FIN511 Notes: 1 Mathematical Base of Continuous Time Stochastic Methods
FIN511 Notes: 1 Mathematical Base of Continuous Time Stochastic Methods
FIN511 Notes: 1 Mathematical Base of Continuous Time Stochastic Methods
Steven Jens
March 20, 2010
This course will concentrate on models in continuous time. Usually an economic model in continuous time
that cant be replicated in discrete time is wrong, but sometimes the continuous model is more tractable.
1.1
Basic Concepts
where
(1)
(
1 if Q() = t
Z(t, ) =
0 otherwise
(2)
(3)
for any open I T . Every stochastic process is equivalent to some separable process.
X(t, ) is continuous if the probability of a continuous sample path is 1.
For a martingale, Es [X(t)] = X(s). For a submartingale, Es [X(t)] X(s) i.e., the current state
is sub. A convex function of a martingale is a submartingale by Jensens inequality. By the Markov
inequality, for separable submartingale X and > 0,
P r (sup X(t) > )
where X + := max{0, X}.
E[X + (T )]
(4)
Steven Jens
FIN511 Notes
Markov process: past and future are independent conditional on the present. A time-homogeneous
Markov process does not depend on absolute time, but may depend on relative time.
Example:
u>t
(5)
A
x r
r
z
-
Figure 1: Chapman-Kolmogorov
A Wiener process can be defined equivalently in any of the following four ways:
a continuous process of independent, normally-distributed increments with mean and variance of an
increment proportional to the time length of the increment
a continuous process such that, for some constants > 0 and and for any 0 s < t,
1. Es [X(t) X(s)] = (t s) and
2. Es {(X(t) X(s) (t s))2 } = 2 (t s)
the limit of a sequence of binomial processes consistent with the two conditions in the previous
definition
the continuous part of any process with independent increments if the distribution of X(t) X(s)
depends only on t s
Brownian motion refers to the standard Wiener process in which = 0 and = 1.
Steven Jens
FIN511 Notes
|wt+1 wt | = ).
P r(lim0 S(s, t, ) = 2 (ts)) = 1 for any 0 s t, where S(s, t, ) is the sum of squared changes
over [s, t] with changes measured between adjacent partition points and > 0 is the maximum
distance between points in the partition.
1.2
Stochastic integrals
(7)
Case 1:
f (t, ) = 1 t, =
Case 2:
f (t, ) = X() t =
dW (t, ) = T
RT
0
Case 3:
Z T
a() 0 t < t1
f (t, )dW (t, ) = a()[W (t1 ) W (0)] + b()[W (t2 ) W (t1 )] + . . .
f (t, ) = b() t1 t < t2 =
Defining
Z
X(t) :=
f (s)dw(s),
X(0) = 0
(8)
s=0
RT
where E{ s=0 f 2 (s)ds} < , (which is a sufficient condition for the convergence of a stochastic integral),
E{X(t)} = 0 t [0, T ]
X is a martingale
Steven Jens
FIN511 Notes
Rt
E{X 2 (t)} = E{ s=0 f 2 (s)ds}
almost all sample paths of X are continuous
X is not differentiable
PN 1
limN i=0 X T (i+1)
X
N
is infinite.
Ti
N
= ; note that if X is a portfolios holdings, trading volume
(9)
It
os formula
If X is an It
o process with dXt = t dt + t dBt and f : <2 < is twice continuously differentiable, then
Y := f (Xt , t) is an It
o process with
1
1
dYt = ft dt + fx dx + fxx (dx)2 + ftt (dt)2 + fxt (dx)(dt) + higher order terms
2
2
(10)
or
or
1
dYt = ft dt + fx dx + fxx (dx)2
2
(11)
1
dYt = ft dt + fx t dt + fx t dBt + fxx (t dt + t dBt )2
2
(12)
1
1
dYt = ft dt + fx t dt + fx t dBt + fxx 2t (dt)2 + fxx t2 (dBt )2 + fxx t t (dt)(dBt )
2
2
This can be simplified using the It
o multiplication rules:
(13)
(dt)2 = 0
(dB)(dt) = 0
(dB)2 = dt
1
dYt = ft dt + fx t dt + fx t dBt + fxx t2 dt
2
(14)
1
dYt = fx (Xt , t)t + ft (Xt , t) + fxx (Xt , t)t2 dt + fx (Xt , t)t dBt
2
(15)
Steven Jens
FIN511 Notes
(16)
(17)
(18)
(19)
(20)
Diffusion processes
A diffusion process is a Markov process with almost surely continuous sample paths.
For diffusion process X(t),
1
E [X(t + t) X(t) | X(t)] = (X(t), t)
t
(21)
1
E (X(t + t) X(t))2 | X(t) = 2 (X(t), t)
t
(22)
lim
t0+
lim +
t0
Steven Jens
FIN511 Notes
Well start with a hybrid discrete-time/continuous-time setup all trades and dividend payments are made
at t0 , t1 , t2 , . . ., but we use continuous time stochastic processes. (Were following the Portfolio Dynamics
handout.)
Dramatis Personae:
X(ti ) is a vector of positions exiting trading date ti . Because of the discrete-time trading restriction,
X(t) is right-continuous.
Well have a riskless loan balance L(t), which is only the principal part of the loan balance, and,
like X, only changes at trading dates.
P (t) N 1 vector of prices at t (ex-dividend if a dividend is paid at time t).
r(tn ) is the riskless rate over (tn , tn+1 ]
D(t) is a right-continuous N 1 process of cumulative dividends paid up to and including t. Dividends
are paid to shareholders entering the period in which they are paid.
C(t) is a right-continuous cash balance.
