Interest Rate Modelling and Derivative Pricing

Sebastian Schlenkrich

HU Berlin, Department of Mathematics

WS, 2019/20
 Part IV

Term Structure Modelling

p. 215
Outline

HJM Modelling Framework

Hull-White Model

p. 216
What are term structure models compared to Vanilla
models?
Vanilla models Term structure models

◮ Specify dynamics for a single ◮ Specify dynamics for

swap rate S(T ) with evolution of all future zero
start/end dates T0 /Tn (and coupon bonds P(T , T ′ )
details). (t ≤ T ≤ T ′ ).

◮ Effectively, only describes ◮ Yields (joint) distribution of

terminal distribution of S(T ). all swap rates S(T ).

◮ Allows pricing of European ◮ Allows pricing of Bermudan

swaptions. swaptions and other complex
derivatives.
◮ Can be extended to slightly
more complex options (with ◮ Typically, computationally
additional assumptions). more expensive than Vanilla
model pricing.

p. 217
Why do we need to model the whole term structure of
interest rates?

Recall

V Berm (t) = MaxEuropean + SwitchOption.

◮ MaxEuropean price is fully determined by Vanilla model.
◮ Residual SwitchOption price cannot be inferred from Vanilla model.
SwitchOption (i.e. right to postpone future exercise decisions) pricing requires
modelling of full interest rate term structure.

p. 218
Outline

HJM Modelling Framework

Hull-White Model

p. 219
Outline

HJM Modelling Framework

Forward Rate Specification
Short Rate and Markov Property
Seperable HJM Dynamics

p. 220
Heath-Jarrow-Morton specify general dynamics of zero
coupon bond prices

Recall our market setting

R t with zero coupon bonds P(t, T ) (t ≤ T ) and bank
account B(t) = exp 0 r (s)ds .
Discounted bond price is martingale in risk-neutral measure.
Martingale representation theorem yields
 
P(t, T ) P(t, T )
d =− · σP (t, T )⊤ · dW (t)
B(t) B(t)

where σP (t, T ) = σP (t, T , ω) is a d-dimensional process adapted to Ft .

We also impose σP (T , T ) = 0 (pull-to-par for bond prices with P(T , T ) = 1).

◮ What are dynamics of (un-discounted) zero bonds P(t, T )?

◮ What are dynamics of forward rates f (t, T )?
◮ How to specify bond price volatility?

p. 221
What are dynamics of zero bonds P(t, T )?

Lemma (Bond price dynamics)

Under the risk-neutral measure zero bond prices evolve according to
dP(t, T )
= r (t) · dt − σP (t, T )⊤ · dW (t).
P(t, T )

Proof.
Apply Ito’s lemma to d (P(t, T )/B(t)) and compare with dynamics of
discounted bond prices.

◮ Zero bond drift equals short rate r (t).

◮ Zero bond volatility σP (t, T ) remains unchanged.
◮ How do we get r (t)?

p. 222
What are dynamics of forward rates f (t, T )?

Theorem (Forward rate dynamics)

Consider a d-dimensional forward rate volatility process σf (t, T ) = σf (t, T , ω)
adapted to Ft . Under the risk-neutral measure the dynamics of forward rates
f (t, T ) are given by
Z T 
df (t, T ) = σf (t, T )⊤ · σf (t, u)du · dt + σf (t, T )⊤ · dW (t)
t

M
and f (0, T ) = f (0, T ). Moreover
Z T
σP (t, T ) = σf (t, u)du.
t

◮ Once volatility σf (t, T ) is specified no-arbitrage pricing yields the drift.

◮ Model is auto-calibrated to initial yield curve via f (0, T ) = f M (0, T ).

p. 223
We prove the forward rate dynamics (1/2)
Recall
∂
ln (P(t, T )) .
f (t, T ) = −
∂T
Exchanging order of differentiation yields
h i
∂ ∂
df (t, T ) = d − ln (P(t, T )) = − d ln (P(t, T )) .
∂T ∂T
Applying Ito’s lemma (to d ln (P(t, T ))) with bond price dynamics yields

d(P(t, T )) σP (t, T )⊤ σP (t, T )

d ln (P(t, T )) = − · dt
P(t, T ) 2
 
σP (t, T )⊤ σP (t, T )
= r (t) − · dt − σP (t, T )⊤ · dW (t).
2

Differentiating d ln (P(t, T )) w.r.t. T gives

h i⊤ h i⊤
∂ ∂
df (t, T ) = σP (t, T ) σP (t, T ) · dt + σP (t, T ) · dW (t).
∂T ∂T

p. 224
We prove the forward rate dynamics (2/2)

h i⊤ h i⊤
∂ ∂
df (t, T ) = σP (t, T ) σP (t, T ) · dt + σP (t, T ) · dW (t).
∂T ∂T
Denote
∂
σP (t, T ).
σf (t, T ) =
∂T
With terminal condition σP (T , T ) = 0 follows integral representation
Z T
σP (t, T ) = σf (t, u)du.
t

Substituting back gived the result

Z T 
df (t, T ) = σf (t, T )⊤ · σf (t, u)du · dt + σf (t, T )⊤ · dW (t).
t

p. 225
It will be useful to have the dynamics under the forward
measure as well

Lemma (Brownian motion in T -forward measure)

Consider our HJM framework with Brownian motion W (t) under the
risk-neutral measure and
dP(t, T )
= r (t) · dt − σP (t, T )⊤ · dW (t).
P(t, T )
Under the T -forward measure the bond price dynamics are
dP(t, T )  
= r (t) + σP (t, T )⊤ σP (t, T ) · dt − σP (t, T )⊤ · dW T (t)
P(t, T )

with W T (t) a Brownian motion (under the T -forward measure). Moreover,

dW T (t) = σP (t, T ) · dt + dW (t).

p. 226
T -forward measure dynamics can be shown by Ito’s lemma
Abbrev. deflated bond prices Y (t) = P(t,T
B(t)
)
, then dY (t)
Y (t)
= −σP (t, T )⊤ dW (t).
Now consider 1/Y (t) and apply Ito’s lemma
  " 2 #
1 dY (t) 1 2 1 dY (t) dY (t)
d =− + [dY (t)]2 = −
Y (t) Y (t)2 2 Y (t)3 Y (t) Y (t) Y (t)
1  
= σP (t, T )⊤ σP (t, T )dt + σP (t, T )⊤ dW (t)
Y (t)
σP (t, T )⊤
= [σP (t, T )dt + dW (t)] .
Y (t)

However,
  1/Y (t) = B(t)/P(t, T ) is a martingale in T -forward measure and
1
d Y (t)
must be drift-less in T -forward measure.
Define W T (t) with

dW T (t) = σP (t, T )dt + dW (t).

Then W T (t) must be a Brownian motion in the T -forward measure.

Substituting dW (t) in the risk-neutral bond price dynamics finally gives the
dynamics under T -forward measure.
p. 227
Outline

HJM Modelling Framework

Forward Rate Specification
Short Rate and Markov Property
Seperable HJM Dynamics

p. 228
Short rate can be derived from forward rate dynamics

Corollary (Short rate specification)

In our HJM framework the short rate becomes

r (t) = f (t, t)
= f (0, t)+
Z t Z t  Z t
σf (u, t)⊤ · σf (u, s)ds du + σf (u, t)⊤ · dW (u).
0 u 0

Proof.
Follows directly from forward rate dynamics and integration from 0 to t.

