⇡ ⌃⇡ .
c,⇡ w 2
The maximum is achieved at c = w and ⇡ = ⌃ 1 (µ r1). Substituting these into the HJB
equation, it reduces to the formula given for K.
(b) We have
✓ ◆
dW ⇤ 0 C⇤
= r + ⇡ (µ r) dt + ⇡ 0 dB
W⇤ W⇤
= (r + 2 ) dt + (µ r)0 ⌃ 1
dB .
Hence, ✓ ◆
⇤ 1
d log W = r + 2 dt + (µ r)0 ⌃ 1
dB .
2
This implies that
h i ✓ ◆
t T T 1 2
E e log WT = e log W0 + e r + T !0
2
as T ! 1.
14.4. Consider an investor with power utility and an infinite horizon. Assume the capital market
line is constant, so we can write J(w) instead of J(w, x) for the stationary value function defined
in Section 14.10.
14 Continuous-Time Portfolio Choice and Beta Pricing 131
(a) Define
(1 ⇢)r (1 ⇢)2
⇠= .
⇢ 2⇢2
Assume (14.32) holds, so ⇠ > 0. Show that
✓ ◆
⇢ 1 1 ⇢
J(w) = ⇠ w
1 ⇢
solves the HJB equation (14.31). Show that c = ⇠w and ⇡ = (1/⇢)⌃ 1 (µ r1) achieve the
maximum in the HJB equation.
(b) Show that, under the assumption ⇠ > 0, the transversality condition
h i
lim E e t
J(WT⇤ ) = 0
T !1
holds, where W ⇤ denotes the wealth process generated by the consumption and portfolio pro-
cesses in part (a).
(a) Substituting J = ⇠ ⇢ w 1 ⇢ /(1 ⇢), wJw = ⇠ ⇢ w1 ⇢ and w2 Jww = ⇢⇠ ⇢ w1 ⇢ , the HJB equation
(14.31) is
⇢ h
1 ci 1
0 = max c1 ⇢
⇠ ⇢
w1 ⇢
+ r + ⇡ 0 (µ r1) ⇠ ⇢
w1 ⇢
⇢⇠ ⇢
w1 ⇢ 0
⇡ ⌃⇡ .
c,⇡ 1 ⇢ 1 ⇢ w 2
The maximum is achieved at c = ⇠w and ⇡ = (1/⇢)⌃ 1 (µ r1). Substituting these into the
HJB equation, it reduces to the formula given for ⇠.
(b) We have
✓ ◆
dW ⇤ C⇤ 0
= r + ⇡ (µ r) dt + ⇡ 0 dB
W⇤ W⇤
✓ ◆
2 1
= r+ ⇠ dt + (µ r)0 ⌃ 1 dB .
⇢ ⇢
Hence, ✓✓ ◆ ◆
Z T
2 2 1
WT⇤ = W0 exp r ⇠+ T+ (µ r) ⌃ 0 1
dBt .
⇢ 2⇢2 ⇢ 0
This implies
✓ ✓ ◆ Z T ◆
2 2 1 ⇢
(WT⇤ )1 ⇢ = W01 ⇢ exp (1 ⇢) r ⇠+ T+ (µ r) ⌃ 0 1
dBt .
⇢ 2⇢2 ⇢ 0
Thus,
132 14 Continuous-Time Portfolio Choice and Beta Pricing
h i ✓⇢ ✓ ◆ ◆
2 2 (1 ⇢)2 2
E e T
(WT⇤ )1 ⇢
= W01 ⇢
exp + (1 ⇢) r ⇠+ + T
⇢ 2⇢2 2⇢2
✓⇢ ◆
(1 ⇢)2
= W01 ⇢ exp + (1 ⇢)(r ⇠) + T .
2⇢
(1 ⇢)2
+ (1 ⇢)(r ⇠) + < 0.
2⇢
Substituting the formula for ⇠ into this and rearranging shows that it is equivalent to
(1 ⇢)r (1 ⇢)2
> 0.
⇢ 2⇢2
where denotes the maximum Sharpe ratio. Assume portfolio processes are constrained to satisfy
Z T
E ⇡t0 ⌃t ⇡t dt < 1 .
0
Maximizing at each date and in each state of the world as in Exercise 13.5, show that V (t, w, x) =
log w + f (t, x), where Z ✓ ◆
T
1
f (t, x) = E rs + 2s ds Xt = x .
t 2
The assumption Z T
E ⇡s0 ⌃s ⇡s ds < 1 ,
0
implies
Z t
⇡s0 s dBs
0
1 0
⇡s0 (µs rs 1) ⇡ ⌃ s ⇡s
2 s
14.6. This exercise completes the proof in Section 14.3 that the tangency portfolio is optimal when
the capital market line is constant.
Assume the capital market line is constant and the horizon is finite. Define Z by (14.9), W ⇤ by
(14.12), by (14.13), and by (14.14).
(b) Use the result of the previous part and the formula dMp /Mp = r dt dZ to calculate
✓ ◆
dMp
(dW ) ,
Mp
and use the definition (14.14) of to show that W , C and satisfy the intertemporal budget
constraint (13.33).
14.7. Consider a power-utility investor with a finite horizon (case (d) in Section 14.10). Assume the
capital market line is constant and the investor is constrained to always have nonnegative wealth.
Let M = Mp . Calculate the optimal portfolio as follows.
for a constant ↵.
(b) Define Ct and WT from the first-order conditions and set
Z T
Ms MT
W t = Et Cs ds + WT .
t Mt Mt
Show that
1/⇢
Wt = g(t)Mt
where
1 2 1
↵= r(⇢ 1)/⇢ (⇢ 1)/⇢ + 2 (⇢ 1)2 /⇢2 .
2 2
(b) The first order conditions are
⇣ ⌘ 1/⇢
t ⇢ t
e Ct = Mt ) Ct = e Mt ,
WT ⇢ = MT ) WT = ( MT / ) 1/⇢
.
Thus,
Z T
1
Wt = Et Ms Cs ds + MT WT
Mt t
Z T ⇣ ⌘ 1/⇢
1 1 1/⇢
= Et es Ms1 1/⇢
ds + ( / ) 1/⇢
MT
Mt t
Z T⇣ ⌘ 1/⇢ h i h i
1 1 1 1/⇢
= es Et Ms1 1/⇢
ds + ( / ) 1/⇢
E t MT
Mt t Mt
Z T⇣ ⌘ 1/⇢
1/⇢ 1/⇢ ↵(T t)
= Mt es e↵(s t)
ds + ( / ) 1/⇢
Mt e .
t
Hence,
1/⇢
W t = Mt g(t) ,
where
Z T ⇣ ⌘ 1/⇢
s
g(t) = e e↵(s t)
ds + ( / ) 1/⇢ ↵(T t)
e .
t
(c) From Itô’s formula,
dW dg dM 1/⇢
= +
W g M 1/⇢
g0 (1/⇢)M 1 1/⇢ dM + (1/2)(1/⇢)(1 + 1/⇢)M 2 1/⇢ (dM )2
= dt + .
g M 1/⇢
136 14 Continuous-Time Portfolio Choice and Beta Pricing
(1/⇢)M 1 1/⇢ dM 1 dM r 1 0
1/⇢
= = dt + p dB .
M ⇢ M ⇢ ⇢
1 0
p dB ,
⇢
implying
1 1
⇡0 = 0
p = (µ r1)0 ⌃ 1
.
⇢ ⇢
Postmultiplying by 0⌃ 1 shows that this has the unique solution
1
⇡0 = (µ r1)0 ⌃ 1
.
⇢
14.8. This exercise derives a linear partial di↵erential equation (PDE) for the optimal wealth of
a power utility investor in complete markets. This specific example is due to Wachter (2002). In
general this approach leads to a PDE for a function of M and the state variables that influence M ,
but only the state variable appears here due to homotheticity (power utility).
Suppose the risk-free rate r is constant. Suppose there is a single risky asset. Suppose the asset
does not pay dividends prior to T and its price S satisfies
dSt
= µt dt + t dBt
St
d t = (✓ t ) dt + dB
for constants , ✓ and , where B is the same Brownian motion that appears in dS/S. Define Ct
and WT by the first-order conditions (14.6), and set
Z T
Ms MT
W t = Et Cs ds + WT
t Mt Mt
for t < T .
14 Continuous-Time Portfolio Choice and Beta Pricing 137
must be zero, derive a PDE that must be satisfied by f and its partial derivatives ft , f and
f .
(d) Define the portfolio weight on the risky asset by
✓ ◆
1 f
⇡= + .
f ⇢
Show that (C, W, ⇡) satisfies the intertemporal budget constraint. Hint: Follow the steps in
Rt
Exercise 14.6, using the fact that the drift of 0 Ms Cs ds + Mt Wt is zero, i.e., f satisfies the
PDE.
WT ⇢ = MT ) WT = ( MT / ) 1/⇢
.
Thus,
Z T
1
Wt = Et Ms Cs ds + MT WT
Mt t
Z T
1/⇢ 1 1 1/⇢
= Et e s/⇢ Ms1 1/⇢ ds + 1/⇢ MT
Mt t
"Z ✓ ◆ ✓ ◆1 #
1/⇢ 1/⇢
T
Ms 1 1/⇢ MT 1/⇢
= Mt Et e s/⇢ ds + 1/⇢
.
t Mt Mt
138 14 Continuous-Time Portfolio Choice and Beta Pricing
We have
dM
= r dt dB ,
M
so, for s > t, ✓ ◆
Z s Z s
Ms 1 2
= exp r(s t) u du u dBu .
Mt 2 t t
depend only on t.
(b) Writing M 1/⇢ = eX , where X = log M/⇢ and using deX /eX = dX + (dX)2 /2, we have
1/⇢
✓ ◆ 2
dM 1 1 2
1/⇢
= r dt dt dB + 2 dt
M ⇢ 2 2⇢
2⇢r + (1 + ⇢) 2
= dt + dB .
2⇢2 ⇢
Also,
1
df = ft dt + f d + f (d )2
2
1 2
= ft dt + f [(✓ ) dt + dB] + f dt
2
1 2
= ft + (✓ )f + f dt + f dB .
2
Hence,
!✓ ◆
dW dM 1/⇢ df dM 1/⇢ df
= 1/⇢
+ + 1/⇢
W M f M f
1 2 ✓ ◆
2⇢r + (1 + ⇢) 2 ft + (✓ )f + 2 f f f
= 2
dt + dt + + dB + dt .
2⇢ f ⇢ f ⇢f
Rt
(c) The di↵erential of 0 Ms Cs ds + Mt Wt is
✓ ✓ ◆✓ ◆◆
C dW dM dW dM
M C dt + M dW + W dM + (dW )(dM ) = M W + + +
W W M W M
t 1/⇢ ✓ ◆ !
e M dW f
= MW 1/⇢ f
+ r dt dB + dt .
M W ⇢ f
Using the previous formula for dW/W and equating the drift to zero yields
t 1/⇢ 2 1 2 ✓ ◆
e 2⇢r + (1 + ⇢) ft + (✓ )f + 2 f f f
+ + + r + = 0.
f 2⇢2 f ⇢f ⇢ f
14 Continuous-Time Portfolio Choice and Beta Pricing 139
Thus,
✓ ◆✓ ◆
dW C dM dW dM
= (⇡ ) dB dt
W W M W M
✓ ◆✓ ◆
C dW dM
= dt + r dt + ⇡ dB
W W M
C
= dt + r dt + ⇡ dt + ⇡ dB
W
C
= dt + r dt + ⇡(µ r) dt + ⇡ dB .
W
14.9. This exercise demonstrates the equivalence between the intertemporal and static budget
constraints in the presence of non-portfolio income when the investor can borrow against the income,
as asserted in Section 14.2.
0 0
dW = rW dt + (µ r1) dt + Y dt C dt + dB .
(a) Suppose that (C, W, ) satisfies the intertemporal budget constraint (14.43), C 0, and the
nonnegativity constraint (14.7) holds.
(i) Suppose the horizon is finite. Show that (C, W ) satisfies the static budget constraint
Z T Z T
W0 + E Mt Yt dt E Mt Ct dt + MT WT
0 0
by showing that
Z t
Ms (Cs Ys ) ds + Mt Wt
0
140 14 Continuous-Time Portfolio Choice and Beta Pricing
is a supermartingale. Hint: Show that it is a local martingale and at least as large as the
martingale Xt , where Z T
X t = Et Ms Ys ds .
0
This implies the supermartingale property (Appendix A.13.)
(ii) Suppose the horizon is infinite and limT !1 E[MT WT ] 0. Assume Y 0. Show that the
static budget constraint
Z 1 Z 1
W0 + E Mt Yt dt E Mt Ct dt
0 0
holds.
(b) Suppose the horizon is finite, markets are complete, C 0, and (C, W ) satisfies the static
budget constraint (14.44) as an equality. Show that there exists such that (C, W, ) satisfies
the intertemporal budget constraint (14.43).
M (C Y ) dt + M dW + W dM + (dM )(dW )
= rM W dt + M 0 (µ r1) dt + M 0
dB rM W dt MW 0
dB M 0
dt
0 0
=M dB MW dB .
Thus, it is a local martingale. By the nonnegative wealth constraint (14.7) and the nonneg-
ativity of C, Z
Z t T
Ms Cs ds + Mt Wt + Et Ms Ys ds 0.
0 t
This implies Z
Z t T
Ms (Cs Ys ) ds + Mt Wt Et Ms Ys ds .
0 0
Hence,
Z t
Ms (Cs Ys ) ds + Mt Wt
0
is a supermartingale, which implies
Z T
W0 E Ms (Cs Ys ) ds + MT WT .
0
(ii) Taking the limit of the finite-horizon static budget constraint as T ! 1, using the nonneg-
ativity of C, Y and M and the Monotone Convergence Theorem yields
Z 1 Z 1 Z 1
W0 + E Mt Yt dt E Mt Ct dt + lim E[MT WT ] E Mt Ct dt .
