Interest rates Models, Lecture Notes

Nicolas Gaussel

December 19, 2024

Abstract
This lecture follows closely the structure and notations of [1]. Work

in progress; made available for students only. Please do not forward. Do

not quote.

Contents
1 Notations and fundamental concepts 2
1.1 Notation review . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Interest rates swaps . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Fundamental law of pricing 6

2.1 European pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Pricing trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Conditional expectation, change of mesures and numeraires 7

3.1 Conditional expectation as a projection in L2 . . . . . . . . . . . 7
3.2 Caracterisation of Q-martingales . . . . . . . . . . . . . . . . . . 9
3.3 Fundamental pricing measures and numeraires . . . . . . . . . . 11

4 The Lognormal Forward-LIBOR Model (LFM) 14

4.1 The Black formula for caplets . . . . . . . . . . . . . . . . . . . . 14
4.2 The Lognormal-Forward-Libor . . . . . . . . . . . . . . . . . . . 15

5 The Vasicek model 18

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.2 Solving the SDE . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.3 The integrated short-term rate . . . . . . . . . . . . . . . . . . . 21
5.4 The Bond price as an expectation . . . . . . . . . . . . . . . . . . 22
5.5 A PDE approach to compute the Bond price . . . . . . . . . . . 23
5.6 A forward rate approach to compute the bond price . . . . . . . 27

6 n-factor Hull-White model 29

6.1 The short-rate dynamics . . . . . . . . . . . . . . . . . . . . . . . 29
6.2 Pricing of a ZC bond . . . . . . . . . . . . . . . . . . . . . . . . . 30

1
1 Notations and fundamental concepts
1.1 Notation review
ˆ instantaneous short-term rate: rt
R 
t
ˆ Bank account: B (t) = exp r ds
0 s

B(t)
ˆ Stochastic Discount Factor: D (t, T ) = B(T )

ˆ Zero-coupon bond: P (t, T ) = E (D (t, T ) |Ft ) ; P (T, T ) = 1.

ˆ Day coounting conventions between as time t and T encapsulated as
τ (t, T ). Example of dierence between act/360 and act/365.

ˆ Continuously compunded rate: R (t, T ) := − ln(P τ(t,T ))

−1
ˆ Simply compounded rate, Libor rate L (t, S) such that P (t, S) = (1 + τ L (t, S))

Denition 1. The zero-coupon curve (sometimes also referred to as yield

curve) at time t is the graph of the function

(
L(t, T ) t < T ≤ t + 1(Short Term)
T →
Y (t, T ) T > t + 1(Long Term)

Note that money market rate (T<1) are simply compunded and others are
continuously compounded.

Denition 2. A Forward Rate Agreement (FRA) is a contract involving three

time instants: The current time t, the expiry time T > t, and the maturity time
S > T. The contract gives its holder an interest-rate payment for the period
between T and S. At the maturity S, a xed payment based on a xed rate K
is exchanged against a oating payment based on the spot rate L(T, S) resetting
in T and with maturity S . Basically, a FRA allows one to lock-in the interest
rate between times T S at a desired value K , with the rates in the contract
and
that are simply compounded. The total value of a FRA at time S is therefore

F RA (T, S) = N τ (T, S) (K − L (T, S))

The pricing of the xed leg, NτK paid at date S is straightforward and
equal to N τ KP (T, S) .
The pricing of the oating leg, τ L, involves a No Arbitrage reasoning. A
strategy to receive the libor rate in S consists in having some have cash at
T and redeem it at S which means that the NPV of N τ L (T, S) is equal to
N (P (t, T ) − P (t, S)) . This is sometimes referred to as a oating leg always
quotes at par.
The level K which ensures the FRA to be fair is called the forward rate
dened by:

2
 τ F (t, T, S) × P (t, S) = P (t, T ) − P (t, S) (1)
 
1 P (t, T )
F (t, T, S) := −1
τ P (t, S)

Obviously, F (T, T, S) = L (T, S)

One can dene the instantaneous forward rate: f (t, T ) := lim+ F (t, T, S) =
S→T
− ∂lnP∂T(t,T ) . Using the previous formula, the zero-coupon bond can be com-
puted/interpreted as the accumulation of forward rates:

!
Z T
P (t, T ) = exp − f (t, u) du
t

1.2 Interest rates swaps

We have just considered a FRA, which is a particular contract whose fairness
can be invoked to dene forward rates. A generalization of the FRA is the
Interest-Rate Swap (IRS). A prototypical Payer (Forward-start) InterestRate
Swap (PFS) is a contract that exchanges payments between two dierently
indexed legs, starting from a future time instant. At every instant Ti in a
prespecied set of dates Tα+1 , ..., Tβ the xed leg pays out the amount

N τi K

, corresponding to a xed interest rate K , a nominal value N and a year fraction

τi between Ti−1 and T_, whereas the oating leg pays the amount

N τi L(Ti−1 , Ti ),

corresponding to the interest rate L(Ti−1 , Ti ) resetting at the previous in-

stant T_ for the maturity given by the current payment instant Ti , with Tα a
given date. Clearly, the oating-leg rate resets at dates Tα , Tα+1 , ..., Tβ−1 and
pays at dates Tα+1 , ..., Tβ . We set T := Tα , ..., Tβ and τ := τα+1 , ..., τβ .
When the xed leg is paid and the oating leg is received the IRS is termed
Payer IRS (PFS), whereas in the other case we have a Receiver IRS (RFS).
discounted payoat a time t < Tα
PβThe of a RFS can be expressed as

i=α+1 D(t, Ti )N τi (K − L(Ti−1 , Ti )).

We can view this last contract as a portfolio of FRAs, we can value each

3
FRA through formulas 1 and then add up the resulting values. We thus obtain

β
X
RF S(t, T, τ, N, K) = F RA(t, Ti−1 , Ti , τi , N, K)
i=α+1
β
X
= τi P (t, Ti )(K − F (t; Ti−1 , Ti )). (2)
i=α+1
β
X
= −N P (t, Tα ) + N P (t, Tβ ) + N τi KP (t, Ti ) (3)
i=α+1

The two legs of an IRS can be seen as two fundamental prototypical contracts.
The xed leg can be thought of as a coupon-bearing bond, and the oating
leg can be thought of as a oating-rate note. An IRS can then be viewed as a
contract for exchanging the coupon-bearing bond for the oating-rate note that
are dened as

Denition 3. 1.5.3. The forward swap rate Sα,β (t) at time t for the sets of
times T and year fractions τ is the rate in the xed leg of the above IRS that
makes the IRS a fair contract at the present time, i.e., it is the xed rate K for
which RF S(t, T, τ, N, K) = 0. We easily obtain :

P (t, Tα ) − P (t, Tβ )
Sα,β (t) = Pβ (4)
i=α+1 τi P (t, Ti )