Rt
◮ Note that integrand in diffusion term D(t) = 0
σf (u, t)⊤ · dW (u)
depends on t.
◮ In general, D(t) is not a martingale.
◮ In general, r (t) is not Markovian unless volatility σf (t, T ) is suitably
restricted.
p. 229
We analyse diffusion term in detail
Z t
D(t) = σf (u, t)⊤ · dW (u).
0
It follows
Z t Z T
⊤
D(T ) = σf (u, T ) · dW (u) + σf (u, T )⊤ · dW (u)
0 t
Z T
= D(t) + σf (u, T )⊤ · dW (u)
t
Z t Z t
+ σf (u, T )⊤ · dW (u) − σf (u, t)⊤ · dW (u)
0 0
Z T Z t
⊤
= D(t) + σf (u, T ) · dW (u) + [σf (u, T ) − σf (u, t)]⊤ · dW (u).
t 0

◮ How is EQ [D(T ) | D(t)] (knowing only last state) related to

EQ [D(T ) | Ft ] (knowing full history)?
◮ If D is Markovian then EQ [D(T ) | D(t)] = EQ [D(T ) | Ft ] (neccessary
condition).
p. 230
Compare EQ [D(T ) | D(t)] and EQ [D(T ) | Ft ]
 Z T 
Q Q ⊤
E [D(T ) | Ft ] = E D(t) + σf (u, T ) dW (u) | Ft
t
Z t 
Q ⊤
+E [σf (u, T ) − σf (u, t)] dW (u) | Ft
0
Z t
= D(t) + 0 + [σf (u, T ) − σf (u, t)]⊤ dW (u) .
|0 {z }
I(t,T )

 Z T 
EQ [D(T ) | D(t)] = EQ D(t) + σf (u, T )⊤ dW (u) | D(t)
t
Z t 
+E Q
[σf (u, T ) − σf (u, t)]⊤ dW (u) | D(t)
0
Z t 
= D(t) + 0 + EQ [σf (u, T ) − σf (u, t)]⊤ dW (u) | D(t) .
0

◮ EQ [D(T ) | D(t)] = EQ [D(T ) | Ft ] only if I(t, T ) is non-random or

deterministic function of D(t).
p. 231
An important separability condition makes D(t) Markovian
Assume
σf (t, T ) = g(t) · h(T )
with g(t) (scalar) process adapted to Ft and h(T ) (scalar) deterministic and
differentiable function.
Then

Z t Z T
D(T ) = g(u) · h(T ) · dW (u) + g(u) · h(T ) · dW (u)
0 t
Z T
h(T )
= · D(t) + h(T ) · g(u) · dW (u).
h(t) t

Thus
h(T )
EQ [D(T ) | D(t)] = EQ [D(T ) | Ft ] = · D(t).
h(t)

Moreover
h′ (t)
d (D(t)) = · D(t) · dt + g(t) · h(t) · dW (t).
h(t)

p. 232
Outline

HJM Modelling Framework

Forward Rate Specification
Short Rate and Markov Property
Seperable HJM Dynamics

p. 233
We describe a very general but still tractable class of
models

◮ We give a general description of a class of term structure models.

◮ Typically, these models are called Cheyette-type or quasi-Gaussian
models; also associated with work by Ritchken and
Sankarasubramanian (1995).
◮ Particular parameter choices will specialise general model to
classical model (e.g. Hull-White model).
◮ More complex parameter choices yield powerfull model instances for
complex interest rate derivatives.

Quasi-Gaussian models are important models in the industry.

p. 234
Separable forward rate volatility

Definition (Separable forward rate volatility)

The forward rate volatility σf (t, T ) of an HJM model is considered of separable
form if
σf (t, T ) = g(t)h(T )
for a matrix-valued process g(t) = g(t, ω) ∈ Rd×d adapted to Ft and a
vector-valued deterministic function h(T ) ∈ Rd .

Corollary
For a separable forward rate volatility σf (t, T ) = g(t)h(T ) the bond price
volatility σP (t, T ) becomes
Z T
σP (t, T ) = g(t) h(u)du.
t

p. 235
Forward rate representation follows directly

Lemma
For a separable forward rate volatility σf (t, T ) = g(t)h(T ) the forward rate
becomes

f (t, T ) = f (0, T )+
Z t Z T 
⊤ ⊤
h(T ) g(s) g(s) h(u)du ds+
0 s
Z t
h(T )⊤ g(s)⊤ dW (s)
0

and
Z t Z t  Z t 
⊤ ⊤ ⊤
r (t) = f (0, t) + h(t) g(s) g(s) h(u)du ds + g(s) dW (s) .
0 s 0

Proof.
Follows directly from definition.

p. 236
We need to introduce new state variables to derive
Markovian representation of short rate

Re-write
Z t Z t  Z t 
⊤ ⊤ ⊤
r (t) = f (0, t) + 1 H(t) g(s) g(s) h(u)du ds + g(s) dW (s)
0 s 0

with
   
1 h1 (t) 0 0
 ..   .. 
1 =  .  and H(t) = diag (h(t)) =  0 . 0 .
1 0 0 hd (t)

Introduce vector of state variables x (t) with

Z t Z t  Z t 
⊤ ⊤
x (t) = H(t) g(s) g(s) h(u)du ds + g(s) dW (s) .
0 s 0

p. 237
We derive the dynamics of the short rate

Theorem (Separable HJM short rate dynamics)

In an HJM model with separable volatility the short rate is given by
r (t) = f (0, t) + 1⊤ x (t). The vector of state variables x(t) evolves according to
x (0) = 0 and

dx (t) = [y (t)1 − χ(t)x (t)] dt + H(t)g(t)⊤ dW (t)

with symmetric matrix of auxilliary variables y (t) as

Z t 
⊤
y (t) = H(t) g(s) g(s)ds H(t)
0

and diagonal matrix of mean reversion parameters χ(t) as

dH(t)
χ(t) = − H(t)−1 .
dt

p. 238
Proof follows straight forward via differentiation (1/3)

We have
Z t Z t  Z t 
⊤ ⊤
x (t) = H(t) g(s) g(s) h(u)du ds + g(s) dW (s) .
0 s 0
| {z }
G(t)

dx (t) = H ′ (t) · G(t) · dt + H(t) · dG(t)

= H ′ (t) · H(t)−1 · H(t) · G(t) · dt + H(t) · dG(t)
= −χ(t) · x (t) · dt + H(t) · dG(t).

p. 239
Proof follows straight forward via differentiation (2/3)

dx (t) = −χ(t) · x (t) · dt + H(t) · dG(t),

Z t Z t  Z t
G(t) = g(s)⊤ g(s) h(u)du ds + g(s)⊤ dW (s).
0 s 0

Leibnitz rule yields

 Z t  Z t Z t  
⊤ ⊤ d
dG(t) = g(t) g(t) h(u)du + g(s) g(s) h(u)du ds dt
t 0
dt s

+ g(t)⊤ dW (t)
 Z t 
= 0+ g(s) g(s) · H(t)1 · ds dt + g(t)⊤ dW (t)
⊤

0
Z t  
⊤
= g(s) g(s)ds H(t)1 dt + g(t)⊤ dW (t).
0

p. 240
Proof follows straight forward via differentiation (3/3)

Combining results gives

dx (t) = −χ(t) · x (t) · dt + H(t) · dG(t)

 Z t  
⊤
= H(t) g(s) g(s)ds H(t)1 − χ(t) · x (t) dt
0

+ H(t) · g(t)⊤ dW (t)

= [y (t) · 1 − χ(t) · x (t)] dt + H(t) · g(t)⊤ dW (t).

◮ Note
R that dx (t) depends on accumulated previous volatility via
t
0
g(s)⊤ g(s)ds.
◮ x (t) is Markovian only if volatility function g(t) is deterministic.
◮ In general, short rate dynamics can be ammended by dynamics of y (t).

p. 241
Short rate dynamics can be written in terms of state and
auxilliary variables
Corollary (Augmented short rate dynamics)
In an HJM model with separable volatility the short rate is given via
r (t) = f (0, t) + 1⊤ x (t) with

dx (t) = [y (t) · 1 − χ(t) · x (t)] dt + σr (t)⊤ dW (t),

 
dy (t) = σr (t)⊤ σr (t) − χ(t)y (t) − y (t)χ(t) dt,

and x (0) = 0, y (0) = 0.

