Finance Notes21
Finance Notes21
Finance Notes21
Péter Kevei
December 6, 2021
Contents
1 Introduction 1
1.1 Forward . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Put–call parity . . . . . . . . . . . . . . . . . . . . . . . . . . 4
7 Stochastic integration 37
7.1 Lévy characterization . . . . . . . . . . . . . . . . . . . . . . . 37
7.2 Girsanov’s theorem . . . . . . . . . . . . . . . . . . . . . . . . 38
9 Black–Scholes model 44
9.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
9.2 Equivalent martingale measure and the fair price . . . . . . . 45
9.3 Black–Scholes formula . . . . . . . . . . . . . . . . . . . . . . 47
9.4 From CRR to Black–Scholes . . . . . . . . . . . . . . . . . . . 48
0 K ST
−K
1 Introduction
These notes are based on the Hungarian lecture notes by Gáll and Pap [2],
on Shiryaev’s monograph [3], and on Elliott and Kopp [1].
There are two type of financial instruments: the basic financial units and
their derivatives.
Underlying:
• bond: risk-free asset, basically money. Its price is deterministic Bt ;
• stock: risky asset. Its price is a random, modeled by a stochastic
process St = (St1 , . . . , Std ).
Derivatives are bets on the underlying. They are used to share or reduce
risk. Here we consider forward contracts and options.
1.1 Forward
A forward contract is an agreement to buy or sell an asset (stock) for a price
previously agreed K in the future time T .
From the buyers point of view, at time T his wealth is ST − K, that is
the payoff function is f (s) = s − K.
We want to determine the fair price of this contract, and to understand
the meaning of ’fair’. Assume B0 = 1.
1
Seller’s point of view: At time 0, we can buy a stock for S0 . Then at time
T selling a stock for K and paying back the loan S0 · BT , we have K − S0 BT .
Therefore,
K ≥ S0 BT .
Buyer’s point of view: At time 0, we sell a stock for S0 . At time T we
pay K for a stock, and the our wealth is S0 BT − K. Thus,
K ≤ S0 BT .
K = S0 B1 = 40 · e0.1 = 44.2.
1.2 Options
An option is right to do something but not an obligation. European option
can be executed only at the expiration date, while American options can be
executed at any time.
The writer of a European call option agrees to sell a stock for a previously
agreed price K. Clearly, the buyer of this option will not use his right if
ST < K. The payoff function for the buyer is f (s) = (s − K)+
In case of a put option the writer agrees to buy a stock for K. The payoff
function of the buyer
2
y
0 K ST
0 K ST
3
1.3 Put–call parity
The aim of the course is to determine the fair price of an option, and under-
stand the fairness. However, there is a simple relation between call and put
prices regardless of the underlying market model.
Let CK be the fair price of the call, and PK be the fair price of the put,
both with strike price K. Then, from the payoff functions it is easy to see
that having put, a stock, and −1 call results at the expiration date (regardless
of the stock price) a wealth K. That is, after discounting
K
= P + S0 − C.
BT
This is the put-call parity.
2.1 Portfolio
An investment portfolio (strategy) is πn = (βn , γn ), where βn ∈ R represents
the amount of bonds in the portfolio at time n, while γn = (γn1 , . . . , γnd ) ∈ Rd ,
where γni represents the amount of type-i stock at time n. The random vari-
ables (βn , γn ) are Fn−1 -measurable, which means the investor has to decide
at time n − 1 how to invest on time n. That is the sequence (βn , γn ) is
predictable. For simplicity
d
X
γn Sn = γni Sni .
i=1
4
The wealth of the investor at time n under the strategy π is
Xnπ = βn Bn + γn Sn .
Proof. We have
∆Xn = Xn − Xn−1
= βn Bn − βn−1 Bn−1 + γn Sn − γn−1 Sn−1
= βn (Bn − Bn−1 ) + (βn − βn−1 )Bn−1 + γn (Sn − Sn−1 ) + (γn − γn−1 )Sn−1
= βn ∆Bn + ∆βn Bn−1 + γn ∆Sn + ∆γn Sn−1 ,
where Gπn is the gain process. So the value of the strategy is the initial
investment plus the gain.
5
2.2 Strategies in a more general market
Previously, we assumed that there are no transaction cost (market is fric-
tionless), shares pay no dividend, and apart from time 0, there is neither
investment, nor consumption. Here we see how to handle this.
2.2.1 Dividend
i
Assume that stock-i pays a dividend δni = Dni − Dn−1 ≥ 0 at n, where δni ,
and Dni are adapted processes. Then the change in the value process is
Xnπ = βn Bn + γn (Sn + δn ).
Bn−1 ∆βn + (1 + λ)Sn−1 ∆γn I(∆γn > 0) + (1 − µ)Sn−1 ∆γn I(∆γn < 0) = 0.
6
2.3 Claim and hedging
Let fN be a nonnegative random variable, which is the payoff function, or
obligation, or contingent claim. A strategy π is an upper (x, fN )-hedge, if
P-almost surely
X0π = x, XNπ ≥ fN .
It is a lower (x, fN )-hedge, if a.s.
X0π = x, XNπ ≤ fN .
For the class of upper (x, fN )-hedge strategies put H ∗ (x, fN , P), and for the
lower H∗ (x, fN , P).
{lemma:hedge}
Lemma 2. For any payoff function fN there exists an x such that there is
an upper (x, fN )-hedge.
Proof. Put
B0
x= max |fN (ω)|.
BN ω∈Ω
Then the (trivial) strategy πn ≡ ( Bx0 , 0) (start with enough money and don’t
do anything) is an upper hedge.
7
Using the strategy π1 = (β1 , γ1 ) we want that
X1π = β1 B1 + γ1 S1 = f a.s.
β1 B0 (1 + r) + γ1 S0 (1 + a) = f0
β1 B0 (1 + r) + γ1 S0 (1 + b) = f1 .
Solution 1. Let
8
S2 = s↑↑ , f = f ↑↑
S1 = s↑
S2 = s↑↓ , f = f ↑↓
S0
S2 = s↓↑ , f = f ↓↑
↓
S1 = s
S2 = s↓↓ , f = f ↓↓
For the filtration let Fn = σ(ρ1 , . . . , ρn ), i.e. the natural filtration generated
by the variables ρ1 , . . . , ρn .
Consider any payoff function fN . A perfect hedge can be constructed
recursively, using the simple one-step market. Indeed, a two-step model can
be seen as 3 one-step markets.
