_
log 1.1 0.1
0.2
_
= (0.4116)(0.02345) = 0.65970.4906 = 0.1691.
I used a statistical package to get the nal normal distribution probabilities, but using tables
will give almost the same answer.
3
3. The per period return X on a certain security is modelled as
X =
_
+ with probability
1
2
,
with probability
1
2
,
where and > 0 are constants. Let A = 1 +X be the corresponding per period accumulation
factor (or price relative; so X is not a log-return).
(a) Show that the mean and variance of A are 1 + and
2
, respectively.
(b) Assuming < 1 + (why??), show that the geometric mean of A is
E
g
(A) =
_
1 + 2 +
2
2
.
(c) In nancial situations and typically are near zero. Use the expansion
1 +z 1 +
1
2
z
1
8
z
2
+ ,
valid for z 0, to derive the approximation
E
g
(A) 1 +
1
2
2
= E(A)
1
2
V ar(A).
Compare the exact and approximate values for the following parameter values: (i)
1
= 13.0%
and = 20.2%, and (ii)
2
= 6.0% and = 8.7%. Express the results in percentage terms.
This approximation appears in texts on investments, and it is discussed at some length by J.D.
McBeth, Whats the long-term expected return to your portfolio, Financial Analysts Journal,
51(5) (1995), 68. It is in fact a rather poor approximation; see O. de La Grandville, The
long-term expected rate of return: setting it right, Financial Analysts Journal, 54(5) (1998),
7580.
Answer. (a) Clearly E(X) =
1
2
( + ) +
1
2
( ) = , and hence E(A) = 1 + . Similarly
E(X
2
) =
1
2
( + )
2
+
1
2
( )
2
=
2
+
2
. Hence V ar(X) =
2
(or, just observe that
(X )
2
=
2
) and V ar(A) = V ar(1 +X) =
2
.
(b) Compute
E(log A) =
1
2
log(1 + +) +
1
2
log(1 + ) =
1
2
log[(1 +)
2
2
] = log
_
1 + 2 +
2
2
.
The restriction < 1 + is needed to ensure that the argument of the second logarithm term
is positive. This isnt a real restriction in nancial situations. From its denition the geometric
expectation is
E
g
(A) = e
E(log A)
=
_
1 + 2 +
2
2
.
(c) Let z = 2 +
2
2
in the square-root approximation given in the question. Then using
all the terms displayed on the RHS,
E
g
(A) 1+
1
2
(2+
2
2
)
1
8
(2+
2
2
)
2
= 1+
1
2
1
8
(
4
+
4
+4
3
4
2
2
2
2
) 1+
1
2
2
.
Here I have neglected all terms of higher than second order on the basis that they contribute
very little in comparison to the terms retained. This is reasonable since and are rarely more
than 0.2.
1
Estimated from S&P 500 Stock Composite Index (1926-2000).
2
Estimated from Salomon Brothers long-term, high-grade corporate bond total return index (1926-2000).
4
Asset type % % E
g
(Y )% Approx.
Common Stocks 12.98 20.17 11.16 10.94
Small Company Stocks 17.3 33.4 12.44 11.72
Long-term Corp. Bonds 6.0 8.7 5.64 5.62
Long-term Govt. Bonds 5.7 9.4 5.28 5.26
4. Let A
n
denote the accumulation factor for period n and assume these factors are iid. For
0, let M() = E(A
n
) denote the moment function, assuming it is nite. The horizon-N
mean per-period accumulation g
N
can be expressed in terms of M() as g
N
= [M(1/N)]
N
.
Recall for the binomial tree model that g
N
= (pu
1/N
+qd
1/N
)
N
.
Show for large N that g
N
(1 + N
1
(p log u + q log d))
N
and hence that lim
N
g
N
=
g
= u
p
d
q
, where the second equality is shown in lectures by direct calculation. [Hint: Write
u
1/N
= exp(N
1
log u) and approximate.]
Answer. You are told that
g
N
=
_
pu
1/N
+qd
1/N
_
N
= exp
_
N log(pu
1/N
+qd
1/N
)
_
.
Now pu
1/N
+qd
1/N
1 as N , so you need an estimate of the dierence D
N
= p(u
1/N
1) +q(d
1/N
1). Use
x
1/N
1 = exp
_
N
1
log x
_
1 = 1 +N
1
log x +O(N
2
) 1,
so D
N
= N
1
(p log u +q log d) +O(N
2
), giving
log
_
pu
1/N
+qd
1/N
_
= log(1 +D
N
) = D
N
+O(D
2
N
) = N
1
(pu
1/N
+qd
1/N
) +O(N
2
).
It follows from this that
lim
N
N log
_
pu
1/N
+qd
1/N
_
= p log u +q log d,
i.e. g
N
exp(p log u +q log d) = u
p
d
q
= g
.