Finally, well use d as a forward-differencing operator; e.g., dC(tn ) := C(tn+1 ) C(tn ), which strikes
me as backward, but Im sure John Long knows what hes doing.
dC(tn ) = C(tn+1 ) C(tn ) = X(tn )0 dD(tn ) + r(tn )L(tn )dtn dX(tn )0 [P (tn ) + dP (tn )] dL(tn )
(23)
Portfolio value
V (t) = X(t)0 P (t) + L(t)
(24)
and
dV (tn ) = V (tn+1 ) V (tn )
(25)
(26)
(27)
Now consider what happens as dt 0+ . If P + D is an Ito process then V + C will be an Ito process
and stochastic differentials dV and dC will be related to X, L, etc. in the same way they were in the
discrete case.
An It
o process is not in general a Markov process, but well assume that P (t) is a diffusion process:
dP (t) = (P, t)dt + 1/2 (P, t)dW (t)
(28)
i.e.,
Z
P (t) P (t0 ) =
(P (s), s)ds +
t0
(29)
t0
where is N 1, 1/2 is N N , and dW is N 1. Note that and are deterministic functions all
randomness in equation 28 comes from dW .
6
Steven Jens
FIN511 Notes
A note on the 1/2 if is positive semi-definite, which covariance matrices are, it has a unique
positive semi-definite square root. This is not the Cholesky decomposition, which is transposed when
multiplied by itself to get . This 1/2 can be found from by diagonalizing into EE 0 , where
E 0 = E 1 and E are orthonormal matrices of eigenvectors of and is a diagonal matrix of eigenvalues.
Then
= EE 0 = (E1/2 E 0 )(E1/2 E 0 ) = 1/2 1/2
(30)
Cumulative dividends should be nondecreasing, which means D cant be a diffusion process. It is an
It
o process, so it almost surely has a continuous sample path, meaning no lump sums. Well assume the
form
dD = (P, t)dt
(31)
where (P, t) is a vector of nonnegative dividend rates per unit time. I.e., dividends are deterministic
conditional on P and t.
Suppose that V is a function of P, t with a continuous first derivative with respect to t and continuous
second derivatives with respect to P . Then Itos formula applies to V (P, t):
1
(32)
dV = Vt + tr(VP P ) dt + VP0 dP
2
Taking
dC = (P, t)dt + (P, t)dP
(33)
(34)
The right hand sides of equations 32 and 34 have to be equal; the unique decomposition principle
says that if they are equal with probability 1, the dt coefficients have to be equal and the dP coefficients
have to be equal.
1
Vt + tr(VP P ) = X 0 + rL
2
VP = X
0
V =X P +L
(35)
(36)
(37)
where equation 37 is equation 24. Equation 36 is really N equations, so we have a total of N +2 equations.
Market characteristics r, , and , are all continuous functions of P and t.
We have 2N +3 unknowns: N each of X and and L, V , and . So the market is not fully characterized
by N + 2 equations. Supposing a self-financing portfolio, i.e., = = 0, would suffice to identify the
other variables. Specifying a strategy determining X and L would also suffice.
Using equations 36 and 37 to eliminate X and L from equation 35 leaves
1
Vt + (rP 0 + 0 )VP + tr(VP P ) rV = (P, t) (P, t)0 (P r )
2
(38)
If we specify , , , r, and , we can find V only with a boundary condition of the form
V (P, T ) = g(P )
7
(39)
Steven Jens
FIN511 Notes
One variant of this is option pricing. Black-Scholes[1] depends on an arbitrage argument. The
continuous-time portfolio insurance version of this argument that was popular in the 80s doesnt quite
work in real life, as many people learned in October of 1987. The argument does work from a buy-and-hold
point of view, but its not risk-free in the first problem set, we used a law of large numbers argument to
find a deterministic sum of squared changes of a Brownian process in the limit as t approaches 0. But
that only works in continuous time. Mertons 1977 piece in the JFE[11] explains this.
Brennan and Schwartz (1988)[2] use a variant of the model in which there are no dividends and stock
prices are log-normal, i.e.,
dP = P dt + P dW
(40)
with constant and . r is assumed constant. The portfolio is assumed to be self-financing and pathindependent, with value V (P, t). Note that equations 36 and 37 in our model imply that X and L are
path-independent.
Our assumptions lead to the differential equation
1
(41)
Vt + rP VP + 2 P 2 VP P rV = 0
2
Note that doesnt show up; its relationship to r should be fixed by the market.
We restrict attention to time-invariant strategies, i.e., those in which leverage ( XP
V ) depends only
on P . So PVVP also depends only on P . This implies that
ln(V )
VP
f (P )
=
=
P
V
P
and
Z
ln(V ) =
P0
f (u)
du + K(t)
u
(42)
(43)
and ultimately
V (P, t) = k(t)g(P )
(44)
If V is of that form and satisfies differential equation 41, then it must have the form
V (P, t) = et (c1 P 1 + c2 P 2 )
(45)
where c1 , c2 , and are free parameters and 1 and 2 are determined by (formulae are on p288 of
Brennan and Schwartz 1988).
For = 0, 1 = 1 and 2 = 2r
2 . c1 = 1, c2 = 0 is buy-and-hold.
The diagram handed out shows V with respect to P for several values of c1 and c2 where c1 + c2 = 1.
Long assigned this paper in large part just to illustrate the mechanics of using equations 3537.