Proof. Rt

Set σr (t) = g(t)H(t) and differentiate y (t) = H(t) 0
g(s)⊤ g(s)ds H(t).

◮ Model class also called Cheyette or quasi-Gaussian models.

◮ Typically σr (t) and χ(t) are specified and σf (t, T ) is reconstructed via
H ′ (t) = −χ(t)H(t), H(0) = 1 and g(t) = σr (t)H(t)−1 .

p. 242
Forward rates and zero bonds can be written in terms of
state/auxilliary variables

Theorem (Forward rate and zero bond reconstruction)

In our HJM model setting we get

f (t, T ) = f (0, T ) + 1⊤ H(T )H(t)−1 [x (t) + y (t)G(t, T )]

and
P(0, T )
n o
1
P(t, T ) = exp −G(t, T )⊤ x (t) − G(t, T )⊤ y (t)G(t, T )
P(0, t) 2
with Z T
G(t, T ) = H(u)H(t)−1 1du.
t

◮ We prove the first part for f (t, T ).

◮ And we sketch the proof for the second part for P(t, T ).

p. 243
We prove the first part for f (t, T ) ...
Z t Z t  Z t 
⊤ −1 T ⊤ ⊤
1 H(T )H(t) x (t) = h(T ) g(s) g(s) h(u)du ds + g(s) dW (s)
| {z } 0 s 0
I1

Z t Z T
⊤ −1 ⊤ ⊤
1 H(T )H(t) y (t)G(t, T ) = h(T ) g(s) g(s)ds h(u)du
| {z } 0 t
I2

Z t Z t  Z t Z T 
I1 + I2 = h(T )T g(s)⊤ g(s) h(u)du ds + g(s)⊤ g(s)ds h(u)du
0 s 0 t
Z t
+ h(T )T g(s)⊤ dW (s)
0
Z t Z t Z T  Z t 
= h(T )T g(s)⊤ g(s) h(u)du + h(u)du ds + g(s)⊤ dW (s)
0 s t 0
Z t Z T  Z t 
= h(T )T g(s)⊤ g(s) h(u)du ds + g(s)⊤ dW (s)
0 s 0
= f (t, T ) − f (0, T )
p. 244
... and sketch the proof for the second part for P(t, T )
 Z T 
P(t, T ) = exp − f (t, s)ds
t
 Z T 


= exp − f (0, s) + 1⊤ H(s)H(t)−1 [x (t) + y (t)G(t, s)] ds
t
 

 


 Z T  

P(0, T )
= · exp − 1⊤ H(s)H(t)−1 ds x (t) ·
P(0, t)  t 

 

 | {z } 
G(t,T )⊤
 Z T 
exp − 1⊤ H(s)H(t)−1 y (t)G(t, s)ds
t

The equality
Z T
1
1⊤ H(s)H(t)−1 y (t)G(t, s)ds = G(t, T )⊤ y (t)G(t, T )
t
2

follows by differentiating both sides w.r.t. T .

p. 245
Outline

HJM Modelling Framework

Hull-White Model

p. 246
We take a complementary view to HJM framework and
consider direct modelling of the short rate r (t)

t
❅
❅
❅
❅
❅
short rate r (t) = f (t, t)

We model short rate of the discount curve as offset point for future rates.

p. 247
Short rate suffices to specify evolution of the full yield
curve
Recall zero bond formula
  Z T  
P(t, T ) = EQ exp − r (s)ds | Ft .
t

◮ Once dynamics of r (t) are specified all zero bonds can be derived.
Libor rates (in multi-curve setting) are
 
P(t, T0 ) 1
L(t; T0 , T1 ) = ET1 [L(T ; T0 , T1 ) | Ft ] = · D(T0 , T1 ) − 1 .
P(t, T1 ) τ

◮ With zero bonds P(t, T ) (and spread factors D(T0 , T1 )) we can also
derive future Libor rates.

Short rate is a natural choice of state variable for modelling evolution of

interest rates.

p. 248
Outline

Hull-White Model
Classical Model Derivation
Relation to HJM Framework
Analytical Bond Option Pricing Formulas
General Payoff Pricing
Summary of Hull-White Pricing Formulas
European Swaption Pricing
Impact of Volatility and Mean Reversion

p. 249
Vasicek model and Ho-Lee model were the first models for
the short rate

Vasicek (1977) assumed Ornstein-Uhlenbeck process

dr (t) = κ (θ − r (t)) dt + σdW (t), r (0) = r0

for positive constants r0 , κ, θ, and σ.

◮ Model is not too different from HJM model representation.
◮ Constant parameters (in particular θ) limit ability to reproduce yield curve
observed today.
Ho and Lee (1986) introduce exogenous time-dependent drift parameter,

dr (t) = θ(t)dt + σdW (t).

◮ Drift parameter θ(t) is used to match today’s zero bonds P(0, T ).

◮ Lack of mean reversion is considered main disadvantage.
◮ Model was historically used with binomial tree implementation.

p. 250
Hull and White (1990) extended Vasicek model by θ(t)
Definition (Hull-White model)
In the Hull-White model the short rate evolves according to

dr (t) = [θ(t) − a(t)r (t)] dt + σ(t)dW (t)

with deterministic scalar functions θ(t), a(t), and σ(t) > 0.

◮ θ(t) is mean reversion level,
◮ a(t) is mean reversion speed, and
◮ σ(t) is short rate volatility.
◮ Original reference is J. Hull and A. White. Pricing interest-rate-derivative
securities.
The Review of Financial Studies, 3:573–592, 1990
◮ To simplify analytical tractability we will assume
◮ constant mean reversion speed a(t) = a > 0, and
◮ piece-wise constant short rate volatility function on a siutable time
grid {t0 , . . . , tk }, k
X
σ(t) = ✶{ti−1 ≤t<ti } · σi .
i=1

p. 251
How do we calibrate the drift θ(t)?

Lemma (Hull-White drift calibration)

In the risk-neutral specification of the Hull-White model the drift term θ(t) is
given by
Z t
∂  2
θ(t) = f (0, t) + a · f (0, t) + e −a(t−u) σ(u) du.
∂T 0

Here f (0, t) = f M (0, t) is exogenously specified and assumed continuously

differentiable w.r.t. the maturity T .

Proof follows along the following steps

◮ Calculate r (s) via integration.
R
◮ Integrate I(t, T ) = tT r (s)ds and calculate distribution of I(t, T ).
 
◮ Derive θ(t) such that EQ e −I(0,t) = P(0, T ).

p. 252
Proof (1/4) - calculate r (s)
We show that for s ≥ t
 Z s 
−a(s−t) a(u−t)
r (s) = e r (t) + e [θ(u)du + σ(u)dW (u)] .
t

 
dr (s) = −ar (s)ds + e −a(s−t) e a(s−t) [θ(s)ds + σ(s)dW (s)]
= [θ(s) − ar (s)] ds + σ(s)dW (s).

