Interest Rate Modelling and Derivative Pricing

Sebastian Schlenkrich

HU Berlin, Department of Mathematics

WS, 2019/20
 Part V

Bermudan Swaption Pricing

p. 313
Outline

Bermudan Swaptions

Pricing Methods for Bermudans

Density Integration Methods

PDE and Finite Differences

American Monte Carlo

p. 314
Outline

Bermudan Swaptions

Pricing Methods for Bermudans

Density Integration Methods

PDE and Finite Differences

American Monte Carlo

p. 315
Let’s have another look at the cancellation option

Interbank swap deal example

Bank A may decide to early terminate deal in 10, 11, 12,..years.

p. 316
What does such a Bermudan call right mean?
L1 Lm
✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻
1
T̃k−1 2
T̃k−1 3
T̃k−1 T̃m
T̃0 ✲
T0 TE1 1
Tl−1 TE2 2
Tl−1 TE3 3
Tl−1 Tn

❄ ❄ ❄ ❄ ❄ ❄
K K

[Bermudan cancellable swap] = [full swap] + [Bermudan option on opposite swap]

K
✻ ✻ ✻
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m

❄ ❄ ❄ ❄ ❄ ❄
Lm

p. 317
What is a Bermudan swaption?

K
✻ ✻ ✻
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m

❄ ❄ ❄ ❄ ❄ ❄
Lm

Bermudan swaption
A Bermudan swaption is an option to enter into a Vanilla swap with fixed rate
K and final maturity Tn at various exercise dates TE1 , TE2 , . . . , TEk̄ . If there is
only one exercise date (i.e. k̄ = 1) then the Bermudan swaption equals a
European swaption.

p. 318
A Bermudan swaption can be priced via backward
induction
K
✻ ✻ ✻
continuation value
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m

❄ ❄ ❄ ❄ ❄ ❄
Lm

K
✻ ✻ ✻
exercise payoff
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m

❄ ❄ ❄ ❄ ❄ ❄
Lm

p. 319
A Bermudan swaption can be priced via backward
induction - let’s add some notation
H1 = . . . H2 = . . . H3 = 0 K
✻ ✻ ✻
continuation value
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m
h i
H0 = B(t)E V1
B(T 1 )
| Ft ❄ ❄ ❄ ❄ ❄ ❄
E
Lm
V3 = max{U3 , H3 }
V2 = max{U2 , H2 }
V1 = max{U1 , H1 }
U1 U2 U3 K
✻ ✻ ✻
exercise payoff
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m

❄ ❄ ❄ ❄ ❄ ❄
Lm

p. 320
First we specify the future payoff cash flows
◮ Choose a numeraire B(t) and corresponding cond. expectations Et [·] = E[· | Ft ].
◮ Underlying payoff Uk if option is exercised

X h i
Xi (Ti )
Uk = B(TEk ) ET k
E B(Ti )
Ti ≥T k
E
 
X X
= B(TEk )  K · τi · P(TEk , Ti ) − Lδ (TEk , T̃j−1 , T̃j−1 + δ) · τ̃j · P(TEk , T̃j )
Ti ≥T k T̃j ≥T k
E E
| {z }
future fixed leg minus future float leg

X
= B(TEk )  K · τi · P(TEk , Ti )
Ti ≥T k
E

X  
−P(TEk , T̃jk ) − P(TEk , T̃j−1 ) · D(T̃j−1 , T̃j ) − 1 + P(TEk , T̃m ) .
T̃j ≥T k
E

p. 321
Then we specify the continuation value and optimal
exercise
◮ Continuation value Hk (t) (TEk ≤ t ≤ TEk+1 ) represents the time-t value of the
remaining option if not exercised.
◮ Option becomes worthless if not exercises at last exercise date TEk̄ . Thus last
continuation value Hk̄ (TEk̄ ) = 0.
◮ Recall that Bermudan option gives the right but not the obligation to enter into
underlying at exercise.
◮ Rational agent will choose the maximum of payoff and continuation at exercise,
i.e. 
Vk = max Uk , Hk (TEk ) .

◮ Vk represents the Bermudan option value at exercise TEk . Thus we also must
have for the continuation value

Hk−1 (TEk ) = Vk .

◮ Derivative pricing formula yields

   
Hk−1 (TEk ) Vk
Hk−1 (TEk−1 ) = B(TEk−1 ) · ET k−1 = B(TEk−1 ) · ET k−1 .
E B(TEk ) E B(TEk )

p. 322
We summarize the Bermudan pricing algorithm

Backward induction for Bermudan options

Consider a Bermudan option with k̄ exercise dates TEk (k = 1, . . . k̄) and underlying
future payoffs with (time-TEk ) prices Uk .

Denote Hk (t) the option’s continuation value for TEk ≤ t ≤ TEk+1 and set


Hk̄ TEk̄ = 0. Also set TE0 = t (i.e. pricing time today).

The option price can be derived via the recursion

  "  #

Hk (TEk+1 ) max Uk+1 , Hk+1 (TEk+1 )
k
Hk TE = B(TEk ) · ET k = B(TEk ) · ET k
E B(TEk+1 ) E B(TEk+1 )

for k = k̄ − 1, . . . , 0. The Bermudan option price is given by

V Berm (t) = H0 (t) = H0 (TE0 ).

p. 323
Some more comments regarding Bermudan pricing ...
◮ Recursion for Bermudan pricing can be formally derived via theory of optimal
stopping and Hamilton-Jacobi-Bellman (HJB) equation.
◮ For more details, see Sec. 18.2.2 in Andersen/Piterbarg (2010).
◮ For a single exercise date k̄ = 1 we get
 
max {U1 , 0)}
H0 (t) = B(t) · Et .
B(TE1 )

This is the general pricing formula for a European swaption (if U1 represents a
Vanilla swap).
  

max Uk+1 ,Hk+1 (T k+1 )
E
◮ In principle, recursion Hk TEk = B(TEk ) · ET k holds
E B(T k+1 )
E
for any payoffs Uk . However, computation

X h i
Xi (Ti )
Uk = B(TEk ) ET k
E B(Ti )
Ti ≥T k
E

might pose additional challenges if cash flows Xi (Ti ) are more complex.

p. 324
How do we price a Bermudan in practice?

◮ In principle, recursion algorithm for Hk () is straight forward.