9
{lemma:arbitrage}
Lemma 3. Assume that π is a weak arbitrage strategy. Then there exists an
arbitrage strategy π 0 .
Proof. If Xnπ ≥ 0 a.s. for all n, then we are ready. Otherwise, there exists
π
m < N such that P(Xm < 0) > 0, and Xnπ ≥ 0 for any n ≥ m + 1. Let
Am = {Xm < 0} ∈ Fm .
It is easy to check that this strategy is predictable, SF, and arbitrage strategy.
Indeed,
(i) predictable: for n ≤ m this is clear, since βn0 = 0 and γn0 = 0, while for
n > m Am is Fm -measurable and thus Fn−1 -measurable as well, and
βn , γn are Fn−1 -measurable by the assumption.
(ii) SF: for n ≤ m this is again clear. For n = m + 1
0 0 π
Bm ∆βm+1 +Sm ∆γm+1 = IAm (Bm βm+1 (ω) − Xm (ω) + Sm γm+1 (ω)) = 0,
since π is SF. For n > m + 1 we have ∆βn0 = IAm ∆βn , and ∆γn0 =
IAm ∆γn , and the result follows, using again that π is SF.
(iii) arbitrage: we have
π
0 Xm Bn
Xnπ = IAm In>m βn Bn + γn Sn − ,
Bm
where the sum of the first two terms in the bracket is nonnegative by
the definition of m and the last is strictly negative on Am , which proves
the statement.
Exercise 2. Assume that a < b < r in the one-step binomial model. Give
an arbitrage strategy.
Assume that an < bn < rn for some n in the N -step binomial model.
Give an arbitrage strategy.
10
3.2 Martingale measures
A probability measure Q is called equivalent martingale measure (EMM) if
P ∼ Q and (Sni /Bn , Fn ) is a Q-martingale for each i = 1, 2, . . . , d.
11
Therefore Sn /Bn is a Q-martingale iff
EQ [ρn |Fn−1 ] = rn .
This condition exactly tells that under the new measure Q the risky asset
behaves as the bond on average. Using that ρn ∈ {an , bn }, we obtain as
above
bn − rn rn − an
Q(ρn = an |Fn−1 ) = , and Q(ρn = bn |Fn−1 ) = .
b n − an b n − an
Note the conditioning on Fn−1 gives a constant, meaning that ρn is indepen-
dent of Fn−1 under the measure Q.
We obtained the following.
{thm:binom-EMM}
Theorem 1. In the binomial market if an < rn < bn for each n then there
exists a unique EMM Q given by the formulas above. Moreover, under Q the
random variables ρ1 , . . . , ρN are independent.
12
Proof. Easily follows from the SF property. Indeed, using that βn , γn are
Fn−1 -measurable
π
Xn Sn
EQ Fn−1 = EQ βn + γn Fn−1
Bn Bn
Sn
= βn + γn EQ Fn−1
Bn
Sn−1
= βn + γn
Bn−1
βn Bn−1 + γn Sn−1
=
Bn−1
π
X
= n−1 ,
Bn−1
where the last equality follow from the self-financing property.
The following main result is the first fundamental theorem of asset pricing.
{thm:emm-arb}
Theorem 2. There exists an EMM if and only if the market is arbitrage-free.
Proof. Let Q be an EMM and π be any strategy with X0π = 0. Then, by the
previous statement
Xπ Xπ
EQ N = EQ 0 = 0.
BN B0
Thus XN ≥ 0 P-a.s., then also Q-a.s., which implies XN ≡ 0 Q-a.s., thus
P-a.s.
We prove the converse later.
Assume that fN is a replicable payoff, i.e. there is a prefect hedge π. This
means that
XNπ = fN a.s.
Then the fair price for fN is the initial cost of the portfolio, X0π = x. By the
martingale property
fN X π mtg Xπ x
EQ = EQ N = EQ 0 = .
BN BN B0 B0
That is, the fair price x for a replicable payoff fN is
B0
x= EQ f.
BN
13
In particular, it also follows that for a replicable f , the value EQ f is the
same for any EMM Q.
Summarizing, we proved the following:
{thm:pricing}
Theorem 3. Consider an arbitrage-free market and let f be a replicable
payoff. Then the fair price of f is
B0
C(f ) = C∗ = C ∗ = EQ f,
BN
where Q is any EMM.
S1 = S0 (1 + ρ),
where ρ > −1 is a random variable, the unique source of randomness in the model. Let
be the distribution function of ρ. Then F induces a probability measure (denoted by P) on the Borel sets
of (−1, ∞) (or R). If F is concentrated on {a, b} then we get back the previous one-step binomial model.
Assume without loss of generality that B0 = 1. Consider a payoff function f : R → R as a function
of the stock price S1 . A strategy π is an upper hedge if
S1 S0
EQ = .
B1 B0
This means
EQ ρ = r.
That is a probability measure Q which is equivalent to P is EMM iff
Z
ρdQ(ρ) = r.
R
f (S0 (1 + ρ))
β + γS0 ≥ EQ .
1+r
14
y
f (x)
a b x
Then
Assume now that ρ ∈ [a, b] for some −1 ≤ a < b < ∞. To ease notation put
15
Indeed, the left-hand side is a linear function equals to f (a) at a, and f (b) at b. Introduce the strategy,
ν
π ∗ = (β ∗ , γ ∗ ) := ,µ .
1+r
Then, by (5)
∗
X1π = ν + µS0 (1 + ρ) ≥ f (ρ),
that is, π ∗ is an upper hedge. Therefore,
ν
C ∗ (f ) = inf X0π ≤ β ∗ + S0 γ ∗ = + µS0 . (6) {eq:C*lower}
πupper hedge 1+r
Assumption (weak limit): In the set P(P) there exists a sequence Pn , such that Pn converges
weakly to a measure Q∗ supported on {a, b}. Since
EPn ρ = r,
b−r r−a
Q∗ ({a}) = , Q∗ ({b}) = .
b−a b−a
Note that Q∗ is, in general, not equivalent to P. In fact, it is only equivalent in the binomial market
setup.
By the convergence of Pn (here we use the continuity of f )
f (ρ) f (ρ)
sup EQ ≥ lim EPn
Q∈P(P) 1+r n→∞ 1+r
f (ρ)
= EQ∗
1+r
f (a) f (b)
= Q∗ ({a}) + (1 − Q∗ ({a}))
1+r 1+r
= β ∗ + γ ∗ S0 ≥ C ∗ (f ).