RT
Use notation [·]′ (t, T ) = ∂
∂T
[·]. Set I(t, T ) = r (s)ds, then
t
∂I(t,T )
I ′ (t, T ) = ∂T
= r (T ). We show

Z T
I(t, T ) = G(t, T )r (t) + G(u, T ) [θ(u)du + σ(u)dW (u)]
t

with Z  
T
1 − e −a(T −t)
G(t, T ) = e −a(u−t) du = .
t
a

p. 253
Proof (2/4) - calculate distribution I(t, T )
Z T
I(t, T ) = G(t, T )r (t) + G(u, T ) [θ(u)du + σ(u)dW (u)] ,
t
Z T
I ′ (t, T ) = G ′ (t, T )r (t) + 0 + G ′ (u, T ) [θ(u)du + σ(u)dW (u)]
t
Z T
= e −a(T −t) r (t) + e −a(T −u) [θ(u)du + σ(u)dW (u)]
t
 Z T 
= e −a(T −t) r (t) + e a(u−t) [θ(u)du + σ(u)dW (u)]
t
= r (T ).

Integral I(t, T ) is normally distributed, I(t, T ) ∼ N(µ, σ 2 ) with

Z T
µ(t, T ) = G(t, T )r (t) + G(u, T )θ(u)du,
t
Z T
σ(t, T )2 = [G(u, T )σ(u)]2 du.
t

p. 254
Proof (3/4) - calculate forward rate
I(t, T ) ∼ N(µ, σ 2 ) with

Z T Z T
µ(t, T ) = G(t, T )r (t) + G(u, T )θ(u)du, σ 2 (t, T ) = [G(u, T )σ(u)]2 du.
t t

  1 2
P(t, T ) = E Q
e −I(t,T ) | Ft = e −µ(t,T )+ 2 σ (t,T )
.

h i
∂ d 1 2
f (t, T ) = − ln [P(t, T )] = µ(t, T ) − σ (t, T )
∂T dT 2
Z T
= G ′ (t, T )r (t) + 0 + G ′ (u, T )θ(u)du
t
 Z T 
1
− 0+ 2G(u, T )G ′ (u, T )σ(u)2 du
2 t
Z T Z T
= G ′ (t, T )r (t) + G ′ (u, T )θ(u)du − G ′ (u, T )G(u, T )σ(u)2 du.
t t

p. 255
Proof (4/4) - derive drift θ(t)
Z T Z T
f (t, T ) = G ′ (t, T )r (t) + G ′ (u, T )θ(u)du − G ′ (u, T )G(u, T )σ(u)2 du.
t t

Use G ′ (t, T ) = e −a(T −t) and G ′′ (t, T ) = −aG ′ (t, T ), then

Z T
f ′ (t, T ) = G ′′ (t, T )r (t) + θ(T ) + G ′ (u, T )θ(u)du − 0
t
Z T
 
− G ′′ (u, T )G(u, T ) + G ′ (u, T )2 σ(u)2 du
t
Z T
 2
= θ(T ) − af (t, T ) − G ′ (u, T )σ(u) du.
t

This finally gives the result (with t = 0)

Z T
′
 2
θ(T ) = f (t, T ) + af (t, T ) + G ′ (u, T )σ(u) du
t
Z T
 2
= f ′ (0, T ) + af (0, T ) + e −a(T −u) σ(u) du.
0
p. 256
Do we really need the drift θ(t)?

◮ Risk-neutral drift representation

Z t
∂  2
θ(t) = f (0, t) + a · f (0, t) + e −a(t−u) σ(u) du
∂T 0

poses some obstacles.

◮ Derivative ∂T∂
f (0, t) may cause numerical difficulties.
◮ In some market situations you want to have jumps in f (0, t).
◮ This is relevant in particular for the short end of OIS curve.

◮ Fortunately, for most applications we don’t need drift term.

◮ HJM representation allows avoiding it alltogether.

p. 257
Now we can also derive future zero bond prices I

Theorem (Zero bonds in Hull-White model)

In the Hull-White model future zero bond prices are given by
 Z t 
P(0, T ) G(t, T )2  2
P(t, T ) = exp −G(t, T ) [r (t) − f (0, t)] − e −a(t−u) σ(u) du
P(0, t) 2 0

with Z  
T
1 − e −a(T −t)
G(t, T ) = e −a(u−t) du = .
t
a

◮ The proof is a bit technical.

◮ We already derived the zero bond representation
  1 2
P(t, T ) = EQ e −I(t,T ) | Ft = e −µ(t,T )+ 2 σ (t,T )
.

p. 258
Now we can also derive future zero bond prices II

We have from the proof of risk-neutral drift that

Z T Z T
f (t, T ) = G ′ (t, T )r (t) + G ′ (u, T )θ(u)du − G ′ (u, T )G(u, T )σ 2 (u)du
t t

and RT RT
−G(t,T )r (t)− G(u,T )θ(u)du+ 12 G(u,T )2 σ 2 (u)du
P(t, T ) = e t t .

We aim at calculating the term

Z T Z T
1
I(t, T ) = − G(u, T )θ(u)du + G(u, T )2 σ 2 (u)du.
t
2 t

p. 259
Now we can also derive future zero bond prices III

Consider

  Z T Z t
P(0, t)
log = [G(0, T ) − G(0, t)] r (0) + G(u, T )θ(u)du − G(u, t)θ(u)du
P(0, T ) 0 0
Z T Z t 
1
− G(u, T )2 σ 2 (u)du − G(u, t)2 σ 2 (u)du
2 0 0
= [G(0, T ) − G(0, t)] r (0)
Z T Z t
+ G(u, T )θ(u)du + [G(u, T ) − G(u, t)] θ(u)du
t 0
Z T Z t 
1 2 2
 2 2
 2
− G(u, T ) σ (u)du + G(u, T ) − G(u, t) σ (u)du .
2 t 0

p. 260
Now we can also derive future zero bond prices IV
We use G(u, T ) − G(u, t) = G(t, T )G ′ (u, t) and re-arrange terms. Then
 
P(0, T )
I(t, T ) = log + G(t, T )G ′ (0, t)r (0)
P(0, t)
Z t
+ G(t, T ) G ′ (u, t)θ(u)du
0
Z t
1
− [G(u, T ) + G(u, t)] [G(u, T ) − G(u, t)] σ 2 (u)du.
2 0
| {z }
[G(u,T )−G(u,t)+2G(u,t)]G(t,T )G ′ (u,t)

We use representation for forward rate f (t, T ) and get

 
P(0, T )
I(t, T ) = log + G(t, T )f (0, t)
P(0, t)
Z t
1
− [G(u, T ) − G(u, t)] G(t, T )G ′ (u, t)σ 2 (u)du
2 0
  Z t
P(0, T ) G(t, T )2
= log + G(t, T )f (0, t) − G ′ (u, t)2 σ 2 (u)du.
P(0, t) 2 0

p. 261
Now we can also derive future zero bond prices V

Finally, we get the result

P(t, T ) = e −G(t,T )r (t)+I(t,T )

2 R t 2
P(0, T ) −G(t,T )[r (t)−f (0,t)]− G(t,T
2
)
e −a(t−u) σ(u) du
= e 0 .
P(0, t)

◮ Future zero coupon bonds depend on:

◮ today’s yield curve f (0, t),
◮ mean reversion parameter a via G(t, T ), and
◮ short rate volatility σ(t).
◮ We see that drift θ(t) is not required for future zero coupon bonds.

p. 262
Outline

Hull-White Model
Classical Model Derivation
Relation to HJM Framework
Analytical Bond Option Pricing Formulas
General Payoff Pricing
Summary of Hull-White Pricing Formulas
European Swaption Pricing
Impact of Volatility and Mean Reversion

p. 263
Recall short rate dynamics in separable HJM model

We consider a one-factor model (d = 1)

r (t) = f (0, t) + x (t)

dx (t) = [y (t) − χ(t) · x (t)] dt + σr (t) · dW (t)
 
dy (t) = σr (t)2 − 2 · χ(t) · y (t) · dt

with
H ′ (t) = −χ(t)H(t), H(0) = 1 and g(t) = H(t)−1 σr (t).

