University Finance Seminar

30 January 2004

Managing Correlation Risk with

Spread Option Models
M A H Dempster
Centre for Financial Research
Judge Institute of Management, University of Cambridge
&
Cambridge Systems Associates Limited
[email protected] http://www-cfr.jims.cam.ac.uk
Co-worker: S S G Hong, UBS, London

! "# " " "$

 Contents

• Introduction
• Spread Option Pricing Review
• Fourier Transform Techniques for Vanilla Options
• Pricing Spread Options with the FFT
• Computational Results
• Market Calibration
• Conclusions and Future Work

! "# " " "$

 1. Introduction
What is a Spread Option ?

• Two Underlying Assets: S1, S2

• Spread (basis): S1 - S2

• Payoff: (S1(T) - S2(T) - K)+

• Price: VT(K) = EQ[ e-rT (S1(T) - S2(T) - K)+ ]

! "# " " "$

 Why are they important?
• Invaluable tools for hedging and speculating...

• ... in almost all markets!

Energy Crack spread, Spark spread
Commodity Crush spread, Cotton calendar spread
Equity Index spread
Bond NOB spread, TED spread
Credit Derivatives Credit spread

• Indispensable for managing “correlation risks”

! "# " " "$

 Hedging Using Spread Options
An oil refinery firm can short a call on the spread of oil
future prices: Fl – Fs

• Fl : long output = Refined product

• Fs : short input = Brent crude
• K : strike = marginal conversion cost
• ( Fl(T) - Fs(T) - K)+ : payoff of the crack spread

! "# " " "$

 Hedging Using Spread Options
• If the spread is greater than the cost the option is
exercised by the holder and the firm meets its
obligation by producing
• If the spread is less than the cost the option expires
worthless and the firm will not produce
• Either way the firm earns the option premium
i.e. a call on the spread replicates the payoff
structure of a firm’s production schedule
• Also used to bridge delivery locations

! "# " " "$

 Speculating Using Spread
Options
A speculator can trade the correlation between two prices,
indices or bond yields (LTCM):
• If we speculate on a correlation drop, we long a call on
spread
• If we speculate on a correlation rise, we short a call on
spread
The reasoning is similar to going long on a vanilla call on a
single asset if we think volatility will rise, with the
variance of the spread replacing the volatility of the single
asset

! "# " " "$

 Speculating Using Spread
Options
The spread variance depends on:
• volatility of the long leg
• volatility of the short leg
• correlation between the two

The first two can be traded by options on individual prices

We need a spread option to trade the third (Mbanefo 1997)

! "# " " "$

 The Problem
• Set up good models for the dynamics of the factors which
accommodate stochasticities in interest rates, volatility...

• Compute the price of a spread option under such models

• Study how the price depends on the model specification

in particular the volatility and correlation structure

• Design appropriate calibration procedures

! "# " " "$

 2. Spread Option Pricing Review
Existing Approaches: I

• Model the spread as a geometric Brownian motion:

X := S1 - S2
dX = X ( µ dt + σ dW)
• Apply the Black-Scholes formula:
VT(K) = EQ [ e-rT [ S1(T) - S2 (T) -K]+ ]
:= EQ [ e-rT [X(T) -K] + ]
• Simple but dangerous!
- spread can go negative
- a multi-factor problem by nature

! "# " " "$

 Existing Approaches: II
• Model S1, S2 as geometric Brownian motions:
dS1 = S1 ( µ 1 dt + σ1 dW1)
dS2 = S2 ( µ 2 dt + σ2 dW2)

where EQ [dW1 dW2] = ρ dt

• ρ is the correlation between the prices

• Apply a conditioning technique to turn the two-dimensional

integral into a single one
(K Ravindran 1993, D Shimko 1994)

! "# " " "$

 ∞ ∞
− rT
VT ( K ) = e ( S1 − S 2 − K ) fT ( S1 , S 2 )dS1dS 2
0 S2 + K

∞ ∞
− rT
=e [ S1 − ( S 2 + K )]+ f1 2 ( S1 S 2 )dS1 f 2 ( S 2 )dS 2
0 S2 + K

∞
− rT
=e C ( S 2 ) f 2 ( S 2 )dS 2
0

where
• fT (·| ·) : joint p.d.f. of S1(T), S2(T) ... bivariate log-normal
• f1|2 (·| ·) : conditional density of S1(T) given S2(T) ... log-normal
• f2 (·) : marginal density of S2(T) .... log-normal
• C (·) : an integral similar to the Black-Scholes call price

! "# " " "$

 Existing Approaches: II
Simple, two-factor, but...