2.0.3
It
o Identities
Rt
0
X(s)ds
(46)
(47)
dy = Xdt
(48)
dZ = Xe dt + 0
(49)
dZ = XZdt
(50)
Steven Jens
FIN511 Notes
Y /X
Y
X
X
XY
2.1
(51)
(52)
Feynman-Kac
(53)
Rt
0
r(s)ds
F (t)A(t) = Et [F (T )A(T )] + Et
then
A(s)h(s)ds
(54)
(55)
with
f (x, T ) = g(x)
(56)
1
Df := ft + fx (x, t) + tr [(x, t)(x, t)0 fxx (x, t)]
2
(57)
then
"Z
f (x, t) = Ex,t
(58)
where
t,s := e
Rs
t
r(Xs ,s)ds
(59)
Alternatively, from Duffie p96 or so: Feynman-Kac says that, given the appropriate conditions,
payoff g(ST ) at time T has price process Yt = C(St , t) where
h RT
i
x,t
C(x, t) = E e t r(Zs )ds g(ZTx,t )
(60)
where
Zsx,t = x for s t
dZsx,t
r(Zsx,t , s)Zsx,t ds
(61)
+
(Zsx,t , s)dW
(62)
Steven Jens
John Long:
FIN511 Notes
we have a smooth function F (x, t) (x <n ) and an Ito process X with stochastic differential
dX(t) = (t)dt + S(t)dW (t)
(63)
RT
RT
where S is n m and dW is m 1 and 0 2 ds and 0 SS 0 ds are bounded.
It
os formula produces
1
0
0
0
dF (X(t), t) = Ft + FX (t) + tr (FXX SS ) dt + FX
SdW (t)
2
i.e.
(64)
Z t
Z t
1
0
0
0
Ft + FX (s) + tr (FXX SS ) ds +
FX
S(s)dW (s)
F (X(t), t) F (X(0), 0) =
2
0
0
(65)
1
0
Ft + FX
(s) + tr(FXX SS 0 ) ds
2
(66)
A special case, important in finance, is that X is a diffusion process, i.e. and S are functions solely
of X(t) and t. In this case, equation 66 is a partial differential equation where and S are given and F
is to be solved for.
Define
G(x, y, t) := F (x, t)ey
(67)
where y is a scalar It
o process such that
dy := a(t)dt
(68)
(69)
The It
o multiplication rules will produce dF dy = 0, so
1
0
0
SdW ) + F adt
dG = ey ((Ft + FX
(t) + tr(FXX SS 0 ))dt + FX
2
1
0
0
= ey (Ft + FX
(t) + tr(FXX SS 0 ) + aF )dt + FX
SdW
2
(70)
(71)
1
E0 [G(X(t), y(t), t)] G0 = E0
Ft +
+ tr(FXX SS 0 ) + aF
2
0
Rt
Suppose (looking back at equation 38) G = V (P , t)ey and y = 0 r(s)ds with
0
FX
(s)
dP := P rdt + SdW
10
e ds
y
(72)
(73)
Steven Jens
FIN511 Notes
(74)
If we are given
V (P (T ), T ) = g(P (T ))
then
"Z
V0 = E0
(75)
#
Rs
r(u)du
i
h
RT
+ E0 g(P (T ))e 0 r(s)ds
(76)
)t
(77)
(78)
is a martingale, i.e.,
h
i
2
2
Et Y (T )e( )T = Y (t)e( )t
2
e( )(tT )
1
=
Et
S(T )
S(t)
(79)
(80)
r(T t)
Et
S(t)
#
(81)
= er(T t)
=
2.2
e(
e(r )(tT )
S(t)
(82)
2r)(T t)
S(t)
(83)
Pricing Kernels
A profit opportunity is a portfolio constructed to hedge the uncertain dW component of returns and
make positive money from the dt component.
One way of characterizing a lack of profit opportunities is to say that there is a positive pricing kernel
which produces a martingale when multiplied by any self-financing portfolio.
Consider a portfolio strategy X(t), L(t) where X is n 1 and X 0 (t)P (t) + L(t) = 0.
We have, as always, accounting identity
dV (t) = X 0 (t)[dP (t) + dD(t)] + rL(t)dt dC(t)
11
(84)
Steven Jens
FIN511 Notes
(85)
(86)
and, because dV = 0,
Let Y (t) be a vector of X and P multiplied element-by-element. Then
dP + dD
0
dC(t) = Y (t)
r1dt
P
(87)
where 1 is a vector of ones, in this case of dimension n 1, and dP/P and dD/P are similarly sloppy but
intuitive element-by-element divisions.
Suppose
dP
= (t)dt (t)dt + S(t)dW (t)
(88)
P
where S has rank n m. Then
dC(t) = Y 0 (t) [(t)dt + S(t)dW r1dt]
0
(89)
(90)
Y 0 ( 1r) > 0
and
(91)
(92)
dZ
= r(t)dt 0 (t)dW
Z
and
12
(93)
Steven Jens
FIN511 Notes
Rt
0
r(s)ds
dA
A
dZ
Z
(94)
= (r(t)dt) (r(t)dt + 0 dW ) + 0
0
= dW
(95)
(96)
dP
P
dZ
Z
(97)
= [ + 1r S]dt + (S + 10 )dW
(98)
= (S + 10 )dW
(99)
Z
2 Z
1 0
= r dt dW
2
Z(t) = Z(0)e
Rt
0
r(s)ds 21
Rt
0
(100)
(101)
0 ds
Rt
0
0 (s)dW
(102)
We can use the pricing kernel to price forward contracts. If Z is a pricing kernel, Pi Z is a martingale,
so Pi (t)Z(t) = Et [Pi (T )Z(T )]; the fair price of
Pi (0) =
E0 [Pi (T )Z(T )]
Z(0)
(103)
similarly,
A(t)Z(t) = Et [A(T )Z(T )]
2.3
(104)
The pricing kernel is one characterization of a no-profit-opportunity situation. Another (equivalent) way
is with a risk neutral probability metric, which can be derived from a pricing kernel.