Theorem 4. Assume that the payoff function is convex and continuous on [a, b], and that the weak limit
assumption holds. Then
Exercise 4. Let ρ be uniform random variable on [a, b]. Show that the weak limit property holds.
Construct Pn explicitly!
Try to weaken the condition on the distribution of ρ.
16
y
f (x)
a r b x
Let’s see the lower price C∗ (f ). Assume again that f in (4) is continuous and convex. Then
f (ρ) f (ρ)
inf EQ ≤ lim EPn
Q∈P(P) 1+r n→∞ 1+r
f (r)
= = β∗ + S0 γ∗
1+r
≤ sup{β + γS0 : (β, γ) lower hedge } = C∗ .
17
Theorem 5. Let f be a convex continuous function on [a, b], and assume that the weak limit-2 assumption
holds. Then
f (ρ) f (r)
C∗ (f ) = inf EQ = ,
Q∈P(P) 1+r 1+r
and the infimum is attained at the measure Q∗ .
Proof. We prove again the easy parts (i) ⇒ (ii), and (iii) ⇔ (i), and postpone
the difficult (ii) ⇒ (i) implication later.
(i) ⇒ (ii): Assume that Q1 and Q2 are EMM’s. Consider any A ∈ F.
We show that Q1 (A) = Q2 (A) implying the uniqueness. Let π be a perfect
hedge to f = IA . Then Xnπ /Bn is both Q1 and Q2 martingale, so
XNπ Xπ
Q1 (A) = EQ1 f = EQ1 XNπ = BN EQ1 = BN 0 = . . . = Q2 (A).
BN B0
(i) ⇒ (iii): Consider a Q-martingale Mn . There exists a strategy πn such
that a.s.
XNπ = BN MN .
18
Using that both Mn and Xnπ /Bn are martingales
π
Xπ
XN Sn
Mn = EQ [MN |Fn ] = EQ |Fn = n = βn + γn .
BN Bn Bn
Thus, using that π is SF
Sn Sn−1
Mn − Mn−1 = ∆βn + γn − γn−1
Bn Bn−1
Sn Sn−1 1
= γn − + (Bn−1 ∆βn + Sn−1 ∆γn )
Bn Bn−1 Bn−1
Sn Sn−1
= γn − ,
Bn Bn−1
as claimed.
(iii) ⇒ (i): Consider a payoff f . We are looking for a strategy π such
that XNπ = f Q-a.s. We know that (Xnπ /Bn )n is a martingale, so this should
be (Mn ). Now the following choice is clear: let
f
Mn = EQ |Fn .
BN
Then Mn is a martingale, therefore by the assumption
n
X Sk
Mn = M0 + γk ∆ .
k=1
Bk
Let
Sn
βn = Mn − γn ,
Bn
and consider the strategy πn = (βn , γn ). To see that this is indeed a strategy
we have to show that it is predictable and SF. The sequence γn is predictable
by the assumption (iii), and βn is predictable because all the terms in Mn
are Fn−1 -measurable except γn Sn /Bn , which is subtracted. To see that it is
SF note that
Bn−1 ∆βn + Sn−1 ∆γn
Sn Sn−1
= Bn−1 Mn − Mn−1 − γn + γn−1 + Sn−1 ∆γn
Bn Bn−1
Sn Sn Sn−1
= Bn−1 γn ∆ − γn + γn−1 + Sn−1 ∆γn = 0,
Bn Bn Bn−1
19
showing that π is SF. It is clearly a perfect hedge since
XNπ = βN BN + γN SN = BN MN = f,
as claimed.
and
V1 = {X : Ω → R r.v. |X ≥ 0, EX ≥ 1}.
We identify a random variable X : Ω → R with a vector in Rk , as X ↔
(X(ω1 ), . . . , X(ωk )). Clearly, V0 is a linear subspace and V1 is convex set in
Rk .
Since there is no arbitrage strategy, V0 ∩V1 = ∅. Therefore, by the Kreps–
Yan theorem, there exists a linear functional ` : Rk → R such that `|V0 ≡ 0
and `(v1 ) > 0 for all v1 ∈ V1 . A linear function in Rk (in any Hilbert space)
is a inner product, thus there exists q ∈ Rk such that
20
Lemma 4. Let (Xn )N n=1 be an adapted process. If for any stopping time
τ : Ω → {0, . . . , N }
EXτ = EX0 ,
then (Xn ) is martingale.
This is indeed a stopping time, since {τA ≤ k} = ∅ for k < n, and A for
k ≥ n, which is Fk -measurable. Then, by the assumption
21
y
V0
V2
Figure 7: Choice of y
so XNπ ∈ V0 . Therefore
0 = EQ XNπ = EQ βN BN + γN SN
Sτ S0 Sτ
EQ I(τ ≤ N − 1) − BN + I(τ = N )BN
Bτ B0 Bτ
Sτ S0
= BN EQ − .
Bτ B0
22
we may choose ε > 0 small enough such that
Exactly as in the previous proof we can show that Q0 is EMM. The uniqueness
of the EMM implies
qi0 qi
Pk 0 = Pk ,
i=1 iq i=1 q i
q = αq 0 = αq − αεy,
qi0 . Thus
P P
with α = qi /
(1 − α)q = −αεy.
Lemma 5. Let X be a random variable on (R, B(R), P) such that P(X > 0)P(X < 0) > 0. Then there
exists a probability measure Q ∼ P such that EQ X = 0. Furthermore, for any a ∈ R
EQ eaX < ∞.
23
Proof. Define the probability measure
2
P1 (dx) = ce−x F (dx),
2
where F (x) = P(X ≤ x) and c−1 = e−x F (dx). That is
R
R
Z
2
P1 (A) = ce−x F (dx).
A
2
Clearly, ϕ(a) < ∞ for any a as the function eax−x is bounded on R. Note that ϕ is convex, because
ϕ00 > 0. Put
eax
Za (x) = .
ϕ(a)
Then
Qa (dx) = Za (x)P1 (dx)
is a probability measure for any a, and Qa ∼ P1 ∼ F . Again, this means
Z Z
c 2
Qa (A) = Za (x)P1 (dx) = eax−x F (dx).
A ϕ(a) A
Let
ϕ∗ = inf ϕ(a).
a∈R
lim ϕ(a) = ∞.
a→±∞
Therefore, the infimum is attained, i.e. there is a∗ such that ϕ(a∗ ) = ϕ∗ . Then ϕ0 (a∗ ) = 0, thus
ea∗ X
0 = ϕ0 (a∗ ) = EP1 Xea∗ X = ϕ(a∗ )EP1 X = ϕ(a∗ )EQa∗ X.
