Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure

Pure Jump

sa
li Hir
yA
eb
Computational Methods in Finance Fin
a n c
n
Lecture – Week 7ethods i
M
nal
t a tio
u
mp
Ali Hirsa
Co n
Industrial Engineering do & Operations Research
b ase
,
o nly Columbia University
ses
urp o
p
al
at ion
c
r edu
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Agenda

sa
li Hir
yA
n c eb
a
Fin
Brownian Bridge s in
t h od
Me
Monte Carlo Integration nal
t a tio
u
Quasi Monte Carlo mp
n Co
d o SDEs
Numerical Integrationaseof
b
ly,
Simulation of SDEs
on
ses
diffusion rpoprocesses
l pu
purea
ion jump processes
cat
r edu
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Brownian Bridge (1 of 3)

To simulate a standard Wiener process @ t sa

li Hir
√ yA
c eb
Wt − W0 ∼ tz where z ∼ N (0, 1) a n
Fin
in
ds
and knowing that W0 = 0 it becomes tho
l Me
a
√ tation
Wt ∼ omtz
p u
C
on
a s ed
b
ly,
s on
suppose we have se simulated Wt1 and Wt2
rp o
l pu
it is now ion
a desired to fill in points in the interval [t1 , t2 ], that is,
cat
toeduinterpolate between the already generated points Wt1 and
r
FoWt2
to do this we use a Brownian bridge that is required to go
through the values Wt1 and Wt2
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Brownian Bridge (2 of 3)

sa
li Hir
yA
n c eb
a
Fin
Definition (Brownian Bridge) s in
t h od
A Brownian bridge x is a process that atal 1 Me t has
timethe value a
n
and at time t2 has the value b. Betweent a tio t1 and t2 , x behaves like
pu
a Brownian motion. A Brownian C om
bridge satisfies
n
do
b ase
bnly−
, xt
dxt = eso dt + dBt , xt1 =a
rp o
s t2 − t
p u
nal
tio a standard Brownian motion
where Btcais
du
re
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Brownian Bridge (3 of 3)

sa
li Hir
yA
the SDE can be solved explicitly n c eb
a
Fin
dt Z s in
t2 − t t − t1 tho dBu
xt = a +b Me
+ (t2 −al t)
t2 − t1 t2 − t1 tio
n t1 t2 − u
uta
o mp
Knowing that the cond. dist.onofC xt is normal
d
ase
,b
n ) = a + (b − a) t − t1
l
Eest1o(x
y
t
rp o
s t2 − t1
p u
n al (t2 − t)(t − t1 )
io Vart1 (xt ) =
ucat t2 − t1
d
re
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Example (1 of 2)

sa
li Hir
construct a Brownian bridge between 0 and T to calculate yA its
n c eb
values at tj for j = 1, . . . , m−1 with i n a
F
in
o ds
th
0 < t1 < t2 < · · · < tm−1al < Metm = T
io n
tat
pu
first generate WT at T = Ctm o m
d on
use a Brownian bridge bas
e to get the entire path at
{t1 , t2 , t3 , . . . , tsm−1ly,
on }
ose
using thel pvalueurp of WT , and Wt0 = W0 = 0, it generates Wt1
i ona
cat
it dgenerates Wt2 using Wt1 and WT , and it generates Wt3
r e u
Fousing Wt2 and WT
the construction proceeds until we reach tm−1
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Example (2 of 2)

sa
li Hir
the discretely sampled Brownian path is generated by yA
determining its values at T , t1 , t2 , t3 , . . . , tm−1 naccording n c eb to
a
n Fi
√ i
ds
WT = T z1 ho
r l Met
t1 − t0 a − t1 )(t1 − t0 )
n(T
Wt1 = Wt0 + (WT − Wt0 ) +atio z2
T − t0put T − t0
o m
C r
otn − t1 (T − t2 )(t2 − t1 )
Wt2 = Wt1 + (WT − aW s etd1 ) 2 + z3
, b T − t1 T − t1
o nly
.. ses
. urp o
p
nal
r
t io= tm−1 − tm−2 (T − tm−1 )(tm−1 − tm−2 )
Wtm−1c a W tm−2 + (W T − W tm−2 ) + zm
edu T − tm−2 T − tm−2
r
Fo
where zi for i = 1, . . . , m are i.i.d. N (0, 1)
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Monte Carlo Integration (1 of 7)