• Only works when distributions are normal

• Prices are the only sources of randomness...
• No stochastic interest rate or convenience yield
• Constant (deterministic) volatility
• Trivial correlation structure

! "# " " "$

 Existing Approaches: III
Variants on the previous approach
• Approximation by piecewise linear payoff function
(N D Pearson 1995)
• Edgeworth series expansion
(D Pilipovic & J Wengler 1998)
• Lattice and PDE methods (Brooks 1995)
• A GARCH model with co-integration is also proposed and the
spread option is valued using a Monte Carlo method (J C Duan, S
R Pliska 1999)
• Gaussian mixture (C Alexander 2003)
• Survey (R Carmona & V Durrleman 2003)

! "# " " "$

 3. Fourier Transform Techniques
for Vanilla Options
What is a Fourier Transform ?

∞
f ( x) φ (υ ) = f ( x) ⋅ eiυx dx
−∞

1 ∞
φ (υ ) f ( x) = φ (υ ) ⋅ e −iυx dυ
2π −∞
• probability density functions → characteristic functions
• differentiation w.r.t. x → multiplication by -iυ and inverting

1 ∞
f ′( x ) = ( − i υ ⋅ φ (υ )) ⋅ e − iυ x d υ
2π −∞
• option pricing = integration of p.d.f. times payoff

! "# " " "$

 … and a Fast Fourier Transform?
• An efficient algorithm for computing the sum
N −1 2πi
− jk
Yk = X j ⋅e N
for k = 1, ,N
j =0
for a complex array X=(Xj) of size N

2
• Reduces the number of multiplications from an order of N to
N log 2 N Strassen (1967)

• Crucial for approximating the Fourier integral as a function ofυ

∞ N −1
iυx iυ x j
f ( x) ⋅ e dx ≈ f ( x j )e ∆x
−∞
j =0

! "# " " "$

 Black-Scholes via Fourier Transform
• S Heston (1993), G Bakshi & D B Madan(1999), P Carr
& D B Madan (1999)

To price a European call under Black-Scholes we need:

• sT := log(ST) : log-price of the underlying at maturity
• qT(.) : risk-neutral density of the log-price sT
• k := log(K) : log of the strike price
• CT(k) : price of a T -maturity call with strike ek
• fT(.) : characteristic function of the risk-neutral density qT

! "# " " "$

 Characteristic function under
Black-Scholes
1 2
d ln S = (r − σ )dt + σ dW
2
(
sT N s0 + (r − 2 σ )T , σ T
1 2 2
)
φT ( u ) : = EQ eiu⋅s T

∞
iu ⋅s
= e qT ( s )ds
−∞

= exp s0 + (r − σ )T − σ T ⋅ u
1
2
2 1
2
2 2

! "# " " "$

 Fourier transform of the (modified) call
(P Carr & D B Madan 1999)

CT (k ) ≡ EQ e − rT ( ST − K )+

(e − e ) qT ( s)ds
∞
− rT
≡ e s k
k

• The call price is not square-integrable since

CT(k) → S0 , k → -∞
• Define the modified call price for some α >0
cT(k) := exp(αk) CT(k)

! "# " " "$

• The Fourier transform of the modified call price cT is given by
∞
ψ T (υ ) := eiυ k cT (k )dk
−∞
∞ ∞
= e− rT e(α +iυ ) k (e s − e k )qT ( s)dsdk
−∞ k
∞ s
= e − rT
qT ( s ) e(α +iυ ) k (e s − e k )dkds
−∞ −∞
∞
− rT (α +1+ iυ ) s
= e qT ( s ) e
(α + iυ )(α +1+ iυ )
ds
−∞

e − rT φT (υ − (α + 1)i )
=
(α + iυ )(α + 1 + iυ )
• But φT is also known in closed form!