0 0 T !1 0
(b) The proof is the same as in Exercise 13.6, replacing C in that exercise by C Y . Specifically,
defining Z T
Ms MT
W t = Et (Cs Ys ) ds + WT ,
t Mt Mt
it follows that
Z t
Ms (Cs Ys ) ds + Mt Wt
0
is a martingale. Therefore,
Z t Z t
0
Ms (Cs Ys ) ds + Mt Wt = W0 + s dBs
0 0
1 1 1 0
= W⌃ (µ r1) + ( ) ,
M
Derivative Securities
15
Option Pricing
rT 1 2 2 1 ⇥ rT
⇤
S n(d1 ) p re K N(d2 ) + rS N(d1 ) + S n(d1 ) p = r SN (d1 ) e K N(d2 ) .
2 T 2 S T
p
It suffices to note that the two terms on the left-hand side involving the reciprocal of T cancel.
15.2. Use put-call parity to show that a European put and a European call with the same strike
and time to maturity have the same gamma.
Letting C denote the call price and P the put price, put-call parity at date 0 is P = C + e rT K S.
Therefore,
@P @C
= 1,
@S @S
and
@2P @2C
= .
@S 2 @S 2
15.3. Consider an asset with a constant dividend yield q. Assume the price S of the asset satisfies
dS
= (µ q) dt + dB ,
S
where B is a Brownian motion under the physical measure, and µ and are constants. Consider a
European call and a European put with strike K on the asset. Assume the risk-free rate is constant,
and adopt the assumptions of Section 15.3.
(a) Let A denote the event ST > K. Show that E[ST 1A ] = e(µ q)T S ⇤
0 N(d1 ), where
1 2
log(S0 /K) + µ q + T
d⇤1 = p 2
.
T
Hint: This can be computed directly under the physical measure or by changing measures using
e(q µ)T S
T /S0 as the Radon-Nikodym derivative.
p
(b) Show that E[K1A ] = K N(d⇤2 ), where d⇤2 = d⇤1 T.
(c) It follows from the previous parts that the expected return of the European call under the
physical measure, if held to maturity, is
where d1 and d2 are defined in (15.29). Assuming T = 1, µ = 0.12, r = 0.04, q = 0.02, and
= 0.20, show that the expected rate of return of a European call that is 20% out of the money
(S0 /K = 0.8) is 118%.
15 Option Pricing 147
(d) Show that the expected return of the European put under the physical measure, if held to
maturity, is
K N( d⇤2 ) e(µ q)T S0 N( d⇤1 )
.
e rT K N( d2 ) e qT S0 N( d1 )
Assuming T = 1, µ = 0.12, r = 0.04, q = 0.02, and = 0.20, show that the expected rate of
return of a European put that is 20% out of the money (S0 /K = 1.2) is 54%.
= e(µ q)T
S0 [1 N( d⇤1 )]
= e(µ q)T
S0 N(d⇤1 ) ,
p
changing variables as y = x T to obtain the third equality. To compute E[ST 1A ] via the
change of measure, note that
" #
e(q µ)T S
T
E[ST 1A ] = e (µ q)T
S0 E 1A = e(µ q)T
S0 E⇤ [1A ] = e(µ q)T
S0 prob⇤ (A) ,
S0
where the asterisks denote the measure defined by the Radon-Nikodym derivative e(q µ)T S
T /S0 .
dS
= (µ q+ 2
) dt + dB ⇤ ,
S
This is equivalent to
B⇤
pT < d⇤1 .
T
Therefore, prob⇤ (A) = N(d⇤1 ), and E[ST 1A ] = e(µ q)T S ⇤
0 N(d1 ).
(b) We have E[K1A ] = K prob(A). As shown in the solution to Part (a), A is the event that x̃ > d⇤2 ,
where x̃ is a standard normal under the physical measure. Thus, prob(A) = 1 N( d⇤2 ) = N(d⇤2 ),
and E[K1A ] = K N(d⇤2 ).
(c) The expected payo↵ of the call under the physical measure is 0.0305K, and its Black-Scholes
price is 0.0140K. Thus, the expected rate of return is 0.0305/0.0140 1 = 118%.
(d) Denote the complement of A by Ac . The expected payo↵ of the put under the physical measure
is
Thus, the expected return is as claimed. For the parameters given, the expected payo↵ of
the put is 0.0080K, and its Black-Scholes price is 0.0173K, so the expected rate of return is
0.0080/0.0173 1= 54%.
15.4. Adopt the Black-Scholes assumptions. Consider an American put and exercise boundary f
with associated exercise time
8
>
<min{t | St f (t)} if St f (t) for some t 2 [0, T ] ,
⌧=
>
:T otherwise .
Show that there is a trading strategy with value max(0, K S⌧ ) at date ⌧ and value (15.26) at
date 0. Hint: Consider replicating the date–T payo↵ X = er(T ⌧ ) max(0, K S⌧ ).
15 Option Pricing 149
From the market completeness result, there exists a self-financing wealth process W such that
WT = X and Wt = e r(T t) ER [X] for each t 2 [0, T ]. Taking t = 0 yields
t
h i ⇥ ⇤
rT
W0 = e ER er(T ⌧)
max(0, K S⌧ ) = E R e r⌧
max(0, K S⌧ ) .
Taking t = ⌧ yields
h i
r(T ⌧ ) R
W⌧ = e E⌧ er(T ⌧)
max(0, K S⌧ ) = E R
⌧ [max(0, K S⌧ )] = max(0, K S⌧ ) .
15.5. Consider a perpetual American call on an asset with price S given by (15.1) and with a
constant dividend yield q > 0. Assume the risk free rate r is constant. Let V (St ) denote the value
of the call (exercised optimally). Assume V is twice continuously di↵erentiable.
(a) Using the fact that M V is a local martingale, derive an ordinary di↵erential equation (ODE)
that V must satisfy in the continuation region.
(b) Show that AS satisfies the ODE for constants A and if and only if is a root of the quadratic
equation ✓ ◆
1 2 2 1 2
+ r q r = 0.
2 2
Show that the positive root of this equation satisfies > 1.
(c) The general solution of the ODE is A1 S 1 + A2 S 2 , where the i are the roots of the quadratic
equation (15.32). Use the fact that limS#0 f (S) = 0 to show that f (S) = AS for some constant
A, where is the positive root of the quadratic equation.
(d) Use the value matching and smooth pasting conditions to show that the optimal exercise point
is S ⇤ = K/( 1), where is the positive root of the quadratic equation, and to derive A.
(e) Show that, if S0 S ⇤ , the value of the call is
✓ ◆✓ ◆ ✓ ◆
K 1 S0
.
1 K
(a) We have
1
dV = V 0 (S) dS + V 00 (S) (dS)2
2
1 2 2 00
= SV 0 (S)[µ dt + dB] + S V (S) dt .
2
150 15 Option Pricing
Also,
dM µ+q r d"
= r dt db + .
M "
Therefore
✓ ◆✓ ◆
d (M V ) dM dV dM dV
= + +
MV M V M V
µ+q r d" SV 0 1 2 2V
00 SV 0
= r dt dB + + [µ dt + dB] + S dt (µ + q r) dt .
" V 2 V V
1
rV (S) + (r q)SV 0 (S) + S V 00 (S) = 0 .
2 2
2
(b) If V (S) = AS , then SV 0 (S) = AS , and S 2 V 00 (S) = ( 1)AS . Hence, the ODE is
1 2
rAS + (r q)AS + ( 1) AS = 0 .
2
This is satisfied by and A 6= 0 if and only if is a root of the quadratic equation (15.32).
(c) If < 0, we have limS#0 S = 1, so the coefficient Ai corresponding to the negative root of the
quadratic equation must be zero.
(d) The value matching condition is A(S ⇤ ) = S ⇤ K. The smooth pasting condition is A(S ⇤ ) 1 =
1. Multiplying the latter equation by S ⇤ and substituting from the former gives (S ⇤ K) = S ⇤ .
Thus, S ⇤ = K/( 1). This implies
✓ ◆
K K
A = (S ⇤ K)(S ⇤ ) = .
1 1
(e) Substituting the values of A and S ⇤ from the previous part, we have
✓ ◆ ✓ ◆ ✓ ◆
K K K 1 S0
V (S0 ) = AS0 = S0 = .
1 1 1 K
15.6. Consider a perpetual American put under the assumptions of Exercise 15.5. Consider the
exercise time ⌧ = min{t | St S ⇤ } for a constant S ⇤ . Let V (St ) denote the value of the put
(exercised optimally). Assume V is twice continuously di↵erentiable and limS!1 V (S) = 0. Show
that the optimal exercise point is
K
S⇤ = ,
1
where is the negative root of the quadratic equation (15.32). Show that the value of the put is,
for S0 S⇤,
15 Option Pricing 151
✓ ◆ ✓ ◆
K K
.
1 1 S0
V (S) = A1 S + A2 S ,
for constants Ai , where and are the roots of the quadratic equation (15.32). The fact that
limS!1 V (S) = 0 implies that the coefficient Ai corresponding to the positive root must be zero.
Thus, V (S) = AS , where is the negative root. The value matching and smooth pasting conditions
give A(S ⇤ ) = K S ⇤ and A(S ⇤ ) 1 = 1. Multiplying the latter equation by S ⇤ and substituting
from the former gives (K S⇤) = S ⇤ . Thus, S ⇤ = K/(1 ). This implies
✓ ◆
K
A = (K S ⇤ )(S ⇤ ) = K .
1 1
Thus, ✓ ◆ ✓ ◆
⇤ ⇤ K K
V (S0 ) = AS0 = (K S )(S ) = .
1 1 S0
15.7. Consider perpetual American calls and puts under the assumptions of Exercise 15.5. Let ER
denote expectation under the infinite-horizon risk neutral probability. Let ⌧ = min{t | St = S ⇤ } for
a constant S ⇤ . For any stochastic process X,
This follows from Revuz and Yor (1991, Proposition VIII.1.3). Thus, setting Xt = 1/Rt = e rt ,
M1 = 0 and e 1 = 0, we have
ER [e r⌧
] = E[M⌧ ] .
Moreover, because the expected rate of change of S is r q under the risk neutral probability,
8
>
< S0⇤
S if S ⇤ S0
R r⌧
E [e ] = ⇣ ⌘
> S⇤
:
S0 if S ⇤ S0 ,
where is the positive root and the negative root of the quadratic equation (15.32). See Borodin
and Salminen (2000, p. 622). Use these facts to derive the values of perpetual American calls and
puts.
152 15 Option Pricing
dZ
= µz dt + z dBz ,
Z
where Bz is a Brownian motion under the physical measure. Let ⇢ denote the correlation process
of Bz and B.
d⇠ dM
= r dt + dB + .
⇠ M
dZ
= (r + ⇢ z ) dt + z dBz⇤ ,
Z
(a) We have
✓ ◆✓ ◆
d⇠ dM dS dM dS
= + +
⇠ M S M S
dM
= + µ dt + dB (µ r) dt
M
dM
= r dt + dB + .
M
(b) We have
✓ ◆ ✓ ◆
d⇠ dM
(dBz ) = (dBz )(dB) + (dBz )
⇠ M
µz r
= ⇢ dt dt ,
z
dZ
= µz dt + dBz z
Z ✓ ◆
⇤ µz r
= µz dt + z dBz + ⇢ dt
z
= (r + ⇢ z ) dt + z dBz⇤ .
15.9. This exercise verifies the assertion (15.6) regarding the dynamics of an SDF process.
dS
= µ dt + dB s .
S
154 15 Option Pricing
= ⇢ dt ,
so ⇢ = (µ r)/ .
15 Option Pricing 155
(d) We want
dM µ r
= r dt dB s + dZ ,
M
which is equivalent to
dM µ r
dZ = + r dt + dB s
M
µ r
= dB m + dB s .
µ r µ r
(dB s )(dZ) = (dB s )(dB m ) + dt = ⇢ dt + dt = 0 .
15.10. A compound option is an option on an option. Suppose there is a constant risk-free rate,
and the underlying asset price has a constant volatility. Consider a European call option with strike
K 0 maturing at T 0 . Assume the underlying asset does not pay dividends during [0, T 0 ]. Consider a
European option maturing at T < T 0 to purchase the first call option at price K. The purpose of
this exercise is to value the “call on a call.” The value at T of the underlying call is given by the
Black-Scholes formula with T 0 T being the time to maturity. Denote this value by V (T, ST ). The
value at T of the call on a call is
max(0, V (T, ST ) K) .
Let S ⇤ denote the value of the underlying asset price such that V (T, S ⇤ ) = K. The call on a call is
in the money at its maturity T if ST > S ⇤ , and it is out of the money otherwise. Let A denote the
set of states of the world such that ST > S ⇤ , and let 1A denote the random variable that equals 1
when ST > S ⇤ and 0 otherwise. The value at its maturity T of the call on a call is
(a) What is the value at date 0 of receiving the payo↵ K1A at date T ?
(b) To value receiving V (T, ST )1A at date T , let C denote the set of states of the world such
that ST 0 > K 0 , and let 1C denote the random variable that equals 1 when ST 0 > K 0 and 0
otherwise. Recall that V (T, ST ) is the value at T of receiving ST 0 1C K 0 1C at date T 0 . Hence,
the value at date 0 of receiving V (T, ST )1A at date T must be the value at date 0 of receiving
(ST 0 1C K 0 1C )1A at date T 0 . Via the following steps, show that this date–0 value is
p rT 0
p
M(d1 , d01 , T /T 0 ) e K 0 M(d2 , d02 , T /T 0 ) ,
156 15 Option Pricing
where M(a, b, ⇢) denotes the probability that ⇠1 < a and ⇠2 < b, where ⇠1 and ⇠2 are standard
normals with correlation ⇢.
(i) Show that the value at date 0 of receiving V (T, ST )1A at date T is
rT 0
S0 probS (D) e K 0 probR (D) ,
where D = A \ C.
(ii) Show that probS (D) is the probability that
B⇤ B⇤ 0
p T < d1 and pT < d01 ,
T T0
where
log(S0 /S ⇤ ) + r + 12 2 T
d1 = p ,
T
log(S0 /K ) + r + 12
0 2 T0
d01 = p ,
T0
and where B ⇤ denotes a Brownian motion under probS . Note that the random variables in
p
(15.35) are standard normals under probS with a correlation equal to T /T 0 .