sa
li Hir
yA
n c eb
a
consider Fin
s in
Z
f (x )dx t h od
Is Me
nal
t a tio
where Is is the s-dimensionalmunit pu cube, Is = [0, 1] × · · · × [0, 1]
C o
want to numerically evaluate n the integral for s = 20
do
b ase
for just 10 discretization , points on each dimension we will
nly
20 seso
have 10 number o of grid points
urp
n a lp
difficulty
io in computing the integral due to the high
cat
edu
dimensionality is referred to as the curse of dimensionality
r
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Monte Carlo Integration (2 of 7)

sa
li Hir
to overcome this hurdle, we employ Monte Carlo integration yA
n c eb
as opposed to try to compute the integral via ina
Fsome numerical
s in
integration scheme o d
th
Me
sample the set x1 , . . . , xN , uniformly n al distributed N vectors and
io
tat
evaluate the function f at aomset pu of points x1 , . . . , xN
nC
evaluate the followingsesum d o as an approximation for the
b a
integral: ly,
s on
se 1 X N
rp o
l pu θ N = f (xi )bIs
i o na N
cat i=1
r edu
Fowhere b I is the volume of the integration domain I
s s
Is = [0, 1] × · · · × [0, 1] so it has volume 1
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Monte Carlo Integration (3 of 7)

the summation is an approximation for the integral sa
li Hir
yA
Z
f (x )dx ≈ θN n c eb
a
Is Fin
s in
law of large numbers t h od
Me
nal N
a tio X
u t 1
lim θN =Complim f (xi )bIs
N↑∞ o n N↑∞ N i=1
d
ase
l y,b = bIs E(f (x ))
on
ses Z
1
urp o = Is f (x ) dx
p b
al
ion I Is
b
c at Z s
r edu
Fo = f (x )dx
Is

use the fact that the probability of xi , that is uniform over Is ,

is b1
Is
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Monte Carlo Integration (4 of 7)

how fast θN approaches the integral as N ↑ ∞ sa

li Hir
define yA
n c eb
a
Fin
Z
δN ≡ f (x )dx − θN s in
t h od
Is
lMe
aN
ionX
t1
Z
t a
= f (x )dxmp−u f (xi )bIs
Is n C o N i=1
d o Z
b aseXN
, 1

=
o nly f (x )dx − f (x )
i sI
b
s
rp ose N i=1 Is
pu
n al N
!
io 1 1
Z
cat
X
edu = f (x ) dx − f (xi ) bIs
r N bIs
Fo i=1 Is
N
!
bIs X Z
1
= f (x ) dx − f (xi )
N i=1 Is bIs
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Monte Carlo Integration (5 of 7)

denote sa
li Hir
yA
1 eb
Z
n c
∆f (xi ) = f (x ) dx − f (xi ) Fin
a
Is bIs in
ds
tho
l Me
na
tio
aN
bIpsutX
δN ≡ om ∆f (xi )
o nC N
d
ase i=1
,b
o nly
∆f (xi ) has zero ses mean
urp o
p
al
at ion E(∆f (xi )) = 0
c
r edu
Fo
also ∆f (xi ) and ∆f (xj ) are uncorrelated.

E(∆f (xi )∆f (xj )) = E(∆f (xi ))E(∆f (xj )) = 0

Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Monte Carlo Integration (6 of 7)

variance of δN sa
li Hir
2
Var(δN ) = E(δN ) − (E(δN ))2 yA
n c eb
a
N
bI 2 X Fin
s in
= E(∆f (xi )2 ) + h0ods
N2 Me
t
i=1 nal
bI 2 Z t a tio !2 
pu 1 1
Z 
s
= fC2om
(x ) dx − f (x ) dx
N d oIsn bIs Is bIs 
ase
,b
define o nly
s
ose
rp !2
pu Z
1
Z
1
n al 2 2
cat
io σ (f ) = f (x ) dx − f (x ) dx
edu Is bIs Is bIs
r
Fo
we get the following for the variance of the error term
bI 2
s
Var (δN ) = σ 2 (f )
N
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Monte Carlo Integration (7 of 7)

sa
li Hir
yA
n c eb
a
the rate of convergence is Fin
s in
t h od
q Me
1al
Var(δN ) ∼ on
ti√
u t a
mp N
Co
d on
from the variance weascan e see how Monte Carlo integration
,b
resolves the curse o nlyof dimensionality
ses
forthe integral
urpo of s = 20, N = 106 the error is approximately
p
al
O c√a1tNion ∼ √ 1 6 = 10−3
du 10
re
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Some Important Remarks (1 of 3)

sa
li Hir
yA
n c eb
magnitude in error term δN is independent of Fthe ina dimension s
 