ϕ(a∗ )
24
As a preliminary result we have to understand how to compute conditional expectation under different
measures.
{lemma:condexp-meas
Lemma 6. Let (Ω, F , (Fn )n=0,1,...,N , P) a filtered probability space, and Z a nonnegative random vari-
able EP Z = 1. Define the new probability measure Q as
dQ = ZdP,
that is Z
Q(A) = ZdP.
A
Proof. Both sides are Fn−1 -measurable. We have to prove that for any A ∈ Fn−1
Z Z
Zn−1 EQ [Xn |Fn−1 ]dP = Xn Zn dP. (8) {eq:cemlemma-0}
A A
which is (8).
25
{lemma:existence-em
Lemma 7. Let (Xn )N
n=1 be an adapted process, and assume that
Then there exists a probabilty measure Q ∼ P such that (Xn ) is a Q-martingale difference.
N N
Z ( ) ( )
X X
c−1 = exp − Xi2 dP = E exp − Xi2 .
Ω i=0 i=0
Let
ϕn (a) = EP1 [eaXn |Fn−1 ].
Note that this is an Fn−1 -measurable random variable. As in the proof of the previous lemma there is a
unique finite an (random!) such that the infimum of ϕn is attained at an . Since ϕn is Fn−1 -measurable
so is an .
Let Z0 = 1, and recursively
ean Xn
Zn = Zn−1 .
EP1 [ean Xn |Fn−1 ]
1
EQ [Xn |Fn−1 ] = EP1 [Zn Xn |Fn−1 ] by Lemma 6
Zn−1
1 Zn−1
= EP1 [Xn ean Xn |Fn−1 ] definition
Zn−1 EP1 [ean Xn |Fn−1 ]
1
= · 0 = 0. choice of an
EP1 [ean Xn |Fn−1 ]
Now we can return to the proof of Theorem 2. The existence of the martingale measure follows from
the previous lemma applied to Xn = ∆Sn .
26
4.2 ARCH processes
Autoregressive conditional heteroscedasticity (ARCH) models were introduced by Robert Engle in 1982
to model log-returns. In 2003 he obtained Nobel prize in economics for this model. The novelty in these
models is the stochastic volatility term.
Let
Sn
Rn = log
Sn−1
denote the log-return of the stock, and assume that
q
Rn = µn + 2
β + λRn−1 ξn ,
where ξn ’s are iid N (0, 1) random variables. Then (Rn ) is an ARCH(1) process. That is conditionally on
Fn−1 the log-return Rn is Gaussian with mean µn , and variance β + λRn−1 2 . Write σn = β + λRn−1 2 .
Then for Sn we obtain
n
( )
X q
Sn = Sn−1 eRn = S0 exp µk + 2
β + λRk−1 ξk
k=1
n
( )
X
= S0 exp (µk + σk ξk ) .
k=1
In what follows we only assume that µn and σn are Fn−1 -measurable, i.e. the sequence (µn , σn )n is
predictable, and (ξn ) is adapted, ξn is independent of Fn−1 , and N (0, 1) distributed. Put hn = µn +σn ξn .
For simplicity we assume that Bn ≡ 1.
We construct a measure Q such that (Sn ) is a Q-martingale. Let
N N
Y Y ean hn
ZN = zn := ,
n=1 n=1
EP [ean hn |F n−1 ]
where
µn 1
an = − 2
− . (10) {eq:disc-girs-0}
σn 2
Introduce the new measure Q as
dQ = ZN dP,
Qn
and let Zn = EP [ZN |Fn ] = i=1 zi .
By Corollary 1, to show that Sn is Q-martingale we have to show that Sn Zn is a P-martingale. We
have
EP [ehn (1+an ) |Fn−1 ]
EP [Sn Zn |Fn−1 ] = Sn−1 Zn−1 .
EP [ean hn |Fn−1 ]
Therefore we have to check that
2 (1+a )2
σn n
EP [ehn (1+an ) |Fn−1 ] = eµn (1+an )+ 2 ,
27
and
2 a2
σn n
EP [ehn an |Fn−1 ] = eµn an + 2 ,
By the choice of an in (10)
2 (1 + a )2
σn σ 2 a2
n
µn (1 + an ) + = µn an + n n .
2 2
Indeed, by (10)
2 1
µn + σn + an = 0.
2
That is, (11) holds.
We proved the following.
Theorem 7 (Discrete Girsanov’s theorem). Let (µn , σn )n be a predictable sequence and assume that the
stock prices are given by Pn
Sn = e k=1 (µk +σk ξk ) ,
where (ξn )n is a adapted sequence of N(0, 1) random variables, ξn is independent of Fn−1 . Further, let
Bn ≡ 1. Then, under the new measure
dQ = ZN dP
(Sn ) is a martingale.
28
For a hedge we need to know not only the fair price C, but also the
strategy π itself. For the given claim fN consider the martingale
fN
Mn = EQ Fn .
BN
γn Sn
βn = Mn − .
Bn
We proved that π = (βn , γn )n is an SF strategy and is a perfect hedge for
fN .
Summarizing, we obtained the following.
XNπ = fN ,
29
where ρk ∈ {a, b}. We proved that this market is arbitrage-free and complete,
and the unique EMM is given by
b−r
Q(ρi = a) = ,
b−a
and ρi ’s are independent. If the claim fN only depends on the final price SN ,
and not on the whole trajectory, i.e.
r−a
where q = b−a
.
6 American options
While European options can be exercised only at the terminal date N , Ameri-
can options can be exercised at any time. Formally, instead of a fixed random
30
payoff function fN , a sequence of payoffs (fn )n=0,1,...,N is given, where fn is
Fn -measurable, i.e. (fn )n is adapted to (Fn )n . So fn is the random payoff
if the option is exercised at time n. Clearly, the exercise time has to be a
stopping time.
Znτ = Zτ ∧n ,
{τ ∗ = n} = ∩k=0
n−1
{Zk > Xk } ∩ {Zn = Xn }.
31
On the event {τ ∗ ≥ n} we have Zn−1 = E[Zn |Fn−1 ] therefore
∗
Proof. Since Z τ is martingale
∗ ∗
Z0 = Z0τ = EZNτ = EZτ ∗ = EXτ ∗ .