! "# " " "$

• Inverting thus yields the call price:

e −α k ∞
CT (k ) = e− iυ kψ T (υ ) dυ
2π −∞

• Approximate this, using trapezoid or Simpson'

s rule, with a finite sum and
then apply the Fast Fourier Transform
e −α km N −1
ψ T (υ j )η
− iυ j km
CT (km ) ≈ e
2π j =0
2π i
e −α km N −1 − jm
= ( −1) j + m ψ υj η ⋅e N
2π j =0
T
for m=0,…,N-1 , where
2π
υ j = ( j − N 2)η k m = ( m − N 2) λ λ ⋅η =
N
Note: With an N grid for the Fourier sum this gives option prices with N
equally spaced strikes

! "# " " "$

 Extending the payoff
By modifying the input function of the inverse transform
ψT(·) we can handle the following instrument with the
same technique:
• e A⋅sT + B − e k : call on bonds ( sT is now the short rate)
+
• [ ( A ⋅ sT + B) − k ]+ : call on yields
• P(sT) : payoff contingent on polynomial in sT
• H(sT) : can even do general payoff in C ∞ via Taylor
series expansion! (G Bakshi & D B Madan 1999)

! "# " " "$

 Extending the distribution
• Normality can be relaxed...
• Explicit expression of the p.d.f. not needed
• Key: Characteristic functions!
• The underlying can evolve as
- O. U. or C. I. R. processes
- Affine diffusion with jumps
- VG (Variance Gamma) process...
• Many of the above have no analytic density but their characteristic
functions are known
• Needed for spreads on prices of pseudo-commodities such as kwH

! "# " " "$

 Extending numbers of factors
• Stochastic volatility, stochastic interest rate... can be incorporated

(
dS = S r dt + v dW1 )
dv = κ v (µ v − v )dt + v dW2
dr = κ rv (µ rv − r )dt + σ r dW3

... as long as the factors have analytic characteristic functions

• This includes pretty much all the diffusion models in the literature:
• Multifactor CIR models (Chen-Scott...)
• General affine diffusion models (Duffie, Kan, Singleton...)
• Gaussian interest rate models (Longstaff-Schwartz...)
• Stochastic volatility models (Heston, Bates, Hull-White...)

! "# " " "$

 Extending the number of assets
• Now consider options whose payoffs are contingent on two assets
S1, S2
• Example (Bakshi & Madan 1999): a generalisation of European
( ) (
call with the following payoff: e s1 (T ) − e k1 + ⋅ e s2 (T ) − e k 2 + )
We can price it in a similar fashion

cT (k1 , k 2 ) := exp(α1k1 + α 2 k 2 ) ⋅ CT (k1 , k 2 )

≡e α1k1 +α 2 k 2 ∞ ∞

k1 k 2
( )( )
e − rT e s1 (T ) − e k1 ⋅ e s2 (T ) − e k2 qT ( s1 , s2 )ds2 ds1

! "# " " "$

• Consider its Fourier transform
∞ ∞
ψ T (υ1 ,υ 2 ) := eiυ1k1 +iυ 2 k2 cT (k1 , k 2 )dk 2 dk1
−∞ −∞

e − rT φT (υ1 − (α1 + 1)i,υ 2 − (α 2 + 1)i )

=
(α1 + iυ1 )(α1 + 1 + iυ1 )(α 2 + iυ 2 )(α 2 + 1 + iυ 2 )

• Inverting thus yields the option price

e −α1k1 −α 2 k2 ∞ ∞
CT (k1 , k 2 ) = e −iυ1k1 −iυ 2k 2ψ T (υ1 ,υ 2 )dυ 2 dυ1
(2π ) 2
−∞ −∞