◮ Computational challenge is calculating conditional expectations
"  #

max Uk+1 , Hk+1 (TEk+1 )
Hk TEk = B(TEk ) · ET k .
E B(TEk+1 )

◮ Note, that this problem is an instance of the general option pricing

problem  
V (T1 )
V (T0 ) = B(T0 ) · E | FT0 .
B(T1 )

We can apply general option pricing methods to roll-back the Bermudan payoff.

p. 325
Outline

Bermudan Swaptions

Pricing Methods for Bermudans

Density Integration Methods

PDE and Finite Differences

American Monte Carlo

p. 326
Note that Uk , Vk and Hk depend on underlying state
variable x (TEk )
H1 = . . . H2 = . . . H3 = 0 K
✻ ✻ ✻
continuation value
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m
h i
H0 = B(t)E V1
B(T 1 )
| Ft ❄ ❄ ❄ ❄ ❄ ❄
E
Lm
V3 = max{U3 , H3 }
V2 = max{U2 , H2 }
V1 = max{U1 , H1 }
U1 U2 U3 K
✻ ✻ ✻
exercise payoff
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 T̃k−1
1 TE2 T̃k−1
2 TE3 T̃k−1
3 T̃m

❄ ❄ ❄ ❄ ❄ ❄
Lm

p. 327
Typically we need to discretise variables Uk , Vk and Hk on
a grid of underlying state variables

Forthcomming, we discuss several methods to roll-back the payoffs.

p. 328
Outline

Bermudan Swaptions

Pricing Methods for Bermudans

Density Integration Methods

PDE and Finite Differences

American Monte Carlo

p. 329
Outline

Density Integration Methods

General Density Integration Method
Gauss–Hermite Quadrature
Cubic Spline Interpolation and Exact Integration

p. 330
Key idea using the conditional density function in the
Hull-White model

Recall that
 
V (T1 )
V (T0 ) = B(T0 ) · E | FT0 .
B(T1 )

We choose the T1 -maturing zero coupon bond P(t, T1 ) as numeraire. Then

V (T0 ) = P(T0 , T1 ) · ET1 [V (T1 ) | FT0 ]

Z +∞
= P(x (T0 ); T0 , T1 ) · V (x ; T1 ) · pµ,σ2 (x ) · dx .
−∞

State variable x = x (T1 ) is normally distributed with known mean and variance.

p. 331
Hull-White model results yield density parameters of the
state variable x (T1 )
Z +∞
V (T0 ) = P(x (T0 ); T0 , T1 ) · V (x ; T1 ) · pµ,σ2 (x ) · dx .
−∞

State variable x = x (T1 ) is normally distributed with mean

µ = ET1 [x (T1 ) | FT0 ] = G ′ (T0 , T1 ) [x (T0 ) + G(T0 , T1 )y (T0 )]

and variance

σ 2 = Var [x (T1 ) | FT0 ] = y (T1 ) − G ′ (T0 , T1 )2 y (T0 ).

Thus  
1 (x − µ)2
pµ,σ2 (x ) = √ · exp −
2πσ 2 2σ 2
and
Z +∞  
V (x ; T1 ) (x − µ)2
V (T0 ) = P(x (T0 ); T0 , T1 ) · √ · exp − dx .
−∞ 2πσ 2 2σ 2

p. 332
Integral against normal density needs to be computed
numerically by quadrature methods
Z +∞  
V (x ; T1 ) (x − µ)2
V (T0 ) = P(x (T0 ); T0 , T1 ) · √ · exp − dx .
−∞ 2πσ 2 2σ 2
◮ We can apply general purpose quadrature rules to function
 
V (x ; T1 ) (x − µ)2
f (x ) = √ · exp − .
2πσ 2 2σ 2

◮ Select a grid [x0 , . . . , xN ] and approximate e.g. via

◮ Trapezoidal rule
Z +∞ N
X 1
f (x ) · dx ≈ [f (xi−1 ) + f (xi )] (xi − xi−1 )
−∞
2
i=1

◮ or Simpson’s rule with equidistant grid (h = xi − xi−1 ) and even

sub-intervalls, then
Z " N/2−1 N/2
#
+∞ X X
h
f (x )·dx ≈ · f (x0 ) + 2 f (x2j ) + 4 f (x2j−1 ) + f (xN ) .
−∞
3
j=1 j=1
p. 333
There are some details that need to be considered - Select
your integration domain carefully

◮ Infinite integral is approximated by definite integral

Z +∞ Z xN
f (x ) · dx ≈ f (x ) · dx ≈ · · · .
−∞ x0

◮ Finite integration boundaries need to be chosen carefully by taking into

account variance of x (t).

◮ One approach is calculating variance to option expiry T1 ,

σ̂ 2 = Var [x (T )] = y (T1 ) and set

x0 = −λ · σ̂ and xN = λ · σ̂.

◮ Note, that ET1 [x (T1 )] = 0, thus we don’t apply a shift to the x -grid.

p. 334
There are some details that need to be considered - Take
care of the break-even point

◮ Note that convergence of quadrature rules depends on smoothness of

integrand f (x ).
◮ Recall that

f (x ) = V (x ) · pµ,σ2 (x ) = max Uk+1 (x ), Hk+1 (x ; TEk+1 ) · pµ,σ2 (x ).

◮ Max-function is not smooth at Uk+1 (x ) = Hk+1 (x ; TEk+1 ).

Determine x ⋆ (via interpolation of Hk+1 (·) and numerical root search) such that

Uk+1 (x ⋆ ) = Hk+1 (x ⋆ ; TEk+1 )

and split integration

Z +∞ Z x⋆ Z +∞
f (x ) · dx = f (x ) · dx + f (x ) · dx .
−∞ −∞ x⋆

p. 335
Can we exploit the structure of the integrand?

Z +∞  
V (x ; T1 ) (x − µ)2
V (T0 ) = P(x (T0 ); T0 , T1 ) · √ · exp − dx .
−∞ 2πσ 2 2σ 2

◮ Integral against normal distribution density can be solved more efficiently:

1. Use Gauss–Hermite quadrature.

2. Interpolate only V (x ; T1 ) by cubic spline and integrate exact.

p. 336
Outline

Density Integration Methods

General Density Integration Method
Gauss–Hermite Quadrature
Cubic Spline Interpolation and Exact Integration

p. 337
Gauss–Hermite quadrature is an efficient integration
method for smooth integrands

◮ Gauss–Hermite quadrature is a particular form of Gaussian quadrature.

◮ Choose a degree parameter d, and approximate
Z +∞ d
X
2
f (x ) · e −x dx ≈ wk · f (xk )
−∞ k=1

with xk (i = 1, 2, ..., d) being the roots of the physicists’ version of the

Hermite polynomial Hd (x ) and wk are weights with
√
2d−1 d! π
wk = .
d 2 [Hd−1 (xk )]2
◮ Roots and weights can be obtained, e.g. via stored values and look-up
tables.

p. 338
Variable transformation allows application of
Gauss–Hermite quadrature to Hull-White model integration
We get
Z +∞   Z +∞
V (x ; T1 ) (x − µ)2 1 √ 2
√ · exp − dx = √ V ( 2σx + µ; T1 ) · e −x dx
−∞ 2πσ 2 2σ 2 π −∞
d
1 X √
≈ √ wk · V ( 2σxk + µ; T1 ).
π
k=1

Some constraints need to be considered:

◮ Payoff V (·, T1 ) is only available on the x -grid at T1 , thus V (·, T1 ) needs

to be interpolated.

◮ Gauss-Hermite quadrature does not take care of non-smooth payoff at

break-even state x ⋆ , thus d needs to be sufficiently large to mitigate
impact.

p. 339
Outline

Density Integration Methods

General Density Integration Method
Gauss–Hermite Quadrature
Cubic Spline Interpolation and Exact Integration

p. 340
If we apply cubic spline interpolation anyway then we can
also integrate exactly
Approximate V (·, T1 ) via cubic spline on the grid [x0 , . . . xN ] as
N−1 d
X X
V (x , T1 ) ≈ C (x ) = ✶{xi ≤x <xi+1 } ck · (x − xi )k .
i=0 k=0

Then
Z +∞ N−1 Z
X xi+1 d
X
V (x ; T1 ) · pµ,σ2 (x ) · dx ≈ ck · (x − xi )k · pµ,σ2 (x ) · dx
−∞ i=0 xi k=0
N−1 d
X X Z xi+1
= ck · (x − xi )k · pµ,σ2 (x ) · dx .
i=0 k=0 xi

Thus, all we need is

Z xi+1
Ii,k = (x − xi )k · pµ,σ2 (x ) · dx .
xi

p. 341
We transform variables to make integration easier
First we apply the variable transformation x̄ = (x − µ)/σ. This yields
pµ,σ2 (x ) = p0,1 (x̄ )/σ and
Z x̄i+1
dx
Ii,k = (σx̄ + µ − xi )k · p0,1 (x̄ ) ·
x̄i
σ
Z x̄i+1  
k 1 k x̄ 2
= σ (x̄ − x̄i ) · √ exp − d x̄
x̄i 2π 2
| {z }
standard normal density

with the shifted grid points x̄i = (xi − µ)/σ.