Let us divide both the numerator and the denominator in 4 by P (t, Tα ) and
notice that the denition of F in terms of P 's implies

k k
P (t, Tk ) Y P (t, Tj ) Y 1
= =
P (t, Tα ) j=α+1 P (t, Tj−1 ) j=α+1 1 + τj Fj (t)

for all k > α, where we have set Fj (t) = F (t; Tj−1 , Tj ). Formula 4 can then be
written in terms of forward rates as
Qβ 1
1− j=α+1 1+τj Fj (t)
Sα,β( t) = Pβ Qi 1
i=α+1 τi j=α+1 1+τj Fj (t)

1.3 Options
1.3.1 Caps, oors
A cap is a contract that can be viewed as a payer IRS where each exchange
payment is executed only if it has positive value. The cap discounted payo is
therefore given by

β
X
D(t, Ti )N τi (L(Ti−1 , Ti ) − K)+
i=α+1

4
Analogously, a oor is equivalent to a receiver IRS where each exchange pay-
ment is executed only if it has positive value. The oor discounted payo is
therefore given by

β
X
D(t, Ti )N τi (K − L(Ti−1 , Ti ))+
i=α+1

Each D(t, Ti )N τi (L(Ti−1 , Ti ) − K)+ denes a contract that is termed caplet.

The oorlet contracts are dened in an analogous way.

1.3.2 Swaptions
A European payer swaption is an option giving the right (and no obligation)
to enter a payer IRS at a given future time, the swaption maturity. Usually
the swaption maturity coincides with the rst reset date of the underlying IRS.
The underlying-IRS length (Tβ − Tα in our notation) is called the tenor of the
swaption. The payer-swaption payo, discounted from the maturity Tα to the
current time, is thus equal to

β
!+
X
N D (t, Tα ) τi P (t, Ti )(K − F (t; Ti−1 , Ti )
i=α+1

Contrary to the cap case, this payocannot be decomposed in more ele-

mentary products, and this is a fundamental dierence between the two main
interest-rate derivatives. Indeed, we have seen that caps can be decomposed
into the sum of the underlying caplets, each depending on a single forward rate.
One can deal with each caplet separately, deriving results that can be nally
put together to obtain results on the cap. The same, however, does not hold for
swaptions. From an algebraic point of view, this is essentially due to the fact
that the summation is inside the positive part operator, (···)+, and not outside
like in the cap case. Since the positive part operator is not distributive with
respect to sums, but is a piece-wise linear and convex function, we have

β
!+
X
P (Tα , Ti )τi (F (Tα ; Ti−1 , Ti ) − K)
i=α+1
β
X
≤ P (Tα , Ti )τi (F (Tα ; Ti−1 , Ti ) − K)+
i=α+1

with no equality in general, so that the additive decomposition is not feasible.

As a consequence, in order to value and manage swaptions contracts, we will
need to consider the joint action of the rates involved in the contract payo.
From a mathematical point of view, this implies that, contrary to the cap case,
correlation between dierent rates could be fundamental in handling swaptions.

5
2 Fundamental law of pricing
2.1 European pricing
2.1.1 Set-up
European pricing refers to pricing claims that entail a single (random) ow at
date T seen from date 0. One can think about the set of all possible claims as all
↕X = L2 (Ω, F, P) .
possible european options payos on all possible underlyings
To be specic, we can restrict ourselves to the single asset case where Ω is
2
generated by all possible values of a stock with value S̃T and L (Ω), in that
 
case, is just the set of all possible f S̃T with nite variance. It is also supposed

that there is a zero-coupon bond i.e. a security which pays 1 at date T and which
value today is equal to P (0, T ) .

2.1.2 Cash and carry and the value of forward contracts

A forward contract is a contract which entails no ow at date 0 and where two
parties agree to exchange S̃T (which is random) against a value which is known
today which we denote FT . At maturity, the payo of that contract is S̃T − FT .
Remarkably, the strategy which consists in buying the stock at date 0 by
borrowing an amount S0 will provide a T value equal to S̃T − S0 /P (0, T ):
1. S̃T , because we can sell the stock that has been bought at date 0,

2. −S0 /P (0, T ) because we have to redeem the loan made to buy the stocks.
This strategy is called the cash and carry strategy.
Now, entering a cash-and-carry strategy and selling a foward contract
would entail no initial price and generate a payo at date T equal to FT −
S0 /P (0, T ) . This strategy is an arbitrage if FT − S0 /P (0, T ) > 0. In the
opposite case, the opposite strategy is an arbitrage.
Hence, if we assume the absence of arbitrage opportunities, the only possible
value for FT is to be equal to S0 /P (0, T ) .

2.1.3 Pricing measure

Interestingly, the fact that forward contract have a known price creates a con-
straint on candidate pricing schemes. If we look for linear pricing functionnals,
on X we are looking for all element of possible element of X ∗. It is well known
that all possible pricing functional can be described as

π0 (XT ) = kT E QT (XT )
where kT is a constant and QT a measure with sum 1, not necessarily positive.

QT is a probability measure: since all claims which are strictly positive

must have a strictly positive price then QT cannot measure negatively
some sets.

6
kT = P (0, T ): by denition of the price of 1

The price of forward contracts is nul:

 
E QT S̃T − S0 /P (0, T ) = 0 The

consequence of the numm price of forward contracts is that, under any candidate
pricing measure:
 
E S̃T = S0 /P (0, T )
= S0 erT
where r is the zero-coupon rate at date T seen from date 0.

2.2 Pricing trajectories

Instead of invoking cash and carry between dates 0 and T, one can think about
a theoretical instaneneous forward contract. That forward contract would give
a nul price to
Bt+dt
St+dt − St
Bt
Since Bt+dt is known at date t (no quadratic term in dB ) then, equivalently,
under the pricing measure Q
 
St+dt St
E − |Ft =0
Bt+dt Bt
 
St St+dt
⇔ = EQ |Ft
Bt Bt+dt
 
St ST
⇔∀T : = EQ |Ft
Bt BT
This new pricing measure does not depend on any specic maturity T. It is
called the risk neutral measure. The key theorem to remind is that:

Xt
Bt is a martingale under the RNP. Prices expressed in units of
bank account are martingale.
As a corollary, whichever non dividend paying price process Xt has to satisfy

Xt = E (D (t, T ) XT |Ft ) (5)

3 Conditional expectation, change of mesures and

numeraires
3.1 Conditional expectation as a projection in L 2

Lp are Banach spaces. F = L2 (Ω, F, P) is even a Hilbert space i.e. a Banach

space with a scalar product

⟨X, Y ⟩ = E (XY )

7
 Let G⊂F and accordingly G = L2 (Ω, G, P) ⊂ F . As F is a Hilbert space,
the projection on G is uniquely dened as

∀X ∈ F : ∃!p (X) ∈ G |
∀Y ∈ G : ⟨X − p (X) , Y ⟩ = 0
⇐⇒ ∀Y ∈ G : ⟨X, Y ⟩ = ⟨p (X) , Y ⟩

In probabilistic terms
p (X) = E (X|G)
dQ
Now let Q be a measure equivalent to Pwith density Z = dP . One can dene
another scalar product