(iii) Show that probR (D) is the probability that
B⇤ B⇤ 0
p T < d2 and pT < d02 ,
T T0
where
p
d2 = d1 T,
p
d02 = d01 T0 ,
(a) This is a digital option. Its value is e rT K probR (A). The calculation of probR (A) is the same
calculation made in Section 15.6, replacing K with S ⇤ . Therefore, the value is e rT K N(d2 ).
(b) (i) The random variable (ST 0 1C K 0 1C )1A equals
ST 0 1C 1A K 0 1C 1A = ST 0 1D K 0 1D .
The random variable ST 0 1D is the payo↵ of a share digital paying on the event D. Following
the same calculation as in Section 15.6, the value at date 0 of this share digital is S0 probS (D).
The random variable K 0 1D is the payo↵ of a digital paying on the event D. Following the
same calculation as in Section 15.6, the value at date 0 of this digital is e rT 0 K 0 probR (D).
15 Option Pricing 157
(ii) We have ✓ ◆
1
d log S = r+ 2
dt + dB ⇤
2
where B ⇤ is a Brownian motion under probS . The event D occurs if and only if log ST >
log S ⇤ and log ST 0 > log K. This is equivalent to
✓ ◆
1 2
log S0 + r + T + BT⇤ > log S ⇤ ,
2
✓ ◆
1 2
log S0 + r + T 0 + BT⇤ 0 > log K .
2
Rearranging shows that D is the event (15.35).
(iii) We have ✓ ◆
1
d log S = r 2
dt + dB ⇤
2
where B ⇤ is a Brownian motion under probR . The event D occurs if and only if log ST >
log S ⇤ and log ST 0 > log K. This is equivalent to
✓ ◆
1 2
log S0 + r T + BT⇤ > log S ⇤ ,
2
✓ ◆
1 2
log S0 + r T 0 + BT⇤ 0 > log K .
2
Rearranging shows that D is the event (15.36).
15.11. Calculate the value of a call on a put assuming a constant risk-free rate and a constant
volatility for the underlying asset price. Assume the underlying asset does not pay dividends during
[0, T 0 ], where T 0 is the time to maturity of the put. Assume
r(T 0 T )
K<e K0 ,
where K is the exercise price of the call on the put, K 0 is the exercise price of the put, and T is the
time to maturity of the call on the put. This assumption ensures there is some S ⇤ > 0 such that
the Black-Scholes value of the put at time T equals K when ST = S ⇤ .
The call on the put will be in the money at T if ST < S ⇤ . Let A denote the event ST < S ⇤ . The
put will be in the money at T 0 if ST 0 < K. Let C denote this event, and set D = A \ C. The value
at date 0 of the call on the put is the value at date 0 of receiving the cash flows K1A at date
T and K 0 1D ST 0 1D at date T 0 . Following the same calculations as in the previous exercise, the
value at date 0 of the call on the put is
158 15 Option Pricing
rT 0
e rT
K probR (A) + e K 0 probR (D) S0 probS (D) .
We have ✓ ◆
1
d log S = r 2
dt + dB ⇤
2
where B ⇤ is a Brownian motion under probR . The event A is the event
✓ ◆
1 2
log S0 + r T + BT⇤ < log S ⇤ .
2
This is equivalent to
B⇤
pT < d2 .
T
Thus, probR (A) = N( d2 ). The event D is the event
✓ ◆
1 2
log S0 + r T + BT⇤ < log S ⇤ ,
2
✓ ◆
1 2
log S0 + r T 0 + BT⇤ 0 < log K .
2
This is equivalent to
B⇤
pT < d2 ,
T
⇤
B 0
pT < d02 .
T0
Let M(a, b, ⇢) denote the probability that two standard normals with correlation ⇢ are less than a
p
and b respectively. Then probR (D) = M( d2 , d02 , T /T 0 ). We also have
✓ ◆
1 2
d log S = r + dt + dB ⇤
2
where B ⇤ is a Brownian motion under probS . The same calculation shows that probS (D) =
p
M( d1 , d01 , T /T 0 ). Thus, the value at date 0 of the call on the put is
rT 0
p p
e rT
K N(d2 ) + e K 0 M( d2 , d02 , T /T 0 ) S0 M( d1 , d01 , T /T 0 ) .
16
16.1. Under the assumption that S1 and S2 have volatilities i and correlation ⇢, show that
q
d(S1 /S2 ) 2 2
= something dt + 2 + 2 2⇢ 1 2 dB
S1 /S2
2
= (µ1 µ2 ⇢ 1 2 + 2 ) dt + 1 dB1 2 dB2 .
Define
q
= 2 2⇢ + 2
1 1 2 2,
B0 = 0, and
1
dB = ( 1 dB1 2 dB2 ) .
dY 2
= (µ1 µ2 ⇢ 1 2 + 2 ) dt + dB .
Y
16.2. This exercise implements Merton’s formula when the discount bond price is given by the
Vasicek model (Section 17.1). is the rate of mean reversion of the short rate process, and is the
(absolute) volatility of the short rate process.
160 16 Forwards, Futures, and More Option Pricing
Suppose the price of a non-dividend-paying stock has a constant volatility . Assume the volatility
at date t of a discount bond maturing at T > t is
⇣ ⌘
(T t)
1 e
for constants > 0 and > 0. Assume the discount bond and stock have a constant correlation ⇢.
(a) Using the result of the previous exercise, write the volatility of S/P (T ) as a function (t).
(b) Define s
Z T
1
avg = (t)2 dt .
T 0
Show that
✓ T
◆ ✓ ◆
2 2 1 2 2 1 e 2 1 e 2T
avg = + 2 2⇢ (2 2⇢ ) + .
T 2T
(c) Use l’Hôpital’s rule to show that for small T , avg ⇡ .
(d) Show that avg > for large T if ⇢ is sufficiently small.
(a) Taking S1 = S and S2 = P in the previous exercise, it follows that the volatility of S/P is
r
2 2⇢
2
(t) = 2+
2
1 e (T t) 1 e (T t) .
(b) We have
Z Z ⇣ Z T⇣ ⌘ ⌘
1 T 2 T
2⇢
(t)2 dt = 2
+ 1 2e (T t)
+e 2(T t)
1 e (T t) dt
dt
T 0 2 T T 0
2
0
2 2 T 1 2T 2⇢ 1
= S + T 1 e + 1 e T 1 e T
2 T 2 T
✓ ◆ ✓ ◆
2 1 2 2 1 e T 2 1 e 2T
= + 2 2⇢ (2 2⇢ ) +
T 2T
(c) By l’Hôpital’s rule,
1 e T 1 e 2T
lim = lim = 1.
T !0 T T !0 2T
Therefore,
Z T
1
lim (t)2 dt = 2
.
T !0 T 0
(d) We have
Z T 2
1 2⇢
lim (t)2 dt = 2
+ .
T !1 T 0 2
This is larger than 2 if and only if
16.3. Consider a forward contract on an asset that pays a single known discrete dividend at a
known date T < u, where u is the date the forward matures. Suppose there are traded discount
bonds maturing at T and u. Prove the following spot-forward parity formula for t < T :
St Pt (T )
Ft (u) = .
Pt (u)
To create a synthetic long forward, buy the asset at date t by shorting units of the discount bond
maturing at T and
St Pt (T )
Pt (u)
units of the discount bond maturing at u. The short sales produce revenue at t equal to
Pt (T ) + St Pt (T ) = St ,
which is sufficient to buy a unit of the asset. Use the dividend paid at T to cover the short position
in the date–T bond. At date u, one will own the asset and owe
St Pt (T )
.
Pt (u)
Thus, this must be the forward price.
16.4. Consider a European call option on an asset that pays a single known discrete dividend at
a known date T < u, where u is the date the option expires. Assume the asset price S drops by
when it goes ex-dividend at date T (i.e., ST = limt"T St ) and otherwise is an Itô process. Suppose
there are traded discount bonds maturing at T and u. Assume the volatility of the following process
8
>
<(St Pt (T ))/Pt (u) if t < T ,
Zt =
>
:S /P (u)
t t if T t u ,
is a constant during [0, u]. Show that the value at date 0 of the call option is
yu
(S0 P0 (T )) N(d1 ) e K N(d2 ) ,
The process Z is the forward price of the asset, for a forward contract maturing at u. A call maturing
at u on a forward maturing at u is equivalent to a call maturing at u on the underlying asset. So,
Black’s formula gives the date–0 value of the call. Substituting u for the time to maturity of the
option, P0 (u) = e yu , and
S0 P0 (T )
F0 (u) =
P0 (u)
in Black’s formula yields the result.
16.5. The purpose of this exercise is to explain why exchange options can be priced by arbitrage
without assuming the existence of a money market account or traded discount bonds. The steps
are similar to those in the market completeness proof in Section 15.5.
and ⇢ can be stochastic processes. However, assume the volatility of S1 /S2 is a constant (this
assumption can be relaxed somewhat). Assume there is an SDF process such that M S1 and M S2
are martingales. Let XT = S2T f (S1T /S2T ) for some nonnegative function f , for example, f (a) =
max(0, a 1) as in (16.8). Assume E[MT XT ] < 1. Define Z = S1 /S2 .
dB ⇤ = (µ1 µ2 ⇢ 2 2 + 2
2 ) dt + 1 dB1 2 dB2 .
On the other hand, Z is a martingale under probS2 with the same volatility as under the physical
measure. Hence,
dZ
= dB ⇤ ,
Z
for a Brownian motion B ⇤ under probS2 , where
q
= 2 + 2 2⇢
1 2 1 2.
(b) We have
2 T /2+ BT⇤
ZT = Z0 e .
This implies that the distribution under probS2 of f (ZT ) conditional on date–t information
depends only on Bt⇤ . Thus, ESt 2 [f (ZT )] is a martingale adapted to B ⇤ . The claim follows from
the martingale representation theorem.
(c) Writing ✓ ◆
W
W = S2
S2
and applying Itô’s formula gives
✓ ◆ ✓ ◆✓ ◆
dW dS2 S2 W S2 dS2 W
= + d + d .
W S2 W S2 W S2 S2
W
d = dB ⇤ = (µ1 µ2 ⇢ 1 2 + 2
2 ) dt + ( 1 dB1 2 dB2 ) .
S2
dW S2 2 S2
= µ2 dt + 2 dB2 + (µ1 µ2 ⇢ 1 2 + 2 ) dt + ( 1 dB1 2 dB2 )
W W W
S2
+ ( 2 dB2 )( 1 dB1 2 dB2 )
✓ W
◆ ✓ ◆
S2 dS1 S2 dS2
= + 1 .
W S1 W S2
164 16 Forwards, Futures, and More Option Pricing
16.6. This exercise derives Margrabe’s formula without relying on the Black-Scholes formula. The
steps are similar to those in the proof of the Black-Scholes formula in Chapter 15.
Adopt the assumptions of Exercise 16.5. Let A denote the event S1T > S2T .
Conclude that the value at date 0 of an option to exchange asset 2 for asset 1 at date T is
1
d log Y = 2
dt + dB ⇤ ,
2
where B ⇤ is a Brownian motion under the probability measure probS1 . Use this fact and the
fact that A is the event log YT < 0 to show that probS1 (A) = N(d1 ), where
1 2
log(S10 /S20 ) + 2 T
d1 = p .
T
1
d log Z = 2
dt + dB ⇤ ,
2
where B ⇤ is a Brownian motion under the probability measure probS2 . Use this fact and the
fact that A is the event log ZT > 0 to show that probS2 (A) = N(d2 ), where
p
d2 = d1 T.
(a) We have
MT SiT
E[MT SiT 1A ] = Si0 E 1A = Si0 ESi [1A ] = Si0 probSi (A) .
Si0
The payo↵ of the exchange option is S1T 1A S2T 1A , and its value at date 0 is
E[MT S1T 1A ] E[MT S2T 1A ] = S10 probS1 (A) S20 probS2 (A) .
(b) Y is a martingale under probS1 with the same volatility under probS1 as under the physical
measure. Thus,
16 Forwards, Futures, and More Option Pricing 165
dY
= dB ⇤ ,
Y
p
where B ⇤ is a Brownian motion under probS1 and = 2
1 + 2
2 2⇢ 1 2. This implies
1
d log Y = 2
dt + dB ⇤ .
2
Thus,
1
log YT = log Y0 2
T + BT⇤ .
2
The event A is the event log YT < 0, which is equivalent to
1 2
B⇤ log Y0 + T
pT < p 2
= d1 .
T T
Thus, probS1 (A) = N(d1 ).
(c) The same reasoning as in the previous part yields
1
log ZT = log Z0 2
T + BT⇤ ,
2
where B ⇤ is a Brownian motion under probS2 . The event A is the event log ZT > 0, which is
equivalent to
B⇤ log Z0 12 2T
pT < p = d2 .
T T
Thus, probS2 (A) = N(d2 ).
16.7. This exercise follows the steps in the derivation of the PDE’s for the Heston (1993) option
pricing model, using a di↵erent volatility process.
Set Vt = log t, where t is the volatility of a non-dividend-paying asset with price S. Assume
dSt
= µt dt + t dB1t ,
St
h p i
dVt = (✓ Vt ) dt + ⇢ dB1t + 1 ⇢2 dB2t ,
where µ, , ✓, and ⇢ are constants, and B1 and B2 are independent Brownian motions under the
physical probability measure. Assume there is a constant risk-free rate.
dMt µt r d"t
= r dt dB1t t dB2t + ,
Mt t "t
for some stochastic process , where " is a local martingale uncorrelated with B1 and B2 .
166 16 Forwards, Futures, and More Option Pricing
dSt ⇤
= r dt + t dB1t ,
St
⇢(µt r) h p i
dVt = ⇤ (✓⇤ Vt ) dt dt + ⇤
⇢ dB1t + 1 ⇤
⇢2 dB2t ,
t
for some constants ⇤ and ✓⇤ , where B1⇤ and B2⇤ are independent Brownian motions under the
risk neutral probability corresponding to M .