:= A(t)Z(t)
Z(t)
(105)
)/Z(t)
tT r(s)ds
tT r(s)ds Z(T )
Pi (t) = Et Pi (t)e
:= Et Pi (T )e
(106)
Z(t)
13
Steven Jens
FIN511 Notes
Now well dump the assumption of no dividends and return to the assumption of smooth dividends.
Then
dPi
= (i i )dt + i0 dW (t)
Pi
(107)
dZ
= rdt 0 dW
Z
(108)
recall
so
d(Pi Z) = dPi Z + Pi dZ + dPi dZ
= [Pi (i i )Z Pi rZ
(109)
Pi i0 Z]dt
+ something dW
(110)
(111)
(112)
(113)
"Z
Pi (t) = Et
t
2.4
#
Z(T )
Z(s)
dD(s) + Et Pi (T )
Z(t)
Z(t)
(114)
Bond Pricing
The zero-coupon yield curve acts much like an E[Z]; it summarizes the pricing of a profit-free market at
a current point in time. Given the yield curve and a set of deterministic cash flows, you can price the set
of cash flows.
Define
B(t, T ) to be the time t price of a default-free zero-coupon bond maturing at time T and A(t)
Rt
r(s)ds
0
to be e
. Suppose that E[dB/B] = r(t). Then
dB = Brdt + BdW
d
(115)
dA = Ardt
B
dB
B
dBdA
B
=
2 dA
+ 3 (dA)2
A
A
A
A2
A
dB
B
=
2 dA
A
A
1
= 2 ((ABrdt + ABdW ) BArdt + 0 0)
A
B(t)
=
dW
A(t)
14
(116)
(117)
(118)
(119)
(120)
Steven Jens
FIN511 Notes
(121)
(122)
=0
(123)
Define yield
y :=
ln B(t, T )
T t
(124)
we have
i
h RT
B(t, T ) = Et e t r(s)ds
e
(125)
Et [r(s)]ds
(126)
if r is a martingale, then
= er(t)(T t)
(127)
(128)
so y < r.
Z(T )
B(t, T ) = Et
Z(t)
#
"
)/A(T )
Z(T
= Et
Z(t)/A(t)
#
"
R
tT r(s)d(s) Z(T )
= Et e
Z(t)
(129)
(130)
(131)
Midterm
dX(t) = dt + tdW (t) where X(0) = 0
2
(132)
2
15
(133)
(134)
(135)
Steven Jens
FIN511 Notes
so,
t
s2 ds + 2
F (X(t), t) F (X(0), 0) =
(136)
0
t
s2 ds =
E0 [F (X(t), t) 0] =
0
1 3
t
3
(137)
(138)
(139)
Almost all sample paths of must be bounded and the integral of the square of S must be bounded. Also,
adapted measurable at each time and nonanticipating, i.e., the future cant be predicted from the
past (e.g., correlation of S and dW is zero).
V (t + ) V (t) = X 0 (t + )P (t + ) X 0 (t)P (t)
0
(140)
(141)
(142)
dV = X dP
(143)
(144)
So V is an It
o process if those terms are adapted, non-anticipating with respect to dW , bounded, etc.
Question 3: you have a single stock price P with log-normal stochastic differential
dP (t) = P (t)dt + P (t)dW (t)
(145)
V (P (t)) = P (t) 2
(146)
2 2
2
which works out.
2r
In this case, X(t) = VP = 2r2 P 2 1 and L = V XP = [1 +
the positions, XP
L , is constant.
16
2r2
2r
2 ]P
(147)
(148)
Steven Jens
FIN511 Notes
Question 4 is to show that no profit opportunities in a bond market implies that all bonds have the
same Sharpe ratio at any point in time.
dB(t, T )
= (t, T )dt + (t, T )dW (t)
B(t, T )
(149)
where is a scalar (i.e., this is a one-factor model), so all bond prices are perfectly correlated.
There are no profit opportunities, so there is a pricing kernel Z:
dZ(t)
= r(t)dt (t)dW (t)
Z(t)
dB
dZ
d(BZ)
=
+
+
BZ
B
Z
dB
B
dZ
Z
(150)
(151)
(152)
Since BZ is a martingale, r = 0, so
(t, T ) r(t)
= (t)
(t, T )
(153)
4
4.1
One of the simpler term structure models is the Vasicek model[12], based on long-term rates being based
on expectations of the future behavior of short term interest rates combined with risk preferences.
The short-term rate is a diffusion process i.e., a continuous Markov process:
dr(t) = (r(t), t)dt + (r(t), t)dW (t)
(154)
Z(T )
Z(t)
(155)
dZ
= rdt dW
Z
(156)
Rt
Rt
17
(157)
(158)
Steven Jens
FIN511 Notes
For constant (or just deterministic) , the first two terms are constant and the last is normal, so B is the
expected value of a lognormal.
In the real world, of course, this has too few degrees of freedom to come particularly close to fitting the
yield curve, and there are other state variables, such as inflation, that affect bond prices. Well consider
an N -dimensional state vector S(t), but retain the assumption that S(t) follows a Markov process.
As we learned back in subsubsection 1.2.2, a diffusion process, which we assume S to be, can be fully
determined by
dS(t) = (S(t), t)dt + (S(t), t)dW (t)
(159)
where is N 1, is N K, and dW is K 1.
A bond price will be B(S(t), t, T ); if the function B is smooth, we can use Itos formula:
dB
= [(S(t), t, T ) + 0 (S(t), t, T )dW ]
B
(160)
where
1
1
Bt + BS0 + tr(BSS 0 )
B
2
1
0 (S(t), t, T ) = (BS0 )
B
(S(t), t, T ) =
(161)
(162)
Obviously this is tedious and messy, but we can use it to find conditions for the absence of profit
opportunities.