⟨X, Y ⟩Q = E Q (XY )
= E (ZXY )
= ⟨ZX, Y ⟩

One can dene the projection on G w.r.t. to the Q-scalar product. In the
same manner. The two projections can be related in the following manner:

∀Y ∈ G : ⟨X − q (X) , Y ⟩Q = 0
⇐⇒ ∀Y ∈ G : ⟨ZX − Zq (X) , Y ⟩ = 0
⇐⇒ ∀Y ∈ G : ⟨ZX, Y ⟩ = ⟨Z, q (X) Y ⟩
⇐⇒ ∀Y ∈ G : ⟨p (ZX) , Y ⟩ = ⟨p (Z) q (X) , Y ⟩

Now since p (Z) q (X) is G measurable and the projection is unique, it is

equal to the projection of ZX which yields the nal result:

p (ZX)
q (X) = .
p (Z)

Put in probabilistic terms:

Theorem 1. Let P be a probabiliy and Q and equivalent probability with density

Z= dQ
dP . One has:
dQ
1. ∀G : dP G = E (Z|G)
E(XZ|G)
2. ∀G : E Q (X|G) = E(Z|G)

Theorem 1 is sometimes referred to as the Bayes rule. It is useful to think to

about in the context of a projection.

8
 Figure 1: Conditional expectation as a projection

3.2 Caracterisation of Q-martingales

3.2.1 Change of measures
Let us now apply the previous result in the context of a a ltered space with
ltration Ft . In this context we introduce

Zt = E (Z|Ft )
ZT = Z
Z0 = 1

Zt is a P martingale. Q measures are generally characterized by their density

process. Since this density process has to be positive it is itself looked after
under the form of a so called Doleans-Dade

dZt = Zt λt dWt
Z0 =1
Rt Rt
Let Yt = 0
λs dW s and ⟨Y ⟩t = λ2s ds
0
 
1
Zt = exp Yt − ⟨Y ⟩t
2
≡ E (λt )

E (λt ) is a convenient notation to avoid lengthy equations. When λt is deter-

ministic, Yt is a gaussian variable with zero mean and variance ⟨Y ⟩t .
Let Xt be a Q-martingale. The Bayes rule yields
E (ZX|Ft )
Xt =
Zt
⇐⇒ Xt Zt = E (ZT XT |Ft )

or equivalently

9
Theorem 2. X. is a Q martingale i X.Z. is a P martingale
This theorem is used often times to caracterise Q martingales. One typically
applies Ito to XZ and equals the nite variation (dt) part to zero. Then for X
to be a Q−martingale

d (XZ) = XdZ + ZdX + d ⟨X, Z⟩

 
d ⟨X, Z⟩
= XdZ + Z dX +
Z

Since Z is a P martingale, we need − d⟨X,Z⟩

Z to be the trend of X under P.
This result can be formulated as one of the dierent versions of the Girsanov
theorem:

Theorem 3 (Girsanov theorem) . Let M a P martingale, Z the density of a

measure Q. Let
Zt = E (Z | Ft )

t
d ⟨M, Z⟩s
Z
Nt ≡ M t −
0 Zs
Z t
= Mt − d ⟨M, ln (Z)⟩s
0

is a Q-martingale.
Proof. Nt ≡ Mt + Λt is a Q martingale i NZ is a P martingale. We have

d (N Z) = d ((M + Λ) Z)
= (M + Λ) dZ + ZdM + ZdΛ + d ⟨M, Z⟩

NZ is a martingale i

ZdΛ + d ⟨M, Z⟩ = 0
d ⟨M, Z⟩
⇔dΛ = −
Z
Z t
d ⟨M, Z⟩s
⇔Nt ≡ Mt −
0 Zs

Corollary 1. Zt being positive; there exists an adapted process λ satisfying the

Novikov condition such that
dZt = Zt λt dWt
Z0 = 1

10
then
Z t
Nt = M t − d ⟨M, ln (Z)⟩s
0
Z t
= Mt − λs ds
0

3.2.2 Change of measures AND numeraire

In asset pricing, changes of measures also come with changes of numeraires so
the following corrollary is often used

Corollary 2. Let N 1 and N 2 two numeraires and Q1 and Q2 two measures with
density Z = dQ2
dQ1 . If X
N1 is a Q1 −martingale and X
N2 is a Q2 −martingale then
N01 Nt2
Zt = (6)
N02 Nt1

3.3 Fundamental pricing measures and numeraires

3.3.1 Forward neutral probabilities and ZC numeraire
Xt
Under the RNP, the dynamics to be modelled is
Bt while the nal payo is
XT
only related to XT . This raises some issues since
BT and XT cannot be easily
related. The T − f orward measure provides a simple and elegant way to handle
this diculty. The fundamental pricing equation 5 at date 0

X0 = E (D (0, T ) XT |F0 ) ,
suggests that prices are indeed related to the expectation of the nal value of
the payo but under a dierent probability QT with density

dQT D (0, T )
ZT = ≡
dQ P (0, T )
Since P (0, T ) = E (D (0, T )), ZT is always positive and has mass 1. One can
check that
D (0, t)
Zt = P (t, T )
P (0, T )
and that, using theorem 1,
 
Xt XT
= ET = E T (XT )
P (t, T ) P (T, T )
This show that prices expressed in numeraire P(t, T ) are martingales under the
Xt
T − f orward measure. Under this measure, the nal value of
P (t,T ) is equal
Xt
to XT . If one disregards the fact that it is unlikely that the volatility
P (t,T ) is
Xt
in general time homogeneous, a B&S model can be used for
P (t,T ) . This is the
path followed to derive the Black formula for caplets.

11
3.3.2 Foreign neutral probability, currency numeraire and the quanto
eect
Framework Let S f a foreign asset denominated in a foreign currency which
value in domestic currency is denoted Xt . All usual quantities with exponent f
refer to foreign quantities. Typically, Btf is the foreign bank account and Bt is
the local bank account.
Sf f
In foreign currency is a Q martingale under the foreign RNP. If we
Bf
make the assumption that the price of any foreign asset can always be changed
f f Sf X f
without any friction, S X can be seen as a local asset and hence, is a
B t
Q-martingale. The density of the local probability with respect to the foreign
one is thus
Bt
Zt =
Xtf Btf

Dynamics of the exchange rate under the RPN Let

Bf X f
Sf = Bf .
Bt
X f
is a Q-martingale so under Q, the drift µ of X is necessarily equal to the
dierence of short term rates:

Bf X f B f dX f X f dB f
 
1
d = + + X f Bf d
Bt Bt Bt Bt
Bf X f  X  Bf X f
= µt + rtf − rt dt + σX dWt
Bt Bt
and hence
f
µX
t = rt − rt

Dynamics of foreign assets Foreign assets can be modelled as

 
dStf = Stf rtf dt + σS dWtf

S0f = 1

with d W f , W = ρdt. Two methods can be followed.