• Compute this with a two-dimensional FFT

! "# " " "$

 Moral of the story
Fourier Transform
Probability density Characteristic
function function

Integrating Multiplication by
the payoff a suitable constant

Option price Inverse Transform

Transform of
(modified) Option price

! "# " " "$

 4. Pricing Spread Options with the FFT
• Let us now try to price a call on the spread S1 - S2
[
VT (k ) = EQ e − rT ( S1 − S 2 − K ) + ]
∞ ∞
− rT
=e (e − e − e )qT ( s1 , s2 )ds1ds2
s1 s2 k
−∞ log( e s2 k
+e )

− rT
≡e (e − e − e k )qT ( s1 , s2 )ds1ds2
s1 s2
Ω

• Big problem: the exercise region Ω to be integrated over

has a curved boundary
Ω := { ( s , s ) ∈ R
1 2
2
| e −e −e ≥0}
s1 s2 k

! "# " " "$

The simple 2-D FFT (Bakshi & Madan 1999) trick
will not work here!

! "# " " "$

 Approximating the Exercise Region
• Approximate it with rectangular strips (Riemann) as

VT (k ) = e − rT (e s1 − e s2 − e k )qT ( s1 , s2 )ds1ds2
Ω
N −1
≈ e − rT (e s1 − e s2 − e k )qT ( s1 , s2 )ds1ds2
Ωu
u =0

• The integral can be computed over each rectangular

region Ωu, u=0,…,N-1

! "# " " "$

 Riemann Approximation

Riemann approximation with rectangular strips

! "# " " "$

• We DON'
T have to do N integrals!!!

• A single 2-D transform will produce N×N of

∞ ∞
(e − e − e )qT ( s1 , s2 )ds1ds2
s1 s2 k
k1 ( m ) k 2 ( n )

for m,n = 0, …, N-1

• These are sufficient for the N components we require since

for different strikes k of the spread option we only need to pick
different components to sum and no additional transform is needed

! "# " " "$

 Why the FFT?
• Consider the following model :

d = (r dt + d )
d = (r dt + d )
d = κ 1 ( µ1 − )dt + d
d = κ 2 (µ 2 − )dt + d
with EQ [dWi dWj] = ρij dt
• Direct generalisation of 1-D stochastic volatility models
with non-trivial correlation!

! "# " " "$

• No existing method can handle this!
- conditioning trick won' t work
- lattice obviously fails...
- a PDE in 4 space variables
- slow convergence for Monte Carlo

• But easy (relatively) with the Fourier transform approach!

- as the number of factors go up the payoff structure based on the
price differences remains the same
- the characteristic function involves more parameters and
complicated expressions (naturally) but is still known in
closed form
- the transform will still be two dimensional

! "# " " "$

 5. Computational Results
• Athlon 650 MHz with 512 MB RAM running Linux

• Code in C++

• Invoke Simpson' s rule for approximation of the

Fourier integral

• Use the award winning FFTW code ("Fastest Fourier Transform in

the West") written by M Frigo and S Johnson from MIT (1999)

! "# " " "$

 Pricing Spread Options under
Two-factor GBM
• First we compute spread option prices with the model
(Existing Approach II):
dS1 = S1 ( µ 1 dt + σ1 dW1)
dS2 = S2 ( µ 2 dt + σ2 dW2)
where EQ [dW1 dW2] = ρ dt

• We compare prices to those obtained by direct 1-D

integration (using conditioning)

! "# " " "$

 Errors in Spread Prices

Errors in Spread Prices across Strikes and Maturities for the FFT Method
with High and Low Volatility N=4096