s in
error convergence rate is O √N makes 1 d
thothe convergence
l Me
pretty slow a
ion
u tat
should employ some kind of p
variance reduction method, which
om
o nC
will be discussed latersed
ba
an error bound s only,
se
rp o
l puZ bIs X N bI 2
a
at ion f (x )dx ≈ f (x i ) ± s 2
σ (f )
c N i=1 N
edu Is
r
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Some Important Remarks (2 of 3)

sa
li Hir
if f (x ) is constant, f (x ) ≡ c, then yA
n c eb
a
n Fin ! 2
1 si
od 1
Z Z
σ 2 (f ) = f 2 (x ) dx − Mefth(x ) dx
Is bIs al I bIs
on s i
tat !2
pu Z
21 1
Z
m
= c Co −
ondx c dx
Is asedb
Is Is bIs
b
ly, Z !2
s on 1
Z
1
rp ose
= c 2
dx − c 2
dx
pu Is Is
n al Is b Is b
io
cat = c −c 2 2
r edu
Fo
= 0
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Some Important Remarks (3 of 3)

sa
li Hir
yA
n c eb
a
that is, what we expect to get if an integrand Fin
s in is a constant
d
and obviously there is no need to do the thointegral
l Me
the way to interpret σ 2 (f ) is how na
iomuch the function f (x )
u tat
deviates from a constant omp
C
on
if we can find a transformation a s ed that changes the coordinate x
y, b
to a new coordinate on
l y in which the transformed function is
rather flat,urthen ses we can reduce the variance of Monte Carlo
po
p
integration al
at ion
c
du
this
r e is the essence of variance reduction
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Quasi-Monte Carlo Methods (1 of 2)

Quasi-Monte Carlo methods can be regarded as a sa

li Hir
deterministic equivalent of classical Monte Carlo e by A
a nc
Fin no
used to evaluate multi-dimensional integralsinwith
s
closed-form solution od
eth M
consider nal
Z
t a tio
u
mfp(x )dx
Co
on Is
a s ed
over the s-dimensional b unit cube, Is = [0, 1] × · · · × [0, 1]
ly,
s on
in classical rMonte se Carlo integration we select set points
po
l pu
x1 , ...,ioxnNa , that is, a sequence of pseudo random numbers, and
cat
approximate the integral by
r edu
Fo
N
1 X
θN = f (xi )bIs
N i=1
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Quasi-Monte Carlo Methods (2 of 2)

sa
li Hir
yA
n c eb
a
Fin
s in
od
in quasi-Monte Carlo methods we selectthpoints
Me
deterministically nal o
ut ati
specifically, quasi-Monte Carlo mp methods produce a
C o
deterministic sequence dofonpoints that provides the best
a s e
possible spread in lyI,s b
on
these deterministic ses sequences are referred to as
urp o
l p
low-discrepancy sequences
a
on i
cat
edu
r
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Example (1 of 3)

sa
li Hir
yA
n c eb
a
Fin
s in
Z 1Z 1 t h od
x 3 (1 + y 2 )dxdyna=
l Me1
io 3
0 0 tat
mpu
Co
n
do
b aNse
y, X
nl1
o seso Uj3 (1 + Vj2 )
rp N
pu j=1
n al
io
cat
duj and Vj are
wherer eU i.i.d. U(0, 1) for j = 1, . . . , N
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Example (2 of 3)

sa
li Hir
yA
n c eb
a
Fin
s in
t h od
Me
nal
t a tio
u
mp
n Co
do
b ase
,
nly
seso
rp o
pu
n al
io
cat
r edu
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Example (3 of 3)

sa
li Hir
yA
n c eb
a
Fin
s in
t h od
Me
nal
t a tio
u
mp
n Co
do
b ase
,
nly
seso
rp o
pu
n al
io
cat
r edu
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Numerical Integration of SDEs (1 of 3)

sa
li Hir
yA
consider the following generic one-dimensional SDE n c eb
a
Fin
s in
dXt = µ(Xt , t)dt + σ(Xt , t)dWtht od t0 ≤ t ≤ T
Me
X (t0 ) = X0 al
t i on
uta
o mp
by Itô-Taylor expansion on C
d
ase
l y,b Z t Z t
on
Xt = Xt0os+ e µ(Xt0 )
s ds + σ(Xt0 ) dW (s)
rp t0 t0
l pu
a
ion 1
 