Sf X f
First one,
Bt is a Q martingale so it has no trend under Q. Let µS the
f
trend of S under Q :

Sf X f S f dX f X f dS f
 
f f 1 1
d = + +X S d + d Sf , X f
Bt Bt Bt Bt Bt
Sf X f  S
µt + rt − rft − rt + ρσS σX dt + σS dWt2 + σdWt


=
Bt
where W2 is a Q-SBM. Necessarily

µSt = rtf − ρσS σX

under Q.

12
 Second one. Let
d W f , Zt
dWt2 = dWtf −
Zt
is a Q-SBM.

 
1
dZt = d
Yt
dYt d ⟨Y ⟩t
= −Zt − Zt
Yt Yt2
2


= Zt −σX dt − σX dWt

and hence
d W f , Zt
= −ρσX dt
Zt
hence dWt2 = dWtf + ρσX dt is a Q-SBM and we can rewrite

 
dStf = Stf rtf dt + σS dWtf
  
= Stf rtf − ρσS σX dt + σS dWt2

and do proper computation under Q.

Example with the value of the Nasdaq paid in 1 year time, but in
EUR Let Yt the value of the Nasdaq paid at date T. One has
Yt = N0 e(rf −r−ρσS σX )(T −t)

A positive covariance between Nasdaq and currency lowers the price. Because
locally if Nasdad goes up, the variation is unhedged. So, if at the same time,
the currency goes up as well, this creates an instantaneous gain which has to be
inputed into the price.

3.3.3 The Libor Swap Models

Remember that the forward swap rate has been dened as

P (t, Tα ) − P (t, Tβ )
Sα,β (t) = Pβ
i=α+1 τi P (t, Ti )

Pβ
Let Cαβ (t) ≡ i=α+1 τi P (t, Ti ) . Cαβ (t) is a portfolio of ZC which provides an
adequate numeraire to value the swap rate. Sα,β (t)Cα,β (t) is indeed a quoted
asset and hence its discounted price shall be a martingale.
To be continued...

13
4 The Lognormal Forward-LIBOR Model (LFM)
The LFM sometimes referred to as BGM (Brace Gatarek Musiela) is a rigor-
ous framework to account for a common practice. Before market models were
introduced, there was no interest-rate dynamics compatible with either Black's
formula for caps or Black's formula for swaptions. These formulas were actually
based on mimicking the Black and Scholes model for stock options under some
simplifying and inexact assumptions on the interest-rates distributions. The
introduction of market models provided a new derivation of Black's formulas
based on rigorous interest-rate dynamics.

4.1 The Black formula for caplets

4.1.1 Pricing of one caplet
A caplet on L (T, S) involves the computation of the expected value of either

B (t) +
× (L (T, S) − K)
B (S)
under the RNP or
+
(L (T, S) − K)
under the S-Forward probability. Remind that

τ P (t, S) × F (t, T, S) = P (t, T ) − P (t, S)

and hence, P (t, S) F (t, T, S) is equal to the dierence of two ZC bonds. As
such F (t, T, S) = P (t,S)F (t,T,S)
P (t,S)
S
is a Q martingale. For notational simplicity

note Ft ≡ F (t, T, S). One can model F as the solution of

dFt = Ft σt dWt
F0 = F (0, T, S)
or equivalently
Ft = F (0, T, S) E (σt )
S
In such model, under Q FT is a Gaussian variable with mean F (0, T, S) and
RT
average variance v (T − t) ≡ t σ 2 ds. Let BS (S, K, r, σ, T ) the standard B&S
formula. One has

+
 √
E S (L (T, S) − K)


= BS F0 , K, 0, v, T

which yields the price of the corresponding caplet as

√

caplet (T, S) = P (0, S) BS F0 , K, 0, v, T
It has to be noted that the uncertainty goes until T and the payment is done
at date S . The time dierence between, depending on the context, is sometimes
called natural. Paying at a dierent time involves another forward measure
and hence necesitates a drift adjustment.

14
4.1.2 Pricing of a cap
It is market practice to price a cap with the following sum of Black's formulas
(at time zero)

β
X
CapBlack (N, T , N, K, σα,β ) = N P (0, Ti )BS (F (0, Ti−1 , Ti ), K, 0, σα,β , τi )
i=α+1

where, the common volatility parameter σα,β is retrieved from market quotes.

4.2 The Lognormal-Forward-Libor

4.2.1 Payos that cannot be handled with Black-formulas
The problem with the Black formula is that it:

1. cannot handle the joint dynamics of two Forward Libor rate and,

2. it can neither account for the expectation of the price of a libor rate paid
at a date dierent from its natural date.

Example1: ratchet payos Ratchet payos are payos which involve the
dierence between two Libor rates. Typically, one might be interested in receiv-
ing a function of

∆ (Ti−1 , Ti−2 ) ≡ L (Ti−1 , Ti ) − L (Ti−2 , Ti−1 )

at date Ti . The problem here is that the dynamics of ∆ involves the knowledge
of both L (Ti−1 , Ti ) and L (Ti−2 , Ti−1 ).
Example2: in arrear pay-os In arrear payos generally involve the pay-
ment of L (T, S) at a date which is not S ! It can be T or any other date but it
involves the dynamics of L (T, S) under a numeraire which is not P (t, S) .
As a matter of fact, it is necessary to handle the joint dynamics of ALL
forward rates dened on a specic ladder.

4.2.2 Model notations and parameters

We assume a given set of dates T and denote Fk (t) ≡ F (t, Tk−1 , Tk ) the k th
k
natural forward LIBOR rate. We have seen before that Fk (t) is a Q martin-
gale where Qk dened by its conditional density w.r.t. to Q, P (t, Tk ).
Under Qk
dFk (t) = Fk (t) σk (t) dZk (t)
k
where Zk is a Q SBM. We introduce β (t) = m if Tm−2 < t < Tm−1 which is
the maturity of the rst forward which has not expired. Reciprocally

 
t ∈ Tβ(t)−2 , Tβ(t)−1 .