Denote Φ(·) the cumulated standard normal distribution function. Then
 
1 x̄ 2
Φ′ (x ) = √ exp − and Φ′′ (x ) = −x Φ′ (x ).
2π 2

As a sub-step we aim at solving the integral

Z x̄i+1
x̄ k · Φ′ (x̄ ) · d x̄ .
x̄i

p. 342
We use cubic splines (d = 3) to keep formulas reasonably
simple I
It turnes out that
Z
F0 (x̄ ) = Φ′ (x̄ )d x̄ = Φ(x̄ ),
Z
F1 (x̄ ) = x̄ Φ′ (x̄ )d x̄ = −Φ′ (x̄ ),
Z
F2 (x̄ ) = x̄ 2 Φ′ (x̄ )d x̄ = Φ(x̄ ) − x · Φ′ (x̄ ),
Z


F3 (x̄ ) = x̄ 3 Φ′ (x̄ )d x̄ = − x̄ 2 + 2 · Φ′ (x̄ ).

This yields for Ii,0

Z x̄i+1
Ii,0 = Φ′ (x̄ ) · dx = F0 (x̄i+1 ) − F0 (x̄i )
x̄i

p. 343
We use cubic splines (d = 3) to keep formulas reasonably
simple II
and for Ii,1
Z x̄i+1
Ii,1 = σ (x̄ − x̄i ) · Φ′ (x̄ ) · dx
x̄i
Z x̄i+1
=σ· x̄ · Φ′ (x̄ ) · dx − σ · x̄i · Ii,0
x̄i
= σ · [F1 (x̄i+1 ) − F1 (x̄i )] − σ · x̄i · Ii,0 .

We may proceed similarly for Ii,2

Z x̄i+1
Ii,2 = σ 2 (x̄ − x̄i )2 · Φ′ (x̄ ) · dx
x̄i
Z x̄i+1
 
= σ 2 x̄ 2 − 2x̄i x̄ + x̄i2 · Φ′ (x̄ ) · dx
x̄i

= σ 2 [F2 (x̄i+1 ) − F2 (x̄i )] − 2σ 2 x̄i [F1 (x̄i+1 ) − F1 (x̄i )] + σ 2 x̄i2 Ii,0

= σ 2 [F2 (x̄i+1 ) − F2 (x̄i )] − 2σx̄i [Ii,1 + σ · x̄i · Ii,0 ] + σ 2 x̄i2 Ii,0
= σ 2 [F2 (x̄i+1 ) − F2 (x̄i )] − 2σx̄i Ii,1 − σ 2 x̄i2 Ii,0
p. 344
We use cubic splines (d = 3) to keep formulas reasonably
simple III
and Ii,3

Z x̄i+1
Ii,3 = σ 3 (x̄ − x̄i )3 · Φ′ (x̄ ) · dx
x̄i
Z x̄i+1
 
= σ 3 x̄ 3 − 3x̄i x̄ 2 + 3x̄i2 x̄ − x̄i3 · Φ′ (x̄ ) · dx
x̄i

= σ 3 [F3 (x̄i+1 ) − F3 (x̄i )] − 3σ 3 x̄i [F2 (x̄i+1 ) − F2 (x̄i )]

+ 3σ 3 x̄i2 [F1 (x̄i+1 ) − F1 (x̄i )] − σ 3 x̄i3 Ii,0 .

Substituting terms as before yields

 
Ii,3 = σ 3 [F3 (x̄i+1 ) − F3 (x̄i )] − 3σx̄i Ii,2 + 2σx̄i Ii,1 + σ 2 x̄i2 Ii,0
+ 3σ 2 x̄i2 [Ii,1 + σ · x̄i · Ii,0 ] − σ 3 x̄i3 Ii,0
= σ 3 [F3 (x̄i+1 ) − F3 (x̄i )] − 3σx̄i Ii,2 − 3σ 2 x̄i2 Ii,1 − σ 3 x̄i3 Ii,0 .

p. 345
Let’s summarise the formulas...

We get
Z +∞
V (T0 ) = P(x (T0 ); T0 , T1 ) · V (x ; T1 ) · pµ,σ2 (x ) · dx
−∞
N−1 3
X X
≈ P(x (T0 ); T0 , T1 ) · ck · Ii,k
i=0 k=0

with

Ii,0 = F0 (x̄i+1 ) − F0 (x̄i )

Ii,1 = σ · [F1 (x̄i+1 ) − F1 (x̄i )] − σ · x̄i · Ii,0
Ii,2 = σ 2 [F2 (x̄i+1 ) − F2 (x̄i )] − 2σx̄i Ii,1 − σ 2 x̄i2 Ii,0
Ii,3 = σ 3 [F3 (x̄i+1 ) − F3 (x̄i )] − 3σx̄i Ii,2 − 3σ 2 x̄i2 Ii,1 − σ 3 x̄i3 Ii,0

and anti-derivative functions Fk (x ) as before.

p. 346
Integrating a cubic spline versus a normal density function
occurs in various contexts of pricing methods

◮ Method already yields good accuracy for smaller number of grid points.
◮ For larger number of grid points accuracy benefit compared to e.g.
Simpson integration seems not too much.

◮ Either way, use special treatment of break-even point x ⋆ .

◮ Computational effort can become significant for larger number of grid
points.
◮ Bermudan pricing requires N 2 evaluations of Φ(·) and Φ′ (·) per
exercise.

p. 347
Outline

Bermudan Swaptions

Pricing Methods for Bermudans

Density Integration Methods

PDE and Finite Differences

American Monte Carlo

p. 348
PDE methods for finance and pricing are extensively
studied in the literature

◮ We present the basic principles and some aspects relevant for

Bermudan bond option pricing.
◮ Further reading:
◮ L. Andersen and V. Piterbarg. Interest rate modelling, volume I
to III.
Atlantic Financial Press, 2010, Sec. 2.
◮ D. Duffy. Finite Difference Methods in Financial Engineering.
Wiley Finance, 2006

p. 349
Outline

PDE and Finite Differences

Derivative Pricing PDE in Hull-White Model
State Space Discretisation via Finite Differences
Time-integration via θ-Method
Alternative Boundary Conditions for Bond Option Payoffs
Summary of PDE Pricing Method

p. 350
We can adapt the Black-Scholes equation to our
Hull-White model setting
◮ Recall that state variable x (t) follows the risk-neutral dynamics
dx (t) = [y (t) − a · x (t)] dt + σ(t) · dW (t).
| {z }
µ(t,x (t))

◮ Consider an option with price V = V (t, x (t)), option expiry time T and
payoff V (T , x (T )) = g (x (T )).