(c) Let W (t, St , Vt ) denote the conditional probability probR
t (ST > K) for a constant K. Show that
dM µ r d
= r dt dB1 + ,
M
and set
d" d
= + dB2 .
"
Then, we have
dM µ r d"
= r dt dB1 dB2 + ,
M "
and
✓ ◆ ✓ ◆
d" d
(dB1 ) = (dB1 ) + (dB2 )(dB1 ) = 0 ,
"
✓ ◆ ✓ ◆
d" d
(dB2 ) = (dB2 ) + (dB2 )2
"
= dt + dt = 0 .
µ r
dB1⇤ = dB1 + dt and dB2⇤ = dB2 + dt
define Brownian motions under the risk neutral probability. Substituting yields
16 Forwards, Futures, and More Option Pricing 167
dS
= r dt + dB1⇤ ,
S
⇢(µ r) p p
dV = (✓ V ) dt dt 1 ⇢2 dt + [⇢ dB1⇤ + 1 ⇢2 dB2⇤ ]
⇢(µ r) p
= (✓⇤ V ) dt dt + [⇢ dB1⇤ + 1 ⇢2 dB2⇤ ] ,
where p
⇤ 1 ⇢2
✓ =✓ .
(c) From Itô’s formula, the drift of dW is
⇢(µ r) 1 1
Wt + rSWS + (✓⇤ V ) WV + 2 2
S WSS + 2
WV V + ⇢SWSV .
2 2
Assume
dSt
= µt dt + t dB1t ,
St
d t = ( t ) dt + ( t ) dB2t ,
for some functions (·) and (·), where B1 and B2 are independent Brownian motions under the
physical probability measure and µ may be a stochastic process. Assume t in (16.26) equals
( t ) for some function (·). Assume there is a constant risk-free rate and the asset does not pay
dividends.
dSt ⇤
= r dt + t dB1t ,
St
⇤ ⇤
d t = ( t ) dt + ( t ) dB2t ,
for some function ⇤ (·), where B1⇤ and B2⇤ are independent Brownian motions under the risk
neutral probability.
(b) Use iterated expectations to show that the date–0 value of a call option equals
⇥ ⇤
ER S0 N(d1 ) e rT
K N(d2 ) ,
168 16 Forwards, Futures, and More Option Pricing
where
1 2
log(S0 /K) + r + 2 avg T
d1 = p ,
avg T
p
d2 = d1 avg T ,
and the risk-neutral expectation in (16.27) is taken over the random “average” volatility
s
Z
1 T 2
avg = dt
T 0 t
on which d1 and d2 depend. This average volatility is explained in a nonrandom context in
Section 15.12.
(c) Implement the Black-Scholes formula numerically. Plot the value of an at-the-money (S0 = K)
call option as a function of the volatility . Observe that the option value is approximately an
affine (linear plus constant) function of .
(d) Explain why the value of an at-the-money call option on an asset with random volatility is
approximately given by the Black-Scholes formula with
2s 3
Z T
1
ER 4 2 dt5
T 0 t
input as the volatility would overstate (understate) the value of the option.
(a) Setting
µ r
dB1⇤ = dB1 + dt and dB2⇤ = dB2 + dt
yields
dS
= r dt + dB1⇤ ,
S
d = ( ) dt + ( )[dB2⇤ ( ) dt]
⇤
= ( ) dt + ( ) dB2⇤ ,
16 Forwards, Futures, and More Option Pricing 169
where
⇤
( )= ( ) ( ) ( ).
(b) Conditioning on avg , log ST is normally distributed under probR with mean
✓ ◆
1 2
log S0 + r T
2 avg
and variance 2T . We know that, given this distribution, the mean of e rT max(0, ST K) is
S0 N(d1 ) e rT K N(d2 ). By iterated expectations,
⇥ ⇤
ER [e rT
max(0, ST K)] = ER ER [e rT
max(0, ST K) | avg ]
= ER [S0 N(d1 ) e rT
K N(d2 )] .
(d) If f is an affine function and x̃ is a random variable, then E[f (x̃] = f (E[x̃]). Because the function
rT
avg 7! S0 N(d1 ( avg )) e K N(d2 ( avg ))
(e) If f is a concave function, then Jensen’s inequality states that E[f (x̃] f (E[x̃]). Thus, if the
function
rT
avg 7! S0 N(d1 ( avg )) e K N(d2 ( avg ))
16.9. This exercise generalizes the price of risk specification in the Heston model. Under the
condition in Part (d), it is a member of the extended affine family defined by Cheridito, Filipović,
and Kimmel (2007), and M R is a martingale. Part (a) transforms the vector process (log S, V ) into
a “standard form.” See Section 16.9 for discussion, including the significance of the condition in
Part (d). The usefulness of generalizing the price of risk process is that it permits more flexible
dynamics under the physical measure (and hence more flexible expected returns) while preserving
the pricing formulas (that depend on dynamics under the risk neutral probability).
10 + 11 Y1t
1t = p ,
Y1t
20 + 21 Y1t
2t = p ,
Y1t
for constants ij . The specification in Section 16.8 is the special case 10 = 0. Derive 20 and
21 as functions of 10 and 11 from the fact that (16.19) must hold for all V .
(c) Assume that M defined in terms of 10 and 11 is such that M R is a martingale. Derive
constants a⇤i and b⇤ij in terms of 10 and 11 such that
0 1 0 1 0 10 1 0p 10 1
dY1 a⇤1 b⇤11 b⇤12 Y1 Y1 0 dB1⇤
@ A = @ A dt + @ A @ A dt + @ A@ A,
p
dY2 a⇤2 b⇤21 b⇤22 Y2 0 Y1 dB2⇤
where the Bi⇤ are independent Brownian motions under the risk-neutral probability.
(d) Assume ✓/ 2 1/2. Under what condition on 10 is a⇤1 1/2?
(a) We have
1
dY1 = 2
dV
p
V
= 2
(✓ V ) dt + dB1
✓ p
= 2
dt Y1 dt + Y1 dB1 ,
⇢
dY2 = d log S dV
✓ ◆ p
1 ⇢ p p
= µ V dt (✓ V ) dt + V [⇢ dB1 + 1 ⇢2 dB2 ] ⇢ V dB1
2
✓ ◆ 2 p p
⇢✓
= µ dt Y1 dt + ⇢ Y1 dt + 1 ⇢2 Y1 dB2 .
2
Thus, 0 1 0 1 0 1 0 1
a1 ✓/ 2 b11 b12 0
@ A=@ A, @ A=@ A,
a2 µ ⇢✓/ b21 b22 2 /2 + ⇢ 0
and = 2 (1 ⇢2 ).
16 Forwards, Futures, and More Option Pricing 171
(b) Given
1/2
Multiplying by Y1 and noting that this must hold for all values of Y1 , we obtain
20 µ r
⇢ 10 + = ,
21
⇢ 11 + = 0.
Thus,
20 =µ r ⇢ 10 ,
21 = ⇢ 11 .
dB1⇤ = dB1 + 1 dt ,
dB2⇤ = dB2 + 2 dt
are Brownian motions under the risk-neutral probability. Making these substitutions yields
0p 10 1 0p 10 1 0 1
Y1 0 dB1 Y1 0 dB1⇤ + Y
@ A@ A=@ A@ A @ 10 11 1
A dt .
p p ⇤
0 Y1 dB2 0 Y1 dB2 20 + Y
21 1
Thus, 0 1 0 1 0 1
a⇤ a ✓/ 2
@ 1A = @ 1 10
A=@ 10
A,
a⇤2 a2 20 r ⇢✓/ + ⇢ 10
and 0 1 0 1 0 1
b⇤11 b⇤12 b11 11 b12 11 0
@ A=@ A=@ A.
b⇤21 b⇤22 b21 21 b22 2 /2 + ⇢ + ⇢ 11 0
16.10. This exercise values an American call on an asset paying a known discrete dividend at a
known date. Part (c) is similar to the valuation of a compound option in Exercise 15.10.
Consider an American call option with strike K on an asset that pays a single known discrete
dividend at a known date T < u, where u is the date the option expires. Assume the asset price
S drops by when it goes ex-dividend at date T (i.e., ST = limt"T St ) and otherwise is an Itô
process. Assume there is a constant risk-free rate.
(a) Show that if < 1 e r(u T ) K, then the call should not be exercised early.
For the remainder of this exercise, assume > 1 e r(u T ) K. Assume the volatility of the
following process 8
>
< St r(T t)
e if t < T ,
Zt =
>
:S
t if T t u ,
is constant over [0, u]. Note that Z is the value of the following non-dividend-paying portfolio:
borrow e r(T t) at any date t < T to partially finance the purchase of the asset and use its
dividend to repay the debt. Let V (t, St ) denote the value of a European call on the asset with strike
K maturing at u. Let S ⇤ denote the value of the stock price just before T such that the holder of
the American option would be indi↵erent about exercising just before the stock goes ex-dividend.
This value is given by S ⇤ K = V (T, S ⇤ ). Exercise is optimal just before T if limt"T St > S ⇤ ;
equivalently, ST > S ⇤ . Let A denote the event ST > S ⇤ and let C denote the set of states
of the world such that ST S ⇤ and Su > K. The cash flows to a holder of the option who
exercises optimally are (ST + K)1A at (or, rather, “just before”) date T and (Su K)1C at
date u.
(b) Show that the value at date 0 of receiving (ST + K)1A at date T is
rT rT
S0 e N(d1 ) e (K ) N(d2 ) ,
where
1 2
log(S0 e rT ) log(S ⇤ )+ r+ 2 T
d1 = p ,
T
p
d2 = d1 T.
(c) Show that the value at date 0 of receiving (Su K)1C at date u is
16 Forwards, Futures, and More Option Pricing 173
p p
S0 e rT
M( d1 , d01 , T /u) e ru
K M( d2 , d02 , T /u) ,
where M(a, b, ⇢) denotes the probability that ⇠1 < a and ⇠2 < b when ⇠1 and ⇠2 are standard
normal random variables with correlation ⇢, and where
rT 1 2
log(S0 e ) log K + r + u
d01 = p 2
,
u
p
d02 = d01 u.
(a) Exercise is optimal just before T if and only if ST + K exceeds the value of the call at T .
The value of the call at T is bounded below by ST e r(u T ) K. Thus, a necessary condition
for exercise to be optimal is that
r(u T ) r(u T )
ST + K > ST e K , > (1 e )K .
(b) The cash flow ST 1A = ZT 1A is a share digital paying in the event A = {ZT > S ⇤ }. Its value
at date 0 is
Z0 probZ (A) = (S0 e rT
) N(d1 ) ,
where
⇣ ⌘
Z0 1 2
log S⇤ + r+ 2 T log(S0 e rT ) log(S ⇤ )+ r+ 1 2
T
p p 2
d1 = = .
T T
The cash flow ( K)1A is a share digital paying in the event A = {ZT > S ⇤ }. Its value at
date 0 is
rT
e ( K) probR (A) = e rT
( K) N(d2 ) .
(c) The cash flow Su 1C = Zu 1C is a share digital paying in the event C = {ZT S ⇤ , Zu > K}.
Its value at date 0 is
Z0 probZ (C) = (S0 e rT
) probZ (C) .
This is equivalent to
1 2
B⇤ log(S ⇤ ) log Z0 r+ 2 T
pT p = d1 ,
T T
1 2
B⇤ log Z0 log K + r + u
pu < p 2
= d01 .
u u
(d) The cash flow K1C = Zu 1C is a digital paying in the event C = {ZT S ⇤ , Zu > K}. Its
value at date 0 is e ru K probR (C). To calculate probR (C), recall that
✓ ◆
1 2
d log Z = r dt + dB ⇤ ,
2
This is equivalent to
1 2
B⇤ log(S ⇤ ) log Z0 r 2 T
pT p = d2 ,
T T
1 2
Bu⇤ log Z0 log K + r u
p < p 2
= d02 .
u u
17
(a) Show that f satisfies the fundamental PDE and the boundary condition f (T, r) = 1 if and only
if
0
+ = 1,
1
↵0 ✓ + 2 2
= 0,
2
1 2
ft + (✓ r)fr + frr = rf .
2
ft = (↵0 + 0
r)f ,
fr = f,
2
frr = f.
1
↵0 + 0
r (✓ r) + 2 2
= r.
2
1
↵0 ✓ + 2 2
= 0,
2
0
+ = 1.
Also, f (T, r) = 1 implies ↵(0) + (0)r = 0. For this to hold for all values of r, we must have
↵(0) = (0) = 0.
(b) From (17.6c), we have
1 e ⌧
(⌧ ) = ,
0 ⌧
(⌧ ) = e ,
0 ⌧ ⌧
(⌧ ) + (⌧ ) = e +1 e = 1,
(0) = 0 .
↵(0) = 0 .
ˆ and ˆ .
for constants ̂, ✓,
17 Term Structure Models 177
(c) The condition ˆ > 0 is necessary and sufficient for Yt 0 for all t in (17.36b) and hence for the
square root to exist. Assuming > 0, what are the corresponding conditions on the coefficients
in (17.35) that guarantee ↵ + rt 0 for all t?
(a) Assume
r= 0 + 1X ,
p
dX = ⌘ dt + kX dt + a + bX dB ⇤ ,
with 1 6= 0. Then
p
dr = 1 ⌘ dt + 1 kX dt + 1 a + bX dB ⇤
✓ ◆ s ✓ ◆
r 0 r 0
= 1 ⌘ dt + 1k dt + 1 a + b dB ⇤
1 1
q
= ( 1⌘ k 0 ) dt + kr dt + a 12 b 0 1 + b 1 r, dB ⇤
p
= dt r dt + ↵ + r dB ⇤ ,
where = 2
1⌘ k 0, = k, ↵ = a 1 b 0 1, and = b 1.
(b) Define ⌘ = ↵/ and Y = r ⌘. Then
p
dY = dr = dt r dt + ↵ + r dB ⇤ ,
p
= dt (⌘ + Y ) dt + ↵ + (⌘ + Y ) dB ⇤ ,
p
= ˆ dt ̂Y dt + ˆ Y dB ⇤ ,
p
where ˆ = ⌘, ̂ = , and ˆ = .