On the other hand for any stopping time τ the process Z τ is supermartingale
(by Doob’s optional sampling), thus
Proposition 4. The stopping time σ is optimal iff the following two condi-
tions hold.
(i) Zσ = Xσ ;
(ii) Z σ is martingale.
Proof. If (i) and (ii) hold than σ is optimal. This follows exactly as the
optimality of τ ∗ .
Conversely, assume that σ is optimal. We have seen that supτ EXτ = Z0
thus
Z0 = EXσ ≤ EZσ ,
by the dominance of Z. By Doob’s optional stopping theorem Z σ is super-
martingale, therefore EZσ ≤ Z0 , implying that
EXσ = EZσ .
32
Since Zn ≥ Xn this implies Xσ = Zσ a.s., proving (i).
By the optimality EZσ = Z0 , while the supermartingale property implies
Z0 ≥ EZσ∧n ≥ EZσ .
Thus
EZσ∧n = EZσ = EE[Zσ |Fn ].
Furthermore, by Doob’s optional stopping
Xnπ ≥ fn , n = 0, 1, . . . , N,
At time N we need
XNπ ≥ fN .
At time N − 1 the holder either exercise the option or continues to time N ,
(in that case we discount the price), therefore
π BN −1
XN −1 ≥ max fN −1 , EQ [fN |FN −1 ] .
BN
33
Dividing by BN −1
XNπ −1
fN −1 fN
≥ max , EQ FN −1 .
BN −1 BN −1 BN
Thus, we see the connection with the Snell-envelope.
For a hedging strategy π we have that
(i) (Xnπ /Bn )n is a Q-martingale (since Q is EMM and π is SF), and
(ii) (Xnπ /Bn ) dominates (fn /Bn ) (since π is a hedge).
Therefore, the value process of a hedge is larger than the Snell-envelope of
(fn /Bn ), i.e.
Xnπ
≥ Zn , n = 0, 1, . . . , N, (12) {eq:di-american-1}
Bn
where (Zn ) is the Snell-envelope of (fn /Bn ). The Snell-envelope (Zn ) is
a supermartingale, therefore by the Doob-decomposition (that’s stated for
submartingale, but multiply by −1) we have
Zn = Mn − An , n = 0, 1, . . . , N, (13) {eq:di-american-2}
where Mn is a Q-martingale, and (An ) is an increasing predictable sequence,
A0 = 0. Comparing (12) and (13) we see that for n ≤ τ ∗
Xnπ
≥ Mn .
Bn
On the other hand, the market is complete, which implies (see the easy parts
of the proof of Theorem 6) that there exists a strategy π such that
Xnπ
= Mn , n = 0, 1, . . . , N.
Bn
This is a minimal hedging strategy with initial cost
x Xπ
= 0 = M0 = Z0 .
B0 B0
{thm:price-di-amer
Theorem 10. Consider an aribtrage-free complete market with unique EMM
Q. Let (fn ) be the nonnegative payoff sequence of an American option. Let
(Zn ) be the Snell-envelope of the discounted payoff sequence (fn /Bn ). The
fair price for this option is
fτ fτ ∗
C = B0 Z0 = B0 sup EQ = B0 EQ ,
τ ∈MN
0
Bτ Bτ ∗
34
where τ ∗ is an (not unique in general) optimal exercise time given by
∗ fn
τ = min n : = Zn .
Bn
Furthermore, there exists a SF strategy π which is an optimal hedge with
initial cost C and
fτ ∗
Xτπ∗ = .
Bτ ∗
Assume that the deterministic sequence (Bn ) is nondecreasing (i.e. the inter-
est rate is nonnegative). Let (Zn ) denote the Snell envelope of (fn /Bn ), that
is
fN fn
ZN = , Zn = max , E [Zn+1 |Fn ] , n = 0, 1, . . . , N − 1.
BN Bn
Using that (Sn /Bn ) is a Q-martingale, by Jensen’s inequality
fN −1 (SN −1 − K)+
=
BN −1 BN −1
SN −1 K
= −
BN −1 BN −1 +
SN K
≤ EQ − FN −1 Jensen’s inequality
BN BN −1 +
SN K
≤ EQ − FN −1 by BN ≥ BN −1
BN BN +
(SN − K)+
= EQ FN −1
BN
= EQ [ZN |FN −1 ].
35
This means that at time N − 1 it is always good to hold the option and
continue to step N .
An induction argument shows that at any time it is better to hold the
option. Indeed, assume for some n
fn
≤ EQ [Zn+1 |Fn ].
Bn
We just proved this for n = N − 1. The same way as above we have
Theorem 11. Assume that the market is arbitrage free and complete, and
the interest rate is nonnegative. Then the price of a European call option
equals to the price of the American call option.
36
7 Stochastic integration
7.1 Lévy characterization
One can define stochastic integral with respect to more general processes.
The process (Xt ) is a continuous semimartingale if
Xt = Mt + At ,
u2 t iu(Mv −Ms )
Z
iu(Mt −Ms )
E e IA = P(A) − E e IA dv.
2 s
37
With A and s fixed, define
u2
g 0 (t) = − g(t), g(s) = P(A).
2
Therefore, the solution
u2
g(t) = P(A) · e− 2
(t−s)
.
38
In what follows, we have more (usually 2) probability measures, therefore
we
R put in the lower indexR of E the corresponding measure. That is EP X =
Ω
XdP, and E Q X = Ω
XdQ. Note that the notion of martingale does
depend on the underlying measure. Therefore, we have P-martingale, and
Q-martingale.
Define the P-martingale
Mt = EP [M∞ |Ft ].
{lemma:p-q-mtg}
Lemma 8. The adapted process (Xt ) is Q-martingale if and only if (Mt Xt )
is P-martingale.
Proof. Since
EP [M∞ Xt |Ft ] = Xt Mt ,
for each A ∈ Ft Z Z
Xt M∞ dP = Xt Mt dP.
A A
Therefore, if A ∈ Fs ⊂ Ft , then
Z Z Z
Xt dQ = Xt M∞ dP = Xt Mt dP
ZA ZA ZA
Xs dQ = Xs M∞ dP = Xs Ms dP.
A A A
Then (Xt ) is Q-martingale if the left-hand sides are equal for each A ∈ Fs ,
s < t, which is obviously equivalent to the equality of the right-hand sides,
which means that (Mt Xt ) is P-martingale.