If there is no profit opportunity, there is a K 1 vector such that
(S(t), t, T ) 1r(t) = 0 (S(t), t, T )(S(t), t)
I.e., expected excess rate of return on any bond is volatility times .
Equivalently,
B(S(t), t, T )Z(S(t), t) is a martingale T
(163)
(164)
where
dZ
= r(t)dt 0 (S(t), t)dW (t)
Z
Plugging equations 161 and 162 into 163 gives
1
Bt + BS0 ( ) + tr(BSS 0 ) rB = 0
2
18
(165)
(166)
(167)
Steven Jens
FIN511 Notes
Default-free zero-coupon unit bonds are pretty simple securities, but some complex models have been
developed for them. All are based on the generic notion that if there are no profit opportunities, theres
a positive pricing kernel that can be multiplied by prices to get a martingale.
If B(t, T )Z(t) is a martingale, then
Z(T )
(168)
B(t, T ) = Et
Z(t)
which means Z is sufficient for bond prices. Similarly, the yield curve provides sufficient information about
bond prices as well, which is obvious to extract in a certain environment and not necessarily as obvious
in a model with uncertainty.
Defining A as usual as the accumulated value of a unit time-zero investment invested continously
in short-term bonds, i.e.,
Rt
A(t) = e 0 r(s)ds
(169)
and noting that AZ is a martingale, we know that the dt coefficient of dZ/Z is r. Also, defining
dB(t, T )
0
=: B (t, T )dt + B
(t, T )dW (t)
B(t, T )
(170)
0
B (t, T ) r(t) = B
(t, T )(t)
(171)
no profits implies
0
(172)
from BZ
So much for review; lets look at Cox, Ingersoll, and Ross (1981)[3]. They examine whether several
term-structure models are consistent with a lack of profit opportunities.
The local expectations hypothesis is
B (t, T ) = r(t) T > t
(173)
i.e., nobody demands a risk premium for holding a long-term bond. This is consistent with = 0 and
Z = 1/A.
The return-to-maturity expectation hypothesis holds that
1
A(T )
= Et
(174)
B(t, T )
A(t)
We can analyze this by analyzing the expected change of both sides.
h
h RT
ii
1
1
= Et dEt e t r(s)ds = r(t)
dt
Et d
B
B(t, T )
The left-hand side can be evaluated with Itos formula to get
dB
2 (dB)2
1
0
Et
+
= [B (t, T ) + B
(t, T )B (t, T )] dt
B2
2 B3
B
(175)
(176)
So,
B (t, T ) r(t) = 0 (t, T )(t, T )
19
(177)
Steven Jens
FIN511 Notes
RT
t
Et [r(s)]ds
(178)
which is sort of like the return-to-maturity expectation hypothesis, but with the expectation brought into
the integral. Under this model,
f (t, T ) =
ln(B(t, T )) = Et [r(T )]
| {z }
T
(179)
forward rate
"Z
ln(B(t, T )) = Et
r(s)ds
(180)
2
1 dB
dB
B
2 B
"
2 #
1 dB
dB
Et [d ln B] = Et
B
2 B
d ln(B) =
(181)
(182)
"Z
##
dEt Et
= r(t)dt
(183)
1 0
(t, T )B (t, T )
2 B
(184)
r(s)ds
t
B (t, T ) r(t) =
For all three hypotheses,
0
(t, T ) r(t) = B
(t, T )B (t, T )
(185)
(186)
where
(187)
implies that B cant depend on T unless = 0. But as t T , should go to 0. But in general, we just
need
0
0
B
(t, T )B (t, T ) = B
(t, T )(t)
(188)
0
and in a multidimensional setting, we cant generally divide out the B
note that the product of nonzero
vectors can be zero.
Digression: if is a covariance matrix, then
|{z}
: |{z}
= 0
N N
(189)
N N
where the lengths of the columns of are volatilities and the cosines of the angles between them are
correlations.
20
Steven Jens
FIN511 Notes
1
2 (t).
(190)
has a solution (t) true for all T . We can then complete the square.
0 (t, T )(t, T ) 2 0 (t, T )(t) + 0 (t)(t) = 0 (t)(t)
(191)
(192)
Suppose and are two-dimensional. The set of (t, T )s that satisfy equation 192 is a circle with
radius equal to the length of (t) (figure 2) in more dimensions, it would be a sphere or hypersphere.
Note that the circular shape and the interpretation of cosines of angles as correlations means that
shorter durations imply lower volatilities imply shorter s imply high correlations.
'$
s
r(t)
&%
Figure 2: and
Any model that has a finite number of sources of variation (in this case two) has to have a lot of
redundancy if maturity of bonds in the market are continous, so that puts a lot of restrictions on the
correlation structure. If you limit yourself to as many bonds as you have dimensions, you can draw a
circle/sphere/hypersphere that goes through those points and the origin.
Theres more good stuff in Fisher and Gilles (1998)[5].
