Anda di halaman 1dari 20

Decisive Drift Terms in Future Bond Prices.

Thomas Gustavsson
August 23, 2009
Abstract. In both arbitrage-free and ane models the bond price is the
product of a drift factor and a term that is related to the variance of the
stochastic disturbance. We identify the drift factor with the solution to the
underlying deterministic system for the short rate. This brings us back to
the beginning of stochastic calculus: How is the solution to a deterministic
system aected by the addition of a stochastic disturbance? (cp Karatzas
and Shreve (1988, Example 3.3.11).
Arbitrage-free bond prices were derived by Heath, Jarrow, and Morton
(1992) and independently by Artzner and Delbaen (1989), and Babbs (1989).
But the basic economic principles leading to these results were not made
clear. In particular, the relationship between arbitrage-free bond pricing,
instantaneous forward rates and ane factor models need more explaining.
Arbitrage-free pricing of bonds makes the short instantaneous interest
rate process redundant. As a result we can solve for the short rate process in
terms of bond price volatilities. This leads to the same bond pricing formulas
as in Heath, Jarrow, Morton (1992) but without any instantaneous forward
rates. Indeed, to highlight the simple economic mechanisms underlying their
result it is better to use just the short instantaneous rate. In addition, the
peculiar form assumed by the short rate process under arbitrage-free pricing
of bonds explains why an exogeneousy chosen short rate process in ane
factor models typically will not t any given initial term structure of bond
prices (or the corresponding volatilities). For this to happen the mean must
match the common drift of the bond prices in the arbitrage-free framework.
This seems to be a forgotten aspect of bond pricing in general.
1 HJM without instantaneous forward rates
The standard approach to arbitrage-free pricing is a model where all relevant
information is public and free. Prices are assumed to be determined in com-
1
petitive trading with plenty of rms, not just a few dominant trend-setters,
in which case strategic behaviour would inuence prices. Furthermore any
liquidity considerations, bid-ask spreads and transactions costs are ignored.
There are zero coupon bonds for all maturities and their prices decrease
smoothly with maturity. The restrictiveness of these underlying economic
assumptions should not be forgotten in the midst of all mathematical for-
mulas.
There is a probability space (, F, Q

) and a ltration {F
t
}, which for
is assumed to be generated by a few Brownian motions. The probability
measure Q

is known as the risk-neutral probability measure. Asset prices


are semimartingales.
Within this framework the arbitrage-free price for any non-dividend pay-
ing asset X can be written symbolically as
dX(t)
X(t)
= r(t)dt +(s, X)dW

s
(1)
where X(t) is the asset price, r(t) the common drift (equal to the short rate
used for discounting), and W

t
is a standard Brownian motion with respect
to the risk-neutral measure Q

. We refer to (s, X) as the volatility of the


asset price. In general, the volatility may be a previsible stochastic process,
possibly path-dependent, and dierent for each asset X.
1
Although the multifactor interpretation is obvious we limit our attention
to one Brownian motion. We shall use the integrated form of (1) which in
view of Itos lemma for the product of a nite variation process and a Doleans
martingale
2
is
X(t) = X
0
exp
__
t
0
r(s)ds
1
2
_
t
0
|(s, X)|
2
ds +
_
t
0
(s, X)dW

s
_
(2)
In addition to the unspecied number of assets X the market consists of
a family of zero coupon bonds {B(, T)}
T0
, also known as discount bonds
or simply bonds. These are special assets with a nite life and the same
known future payo, say , upon expiry. According to standard notation
B(t, T) is the price at time t of a discount bond that matures at time T, so
B(T, T) = 1. The rst argument represents the time index of a stochastic
process while the second argument stands for a member of a function family.
As bonds are the current value of a future 1 we must have 0 < B(t, T) 1
1
The existence of such a process is guaranteed by martingale representation for a
Brownian ltration, see Rogers and Williams (1987, Theorem IV.36.11).)
2
see Rogers and Williams (1987, Theorem IV.37.1 and Lemma IV.32.1)
2
for all t and all T. Furthermore, economic arguments lead us to assume that
bond prices decrease with maturity in a smooth manner: for each xed t the
function T B(t, T) is dierentiable. The derivative with respect to the
second argument is denoted
2
B(t, T). Thus, the time indeces represent two
extremes with respect to smoothness: while the rst calender time index is a
jagged stochastic process, not likely to be dierentiable anywhere, the other
index, maturity, can be as smooth as one wants. In practice only a nite
number of maturities are ever traded so typically intermediate maturities
are interpolated.
Focusing our attention on the family of discount bonds we can show that
any common drift becomes redundant. (Provided of course, that there are
bonds expiring at any maturity, which was assumed at the outset) To see
this we use an argument from El Karoui and Geman (1991), also used by
Flesaker and Houghston (1995). The arbitrage-free price of a bond maturing
at time T according to (2) is
B(t, T) = B
0,T
exp
__
t
0
r(s)ds
1
2
_
t
0
|(s, t)|
2
ds +
_
t
0
(s, t)dW