Let Z t Z t
1
ζts = θu dWu − θu2 du, ζt = ζt0 ,
s 2 s
ζt
where θt is adapted. Then Zt = e satisfies the SDE
Z t
Zt = 1 + Zs Xs dWs . (14) {eq:Gir-sde}
0
We use this formula in the proof of Girsanov’s theorem. We can write the
SDE above as
dZt = Zt Xt dWt , Z0 = 1.
39
Indeed, rewriting ζ as an Itô process
Z t Z t
1 2
ζt = − θu du + θu dWu .
0 2 0
Exercise 10. Let ζt as above. Prove that Yt = e−ζt satisfies the SDE
dQθ
= ΛT .
dP FT
Rt
Then W
ft = Wt +
0
θs ds is Q-SBM.
40
Proof. First we show that Q is indeed a probability measure. By (14)
Z t
Λt = 1 − Λs θs dWs ,
0
which is martingale, so
EP ΛT = EP Λ0 = 1.
Since ΛT > 0 we see that Q is probability measure.
Next we show that Wf satisfies the conditions of the Lévy characterization.
The continuity is clear, since W is SBM and Q P. By Lemma 8 (W ft )
is Q-martingale iff (W
ft Λt ) is P-martingale. We apply the Itô formula with
f (x, y) = xy and the Itô process
Z t Z t
W
ft = θs ds + 1dWs
0 0
Z t
Λt = 1 − Λs θs dWs .
0
Then
Z t Z t Z t
Λt W
ft = W
fs dΛs + Λs dW
fs + −Λs θs ds
0 0 0
Z t Z t Z t
=− W
fs Λs θs dWs + Λs (θs ds + dWs ) − Λs θs ds
0 0 0
Z t
= Λs (1 − θs W
fs )dWs ,
0
41
ft2 − t) is Q-martingale, and the proof is
which is P-martingale. Thus (W
complete.
Finally, we state without proof (and precise statement) the martingale
representation theorem.
{thm:martingale-re
Theorem 15 (Martingale representation theorem). Let (Wt ) SBM on (Ω, A, P),
and let (Ft ) the generated filtration, together with the P-zero sets. If (Mt ) is
continuous square integrable martingale with M0 = 0 a.s., then there exists
an adapted (Yt ) such that
Z t
Mt = Ys dWs .
0
42
The continuous time analogue of the above is the SDE
as claimed.
For the reverse direction, we have
π
dX t = γt dS t .
43
An SF strategy π is arbitrage, if X0π = 0 a.s., XT ≥ 0 a.s., and P(XTπ >
0) > 0. The market is arbitrage free if there exists no arbitrage strategy.
A probability measure Q is equivalent martingale measure (EMM) if P ∼
Q (that is P Q and Q P), and (S t ) is Q-martingale.
We have seen in the discrete time setup that the existence of EMM is
equivalent to the arbitrage free property. One of the implications is rather
simple in the continuous time setup. Assume that Q is EMM, and let π
be an (SF) strategy. By Proposition 5 the discounted value process has the
representation Z t
π
X t = X0π + γs dS s .
0
π
Since (S t ) is Q-martingale, and X t is a stochastic integral with respect to
π
S, we see that (X t ) is Q-martingale. (Recall the discrete time analogue of
this statement.) Therefore
π
EQ X T = EQ X0π .
π
Since P ∼ Q, X0π = 0, XTπ ≥ 0 P-a.s., implies Q-a.s. Then EQ X T =
EQ X0π = 0, from which XTπ ≡ 0 Q-a.s., and so P-a.s.
We proved the following.
Theorem 16. Assume that on the market (Ω, A, P, (St ), (Bt = ert ), (Ft ))
there exists EMM. Then the market is arbitrage free.
9 Black–Scholes model
In a special model we explicitly construct the EMM via Girsanov’s theorem,
and compute the fair price of a payoff. In particular, we prove the Nobel-
prize winner Black–Scholes pricing formula, which gives the fair price of a
European call option.
44
From the form of St we immediately see that St is a martingale if and
only if µ = 0.
The bond price is simply Bt = ert .
Writing St as an Itô process
Z t Z t
St = S0 + µSs ds + σSs dWs .
0 0
From which 2
σWt + µ− σ2 t
St = S0 · e . (19) {eq:exp-BM}
This is the exponential Brownian motion.
Note that the proof is not complete, because the logarithm is not smooth
at 0. The argument above only helps to find out the solution. (A more
constructive approach is to apply Itô with a general f , and then choose f to
obtain a solvable equation.)
45
µ−r
Let θt ≡ θ = σ
, and
Z T Z T
dQ 1 θ2 T
= ΛT = exp − θdWs − 2
θ ds = e−θWT − 2 .
dP FT 0 2 0
Nt = EQ e−rT fT |Ft , 0 ≤ t ≤ T.
Yt Yt ert
βt = Nt − , γt = .
σ σSt
π
Lemma 9. The strategy (πt = (βt , γt )) is self-financing and X t = Nt .
46
Since
XTπ = erT NT = erT EPµ e−rT fT |FT = fT ,
Yt Yt ert
βt = Nt − , γt = ,
σ σSt
Rt
is a perfect hedge, where Nt = EQ [e−rT fT |Ft ], and Nt = N0 + 0
fµ .
Ys dW s
By (22)
fµ σ2
ST = S0 erT eσWT − 2
T
,
f µ ∼ N(0, T ) under Q. Therefore, writing Z for a standard normal
where WT
f µ σ2
= EQ S0 eσWT − 2 T − e−rT K
+
√ σ2
= E S0 eσ T Z− 2 T − e−rT K
+
Z ∞ √
1 σ2
x2
(24) {eq:BS-calc}
=√ S0 eσ T x− 2 T − e−rT K e− 2 dx
2π γ
Z ∞ √
1 −
(x−σ T )2
= S0 √ e 2 dx − e−rT K(1 − Φ(γ))
2π γ
√
= S0 1 − Φ(γ − σ T ) − e−rT K(1 − Φ(γ)),
47
where 2
1 K σ
γ= √ log + −r T .
σ T S0 2
The pricing formula
√
CT (K) = S0 1 − Φ(γ − σ T ) − e−rT K(1 − Φ(γ))
which in fact suggests the choice of rN . Similar, but more complicated cal-
culations gives that with the choice above VarSτNN converges.