15
The volatility σk (t) is chosen to be piecewise-constant i.e.:

σk (t) = σk,β(t)

This correlation matrix has to be specied (see discussion Brigo-Mercurio) page

210-211. As a popular example

V ols
F1 ϕ1 ψ1 dead ... dead
F2 ϕ2 ψ2 dead dead
...
FM ϕ M ψ 1 ϕM ψM

d ⟨Zi , Zj ⟩ = ρij dt
One can use a Black formula to price each related caplets but what about
pricing all maturities in the same model ?
4.2.3 Drift adjustments
Choose one given Qk under which Fk is a martingale. What is the dynamics of
Fk under Qi ?
Let start with the case Tk < Ti . We have

dQi D (0, Ti )
=
dQ P (0, Ti )
dQk D (0, Tk )
=
dQ P (0, Tk )
dQi P (0, Tk )
⇒ HT ≡ = × D (Tk , Ti )
dQk P (0, Ti )

Now we have to compute Ht ≡ E k (HT |Ft ). For this we recognize that

P (t, Ti ) = E (D (t, Ti ) |Ft )

= E (D (t, Tk ) D (Tk , Ti ) |Ft )
= P (t, Tk ) E k (D (Tk , Ti ) |Ft )

from which we obtain that:

P (0, Tk ) P (t, Ti )
Ht =
P (0, Ti ) P (t, Tk )

P (0, Tk ) 1
= × Qi
P (0, Ti ) j=k+1 (1 + τj Fj )

Now Girsanov theorem states that

dZki = dZk − mi dt

16
is a Qi -SBM i

mi dt = d ⟨Zk , ln (H)⟩t
* i
+
X
= −d Zk , ln (1 + τj Fj )
j=k+1
i
X τj
=− d ⟨Zk , Fj ⟩
1 + τj Fj
j=k+1
i
X τj Fj (t)
=− ρjk σj dt
1 + τj Fj (t)
j=k+1

where all variables are parameters except Fj (t) which is known at date t. All
can be seen as quanto adjustments.
As a result, under Qi
 
i
X τj Fj (t)
dFk = σk Fk × dZki + ρjk σj dt
1 + τj Fj (t)
j=k+1

where Zki is a Qi − SBM.

In the case when , we have

k
P (0, Tk ) Y
Ht = × (1 + τj Fj )
P (0, Ti ) j=i+1

τj Fj (t)
Pk
and hence the drift is equal to −
j=i+1 1+τj Fj (t) ρjk σj dt.
Note that the drift is time dependent due to the presence of Fj (t) so no
closed form solutions can be expected and all expectations have to be computed
by Monte-Carlo simulations.
Yet, some good approximation can be obtained by freezing Fj (t) to its initial
value Fj (0). We will see that in the following example.

4.2.4 Pricing of specic claims in LFM

Ratchet payos
+
E (D (0, Ti ) [L(Ti−1 , Ti ) − L(Ti−2 , Ti−1 ) + X)]) =
+
P (0, Ti ) E i [Fi (Ti−1 ) − Fi−1 (Ti−2 ) − X]

where all has been written under P (t, Ti ) numeraire and the Ti forward
probability. Under this probability, using the drift adjustment formulas:

dFi (t) = Fi (t) σi (t) dZi (t)

17
and

dFi−1 (t) = Fi−1 (t) σi−1 (t) dZi−1 (t)

i
 
= Fi−1 (t) mi−1 dt + σi−1 (t) dZi−1 (t)
i
where Zi−1 is not a Qi − SBM but Zi−1 is and

ρi−1,i τi σi Fi (t)
mi−1 (t) = −
1 + τi Fi (t)

At that stage, the only possible way to move forward is to envisage a Monte-
Carlo simulation of (Fi−1 , Fi ) . It is also possible to envisage an approximation
of the drift which consists in freezing its value at date 0. The approximation is
likely to be good as long as ρi−1,i τi σi is not too large. under this approximation:

mi−1 (t) ≈ mi−1 (0)

ρi−1,i τi σi Fi (0)
=−
1 + τi Fi (0)

and Fi−1 also follows a log-normal dynamics. Ratchet payos can be approxi-
mated by spread payos and formula similar to Margrabe can be used.

In arrear payos and convexity adjustments So called in arrear payos

involve the payment of L (T, S) at date T instead of date S. This can be
expressed as an S-payo:

    
B (t) B (t) 1
E τ L (T, S) E −1
B (T ) B (T ) P (T, S)
  
B (t) 1
=E − P (t, T )
B (S) P (T, S) P (T, S)
 
1
= P (t, S) E S − P (t, T )
P 2 (T, S)
 
2
= P (t, S) E S (1 + τ FS (T )) − P (t, T )
= P (t, S) − P (t, T ) + P (t, S) 2τ FS (t) + τ 2 E S FS2 (T )

Now, FS being log-normal, so is FS2 and E S FS2 (T ) = FS (t) exp vS2 where
RS
vS = t
σu2 du

5 The Vasicek model

5.1 Introduction
Historically, the objective of interest rates model was to relate the bond prices
to the dynamics of the interest rates since, by NA, the price of 1 paid at T is

18
the expected value of the bank account (equation 5 with XT = 1). The puzzle
here is not only to price derivatives but also to provide a normative answer to
the possible shapes of the yield curve.
There was a series of model developped in years 80, see Table 3.1 of [1]
provides with a nice overview of dierent models. Before presenting the most
famous We present here the most famous and infuential model developped by
Vasicek, [3], we expose the solution of an ane family of SDEs which might be
useful to tackle other concurrent models.

5.2 Solving the SDE

Vasicek assumed that the instantaneous spot rate under the real-world measure
evolves as an Ornstein-Uhlenbeck process with constant coecients. This does
not say anything a priori about the dynamics under the risk neutral probability
but Vasicek is looking for a model which is mean-reverting under both measures.
Under the RNP

dr (t) = k (θ − r (t)) dt + σ dWt (7)

r (0) = r0 (8)

where Wt is a SBM. The standard deviation parameter, σ , determines the

volatility of the interest rate. The typical parameters θ, k and σ , together
with the initial condition r0 , completely characterize the dynamics, and can be
quickly characterized as follows, assuming a to be non-negative:

ˆ θ: "long term mean level". All future trajectories of r will evolve around
a mean level θ in the long run;

ˆ k: "speed of reversion". a characterizes the velocity at which such trajec-

tories will regroup around θ in time;

ˆ σ: "instantaneous volatility", measures instant by instant the amplitude

of randomness entering the system.

To solve the OU process one rst introduces

X (t) ≡ r (t) − θ
dX (t) = −kX (t) dt + σdWt
X (0) = r0 − θ

and then

Y (t) = X (t) ekt

dY (t) = σekt dWt
Y (0) = r0 − θ

19
 Rt
Dene I (t) ≡ e−kt t eks dWs . As a stochastic integral with deterministic inte-
grand, I(t) is a Gaussian process with zero mean and variance
1
1 − e−2kt


V (t) ≡ V (I (t)) =
2k
One has

Y (t) = r0 − θ + σI (t)
r (t) = θ + (r0 − θ) e−kt + σI (t) (9)
Z t
= θ + (r0 − θ) e−kt + σ e−k(t−s) dWs (10)
0

which solves the SDE. The nice thing with introducing I (t) is that it contains
all the randomness of r (t) and hence all calculations related to variance can be
made on I (t).
Thanks to equations 9 and 10. One can easily the following proposition:

Z t
rt − θ = (r0 − θ) e−kt + σ e−k(t−s) dWs
0
Z T
rT − θ = (r0 − θ) e−kT + σ e−k(T −s) dWs
0
 Z t  Z T
k(T −t) −kt −k(t−s)
=e (r0 − θ) e +σ e dWs + σ e−k(T −s) dWs
0 t
Z T
= e−k(T −t) (rt − θ) + σ e−k(T −s) dWs
t
| {z }
I(t,T )

Proposition 1. r (t) is a Gaussian process with the following properties

r (T ) = θ + (rt − θ) e−k(T −t) + σI (t, T ) (11)
−k(T −t)
E (rT |rt ) = θ + (rt − θ) e (12)
2
σ  
V ar (rT |rt ) = V (t, T ) = 1 − e−2k(T −t) (13)
2k
γ (t, s, u) = cov (rs , ru |rt ) (14)

= E (I (t, s) I (t, u) |r (t)) (15)

2
σ −k(s+u) 2ku∧s 2kt


= e e −e (16)
2k
where u ∧ s, the covariance function of the brownian motion, is a short cut for
min (u, s) .