5. Suppose that the accumulation factors A
n
have the gamma law with pdf
f
A
(x) =
x
1
e
x
/(), (x > 0),
where () =
_
0
v
1
e
v
dv is the gamma function. You can see almost by inspection (a minds
eye integration by parts) that (n + 1) = n!, i.e. the gamma function smoothly interpolates
factorials. (In fact it is the unique log-convex function to do so, the Bohr-Mollerup theorem.)
Gamma functions can be evaluated using the factorial button on some calculators, () =
( 1)!.
(a) Show that the moment function
M() := E(A
) =
( +)
()
.
5
Hence show that a = /, s
2
= /
2
, and that
g
N
=
1
_
( + 1/N)
()
_
N
.
The logarithmic derivative of the gamma function is called the psi-function:
() =
d
d
log () =
()
()
, ( > 0).
[Numerical values of the gamma and psi functions can be calculated using the Maple package
with the commands
evalf(GAMMA(x)); and evalf(Psi(x));
where x represents a specic numerical value. To execute the command, use the enter button
and not return. Other packages can be used too.]
(b) Use a MacLaurin expansion to show for large N that
g
N
1
(1 +()/N)
N
,
Hence show that g
=
1
exp(()).
Answer. (a). Compute the moment function
M() := E(A
) =
_
x
f(x)dx = (
/())
_
0
x
+1
e
x
dx
= (
/())
_
0
(x)
+1
e
x
d(x) =
( +)/().
This gives a = M(1) = /, a
2
= M(2) = ( + 1)/
2
= /
2
+a
2
, whence s
2
= /
2
.
(b) Since M() =
(), whence
g
N
1
_
1 +N
1
()
()
_
N
=
1
_
1 +
()
N
_
N
.
Let N to get g
=
1
exp(()), using lim
N
(1 +z/N)
N
= e
z
.
6. Recall the lecture note stu in 3.3,6 about the mean, variance and geometric mean
for the binomial tree model, and read 3.7 about parameter estimation. There are three free
parameters, p, u > 1 and d < 1, which must be estimated using observed data. Often this
problem is simplied by setting d = u
1
.
(a) Assume this holds, and write down expressions for a and a
2
in terms of p and u. By
eliminating p, show that u solves the quadratic equation au
2
(1 +a
2
)u +a = 0. [This can be
obtained in a few lines. You should not need pages of algebra.]
6
Substituting the sample mean a of observed one-period accumulations for a, and the corre-
sponding sample second-order moment a
2
for a
2
, and solving the resulting quadratic equation
gives a number u called the method of moments estimate of u.
(b) Use the S&P values from the example in printed notes 3.7 to compute u and a correspond-
ing estimate p from a. Hence estimate the 75 year mean return
75
and
. Compare
75
with
the empirical and the lognormal estimates in Table 3, 3.7.
Answer. Lecture calculations have shown you that a = pu + qd and a
2
= pu
2
+ qd
2
(*).
Multiply the rst equation by u and by d to obtain the two equations au = pu
2
+ q and
ad = p+qd
2
. Add these equations and compare the result with (*) to get au+ad = pu
2
+qd
2
+
ud = a
2
+ud. Taking d = 1/u gives the quadratic equation
au
2
(1 +a
2
)u +a = 0.
This derivation involves a bit of algebraic trickery to avoid cubic and quartic terms.
Alternatively, solving the rst equation for p and using d = 1/u leads to
p =
a d
u d
=
au 1
u
2
1
. (1)
This gives
q = 1 p =
u
2
au
u
2
1
so qd
2
=
u a
u(u
2
1)
.
Substituting these into the identity for a
2
and putting the right-hand side over the common
denominator u(u
2
1) (do it!) leads to
a
2
u =
u
3
(au 1) +u a
u
2
1
=
a(u
4
1) u(u
2
1)
u
2
1
= a(u
2
+ 1) u,
as above.
The solution is
u =
1
2a
_
1 +a
2
_
(a
2
+ 1)
2
4a
2
_
.
The rst line of Table 3 (for the S & P 500 composite index) in Chapter 3 gives r = 12.98%,
i.e. a = 1.1298. Similarly s = 0.2017, so a
2
= s
2
+ a
2
= 1.317131. Substituting these sample
estimates into the quadratic solution gives a method-of-moments estimate of u. One solution for
u exceeds unity, and the other < 1. Use the larger: u = 1.252550. Next, (1) yields p = 0.72973.
(If you round to u = 1.2525 then p = 0.7298, and if u = 1.2526 then p = 0.7297.)
Use these values to compute (hats omitted) E(A
1/75
) = pu
1/75
+qu
1/75
= 1.001384, giving
g
75
= [E(A
1/75
)]
75
= 1.109300, i.e.
75
= 10.93%. Also, the assumption about d implies that
g
= u
2p1
= 1.109004, i.e.