In the homogeneous binomial model the EMM was given by the upwards
step probability
r N − aN
p∗N = .
b N − aN
48
Under the EMM
YN
N −YN 1 + bN
SτNN = S0 (1 + bN ) YN
(1 + aN ) = S0 (1 + aN )N , (26) {eq:crrS_N}
1 + aN
where YN ∼ Binomial(N, p∗N ).
The CRR pricing formula gives
(SτNN − K)+
CN (K) = E∗N . (27) {eq:crr-ar}
BτNN
49
Similarly,
√ σ2
aN = −σ h + + r h + O(h3/2 ).
2
From this
√ 2
∗ r N − a N σ h − σ2 h + O(h3/2 )
pN = = √
b N − aN 2σ h + O(h3/2 )
√
1 σ h + O(h)
= −
2 + O(h) 4 + O(h)
1 σ √
= − h + O(h).
2 4
Substituting back, and using the second order expansion log(1 + x) = x −
x2 /2 + O(x3 ), x → 0, we obtain
q
1 + bN
q √ √
∗ ∗
lim N pN (1 − pN ) log = lim p∗N (1 − p∗N )2σ T = σ T ,
N →∞ 1 + aN N →∞
and
∗ 1 + bN
lim N pN log + log(1 + aN )
N →∞ 1 + aN
" r # r r !
1 σ T T T T
= lim N − + O(N −1 ) 2σ −σ + r + O(N −3/2 )
N →∞ 2 4 N N N N
2
σ
= r− T.
2
Substituting back into (27)
√ 2
−rT σ T Z+T r− σ2
∗
lim CN (K) = e E S0 e −K
N →∞
+
√ 2
∗ σ T Z− σ2 T −rT
=E S0 e −e K ,
+
50
10 Interest rate models
10.1 The general setup
In what follows we are interested in options on bonds instead of stocks.
Therefore, we assume that the stock price Bt is also random. The bond price
is given by Z t
Bt = exp ru du , (29) {eq:bond}
0
where rt , the interest rate is an adapted stochastic process. The time interval
is [0, T ]. The stock price is given by
Z t Z t
St = S0 + µ(u)Su du + σu Su dWu , (30) {eq:stock}
0 0
with some adapted process µ and σ. Note that the bond price Bt is a stochas-
tic process too, but it is much smoother than the stock price St , as it is the
exponential of the Lebesgue integral of a stochastic process. In particular,
the path of Bt are of bounded variation, while the path of St are not. (Re-
call that an Itô process is of bounded variation if and only if the stochastic
integral part vanishes.)
We want to find an equivalent martingale measure. For the discounted
stock price S t = St /Bt
St
− 0t ru du
R
d = d St e
Bt
Rt Rt
= e− 0 ru
dSt + St (−rt )e− 0 ru du
dt
= S t ((µt − rt )dt + σt dWt )
= S t σt dW
ft ,
where Z t
W
ft = θs ds + Wt ,
0
with θs = µsσ−r
s
s
. Applying Girsanov’s theorem W
ft is SBM under the measure
Qθ , where Z T
1 T 2
Z
dQθ
= exp − θs dWs − θ ds .
dP 0 2 0 s
51
Therefore, under Qθ the discounted stock price S t is a martingale, i.e. Qθ is
an equivalent martingale measure.
We are not interested in the specific form of the underlying risky asset
(St ) in (30), but we assume that there exists a unique equivalent martingale
measure (that is (St /Bt ) is martingale). This will be the only measure on
the probability space, therefore it is denoted by P (instead of Qθ ).
Formally, let (Ω, A, (Ft ), P) be a filtered probability space, (ru ) an adapted
stochastic process, and (Bt ) is given by (29). We assume that the risky as-
set (St ) is an adapted stochastic process, such that (St /Bt )t is a martingale
under P, and P is the unique such measure.
A zero coupon bond (elemi kötvény) maturing at time T is a claim that
pays 1 at time T . Its value at time t ∈ [0, T ] is denoted by P (t, T ), 0 ≤ t ≤
T ≤T.
From the pricing theorem we see that the fair price of the zero coupon
bond at time 0 is
1
P (0, T ) = E ,
BT
thus at time 0 ≤ t ≤ T
Z T
1
P (t, T ) = Bt E Ft = E exp − ru du Ft . (31) {eq:P(tT)}
BT t
52
where µ > 0, σ > 0, and Y0 is independent of σ(Ws : s ≥ 0).
The solution of the homogeneous equation is e−µt . Taking the derivative
of eµt Yt we obtain
which gives Z t
−µt µs
Yt = e Y0 + e σ dWs .
0
This is the Ornstein–Uhlenbeck process. The integral of a deterministic func-
tion with respect to SBM is Gaussian, thus
Yt − e−µt Y0
Taking the limit for the initial distribution Y0 we see that (Yt ) is Gaussian
and
σ2
Yt ∼ N 0, .
2µ
Next we determine the covariance function of Y . Since
Z t
−µt µu
Yt = e Y0 + σ e dWu
0
we get Z t
−µ(t−s) −µt
Yt − e Ys = e σ eµu dWu , t > s, (33) {eq:ou-fgt}
s
which is independent of σ(Wu : u ≤ s) σ. Therefore,
σ 2 −µ(t−s)
= e−µ(t−s) EYs2 = e ,
2µ
53
which depends only on t − s. That is (Yt ) is stationary.
Using formula (33) for A ∈ B(R)
P(Yt ∈ A|Yu : u ≤ s, Ys = x)
= P(Yt − e−µ(t−s) Ys ∈ A − e−µ(t−s) x|Yu : u ≤ s, Ys = x)
= P(Yt − e−µ(t−s) Ys ∈ A − e−µ(t−s) x).
σ2
−µt −2µt
pt (·|x) ∼ N e x, 1−e ,
2µ
∂ ∂ σ2 ∂ 2
ρt (y|x) = −µx ρt (y|x) + ρt (y|x),
∂t ∂x 2 ∂x2
which is called Fokker–Planck equation. The forward is
∂ ∂ σ2 ∂ 2
ρt (y|x) = − (−µyρt (y|x)) + ρt (y|x).
∂t ∂y 2 ∂y 2
54
where Wt is a standard Brownian motion. Thus rt is a translated Ornstein–
Uhlenbeck process. Indeed, Xt = rt − b satisfies
thus Z t
−at as
Xt = e X0 + e σdWs ,
0
from which Z t
−at as
rt = b + e r0 − b + e σdWs .
0
and variance
σ2
Var(rt ) = (1 − e−2at ).
2a
This implies that rt can take arbitrarily large negative values, which is not
very realistic.