! "# " " "$

 Error Variation With Strikes
Strike Analytic FFT Error (b.p.)
0 6.56469 6.564078 0.932488
0.1 6.52267 6.522448 0.341628
0.2 6.480852 6.480436 0.641932
0.3 6.439226 6.439017 0.324435
0.4 6.397804 6.397531 0.426712
Maturity = 1.0
0.5 6.356578 6.356316 0.410849
Interest Rate = 0.1
0.6 6.315548 6.315321 0.359201
Initial price of Asset 1 = 100
0.7 6.27472 6.27449 0.367451
0.8 6.234087 6.233878 0.335393
Initial price of Asset 2 = 100
0.9 6.193652 6.19345 0.325424 Dividend of Asset 1 = 0.05
1 6.153411 6.153223 0.306302 Dividend of Asset 2 = 0.05
1.1 6.113369 6.113193 0.288202 Volatility of Asset 1= 0.2
1.2 6.07352 6.073361 0.261818 Volatility of Asset 1= 0.1
1.3 6.03387 6.033721 0.247201 Correlation = 0.5
1.4 5.994414 5.994279 0.2244
1.5 5.955153 5.95503 0.205267 Number of Discretisation N = 4096
1.6 5.916084 5.915977 0.181615 Integration step η = 1.0
1.7 5.877211 5.877117 0.161329 Scaling factor α = 2.5
1.8 5.838531 5.83845 0.138798
1.9 5.800047 5.79998 0.115989
2 5.761753 5.761697 0.098485 Table 1. Two-factor spread option prices
across strikes

! "# " " "$

 Accuracy of Alternative Methods
( Athlon 650 MHz with 512 MB RAM )

Fast Fourier Transform Monte Carlo

Number of Number of Time Steps
Discretisation Lower Upper Simulations 1000 2000
512 4.44 25.6 10000 129.15 0.051839 70.81 0.050949
1024 1.13 13.9 20000 22.34 0.036225 40.67 0.035899
2048 0.32 7.2 40000 7.44 0.025737 7.63 0.025733
4096 0.1 3.65 80000 18.34 0.018076 4.94 0.018184

Table 2. Accuracy of alternative methods for the two-factor geometric Brownian

motion model in which the analytic price is available using direct
integration: Error in basis points

! "# " " "$

 Impact of Volatility and Correlation
Correlation

0.5 0 -0.5
6.675496 8.494941 9.979849
0.1 6.675800 8.495493 9.981407
Volatility of asset 2

(0.454684) (0.649928) (1.561482)

7.510577 10.549590 12.870614 Maturity = 1.0

Interest Rate = 0.1
0.2 7.511055 10.550798 12.873037
Initial price of Asset 1 = 100
(0.636531) (1.145356) (1.882598) Initial price of Asset 2 = 95
Dividend of Asset 1 = 0.05
9.712478 13.261339 16.01200 Dividend of Asset 2 = 0.05
0.3 9.714326 13.264996 16.18352 Volatility of Asset 1= 0.2
(1.901766) (2.757088) (3.965540)
Strike of the spread option = 5.0
The first value is computed using the Fast Fourier Transform method.
The second value is the analytic price computed using the conditioning technique (the one-
dimensional integral is evaluated using the qromb.c routine in Numerical Recipes in C ).
The third value is the error of the FFT method in basis points.

Table 3. 2-factor spread option prices across volatilities and

correlations

! "# " " "$

 Pricing Spread Options under Three-factor
Stochastic Volatility Models

(
dS1 = S1 r dt + σ 1 vdW1 )
dS2 = S ( r dt + σ
2 2 vdW ) 2

dv = κ ( µ − v )dt + σ ν vdWν

EQ[dW1 dW2] = ρdt EQ[dW1 dWv] = ρ1dt EQ[dWv dW2] = ρ2dt

• Characteristic function is known in closed-form so that the FFT
method is applicable
• Benchmark with Monte Carlo and finite difference methods

! "# " " "$

Characteristic Function of the 3-Factor Model

φT (u1 , u2 ) := E [ exp(iu1s1 (T ) + iu2 s2 (T ))]

= exp iu1 ( rT + s1 ( 0 ) ) + iu2 ( rT + s2 ( 0 ) )

κµ (θ − Γ)(1 − e −θ T )
− 2 2 ln 1 − + (θ − Γ ) T
σv 2θ
2ζ (1 − e −θ T )
+ v(0)
2θ − (θ − Γ)(1 − e ) −θ T