0
cat +
2
d u σ(X t0 )σ (X t 0 ) [W (t) − W (t 0 )] − (t − t0 ) +R
re 2
Fo
where R is the remainder
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Numerical Integration of SDEs (2 of 3)

when we simulate an SDE we generate samples of the discretized version sa
of SDE at a finite number of points li Hir
yA
n c eb
X̂∆t , X̂2∆t , . . . , X̂m∆t a
Fin
s in
d
m is number of time steps & ∆t is time step assuming tho equidistant
subintervals, ∆t = Tm−0 l Me
a
ion
u tat
, X̂p tj , . . . , X̂tm
X̂t1 , X̂t2 , . . .om
o nC
tj = t0 + j∆t = j∆t for aj s= ed 1, . . . , m
l y,b
∆t → 0 our discretized on path will converge toward the continuous path
o ses
for the interval r p
pu [tj , ti+1 ] , by choosing
i onal
cat t0 = tj ,
r edu
Fo t = tj+1 ,
∆t = tj+1 − tj ,
∆Wj = W (tj+1 ) − W (tj ),
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Numerical Integration of SDEs (3 of 3)

sa
li Hir
yA
n c eb
a
Fin
s in
t h od
we get the following expression Me
nal
t a tio
Xtj+1 = Xtj + µ(Xtj )∆t pu
C om + σ(Xtj )∆Wj +
1 on
0 ased

2
σ(Xtj )σ b(X t ) (∆W j ) − ∆t +R
2 ly, j
s on
se
rp o
l pu
a
ion
ucat
r ed
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Simulating SDEs

sa
li Hir
yA
n c eb
a
Fin
s in
t h od
There are various schemes for simulating SDEs
Me of this form, and
al
the most common ones are ion
tat
pu
Euler scheme C om
n
do
Milstein scheme , base
ly
on
Runge–Kuttaosscheme es
rp
pu
n al
io
cat
edu
r
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Euler scheme

sa
li Hir
yA
n c eb
Euler scheme is the simplest discretization scheme Fin
a available
for discretizing SDEs s in
d
tho
keeping the first three terms gives us l Me explicit Euler method
the
a
ion
u tat
mp
X̂tj+1 = X̂tj + µ( Cotj , tj )∆t + σ(X̂tj , tj )∆Wj
n X̂
d o √
= lX̂ ase
y, tbj + µ(X̂tj , tj )∆t + σ(X̂tj , tj ) ∆tZj
on
o ses
where Zjl pare p
ur i.i.d. N (0, 1)
i ona
t
thisucaapproximation expands the drift
ed √ term to O(∆t) but only
r
Foexpands the diffusion term to O( ∆t)
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Milstein scheme

sa
li Hir
yA
n c eb
Milstein scheme improves upon the Euler discretizationFinby a adding a
s n
ito
second diffusion term, expanding the diffusion term od O(∆t)
Meth
it is obtained by simply keeping all terms of a l O(∆t)
ion
u tat 1
p
X̂tj+1 = X̂tj +µ(Xtj , tj )∆t +σ(X C otm
j , tj )∆Wj + σ(Xtj , tj )σ 0 (Xtj , tj )[(∆Wj )2 − ∆t]
o n 2
ed √ 1
bas +σ(X̂tj , tj ) ∆tZj + σ(Xtj , tj )σ 0 (Xtj , tj )∆t(Zj2 −1)
= X̂tj +µ(X̂tj , ty,j )∆t
nl 2
s o
ose ∂
where σ 0 (x ,pt)
urp= ∂x (σ(x , t))
al
while athe
t ion Milstein scheme has a higher order in discretization, it requires
c
du
knowing the first derivative of the the volatility function
re
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Runge-Kutta Scheme

sa
li Hir
Milstein requires knowledge of the first derivative cof yA
n e b the
volatility function, which may not be available na all or may be
Fiat
s in
expensive to compute d
tho
l Me
Runge-Kutta scheme allows us totioavoid na using the first
u ta
derivative of the volatility function, mp by using the Runge-Kutta
Co
nkeeping
approximation, while still d o the same order of accuracy
b ase
,
nly √
seso
Xi =rpo Xi + µ(Xi )∆t + σ(Xi ) ∆t
b
pu
al
ion = Xi + µ(Xi )∆t + σ(Xi )∆Wi
Xati+1
c
du 1 h b
re
i
Fo + √ σ(X ) − σ(X ) ((∆W )2 − ∆t)
i i i
2 ∆t
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Simulating SDEs under Different Models

sa
li Hir
yA
n c eb
a
Fin
s in
geometric Brownian motion (GBM) t h od
Me
Ornstein-Uhlenbeck process nal
t a tio
u
Cox-Ingersoll-Ross (CIR) mp
Co n
GBMSA (Heston) do
b ase
,
nly
variance gamma ses o(VG) model
rp o
pu
variancealgamma w/ stochastic arrival (VGSA)
ion
c at
du
re
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Geometric Brownian Motion