(c) We have
↵ + r = ↵ + (⌘ + Y ) = Y .
17.3. Consider a two-factor affine model with Gaussian factors — i.e., (Xt ) in (17.15) is a constant
matrix. Show that the two factors can be taken to be the short rate and its drift in the sense that
⇤
drt = Yt dt + dZ1t ,
⇤ ⇤
dYt = (a + brt + cYt ) dt + ⇣1 dZ1t + ⇣2 dZ2t
178 17 Term Structure Models
for constants , a, b, c, ⇣1 and ⇣2 and independent Brownian motions Zi⇤ under the risk neutral
probability.
Assume
r= 0 + 1 X1 + 2 X2 ,
and 0 1 0 1 0 10 1 0 10 1
X1 ✓1 k11 k12 X1 dB ⇤
11 12
d @ A = @ A dt + @ A @ A dt + @ A @ 1A .
X2 ✓2 k21 k22 X2 21 22 dB2⇤
Define 0 1 0 10 1
⇣ ⌘ ✓ ⇣ ⌘ k11 k12 X
Y = 1 2
@ 1A + 1 2
@ A @ 1A .
✓2 k21 k22 X2
Then
0 10 1
⇣ ⌘ 11 12 dB1⇤
dr = Y dt + 1 2
@ A@ A
21 22 dB2⇤
⇤ ⇤
= Y dt + 1 dB1 + 2 dB2
Furthermore, 0 1 0 1
r 0 X1
@ A = L@ A,
P2
Y i=1 i ✓i X2
where 0 1
1 2
L = @P P2
A.
2
i=1 i ki1 i=1 i ki2
If L is singular, then 0 1
2
X ⇣ ⌘ X1
Y i ✓i = 1 2
@ A = (r 0)
i=1 X2
for some , which implies
17 Term Structure Models 179
0 10 1
⇣ ⌘ 11 12 dB1⇤
dY = dr = Y dt + 1 2
@ A@ A.
dB2⇤
21 22
Hence,
0 10 1 0 10 1 0 1
⇣ ⌘ k11 k12 ✓ ⇣ ⌘ k11 k12 k k r
dY = @ A @ 1 A dt + @ A @ 11 12 A L 1 @ 0
A dt
1 2 1 2 P2
k21 k22 ✓2 k21 k22 k21 k22 Y i=1 i ✓i
0 10 10 1
⇣ ⌘ k11 k12 11 12 dB1⇤
+ 1 2 @ A@ A@ A.
k21 k22 dB2⇤
21 22
In either case,
dY = (a + br + cY ) dt + ⌘1 dB1⇤ + ⌘2 dB2⇤
Then Z1⇤ is a Brownian motion under the risk neutral probability, and
dr = Y dt + dZ1⇤ .
Then, under the risk neutral probability, Z2⇤ is a Brownian motion that is independent of Z1⇤ .
Defining
0 1 0 1 10 1
⇣ / ⇠1 ⌘1
@ 1A = @ 1 A @ A,
⇣2 2/ ⇠2 ⌘2
we have
180 17 Term Structure Models
0 1
⇣
dB1⇤ ⌘
dY = (a + br + cY ) dt + ⌘1 ⌘2 @ A
dB2 ⇤
0 10 1
⇣ ⌘ 1/ 2/ dB1⇤
= (a + br + cY ) dt + ⇣1 ⇣2 @ A@ A
⇠1 ⇠2 dB2⇤
0 1
⇣ ⌘ dZ ⇤
1A
= (a + br + cY ) dt + ⇣1 ⇣2 @ .
dZ2⇤
17.4. This exercise shows that the factors in a two-factor CIR model can be taken to be the short
rate and its volatility. This idea is developed by Longsta↵ and Schwartz (1992).
Suppose r is the sum of two independent square-root processes X1 and X2 as in Section 17.3. Define
Yt = 2 2
1 X1t + 2 X2t . Note that the instantaneous variance of r = X1 + X2 is
⇣ p p ⌘2
⇤ ⇤
1 X1t dB1t + 2 X2t dB2t = Yt dt .
for a constant vector , and constant matrices K and ⌃, where S(r, Y ) is a diagonal matrix the
squared elements of which are affine functions of (r, Y ).
It is without loss of generality to take 1 > 0 and 2 > 0 (because Bi⇤ can be used as the Brownian
motion if necessary). It is also with loss of generality to take 1 > 2 (because X1 and X2 can be
reordered if necessary). From r = X1 + X2 and Y = 2 2
1 X1 + 2 X2 , we obtain
Y 2 2
2r 1r Y
X1 = 2 2 , and X2 = 2 2 .
1 2 1 2
We have
0 1 0 10 1 0 10 1 0 p 10 1
X1 1 0 ✓1 1 0 X1 1 X1 0 dB1⇤
d@ A = @ A @ A dt @ A @ A dt + @ p A @ ⇤A
X2 0 2 ✓2 0 2 X2 0 2 X2 dB2
0 10 1 0 10 1
1 0 ✓ 2 1 r
A @ 1 A dt + 1 @ 2
=@ 2 2
A @ A dt
0 2 ✓2 2
1 2 1 1 Y
0 p 10 1
Y 2r 0 dB ⇤
1 @ 1 2 A @ 1A .
+p 2 2 p
1 2 0 2
2
1r Y dB2⇤
17 Term Structure Models 181
Thus,
0 1 0 10 1
r 1 1 dX1
d@ A = @ A@ A
Y 2 2 dX
1 2 2
0 10 10 1 0 10 10 1
1 1 1 0 ✓ 1 1 2 1r
A @ 1 A dt + 1 2
=@ A@
2 2
@ A@ A @ A dt
2 2 0 2 ✓2 2 2 2
1 2 1 2 1 2 1 1 Y
0 10 p 10 1
1 1 1 Y 2
2r 0 dB1⇤
+p 2 @ A@ 1 p A @ A.
2 2 2 2r ⇤
1 2 1 2 0 2 1 Y dB 2
17.5. This exercise develops the completely affine version of the Vasicek model.
Assume the short rate is an Ornstein-Uhlenbeck process under the physical measure; i.e.,
dr = ̂(✓ˆ r) dt + dB ,
for constants ̂, ✓ˆ and , where B is a Brownian motion under the physical measure. Assume there
is an SDF process M with
dM d"
= r dt dB + ,
M "
where is a constant and " is a local martingale uncorrelated with B.
(a) Show that the short rate is an Ornstein-Uhlenbeck process under the risk neutral probability
corresponding to M (i.e., the Vasicek model holds).
(b) Show that the risk premium of a discount bond depends only on its time to maturity and is
independent of the short rate.
(a) Girsanov’s theorem states that B ⇤ is a Brownian motion under the risk neutral probability,
where dB ⇤ = dB + dt. We have
= (✓ r) dt + dB ⇤ ,
↵(⌧ ) (⌧ )rt
e ,
182 17 Term Structure Models
where ⌧ = T t, ↵(⌧ ) = ⌧ a(⌧ ), (⌧ ) = ⌧ b(⌧ ), and a(·) and b(·) are the functions given in (17.6).
Setting Xt = ↵(T t) (T t)rt , Itô’s formula implies that the return of a discount bond is
1 1
dX + (dX)2 = ↵0 (⌧ ) dt + 0
(⌧ )r dt (⌧ ) dr + 2
(⌧ )(dr)2
2 2
1
= ↵0 (⌧ ) dt + 0
(⌧ )r dt (⌧ )(✓ˆ r) dt (⌧ ) dB + 2
(⌧ ) 2
dt .
2
An alternative calculation of the risk premium uses the the fundamental PDE (Exercise 17.1).
The fundamental PDE is
1
↵0 (⌧ ) ✓ (⌧ ) + 2 2
(⌧ ) = 0 ,
2
0
(⌧ ) + (⌧ ) = 1 .
1
↵0 ✓ˆ (⌧ ) + (⌧ ) + 2 2
= 0.
2
1
↵0 (⌧ ) dt + 0
(⌧ )r dt (⌧ )(✓ˆ r) dt (⌧ ) dB + 2
(⌧ ) 2
dt = r dt (⌧ ) dt (⌧ ) dB .
2
17.6. This exercise develops the completely affine version of the multi-factor CIR model.
Assume r = X1 + X2 where the Xi are independent square-root processes under the physical
measure; i.e.,
p
dXi = ̂i (✓ˆi Xi ) dt + i Xi dBi ,
for constants ̂i , ✓ˆi and i, where the Bi are independent Brownian motions under the physical
measure. Assume there is an SDF process M with
dM p p d"
= r dt 1 X1 dB1 2 X2 dB2 + ,
M "
where 1 and 2 are constants and " is a local martingale uncorrelated with B.
17 Term Structure Models 183
(a) Show that the Xi are independent square-root processes under the risk neutral probability
corresponding to M .
(b) Show that the risk premium of a discount bond is a linear function of the factors X1 and X2 ,
with coefficients depending on the time to maturity.
p
(a) By Girsanov’s theorem, setting dBi⇤ = dBi + i X i dt produces independent Brownian motions
Bi⇤ under the risk neutral probability. We have
p p
dXi = ̂i (✓ˆi Xi ) dt + Xi (dBi⇤
i i X i dt)
p
= i (✓i Xi ) dt + i Xi dBi⇤ ,
Pt (T ) = eYt ,
where
2
X
Yt = [↵i (⌧ ) + i (⌧ )Xit ]
i=1
with ⌧ = T t and ↵i (·) and i (·) being defined in (17.14). By Itô’s formula,
dP 1
= dY + (dY )2
P 2
X2
1
= ↵i0 (⌧ ) dt + 0
i (⌧ )Xi dt i (⌧ ) dXi + 2
i (⌧ )(dXi )
2
2
i=1
X2 p
ˆ 1
= ↵i0 (⌧ ) dt + 0
i (⌧ )Xi dt i (⌧ )̂i (✓i Xi ) dt i (⌧ ) i X i dBi + 2 2
i (⌧ ) i Xi dt .
2
i=1
dM ⇥ 1
⇤0 d"
= r dt S(Xt ) + S(Xt ) ⇤Xt dB + ,
M "
where B is a vector of independent Brownian motions under the physical measure, " is a local
martingale uncorrelated with B, S(X) is a diagonal matrix the squared elements of which are
184 17 Term Structure Models
affine functions of X, S(X) 1 denotes the inverse of S(X), is a constant vector, and ⇤ is a
constant matrix. Assume M R is a martingale, so there is a risk neutral probability corresponding
to M . [Warning: This assumption is not valid in general. See the end-of-chapter notes.]
Brownian motions under the risk neutral probability, 0 is a constant, and are constant
vectors, and K and are constant matrices. Show that
dX = ( ˆ + K̂X) dt + S(X) dB
⇥ ⇤
dB ⇤ = dB + S(X) + S(X) 1
⇤X dt .
Thus,
dX = ( + KX) dt + S(X) dB ⇤
⇥ 1
⇤
= ( + KX) dt + S(X) dB + S(X) + S(X) ⇤X dt
The vector S(X)S(X) is an affine function of X, say A + CX for a vector A and matrix C.
Set ˆ = + A and K̂ = K + C + ⇤.
(b) Let k denote the dimension of X. The price at t of a discount bond maturing at T is Pt = eYt ,
where
k
X ⌧
X
Yt = ↵i (⌧ ) i (⌧ )Xi ,
i=1 i=1
17 Term Structure Models 185
for some functions ↵i (·) and i (·) and where ⌧ = T t. By Itô’s formula,
dP 1
= dY + (dY )2
P 2
Xk k k
1 XX
= [↵i0 (⌧ ) dt + 0
i (⌧ )Xi dt
i (⌧ ) dXi ] + i (⌧ ) j (⌧ )(dXi )(dXj ) .
2
i=1 i=1 j=1
P
The stochastic part of dP/P comes from i dXi and hence is
(⌧ )> S(X) dB ,
17.8. This example is from Du↵ee (2002). The model in (b) is an essentially affine model.
Assume 0 1 0 1 0 10 10 1
dr (✓ r) 0 1 0 dB1⇤
@ A=@ A dt + @ A@ p A@ A
dY ( Y) 0 ⌘ 0 Y dB2⇤
for constants , ✓, , , and ⌘, where the Bi⇤ are independent Brownian motions under a risk
neutral probability.
(a) Given the completely affine price of risk specification (17.21), where X = (r Y )0 and
0 1
1 0
S(X) = @ p A ,
0 Yt
show that
dr = (✓ˆ r) dt + dB1 ,
ˆ where B1 is a Brownian motion under the physical measure. Show that the
for some constant ✓,
risk premium of a discount bond depends only on its time to maturity and does not depend on
r or Y .
186 17 Term Structure Models
Show that the risk premium of a discount bond can depend on r and Y .
dB ⇤ = dB + S(X) dt
0 10 1
1 0 1
= dB + @ p A @ A dt
0 Y 2
0 1
1
= dB + @ p A dt .
2 Y
Thus,
dr = (✓ r) dt + dB1⇤
= (✓ˆ r) dt + dB1 ,
where ✓ˆ = ✓+ 1 /. Note that r satisfies the Vasicek model (under the risk neutral probability).
Thus, the price at t of a discount bond maturing at T is Pt = eZt , where
Zt = ↵(⌧ ) (⌧ )rt ,
dP 1
= dZ + (dZ)2
P 2
1
= ↵0 (⌧ ) dt + 0
(⌧ )r dt (⌧ )[(✓ˆ r) dt + dB1 ] + 2 2
(⌧ ) dt
2 0 1
⇣ ⌘
1 dB1
= ↵0 (⌧ ) + 0
(⌧ )r (⌧ )(✓ˆ r) + 2 2
(⌧ ) dt + (⌧ ) 0 @ A.
2 dB2
Hence,
17 Term Structure Models 187
0 1
✓ ◆✓ ◆
dM dP 0
(⌧ )
= S(X)0 (dB)(dB)0 @ A
M P 0
0 10 1
⇣ ⌘ 1 0 (⌧ )
= 1 2
@ p A@ A dt
0 Yt 0
= 1 (⌧ ) dt .