◮ How does this relate to Hull-White model with

dr (t) = [θ(t) − a · r (t)] · dt + σ(t) · dW (t)?

p. 264
Differentiate short rate in HJM model

dr (t) = f ′ (0, t)dt + dx (t)

= f ′ (0, t)dt + [y (t) − χ(t)x (t)] dt + σr (t)dW (t)
 
= f ′ (0, t) + y (t) − χ(t) (r (t) − f (0, t)) dt + σr (t)dW (t)
 

= f ′ (0, t) + χ(t)f (0, t) + y (t) − χ(t) r (t) dt + σr (t) dW (t)

 
| {z } |{z} |{z}
θ(t) a σ(t)

HJM volatility parameters become

H ′ (t) = −aH(t), H(0) = 1 ⇒ h(t) = H(t) = e −at ,

g(t) = σr (t) · H(t)−1 = σ(t)e at .

p. 265
Deterministic volatility allows calculation of auxilliary
variable y (t)

We have
y ′ (t) = σ(t)2 − 2 · a · y (t), y (0) = 0.
Solving initial value problem yields
Z t
y (t) = σ(u)2 · e −2a(t−u) du.
0

p. 266
Hull-White model in HJM notation

In the HJM framework the Hull-White model becomes

r (t) = f (0, t) + x (t),

Z t 
dx (t) = σ(u)2 · e −2a(t−u) du − a · x (t) · dt + σ(t) · dW (t),
0

x (0) = 0.

We will use this representation of the Hull-White model for our

implementations.

p. 267
This also gives HJM representation of Hull-White model
Corollary (Forward rate dynamics in Hull-White model)
In a Hull-White model the dynamics of the forward rate f (t, T ) become

1 − e −a(T −t)
df (t, T ) = σ(t)2 e −a(T −t) dt + σ(t)e −a(T −t) dW (t).
a

Proof.
Z T 
df (t, T ) = σf (t, T ) · σf (t, u)du · dt + σf (t, T ) · dW (t)
t
Z T 
= g(t)h(T ) g(t)h(u)du · dt + g(t)h(T ) · dW (t)
t
Z T 
2 −a(T −t) −a(u−t)
= σ(t) e e du ·dt + σ(t)e −a(T −t) · dW (t).
t
| {z }
1−e −a(T −t)
a

p. 268
Zero bond prices may also be computed in terms of x (t)

Corollary (Zero bonds in Hull-White model)

In the Hull-White model future zero coupon bonds are
 Z t 
P(0, T ) G(t, T )2  −a(t−u)
2
P(t, T ) = exp −G(t, T )x (t) − e σ(u) du
P(0, t) 2 0

with Z  
T
1 − e −a(T −t)
G(t, T ) = e −a(u−t) du = .
t
a

Proof.
Result follows either from Hull-White model zero bond formula with
x (t) = r (t) − f (0, T ) or from zero bond formula for the separable HJM model
with Hull-White results for G(t, T ) and y (t).

p. 269
Outline

Hull-White Model
Classical Model Derivation
Relation to HJM Framework
Analytical Bond Option Pricing Formulas
General Payoff Pricing
Summary of Hull-White Pricing Formulas
European Swaption Pricing
Impact of Volatility and Mean Reversion

p. 270
First we need the distribution of the state variable x (t)

We have
dx (t) = [y (t) − a · x (t)] · dt + σ(t) · dW (t).
This yields for t ≥ s
 Z t 
x (t) = e −a(t−s) x (s) + e a(u−s) (y (u)du + σ(u)dW (u)) .
s

Lemma (State variable distribution)

In the HJM version of the Hull-White model we have that under the
risk-neutral measure the state variable x (t) is normally distributed with
 Z t 
EQ [x (t) | Fs ] = e −a(t−s) x (s) + e a(u−s) y (u)du and
s
Z t
 2
Var [x (t) | Fs ] = e −a(t−u) σ(u) du.
s

p. 271
Result follows directly from state variable representation
for x (t)
Proof.
Result for E [x (t) | Fs ] follows from martingale property of Ito integral.
Variance follows from Ito isometry
Z t
 2
Var [x (t) | Fs ] = e −2a(t−s) e −a(u−s) σ(u) du
s
Z t
 2
= e −a(t−u) σ(u) du.
s

◮ We will have a closer look

 at Rt 
EQ [x (t) | Fs ] = e −a(t−s) x (s) + s
e a(u−s) y (u)du later on.
◮ Note, that we can also write
Var [x (t) | Fs ] = y (t) − G ′ (s, t)2 y (s).

Auxilliary variable y (t) represents the (co-)variance process of x (t).

p. 272
Zero coupon bond options are important building blocks
✻
1

TE
✲
t TM

❄
K

Definition (Zero coupon bond (ZCB) option)

A zero coupon bond option is defined as an option with expiry time TE , ZCB
maturity time TM with TM ≥ TE , strike K , call/put flag φ ∈ {1, −1} and
payoff
V ZBO (TE ) = [φ (P(TE , TM ) − K )]+ .

◮ We are interested in present value V ZBO (t).

◮ We use TE -forward measure for valuation
 
V ZBO (t) = P(t, TE ) · ETE [φ (P(TE , TM ) − K )]+ | Ft .

p. 273
P(TE , TM ) is log-normally distributed with known
parameters

◮ We have for the forward bond price

ETE [P(TE , TM ) | Ft ] = P(t, TM )/P(t, TE ).
◮ From
2 R TE  2
P(t, TM ) −G(TE ,TM )x (TE )− G(TE ,T
2
M) e −a(TE −u) σ(u) du
P(TE , TM ) = e t
P(t, TE )

we get
◮ P(TE , TM ) is log-normally distributed with log-normal variance
Z TE
 2
ν 2 = Var [G(TE , TM )x (TE ) | Ft ] = G(TE , TM )2 e −a(TE −u) σ(u) du,
t

◮ we can apply Black’s formula for option pricing.

p. 274
ZCO prices are given by Black’s formula

Theorem (ZCO pricing formula)

The time-t price of a zero coupon bond option with expiry time TE , ZCB
maturity time TM with TM ≥ TE , strike K , call/put flag φ ∈ {1, −1} and
payoff
V ZBO (TE ) = [φ (P(TE , TM ) − K )]+
is given by

V ZBO (t) = P(t, TE ) · Black (P(t, TM )/P(t, TE ), K , ν, φ)

with log-normal bond price variance

Z TE
2 2
 2
ν = G(TE , TM ) e −a(TE −u) σ(u) du.
t

Proof.
Result follows from log-normal distribution property.

p. 275
Coupon bond options are further building blocks

✻
Cn

TE ✻
C1 ✻ ✻ ✻
✲
t T1 Tn

K
❄
Payoff at option expiry TE
" n
! #+
X
V (TE ) = Ci · P(TE , Ti ) −K .
i=1

p. 276
Coupon bond options are options on a basket of future
cash flows

Definition (Coupon bond option (CBO))

A coupon bond option is defined as an option with expiry time TE , future cash
flow payment times T1 , . . . , Tn (with Ti > TE ), corresponding cash flow values
C1 , . . . , Cn , a fixed strike price K , call/put flag φ ∈ {1, −1} and payoff
" " n
! #!+ #
CBO
X
V (TE ) = φ Ci P(TE , Ti ) −K .
i=1

◮ We cannot price CBO directly due to the basket structure.

◮ However, with some (not too strong) assumptions we can represent the
’option on a basket’ as a ’basket of options’.
◮ We use monotonicity of bond prices (for t < T )
∂
P(x (t); t, T ) = −G(t, T ) · P(x (t); t, T ) < 0.
∂x

p. 277
CBO’s are transformed via Jamshidian’s trick I

W.l.o.g. set φ = 1 (method works for φ = −1 as well).