◮ Derivative pricing formula yields that discounted option price is a

martingale, i.e.
 
V (t, x (t))
d = σV (t, x (t)) · dW (t).
B(t)

How can we use this to derive a PDE?

p. 351
Apply Ito’s Lemma to the discounted option price
We get    
V (t, x (t)) dV (t, x (t)) 1
d = + V (t)d .
B(t) B(t) B(t)

With d B(t)−1 = −r (t) · B(t)−1 · dt follows
 
V (t, x (t)) 1
d = [dV (t, x (t)) − r (t) · V (t) · dt] .
B(t) B(t)

Applying Ito’s Lemma to option price V (t, x (t)) gives

1
dV (t, x (t)) = Vt · dt + Vx · dx (t) + Vxx · [dx (t)]2
2
h i
1
= Vt + Vx · µ (t, x (t)) + Vxx · σ(t)2 dt + Vx · σ(t) · dW (t)
2
with partial derivatives Vt = ∂V (t, x (t)) /∂t, Vx = ∂V (t, x (t)) /∂x and
Vxx = ∂ 2 V (t, x (t)) /∂x 2 .

p. 352
Combining results yields dynamics of discounted option
price

  h i
V (t, x (t)) 1 1
d = Vt + Vx · µ (t, x (t)) + Vxx · σ(t)2 − r (t) · V dt
B(t) B(t) 2
| {z }
µV (t,x (t))

Vx · σ(t)
+ ·dW (t).
B(t)
| {z }
σV (t,x (t))

V (t,x (t))
Martingale property of B(t)
requires that drift vanishes. That is

1
µV (t, x (t)) = Vt + Vx · µ (t, x (t)) + Vxx · σ(t)2 − r (t) · V = 0.
2

Substituting µ (t, x (t)) = y (t) − a · x (t) and r (t) = f (0, t) + x (t) yields pricing
PDE.

p. 353
We get the parabolic pricing PDE with terminal condition
Theorem (Derivative pricing PDE in Hull-White model)
Consider our Hull-White model setup and a derivative security with price
process V (t, x (t)) that pays at time T the payoff V (T , x (T )) = g (x (T )).
Further assume V (T , x (T )) has finite variance and is attainable.
Then for t < T the option price
 
Q V (T , x (T ))
V (t, x (t)) = B(t) · E | Ft
B(T )

follows the PDE

σ(t)2
Vt (t, x ) + [y (t) − a · x ] · Vx (t, x ) + · Vxx (t, x ) = [x + f (0, t)] · V (t, x )
2
with terminal condition
V (T , x ) = g(x ).

Proof.
Follows from derivation above.

p. 354
How does this help for our Bermudan option pricing
problem?

◮ We need option prices on a grid of state variables [x0 , . . . xN ]

Solve Hull-White option pricing PDE backwards from exercise to exercise.

p. 355
Pricing PDE is typically solved via finite difference scheme
and time integration

◮ Use method of lines (MOL) to solve parabolic PDE:

◮ First discretise state space.
◮ Then integrate resulting system of ODEs with terminal condition in
time-direction.

◮ We discuss basic (or general purpose) approach to solve PDE; for a

detailed treatment see Andersen/Piterbarg (2010) or Duffy (2006).

◮ Some aspects may require special attention in the context of Hull-White

model:
◮ more problem-specific boundary discretisation,
◮ non-equidistant grids with finer grid around break-even state x ⋆ ,
◮ upwinding schemes to deal with potentially dominant impact of
convection term [y (t) − a · x ] · Vx (t, x ) at the grid boundaries of x .

p. 356
Outline

PDE and Finite Differences

Derivative Pricing PDE in Hull-White Model
State Space Discretisation via Finite Differences
Time-integration via θ-Method
Alternative Boundary Conditions for Bond Option Payoffs
Summary of PDE Pricing Method

p. 357
How do we discretise state space?

◮ PDE for V (t, x (t)) is defined on infinite domain (−∞, +∞).

◮ We don’t get explicit boundary conditions from PDE derivation.
◮ Thus, we require payoff-specific approximation.
◮ Finally, we are interested in V (0, 0).

◮ We use equidistant x -grid x0 , . . . , xN with grid size hx centered around

zero and scaled via standard deviation of x (T ) at final maturity T ,

x0 = −λ · σ̂ and xN = λ · σ̂

with σ̂ 2 = Var [x (T )] = y (T ) and λ ≈ 5.

◮ Why not shift the grid by expectation E [x (T )] (as suggested in the

literature)?
◮ Pricing PDE is independent of pricing measure (used for derivation).
◮ There is no natural measure under which E [x (T )] should be
calculated.
◮ In T -forward measure ET [x (T )] = 0 anyway.

p. 358
Differential operators in state-dimension are discretised via
central finite differences
For now leave time t continuous. We use notation V (·, x ).
For inner grid points xi with i = 1, . . . , N − 1

V (·, xi+1 ) − V (·, xi−1 )

Vx (·, xi ) = + O(hx2 ) and
2hx

V (·, xi+1 ) − 2V (·, xi ) + V (·, xi−1 )

Vxx (·, xi ) = + O(hx2 ).
hx2

At the boundaries we impose condition

Vxx (·, x0 ) = λ0 · Vx (·, x0 ) and Vxx (·, xN ) = λN · Vx (·, xN ).

This yields one-sided first order partial derivative approximations

2 [V (·, x1 ) − V (·, x0 )] 2 [V (·, xN ) − V (·, xN−1 )]

Vx (·, x0 ) ≈ and Vx (·, xN ) ≈ .
(2 + λ0 hx ) hx (2 − λN hx ) hx

p. 359
Some initial comments regarding choice of λ0,N

◮ Often, λ0,N = 0 (also suggested in the literature).

◮ With λ0,N = 0 we have Vxx (·, x0 ) = Vxx (·, xN ) = 0 and

V (·, x1 ) − V (·, x0 )
Vx (·, x0 ) = + O(hx2 ) and
hx
V (·, xN ) − V (·, xN−1 )
Vx (·, xN ) = + O(hx2 ).
hx

◮ However, for bond options the choice Vxx (·, x0 ) = Vxx (·, xN ) = 0 might be
a poor approximation.

◮ We will discuss an alternative choice for λ0,N later.

p. 360
Now consider PDE for each grid point individually
Define the vector-valued function v (t) via

v (t) = [v0 (t), . . . , vN (t)]⊤ = [V (t, x0 ), . . . , V (t, x0 )]⊤ ∈ RN+1 .

Then state discretisation yields for inner points xi with i = 1, . . . , N − 1,

vi+1 (t) − vi−1 (t) σ(t)2 vi+1 (t) − 2vi (t) + vi−1 (t)
vi′ (t)+[y (t) − axi ] + = [xi + f (0, t)] vi (t)
2hx 2 hx2

and for the boundaries

 
σ(t)2 2 [v1 (t) − v0 (t)]
v0′ (t) + y (t) − ax0 + λ0 = [x0 + f (0, t)] v0 (t),
2 (2 + λ0 hx ) hx
 
σ(t)2 2 [vN (t) − vN−1 (t)]
vN′ (t) + y (t) − axN + λN = [xN + f (0, t)] vN (t).
2 (2 − λN hx ) hx

As before, we have the terminal condition

vi (T ) = g(xi ).