The affine model (or class of models) is based again on the assumption that long-term rates are based
on short-term rates and that the short-term rate is a Markov (diffusion) process:
dr(t) = r (r, t)dt + r0 (r, t)dW (t)
(193)
Recall
Z(t) = Z(0)e
=
Rt
0
Z(0) 1
e 2
A(t)
1
A(t)Z(t) = Z(0)e 2
r(s)ds 21
Rt
0
Rt
0
Rt
0
R
0 (s)(s)ds 0t (0 s)dW
0 (s)(s)ds
0 (s)(s)ds
21
Rt
0
Rt
0
0 (s)dW
0 (s)dW
(194)
(195)
(196)
Steven Jens
FIN511 Notes
is a martingale on the left, hence a martingale on the right. So the value of a self-financing portfolio
A(T )Z(T )
V (t) = Et V (T )
(197)
A(t)Z(t)
#
"
Z0
thing
A(T ) e
= Et V (T ) Z0 thing
(198)
A(t) e
i
h
RT 0
RT 0
1
(199)
= Et V (T )e 2 t (s)(s)ds t (s)dW
Girsanovs Theorem holds that under the risk neutral probability measure,
Z t
W (t) = W (t) +
(s)ds
(200)
is a standard Wiener process, where W (t) is an actual standard Wiener process. Alternatively,
(t)dt
dW = dW
(201)
(202)
(t)
dX = ((t) 0 (t)(t))dt + 0 (t)dW
(203)
So
can be rewritten
So, back to the single-factor affine model. Assuming r is a time-homogeneous diffusion process,
dr(t) = r (r(t))dt + r (r(t))dW (t)
(204)
(205)
(206)
Example: homework 5
dr(t) = f (r(t) r0 )dt + dW (t)
(207)
(208)
satisfying this, there is a risk-neutral measure under which the expected excess return on all bonds is 0.
By It
o,
dB
1
= (a0 ( ) b0 ( )r(t)) + b( )dr + b2 ( )(dr)2
B
2
1 2 2
0
0
= a ( ) b ( )r(t) + b ( ) + b( )f (r r0 ) dt + b( )dW (t)
2
22
(209)
(210)
Steven Jens
FIN511 Notes
(211)
1
a0 ( ) (b0 ( ) + 1)r + 2 b2 ( ) + b( )[f (r r0 ) ] = 0
(212)
2
f (r r0 ) is the risk-neutral expected growth rate of r. a0 ( ) (b0 ( ) + 1)r + 21 2 b2 ( ) is affine in
r, so f (r r0 ) has to be affine in r:
f (r r0 ) (r) =: 1 + 2 r
(213)
(214)
(215)
4.2
One of the downsides of the affine model is that fewer factors than maturities doesnt fit any set of
empirical data. One kludge is to make the Markov process nonhomogeneous (i.e., dependant on time).
The Heath-Jarrow-Morton model fits the whole current yield curve and projects the future based on
that.
One way of describing todays yield curve is in terms of forward short-term rates:
B(t, T ) = e
RT
t
f (t,s)ds
(217)
which is the same as our original definition of B from the deterministic case (e.g., equation 125) except
known future rates (r) are replaced with forward rates f .
ln B(t, T ) = f (t, T )
T
(218)
(219)
Steven Jens
FIN511 Notes
Presumably
lim f (t, s) = r(t)
st
(220)
The reason we need to related expected excess returns to volatility in a no-profit-opportunity setting
is that excess returns are free. Given our specification and the no-profit-opportunity condition, can we
get dB/B?
We can apply It
os formula to equation 217, but we have to be careful that our derivatives are valid.
The partial derivative with respect to time assumes the Ito processes f dont change, so the derivative is
f (t, t). The next term comes from the insight that the sum of a bunch of changes is the change in the
sum.
"Z
#2
Z T
T
1
dB
df (t, s)ds
(221)
= f (t, t)dt +
df (t, s)ds +
B
2 t
t
"Z
# "Z
#!
!
Z T
Z T
T
T
1
0
0
= r(t)dt
(t, s)ds
(t, s)ds
(t, s)ds
(t, s)ds dW
dt
(222)
2 t
t
t
t
No profit opportunities implies the existence of a (t) of the same dimension as such that expected
excess return is 0 for all bonds at all dates:
"Z
# "Z
#!
!
Z T
Z T
T
T
1
0
0
(t, s)ds
(t, s)ds
(t, s)ds
=
(t, s)ds (t)
T
(223)
2 t
t
t
t
This doesnt work for arbitrary and (e.g., scalar generally allows you to solve for T ).
"Z
# "Z
#
Z T
T
T
1
0 (t, s)ds
(t, s)ds = 0
T
((t, s) 0 (t, s)(t))ds
2 t
t
t
Since this is constant for all T , the derivative with respect to T is 0:
Z T
0
0
(t, s)ds = 0
((t, T ) (t, T )(t)) (t, T )
{z
}
|
t
(224)
(225)
So you pick a plausible or convenient for forward rates, replace the actual mean with that integral term
on the right, and simulate bond price paths. This isnt trivial in practice, and some volatility assumptions
are easier than others, but what you get is a simulator that agrees with current prices and simulates
possible future yield curves (or functions thereof) under the risk-neutral measure, which is useful for
pricing interest-rate derivatives.
The big advantage is that it agrees with current prices. The disadvantage is that its difficult or
impossible to put in a Markov framework.
This applies to any bond market with no profit opportunities. In particular, it ought to be consistent
with our affine model.
4.2.1
Generally (not quite always), we observe interest rates to be nonnegative. Most of our models so far have
included the possibility of negative interest rates.
24
Steven Jens
FIN511 Notes
i
h RT
B(t, T ) = EQ e t r(s)ds
(226)
EQ [r(T )ething ]
ln B(t, T ) =
T
EQ [ething ]
(227)
is a weighted average of possible r(T )s. Similarly, positive forward rates imply positive yields. Generally,
positive r implies positive interest rates by any reasonable measure.
Based on the fact that B(t, T )Z(t) is a martingale for fixed T ,
Z(T )
B(t, T ) = Et
(228)
Z(t)
This suggests that another way to simulate bond prices is to find an appropriate pricing kernel and simulate
that. Positive interest rates Z is a supermartingale.