Assume underlying bond is monotone in state variable x (TE ), i.e.

n n
∂ X X ∂
Ci P(x (TE ); TE , Ti ) = Ci P(x (TE ); TE , Ti )
∂x ∂x
i=1 i=1
n
X
=− Ci G(TE , Ti )P(x (TE ); TE , Ti ) < 0.
i=1

◮ Condition is satisfied, e.g. if Ci ≥ 0.

◮ Small negative cash flows typically don’t violate the assumption since last cash
flow Cn is typically a large positive cash flow.

p. 278
CBO’s are transformed via Jamshidian’s trick II
Then find x ⋆ such that

n
!
X
Ci P(x ⋆ ; TE , Ti ) −K =0
i=1

and set Ki = P(x ⋆ ; TE , Ti ).

We get (using monotonicity assumption)
" n
! #+ " n
! #
X X
Ci P(TE , Ti ) −K = ✶{x (TE )≤x ⋆ } Ci P(TE , Ti ) −K
i=1 i=1
" n n
#
X X
= ✶{x (TE )≤x ⋆ } Ci P(TE , Ti ) − Ci Ki
i=1 i=1
" n
#
X
= Ci [P(TE , Ti ) − Ki ] ✶{x (TE )≤x ⋆ }
i=1
" n
#
X
+
= Ci [P(TE , Ti ) − Ki ] .
i=1
p. 279
CBO’s are transformed via Jamshidian’s trick III

This gives
"" n
! #+ # n
X X  
E TE
Ci P(TE , Ti ) −K = Ci ETE [P(TE , Ti ) − Ki ]+
i=1 i=1
| {z }
Black’s formula

or

n
X
V CBO (t) = Ci · ViZBO (t)
i=1
n
X
= Ci · P(t, TE ) · Black (P(t, Ti )/P(t, TE ), K , νi , φ) ,
i=1
Z TE
 2
νi2 = G(TE , Ti )2 e −a(TE −u) σ(u) du.
t

p. 280
CBO’s are prices as sum of ZBO’s
Theorem (CBO pricing formula)
Consider a CBO with expiry time TE , future cash flow payment times
T1 , . . . , Tn (with Ti > TE ), corresponding cash flow values C1 , . . . , Cn , fixed
strike P price K and call/put flag φ ∈ {1, −1}. Assume that the underlying bond
n
price C P(x (TE ); TE , Ti ) is monotonically decreasing in the state variable
i=1 i
x (TE ). Then the time-t price of the CBO is
n
X
V CBO (t) = Ci · ViZBO (t)
i=1

where ViZBO (t) is the time-t price of a corresponding ZBO with strike
Ki = P(x ⋆ ; TE , Ti ) where the break-even state x ⋆ is given by
n
!
X
Ci P(x ⋆ ; TE , Ti ) − K = 0.
i=1

Proof.
Follows from derivation above.

p. 281
Outline

Hull-White Model
Classical Model Derivation
Relation to HJM Framework
Analytical Bond Option Pricing Formulas
General Payoff Pricing
Summary of Hull-White Pricing Formulas
European Swaption Pricing
Impact of Volatility and Mean Reversion

p. 282
We have another look at the expectation(s) of x (t)

◮ For general option pricing we also need expectation ET [x (T ) | Ft ].

◮ Then we can price
Z +∞
T
V (t) = P(t, T ) · E [V (x (T ); T ) | Ft ] = P(t, T ) · V (x ; T ) · pµ,σ2 (x ) · dx .
−∞

◮ Here pµ,σ2 (x ) is the density of a normal distribution N µ, σ 2 with

µ = ET [x (T ) | Ft ] and σ 2 = Var [x (T ) | Ft ] .

R +∞
◮ Integral −∞
V (x ; T ) · pµ,σ2 (x ) · dx is typically evaluated numerically
(i.e. quadrature).

◮ We first calculate EQ [x (T ) Ft ] and then derive ET [x (T ) Ft ].

p. 283
We calculate expectation in risk-neutral measure I
Recall
dx (t) = [y (t) − a · x (t)] · dt + σ(t) · dW (t).

This yields for T ≥ t

 Z T 
x (T ) = e −a(T −t) x (t) + e a(u−t) (y (u)du + σ(u)dW (u))
t

and Z T
EQ [x (T ) | Ft ] = e −a(T −t) x (t) + e −a(T −u) y (u)du.
t

We get
Z T Z T Z u 
−a(T −u) −a(T −u) 2 −2a(u−s)
e y (u)du = e σ(s) e ds du
t t 0
Z T Z t 
= e −a(T −u) σ(s)2 e −2a(u−s) ds du
t 0
Z T Z u 
+ e −a(T −u) σ(s)2 e −2a(u−s) ds du.
t t

p. 284
We calculate expectation in risk-neutral measure II
We analyse the integrals individually,
Z T Z t 
I1 (t, T ) = e −a(T −u) σ(s)2 e −2a(u−s) ds du
t 0
Z T Z t 
−a(T −u) 2 −2a(u−s)
= e σ(s) e ds du
t 0
Z t Z T 
= e −a(T −u) σ(s)2 e −2a(u−s) du ds
0 t
Z t Z T 
= σ(s)2 e −a(T −u) e −2a(u−s) du ds
0 t
Z t  T
e −a(T −u) e −2a(u−s)
= σ(s)2 ds
0
−a
u=t
Z t
σ(s)2  −a(T −t) −2a(t−s) 
= e e − e −a(T −T ) e −2a(T −s) ds.
0
a

Exponential terms can be further simplified as

 
e −a(T −t) e −2a(t−s) − e −2a(T −s) = e −a(T −t) 1 − e −a(T −t) e −2a(t−s) .
p. 285
We calculate expectation in risk-neutral measure III
This gives Z t
−a(T −t) 1 − e −a(T −t)
I1 (t, T ) = e σ(s)2 e −2a(t−s) ds.
a 0

For the second integral we get

Z T Z u 
I2 (t, T ) = e −a(T −u) σ(s)2 e −2a(u−s) ds du
t t
Z T Z u 
−a(T −u) 2 −2a(u−s)
= e σ(s) e ds du
t t
Z T Z T 
= e −a(T −u) σ(s)2 e −2a(u−s) du ds
t s
Z T Z T 
= σ(s)2 e −a(T −u) e −2a(u−s) du ds
t s
Z T  T
e −a(T −u) e −2a(u−s)
= σ(s)2 ds
t
−a
u=s
Z T
σ(s)2  −a(T −s) −2a(s−s) 
= e e − e −a(T −T ) e −2a(T −s) ds.
t
a
p. 286
We calculate expectation in risk-neutral measure IV

Again we simplify exponential terms

 
e −a(T −s) − e −2a(T −s) = e −a(T −s) 1 − e −a(T −s) .

This gives
Z T
1 − e −a(T −s)
I2 (t, T ) = σ(s)2 e −a(T −s) ds.
t
a

In summary, we get

EQ [x (T ) | Ft ] = e −a(T −t) x (t) + I1 (t, T ) + I2 (t, T )

 Z t 
−a(T −t) 1 − e −a(T −t) 2 −2a(t−s)
=e x (t) + σ(s) e ds
a 0
Z T
1 − e −a(T −s)
+ σ(s)2 e −a(T −s) ds.
t
a

p. 287
We calculate expectation in terminal measure I
Recall change of measure

dW T (t) = dW (t) + σP (t, T )dt.