Parabolic PDE is transformed into linear system of ODEs with terminal condition.
p. 361
It is more convenient to write system of ODEs in
matrix-vector notation
We get  
c0 u0
 .. .. 
 l1 . .  · v (t)
v ′ (t) = M(t) · v (t) =  
 .. .. 
. . uN−1
lN cN

with time-dependent components ci , li , ui (i = 1, . . . N − 1),

σ(t)2 σ(t)2 y (t) − axi σ(t)2 y (t) − axi

ci = + xi + f (0, t), li = − + , ui = − −
hx2 2hx2 2hx 2hx2 2hx

and
h i h i
σ(t)2 σ(t)2
2 y (t) − ax0 + λ0 2
2 y (t) − axN + λN 2
c0 = +x0 +f (0, t), cN = − +x0 +f (0, t),
(2 + λ0 hx ) hx (2 − λN hx ) hx
h i h i
σ(t)2 σ(t)2
2 y (t) − ax0 + λ0 2
2 y (t) − axN + λN 2
u0 = − , lN = .
(2 + λ0 hx ) hx (2 − λN hx ) hx

p. 362
Outline

PDE and Finite Differences

Derivative Pricing PDE in Hull-White Model
State Space Discretisation via Finite Differences
Time-integration via θ-Method
Alternative Boundary Conditions for Bond Option Payoffs
Summary of PDE Pricing Method

p. 363
Linear system of ODEs can be solved by standard methods

We have
v ′ (t) = f (t, v (t)) = M(t) · v (t).
We demonstrate time discretisation based on θ-method. Consider equidistant
time grid t = t0 , . . . , tM = T with step size ht and approximation
v (tj+1 ) − v (tj )
≈ f (tj+1 − θht , (1 − θ)v (tj+1 ) + θv (tj ))
ht
for θ ∈ [0, 1].
◮ In general approximation yields method of order O(ht ).
◮ For θ = 21 approximation yields method of order O(ht2 ).
For our linear ODE we set v j = v (tj ), Mθ = M(tj+1 − θht ) and get the scheme

v j+1 − v j  
= Mθ (1 − θ)v j+1 + θv j .
ht

p. 364
We get a recursion for the θ-method
Rearranging terms yields

[I + ht θMθ ] v j = [I − ht (1 − θ) Mθ ] v j+1 .

If [I + ht θMθ ] is regular then we can solve for v j via

v j = [I + ht θMθ ]−1 [I − ht (1 − θ) Mθ ] v j+1 .

Terminal condition is
v M = [g(x0 ), . . . , g(xN )]⊤ .

◮ Unless θ = 0 (Explicit Euler scheme) we need to solve a linear equation

system.
◮ Fortunately, matrix [I + ht θMθ ] is tri-diagonal; solution requires O(M)
operations.
◮ θ-method is A-stable for θ ≥ 12 .
◮ However, oscillations in solution may occur unless θ = 1 (Implicit Euler
scheme, L-stable).

p. 365
Outline

PDE and Finite Differences

Derivative Pricing PDE in Hull-White Model
State Space Discretisation via Finite Differences
Time-integration via θ-Method
Alternative Boundary Conditions for Bond Option Payoffs
Summary of PDE Pricing Method

p. 366
Let’s have another look at the boundary condition ...
We look at an example of a zero coupon bond option with payoff
 +
V (x , T ) = P(x , T , T ′ ) − K .

For x ≪ 0 option is far in-the-money and V (x , t) can be approximated by

intrinsic value
 +
 + P(0, T ′ ) −G(t,T )x − 21 G(t,T )2 y (t)
V (x , t) ≈ Ṽ (x , t) = P(x , t, T ′ ) − K = e −K .
P(0, t)

This yields
∂  
Ṽ (x , t) = −G(t, T ) Ṽ (x , t) + K
∂x
and
∂2 ∂
Ṽ (x , t) = −G(t, T ) Ṽ (x , t).
∂x 2 | {z } ∂x
λ

Alternatively, for x ≫ 0 option is far out-of-the-money and

∂2 ∂
Ṽ (x , t) = Ṽ (x , t) = 0.
∂x 2 ∂x

p. 367
We adapt that approximation to our general option pricing
problem
◮ In principle, for a coupon bond underlying we could estimate λ = λ(t) via
option intrinsic value Ṽ (x , t) and
 
∂2 ∂ ∂
λ(t) = Ṽ (x , t) / Ṽ (x , t) for Ṽ (x , t) 6= 0,
∂x 2 ∂x ∂x

otherwise λ(t) = 0.

◮ We take a more rough approach by approximating λ based only on

previous solution
   
∂2 ∂ ∂2 ∂
λ0,N = V (x , t) / V (x , t) ≈ V (x1,N−1 , t + ht ) / V (x1,N−1 , t + ht )
∂x 2 ∂x ∂x 2 ∂x
j+1 j+1 j+1 j+1 j+1
v0,N−2 − 2v1,N−1 + v2,N v2,N − v0,N−2
≈ /
hx2 2hx

j+1 j+1
for v2,N − v0,N−2 /(2hx ) 6= 0, otherwise λ0,N = 0.

p. 368
It turns out that accuracy of one-sided first order
derivative approximation is of order O(hx2 ) I
Lemma
Assume V = V (x ) is twice continuously differentiable. Moreover, consider grid points
x−1 , x0 , x1 with equal spacing hx = x1 − x0 = x0 − x−1 . If there is a λ0 ∈ R such that

V ′′ (x0 ) = λ0 · V ′ (x0 )

then
2 [V (x1 ) − V (x0 )]
V ′ (x0 ) = + O(hx2 ).
(2 + λ0 hx ) hx

Proof:
Denote vi = V (xi ). We have from standard Taylor approximation

v−1 − 2v0 + v1 v1 − v−1

V ′′ (x0 ) = + O(hx2 ) and V ′ (x0 ) = + O(hx2 ).
hx2 2hx

From V ′′ (x0 ) = λ · V ′ (x0 ) follows

h i
v−1 − 2v0 + v1 v1 − v−1
+ O(hx2 ) = λ0 + O(hx2 ) .
hx2 2hx
p. 369
It turns out that accuracy of one-sided first order
derivative approximation is of order O(hx2 ) II
Multiplying with 2hx2 gives the relation

2 (v−1 − 2v0 + v1 ) + O(hx4 ) = λ0 hx (v1 − v−1 ) + O(hx4 ).

Reordering terms yields

(2 + λ0 hx ) v−1 = 4v0 + (λ0 hx − 2) v1 + O(hx4 ).

And solving for v−1 gives v−1 = [4v0 + (λ0 hx − 2) v1 ] / (2 + λ0 hx ) + O(hx4 ).

Now, we substitute v−1 in the approximation for V ′ (x ). This gives
 
v1 − [4v0 + (λ0 hx − 2) v1 ] / (2 + λ0 hx ) + O(hx4 )
′
V (x0 ) = + O(hx2 )
2hx
(2 + λ0 hx ) v1 − [4v0 + (λ0 hx − 2) v1 ]
= + O(hx2 ) + O(hx3 )
2 (2 + λ0 hx ) hx
2v1 − 4v0 + 2v1
= + O(hx2 )
2 (2 + λ0 hx ) hx
2 (v1 − v0 )
= + O(hx2 ).
(2 + λ0 hx ) hx

p. 370
It turns out that accuracy of one-sided first order
derivative approximation is of order O(hx2 ) III

◮ With constraint V ′′ (x0 ) = λ · V ′ (x0 ) we can eliminate explicit dependence

on second derivative V ′′ (x0 ) and outer grid point v−1 = V (x−1 ).