Comments Proposition 1 allows to draw some insights on the short-term

rates dynamics as modelled by a OU process.

20
 ˆ When
1
k then V ar (rT |rt )
kt << 1 i.e. t << ≈ T − t which shows that the
process behaves like a brownian motion.

σ2
ˆ When kt >> 1 i.e.
1
k then V ar (rT |rt ) → 2k which shows that the
t >>
process variance is capped and the return to the mean eect dominates
the diusion part.

ˆ The initial value of the short-term rate, r (t) is forgotten at an expo-

nential speed and the long-term expected value of the process is equal to
θ.
ˆ r can be negative

5.3 The integrated short-term rate

The bond price P (t, T ) is the conditional expected value of the integrated short-
term rate:
 
P (t, T ) = E e−ξ(t,T ) |Ft
Z T
ξ (t, T ) = r (s) ds
t

Since r (s) is a Gaussian process, ξ (t, T ) is also a Gaussian process. The bond
price will be obtained once the expected value and variance of ξ (t, T ) have been
computed.
The expected value calculation is a straightforward integration of equation
11:

ξ¯ (t, T ) = E (ξ (t, T ) |rt ) = θ (T − t) + (r (t) − θ) A (t, T )

1 
A (t, T ) = 1 − e−k(T −t)
k
The computation of the variance of R (t) is a little more involved:

V ar (ξ (t, T ) |rt ) = cov (ξ (t, T ) , ξ (t, T ))

Z T Z T !
=E I (t, s) ds I (t, u) du|rt
t t
Z T Z T
= E (I (t, u) I (t, s) |rt ) duds
t t
Z T Z T
= γ (t, s, u) dsdu
t t
2 Z T Z T
σ
e−k(s+u) e2ku∧s − e2kt dsdu


=
2k t t

The u ∧ s function has to be determined. If one representes the square in which

u and s evolve, s being the x-axis and u the y-axis, then u ∧ s = s below the

21
rst diagonal and u∧s = u above. The integrand being perfectly symmetric on
u and s, the integral is equal to twice the value of the integration over one of
the half spaces. Hence:

T u
σ2
Z Z 
e−k(s+u) e2ks − e2kt ds du


V (t, T ) = V ar (ξ (t, T ) |rt ) =
k t t
T u
σ2
Z Z 
−ku ks −ks 2kt


= e e −e e ds du
k t t
T
σ2
Z
e−ku eku − ekt + e2kt e−ku − e−kt du



=
k2 t
2 Z T
σ h i
= 1 − 2e−k(u−t) + e−2k(u−t) du
k2 t

The 3 terms within this integral integrate respectively as:

1
1.
2k × 2k (T − t)
1
4e−k(T −t) − 4


2.
2k
1
1 − e−2k(T −t)


3.
2k

and hence

σ2 h −k(T −t) −2k(T −t)

i
V (t, T ) = 2k (T − t) − 3 + 4e − e (17)
2k 3

5.4 The Bond price as an expectation

The bond price involves the calculation of the expected value of the exponential
of ξ (t, T ) . Reminding proposition 3 in appendix, one has that

1
lnP (t, T ) = −ξ¯ (t, T ) + V (t, T )
2
1
The zero-coupon rate, R (t, T ) ≡ T −t ln (P (t, T )) can then be obtained in a
closed form manner.

Proposition 2. In the Vasicek model, the ZC rate yields:

1
R (t, T ) = (A (t, T ) r (t) + B (t, T )) (18)
T −t
with
1 
A (t, T ) = 1 − e−k(T −t)
k
σ2 σ2 2
 
B (t, T ) = θ − 2 (A (t, T ) − (T − t)) − A (t, T )
2k 4k

22
Comments The above proposition makes explicit the ane relationship be-
tween the short-term rate and any zero-coupon rate. It illustrates the rigid
relationship across the entire yield curve. Equation 18 show that

R (t, T ) = f (t, rt )
As a result
 2
∂f
d ⟨Rt ⟩ = d ⟨rt ⟩
∂r
 
∂f
d ⟨Rt , rt ⟩ = d ⟨rt ⟩
∂r
d ⟨Rt , rt ⟩
ρ (Rt , rt ) ≡ p
d ⟨Rt ⟩ d ⟨rt ⟩
=1
the correlation of the variation of the entire yield curve are equal to 1.
Whenever the short-term rates moves by dr, the zero-coupon curve moves
1
by
T −t A (t, T )

5.5 A PDE approach to compute the Bond price

5.5.1 About the relationship between martingales and PDEs
Dene a process

dX (t) = µ (t, X) dt + σ (t, X) dW (t) (19)

X (0) = X0
with µ and σ satisfying Lipshitz condition so that X has a unique strong solu-
tion. This diusion is characterized by its innitesimal linear generator

∂f ∂f 1 ∂2f
f (t, x) → Lf (t, x) = + µ (t, x) + σ 2 (t, x) 2
∂t ∂x 2 ∂x
Note that:
∂f
df (t, X) = Lf dt + σ (t, X) dWt
∂x
Now, consider a function F. We want to compute

E (F (XT ) |Ft )
Dene
Zt ≡ E (F (XT ) |Ft ) ≡ f (t, Xt )
By the tower property, for any u>t
Z (t) = E (F (XT ) |Ft )
= E (E (F (XT ) |Fu ) |Ft )
= E (Zu |Ft )

23
Z (t) is a martingale. Now, using Ito
Z u  
∂f
Z (u) − Z (t) = Lf ds + σ (s, X) dW (s)
t ∂x
and hence Z t 
E (Z (t) |Fu ) = Z (u) + E Lf ds|Fu
u
So, the only way for Z to be a martingale is to ensure that
Z t 
∀t ≤ u : E Lf ds|Fu =0
u

which can be proven to be equivalent to

Lf = 0
Now dene the Dirichlet problem:
(
Lf =0
P (L, F ) =
f (T, x) = F (x)
Clearly, for f (t, Xt ) to be a martingale then f has to be a solution of the
Dirichlet problem P (L, F ). Reciprocally, if f is a solution of the Dirichlet
problem P (L, F ) then f (t, Xt ) = E (F (XT ) |Ft ) is a martingale.

