Upload
tu-shirota
View
225
Download
0
Embed Size (px)
Citation preview
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 1/17
7. Equivalent Martingale Measures
So far we have considered derivative asset
pricing exploiting PDEs implied by arbitrage-
free portfolios.
Another approach is to change the probabil-ity measure to another probability measure
implied by arbitrage-free markets such that
under that the (risk-free return discounted)
prices become martingales.
1
As for background, consider pricing an Euro-
pean call option.
The aim is to find the fair price for the option
given the available information.
To price the option (C t), we use the bestprediction of the end value in the light of
available information, such that
C t = Et [ρ max(S T − K, 0)],(1)
where Et
is the conditional expectation given
information up to time t, and ρ is a discount
factor.
2
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 2/17
The no arbitrage theory implies that if the
option is replicable, then the discount factor
will be the riskfree rate, and the probability
measure with respect to which the expec-
tation must be calculated is such that the
discounted price process
S t = e−r tS t,(2)
where r is the riskfree return, is martingale.
3
To illustrate the situation, consider the fol-
lowing single period discrete world.
Example 7.1: Suppose we have a call option C on
stock S and a bank account B. Let the exercise price
of the option be K , and assume that there are two
possible end values S 1 > K > S 2 of the stock.
So
S 1 = u S 0, with u > 1
S 0
S 2 = d S 0, with d < 1,
p
1 − p
where p is the probability that the price goes up to
S = u S 0, and S 0 is the current price of the stock.
Then the option with initial cost C has the end value
max
{S 1
−K, 0
}. To replicate this with the stock and
bank account with (riskfree) interest rate, r, we may
construct the following strategy:
4
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 3/17
Buy one share of the stock by financing it with cash
and S 2/(1+r) borrowed at rate r from the bank (after
one period the repayment is accordinly S 2). The value
of the initial position is then
S 0 − 1
1 + rS 2.(3)
The end value of the position is according to the stockprice as
S = S 1 S = S 2
Stock value S 1 S 2Loan repayment −S 2 −S 2Total payoff S 1 − S 2 0
5
We observe that in the case of S = S 2 the payoff is
0, the same as with the call options, and in the case
S = S 1 the total payoff is S 1 − S 2 = a(S 1 − K ), where
a = (S 1 − S 2)/(S 1 −K ). Thus in all, the payoff of the
strategy is exactly the same as the payoff of a call
options.
This implies that in the absence of arbitrage the cost
of the investment must be the same in both cases.
That is, buying a call options must have the same
value as the other strategy based on one stock and
bank loan.
So
aC = S 0 − 1
1 + rS 2
or
C = (S 0−
1
1 + r
S 2)/a = 1
1 + r
p∗(S 1−
K ),
where
p∗ = 1 + r − d
u − d .
6
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 4/17
We observe that 0 < p∗ < 1 (provided that 1 + r < u
Exercise: What is the rationale that this also holds?),
so that p∗ can be considered as a conditional proba-
bility given the initial price S 0 of the stock (the prob-
ability depends on u and d which are dependent on
S 0, i.e., how much the price should go up to reach
the given value S 1, or to decrease to go down to the
other possible given value S 2). These probabilities are
called risk neutral, hedging or martingale probabilities
or probability measures. The last name is because
they make the discounted price process S a martin-
gale.
7
This is seen as follows: We easily find that
(4)
S 0 = 1
1 + r ( p∗S 1 + (1 − p∗)S 2) = p∗S 1 + (1− p∗)S 2,
where S = S/(1 + r) is the discounted price pro-
cess. That is, the conditional expectations of the
discounted price process S given the information I 0 =
{S 0} is
E∗[S |I 0] = p∗S 1 + (1− p∗)S 2 = S 0 = S 0,(5)
so that S is martingale with respect to the risk neutral
probability measure, as stated above.
8
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 5/17
We observe that the option value does not at all de-
pend on the true probabilities p and q = 1− p!