◮ Analogous result can be derived for upper boundery and down-ward

approximation of first derivative.

◮ Resulting scheme is still second order accurate in state space direction.

p. 371
Outline

PDE and Finite Differences

Derivative Pricing PDE in Hull-White Model
State Space Discretisation via Finite Differences
Time-integration via θ-Method
Alternative Boundary Conditions for Bond Option Payoffs
Summary of PDE Pricing Method

p. 372
We summarise the PDE pricing method

1. Discretise state space x on a grid [x0 , . . . , xN ] and specify time step size
ht and θ ∈ [0, 1].

2. Determine the terminal condition v j+1 = max {Uj+1 , Hj+1 } for the current
valuation step.

3. Set up discretised linear operator Mθ of the resulting ODE system

d
dt
v = Mθ · v .

4. Incorporate appropriate product-specific boundary conditons.

5. Set up linear system [I + ht θMθ ] v j = [I − ht (1 − θ) Mθ ] v j+1 .

6. Solve linear system for v j by tri-diagonal matrix solver.

7. Repeat with step 3. until next exercise date or tj = 0.

p. 373
Outline

Bermudan Swaptions

Pricing Methods for Bermudans

Density Integration Methods

PDE and Finite Differences

American Monte Carlo

p. 374
Monte Carlo methods are widely applied in various finance
applications

◮ We demonstrate the basic principles for

◮ path integration of Ito processes
◮ exact simulation of Hull-White model paths
◮ There are many aspects that should also be considered, see
e.g.
◮ L. Andersen and V. Piterbarg. Interest rate modelling, volume I
to III.
Atlantic Financial Press, 2010, Sec. 3.
◮ P. Glasserman. Monte Carlo Methods in Financial Engineering.
Springer, 2003

p. 375
Outline

American Monte Carlo

Introduction to Monte Carlo Pricing
Monte Carlo Simulation in Hull-White Model
Regression-based Backward Induction

p. 376
Monte Carlo (MC) pricing is based on the Strong Law of
Large Numbers

Theorem (Strong Law of Large Numbers)

Let Y1 , Y2 , . . . be a sequence of independent identically distributed (i.i.d.)
random P variables with finite expectation µ < ∞. Then the sample mean
n
Ȳn = n1 i=1 Yi converges to µ a.s. That is

lim Ȳn = µ a.s.

n→∞

◮ We aim at calculating V (t) = N(t) · EN [V (T )/N(T ) | Ft ].

n o
(T ;ωi )
◮ For MC pricing simulate future discounted payoffs VN(T ;ω )
.
i
i=1,2,...n
◮ And estimate
n
1 X V (T ; ωi )
V (t) = N(t) · .
n N(T ; ωi )
i=1

p. 377
Keep in mind that sample mean is still a random variable
governed by central limit theorem
Theorem (Central Limit Theorem)
Let Y1 , Y2 , . . . be a sequence of i.i.d. random variables with finite expectation
µ<∞P and standard deviation σ < ∞. Denote the sample mean
n
Ȳn = n1 i=1 Yi . Then
Ȳn − µ d
√ −→ N(0, 1).
σ/ n
Pn
2
Moreover, for the variance estimator sn2 = 1
n−1 i=1
Yi − Ȳn we also have

Ȳn − µ d
√ −→ N(0, 1).
sn / n

◮ Here, N(0, 1) is the standard normal distribution.

d
◮ −→ denotes convergence in distribution, i.e. limn→∞ Fn (x ) = F (x ) for
the corresponding cumulative distribution functions and all x ∈ R at
which F (x ) is continuous.
√
◮ sn / n is the standard error of the sample mean Ȳn .
p. 378
How do we get our samples V (T ; ωi )/N(T ; ωi )?
1. Simulate state variables x (t) on relevant dates t:

2. Simulate numeraire N(t) on relevant dates t:

3. Calculate payoff V (T , x (T )) at observation/pay date T .

p. 379
We need to simulate our state variables on the relevant
observation dates
Consider the general dynamics for a process given as SDE

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

◮ Typically, we know initial value X (t) (t = 0).

◮ We need X (T ) for some future time T > t.

◮ In Hull-White model and risk-neutral measure formulation we have

µ(t, X (t)) = y (t) − a · X (t), and, σ(t, X (t)) = σ(t).

There are several standard methods to solve above SDE. We will briefly discuss
Euler method and Milstein method.

p. 380
Euler method for SDEs is similar to Explicit Euler method
for ODEs
◮ Specify a grid of simulation times t = t0 , t1 , . . . , tM = T .

◮ Calculate sequence of state variables

Xk+1 = Xk + µ(tk , Xk ) (tk+1 − tk ) + σ(tk , Xk ) [W (tk+1 ) − W (tk )] .

◮ Drift µ(tk , Xk ) and volatility σ(tk , Xk ) are evaluated at current time tk

and state Xk .

◮ Increment of Brownian motion W (tk+1 ) − W (tk ) is normally distributed,

i.e.
√
W (tk+1 ) − W (tk ) = Zk · tk+1 − tk with Zk ∼ N(0, 1).

p. 381
Milstein method refines the simulation of the diffusion term
◮ Again, specify a grid of simulation times t = t0 , t1 , . . . , tM = T .
◮ Calculate sequence of state variables

Xk+1 = Xk + µ(tk , Xk ) (tk+1 − tk ) + σ(tk , Xk ) [W (tk+1 ) − W (tk )]

1 ∂  
+ · σ(tk , Xk ) · σ(tk , Xk ) · (W (tk+1 ) − W (tk ))2 − (tk+1 − tk ) .
2 ∂x

◮ Drift µ(tk , Xk ) and volatility σ(tk , Xk ) are evaluated at current time tk and
state Xk .
◮ ∂
Requires calculation of derivative of volatility ∂x
σ(tk , Xk ) w.r.t. state variable.
◮ Increment of Brownian motion W (tk+1 ) − W (tk ) is normally distributed, i.e.
p
W (tk+1 ) − W (tk ) = Zk · tk+1 − tk with Zk ∼ N(0, 1).

◮ With ∆k = tk+1 − tk iteration becomes

p 1 ∂σ(tk , Xk )

Xk+1 = Xk + µ(tk , Xk )∆k + σ(tk , Xk )Zk ∆k + σ(tk , Xk ) Zk2 − 1 ∆k .
2 ∂x

p. 382
How can we measure convergence of the methods?
◮ We distinguish strong order of convergence and weak order of
convergence.
 M
◮ Consider a discrete SDE solution Xkh with Xkh ≈ X (t + kh),
k=0
T −t
h= M
.

Definition (Strong order of convergence)

h
The discrete solution XM at final maturity T = t + hM converges to the exact
solution X (T ) with strong orderβ if there exists a constant C such that
 
E h
XM − X (T ) ≤ C · hβ .

◮ Strong order of convergence focuses on convergence on the individual

paths.

◮ Euler method has strong order of convergence of 1

2
(given sufficient
conditions on µ(·) and σ(·)).

◮ Milstein method has strong order of convergence of 1 (given sufficient

conditions on µ(·) and σ(·)).
p. 383
For derivative pricing we are typically interested in weak
order of convergence
We need some context for weak order of convergence
◮ A function f : R → R is polynomially bounded if |f (x )| ≤ k (1 + |x |)q for
constants k and q and all x .
◮ The set CPn represents all functions that are n-times continuously
differentiable and with 1st to nth derivative polynomially bounded.