Now we determine the distribution of P (t, T ). By (31)
Z T
P (t, T ) = E exp − ru du Ft
t
Z T
−b(T −t)
=e E exp − Xu du Ft ,
t
where X
e is the solution to the Langevin equation
es = −aX
dX es + σdWs , e0 = x0 = rt − b.
X (36) {eq:vasicek-initia
55
We have seen that (Xu ) is a continuous Gaussian process, therefore its inte-
gral is Gaussian too. Since EXeu = e−at x0 , we have
Z t Z t
x0
E X
eu du = x0 e−au du = (1 − e−at ).
0 0 a
Furthermore, for t ≥ s
Z t Z s
−at au −as
Cov(X
et , X
es ) = Ee σe dWu e σeau dWu
0 0
Z s 2
2 −a(t+s) au
=σ e E e dWu
0
Z s
2 −a(t+s)
=σ e e2au du
0
σ2
= e−a(t+s) e2as − 1 .
2a
Therefore
Z t Z t Z t
Var X
eu du = Cov X
eu du, Xeu du
0 0 0
Z t Z t
= E (Xu − EXu )dv (X
e e e v − EX ev )dv
Z t 0Z t 0
56
Substituting back into (35) and using the initial condition (36), we obtain
n rt − b
P (t, T ) = exp − b(T − t) − (1 − e−a(T −t) )
a
σ2 o
+ 3 a(T − t) − 3 + 4e−a(T −t) − e−2a(T −t) .
4a
The fair price of a European call option with strike price K at T1 for a
zero coupon bond with expiry T2 > T1 is
R T1
C(K; T1 , T2 ) = Ee− 0 rt dt
(P (T1 , T2 ) − K)+ . (37) {eq:vasicek-eucall
C(K; , T1 , T2 )
n V −b
= Ee−U exp − b(T2 − T1 ) − (1 − e−a(T2 −T1 ) )
a
!
σ2 o
+ 3 a(T2 − T1 ) − 3 + 4e−a(T2 −T1 ) − e−2a(T2 −T1 ) −K ,
4a
+
The main point here is that there exists an explicit formula, which can
be computed numerically easily.
57
10.2.3 Hull–White model
This is a simple generalization of the Vasicek model, where we allow the pa-
rameters to be time dependent. Assume that for some deterministic functions
a, b, and σ
58
The process
n Z t
X Xi (u)
Wt = √ dWi (u)
i=1 0 ru
is a continuous martingale, such that
Z t n Z t
2
X Xi (u)2
Wt = 2 Wu dWu + du
0 i=1 0 ru
Z t
=2 Wu dWu + t,
0
where
2γet(b+γ)/2
2
φu,v (t) = − 2 log
σ σ 2 u(eγt − 1) + γ − b + eγt (γ + b)
u(γ + b) + eγt (γ − b) + 2v(eγt − 1)
ψu,v (t) = 2 γt ,
σ u(e − 1) + γ − b + eγt (γ + b)
59
√
where γ = b2 + 2σ 2 v.
Therefore, using the result above and the Markov property the value of
the zero coupon bound
h RT i
P (t, T ) = E e− t ru du Ft
h R T −t i
= E e− 0 ru du r0 = rt
= exp {−aφ0,1 (T − t) − rt ψ0,1 (T − t)}
The price of a European call option with strike price K at T1 for a zero
coupon bound with expiry T2 > T1
C(K; T1 , T2 )
h R T1 i
− 0 ru du
=E e (exp {−aφ0,1 (T2 − T1 ) − rT1 ψ0,1 (T2 − T1 )} − K)+ .
This is not Ran explicit formula, but we now the joint Laplace transform of
T
the vector ( 0 1 ru du, rT1 ), therefore it is numerically computable.
P (t, T )
P (t, T ) − P (t, T + ε) = 0,
P (t, T + ε)
P (t, T )
= eεR(t,T,T +ε) ,
P (t, T + ε)
that is
1
R(t, T, T + ε) = − (log P (t, T + ε) − log P (t, T )) .
ε
60
Thus the instantaneous forward interest rate at time T calculated at time t,
called forward rate is
∂
f (t, T ) = lim R(t, T, T + ε) = − log P (t, T ). (40)
ε↓0 ∂T
Intuitively, it is clear that at time t we predict the interest at time t to
equal the short rate rt , that is rt = f (t, t). In what follows we prove this
statement.
{lemma:forward-sho
Lemma 10. For any t ∈ [0, T ]
f (t, t) = rt .
we obtain RT
P (t, T ) = e− t f (t,u)du
. (41) {eq:P-f}
Differentiating
∂
P (t, T ) = −f (t, T )P (t, T ),
∂T
which at t = T
∂
P (t, T ) = −f (t, t),
∂T T =t
On the other hand, differentiating
h RT i
P (t, T ) = E e− t ru du Ft
we obtain
∂ h
− tT ru du
R i
P (t, T ) = E −rT e Ft
∂T
which at T = t
∂
P (t, T ) = −rt ,
∂T T =t
and the statement follows.
61
10.3.2 The Heath–Jarrow–Morton model
The Heath–Jarrow–Morton (HJM) model describes the dynamic of the for-
ward rate f (t, T ) with the SDE
where Z T Z T
∗ ∗
α (t, T ) = α(t, u)du, σ (t, T ) = σ(t, u)du.
t t
Here we use a stochastic version of Fubini’s theorem, which we did not even
formulate. Put Z T
Xt = log P (t, T ) = − f (t, u)du.
t
Then the above calculation gives
Thus
1 ∗
dP (t, T ) = eXt ∗
rt − α (t, T ) + σ (t, T ) dt − eXt σ ∗ (t, T )dWt
2
2
∗ 1 ∗ 2 ∗
= P (t, T ) rt − α (t, T ) + σ (t, T ) dt − σ (t, T )dWt .
2
62
is a martingale. Since
Rt Rt Rt
d e− 0 ru du P (t, T ) = e− 0 ru du dP (t, T ) − rt e− 0 ru du P (t, T )dt
− 0t ru du
R
∗ 1 ∗ 2 ∗
=e P (t, T ) −α (t, T ) + σ (t, T ) dt − σ (t, T )dWt ,
2
Theorem 19. If the HJM model is determined by the SDE (42) then neces-
sarily (43) holds.
References
[1] R. J. Elliott and P. E. Kopp. Mathematics of financial markets. Springer
Finance. Springer-Verlag, New York, second edition, 2005.
63