However, if we write the above martingale equaton as
(q∗ = 1− p∗)
E∗[S |I 0] = p∗S 1 + q∗S 2 = p∗ p
pS 1 + q∗
q qS 2,(6)
where p∗/p and q∗/q can be considered kinds of like-
lihood ratios or odds ratios, judged by the markets
for the events that the stock price will be S 1 and S 2,
respectively.
So the market expected value of the future stock price
is a kind of likelihood weighted value of the possible
future outcomes.
9
Translations of Probabilities
Probability Measure
As an illustration, consider the probability
density f (z) of a standard normal distribu-
tion,
f (z) = 1√ 2π
e−12z2
.(7)
Then probability of that the random variable
Z is near a specific value z is
P
z − 1
2∆ < Z < z +
1
2∆
=
z+
12 ∆
z−12
∆
1√ 2π
e−12
z2
dz,(8)
which is a real number (between zero and
one).
Thus, the probability associates a real num-ber (in this case between zero and one) to
intervals on real line, or more generally to
(Borel) sets.
Such functions are called measures in math-
ematics or measure functions. 10
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 6/17
Because ∆ is small
z+12∆
z−12∆
1√ 2π
e−12z2
dz ≈ 1√ 2π
e−12z2
z+12∆
z−12∆
dz
= 1√
2πe−
12z2
∆.
(9)
For infinitesimal ∆, denoted as dz, we desig-
nate the associated measure by symbol dP (z),
or simply by dP .
Thus, in the above case we have
dP (z) = 1√
2πe−
12z2
dz.(10)
11
Generally, if P is a probability measure, we
have ∞−∞
dP = 1.(11)
With these notations, e.g.,
E[X ] = ∞
−∞x dP (x).(12)
So the expected value is mathematically an
integral with respect to probability measure.
dP is called sometimes the density of the
probability measure P .
12
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 7/17
Changing Probability Measure
Martingale model is a central tool for mod-
eling fair prices of derivative securities.
However, generally, if S t is a risky asset, then
given information up to time point t, we have
Et[S t+h] > (1 + rf )S t, (h > 0),(13)
because investors want some compensation
for the risk, where Et is the conditional ex-
pectation, and rf
is the risk-free rate.
13
However, we observed in the PDE approach
that under the arbitrage-free pricing the risk-
free rate should be a proper discounting fac-
tor in pricing risky derivative assets.
More importantly, the fundamental theorem
of asset pricing establishes the equivalence of the absence of arbitrage opportunity and ex-
istence of martingale measure in (the stochas-
tic model of) financial markets.
14
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 8/17
A probability measure, P , is a martingale mea-
sure for the discounted price process S t =
e−rtS t, if S t is martingale under P , i.e.
EP t [S s] = S t, whenever s ≥ t,(14)
where r is the riskfree rate, and EP t indicates
that the conditional expectation is taken withrespect to probability measure P .
This theory implies that the no-arbitrage price
of a contingent claim with underlying secu-
rity S t and (random) payoff X at maturity T
is obtained by
C t = EP t
e−rτ X
,(15)
where τ = T − t, and P is the martingale
measure for the discounted price process S t.
We say that S t is P -martingale.
15
Thus a martingale measure can be viewed
as a representation of the market’s current
opinion on the evolution of values of under-
lying assets and the prices of all derivatives
contingent to them.
Consequently, the knowledge of the martin-gale measure is all that is needed, in principle,
to value whatever derivative securities by the
formula of the form (15).
16
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 9/17
Then given a stock price process S t with
probability measure P , the goal is to find the
martingale measure P .
This can be accomplished if there is an in-
vertible function (one-to-one) ξ(z) such that
dP (z) = ξ(z) dP (z).(16)
Now recalling from standard calculus; if G(z) = g(z) dz then g(z) = dG(z)/dz = G(z), i.e.,
g(z) is the (mathematical) derivative of G(z).
Alternatively, we can write dG(z) = g(z)dz,
and hence
G(z) =
dG(z) =
g(z)dz.(17)
17
In the same manner, because
P (z) = z
−∞dP (t) =
z
−∞ξ(t) dP (t),(18)
so that we can adopt notation
dP (z)
dP (z) = ξ(z),(19)
and call ξ as a derivative of P with respect
to P .