Thus,
dr = (✓ r) dt + dB1⇤
Note that r satisfies the Vasicek model (under the risk neutral probability). Thus, the price at t of
a discount bond maturing at T is Pt = eZt , where
Zt = ↵(⌧ ) (⌧ )rt ,
Hence,
2 0 1 30 0 1
✓ ◆✓ ◆
dM dP 1 0 (⌧ )
= 4S(X) + @ A ⇤X 5 (dB)(dB)0 @ A
M P 0 0 0
2 0 1 0 10 13 0 1
⇣ ⌘ 1 0 ⇣ ⌘ ⇤11 ⇤21 1 0 (⌧ )
=4 1 2 @ p A+ r Y @ A@ A5 @ A dt
0 Yt ⇤12 ⇤22 0 0 0
0 1
⇣ p ⌘ (⌧ )
= 1 + ⇤11 r + ⇤12 Y 2 Y
@ A dt
0
= (⌧ )[ 1 + ⇤11 r + ⇤12 Y ] dt .
17.9. This exercise verifies that the two-factor version of the model of Constantinides (1992) is a
quadratic model.
dX1 = µ dt + dB1t ,
with µ, , and being constants and with B1 and B2 being independent Brownian motions under
the physical measure.
(a) Derive dM/M , and deduce that r is a quadratic function of X1 and X2 and the market prices
of risk are affine functions of X1 and X2 .
(b) Given the prices of risk calculated in the previous part, find Brownian motions B1⇤ and B2⇤ under
the risk neutral probability and show that the dX satisfy (17.24b).
and
17 Term Structure Models 189
dM 1
= dY + (dY )2
M 2
2 1 2 2
= [µ + 2(X2 a)X2 ] dt + dB1 + 2 (X2 a) dB2 + dt + 2 (X2 a)2 dt
2
2 1 2 2
= µ+ + 2(X2 a)X2 + 2 (X2 a)2 dt + dB1 + 2 (X2 a) dB2 .
2
2 1 2 2
r= µ + 2(X2 a)X2 2 (X2 a)2 .
2
Also,
dM
= r dt 1 dB1 2 dB2 ,
M
where
1 = ,
2 = 2 (X2 a) .
(b) By Girsanov’s theorem, B ⇤ is a vector of independent Brownian motions under the risk neutral
probability associated to M , where
dB1⇤ = dB1 dt ,
We have
= (µ + ) dt + dB1⇤ ,
= 2a 2
dt + (2 2
)X2 dt + dB2⇤ .
17.10. This exercise develops the Vasicek model with time-dependent parameters studied by Hull
and White (1990).
= (t)[✓(t) g(t)] .
Thus,
does not involve the function ✓(·). Thus, one can match the market price Qt (T ) by setting
✓ Z T ◆
Q (T )
exp g(u) du = h ⇣t R ⌘i .
T
t ER
t exp t r̂ u du
Taking another derivative in T and using the previous formula for g 0 yields
✓ Z T ◆
d2 d2 R
(T )[✓(T ) g(T )] = log Qt (T ) + log Et exp r̂u du .
dT 2 dT 2 t
Thus, to fit the yield curve at t, one should choose ✓(T ) for T > t by
✓ Z T ◆
1 d2 1 d2 R
✓(T ) = g(T ) log Qt (T ) + log Et exp r̂u du .
(T ) dT 2 (T ) dT 2 t
17.11. This exercise asks for the Hull-White model to be written in the Heath-Jarrow-Morton
form. Assume the short rate is rt = r̂t + g(t), where
dr̂ = r̂ dt + dB ⇤ ,
for constants and and g(·) is chosen to fit the current yield curve.
(a) Calculate the forward rates fs (u) using the Vasicek bond pricing formula.
(b) Calculate ↵s (u) and s (u) such that, as s changes,
⇤
dfs (u) = ↵s (u) ds + s (u) dBs .
where
Ys = ↵(⌧ ) (⌧ )r̂s
with ⌧ = u s, ↵(⌧ ) = ⌧ a(⌧ ), (⌧ ) = ⌧ b(⌧ ) and a(·) and b(·) being defined in (17.6), taking
✓ = 0. Thus,
2 2 2
⌧ 2⌧
↵(⌧ ) = 2
⌧+ 1 e 1 e ,
2 3 43
1 ⌧
(⌧ ) = 1 e .
d log Ps (u)
= g(u) + ↵0 (⌧ ) + 0
(⌧ )r̂s
du
2 2 2
⌧ 2⌧ ⌧
= g(u) + e e +e r̂s .
22 2 22
(c) We have
Z u Z u
(u s) (t s)
s (u) s (t) dt = e e dt
s s
2 ⇣ ⌘
(u s) (u s)
= e 1 e
= ↵s (u) .
17.12. This exercise derives an option pricing formula for discount bonds in the Vasicek/Hull-
White model.
17 Term Structure Models 193
dr̂ = r̂ dt + dB ⇤ ,
(a) Consider a forward contract maturing at T on a discount bond maturing at u > T . Let Ft
denote the forward price for t T . What is the volatility of dFt /Ft ?
(b) What is the average volatility between 0 and T of dFt /Ft in the sense of (16.15)?
(c) Consider a call option maturing at T on a discount bond maturing at u > T . Derive a formula
for the value of the call option at date 0.
(a) The forward price is Ft = Pt (u)/Pt (T ). The volatility of dF/F is the volatility of d log F , so it
is the volatility of d log Pt (u) d log Pt (T ). From the previous exercise, the price Pt (x) at t of
a discount bond maturing at x is
✓ Z x ◆
exp g(s) ds eYt (x) ,
t
where
Yt (x) = ↵(x t) (x t)r̂t
with
2 2 2
⌧ 2⌧
↵(⌧ ) = 2
⌧+ 1 e 1 e ,
2 3 43
1 ⌧
(⌧ ) = 1 e .
The volatility of d log Pt (u) d log Pt (T ) is the volatility of dYt (u) dYt (T ), which is the volatility
of
[ (u t) (T t)]dr̂t .
(c) Because the risk-free rate is stochastic, we need to use Merton’s (equivalently, Black’s) formula.
The value at date 0 is
yT
P0 (u) N(d1 ) e K N(d2 ) ,
where
1 2
log(P0 (u)/K) + y + 2 avg T
d1 = p ,
avg T
p
d2 = d1 avg T ,
Topics
18
Heterogeneous Priors
18.1. Suppose each investor h has CARA utility with absolute risk aversion ↵h . Assume the infor-
mation in the economy is generated by w̃m . Assume investor h believes w̃m is normally distributed
with mean µh and variance 2, where is the same for all investors.
(a) Show that the Radon-Nikodym derivative of investor h’s probability Ph with respect to the
average probability P is ⇣ ⌘
(w̃m µh )2
exp 2 2
z̃h = PH ⇣ ⌘.
1 (w̃m µj )2
H j=1 exp 2 2
(c) Show that if investors also disagree about the variance of w̃m , then the sharing rule (18.4) is
quadratic in w̃m .
(a) We have z̃h = gh (w̃m ) for some function gh . Consider a random variable x̃ = f (w̃m ) for some
function f . By the normal distribution assumption and the definition of a Radon-Nikodym
derivative,
Z 1 ✓ ◆
1 (w µh )2
p f (w) exp dw = Eh [x̃]
2⇡ 1 2 2
= E[x̃z̃h ]
H Z ✓ ◆
1 X 1 1
(w µj )2
= p f (w)gh (w) exp dw .
H 2⇡ 1 2 2
j=1
✓ ◆ H ✓ ◆
(w µh )2 1 X (w µj )2
exp = gh (w) exp ,
2 2 H 2 2
j=1
implying ⇣ ⌘
(w µh )2
exp 2 2
gh (w) = PH ⇣ ⌘.
1 (w µj )2
H j=1 exp 2 2
Hence, ✓ ◆
z̃h j (w̃m µj )2 (w̃m µh )2
log = log + .
z̃j h 2 j2 2 h2
Now, one can follow the reasoning in Part (b), producing a sharing rule that includes terms
2 /
involving the w̃m 2
j.
18.2. Assume all investors have constant relative risk aversion ⇢ and the same discount factor .
Solve the social planning problem in a finite-horizon discrete-time model to show that the social
planner’s utility is " #
T
X
t C1 ⇢
E Zt t
1 ⇢
t=0
18 Heterogeneous Priors 199
for some stochastic process Z. Show that Z is a supermartingale relative to the average beliefs if
⇢ > 1. Hint: For the last statement, use a conditional version of the Minkowski inequality. The
Minkowski inequality states that for random variables x̃h and any ⇢ > 1,
" H
!⇢ #1/⇢ H
X X ⇥ ⇤1/⇢
E x̃h Et x̃⇢h .
h=1 h=1
for each t and in each state of the world. Following the calculations in Exercise 3.9 yields
( h Zht )1/⇢
Cht = PH Ct .
1/⇢
j=1 ( j Zjt )
Thus,
( 1/⇢
1 ⇢ h Zht ) 1 ⇢
ht Zht Cht = ⇣P ⌘1 ⇢ Ct ,
H 1/⇢
j=1 ( j Zjt )
where !⇢
H
X
1/⇢
Zt = ( h Zht ) .
h=1
" H
!⇢ #1/⇢
X
Et [Zt+1 ]1/⇢ = Et x̃h
h=1
H
X ⇥ ⇤1/⇢
Et x̃⇢h
h=1
XH
1/⇢
= Et [ h Zh,t+1 ]
h=1
H
X
1/⇢
= ( h Zht )
h=1
1/⇢
= Zt .
18.3. Consider an infinite-horizon version of the model in Section 18.5 in which both investors
agree the dividend process is a two-state Markov chain, with states D = 0 and D = 1. Suppose the
investors’ beliefs Ph satisfy, for all t 0,
to calculate Vh (0) and Vh (1). Show that investor 2 has the highest fundamental value in both states
[V2 (0) > V1 (0) and V2 (1) > V1 (1)] but investor 1 is the most optimistic in state D = 0 about
investor 2’s future valuation, in the sense that
4 1 1
x= x+ (1 + y) ,
3 2 2
4 2 1
y= x+ (1 + y) .
3 3 3
The solution is V1 (0) = x = 4/3 = 1.33 and V1 (1) = y = 11/9 = 1.22. Now for investor 2, set
x = V2 (0) and y = V2 (1). We want to solve
4 2 1
x= x+ (1 + y) ,
3 3 3
4 1 3
y= x+ (1 + y) .
3 4 4
The solution is V2 (0) = x = 16/11 = 1.45 and V2 (1) = y = 21/11 = 1.91 Hence, investor 2 has the
highest fundamental value in both states. The last claim is equivalent to
1 16 1 21 2 16 1 21
· + · 1+ > · + · 1+ .
2 11 2 11 3 11 3 11
Asymmetric Information
19.1. In the economy of Section 19.4, assume the uninformed investors are risk neutral. Find a
fully revealing equilibrium, partially revealing equilibria in which the price reveals s̃ + bz̃ for any
b, and a completely unrevealing equilibrium (an equilibrium in which the price is constant rather
than depending on s̃ and/or z̃).
For any b, we can define a0 and a1 by the first two equations and the third equation will hold. Thus,
there is an equilibrium revealing s̃ + bz̃ for any b. This includes b = 0, which is the fully revealing
equilibrium.
19.2. Consider the model of Section 19.5, but assume there is a continuum of investors indexed by
h 2 [0, 1] with possibly di↵ering risk aversion coefficients ↵h and possibly di↵ering error variances
var(˜
"h ). Suppose, for some b, that each investor observes x̃ + bỹ in addition to his private signal s̃h .
The market-clearing condition is
Z 1
✓h (x̃ + bỹ, s̃h ) dh = ỹ ,
0
where ✓h is the number of shares demanded by investor h. Let 2 denote the variance of x̃ conditional
h
(a) Show that the equilibrium price is a discounted weighted average of the conditional expectations
of x̃ minus a risk premium term, where the weight on investor h is ⌧h h /(⌧ ).
(b) Define
var(x̃) (1 ) var(x̃)
= and h = .
var(x̃) + b2 var(ỹ) (1 ) var(x̃) + var(˜
"h )
Show that
1
⌧h h h = .
↵h var(˜
"h )
(c) Assume the strong law of large numbers holds in the sense that
Z 1
⌧h h h "˜h dh = 0 .
0
Define
Z 1
1
= ⌧h h h dh .
⌧ 0
Show that the equilibrium price equals a0 + a1 (x̃ + bỹ) if and only if
(1 )[(1 )x̄ bȳ]
a0 = ,
Rf
(1 ) +
a1 = ,
Rf
Z 1
1
b= 1 dh .
0 ↵h var(˜
"h )
19 Asymmetric Information 205
This is equivalent to
Z 1 ✓ ◆
⌧h h E[x̃ | x̃ + bỹ, s̃h ] ỹ
p(x̃, ỹ) = dh .
0 ⌧ Rf ⌧ Rf
2
h = (1 h )(1 ) var(x̃) .
This implies
1 1 h var(˜"h )
= (1 ) var(x̃) = (1 ) var(x̃) = var(˜
"h ) .
h h h (1 ) var(x̃)
Thus,
h h 1
⌧h h h = = .
↵h ↵h var(˜
"h )
(c) From (19.4a),
Thus,
Z 1 ✓ ◆
⌧h h E[x̃ | x̃ + bỹ, s̃h ] ỹ
p(x̃, ỹ) = dh
0 ⌧ Rf ⌧ Rf
Z 1 Z 1
1 1
= ⌧h h E[x̃ | x̃ + bỹ] dh ⌧h h h E[x̃ | x̃ + bỹ] dh
⌧ Rf 0 ⌧ Rf 0
Z 1
1 ỹ
+ ⌧h h h s̃h dh
⌧ Rf 0 ⌧ Rf
E[x̃ | x̃ + bỹ] E[x̃ | x̃ + bỹ] x̃ ỹ
= + .