s
_
(3)
We refer to (s, t) as the volatility at time s of a bond maturing at the xed
date T. As for any asset X the volatility of a bond must be previsible but
it may depend on current and past bond prices in several dierent ways.
By assumption the value of the bond upon maturity must equal one. So
for t = T we get
1 = B
0,t
exp
__
t
0
r(s)ds
1
2
_
t
0
|(s, t)|
2
ds +
_
t
0
(s, t)dW

s
_
(4)
As this holds for all t we can take the logarithm and solve for the common
drift which must be
_
t
0
r(s)ds = ln B
0,t
+
1
2
_
t
0
|(s, t)|
2
ds
_
t
0
(s, t)dW

s
(5)
In other words, the common drift is a function of the initial term structure of
bond prices {B
0,t
}
t0
and future and initial bond price volatilities {(s, t) :
0 s t} for all t 0. This is a very special nite variation process. It is
derived point by point from the bond price volatilities. At each time t the
common drift picks up a whole new set of bond price volatilities {(s, t), 0
s t}, and a new starting value B
0,t
. The time index t plays a double role
in the expression for the common drift: on the one hand it denes the range
of integration, and on the other hand it denes which bond price is being
3
integrated. This happens because on the maturity date of each bond the
drift and volatility parts must exactly cancel out to force the bond price
process to one. During the nal moment of maturing this introduces a
redundancy which must be resolved in order to avoid arbitrage between the
bond and the short rate.
Thus, replacing the common drift in (2) with the bond price volatilities
(5) the arbitrage free price of any (non-dividend) paying asset X must be
X(t) =
X
0
B
0,t
exp
_

1
2
_
t
0
{|(s, X)|
2
|(s, t)|
2
}ds +
_
t
0
{(s, X) (s, t)}dW

s
_
(6)
In particular, for the arbitrage-free bond price we get
B(t, T) =
B
0,T
B
0,t
exp
_

1
2
_
t
0
{|(s, T)|
2
|(s, t)|
2
}ds +
_
t
0
{(s, T) (s, t)}dW

s
_
(7)
This is, basically, the bond pricing result in Heath, Jarrow, and Morton
(1992). In fact, assuming as they do (their Condition C.1) that bond price
volatilities increase with maturity, so that the derivative with respect to
maturity exists, we can solve for the short rate process. Equivalently, if all
short rates are positive then the nite variation process (5) is an increasing
process. In this case the derivative with respect to maturity exists, see Cohn
(1980, Theorem 6.3.3), and can be written
r(t) =
2
ln B
0,t
+
1
2
_
t
0

2
|(s, t)|
2
ds
_
t
0

2
(s, t)dW

s
(8)
where
2
|(s, t)|
2
= 2(s, t)
2
(s, t). Disregarding the dierence in notation
this is precisely formula (26) in Heath, Jarrow, and Morton (1992, p. 90).
The dierence in sign on the Brownian motion occurs because they choose
to start with forward rates instead of bond prices. Implicit here are the
exact conditions on
2
(s, t) for which r(t) really is positive. But there was
no need for any instantaneous forward rates.
Using the volatility derivatives we can write bond prices in (7) a bit more
compactly as
B(t, T) =
B
0,T
B
0,t
exp
_

1
2
_
t
0
_
T
t

2
|
s,u
|
2
duds +
_
t
0
_
T
t

s,u
dudW

s
_
(9)
Interchanging the order of integration we can identify the future instanta-
neous forward rates. Then the bond price in (9) simply becomes
B(t, T) = exp
_

_
T
t
f(t, u)du
_
4
where the forward rates are dened as f(t, u) =
2
ln B(t, u) and
f(t, u) =
2
ln B
0,u
+
1
2
_
t
0

2
|
s,u
|
2
ds
_
t
0

s,u
dW

s
Thus, provided short rates are positive and bond price volatilities increase
with maturity the arbitrage-free pricing of bonds can be written using the
instantaneous forward rates in HJM (1992).
However, these forward rates are completely unnecessary. The crucial
feature is that any common drift for a family of (zero coupon) bonds becomes
uniquely determined by the bond equal to one on maturity requirement.
2 The Simplest Example
When the initial term structure of bond prices is considered as a given
parameter the only thing left is to model the bond price volatilities. Fur-
thermore, if the initial term structure of volatilities too is taken as given
there remains only the future bond price volatilities. We now consider the
simplest possible case: a constant derivative of volatility at all times s and
all positive maturities u > s, i e