In mathematical measure theory this is known
as the Radon-Nikodym derivative .
18
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 10/17
The existence of ξ is guaranteed if P and P
satisfy
P (dz) > 0 if and only if P (dz) > 0.(20)
Actually this condition guarantees besides the
existence of ξ (known as Radon-Nikodym The-
orem), also the equivalence of P and P in thesense
dP (z) = ξ(z) dP (z)(21)
and
dP (z) = ξ(z)−1dP (z).(22)
In this sense P and P are equivalent proba-
bility measures .
19
Example 7.2: (GBM) Consider the geometric Brown-
ian motion
dS t = µS t dt + σS t dW t, t ≥ 0(23)
so that
Z t = log (S t/S 0) =
µ− 1
2σ2
t + σ W t(24)
and
Z t ∼ N (µ∗t, σ2t),(25)
where µ∗ = µ − σ2/2.
Thus
dP (z) = 1√ 2π σ2t
e−(z−µ∗t)2
2σ2t dz.(26)
20
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 11/17
Let
Z t =
rf − 12
σ2
t + σ W t,(27)
where rf is a riskfree rate, and W t is another Wiener
process (the next example shows the relationship be-
tween W and W ).
Then
dP (z) = 1√ 2π σ2t
e−(z−r∗t)2
2σ2t dz,(28)
where r∗ = rf − 12
σ2.
Because the density function of the normal distribu-
tion is always positive, trivially
P (dz) > 0 ⇐⇒ P (dz) > 0.(29)
21
The transformation between these two probability mea-
sures is
ξ(z) = e−(µ∗−r∗)z−1
2(µ∗2−r∗2
)t
σ2 ,(30)
so that
dP (z) = ξ(z) dP (z).(31)
22
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 12/17
The Girsanov Theorem
The Radon-Nikodym theorem gives the con-
ditions under which the derivative ξ exist (i.e.,
that we can move to another probability mea-
sure without losing any information, which
means that those events that have positiveprobability have also positive probability un-
der the other measure).
The Girsanov Theorem provides the condi-
tions under which the Radon-Nikodym deriva-
tive exists for cases where Z t is a continuous
stochastic process.
23
For the purpose, let
{I t
}, t
∈[0, T ] be a family
of information sets (T < ∞). Define
ξt = e t
0 X u dW u−12
t0 X 2u du, t ∈ [0, T ],(32)
where X t is an I t-measurable process (that is
once I t is given, the value of X t is known),
and W t is a Wiener process with probabilitymeasure P .
24
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 13/17
It is assumed that X t does not increase too
fast, so that
E
e t
0 X u du
< ∞, t ∈ [0, T ],(33)
called Novikov condition (after a Russian math-
ematician).
Using Ito
dξt = ξtX t dW t,(34)
from which we immediately see that ξt is a
martingale, because W t is a Wiener process
and there is no drift component in (34).
25
This is also easy to see formally.
Obviously, from (32)
ξ0 = 1.(35)
Thus,
ξt = ξ0 +
t
0ξsX s dW s
= 1 + t
0ξsX s dW s.
(36)
E
t
0
ξsX s dW s | I u
=
u
0
ξsX s dW s, u < t,(37)
i.e., t
0 ξsX s dW s is a martingale, and hence
E[ξt|I u] = 1 + u
0ξsX s dW s = ξu,(38)
implying that ξt is a martingale.
26
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 14/17
Theorem. (Girsanov) Let W t be a Wiener
process w.r.t probability measure P and w.r.t
information sets I t, and let X t be as defined
above. Then if the process ξt is a martingale
w.r.t information sets I t, then W t defined by
W t = W
t − t
0X
udu, t
∈[0, T ](39)
is a Wiener process w.r.t information sets I tand w.r.t probability measure
P (A) = EP [1AξT ],(40)
where A∈ I T
and 1A
is the indicator function
of the event A.