! "# " " "$

 Computing Time of Alternative Methods
( Athlon 650 MHz with 512 MB RAM )
Monte Carlo: 1000 Time Steps
Number of 10 Strikes 100 Strikes
Simulations GBM SV GBM SV
10000 38.2 144.87 41.95 151.75
Fast Fourier Transform 20000 76.22 288.09 83.81 303.31
Number of 10 Strikes 100 Strikes 40000 152.5 576.25 168.48 606.53
Discretisation GBM SV GBM SV 80000 304.95 1152.9 335.2 1212.76
512 1.04 1.11 1.1 1.2
1024 4.28 4.64 4.48 4.83
Monte Carlo: 2000 Time Steps
2048 18.46 19.54 18.42 19.74
4096 74.45 81.82 76.47 81.27 Number of 10 Strikes 100 Strikes
Simulations GBM SV GBM SV
10000 75.57 287.41 79.83 295.21
20000 157.28 574.18 159.08 590.23
40000 303.37 1149.25 317.49 1184.32
80000 606.4 2298.37 636.33 2359.05

Table 4. Computational time (seconds) of alternative methods for the two-factor

Geometric Brownian motion model and the three-factor Stochastic
Volatility model

! "# " " "$

Spread Option Prices by Alternative Methods
Fast Fourier Transform
N Lower Upper
512 5.059379 5.068639
1024 5.062695 5.067405
2048 5.063545 5.065897
4096 5.063755 5.064492

Explicit Finite Difference

No. of Discretisation Monte Carlo Simulation
Space Time Price Nu m ber of
100 * 100 * 100 400 5.0845 Simu lation Step s Price error)
100 * 100 * 100 1600 5.0769 1280000 1000 5.052372 0.004301
100 * 100 * 100 2500 5.076 1280000 2000 5.053281 0.004297
100 * 100 * 100 10000 5.0748 1280000 4000 5.037061 0.004286
200 * 200 * 100 1600 5.0703 2560000 1000 5.04989 0.003039
200 * 200 * 200 1600 5.0703 2560000 2000 5.051035 0.003039
200 * 200 * 100 2500 5.0694 2560000 4000 5.042114 0.003037
200 * 200 * 100 10000 5.0682 5120000 1000 5.047495 0.002148
300 * 300 * 100 4000 5.0668 5120000 2000 5.046263 0.002148

Table 2. Accuracy of alternative methods for the three-factor

Stochastic Volatility model

! "# " " "$

Price Variation With the Volatility of the
Stochastic Volatility Stochastic Volatility
parameters

6.65 T = 1.0
r = 0.1
6.6
K = 2.0
Sp r e a d O p tio n p r ic e

6.55 S1 (0) = 100

S1 (0) = 98
6.5 δ 1 = δ 2 = 0.05
σ 1 = 1.0
6.45
σ 2 = 0.5
6.4 ρ = 0.5
v(0) = 0.04
6.35 κ = 1.0
µ = 0.04
6.3 ρ1 = −0.25
0.01 0.02 0.03 0.04 0.1 0.2 ρ 2 = −0.5
Volatility of Volatility

Figure 2. Spread option prices under Three-factor Stochastic Volatility Model

with varying volatility σ v of the stochastic volatility v

! "# " " "$

 Price Variation With the Mean Reversion
Rate of Volatility Stochastic Volatility
parameters
T = 1.0
6.62 r = 0.1
K = 2.0
6.6
S1 (0) = 100
Spread Option Price

6.58 S1 (0) = 98
6.56 δ 1 = δ 2 = 0.05
σ 1 = 1.0
6.54 σ 2 = 0.5
6.52 ρ = 0.5
v(0) = 0.04
6.5
σ v = 0.05
6.48 µ = 0.04
0.125 0.25 0.5 1 2 4 8 ρ1 = −0.25
Mean Reversion Rate of Volatility ρ 2 = −0.5