GBM is governed by the following SDE: sa

li Hir
yA
dXt = µXt dt + σXt dWt n c eb
a
Fin
s in
the solution via Itô’s lemma t h od
Me
(
2io nal )
tat
σ
XT = X0 exp (µom−pu )T + σWT
on
C 2
d
ase
l y,b
the distribution of
on XT is
known and we can simulate XT
ses is no need for discretization
po
directly, sourthere
p
al
at ion
c
r edu
Fo √
( )
σ2
XT = X0 exp (µ − )T + σ T Z
2

where Z ∼ N (0, 1)
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Ornstein-Uhlenbeck (OU) process

sa
li Hir
yA
c eb
OU process is governed by a n
Fin
in
ds
tho
dXt = κ(θ − Xt )dt + σdW
Me t
nal
t a tio
solution via Itô’s lemma u
mp
Co
on Z T
sed
XT = e −κT X0ly+ a
, b θ(1 − e
−κT
) + σe κT e κs dWs
n
so 0
rp ose
pu
XT nownaldepends on the entire path of the Brownian motion
o
uc ati
however,
d the distribution of XT is known and we can simulate
re
FoXT directly without having to discretize the SDE
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

CIR process

sa
li Hir
the CIR process is governed by the following SDE: yA
n c eb
a
p Fin
dXt = κ(θ(t) − Xt )dt + σ Xt dW s tin (1)
d
e tho
M
There is no explicit solution to the SDE. nal
t a tio We do not necessarily
need an explicit solution to determine pu the distribution of X . In
C om T
case of θ(t) = θ we know sthat d on XT is distributed as a non-central
ae
chi-squared from which ,b
o nly we can easily simulate. Unfortunately, once
we move to a CIR es
swith a time varying θ(t), the distribution of XT
urp o
is not available, l p and one method of simulating XT indirectly is
a
ion
cat and simulating the SDE. This situation, where we do
discretizing
u
d
re
notFoknow the distribution of X , is typical. Therefore, it is often
T
necessary to simulate an SDE.
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Vesicek vs. CIR

sa
li Hir
yA
0.25
Vasicek
CIR
0.18
n c eb
a
0.16 Fin
s in
t h od
0.2 0.14 Me
nal
t a tio
0.12
u
mp
Co
level xt

level xt
0.15 n
do 0.1

b ase
,
nly 0.08

0.1 seso
rp o
pu 0.06

n al
io
cat 0.04

r edu 0.05
Fo 0.02
Vasicek
CIR

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

calendar time calendar time
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Heston stochastic volatility model (GBMSA) (1 of 3)

Heston SDE is given vy sa

li Hir
√ yA
dSt = rSt dt + vt St dWS (t), n c eb
a
√ Fin
dvt = κ(θ − vt )dt + σ vt ddWs in v (t),
tho
l Me
two Brownian components WS (t)tioand na Wv (t) are correlated
u ta
with rate ρ p
om
o nC
conditional on time s saed Euler discretization of the variance
ba
process for t > sonreads ly,
ses
urp o q √
p
vo(t)
n al = v (s) + κ(θ − v (s))∆t + σ v (s) ∆tzv
i
cat
edu
r
Fowhere ∆t = t − s with zv ∼ N (0, 1)
can show the above scheme can go negative with positive
probability
that is the main issue with this scheme
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Heston stochastic volatility model (GBMSA) (2 of 3)

sa
several fixes in the literature on this issue
li Hir
yA
Lord et al unify several Euler schemes in the following n c eb
a
framework: Fin
s in
d
thoq √
v (t) = f1 (v (s)) + κ(θ − f2 (v (s)))∆t l Me σ f (v (s)) ∆tz
+
n a 3 v
io
u tat
p
all should satisfy fi (x ) = nx Cfor om x ≥ 0 and f3 (x ) ≥ 0 for all x
o
ed
the one that seems, to
y bas work the best is produced by full
l
truncation scheme on
o ses
r p
it chooses puf (x ) = x , f2 (x ) = f3 (x ) = x + where
+ i onal 1
x u= cat max(x , 0)
r ed
Fothe resulting scheme is
q √
v (t) = v (s) + κ(θ − v (s)+ )∆t + σ v (s)+ ∆tzv
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Heston stochastic volatility model (GBMSA) (3 of 3)

either directly apply an Euler discretization scheme to the sa

li Hir
A
price proces or to simulate it from its exact distribution
by
n ce
ina
direct discretization yields the following Euler Fscheme
in
q thods √
M + ∆tz e
S(t) = S(s) + rS(s)∆t + S(s)
nal v (t) S
io
tat
mpu
alternatively the exact solution Co by applying Itô’s lemma yields
d on
Z tbase Z tq
1