27
In heuristic terms: If W t is a Wiener process
with probability measure P , then
d W t = dW t − X t dt(41)
is a Wiener process with probability measure
P , such that dP = ξT dP .
Remark 7.1: Generally the theorem gives us a method
to find the (equivalent) probability measures with re-
spect to which a drifting process can be turned to a
martingale.
Remark 7.2: We only change the drift and live the
volatility intact.
28
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 15/17
Example 7.3: Consider the general diffusion process
dS = a(S, t)dt + b(S, t)dW ,(42)
where W is the Wiener process w.r.t the probabilitymeasure P ,
dP (w) = 1√
2πte−
w2
2t dw,(43)
the normal distribution N (0, t).
Define
X t = a(S, t)
b(S, t)(44)
and assume that the drift a(S, t) and diffusion b(S, t)
are such that the Novikov condition (33) holds for X t.
Then defining
W = W − t
0
a(S, u)
b(S, u)du(45)
is a Wiener process with respect to the probability P
given by (40) and
dS = b(S, t)d W (46)
is a martingale w.r.t the probability measure P .
29
Example 7.4: As in Example 7.2, consider again the
geometric Brownian motion
dS = µSdt + σSdW,(47)
where µ and σ are constants. The probability measurefor W is again (43).
Let r be the risk-free rate and consider the discountedprice series
S t = e−rtS t.(48)
Using Ito,
dS = (µ − r)Sdt + σSdW.(49)
Now (44) becomes simply
X t = (µ− r)S
σS
= µ− r
σ
,(50)
W = W − µ− r
σ t(51)
is Wiener process w..r.t the probability measure P ,and (w.r.t. this measure)
dS = σSd W (52)
is martingale.
P ?
30
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 16/17
In the Girsanov theorem
ξt = e
t
0X u dW u−1
2
t
0X 2u du
= e1
σ(µ−r)W t− 1
2σ2 (µ−r)2t.(53)
Furthermore, we can consider only the events A =
{W t ≤ w}, w ∈ IR (because here the information sets
on the real line are Borel sets that are essentially open
intervals).
Because W t ∼ N (0, t) the associated probability mea-sure is (43). Then
P (A) = EP [1Aξt] =
w
−∞ξt(u)
1√ 2π t
e−u2
2t du =
w
−∞ξ(u)dP (u),
(54)i.e.,
dP (w) = ξt(w)dP (w)
= e1
σ(r−µ)w− 1
2σ2 (r−µ)2tdP (w)
= e1
σ(r−µ)w− 1
2σ2 (r−µ)2t 1√ 2π t
e−w2
2t dw
=
1
√ 2π te−1
2t(w
−r−µ
σ t)2
dw.
(55)
31
Denoting
w = w − µ− rσ
t,
we have finally
dP (w) = 1√
2π te−
1
2t w2
dw,(56)
which is again the density of the N (0, t) distribution.
The end result is that discounted price process (48)
is martingale with respect to the probability measure
P .
The solution of (48) is
S t = S 0e−1
2σ2t+σd W t.(57)
In terms of the original process from (48) S t = ertS ,we would have
S t = S 0e(r−1
2σ2)t+σ W t,(58)
or
dS = rS dt + σSd W ,(59)
i.e., we have essentially replaced the original drift µ
with the risk free rate r, and the end result is a pro-
cess whose discounted price process, S t = e−rtS t is a
martingale. Process (58) is usally called the risk neu-
tral process.
32
7/27/2019 mfdc7d
http://slidepdf.com/reader/full/mfdc7d 17/17
Remark 7.3: When pricing options, we calculate the
expected values with respect to the distribution of
risk neutral process S t given in (58). Its distribution
is log-normal with density
f S t(
y) =
1
√ 2πt σy e−(log(y)−θt)2
2σ2t
, y > 0
,(60)
where
θt = log S 0 + (r − 1
2σ2)t.(61)
Remark 7.4: The original distribution of S t is log-normal with density
f S t(y) = 1√ 2πt σy
e−(log(y)−θt)2
2σ2t , y > 0,(62)
where
θt = log S 0 + (µ − 1
2σ2)t.(63)
33