2
(s, u) = c > 0 (10)
where 0 s < u. The resulting bond pricing model is sometimes called the
continuous time limit of the basic Ho and Lee (1985) model. The derivative
of the bond price volatility is the volatility of the instantaneous forward rate
used by HJM. This explains why they choose to use instantaneous forward
rates as state variables. But we prefer to avoid them because the basic
economic principles get all mixed up. The constant derivative of volatility c
must be greater than zero because a necessary condition for positive interest
rates is that bond price volatilities increase with maturity. (If the volatilities
were constant their future value could never become one upon maturity as
required by the very denition of the bond asset.) To nd the bond price
volatilities we integrate (10) with respect to maturity
(s, t) =
_
t
s

2
(s, u)du = c(t s) (11)
where (s, s) = (t, t) = 0. Inserting this into (7) we can calculate the future
arbitrage-free bond price
B(t, T) =
B
0,T
B
0,t
exp
_

c
2
2
_
t
0
{|T s|
2
|t s|
2
}ds +
_
t
0
c(T t)dW

s
_
5
Note that the integrand in the stochastic integral is independent of s and
can be factored out. So evaluating the integrals we get
B(t, T) =
B
0,T
B
0,t
exp
_

c
2
2
tT(T t) +c(T t)W

t
_
(12)
as W

0
= 0. This is formula (30) in HJM (1992), disregarding the dierence
in sign for W

t
as before. The future bond price equals the initial forward
price times an exponential term related to the maturity of the bond, (T t),
the variance and the volatility of the short rate, c
2
t and c, respectively. The
volatility of the bond price, c(T t), equals the volatility of the short rate
times the time to maturity of the bond . Indeed, the time to maturity,
(T t), factors out in the entire exponential expression. This provides the
key link to another way to express arbitrage-free bond prices. When the
volatility derivatives are constant the short rate process (8) becomes
r(t) = f
0,t
+c
2
t
2
2
+cW

t
(13)
where f
0,t
=
2
ln B
0,t
is the initial forward rate. (We succumb to using
initial forward rates in the interest of a clear economic interpretation of the
initial slope of the term structure of bond prices.) Using this description of
the only possible shot rate process consistent with arbitrage-free pricing of
bonds we can rewrite the arbitrage-free bond price in terms of r(t), f
0,t
, and
c. Adding c
2
tT/2 to (13) we get
r(t) f
0,t
+
c
2
2
t(T t)
c
2
2
tT = cW

t
and inserting into (12) we get
B(t, T) =
B
0,T
B
0,t
exp
_
(T t)[r(t) f
0,t
]
c
2
2
t(T t)
2
_
(14)
This is the same expression for the arbitrage-free bond price as used by Hull
and White (1994, p.8).
3 An Alternative Derivation
The same arbitrage-free bond price formula can, of course, also be derived in
the more conventional way starting from the bond price as the conditional
expected value
B(t, T) = E

_
exp
_
T
t
r(u)du | F
t
_
(15)
6
By plugging in the arbitrage-free short rate process (13) and integrating
from t to T we get
_
T
t
r(u)du =
_
T
t
_
f
0,u
+c
2
u
2
2
+cW

u
_
du
Integrating and using the denition of the initial forward rate this equals
_
T
t
r(u)du = ln
B
0,T
B
0,t
+
c
2
2

(T
3
t
3
)
3
+c
_
T
t
W

u
du
Inserting this into (28) we can write the bond price as
B(t, T) =
B
0,T
B
0,t
exp
_

c
2
6
(T
3
t
3
)
_
E

_
exp
_
c
_
T
t
W

u
du
__
(16)
In words, the future bond price equals the implied forward price less two
expressions related to the stochastic disturbance. Next, to evaluate the
expected value we note that
c
_
T
t
W

u
du = c
_
Tt
0
W

u
du +c (T t)W

t
and according to Karlin and Taylor (1975, p.385) we have
E
_
exp
_
c
_
Tt
0
W

u
du
__
= exp
_

c
2
2
(T t)
3
3
_
Expanding the exponential term we get
(T t)
3
3
=
T
3
t
3
3
tT(T t)
Inserting these expressions in(16) we get precisely the same bond price as
in (12). As we shall see in the next section this kind of derivation can be
used to describe the ane models approach to bond pricing.
4 Ane Models
The arbitrage-free approach to bond pricing starts with a given initial term
structure of bond prices and models their future volatilities. In contrast,
ane models start by specifying a model for the short rate process and then
calculates bond prices and their volatilities. In general the resulting prices
7
will not t any given initial term structure. However, if the parameters for
the short rate process depend on time in a special way a perfect match can
be accomplished. This result was properly described and proved by Hull and
White (1990). Here we provide a new argument based on our reformulation
of the basic arbitrage-free bond pricing result in the previous section.
Typically ane models use a linear stochastic dierential equation to
dene a vector of interest rates or factors (X
t
)
t0
dX
t
= (A(t)X
t
+b(t))dt +c(t)dW