Figure 3. Spread option prices under Three-factor Stochastic Volatility Model

with varying mean reversion rate κ, of the stochastic volatility v

! "# " " "$

 Comparison of Two-factor and Three-
factor Prices Stochastic Volatility
parameters
T = 1.0
0.3 r = 0.1
K = 2.0
0.2 S1 (0) = 100
0.1 S1 (0) = 98
Price δ 1 = δ 2 = 0.05
0
Differentials σ 1 = 1.0
-0.1 σ 2 = 0.5
-0.2 ρ = 0.5
v(0) = 0.04
-0.3
σ v = 0.05
0.8

10 κ = 1.0
Correlation of 7
0.2

4 Correlation of Asset µ = 0.04

-0.4

Asset 1 and
1 2 and Volatility
-1

Volatility

Figure 5. Price difference between Three-factor Stochastic Volatility Model and

the Two-factor Geometric Brownian motion model (with implied
constant volatilities and correlation)

! "# " " "$

Comparison of Two-factor and Three-
factor State Price Densities

! "# " " "$

 State-Price Densities

State-Price Densities of Constant and Stochastic Volatility models,

with volatility of (a) 5% (b) 50%

! "# " " "$

 A Stochastic Volatility Model with
Jumps
1 2
ds1 = (r − δ 1 − σ 1 − λ1µ1 )dt + σ 1 dW1 + log(1 + J1 )dN1
2
1 2
ds2 = (r − δ 2 − σ 2 − λ1µ 2 )dt + σ 2 dW2 + log(1 + J 2 )dN 2
2
dv = κ ( µ − )dt + dWν
where N1 , N 2 are orthogonal Poisson processes independent of
W1 ,W2 ,Wν with constant arrival rates λ1 , λ2

• Specifically ( dN i (t ) = 1) = λi dt with probability generating

functions [ s N i ( t ) ] = exp(λi ( s − 1)t ]

! "# " " "$

 Market Calibration
Difficult to obtain OTC spread option data

A challenging econometric task because…

• Absence of closed-form expressions for risk-neutral and objective probability
density functions
• Presence of a latent variable the stochastic volatility process
• Difficult to combine the spot price and panel option price data (a highly non-
linear function in the log-spot price) optimally
• Neither the Maximum Likelihood (ML) method nor Kalman filters can be
applied (the common obstacle faced by SV model estimation)

Furthermore, there are two underlying assets now requiring

• Simultaneous estimation of two combined stochastic volatility models
• Calibration of a correlation surface
• Possibility of “correlation smiles” and “correlation skews”

! "# " " "$

 Recent Advances in the Estimation of the
Stochastic Volatility Models
Extensive interest in estimating single-asset SV models and some of the most
promising econometric techniques include:

• Generalized Method of Moments (GMM):

Hansen (1985), Pan (2000), Bollerslev & Zho (2000)
• Simulated Method of Moments (SMM):
Duffie & Singleton (1993), Bakshi,Cao, Chen (1997, 2000)
• Efficient Method of Moments (EMM):
Gallant & Tauchen (1996), Chernov & Ghysels (2000)

However, due to the complexity and computational burden, very few have
considered the problem of estimating a multi-asset SV model!

! "# " " "$

 Extension to Multiple Assets
Data assumed available:
• spot prices Sˆ1 , Sˆ 2
• a panel of vanilla option prices on individual assets across strikes and
maturities e.g.
ˆ n1 n = 1,.., N
- calls on asset 1: C1 1 1

- calls on asset 2: Cˆ 2n2 n2 = 1,.., N 2

Model:
• Three-factor Stochastic Volatility: (S1 , S 2 ,ν )
Parameters to be estimated:
Θ1 := {δ 1 , σ 1 , ρ1 , κ , µ , σ ν } Θ 2 := {δ 2 , σ 2 , ρ 2 , κ , µ , σ ν } Θ3 := {ρ }
• the first two sets Θ1 , Θ 2 contain structure parameters when the model is each
viewed as single asset SV models
• Θ 0 contains the parameters which are crucial in determining the state-
dependent correlation structure between the underlying assets