nly,
S(t) = S(s) exp o (r − v (u))du + v (u)dWS (u)
ses 2
urp o s s
p
al
by taking
at ion the log and utilizing Euler discretization we obtain
c
du
re
Fo 1 q √
log(S(t)) = log(S(s)) + [r − v (s)+ ]∆t + v (s)+ ∆tzS
2
p
where zS = ρzv + 1 − ρ2 z with z ∼ N (0, 1)
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Full Truncation Algorithm

sa
li Hir
yA
n c eb
a
full truncation scheme for Heston can be summarized Fin by:
s in
generate a random sample zv from a standard o d normal
th
distribution Me
n al
o
having v (s), compute v (t) utati
mp
generate a random sample o z from
p the standard normal
nC
distribution and setsedzSo = ρzv + 1 − ρ2 z
a
having log S(s), ,b
o nly compute log S(t)
ses
po
it producesurbiased estimates for coarse time intervals
p
n al
shown tiothat at least one has to use 32 time steps per year to
ed uca
obtain
r reasonable small biases
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Variance Gamma Process (1 of 3)

sa
li Hir
yA
n c eb
VG process X (t; σ, ν, θ) is obtained by evaluating Brownian
Fin
a motion
in
with drift θ and volatility σ at a random time ogiven
ds by a gamma
e t h
process γ(t; 1, ν) with mean rate unity and
al variance rate ν as
M
ion
u tat
X (t; σ, ν, θ) = θγ(t;Co1, mpν) + σW (γ(t; 1, ν))
o n
d
ase
with the characteristicnlfunction y, b
o
o ses
r p t/ν
1

pu
φ(u)
n al = E(e iuXt ) =
io 1 − iuθν + σ 2 u 2 ν/2
cat
edu
r
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Variance Gamma Process (2 of 3)

sa
li Hir
yA
c eb
suppose the stock price process is given by the geometric a n VG law
n Fin
with parameters σ, ν, θ and the log price at time i
ds t is given by
e tho
M
ln St = ln S0 + (r − q + ω)t nal
t a tio + X (t; σ, ν, θ)
pu
C om
where do
n
b ase
,
nly 1
sω
es o = − log(φ(−i))
r p o t
pu
n al 1
io = ln(1 − θν − σ 2 ν/2)
cat
edu ν
r
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Variance Gamma Process (3 of 3)

sa
li Hir
yA
c eb
In addition to the volatility of the normal distribution a nσ, there are
n Fin
parameters that control for i
ds
e tho
(i) kurtosis, ν (long tailedness, a symmetric increase in the left
M
nal
and right tail probabilities relative
t a tioto the normal for the return
pu
distribution) C om
n
do
(ii) skewness, θ, that allows ase for the asymmetry of the left and
,b
o nly
right tails of the
ses return density
rp o
An additional pu
attractive feature of the model is that it nests the
i onal
lognormal at
c density and the Black-Merton-Scholes formula as a
du
re
Fo
parametric special case (ν = 0 and θ = 0)
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Sampling VG process (1 of 2)
T
assume N equidistant time intervals of length h where h = sa
Hir
N
sample from a gamma dist. w/ mean h and variance νh A li
by
eβ:
a gamma process w/ shape parameter α & scale parameter a n c
x in
Fin
1 α−1 − sβ
f (x , α, β) = x ed
Γ(α)β α tho
l Me
a
its mean and variance are ion
u tat
mp
Co = αβ
nµ
d o 2
ase σ = αβ 2
l y,b
on
in the case of the
o sesvariance gamma process
r p
pu αβ = h
i onal
at
c αβ 2 = νh
r edu
Fo
implies
h
α =
ν
β = ν
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Sampling VG process (2 of 2)

sa
li Hir
yA
n c eb
a
Fin
s in
t h od
Therefore a sample for the VG process, X (h; Me σ, ν, θ), is
al
s tation
h pu
C om h
θg( , ν) + n σ g( , ν)z
ν ased o ν
l y,b
n
where z ∼ N (0, 1) s o g( h , ν) ∼ gamrand( h , ν)
seand
r p o ν ν
pu
n al
io
cat
edu
r
Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

algorithm

sa
li Hir
an algorithm for simulating a VG process yA
n c eb
a
for i = 1, . . . , N Fin
s in
z ∼ N (0, 1)
t h od
g ∼ gamrand(h/ν, ν) Me
√ nal
Xi = θg + σ gz t a tio
u
end mp
Co
d on
for the log of stock price
ase process b
for i = 1, . . .o,nN ly,
ses
logrpSoi = log Si−1 + (r − q)h + ωh + Xi
pu
end n al
i o
cat
edu
where
r ω = ν1 log(1 − θν − σ 2 ν/2)
Fo
note X (t; σ, ν = 0, θ = 0) = σWt & ω = − 21 σ 2
 Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