t
(17)
where, A(t) is a matrix which may depend on time, b(t) and c(t) are vectors,
and W

t
is a multidimensional standard Brownian motion (with respect to
the risk-neutral probability measure). This particular stochastic dierential
equation has a known solution which is
X
t
= G(t)
_
X
0
+
_
t
0
G
1
(s)b(s)ds +
_
t
0
G
1
(s)c(s)dW

s
_
(18)
where G(t) = e
A(t)
is the (fundamental) solution to the corresponding ho-
mogeneous rst order deterministic dierential equation system
dG(t)
dt
= A(t)G(t) and G(0) = I (19)
see Karatzas and Shreve (1988, section 5.6). The short rate is then identi-
ed as one of the components of the vector X, as in Brennan and Schwartz
(1982), or as the sum of some of its components (the factors), see e g
Longsta and Schwartz (1992). Either way we can write down the short
rate process in the more or less explicit form (18). For more information see
El Karoui and Lacoste (1992), Bajeux and Portait (1993). To keep things
simple we shall limit our attention to the case of one dimension when the
short rate can be identied with the single factor X.
Compare the explicit solution for r(t) = X
t
in (18) to the arbitrage-free
expression (8) for the short rate. In order for the two processes to match
exactly the drift terms must be the same
G(t)
_
r
0
+
_
t
0
G
1
(s)b(s)ds
_
= f
0,t
+
_
t
0

2
(s, t)(s, t)ds (20)
and the variances must be the same
G(t)
2
_
t
s
_
G
1
(u)c(u)
_
2
du = |(s, t)|
2
(21)
8
or G(t)G
1
(s)c(s) =
2
(s, t). So for the ane approach the bond price
volatility derivatives must be separable in s and t. Thus, ane models can-
not accomodate any functional form for future bond price volatilities. In
comparison, the arbitrage-free pricing approach does not constrain volatili-
ties in this way.
5 Classical Vasicek
At the end of his article Vasicek (1977) considered a specic case with only
one factor, the short rate, and A = a, b, and c are constant numbers. In
this case the linear equation system in (17) becomes just the one equation
dr(t) = a
_
r(t)
b
a
_
dt +c dW

t
(22)
for which the solution (18) is the famous Ornstein-Uhlenbeck process
r(t) = e
at
_
r
0
+
_
t
0
e
as
bds +
_
t
0
e
as
c dW

s
_
(23)
see Karatzas and Shreve (1988, Example 5.6.8). The mean and variance of
this process are well-known from statistics. Taking the expected value in
(23) we get
E

[r(t)] =
b
a
+e
at
(r
0

b
a
) (24)
and the variance is
V ar

[r(t)] =
c
2
2

1 e
2at
a
(25)
However, we already know from (20) and (21) what the mean and variance
must be to match the initial term structure of bond prices. In this case
G(t) = e
at
and the derivative of the bond price volatility must be

2
(s, t) = G(t)G
1
(s)c = ce
a(ts)
(26)
Integrating we get
(s, t) =
_
t
s

s,u
du = c
1 e
a(ts)
a
The mean on the other hand is
G(t)
_
r
0
+
_
t
0
G
1
(s)b(s)ds
_
=
b
a
+e
at
(r
0

b
a
) (27)
9
which in general cannot be equal to
f
0,t
+
_
t
0

2
(s, t)(s, t)ds = f
0,t
+c
2
_
t
0
e
a(ts)
1 e
a(ts)
a
ds =
= f
0,t
+
c
2
2a
2
(1 e
at
)
2
This shows that volatility is not the problem here. To match any initial
term structure of bond prices the mean, i e the drift term in the short rate
process, must have a special form. In particular, it must dependend on time.
6 The Deterministic Solution
The drift and volatility of the short rate process determines bond prices.
According to the arbitrage-free approach the deterministic part of the system
is given by the initial forward prices. In ane models it is given by the
exogeneously specied (or calibrated) drift of the short rate process. In both
cases we can nd the bond prices by calculating the conditional expected
value
B(t, T) = E

_
exp
_
T
t
r(u)du | F
t
_
(28)
For the ane model X(t, T) =
_
T
t
r(u)du is normally distributed with mean
M(t, T) and variance V (t, T)) say. In this case the expected value can be
found from the moment generating function for the normal distribution. We
have
B(t, T) = E