Rf Rf Rf ⌧ Rf
206 19 Asymmetric Information
From (19.2a),
E[x̃ | x̃ + bỹ] = x̄ + (x̃ x̄ + bỹ bȳ) .
Therefore,
(1 )[x̄ + (x̃ x̄ + bỹ bȳ)] x̃ ỹ
p(x̃, ỹ) = + .
Rf Rf ⌧ Rf
This is equal to a0 + a1 (x̃ + bỹ) if and only if
(1 )x̄)[(1
bȳ]
a0 = ,
Rf
(1 ) +
a1 = ,
Rf
(1 ) b 1
a1 b = .
Rf ⌧ Rf
19.3. In the single-period Kyle model, assume the informed investor has CARA utility. There is
a linear equilibrium. Derive an expression for as a root of a fifth-order polynomial, assuming
the informed investor observes s̃ = x̃ + "˜, where "˜ is normally distributed with zero mean and is
independent of x̃.
A strategy for the informed trader is affine in the signal s̃ if and only if it is affine in ṽ, so we are
looking for an equilibrium in which the informed trader plays ✓(v) = ↵ + v for some ↵ and .
With ỹ = ↵ + ṽ + z̃, the normal-normal updating rule implies E[x̃ | ỹ] = E[ṽ | ỹ] = + ỹ, where
✓ 2 var(ṽ)
◆
= 1 2 var(ṽ) + var(z̃)
v̄ , (19.1a)
var(ṽ)
= 2 . (19.1b)
var(ṽ) + var(z̃)
↵ ↵ 2 2 2
E[✓(x̃ (✓ + z̃)) | s̃] var(✓(x̃ (✓ + z̃)) | s̃) = ✓(ṽ ✓) ✓ [ + var(z̃)] .
2 2
The optimum is attained at
ṽ
✓= 2 2 var(z̃)
.
2 +↵ +↵
Thus,
↵= 2 2 var(z̃)
, (19.1c)
2 +↵
+↵
1
= . (19.1d)
2 +↵ 2+↵ 2 var(z̃)
19.4. In the continuous-time Kyle model, assume log ṽ is normally distributed instead of ṽ being
normally distributed. Denote the mean of log ṽ by µ and the variance of log ṽ by 2. Set = v/ z.
dPt = Pt dYt
(log ṽ µ)/ Yt
d✓t = dt
1 t
form an equilibrium by showing the following:
(a) Define Wt = Yt / z. Show that, conditional on ṽ, W is a Brownian bridge on [0, 1] with terminal
value (log ṽ µ)/ v. Use this fact to show that P satisfies P1 = ṽ and is a martingale relative
to the market makers’ information.
(b) For v > 0 and p > 0, define
p v + v(log v log p) 1
J(t, p) = + v z (1 t)v .
2
In the definition of the class of allowed strategies in Section 19.7, modify condition (ii) to (ii0 )
hR i
1
E 0 Pt2 dt < 1. Prove the verification theorem.
208 19 Asymmetric Information
(a) We have
1 (log ṽ µ)/ v W
dW = (d✓ + z dB) = dt + dB .
z 1 t
Hence, W is a Brownian bridge on [0, 1] with terminal value (log ṽ µ)/ v. Moreover,
dP
= dY = v dW ,
P
so P is a geometric Brownian motion relative to the market makers’ information. We also have
1 2 1 2 1 2
v W1
P1 = P0 e 2 v = eµ+ 2 v elog ṽ µ 2 v = elog ṽ = ṽ ,
so Pt is the conditional expectation of ṽ given the market makers’ information, for each t.
Furthermore, P being a geometric Brownian motion implies that ✓ satisfies condition (ii0 ).
(b) We have
v zv
Jt = ,
2
1 v
Jp = ,
p
v
Jpp = ,
p2
(dP )2 = 2 2
v P dt .
Hence,
Z 1
1
J(1, P1 ) = J(0, P0 ) + Jt dt + Jp dP + Jpp (dP )2
0 2
Z 1✓ ◆ 2
v zv 1 v 1v v
= J(0, P0 ) + dP +
2 0 P 2
Z 1
P v dP
= J(0, P0 ) +
0 P
Z 1
= J(0, P0 ) + (P v)(d✓ + z dB) .
0
Rearranging gives
Z 1 Z 1
(v P ) d✓ = J(0, P0 ) J(1, P1 ) + (P v) z dB .
0 0
the inequality following from J(1, p) 0, which can be deduced from the fact that J(1, p) is
convex in p and hence has a unique minimum at p = v. Thus, any strategy that implies P1 = ṽ
is optimal.
20
(a) Show that an expected utility maximizer who prefers A to B must also prefer C to D.
(b) Show that the preferences A B and D C violate the independence axiom by showing that
C = ↵A + (1 ↵)Q and D = ↵B + (1 ↵)Q for some 0 < ↵ < 1 and some gamble Q.
(c) Plot the gambles A, B, C, and D in the probability simplex of Figure ??, taking p1 to be the
probability of $0 and p3 to be the probability of $4,000. Show that the line connecting A with
B and the line connecting C with D are parallel.
Note: the preferences A B and D C are common. This example is also due to Allais (1953)
and is a special case of the common ratio e↵ect. See, e.g., Starmer (2000).
(b) Let Q be the gamble that pays 0 for sure, and set ↵ = 0.25.
(c) In the probability simplex, A = (0, 0), B = (0.2, 0.8), C = (0.75, 0) and D = (0.8, 0.2). The
slope of the line connecting A and B is 0.8/0.2 = 4, and the slope of the line connecting C and
D is 0.2/0.05 = 4.
210 20 Alternative Preferences in Single-Period Models
Show that an expected utility maximizer who prefers A to B must also prefer C to D. Note: the
preferences A B and D C are common. This example is due to Kahneman and Tversky (1979).
We have
and
20.3. Consider weighted utility. Let "˜ have zero mean and unit variance. For a constant , denote
the certainty equivalent of w + "˜ by w ⇡( ). Assume ⇡(·) is twice continuously di↵erentiable.
By di↵erentiating
v(w ⇡( ))E[ (w + "˜)] = E[ (w + "˜)v(w + "˜)] ,
assuming di↵erentiation and expectation can be interchanged, show successively that ⇡ 0 (0) = 0 and
v 00 (w) 2 0 (w)
⇡ 00 (0) = .
v 0 (w) (w)
Note: This implies that for CRRA weighted utility and small , ⇡( )/w ⇡ (⇢ 2 ) var( "˜/w)/2.
We have
= E[ (w + "˜)v 0 (w + "˜)˜
"] + E[ 0 (w + "˜)v(w + "˜)˜
"] .
20 Alternative Preferences in Single-Period Models 211
which implies
v 00 (w) 2 0 (w)
⇡ 00 (0) = .
v 0 (w) (w)
(a) Show that g in (20.11) is strictly monotone in y > 0 — so the preferences are monotone with
regard to stochastic dominance — if and only if 0 and ⇢ + 1 with at least one of these
being a strict inequality.
(b) Show that g in (20.11) is strictly monotone and concave if and only if 0 and ⇢ +1
with either < 0 or ⇢ < + 1.
(c) Consider a lognormal gamble: w̃ = w(1+˜
") where log(1+˜
") is normally distributed with variance
2 and mean 2 /2 (implying E[˜
"] = 0). Show that the certainty equivalent is w(1 ⇡) where
(⇢ 2 ) 2 /2
⇡=1 e .
g 0 (y) = ( + 1 ⇢)y ⇢
y 1
⇥ ⇤
=y 1 ( +1 ⇢)y 1 ⇢
.
This is positive for all y > 0 if and only if 0 and +1 ⇢ 0 with at least one of these
being a strict equality.
(b) We have
⇥ ⇤
=y 2 ( +1 ⇢)( ⇢)y 1 ⇢
( 1) .
(c) Let z̃ = log(1 + "˜), and let P denote the distribution of wez̃ . We are assuming
h ⇣ ⌘ i
z̃ 1 1 ⇢
1 E we 1 ⇢ wez̃
[w(1 ⇡)]1 ⇢ = v(w(1 ⇡)) = v(U (P )) = .
1 ⇢ E [(wez̃ ) ]
Equivalently,
⇥ ⇤
1 ⇢ E e[ +(1 ⇢)]z̃
(1 ⇡) =
E [e z̃ ]
[ +(1 ⇢)] 2 /2+[ +(1 ⇢)]2 2 /2
e
= 2 /2+ 2 2 /2
e
(⇢ 2 )(1 ⇢) 2 /2
=e .
Hence,
(⇢ 2 ) 2 /2
1 ⇡=e .
20.5. Consider CRRA disappointment-averse utility and a random wealth w̃ = ez̃ , where z̃ is
normally distributed with mean µ and variance 2. Let ⇠ denote the certainty equivalent of w̃, and
set ✓ = log ⇠.
where N denotes the standard normal distribution function. Hint: See the calculation of
⇥ ⇤
E ebz̃ 1{z̃<a} for a normal random variable z̃ in Section 7.4.
(b) Let µ = 2 /2 + log w for a constant w. [ Defining the standard normal x̃ = (z̃ µ)/ , we then
2 /2+ x̃
have w̃ = w(1 + "˜), where "˜ ⌘ e 1 has mean zero.] Define ⇡ = (w ⇠)/w. [ Then
w(1 ⇡) is the certainty equivalent of w(1+ "˜).] Show numerically that ⇡/ 2 appears to increase
without bound and ⇡/ appears to converge to a positive constant as # 0.
This is equivalent to ⇥ ⇤ ⇥ ⇤
E w̃1 ⇢ + E 1{w̃<⇠} w̃1 ⇢
⇠1 ⇢
= . (⇤)
1 + P (w̃ < ⇠)
20 Alternative Preferences in Single-Period Models 213
To evaluate the numerator, define the standard normal x̃ = (z̃ µ)/ . We have
w̃1 ⇢
= e(1 ⇢)µ+(1 ⇢) x̃
,
and
✓ µ
w̃ < ⇠ , x̃ < .
Therefore,
⇥ ⇤ h i
E 1{w̃<⇠} w̃1 ⇢
= e(1 ⇢)µ
E 1{x̃<(✓ µ)/ } e
(1 ⇢) x̃
Z (✓ µ)/
1 1 2
= e(1 ⇢)µ
p e(1 ⇢) x 2 x dx
1 2⇡
Z (✓ µ)/
⇢)µ+ 12 (1 ⇢)2 2 1 1
(1 ⇢) ]2
= e(1 p e 2 [x dx
1 2⇡
✓ ◆
⇢)µ+ 12 (1 ⇢)2 2 ✓ µ
= e(1 prob ỹ < ,
and we have 2 ⇣ 2
⌘3
✓ µ (1 ⇢)
1+ N
⇢)µ+ 12 (1 ⇢)2 2
4 5.
⇠1 ⇢
= e(1 ⇣ ⌘
✓ µ
1+ N
Taking logs and dividing by 1 ⇢ yields
2 ⇣ 2
⌘3
✓ µ (1 ⇢)
1 1 1+ N
✓ = µ + (1 ⇢) 2
+ log 4 ⇣ ⌘ 5.
2 1 ⇢ 1+ N ✓ µ
214 20 Alternative Preferences in Single-Period Models
= 0.3 ) ⇡/ = 0.53 ,
= 0.1 ) ⇡/ = 0.37 ,
= 0.01 ) ⇡/ = 0.286 ,
= 0.001 ) ⇡/ = 0.277 ,
= 0.0001 ) ⇡/ = 0.276 .
21
21.1. Consider consumption processes (ii) and (iii) in the introduction to this chapter. Take T = 2.
Suppose consumption C0 is known at date 0 (before any coins are tossed). Assume the power
certainty equivalent and the CES aggregator.
(a) Assume two coins are tossed at date 0 determining C1 and C2 . Calculate the utility U0 of the
person before the coins are tossed.
(b) Assume a coin is tossed at date 1 determining C1 , and a coin is tossed at date 2 determining
C2 . Calculate the utility U0 .
(c) Show numerically that the utility is higher in part (a) than in part (b) — i.e., early resolution
of uncertainty is preferred — if ⇢ > ↵. Show that late resolution is preferred if ⇢ < ↵, and show
that the person is indi↵erent about the timing of resolution of uncertainty if ⇢ = ↵.
This implies
1 X X 1 ⇢
⇠01 ⇢
= E[U11 ⇢
]= x1 ↵
+ y1 ↵ 1 ↵
.
4
x2{a,b} y2{a,b}
It follows that
8 2 31 ↵
9 1
1 ↵
>
< 1 ⇢ >
=
1 X X 1 ⇢
U0 = V (C0 , ⇠0 ) = C01 ↵
+ 4 x1 ↵
+ y1 ↵ 1 ↵ 5 .
>
: 4 >
;
x2{a,b} y2{a,b}
(b) We have
216 21 Alternative Preferences in Dynamic Models
1 X 1
⇠11 ⇢
= E1 [U21 ⇢
]= y ⇢
.
2
y2{a,b}
Therefore,
2 0 11 ↵ 3 1 1↵
1 ⇢
6 1 X 7
U1 = V (C1 , ⇠1 ) = 4C11 ↵
+ @ y1 ⇢A
5 .
2
y2{a,b}
Hence,
2 0 11 ↵ 3 11 ⇢
↵
1 ⇢
1 X 6 1 1 X 7
⇠01 ⇢
= E[U11 ⇢
]= 4x
↵
+ @ y1 ⇢A
5 .
2 2
x2{a,b} y2{a,b}
It follows that
0 1 1
8 2 3 11 ⇢ 9 11 ↵ 1 ↵
> 0 1 1 ↵ ↵ > ⇢
B >
< 1 ⇢ >
= C
B 1 1 X 6 1 1 X 1 7 C
U0 = V (C0 , ⇠0 ) = B
BC0
↵
+ 4x
↵
+ @ y ⇢A
5 C
C .