GBM vs. VG

sa
li Hir
yA
n c eb
a
Fin
150

s in
115 VG VG

od
GBM GBM

140 t h
Me
nal
110

t a tio
130
u
mp
stock price

stock price
105

n Co
do 120

b ase
,
nly
100

seso 110

rp o
95
pu
n al 100

io
cat
0.1 0.2

r edu
0.3 0.4 0.5 0.6
calendar time
0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5
calendar time
0.6 0.7 0.8 0.9 1

Fo
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Variance Gamma w/ Stochastic Arrival (VGSA) (1 of 4)

define the CIR process y (t) as the solution to SDE differential
sa
li Hir
equation A
√ by
dyt = κ(η − yt )dt + λ yt dWt inance
F
in
o ds
th
• η is the long-term rate of time change Me
n al
• κ is the rate of mean reversion io
tat
• λ is the volatility of the time mpu change
Co
on
process y (t) is the instantaneous e d rate of time change and so
y, bas
the time changeois n given by Y (t) where
l
o ses
r p Z t
pu
n al Y (t) = y (u)du
io
ucat 0
d
re
Fothe characteristic function for Y (t) is given by

E(e iuY (t) ) = φ(u, t, y (0), κ, η, λ)

= A(t, u)e B(t,u)y (0)
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Variance Gamma w/ Stochastic Arrival Process (2 of 4)

sa
li Hir
yA
where n c eb
a
 n Fin
κ2 ηt ds i

exp λ
e t2ho
A(t, u) = M
al 2κη/λ2
on κ

i
cosh(γt/2)tat+ sinh(γt/2)
pu γ
om
o n C2iu
B(t, u) = ased
ly, κ + γ coth(γt/2)
b
s on
se
rp o
l pu
with tiona
ca
r edu
Fo
p
γ= 2 2
κ − 2λ iu
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Variance Gamma w/ Stochastic Arrival Process (3 of 4)

the VGSA process is defined by sa

li Hir
yA
Z (t) = XVG (Y (t); σ, ν, θ) n c eb
a
Fin
= θγ(Y (t); 1, ν) + σW (γ(Ys in(t); 1, ν))
od
eth
n a lM
σ, ν, θ, κ, η, and λ are the six parameters io defining the process
tat
pu
characteristic function is given m
Co by
d on
ase 1
E(e iuZVGSA ,b
(t)
o nly ) = φ(−iΨVG (u), t, , κ, η, λ)
ses ν
rp o
l pu
wheretioφna is the characteristic function of Y (t) and ΨVG is the
ca
edu characteristic function of the variance gamma process at
log
r
Founit time, namely,
1  
ΨVG (u) = − log 1 − iuθν + σ 2 νu 2 /2
ν
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Variance Gamma w/ Stochastic Arrival Process (4 of 4)

we define the stock process at time t by the random variableHirsa

Ali
by
e (r −q)t+Z (t) anc
e
S(t) = S(0) Fin
E[e Z (t) ] s in
t h od
Me
we note that nal
t a tio
u
mp 1
Co
E[e Z (t) ] = φ(−iΨ onVG (−i), t, , κ, η, λ)
ed ν
bas
ly,
on
s e −ωt in the VG case. Therefore the
which is equivalent seto
urp o
characteristic p
al function of the log of the stock price at time t is
at ion
given by
du
c
re
Fo
φ(−iΨVG (u), t, ν1 , κ, η, λ)
E[e iu log St ] = exp(iu(log S0 + (r − q)t)) ×
φ(−iΨVG (−i), t, ν1 , κ, η, λ)iu
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Characteristic function of VGSA (1 of 3)

to simulate the VGSA process, as before, we assume N i Hirsa

equidistant time intervals of length h where h = Te/N. Al
by We
nc
ina
wish to simulate the VGSA process over each Fsub-interval h at
s i n
time t + h ho
d
t
Me
we can write n al
io
tat
mpu
∆Zt = Z (t) − Z (ton−Coh)
ed
bas 1, ν) + σW (γ(Y (t); 1, ν))
= θγ(Yly,(t);
on
ses
− po(θγ(Y (t − h); 1, ν) + σW (γ(Y (t − h); 1, ν)))
r
l pu
ion = θ(γ(Y
a (t); 1, ν) − γ(Y (t − h); 1, ν))
ucat q
r ed + σ γ(Y (t); 1, ν) − γ(Y (t − h); 1, ν)z
Fo

where z is N (0, 1)
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Characteristic function of VGSA (2 of 3)

gamma process γ(Y (t); 1, ν) with mean Y (t) and variance Y (t)ν has the
sa
following shape and scale parameters li Hir
yA
Y (t) n c eb
α = a
ν Fin
s in
β = ν d
tho
l Me
a
that implies ion
u tat 
p
om
  