Definition (Weak order of convergence)

h
The discrete solution XM at final maturity T = t + hM converges to the exact
solution X (T ) with weak orderβ if there exists a constant C such that


h
E f XM − E [f (X (T ))] ≤ C · hβ 2β+2
∀f ∈ CP

for sufficiently small h.

◮ Think of f as a payoff function, then weak order of convergence is related
to convergence in price.
◮ Euler method and Milstein method can be shown to have weak order 1
convergence (given sufficient conditions on µ and σ).

p. 384
Some comments regarding weak order of convergence

Error estimate 

h
E f XM − E [f (X (T ))] ≤ C · hβ
requires considerable assumptions regarding smoothness of µ(·), σ(·) and test
functions f (·).

◮ In practice payoffs are typically non-smooth at the strike.

◮ This limits applicability of more advanced schemes with theoretical higher
order of convergence.

◮ A fairly simple approach of a higher order scheme is based on Richardson

extrapolation:
◮ this method is also applied to ODEs,
◮ see Glassermann (2000), Sec. 6.2.4 for details.

◮ Typically, numerical testing is required to assess convergence in practice.

p. 385
The choice of pricing measure is crucial for numeraire
simulation
Consider risk-neutral measure, then
Z T  Z T 
N(T ) = B(T ) = exp r (s)ds = exp [f (0, s) + x (s)] ds
0 0
Z T 
= P(0, T )−1 exp x (s)ds .
0
RT
Requires simulation or approximation of 0 x (s)ds.
Suppose x (tk ) is simulated on a time grid {tk }M
k=0 then we approximate integral
via Trapezoidal rule
Z T M
X x (tk−1 ) + x (tk )
x (s)ds ≈ (tk − tk−1 ) .
0
2
i=1

Numeraire simulation is done in parallel to state simulation

 
P(0, tk−1 ) x (tk−1 ) + x (tk )
N(tk ) = · N(tk−1 ) · exp (tk − tk−1 ) .
P(0, tk ) 2

p. 386
Alternatively, we can simulate in T -forward measure for a
fixed future time T

Select a future time T̄ sufficiently large. Then N(0) = P(0, T̄ ).

At any pay time T ≤ T̄ numeraire is directly available via zero coupon bond
formula
P(0, T̄ ) −G(T ,T ′ )x (T )− 21 G(T ,T ′ )2 y (T )
N(T ) = P(x (T ), T , T̄ ) = e .
P(0, T )

However, T̄ -forward measure simulation needs consistent model formulation or

change of measure.
In particular

dW T̄ (t) = σP (t, T̄ ) ·dt + dW (t) .

| {z } | {z } | {z }
B.M. in T̄ -forward measure ZCB volatility B.M. in risk-neutral measure

p. 387
Another commonly used numeraire for simulation is the
discretely compounded bank account
◮ Consider a grid of simulation times t = t0 , t1 , . . . , tM = T .
◮ Assume we start with 1 EUR at t = 0, i.e. N(0) = 1.
◮ At each tk we take numeraire N(tk ) and buy zero coupon bond maturing
at tk+1 . That is
N(tk )
N(t) = P(t, tk+1 ) · for t ∈ [tk , tk+1 ] .
P(tk , tk+1 )

Explicitly, define discretely compounded bank account as B̄(0) = 1 and

Y P(t, tk+1 )
B̄(t) = .
P(tk , tk+1 )
tk <t

We get
   
B̄(t) Y 1 P(t, tk+1 )
d = ·d =0 for t ∈ [tk , tk+1 ] .
P(t, tk+1 ) P(tk , tk+1 ) P(t, tk+1 )
tk <t

Simulating in B̄-measure is equivalent to simulating in rolling tk+1 -forward

measure.
p. 388
Outline

American Monte Carlo

Introduction to Monte Carlo Pricing
Monte Carlo Simulation in Hull-White Model
Regression-based Backward Induction

p. 389
Do we really need to solve the Hull-White SDE
numerically?
Recall dynamics in T -forward measure
 
dx (t) = y (t) − σ(t)2 G(t, T ) − a · x (t) · dt + σ(t) · dW T (t).

That gives
 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

As a result x (T ) ∼ N(µ, σ 2 ) (conditional on t) with

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

2 ′ 2
σ = Var [x (T ) | Ft ] = y (T ) − G (t, T ) y (t).
We can simulate exactly

x (T ) = µ + σ · Z with Z ∼ N(0, 1).

p. 390
Expectation calculation via µ = ET [x (T ) | Ft ] requires
carefull choice of numeraire
Consider grid of simulation times t = t0 , t1 , . . . , tM = T .
We simulate
x (tk+1 ) = µk + σk · Zk
with

µk = G ′ (tk , tk+1 ) [x (tk ) + G(tk , tk+1 )y (tk )] ,

σk2 = y (tk+1 ) − G ′ (tk , tk+1 )2 y (tk ), and
Zk ∼ N(0, 1).

Grid point tk+1 must coincide with forward measure for Etk+1 [·] for each
individual step k → k + 1.
Numeraire must be discretely compounded bank account B̄(t) and

B̄(tk )
B̄(tk+1 ) = .
P(x (tk ), tk , tk+1 )

Recursion for x (tk+1 ) and B̄(tk+1 ) fully specifies path simulation for pricing.
p. 391
Some comments regarding Hull-White MC simulation ...

◮ We could also simulate in risk-neutral measure or T̄ -forward measure.

◮ This might be advantegous if also FX or equities are
modelled/simulated.
◮ Requires adjustment of conditional expectation µk and numeraire
N(tk ) calculation.
◮ Variance σk2 is invariant to change of meassure in Hull-White model.

◮ Repeat path generation for as many paths 1, . . . , n as desired (or

computationally feasible).

◮ For Bermudan pricing we need to simulate x and N (at least) at exercise

dates TE1 , . . . , TEk̄ .

◮ For calculation of Zk use

◮ pseudo-random numbers or
◮ Quasi-Monte Carlo sequences.
as proxies for independent N(0, 1) random variables accross time steps
and paths.

p. 392
We illustrate MC pricing by means of a coupon bond
option example
Consider coupon bond option expiring at TE with coupons Ci paid at Ti
(i = 1, . . . , u, incl. strike and notional).
◮ Set t0 = 0, t1 = TE /2 and t2 = TE (two steps for illustrative purpose).
◮ Compute 2n independent N(0, 1) pseudo random numbers Z 1 , . . . , Z 2n .
◮ For all paths j = 1, . . . , n calculate:
◮ µj0 , σ0 and B̄ j (t1 ); note µj0 and B̄ j (t1 ) are equal for all paths j since
x (t0 ) = 0,
◮ x1j = µj0 + σ0 · Z j ,
◮ µj1 , σ1 and B̄ j (t2 ); note now µj1 and B̄ j (t2 ) depend on x1j ,
◮ x2j = µj1 + σ1 · Z n+j ,
Pu +
◮ payoff V j (t2 ) = C · P(x2j , t2 , Ti ) at t2 = TE .
i=1 i
◮ Calculate option price (note B̄(0) = 1)
n
1 X V j (t2 )
V (0) = B̄(0) · .
n B̄ j (t2 )
j=1

p. 393
Outline

American Monte Carlo

Introduction to Monte Carlo Pricing
Monte Carlo Simulation in Hull-White Model
Regression-based Backward Induction

p. 394
Let’s return to our Bermudan option pricing problem
H1 = . . . H2 = . . . H3 = 0 K
✻ ✻ ✻
continuation value
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m
h i
H0 = B(t)E V1
B(T 1 )
| Ft ❄ ❄ ❄ ❄ ❄ ❄
E
Lm
V3 = max{U3 , H3 }
V2 = max{U2 , H2 }
V1 = max{U1 , H1 }
U1 U2 U3 K
✻ ✻ ✻
exercise payoff
1
Tl−1 2
Tl−1 3
Tl−1 Tn
✲
TE1 1
T̃k−1 TE2 2
T̃k−1 TE3 3
T̃k−1 T̃m