5.5.2 Quick reminder of general option pricing

Consider a diusion under the RNP

dX (t) = X (t) µ (t, X (t)) dt + X (t) σ (t, X) dW (t)

X (0) = X0
and the bank account
dB (t) = B (t) r (t) dt
Let f (t, Xt ) any possible price Necessarily

f (t, X (t)) /B (t)

is a martingale. In the particular of (european) option pricing:

f (t, X (t)) = F (X (T ))

Underlying drift
 
X (t) 1 1
d = dX − 2 dB
B (t) B B
X
= ((µ − r) dt + σdW )
B
and hence
µ = r P − a.s.

24
B&S option price PDE
 
f (t, X (t)) Lf f
d = dt − 2 dB (t) + ...dW (t)
B (t) B B
1
= (Lf − f r (t) dt) + ...dW (t)
B
1
= La (f ) dt + ...dW (t)
B
hence the PDE is

La f = 0
f (T, x) = F

with
∂f ∂f 1 ∂2f
La f = + xr (t) + x2 σ 2 (t, x) 2 − r (t) f (t, x)
∂t ∂x 2 ∂x

5.5.3 The bond price in the Vasicek model

Another way to compute the bond price in the Vasicek model consists in making
prot of the markov property of the short term rate (see equation 11)

 RT 
P (t, T, rt ) = E e t −ru du |rt (20)

and calculate it as a function of rt . Along the same lines as precedently, but

with the particular dynamics of r

∂P ∂P 1 ∂2P
+ k (θ − r) + σ 2 2 (t, r) = rP (t, r)
∂t ∂r 2 ∂r
P (T, r) = 1

This equation is in general dicult to solve but we can use equation 2 to better
understand the structure of P. Remember that equation 11 states:

  Z u
r (u, rt ) = rt e−k(u−t) + θ 1 − e−k(u−t) + σ e−k(u−s) dWs .
t
and hence
∂r (u, rt )
= e−k(u−t)
∂rt
and then
Z T
∂ru 1 
du = 1 − e−k(T −t)
t ∂rt k
= A (t, T )

25
Now, the derivation of 20 yields

∂P 1 
=− 1 − e−k(T −t) P
∂rt k
= −AP
and hence, by integrating

P = C (t, T ) e−A(t,T )rt (21)

Now:

ˆ ∂P
∂rt = −AP,
2
ˆ ∂ P
∂rt2
= A2 P ,

ˆ ∂P ∂C 1 ∂A


∂t = ∂t C − ∂t r P
With the assumption that P cannot be null the previous equation can be rewrit-
ten as
1 ∂C 1
− Ak (θ − r) + σ 2 A2 − Akr = 0
C ∂t 2
which can be simplied as:

σ2 2
 
∂C
+ −kθA + A C=0
∂t 2
C (T, T ) = 1
C is the solution of a backward ODE which can be integrated as

T  !
σ2 2
Z 
C (t, T ) = exp −kθAu + A du
t 2 u

Now
! !
T T
σ2
Z Z  
2 −k(T −t) 2k(T −t)
exp σ A2u = exp 1 − 2e +e dt
t k2 t

= exp (V (t, T ))
 2 h i
σ −k(T −t) −2k(T −t)
= exp 2k (T − t) − 3 + 4e − e
2k 3
and
!
Z T
exp −kθAu du × exp (−A (t, T ) r) = exp (−θ (T − t) + (θ − r) A (t, T ))
t

= exp −ξ¯ (t, T )

and the price of the bond is recovered as

 
¯ 1
P (t, T ) = exp −ξ (t, T ) + V (t, T )
2

26
5.6 A forward rate approach to compute the bond price
We have seen previously that the forward rate can be dened as

 
1 P (t, T )
F (t, T, S) := −1
τ P (t, S)

One can introduce the instantaneous forward rate: f (t, T ) := lim+ F (t, T, S) =
S→T
− ∂lnP∂T(t,T ) . Using the previous formula, the zero-coupon bond can be com-
puted/interpreted as the accumulation of forward rates:

!
Z T
P (t, T ) = exp − f (t, u) du
t

Equivalentlty
∂P
= −f (t, T ) (22)
P ∂T
We have already seen that the forward rate is a martingale under the T-forward
measure. Let us prove that again. First

! !
Z T
P (t, T ) = E −exp rs ds |Ft
t
! !
Z T
∂P
⇒ =E −exp rs ds rT |Ft
∂T t

= P (t, T ) E T (−ru |Ft ) (23)

where the T-forward measure is the measure which density w.r.t. to the risk
exp( 0T rs ds)
R
dQT
neutral is the normalised stochastic discount factor ZT = dQ ≡ P (0,T ) .
The corresponding density process yields

R 
t
exp r ds
0 s
Zt = E (ZT |Ft ) = P (t, T ) (24)
P (0, T )

Now, by equalizing terms in equations (22) and (23) one has

f (t, T ) = E T (rT |Ft ) (25)

where E T (.) stands for the expectation under the T-forward measure.
Equation (25) provides a third route to solve the bond price: computing the
dynamics of the short term rate under the T-forward measure, computing its
expected value and integrating over time.

27
5.6.1 Dynamics of the short-term rate under the T-forward measure
To apply Girsanov, one needs to calculate the covariation between Zt and rt .
Dierentiating equation (24) yields:

R 
t
exp r ds
0 s ∂P
dZt = dr + (...)dt
P (0, T ) ∂r

As in the previous subsection, we are looking for the bond price as

P (t, T ) = exp (−A (t, T ) rt + ...)

and hence
R  R 
t t
exp r ds
0 s ∂P exp r ds
0 s
dr = − P A (t, T ) dr
P (0, T ) ∂r P (0, T )
= −Z × A (t, T ) dr
= −Z × A (t, T ) σdW + ...
1
1 − e−k(T −t) .


with A (t, T ) =
k
Now one has

d ⟨Z, W ⟩
= −A (t, T ) σd ⟨W ⟩
Z
= −A (t, T ) σdt

and, by a direct application of Girsanov, the process dened as:

d ⟨Z, W ⟩
dW T = dW −
Z
= dW + A (t, T ) σdt

is a T-SBM. Now, the dynamics of the short-term rates under the T-foward
measure yields

drt = k (θ − rt ) − σ 2 A (t, T ) dt + σdW T

 

As the expected value of r(T ) the foward price f (t, T, rt ) is a solution of the
following ODE:

∂f
+ kf = θk − σ 2 A (t, T )
∂t
f (t, t, r) = rt

σ2
This is a linear ODE with a forcing term g (t, T ) = θk − σ 2 A (t, T ) = θk − k +
σ 2 −k(T −t)
k e .
.....