C Y (t) Y (t − h)
γ(Y (t); 1, ν) − γ(Y (t − h); 1, ν)) on= gamma , ν − gamma ,ν
d ν ν
b ase
,
nly
 
so Y (t) − Y (t − h)
o se = gamma ,ν
r p ν
l pu
a
ion
ucat
by the summation property of gamma processes. Hence
r ed
Fo   s  
Y (t) − Y (t − h) Y (t) − Y (t − h)
∆Zt = θ gamma , ν + σ gamma ,ν z
ν ν
Rt
where Y (t) − Y (t − h) = t−h
y (u)du
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Characteristic function of VGSA (3 of 3)

for simulation, start by discretization of the CIR process irsa

H
Milstein discretization of the CIR process gives Ali
by
nce
√ F
2 ina
√ s inλ h(z 2 − 1)
yj = yj−1 + κ(η − yj−1 )h + λ yj−1 hz t h od+
Me 4
al
t i on
where yj is an approximation to utay (tj ) with tj = jh for
o mp
j = 0, . . . , N and z ∼ N o(0, n C1)
a s ed
for the interval (tj−1 b, tj ) the new clock is given by the integral
ly,
s on
se Z tj
rp o
l pu y (u)du
a
at ion tj−1
c
du
re
Foby the trapezoidal rule we get
Z tj
h
y (u)du = (yj−1 + yj )
tj−1 2
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Simulating VGSA

sa
li Hir
yA
c eb
to simulate a VGSA process we do a n
Fin
for j = 1, . . . , N s in
t h od
z ∼ N (0, 1) l Me
√ ona √ λ2
yj = yj−1 + κ(η − yj−1 )h + λutatiyj−1 hz + 4 h(z
2
− 1)
p
t̂j = h2 (yj + yj−1 ) om
nC o
t̂ d
g = gamrand( νj ,bν) ase
l y,
z ∼ N (0, 1)s on
Xj = θg ose √
urp+ σ gz
p
al
end at ion
c
r edu
where
Fo Xj is an approximation to ∆Ztj
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Simulating VGSA (alternative)

sa
li Hir
an alternative using the scaling property of gamma processes yA
n c eb
for j = 1, . . . , N a
Fin
s in
z ∼ N (0, 1) d
√ lM √etho λ2
yj = yj−1 + κ(η − yj−1 )h + λ yj−1 na hz + 4 h(z 2 − 1)
btj = h2 (yj + yj−1 ) t a tio
pu
q
C om
σb = σ btj n
do
b ase
νb = ν/btj ly,
on
θ̂ = θbtj poses
r
pu
g =algamrand( 1
n ν̂ , ν̂)
i o
cazt ∼ N (0, 1)
r edu √
Fo Xj = θgb +σ b gz
end
Agenda Brownian Bridge MC Integration Num Int of SDEs Sim SDEs Sim SDEs Diff Models Sim of Pure Jump

Simulating VGSA
for the logarithmic of stock price we have
sa
li Hir
Z (t)
log St = log S0 + (r − q)t + Z (t) − log(E(e )) yA
n c eZb(t−h)
log St−h = log S0 + (r − q)(t − h) + Z (t − h) − log(E(e a ))
Fin
s in
d
saubtracting tho
l Me
na
log St = log St−h + (r − q)h + Z (t) − Zta(t tio− h) + log(E(e Z (t−h) )) − log(E(e Z (t) ))
p u
= log St−h + (r − q)h + ∆Z om+ log(E(e Z (t−h) )) − log(E(e Z (t) ))
o n Ct
d
= log St−h + (r − q)h ase+ ∆Zt + ∆ωt
l y,b
on
thus
o ses
ur p
for j = 1,na.l .p. , N
t i o
cad 1
r edu ∆ω j = log(φ(−iΨVG (−i), (j − 1)h, ν , κ, η, λ)) −
Fo log(φ(−iΨ (−i), jh, 1 , κ, η, λ))
VG ν

log Sj = log Sj−1 + (r − q)h + ∆ω

dj + Xj
end