@ >
> 2 2 >
> A
: x2{a,b} y2{a,b} ;
(c) For example, take a = 100, b = 200, C0 = 150, = 0.9, and ↵ = 4. The utility in part (a) is
21.2. Derive (21.12a)–(21.12f) in the order given. To compute B, use (21.8), (21.9) and the fact
that ✓ ◆
Wt+1 1 Ct+1
Rm,t+1 = = .
Wt C t 1 ✓ Ct
h i
1 ⇢
B1 ⇢
= E Rm,t+1
"✓ ◆ #
1 ↵
1 ⇢ Ct+1 1 ⇢
= B ↵
E
Ct
✓ ◆
1 ↵
1 ⇢ 1 2 2
= B ↵
exp (1 ⇢)µ + (1 ⇢) .
2
defined in (21.12f).
21.3. Consider the continuous-time portfolio choice problem with linear habit (21.15). Assume
(21.16) holds with strict inequality. Repeating the argument at the end of Section 21.5 shows that,
for any date t,
Z T Z T Z T
(a b)(u t)
Et Mu Cu du = Et M̂u Ĉu du + Xt Et Mu e du .
t t t
1
(Section 21.5 considers t = 0.) Assume power utility: u(c x) = 1 ⇢ (c x)1 ⇢. Assume the
information in the economy is generated by a single Brownian motion B, there is a constant risk-
free rate r, and there is a single risky asset with constant expected rate of return µ and constant
volatility .
218 21 Alternative Preferences in Dynamic Models
for a constant K.
(b) Define
1 h i
(r+a b)(T t)
t = 1 e .
r+a b
Show that
Z T
Xt (a b)(u t)
Et Mu e du = t Xt ,
Mt t
M̂t = (1 + b t )Mt ,
Z T 1
1 ⇢
Et M̂u Ĉu du = t Mt ,
Mt t
Show that
1
✓ ◆
⇢ µ r
dWt = t Mt dBt + something dt
⇢
✓ ◆
µ r
= (Wt t Xt ) dBt + something dt .
⇢
⇢
Ĉt = M̂t .
yields the first two formulas. For the third, note that
1 1
( /⇢)u 1 ⇢ ( /⇢)u 1 1 1 ⇢
M̂u Ĉu = Ke M̂u = Ke (1 + b u) ⇢ Mu .
Therefore,
21 Alternative Preferences in Dynamic Models 219
Z Z "✓ ◆1 1 #
T 1 T
1 ⇢ ( /⇢)u 1 1 Mu ⇢
Et M̂u Ĉu du = KMt e (1 + b u) ⇢ Et du .
Mt t t Mt
Also, from
dM µ r
= r dt dB ,
M
we obtain
"✓ ◆1 1 # ✓ ◆ ✓ ◆2 ✓ ◆ !
Mu ⇢ ⇢ 1 1 µ r ⇢ 1
Et = exp r (u t) (u t) .
Mt ⇢ 2 ⇢2
21.4. Consider the model with a single risky asset described in Section 21.4. Assume the investor
is a representative investor, there is a single unit of the risky asset, and the risk-free asset is in zero
net supply. Assume Bayesian updating of the set of priors, and assume the representative investor
is risk neutral (but ambiguity averse). Let Pt denote the price of the risky asset at date t, with
P2 = x̃. Take Rf = 1. The intertemporal budget equation is
Wt+1 = Wt + ✓t (Pt+1 Pt ) ,
where ✓t denote the number of shares of the risky asset chosen at date t. The distribution of x̃
conditional on s̃ is normal with mean µ + (s̃ µ) and variance (1 ) 2, and the marginal
distribution of s̃ is normal with mean µ and variance 2/ , for
a b. Take the backward
induction (dynamic programming) approach.
220 21 Alternative Preferences in Dynamic Models
if and only if
P1 = µ + b (s̃ µ) .
if and only if
P1 = µ + a (s̃ µ) .
(c) Suppose that P1 depends on s̃ as described in the previous parts. Show that ✓0 = 1 maximizes
min E [W1 ]
if and only if
( b a)
P0 = µ p .
2⇡ a
Hint: The function P1 is concave in s̃. Hence, E [P1 (s̃)] < P1 (µ), and the di↵erence P1 (µ)
E [P1 (s̃)] is maximized at the maximum variance for s̃.
(d) Now take the forward planning approach. Let P0 and P1 (s) be as described in the previous
parts. The investor chooses ✓0 and a plan s 7! ✓1 (s) at date 0 to maximize
Show that the investor can achieve unbounded worst-case expected wealth. In particular, choos-
ing ✓0 = 1 and ✓s (1) = 1 for all s is not optimal.
= W1 + ✓1 [µ + a (s̃ µ) P1 ] .
21 Alternative Preferences in Dynamic Models 221
µ+ a (s̃ µ) P1 0.
= W1 + ✓1 [µ + b (s̃ µ) P1 ] .
µ+ b (s̃ µ) P1 0 .
µ+ b (s̃ µ) P1 µ + a (s̃ µ) .
If the first inequality is an equality, then any long position is optimal. If the second is an
equality, then any short position is optimal. If both are strict inequalities, then ✓1 = 0 is the
unique optimum. Thus, ✓1 = 1 is optimal if and only if
P1 = µ + b (s̃ µ) .
= W1 + ✓1 [µ + b (s̃ µ) P1 ] .
µ+ b (s̃ µ) P1 0.
= W1 + ✓1 [µ + a (s̃ µ) P1 ] .
µ+ a (s̃ µ) P1 0 .
µ+ a (s̃ µ) P1 µ + b (s̃ µ) .
If the first inequality is an equality, then any long position is optimal. If the second is an
equality, then any short position is optimal. If both are strict inequalities, then ✓1 = 0 is the
unique optimum. Thus, ✓1 = 1 is optimal if and only if
P1 = µ + a (s̃ µ) .
(c) We have
E [W1 ] = W0 ✓0 P0 + ✓0 E [P1 (s̃)] .
If ✓0 < 0, then E [W1 ] is minimized by maximizing E [P1 (s̃)], which is equivalent to minimizing
the variance of s̃; i.e., taking = b. Thus, for ✓0 < 0,
⇣ ⌘
min E [W1 ] = W0 + ✓0 E b [P1 (s̃)] P0 .
E b [P1 (s̃)] P0 0.
If ✓0 > 0, then E [W1 ] is minimized by minimizing E [P1 (s̃)], which is equivalent to maximizing
the variance of s̃; i.e., taking = a. Thus, for ✓0 > 0,
⇣ ⌘
min E [W1 ] = W0 + ✓0 E a [P1 (s̃)] P0 .
E a [P1 (s̃)] P0 0 .
If the first inequality is an equality, then any long position is optimal. If the second is an
equality, then any short position is optimal. If both are strict inequalities, then ✓1 = 0 is the
unique optimum. Thus, ✓1 = 1 is optimal if and only if
21 Alternative Preferences in Dynamic Models 223
P0 = E a [P1 (s̃)] .
Thus,
⇥ ⇤
E a [P1 (s̃)] = µ ( b a )E
a
(s̃ µ)1{s̃>µ}
Z 1 ✓ ◆
x x2
=µ ( b a) p exp dx .
0 2⇡ 2/ 2 2/
where n denotes the normal density with mean zero and variance 2. We have
Z 1
1
n0 (x) dx = n(1) n(0) = n(0) = p .
0 2⇡ 2
Thus, ✓ ◆
Z 1
x x2
p exp dx = p ,
0 2⇡ 2 2 2 2⇡
which yields the result.
(d) For any positive number n, consider ✓0 = n and ✓1 (s) = n for all s. Then, for each ,
⇣ ⌘
E [W2 ] = W0 + n E [x̃] P0 = W0 + n(µ P0 ) ! 1
as n ! 1.
22
Production Models
22.1. Assume there is costless adjustment of capital, i.e. Kt+1 = Kt + It for any It .
(a) Show that a capital stock process is optimal if and only if Kt solves
Mt
max Et 1 [⇡(K, Xt ) + K] K
K Mt 1
for each t.
(b) Show that the first-order condition for the maximization problem in the previous part is a
special case of (22.4a).
This equals
" 1
#
X
M0 [⇡(X0 , K0 ) + K0 ] + E Mt ⇡(Xt , Kt ) Mt 1 Kt + M t Kt ] .
t=1
E[Mt ⇡(Xt , Kt ) Mt 1 Kt + M t Kt ] .
22.2. Assume a firm combines labor L with capital K to produce output Qt = Kt↵ Lt , where ↵
and are constants with ↵ + 1. The firm’s operating cash flow is Pt Qt Wt Lt , optimized over
Lt 0, where P and W are regarded as exogenous stochastic processes. Show that operating cash
flow equals Xt Kt for some exogenous stochastic process X and constant 1, and show that
= 1 if ↵ + = 1.
W L = P K ↵L ,
22.3. Assume a firm can produce any output Qt up to a maximum capacity Zt Kt , where Z is
1/
a positive stochastic process. Assume the output price is Pt = Yt Qt for a positive stochastic
process Y and constant > 1. Assume operating cash flow equals revenue Pt Qt .
22 Production Models 227
(a) Show that it is optimal to produce at full capacity Qt = Zt Kt , and operating cash flow is
1 1/
X t Kt , where X = Y Z 1 1/ .
(b) Assume investment is irreversible and there is no depreciation. Assume the risk-free rate r is
constant, the market price of risk is constant, and X is a geometric Brownian motion with drift
µ and volatility . Assume the correlation of X and M is a constant ⇢, and assume µ ⇢ < r.
Show that the value (22.22) of the asset paying marginal cash flow is
✓ ◆
1 1
k 1/ Xt ,
r µ+ ⇢
(r + ⇢) dt + dB ,
Integrate the values of the growth options over k 2 [K0 , 1) to compute the value of the firm.
What condition on and is needed for the value of the firm to be finite?
Mu X u (r µ+ ⇢)(u t)
Et =e .
Mt X t
It follows that the value of the asset paying marginal cash flow is
✓ ◆✓ ◆
1 Xt
k 1/ .
r µ + ⇢)
(r µ+ ⇢) dt + µ dt + dB = (r + ⇢) dt + dB .
(d) Observing that the strike of the option is 1, Exercise 15.5 shows that the optimal exercise time
is /( 1), where is the positive root of the quadratic equation (15.32). In that equation,
q denotes the dividend yield, which here is r µ+ ⇢. Substituting this value of q, (15.32) is
the same as (22.27). The condition St (k) = /( 1) is equivalent to
✓ ◆✓ ◆
1/
k Xt = (r µ+ ⇢) ,
1 1
and
1/ 1/ 1 1/ 1/
Kt X t = Kt Yt Zt = Qt Y t Z t = Pt Z t .
for all t.
22 Production Models 229
The integral of k / over [K0 , 1) is finite if and only if > . Assuming > ,
Z 1
/ ( )/
k dk = K0 ,
K0
22.4. Under the assumptions of Section 22.8, marginal q (22.23) is a function of Xt . Denote this
function by f . The purpose of this exercise is to derive the formula (22.26) for f .
is a martingale.
(b) Using Itô’s formula and the result of the previous part, show that f satisfies the di↵erential
equation
1
x + (µ ⇢)xf 0 (x) + 2 2 00
x f (x) = rf (x)
2
for 0 < x < x⇤ .
(c) Show that the function
x
r µ+ ⇢
satisfies the di↵erential equation.
230 22 Production Models
1
(µ ⇢)xf 0 (x) + 2 2 00
x f (x) = rf (x)
2
is satisfied by Ax for constants A and if and only if is a root of the quadratic equation
(22.27).
(e) The general solution of the di↵erential equation in Part (b) is
x
+ A1 x 1
+ A2 x 2
r µ+ ⇢
for constants Ai , where the i are the roots of the quadratic equation. From the definition of f ,
it follows that limx#0 f (x) = 0. Use this fact to show that
x
f (x) = + Ax ,
r µ+ ⇢
(a) We have
Z t Z t Z 1
Mu
Mu Xu du + Mt f (Xt ) = Mu X u + Mt E t Xu du
0 0 u Mt
Z 1
= Et Mu Xu du .
0
Also,
1
df = f 0 (X) dX + f 00 (X) (dX)2
2
✓ ◆
0 dX 1 2 00 dX 2
= Xf (X) + X f (X) ,
X 2 X
and
dX dW 1 dKb
= .
X W b
K
22 Production Models 231
Hence, ✓ ◆2 ✓ ◆2
dX dW 2
= = dt ,
X W
and
✓ ◆ ✓ ◆✓ ◆ ✓ ◆✓ ◆
dM 0 dX dM 0 dW dM
(df ) = Xf (X) = Xf (X) = ⇢Xf 0 (X) dt .
M X M W M
Therefore, the coefficient of dt in the di↵erential is
1 2 2 00
M X + µXf 0 (X) + X f (X) rf ⇢Xf 0 (X) .
2
Equating this to zero yields the di↵erential equation.
(c) For this function f ,
1 (µ ⇢)x rx
x + (µ ⇢)xf 0 (x) + 2 2 00
x f (x) = x + = = rf (x) .
2 r µ+ ⇢ r µ+ ⇢
(d) Given f (x) = Ax , we have
1 1
(µ ⇢)xf 0 (x) + 2 2 00
x f (x) rf (x) = (µ ⇢)Ax + ( 1) 2
Ax rAx .
2 2
This is equal to zero for all x 6= 0 if and only if
1 2
(µ ⇢) + ( 1) r = 0,
2
which is the quadratic equation (22.27).
(e) If 2 < 0, then f (x) ! 1 as x ! 0, unless A2 = 0.
(f) The term involving dK̂ in the calculation in Part (b) comes from M df and is equal to
1 dKb
M Xf 0 (X)
Kb
This is equal to zero when Xt = x⇤ if and only if f 0 (x⇤ ) = 0. We have
1
f 0 (x) = + Ax 1
.
r µ+ ⇢
Setting this equal to zero at x = x⇤ yields
✓ ◆
1 1
A= (x⇤ )1 .
r µ+ ⇢
Hence,
✓ ◆
1 1
f (x) = x (x⇤ )1 x
r µ+ ⇢
✓ ◆
x 1⇣x⌘ 1
= 1 .
r µ+ ⇢ x⇤