We have
1 − e −a(T −t)
σP (t, T ) = σ(t)G(t, T ) = σ(t) · .
a
This gives
 
dx (t) = y (t) − σ(t)2 G(t, T ) − a · x (t) · dt + σ(t) · dW T (t)

and
 Z T 
−a(T −t) a(u−t)
 2
 T


x (T ) = e x (t) + e y (u) − σ(u) G(u, T ) du + σ(u)dW (u) .
t

We find that
Z T
ET [x (T ) | Ft ] = EQ [x (T ) | Ft ] − σ(u)2 e −a(T −u) G(u, T )du.
t

p. 288
We calculate expectation in terminal measure II

It turns out that

Z T Z T
1 − e −a(T −u)
σ(u)2 e −a(T −u) G(u, T )du = σ(u)2 e −a(T −u) du = I2 (t, T ).
t t
a

As a result, we get
 Z t 
1 − e −a(T −t)
ET [x (T ) | Ft ] = e −a(T −t) x (t) + σ(s)2 e −2a(t−s) ds
a 0

or more formally

ET [x (T ) | Ft ] = G ′ (t, T ) [x (t) + G(t, T )y (t)] .

p. 289
Outline

Hull-White Model
Classical Model Derivation
Relation to HJM Framework
Analytical Bond Option Pricing Formulas
General Payoff Pricing
Summary of Hull-White Pricing Formulas
European Swaption Pricing
Impact of Volatility and Mean Reversion

p. 290
All the formulas serve the purpose of model calibration and
derivative pricing

Model Calibration Derivative Pricing

zero bond option (ZBO) future zero bonds P(x (t); t, T )

expectation ET [x (T ) | Ft ] and
coupon bond option (CBO)
variance Var [x (T ) | Ft ]

payoff pricing V (t) =

European swaption P(t, T ) · ET [V (x (T ); T ) | Ft ]

p. 291
Bond option pricing is realised via ZBO’s and CBO’s

Zero Bond Option (ZBO)

Zero bond with expiry TE , maturity TM , strike K and call/put flag φ

V ZBO (0) = P(0, TE ) · Black (P(0, TM )/P(0, TE ), K , ν, φ) ,

ν 2 = G(TE , TM )2 y (TE ).

Coupon Bond Option (CBO) Pn

Coupon bond option with strike K and underlying bond i=1
Ci · P(TE , Ti ),

n
X
V CBO (t) = Ci · ViZBO (t)
i=1

where ZBO’s ViZBO (t) with expiry TE , maturity Ti , and strike Ki = P(x ⋆ , TE , Ti ) and
x ⋆ s.t.
n
X
Ci · P(x ⋆ ; TE , Ti ) = K .
i=1

p. 292
General derivative pricing requires state variable
expectation and variance
Zero Bonds (as building blocks for payoffs V (x (T ); T ))
 
P(0, S) G(T , S)2
P(x (T ); T , S) = exp −G(T , S)x (T ) − y (T ) .
P(0, T ) 2

General Derivative Pricing

Z +∞
V (t) = P(t, T ) · ET [V (x (T ); T ) | Ft ] = P(t, T ) · V (x ; T ) · pµ,σ2 (x ) · dx
−∞

with pµ,σ2 (x ) density of a Normal distribution N µ, σ 2 with

µ = ET [x (T ) | Ft ] = G ′ (t, T ) [x (t) + G(t, T )y (t)]

and
σ 2 = Var [x (T ) | Ft ] = y (T ) − G ′ (t, T )2 y (t).

p. 293
Fortunately, we only need a small set of model functions
for implementation

◮ Discount factors P(0, T ) from input yield curve.

◮ Function G(t, T ) with
1 − e −a(T −t)
G(t, T ) = .
a
◮ Function G ′ (t, T ) with
G ′ (t, T ) = e −a(T −t) .
◮ Auxilliary variable y (t) with
Z t k
 2 X e −2a(t−tj ) − e −2a(t−tj−1 )
y (t) = e −a(t−u) σ(u) du = σj2
0
2a
j=1

where we assume σ(t) piece-wise constant on a grid 0 = t0 , t1 , . . . , tk = t.

p. 294
Outline

Hull-White Model
Classical Model Derivation
Relation to HJM Framework
Analytical Bond Option Pricing Formulas
General Payoff Pricing
Summary of Hull-White Pricing Formulas
European Swaption Pricing
Impact of Volatility and Mean Reversion

p. 295
It remains to show how Hull-Wite model is applied to
European swaptions

Model Calibration Derivative Pricing

zero bond option (ZBO) future zero bonds P(x (t); t, T )

expectation ET [x (T ) | Ft ] and
coupon bond option (CBO)
variance Var [x (T ) | Ft ]

payoff pricing V (t) =

European swaption P(t, T ) · ET [V (x (T ); T ) | Ft ]

p. 296
Recall that Swaption is option to enter into a swap at a
future time
K
✻ ✻
T0 Tn ✲
TE T̃0 T̃m

❄ ❄ ❄ ❄
Lm
◮ At option exercise time TE present value of swap is

n
X m
X
V Swap (TE ) = K · τi · P(TE , Ti ) − Lδ (TE , T̃j−1 , T̃j−1 + δ) · τ̃j · P(TE , T̃j ) .
i=1 j=1
| {z } | {z }
future fixed leg future float leg

◮ Option to enter represents the right but not the obligation to enter swap.
◮ Rational market participant will exercise if swap present value is positive, i.e.

V Swpt (TE ) = max V Swap (TE ), 0 .

p. 297
How do we get the swaption payoff compatible to our
Hull-White model formulas?

n m
X X
V Swap (TE ) = K · τi · P(TE , Ti ) − Lδ (TE , T̃j−1 , T̃j−1 + δ) · τ̃j · P(TE , T̃j )
i=1 j=1
| {z } | {z }
future fixed Leg future float leg

◮ Fixed leg can be expressed in terms of future state variable x (TE ) via
P(x (TE ); TE , Ti )

◮ Float leg contains future forward Libor rates Lδ (TE , T̃j−1 , T̃j−1 + δ) from
(future) projection curve

◮ However, Hull-White model only provides representation of discount

factors, i.e. P(TE , T̃j )

We need to model the relation between future Libor rates

Lδ (TE , T̃j−1 , T̃j−1 + δ) and discount factors P(TE , T̃j ).

p. 298
We do have all ingredients from our deterministic
multi-curve model
Recall the definition of (future) forward Libor rate
 
Lδ (TE , T̃j−1 , T̃j−1 + δ) = ET̃j−1 +δ Lδ (T̃j−1 , T̃j−1 , T̃j−1 + δ) | FTE
 
P(TE , T̃j−1 ) 1
= · D(T̃j−1 , T̃j−1 + δ) − 1
P(TE , T̃j−1 + δ) τ (T̃j−1 , T̃j−1 + δ)

with tenor basis spread discount factor

Q(TE , T̃j−1 )
D(T̃j−1 , T̃j−1 + δ) =
Q(TE , T̃j−1 + δ)

and discount factors Q(TE , T ) arising from credit (or funding) risk embedded in Libor
rates Lδ (·).
◮ Key assumption is that D(T̃j−1 , T̃j−1 + δ) is deterministic or independent of TE .
◮ Then

Q(0, T̃j−1 ) P δ (0, T̃j−1 ) P(0, T̃j−1 + δ)

D(T̃j−1 , T̃j−1 + δ) = = · .
Q(0, T̃j−1 + δ) P δ (0, T̃j−1 + δ) P δ (0, T̃j−1 )

p. 299
We use basis spread model to simplify Libor coupons

◮ Basis spread discount factor

P δ (0, T̃j−1 ) P(0, T̃j−1 + δ)

D(T̃j−1 , T̃j−1 + δ) = ·
δ
P (0, T̃j−1 + δ) P δ (0, T̃j−1 )

is calculated from today’s projection curve P δ (0, T ) and discount curve

P δ (0, T ).
◮ Further assume natural Libor payment dates and consistent year fractions

T̃j = T̃j−1 + δ, τ (T̃j−1 , T̃j−1 + δ) = τ̃j .