❄ ❄ ❄ ❄ ❄ ❄
Lm

p. 395
In this setting we need to calculate future conditional
expectations
◮ Assume we already simulated paths for state variables xk , underlyings Uk
and numeraire Bk for all relevant dates tk .
◮ We need continuation values Hk defined recursively via Hk̄ = 0 and
 
max {Uk+1 , Hk+1 }
Hk = Bk Ek .
Bk+1

◮ In principle, we could use nested Monte Carlo:

◮ In practice, nested Monte Carlo is typically computationally not feasible.

p. 396
A key idea of American Monte Carlo is approximating
conditional expectation via regression

Conditional expectation
 
Bk
Hk = Ek max {Uk+1 , Hk+1 }
Bk+1

is a function of the path x (t) for t ≤ tk .

For non-path-dependent underlyings Uk , Hk can be written as function of
xk = x (tk ), i.e.
Hk = Hk (xk ).

We aim at finding a regression operator

Rk = Rk [Y ]

which we can use as proxy for Hk .

p. 397
What do we mean by regression operator?
Denote ζ(ω) = [ζ1 (ω), . . . , ζq (ω)]⊤ a set of basis functions (vector of random
variables).

Let Y = Y (ω) be a target random variable.

Assume we have outcomes ω1 , . . . , ωn̄ with control variables ζ(ω1 ), . . . , ζ(ωn̄ )

and observations Y (ω1 ), . . . , Y (ωn̄ ).

A regression operator R [Y ]is defined via

R [Y ] (ω) = ζ(ω)⊤ β

where the regression coefficients β solve linear least squares problem

  2
ζ(ω1 )⊤ β − Y (ω1 )
 .. 
 .  → min .
⊤
ζ(ωn̄ ) β − Y (ωn̄ )

Linear least squares system can be solved e.g. via QR factorisation or SVD.

p. 398
A basic pricing scheme is obtained by replacing conditional
expectation of future payoff by regression operator
Approximate H̃k ≈ Hk via H̃k̄ = Hk̄ = 0 and
 
Bk 
H̃k = Rk max Uk+1 , H̃k+1 for k = k̄ − 1, . . . , 1.
Bk+1

◮ Critical piece of this methodology is (for each step k)

◮ choice of regression variables ζ1 , . . . , ζq and
◮ calibration of regression operator Rk with coefficients β.
◮ Regression variables ζ1 , . . . , ζq must be calculated based on information
up to tk .
◮ They must not look into the future to avoid upward bias.
◮ Control variables ζ(ω1 ), . . . , ζ(ωn̄ ) and observations Y (ω1 ), . . . , Y (ωn̄ ) for
calibration should be simulated on paths independent from pricing.
◮ Using same paths for calibration and payoff simulation also
incorporates information on the future.

p. 399
What are typical basis functions?

State variable approach

Set ζi = x (tk )i−1 for i = 1, . . . , q. Typical choice is q ≈ 4 (i.e. polynomials of
Qd
order 3). For multi-dimensional models we would set ζi = j=1 xj (tk )pi,j with
Pd
j=1
pi,j ≤ r .
◮ Very generic and easy to incorporate.

Explanatory variable approach

Identify variables y1 , . . . yd̄ relevant for the underlying option. Set basis
functions as monomials
d̄
Y d̄
X
ζi = yj (tk )pi,j with pi,j ≤ r .
j=1 j=1

◮ Can be chosen option-specific and incorporate information prior to tk .

◮ Typical choices are co-terminal swap rates or Libor rates (observed at tk ).

p. 400
Regression of the full underlying can be a bit rough - we
may restrict regression to exercise decision only
For a given path consider
Bk
Hk = max {Uk+1 , Hk+1 }
Bk+1
h   i
Bk
= ✶{Uk+1 >Hk+1 } Uk+1 + 1 − ✶{Uk+1 >Hk+1 } Hk+1 .
Bk+1

Use regression to calculate ✶{Uk+1 >Hk+1 } .

Calculate Rk = Rk [Uk+1 − Hk+1 ], set Hk̄ = 0 and

Bk 

Hk = ✶{Rk >0} Uk+1 + 1 − ✶{Rk >0} Hk+1 for k = k̄ − 1, . . . , 1.
Bk+1

◮ Think of ✶{Rk >0} as an exercise strategy (which might be sub-optimal).

◮ This approach is sometimes considered more accurate than regression on
regression.
◮ For further reference, see also Longstaff/Schwartz (2001).
p. 401
We summarise the American Monte Carlo method
1. Simulate n paths of state variables xkj , underlyings Ukj and numeraires Bkj
(j = 1, . . . , n) for all relevant times tk (k = 1, . . . k̄).
2. Set H j = 0.
k̄
3. For k = k̄ − 1, . . . 1 iterate:
 j=1,...,n̂
3.1 Calculate control variables ζij = ζi (ωj ) i=1,...,q
and regression variables
Y j = Ukj − Hkj for the first n̂ paths (n̂ ≈ 1
4
n).
3.2 Calibrate regression operator Rk = Rk [Y ] which gives coefficients β.
 j=n̂+1,...n
3.3 Calculate control variables ζij = ζi (ωj ) i=1,...,q
for remaining paths
and (for all paths)

Bkj
h   i
Hkj = j
✶{Rk (ωj )>0} Uk+1
j
+ 1 − ✶{R
k (ωj )>0}
j
Hk+1 .
Bk+1

4. Calculate discounted payoffs for the paths j = n̂ + 1, . . . n not used for regression

Bkj 
H0j = j
max U1j , H1j .
Bk+1

Pn
5. Derive average V (0) = 1
n−n̂ j=n̂+1
H0j .
p. 402
Some comments regarding AMC for Bermudans in
Hull-White model

◮ AMC implementations can be very bespoke and problem specific.

◮ See literature for more details.

◮ More explanatory variables or too high polynomial degree for regression

may deteriorate numerical solution.
◮ This is particularly relevant for 1-factor models like Hull-White.
◮ Single state variable or co-terminal swap rate should suffice.

◮ AMC with Hull-White for Bermudans is not the method of choice.

◮ PDE and integration methods are directly applicable.
◮ AMC is much slower and less accurate compared to PDE and
integration.

AMC is the method of choice for high-dimensional models and/or

path-dependent products.

p. 403
Contact

Dr. Sebastian Schlenkrich

Office: RUD25, R 1.211
Mail: sebastian.schlenkrich@hu-berlin.de

d-fine GmbH
Mobile: +49-162-263-1525
Mail: sebastian.schlenkrich@d-fine.de