Theorem:
m(t, s)ds
(229)
where Z( ) is a positive random variable and for each s [t, ], m(t, s) is positive martingale.
s 6
s=t
m(t, s2 )
m(t, s1 )
R
m(t, s)ds + Et [Z( )]
B(t, T ) = RT
m(t, s)ds + Et [Z( )]
t
(230)
(231)
Steven Jens
FIN511 Notes
The terms being summed in the numerator are a subset of those in the denominator, so the price is less
than 1.
R
The scale of the ms can be chosen arbitrarily; one option is 0 m(t, s)ds = 1.
You can simulate this just by simulating martingales, which is easier.
4.2.2
Homework 6
HJM is pretty general, assuming that individual forward rates follow continuous stochastic processes, but
not assuming anything Markovian. Given no profits, you can still relate growth rates to volatility (as we
did).
The affine model is a lot more specific. Given the right conditions (i.e., if the short-term rate has affine
volatility) under the risk-neutral distribution, we also get affine drift and an arbitrage-free bond-pricing
model.
Affine bond prices plus no profit opportunities imply
d
r(t) = (1 (t) + 2 (t)r(t))dt + (1 (t) + 2 (t)r(t))1/2 dW
(232)
ln B(r, t, T )
f (t, T ) =
T
= aT bT r
f :=
(233)
(234)
(235)
(236)
f (r,t,T ):=
f (r, t, T ) = f (r, t, T )
f (r, t, u)du
(237)
Derivative Assets
(238)
Rt
26
r(s)ds
(239)
Steven Jens
FIN511 Notes
note that A is stochastic so it cant be taken outside of the expectation operator, which means this equation
is not always convenient to use.
It can be helpful to consider the origin of this identity: subsection 2.3, page 13 NPO positive,
adapted, scalar pricing kernel process Z(t) such that Z(t)V (t) is a martingale for any self-financing
portfolio. In particular, Z(t)A(t) =: ZA (t) is a positive martingale. So
Z(t)V (t) = Et [Z(T )V (T )]
V (T )
V (t)
ZA (t) = Et
ZA (T )
A(t)
A(T )
"
#
V (t)
V (T ) ZA (T )
= Et
A(t)
A(T ) ZA (t)
(240)
(241)
(242)
so we can define
"
A
EQ
t [y(t)]
:= Et
ZA (T )
y(T )
ZA (t)
#
(243)
But we didnt have to choose A as our numeraire; we can define Z based on any always-positive (i.e.,
something we can divide by) self-financing portfolio. Suppose N is a long-term bond i.e., it goes to one
at time T . Then
V (t)
N
= EQ
(244)
t [V (T )]
N (t)
Unlike our A numeraire, we can use a known discount rate N (t) = B(t, T ). This probability measure is
called a T -forward measure.
Define F (t, T ) as the time t forward price of a fleem at time T and S(T ) as the spot price of the fleem
at time T . Then
QN S(T ) F (t, T )
0 = Et
(245)
N (T )
If we define N such that we know N (T ) to be 1 at time t i.e., B(t, T ) we get
N
F (t, T ) = EQ
t [S(T )]
(246)
h
i
RT
A
EQ
S(T ) exp( t r(s)ds)
t
F (t, T ) =
RT
Et [exp( t r(s)ds)]
(247)
If N = A,
B(t, T ) B(t, )
ln B(t, T ) = lim+
T
( T )B(t, T )
T
(248)
as t increases, ( T ) is constant and B(t, T )/B(t, T ) is constant (1). Under the T -forward measure,
B(t, )
B(t,T ) is a martingale. So under the T -forward measure, f (t, T ) is a martingale whose final value is r(T ).
27
Steven Jens
5.1
FIN511 Notes
Futures
Futures are set up with a clearinghouse so as to avoid counterparty risk by having the clearinghouse become
the counterparty and requiring collateral. The amount of the collateral changes over time as underlying
prices go up and down. So instead of a lump-sum settlement on the delivery date, theres effectively a
series of cash flows leading up to the delivery date. So a futures contract is not a self-financing portfolio.
One way to turn it into one is to balance the cash flows (the variation margin) with interest-bearing
securities and loans. You also increase the futures position each day by the interest rate. The futures
contract isnt self-financing, but the whole portfolio is.
dV (t) = rV (t)dt + A(t)df
(249)
d
V
A
=
1
[dV rV dt] = df
A
(250)
(251)
5.2
Stochastic Volatilities
(252)
(253)
28
Steven Jens
FIN511 Notes
What is usually meant by a stochastic volatility model these days is a multi-factor model in which
volatility is one of the factors. For example, from Heston (1993)[8].
dS = rSdt + vdWs
(254)
dV = k(
u v)dt + v vdWv
(255)
This can generate the volatility smile.
American-style put options can be solved numerically. Trying to solve them analytically is the kind
of problem that can suck your brain out. The optimal solution is to exercise the option if the value of
the underlier goes below a time-dependent boundary, where the boundary is below the strike price and
approaches the strike price as time approaches expiration. Solving the price for a given exercise strategy
is feasible, but calculating the optimal boundary is hard.
Duffie uses the term superreplicating portfolio for a self-financing portfolio that is always worth at
least the value of a particular option.
The nature of the replicating portfolio for an American put is that of a European put with the same
strike price and expiration date, plus a security that pays interest on the strike price iff the underlier price
is below the boundary. So the maximum value of the difference between the value of the American and
European puts is the value of interest paid on the strike price between now and expiration.
Almost any pricing model can be stated as the covaraiance of the returns being equal to something (see
ps 7). Sometimes it doesnt matter what numeraire is chosen, but sometimes its handy to choose one
whose value you will know in the future (e.g. a long-term bond) or a bank account numeraire (our A).