◮ Libor coupon becomes

 
P(TE , T̃j−1 ) 1
Lδ (TE , T̃j−1 , T̃j ) · τ̃j · P(TE , T̃j ) = · D(T̃j−1 , T̃j ) − 1 · τ̃j · P(TE , T̃j )
P(TE , T̃j ) τ̃j
= P(TE , T̃j−1 ) · D(T̃j−1 , T̃j ) − P(TE , T̃j ).

p. 300
We can write the float leg ... I

n m
X X
V Swap (TE ) = K · τi · P(TE , Ti ) − Lδ (TE , T̃j−1 , T̃j−1 + δ) · τ̃j · P(TE , T̃j )
i=1 j=1
| {z } | {z }
future fixed leg future float leg
n m
X X
=K· τi · P(TE , Ti ) − P(TE , T̃j−1 ) · D(T̃j−1 , T̃j ) − P(TE , T̃j )
i=l j=1
n
X
=K· τi · P(TE , Ti )
i=1
" m
#
X  
− P(TE , T̃0 ) · D(T̃0 , T̃1 ) − P(TE , T̃m ) + P(TE , T̃j−1 ) · D(T̃j−1 , T̃j ) − 1
j=2
n
X
=K· τi · P(TE , Ti )
i=1
" m
#
X  
− P(TE , T̃0 ) − P(TE , T̃m ) + P(TE , T̃j−1 ) · D(T̃j−1 , T̃j ) − 1 .
j=1
p. 301
We can write the float leg ... II
Reordering terms yields
n
X
V Swap (TE ) = − P(TE , T̃0 ) + K · τi · P(TE , Ti )
| {z }
i=1
strike paid at T0 | {z }
fixed rate coupons
m
X  
− P(TE , T̃j−1 ) · D(T̃j−1 , T̃j ) − 1 + P(TE , T̃m )
| {z }
j=1
notional payment
| {z }
negative spread coupons
n+m+1
X
= Ck · P(TE , T̄k )
k=0

with
 
C0 = −1, Ci = K · τi (i = 1, . . . , n), Cn+j = − D(T̃j−1 , T̃j ) − 1 , (j = 1, . . . , m),

and Cn+m+1 = 1,

and corresponding payment times T̄k .

p. 302
Swaptions are equivalent to coupon bond options
Corollary (Equivalence between Swaption and bond option)
Consider a European Swaption with receiver/payer flag φ ∈ {1, −1} payoff
" ( n m
)#+
X X
Swpt δ
V (TE ) = φ K· τi · P(TE , Ti ) − L (TE , T̃j−1 , T̃j−1 + δ) · τ̃j · P(TE , T̃j ) .
i=1 j=1

Under our deterministic basis spread assumption the swaption payoff is equal to a
call/put bond option payoff
" (n+m+1 )#+
X
V CBO (TE ) = φ Ck · P(TE , T̄k )
k=0

with zero strike and cash flows Ck and times T̄k as elaborated above. Moreover, if the
underlying bond payoff is monotonic then

n+m+1
X
V Swpt (t) = V CBO (t) = Ck · VkZBO (t)
k=0

with corresponding zero bond option parameters.

p. 303
We give some comments regarding the CBO maping
◮ Note that C0 = −1 is a large negative cash flow.
∂
 
◮ However, ∂x
−P(TE , T̃0 ) ≈ −G(TE , T0 ) is small because TE − T0 is
small.

◮ If TE = T̃0 , i.e. no spot offset between option expiry and swap start time,
then
◮ set CBO strike K = D(T̃0 , T̃1 ),
◮ remove first negative spread coupon Cn+1 from cash flow list.

◮ In practice monotonicity assumption

"n+m+1 #
∂ X
Ck · P(TE , T̄k ) < 0
∂x
k=0

is typically no issue.

In Hull-White model calibration we will use CBO formula to match Hull-White

model prices versus Vanilla model swaption prices.

p. 304
Outline

Hull-White Model
Classical Model Derivation
Relation to HJM Framework
Analytical Bond Option Pricing Formulas
General Payoff Pricing
Summary of Hull-White Pricing Formulas
European Swaption Pricing
Impact of Volatility and Mean Reversion

p. 305
How do the simulated paths look like?
◮ Model short rate volatility σ calibrated to 100bp flat volatility at 5y and 10y ,
mean reversion a ∈ {−5%, 0%, 5%} 5

◮ Higher mean reversion yields more forward volatility.

5
Zero mean reversion is effectively approximated via a = 1bp. This does not change the overall behavior and
avoids special treatment in formulas. p. 306
Forward volatility dependence on mean reversion can also
be derived analytically
Denote forward volatility as
 
Var x (T1 ) | FT0 y (T1 ) − G ′ (T0 , T1 )2 y (T0 )
σFwd (T0 , T1 ) = √ = √
T1 − T0 T1 − T0

◮ Suppose spot volatilities σFwd (0, T1 ) and σFwd (0, T0 ) (and thus y (T0 ) and
y (T1 ) are fixed)
◮ If mean reversion a increases then G ′ (T0 , T1 ) = e −a(T1 −T0 ) decreases
◮ Thus forward volatility σFwd (T0 , T1 ) increases

p. 307
Which kind of curves can we simulate with Hull-White
model?
◮ Models use flat short rate volatility σ = 100bp and mean reversion
a ∈ {−5%, 0%, 5%} 6

◮ Model works with negative mean reversion - however, yield curves are exploding

6
Zero mean reversion is effectively approximated via a = 1bp. This does not change the overall behavior and
avoids special treatment in formulas. p. 308
What are relevant properties of a model for option pricing?

◮ Vanilla models require input (ATM volatility) parameters for

expiry-tenor-pairs.
◮ Which shape of ATM volatilities for expiry-tenor-pairs are predicted
by Hull-White model?

◮ SABR model allows modelling of volatility smile.

◮ Which shapes of volatility smile can be modelled with Hull-White
model?
◮ How does the smile change if we change the model parameters?

◮ We aim at applying the Hull-White model to price Bermudan swaptions.

◮ How do the model parameters impact prices of exotic derivatives?

For now we focus on model-implied volatilities (ATM and smile). The impact
of model parameters on Bermudans is analysed later.

p. 309
Model properties for option pricing are assessed by
analysing model-implied volatilities

Model-implied normal volatility

Consider a swaption with expiry/start/end-dates TE /T0 /Tn and strike rate K .
For a given Hull-White model the model-implied normal volatility is calculated
as

p
σ(T0 , Tn , K ) = Bachelier−1 S(t), K , V CBO (t)/An(t), φ / TE − t.

Here, S(t) and An(t) are the forward swap rate and annuity of the underlying
swap with start/end-date T0 /Tn . V CBO (t) is the Hull-White model price of a
coupon bond option equivalent to the input swaption.

p. 310
Which shapes of volatility smile can be modelled and how
does the smile change if we change the model parameters?
◮ Models use flat short rate volatility σ ∈ {50bp, 75bp, 100bp, 125bp} and mean
reversion a ∈ {−5%, 0%, 5%}:

◮ We can only model flat smile - this is a major model limitation!

◮ Model-implied volatility decreases if mean reversion increases.

p. 311
Which shape of ATM volatilities for expiry-tenor-pairs are
predicted by Hull-White model?
◮ Models use flat short rate volatility σ - calibrated to 10y-10y swaption with
100bp volatility
◮ Mean reversion a ∈ {−5%, 0%, 5%}:

◮ Mean reversion impacts slope of ATM volatilities in expiry and swap term
dimension.

p. 312
Contact

Dr. Sebastian Schlenkrich

Office: RUD25, R 1.211
Mail: [email protected]

d-fine GmbH
Mobile: +49-162-263-1525
Mail: [email protected]