28
6 n-factor Hull-White model
Introduction Damiano-brigo :
In this section we consider an interest-rate model where the instantaneous
short-rate process is given by the sum of two correlated Gaussian factors plus
a deterministic function that is properly chosen so as to exactly t the current
term structure of discount factors. The model is quite analytically tractable in
that explicit formulas for discount bonds, European options on pure discount
bonds, hence caps and oors, can be readily derived. Gaussian models like this
G2++ model are very useful in practice, despite their unpleasant feature of
the theoretical possibility of negative rates. Indeed, their analytical tractability
considerably ease the task of pricing exotic products. The Gaussian distribution
allows the derivation of explicit formulas for a number of non-plain-vanilla in-
struments and, combined with the analytical expression for zero-coupon bonds,
leads to ecient and fairly fast numerical procedures for pricing any possible
payo. Also, nite spot and forward rates at a given time for any maturity and
accrual conventions can be given an explicit analytical expression in terms of
the short-rate factors at the relevant instant. This allows for easy propagation
of the whole zero-coupon curve in terms of the two factors. Another conse-
quence of the presence of two factors is that the actual variability of market
rates is described in a better way: Among other improvements, a non-perfect
correlation between rates of dierent maturities is introduced. This results in a
more precise calibration to correlation-based products like European swaptions.
These major advantages are the main reason why we devote so much attention
to a two-factor Gaussian model.
For pedagogical reasons, a n-factor Hull-White model is presented even-if,
in most practical applications, n=2 is used.

6.1 The short-rate dynamics

The short rate is modelled as

n
X
r (t) = xi (t) + φ (t)
i=1

where

dxi (t) = −ai x (t) dt + σi dWi (t)

xi (0) = 0

and d ⟨W1 , W2 ⟩ = ρdt. φ is a deterministic function such that φ (0) = r (0) .

In the Vasicek model, one had:

φ (t) = r0 e−kt + θ 1 − e−kt

29
 Important relationships:
Z t
xi (t) = σi e−ai (t−u) dWi (u)
0
 
E (xi (t) |Fu ) = xi (u) 1 − e−ai (t−u)
σi2
1 − e−2ai t


⟨xi (t)⟩ =
2ai
ρσi σj  
⟨xi (t) , xj (t)⟩ = 1 − e−(ai +aj )t
ai + aj
γij (t, s, u) = cov (xi (s) , xj (u) |Ft )
ρσi σj −ai s −aj u  (ai +aj )s∧u 
= e e e − e(ai +aj )t
ai + aj
Xn Xn
⟨r (t)⟩ = ⟨xi (t)⟩ + 2 ⟨xi (t) , xj (t)⟩
i=1 i<j

6.2 Pricing of a ZC bond

Lemma 1. is a gaussian variable with mean
RT
Ii (t, T ) ≡ t
xi (u)
E (Ii (t, T )) = Mi (t, T ) xi (t)
1 − e−ai (T −t)
with Mi (t, T ) =
ai
and
σi2
 
2 −ai (T −t) 1 3
⟨Ii (t, T )⟩ = T −t+ e − e−2ai (T −t) −
a2i ai ai 2ai
(26)
−ai (T −t) −aj (T −t)
ρσi σj e −1 e −1 e−(aj +ai )(T −t) − 1
⟨Ii (t, T ) , Ij (t, T )⟩ = (T − t) + + −
ai aj ai aj aj + ai
(27)
* n
+ n n
X X X
V (t, T ) ≡ Ii (t, T ) = ⟨Ii (t, T )⟩ + 2 ⟨Ii (t, T ) , Ij (t, T )⟩
i=1 i=1 i<j

Proof. Equation (26 is another version of equation (17). Equation (27) is ob-
tained along the same lines as equation (17) but the demonstration has to be
adapted a little bit.

Theorem 4. The price of the zero-coupon bond is equal to

( n
)
X 1
P (t, T ) = exp −ϕ (t, T ) − Mi (t, T ) xi (t) + V (t, T )
i=1
2

where ϕ (t, T ) ≡
RT
t
φ (u) du.

30
 The nice thing with the previous expression is that it can be easily t to the
spot ZC curve since :

 
1
P (0, T ) = exp −ϕ (0, T ) + V (0, T )
2
 
1
P (0, t) = exp −ϕ (0, t) + V (t, T )
2
 
P (0, T ) 1
⇔ = exp −ϕ (t, T ) + (V (0, T ) − V (0, t))
P (0, t) 2
 
P (0, T ) 1
⇔ exp (−ϕ (t, T )) = exp − (V (0, T ) − V (0, t))
P (0, t) 2

Now dene The price of the zero-coupon bond can be expressed for a non spec-
ied shape of the spot yield curve:

Theorem 5. The price of the of the zero-coupon bond is equal to

n
!
P (0, T ) X 1
P (t, T ) = exp − Mi (t, T ) xi (t) + ∆V (t, T )
P (0, t) i=1
2

with ∆V (t, T ) ≡ V (t, T ) + V (0, t) − V (0, T ) .

The forward rate can be obtained easily as

∂
f (t, T ) = − ln (P (t, T ))
∂T
∂ ∂M1 (t, T ) ∂M2 (t, T )
=− ln (A (t, T )) + x1 (t) + x2 (t)
∂T ∂T ∂T
One nd that
∂Mi (t, T )
mi (t) = = e−ai (T −t)
∂T
and

df (t, T ) = ...dt + m1 (t, T ) σ1 dW1 (t) + m2 (t, T ) σ2 dW2 (t)

Appendix
Laplace transform of Gaussian variable
Proposition 3. If Y then


= N µ, σ 2
λ2 σ 2
∀λ : E eλY = eλµ+ 2

31
Solution of SDEs with ane coecients
The following is borrowed from exercise 6.15 of [2]. Consider the following family
of SDEs

dXt = [A (t) Xt + a (t)] dt + [σ1 (t) Xt + σ2 (t)] dWt (28)

X0 = X0

with time dependant stochastic adapted ane coecients satisfying a.s.

bounded.
Dene

t
1 t 2
Z Z
Mt ≡ σ1 (s) dWs − σ (s) ds
0 2 0 1
 Z t 
Zt ≡ exp Mt + A (u) du
0

then the unique solution of equation 28 is provided by

 t t 
a (u) − σ1 (u) σ2 (u)
Z Z
σ2 (u)
Xt = Zt X0 + du + dWu
0 Z (u) 0 Zu

This can be checked as an exercise.

References
[1] Damiano Brigo and Fabio Mercurio. Interest rate models: theory and prac-
tice, volume 2. Springer, 2001.

[2] Ioannis Karatzas, Steven E Shreve, I Karatzas, and Steven E Shreve. Meth-
ods of mathematical nance, volume 39. Springer, 1998.

[3] Oldrich Vasicek. An equilibrium characterization of the term structure.

Journal of nancial economics, 5(2):177188, 1977.

32