_
e
X(t,T)
| F
t
_
= e
M(t,T)V (t,T)/2
see Dybvig (1988) or Grimmet and Stirzaker (1995 p.168). Thus the bond
price is the product of a factor related to the mean times a factor related to
the variance of the integrated short rate. This decomposition is extremely
interesting because the mean does not involve the stochastic disturbance at
all. In fact, as we shall see the mean is the solution to the corresponding
deterministic system, for which the stochastic disturbance is zero.
Setting c = 0 in (22) we get an ordinary dierential equation
dr(u)
du
= a
_
r(u)
b
a
_
(29)
starting at time t with the deterministic solution
r(u)
b
a
= r
t

b
a
e
a(ut)
10
To nd the mean we integrate the short rate over the life-time of the bond
M(t, T) =
_
T
t
E

[r(u) | F
t
]du =
_
T
t
e
a(ut)
(r
t

b
a
)du +
b
a
_
T
t
du
or
M(t, T) =
1 e
a(Tt)
a
(r
t

b
a
) +
b
a
(T t)
Inserting this into (28) we can write stochastic bond prices as
B(t, T) = e
M(t,T)
E

_
exp
_
c
_
T
t
e
au
_
u
t
e
as
dW

s
du
_
| F
t
_
(30)
This decomposition makes it possible to evaluate the size of the eect on the
bond price of adding the stochastic distubance. If there were no disturbances
the expected value would equal one so it is correct to identify e
M(t,T)
as the
deterministic solution. The stochastic bond price equals the deterministic
bond price times a factor (the expected value) that is related to the vari-
ance of the stochastic disturbance. The addition of stochastic disturbances
to the deterministic system (19) translates into multiplication with a fac-
tor less than one. The stochastic bond price is lower than the deterministic
bond price. Typically, as can be veried by direct calculations using esti-
mated parameters, the deterministic solution accounts for over 90% of the
stochastic bond price! The exact relation obviously depends on the size of
the variance of the stochastic disturbance in relation to the speed of mean
reversion, i e the so-called diusion coecient.
7 The Stochastic Disturbance Term
We now proceed to calculate the expected value of the disturbance in (30).
First we change the order of integration
_
T
t
e
au
_
u
t
e
as
dW

s
du =
_
T
t
e
as
_
T
s
e
au
dudW

s
=
_
T
t
1 e
a(Ts)
a
dW

s
Then we calculate the quadratic variation of the latter which is
V (t, T) = c
2
_
T
t
_
1 e
a(Ts)
a
_
2
ds =
c
2
a
2
_
T
t
_
1 +e
2a(Ts)
2e
a(Ts)
_
ds =
=
c
2
a
2
_
(T t)
1 e
a(Tt)
a

(1 e
a(Tt)
)
2
2a
_
=
=
c
2
a
2
_
(T t)
Tt

a
2

2
Tt
_
11
where we have introduced the abbreviation

Tt
=
1 e
a(Tt)
a
Using the characteristic function for the remaining normally distributed ran-
dom variable with zero mean the conditional expected value equals
E

_
exp
_
c
_
T
t
e
au
_
u
t
e
as
dW

s
du
_
| F
t
_
= exp
_

1
2
V (t, T)
_
Inserting and doing purely algebraic transformations of the exponential in
(30) we get the same bond price as did Vasicek (1977)
B(t, T) = exp
_

Tt
r(t) +
_
b
a

c
2
4a
2
_
(
Tt
(T t))
c
2
4a

2
Tt
_
When the stochastic disturbance is taken into account the bond price be-
comes lower than that of the deterministic system. The eect can be de-
scribed as the multiplication of a term related to the variance of the distur-
bance and the maturity of the bond. This term subtracts half the variance
and depresses the asymptotic level b/a for the deterministic system (29) into
the new asymptotic level of the continuously compounded yield to maturity
y() lim
T
1
T t
ln B(t, T) =
b
a
_
1
c
2
4ab
_
for the disturbed stochastic system (22). By assumption the mean-reversion
level b/a is unaected by the stochastic disturbance. The relation between
the initial short rate and these two dierent levels determines the shape of
the yield curve. For an initial short rate less than the asymptotic level the
yield curve is rising with maturity. For a short rate starting between the
two levels the yield curve is humped. And, for a short rate higher than the
mean level all yields are falling with maturity. These results were described
in dierent terms by Vasicek (1977). In particular, he did not distinguish
between the mean level and the asymptotic level. The distinction between
the deterministic system and the stochastic (disturbed) system is standard
from systems theory although I have not seen it in the context of bond
pricing before.
8 The Market Price of Risk
Vasicek also considered a constant market price of risk . Depending on
the individual investors attitude to risk this constant may be positive or
12
negative, and as a consequence, may either increase or depress the stochastic
bond price in the general equilibrium framework. However, this does not
matter for pricing purposes. As all calculcations were performed with respect
to the risk-neutral probability measure Q