6.1
Control Theory
A classic control problem is the consumption and investment plan of a consumer. A control problem
has an objective function
Z T
(256)
J(S(t), t) := max Et
U (C(s), S(s), s)ds + F (S(T ))
C
{z
}
|
t
if T <
(257)
C(t) A(S(t), t)
(258)
29
Steven Jens
FIN511 Notes
Suppose t < < T and define C to be the optimal plan (the argmax of the control problem) and S
the corresponding states. Then
Z
U (C (s), S (s), s)ds + J(S ( ), )
J(S(t), t) = Et
(259)
Zt
U (C(s), S(s), s)ds + J(S( ), )
(260)
Et
t
(261)
Weve assumed S is an It
o process specifically, a controlled diffusion process. Now well assume that
J is twice continuously differentiable so Itos lemma applies:
1
0
0
dJ(S(t), t) = Jt + JS + tr(JSS ) dt + JS0 dW
(262)
2
or
Z
J(S( ), ) J(S(t), t) =
t
Combining, we get
Z
0 Et
t
Z
1
0
0
Jt + JS + tr (JSS ) ds +
JS0 dW (s)
2
t
1
0
[U + Jt + JS + tr(JSS )]ds
2
0
(263)
(264)
with equality for the optimal C. was chosen arbitrarily this equation holds for all [t, T ]. So, taking
a derivative with respect to ,
1
U (C(t), S(t), t) + Jt (S(t), t) + 0 (C(t), S(t), t)JS (S(t), t) + tr(JSS (S(t), t) 2 (C(t), S(t), t)) 0 (265)
2
again with equality for C ; this is the Bellman Equation.
Note that C doesnt appear in the derivatives of J. This is one of those times when people wave their
hands and say that in principle you can do this or that. In principle, you can take the derivative with
respect to C, set it to 0, and solve for C as a function of the derivatives of the value function J. But you
dont know what J is yet. If you plug back in the C, you have a nonlinear partial differential equation
of J with a natural boundary condition (F ), so you can find J. In practice, this isnt usually feasible, so
this is solved by guessing a form for J.
6.2
30
(266)
Steven Jens
FIN511 Notes
(267)
or
UC JW = 0
(269)
UCi
= JW
i
(270)
i.e., the marginal utility of spending money on each commodity equals the marginal utility of keeping
money.
First order condition with respect to X:
(R 1r)JW + R X JW W + R JW = 0
(271)
X =
JW
JW W
{z
}
1
( 1r)
| R {z
}
tangent portfolio
1
R
| R {z }
JW
JW W
(272)
Regression of on R
inverse risk-aversion
(273)
which John Long is referring to as a real bond it grows at a constant rate in terms of commodity i.
dPi (t)
di (t)
= i dt +
Pi (t)
i (t)
(274)
31
Steven Jens
FIN511 Notes
(275)
So the sum of the values of the hedges in the real bond scenario is
0 JW
=W
JW W
JW
W JW W
|
{z
}
(276)
inverse RRA
So, generally, more risk-averse consumers (RRA> 1) will hedge against consumption-side inflation and
less risk-averse consumers will provide these hedges.
If consumers have the same beliefs but different preferences, they differ in equation 272 only in the J
terms
In our set-up, state variables besides wealth were just commodity prices. There could, of course, be
others interest rates are not actually constant. Yield curves mostly move up and down, with slope and
curvature explain most of the rest of their movements. There seem to be significant time-series changes
in expected excess returns, too.
When dealing with highly auto-correlated time series (like interest rates), the nave expected volatility
(i.e., just differencing) is a lot different from the expected difference from what you can predict from the
time series.
General Equilibrium
We can find a general equilibrium by applying market-clearing conditions to the plans of the consumers.
You can also find an allocation that you know to be efficient and find prices that support that as an
equilibrium, but the other method is more in line with what were doing.
[10]
32
Steven Jens
FIN511 Notes
References
[1] Fischer Black and Myron Scholes. The pricing of options and corporate liabilities. The Journal of
Political Economy, 81(3):637654, 1973.
[2] Michael J. Brennan and Eduardo S. Schwartz. Time-invariant portfolio insurance strategies. The
Journal of Finance, 43(2):283299, 1988.
[3] John C. Cox, Jr. Ingersoll, Jonathan E., and Stephen A. Ross. A re-examination of traditional
hypotheses about the term structure of interest rates. The Journal of Finance, 36(4):769799, 1981.
[4] Darrell Duffie and Kenneth J. Singleton. Modeling term structures of defaultable bonds. The Review
of Financial Studies, 12(4):687720, 1999.
[5] Mark Fisher and Christian Gilles. Around and around: The expectations hypothesis. The Journal
of Finance, 53(1):365383, 1998.
[6] Bjorn Flesaker and Lane Hughston. Positive interest. Risk, 9(1):4649, January 1996.
[7] David Heath, Robert Jarrow, and Andrew Morton. Bond pricing and the term structure of interest
rates: A new methodology for contingent claims valuation. Econometrica, 60(1):77105, 1992.
[8] Steven L. Heston. A closed-form solution for options with stochastic volatility with applications to
bond and currency options. The Review of Financial Studies, 6(2):327343, 1993.
[9] Francis A. Longstaff and Eduardo S. Schwartz. Interest rate volatility and the term structure: A
two-factor general equilibrium model. The Journal of Finance, 47(4):12591282, 1992.
[10] Robert C. Merton. An intertemporal capital asset pricing model. Econometrica, 41(5):867887, 1973.
[11] Robert C. Merton. On the pricing of contingent claims and the modigliani-miller theorem. Journal
of Financial Economics, 5(2):241249, November 1977.
[12] Oldrich Vasicek. An equilibrium characterization of the term structure. Journal of Financial Economics, 5(2):177 188, 1977.
33