! "# " " "$

 Outline of the Calibration
Procedure
1. For i = 1,2 , conduct parallel estimation of structural parameters
in Θi based on panel data of call prices Cˆ ini ni = 1,..., N using
either SMM, Ordinary Least Squares or any previously
mentioned single-asset SV model estimation procedure

2. Based on the structural parameters obtained filter the unobserved

volatility process ν t with an OLS procedure on the combined
{ }
option data Cˆ1n1 (t ), Cˆ 2n2 (t )

3. Estimate the correlation-dependent parameters in Θ 0 by…

! "# " " "$

 Estimation of correlation
sensitive parameters I
If reliable market data on spread option prices are unavailable,
{
• Use the spot price time-series Sˆ (t ), Sˆ (t )
1 2 }
• Compute the correlation function between the terminal spot prices S1 (T ), S 2 (T )
conditional on S (t ), S (t ),ν (t ) by differentiating the characteristic
1 2
functions and evaluating the second moments

• Fix the parameters in Θ1 , Θ 2 which appear in this correlation function to the

estimates obtained so far so that the spot price correlation is a univariate function
in ρ ∈ Θ 0

• Manipulate ρ to take a forward view on the spot price correlation using this
function or evaluate its inverse at the historical value of the spot price correlation
for an optimal estimate ρ̂ of ρ

! "# " " "$

 Correlation structure between
asset prices under SV
• This refers to the terminal correlation between S1(T), S2(T)
conditional on the time-t state xt=(s1(t), s2(t), v(t)) over the horizon
[t,T] v (T )
ρ (T ) := 12

v11 (T )v22 (T )
v (T ) := ( s(T ) − [ s (T )]) ( s (T ) − [ s (T )]Τ )
2
∂ 2φT ∂φT ∂ 2φT ∂φT ∂φT
− −
∂u12 ∂u1 ∂u1∂u2 ∂u1 ∂u2
=− 2 u =0
∂ φT
2
∂φT ∂φT ∂ φT
2
∂φT
− −
∂u1∂u2 ∂u2 ∂u1 ∂u22 ∂u2

• Can be calculated by differentiating the characteristic function

using Maple

! "# " " "$

 Correlation between Asset Prices as a Function of
Correlation Parameters between Asset Prices and Volatility

T = 1.0
r = 0.1
S1 (0) = 100
S1 (0) = 96
δ 1 = δ 2 = 0.05
σ 1 = 1.0
σ 2 = 0.5
ρ = 0.9
v (0) = 0.04
σ v = 0.05
κ = 1.0
µ = 0.04

! "# " " "$

Excess Correlation at Maturity

Terminal Correlation less the Instantaneous Correlation Parameter ρ

! "# " " "$

 Correlation at Maturity

Terminal Correlation between Asset Prices (a) ρ=-0.8, (b) ρ=0.8

! "# " " "$

 Estimation of correlation
sensitive parameters II
If spread option price data Vˆ n (t ) are observed for a range of strikes
and maturities

• Take the parameter estimates and filtered volatility obtained above

as fixed and compute the dependence of the theoretical spread
option price V n (t ; ρ ) on ρ ∈ Θ 0 using the FFT pricing method

• Minimize the sum-of-square pricing errors to obtain an estimate

T N 2
ρˆ = arg min V (t ; ρ ) − Vˆ n (t )
n

ρ∈Θ 0 t = 0 n =1

! "# " " "$

 Conclusions and Future Work
• Existing approaches are unable to price spread options beyond
two-factor GBM models
• The Fast Fourier Transform provides a robust method for pricing
spread options with more factors under stochastic volatility and
correlation, general affine models, etc
• Computation times do not increase with the number of random
factors in the diffusion model
• Method is applicable to other exotic options
• Also have a 4-factor model which allows full freedom to the future
correlation surface…
• Fast wavelet transform O(N) versus O(N log N) for FFT

! "# " " "$