its value is already present in the


parameter b. The market price of risk can only be separated from b if we
introduce a model for the drift in the bond price under some other equivalent
probability measure, e g the actual probabilities. Only then is it relevant to
consider
(t, T) r(t) = (t, T) = c
Tt
where (t, T) and (t, T) are the drift and volatility of a bond maturing
at time T under some other, say the actual, probabilities. As the volatility
is the same under actual and risk-neutral probabilities for any equivalent
change of measure (and a Brownian ltration) the rightmost equality holds
for model parameters. The market price of risk provides an interesting an-
swer to the classical question whether the slope of the yield curve reects
investors attitudes towards risk or expectations about future short rates.
For each market price of risk we get a dierent curve of expected future
short rates. Thus, any given term structure can reect several dierent
expected future courses of short rates depending on , the individual in-
vestors attitude towards risk. Even more varying expectations about the
future could be consistent with the given yield curve if limited information
is taken into account. By assumption, all information in these models were
asumed to be public and known to all market participants. In particular, all
investors were assumed to know the true value of volatilities, which clearly
is an unrealistic assumption.
In the single factor Vasicek model the dierence between any (discrete)
forward rate and the expected future short rate (the term premia) is a
function of the maturity and model parameters
_
c +
c
2
4a
2

Tt
_

Tt
9 Extended Vasicek
Hull and White (1990) showed how to match the initial prices in ane
models to any given initial term structure of bond prices by choosing the
parameter b as a particular function of time. However, their approach was
not very transparent so we choose a dierent route here. In view of our
description of arbitrage-free bond prices we can get the same result more
13
quickly from the conditions (20) and (21). Simply plug in the bond price
volatility for the specic case considered by Vasicek in the arbitrage-free
bond price formula (7). This will automatically result in the matching drift
in prices! The volatility coecients are
(s, T) (s, t) = c
Ts
c
ts
=
Tt
ce
a(ts)
i e the bond price volatility equals the ever appearing function
Tt
times
the short rate volatility. Furthermore the bond price drift in (7) becomes

1
2
_
(s, T)
2
(s, t)
2
_
=
c
2
2a
2
_
e
2a(Tt)
e
2a(ts)
2(e
a(Tt)
e
a(ts)
)
_
where
_
t
0
(s, t)
2
(s, t)ds =
c
2
a
_
t
0
e
a(ts)
(1 e
a(ts)
) ds
Plugging these into the arbitrage-free volatility bond pricing formula (7)
gives a result that does not look very familiar. To simplify the formula we
use a little trick. Using the Vasicek volatilities the short rate process (8)
can be written
r(t) f
0,t
=
c
2
a
_
t
0
[e
a(ts)
e
2a(ts)
]ds c
_
t
0
e
a(ts)
dW

s
As the bond price volatility equals the short rate volatility times the factor

Tt
we can replace the stochastic dW

s
-term in the formula with

Tt
(r(t) f
0,t
)
Inserting into the arbitrage-free bond pricing formula (7) and using standard
algebra we get a recognizable expression
B(t, T) =
B
0,T
B
0,t
exp
_

Tt
(r(t) f
0,t
)
c
2
2
1 e
2at
2a

2
Tt
_
(31)
This is the arbitrage-free price in the Vasicek model that matches the initial
term structure, cp Hull and White (1994, p.9). For this so called extended
Vasicek model the traditional duration and convexity are replaced by
Tt
and
2
Tt
. The reason is that mean-reversion in the short rate dampens its
local variance c
2
(T t) over time in a rather special way
V ar

[r(T) | F
t
] =
c
2
2
1 e
2a(Tt)
a
<
c
2
2
1
a
14
This approach to the extended Vasicek model is simpler than the pi-
oneering approach in Hull and White (1990). They identify terms in the
dr-form not in the integrated form r(t). That approach is more compli-
cated. But by making b time dependent in a special way one can t the
initial term structure anyway. The trick is that when integrating one term
cancels out.
Hull and White start with
dr(t) = (ar(t) +b(t))dt +c dW
t
and specify the time dependence as
b(t) =
t
f
0,t
+af
0,t
+
c
2
2
1 e
2at
a
Note that the rightmost term is V ar

[r(t)]. The two other terms are the


initial forward rate and its derivative (sic!) corresponding to the given inital
term structure of bond prices. This does not look very intuitive. However,
we are now in a position to explain their magic trick. As we are still dealing
with the same process as in (22) the solution too is the same. Indeed,
integrating the drift term in the (well-known) solution (23) we get
_
t
0
e
as
b(s)ds =
_
t
0
_
e
as
(
s
f
0,s
+af
0,s
) +
c
2
2a
(e
as
e
as
)
_
ds =
=
_
t
0

s
(e
as
f
0,s
) ds +
c
2
2a
2
(e
at
1 +e
at
1) =
= e
at
f
0,t
f(0, 0) +
c
2
2a
2
(e
at
+e
at
2)
Inserting these values in the known solution (23) starting from r
0
the drift
in (24) becomes
E

[r(t)] = e
at
r
0
+f
0,t
e
at
f(0, 0) +
c
2
2a
2

_
1 +e
2at
2e
at
_
=
= f
0,t
+
c
2
2

_
1 e
at
_
2
where we have used the equality r(0) = f
0,0
. As a result two terms cancel
out very conveniently. Integrating over the time to maturity we get the same
expression for the bond price as before, i e formula (31).
10 Stacked Factor Models
After describing the simple analytics of the one-factor case we now proceed to
the general case of several factors. Most of the multifactor models are based
on the single factor models of either Cox, Ingersoll, Ross (1985) or Vasicek
15
(1977). In the rst type of model the distribution of future interest rates is
a non-central chi-square distribution, while in the latter interest rates follow
an Ornstein-Uhlenbeck process and become normally distributed. While
Due and Kan (1995), and Jamshidian (1995) use the chi-square distribu-
tion, El Karoui and Lacoste (1992) analyse models based upon the normal
distribution, and for a bit of both see Rogers (1994). The common feature
in these multifactor models is that yields are ane functions of a set of
stochastic factors.
The continuously compounded yield to maturity at time t for a pure
discount bond maturing at time T is dened as
y(t, T) =
1
T t
ln B(t, T)
Assume that each yield is an ane combination of some uncorrelated stochas-
tic factors X
k
, k = 1, 2, ..., d
y(t, T) = b
T
+
d

k=1
a
Tk
X
k
(t) (32)
where the coecients a
T1
, a
T2
, ..., a
Td
, and b
T
are constant over time but
may dier among maturities. We assume that all coecients and factors
are positive. Thus the maturity dependent constant b
T
may be interpreted
as the minimum level for each yield (with respect to the risk-neutral prob-
ability measure). The formula relates the current value of the factors to
the yields. Thus yields satisfy the Markov property. Typically we cannot
directly observe the uncorrelated factors. We also dene the short term rate
of interest as the yield of almost instant maturity
r(t) = lim
h0
y(t, t +h) = b
0
+
d

k=1
a
0k
X
k
(t) (33)
It is important to note that as the discount bonds decrease with time to
maturity on any date t they must all be ane functions of the same factors.
According to the basic theorem of no arbitrage all current prices equal
the expected value of their future pay-os discounted at the short rate r(t).
For a pure discount bond maturing at time T this means
B(t, T) = E

_
exp
_
T
t
r(u)du | F
t
_
16
where the conditional expectation (as before) is taken with respect to the
risk-neutral probability measure. Evaluating the integral of future short
rates using (33) we get
_
T
t
r(u)du = b
0
(T t) +
d

k=1
_
T
t
a
0k
X
k
(u)du
And as the factors are uncorrelated by assumption we have
B(t, T) = e
b
0
(Tt)

k=1
E

_
exp
_

_
T
t
a
0k
X
k
(u)du
_
| F
t
_
(34)
Thus, the arbitrage-free bond price equals the product of a series of ex-
pected values, one for each factor. Obviously this is just the probabilistic
counterpart to a separation of variables.
The product formula is very practical even if we cannot observe the
uncorrelated factors X
k
(t) directly. We are free to choose how to express
the factor model in terms of observable variables. This is also known as
choosing a measurement model in statistics.
For d = 2 choose two yields, y(t, t + 1) = z
1
, and y(t, t + 2) = z
2
say, as
observable variables. Rewriting (32) we get
_
y
1
y
2
_
=
_
a
11
a
12
a
21
a
22
__
X
1
X
2
_
+
_
b
1
b
2
_
or more compactly
y = AX +b
Solving for X (provided A is non-singular) in terms of the observable vari-
ables chosen here we get
X = A
1
(y b)
Note that the observable variables are correlated although the factors are
not. For more information see Pang and Hodges (1995) and El Karoui and
Lacoste (1992, Appendix).
Another example on the choice of observable variables was provided by
Longsta and Schwartz (1992). They choose to observe the spot rate r and
its conditional variance V . This is a very cleaver choice because
V = V ar(r) = V ar(
2

k=1
a
0k
X
k
) =
2

k=1
a
2
0k
V ar(X
k
)
17
Thus, the number of parameters is minimal and the transformation is given
by
_
r
V
_
=
_
1 1
a
01
a
02
__
a
01
X
1
a
02
X
2
_
Obviously this simplication will work for any yield and its volatility. But
the identication of parameters will not be as simple as before. When the
rate is a yield of xed maturity we have to invert the bond pricing formula
(34) to nd the parameters.
11 Conclusions
As we have seen the decisive drift terms referred to in the title of this paper
were: the discounted price is a martingale so it must have a zero drift, the
common drift in bond prices can only be one, the drift in the short rate
itself can be solved for, the deterministic drift in ane factor models has a
decisive impact on stochastic bond prices, and the implied forward prices
consititute the deterministic drift system for the arbitrage-free approach.
Also, we showed, in a simple case, how to adjust expectations about
future short rates in order to take into account risk premia in the term
structure of bond prices. Since expectations depend on the individual in-
vestors attitude towards risk each individual may have his own opinion of
the alleged drift in future short rates despite the fact that they all observe
the same term structure of bond prices.
Although most of our calculations were performed for the simplest one-
factor case similar calculations can be performed for several factors using
the standard product formula
E[X Y ] = E[X] E[Y ] + Cov(X, Y )
Indeed there are many ways to construct a measurement model of observable
variables from a model of several uncorrelated factors.
References
[1] Artzner, P, and F.Delbaen, 1989, Term structure of interest rates: the
martingale approach, Advances in Applied Mathematics 10:95-129.
[2] Babbs, S., 1990, A family of Ito process models for the term structure
of interest rates, University of Warwick, preprint 90/24.
18
[3] Bajeux, I. and R.Portait, 1993, The martingale approach in a state
variable framework, and the derivation of closed-form solutions of a
multifactor model of the yield curve, working paper, ESSEC, no 93014.
[4] Brennan, M. and E.S. Schwartz, 1982, An Equilibrium Model of Bond
Pricing and a Test of Market Eciency, Jour. Fin. Quant. An. 17,
301-329.
[5] Cohn, D. L., Measure theory, Birkhauser, Boston 1980.
[6] Cox, J. C., J. E. Ingersoll, Jr., and S. A. Ross. A Theory of the Term
Structure of Interest Rates. Econometrica, 53 ( March 1985), pp. 385-
408.
[7] Due, D. and R. Kan. A Yield-Factor Model of Interest Rates. Work-
ing Paper, 1992, revised vers Jan 1995.
[8] Dybvig, P. Inecient Dynamic Portfolio Strategies, or How To Throw
Away a Million Dollars. Review of Financial Studies 1, 1988, pp. 67-
88.
[9] El Karoui, N. and H. Geman, 1991, Valuation of Floating-Rate In-
struments: The Eciency of a Probabilistic Approach, unpublished,
University of Paris.
[10] El Karoui, N. and V. Lacoste, 1992, Multifactor Models of the Term
Structure of Interest Rates. Working Paper, CERESSEC, Cergy Pon-
toise, France.
[11] Flesaker, B. and L. P. Houghston, 1995, Positive Interest, Merrill
Lynch preprint.
[12] Grimmett, G.R., and D.R. Stirzaker, Probability and Random Pro-
cesses. Clarendon Press, Oxford, 2nd ed., 1992.
[13] Heath, D., Jarrow, R. and A.Morton, 1992, Bond Pricing and the Term
Structure of Interest Rates, Econometrica 60:77-105.
[14] Ho, T.S.Y. and S.-B. Lee, 1986,Term Structure Movements and Pric-
ing Interest Rate Contingent Claims, Jour.Fin 41, 1011-1028.
[15] Hull, J. and A.White, 1990, Pricing Interest-Rate Derivative Securi-
ties, Rev.Fin.Stud. 3, 573-592.
19
[16] Hull, J. and A.White,1994,Numerical Procedures for Implementing
Term Structure Models I: Single Factor Models, Jour.of Der., Fall,
7-16.
[17] Jamshidian, F. A Simple Class of Square-Root Interest Rate Models.
Applied Mathematical Finance, March 1996, pp. .
[18] Karatzas, I. and S. Shreve, 1988. Brownian Motion and Stochastic Cal-
culus. New York Springer-Verlag.
[19] Karlin, S. and H. M. Taylor, 1975. A First Course in Stochastic Pro-
cesses (2nd edn). Academic Press, New York.
[20] Longsta, F., and E. S. Schwartz. Interest Rate Volatility and the
Term Structure: a Two-Factor General Equilibrium Model. Journal
of Finance, 47, 1992, pp.1259-1282.
[21] Pang, K., and S. Hodges. Non-Negative Ane Yield Models of the
Term Structure. Preprint 95/62, 1995, Financial Options Research
Centre, University of Warwick.
[22] Rogers, C. Which Model for Term Structure of Interest Rates should
one use? Proc. IMA workshop on math.nance, (D.Due and
S.Shreve eds) Springer Verlag,1994, pp. 93-116.
[23] Rogers, L.C.G., and D.Williams, Diusions, Markov Processes, and
Martingales, Vol. 2, Wiley, Chichester 1987.
[24] Vasicek, O. An Equilibrium Characterization of the Term Structure.
Journal of Financial Economics, 5, 1977, pp.177-188.
20

Anda mungkin juga menyukai