Upload
meko1986
View
242
Download
1
Embed Size (px)
Citation preview
8/10/2019 Euler and Milstein Discretization
1/8
Euler and Milstein Discretizationby Fabrice Douglas Rouah
www.FRouah.comwww.Volopta.com
"Monte Carlo simulation" in the context of option pricing refers to a setof techniques to generate underlying valuestypically stock prices or interestratesover time. Typically the dynamics of these stock prices and interest ratesare assumed to be driven by a continuous-time stochastic process. Simulation,however, is done at discrete time steps. Hence, the rst step in any simulationscheme is to nd a way to "discretize" a continuous-time process into a dis-crete time process. In this Note we present two discretization schemes, Eulerand Milstein discretization, and illustrate both with the Black-Scholes and theHeston models.
We assume that the stock price St is driven by the stochastic dierentialequation (SDE)
dSt= (St; t) dt + (St; t) dWt (1)
where Wt is Brownian motion. We simulate St over the time interval [0; T],which we assume to be is discretized as 0 = t1 < t2
8/10/2019 Euler and Milstein Discretization
2/8
identical fashion, the second integral is approximated as
Z t+dt
t
(Su; u) dWu (St; t)Z t+dt
t
dWu
= (St; t) (Wt+dt Wt)= (St; t)
pdtZ;
sinceWt+dtWt andp
dtZare identical in distribution, where Zis a standardnormal variable. Hence, Euler discretization of (2) is
St+dt = St+ (St; t) dt + (St; t)p
dtZ: (3)
1.1 Euler Scheme for the Black-Scholes Model
The Black-Scholes stock price dynamics under the risk neutral measure are
dSt= rStdt + StdWt: (4)
An application of Equation (3) produces Euler discretization for the Black-Scholes model
St+dt= St+ rStdt + Stp
dtZ: (5)
Alternatively, we can generate log-stock prices, and exponentiate the result. ByItos lemma ln St follows the process
d ln St =
r 1
22
dt + dWt: (6)
Euler discretization via Equation (3) produces
ln St+dt= ln St+
r 1
22
dt + p
dtZ
so that
St+dt= Stexp
r 1
22
dt + p
dtZ
: (7)
where dt= ti ti1.
1.2 Euler Scheme for the Heston Model
The Heston model is described by the bivariate stochastic process for the stock
priceS
t and its variancevt
dSt = rStdt +p
vtStdW1;t (8)
dvt = ( vt) dt + p
vtdW2;t
where E[dW1;tdW2;t] = dt.
2
8/10/2019 Euler and Milstein Discretization
3/8
1.2.1 Discretization of vt
The SDE for vt in (8) in integral form is
vt+dt= vt+
Z t+dtt
( vu) du +Z t+dtt
p
vudW2;u: (9)
The Euler discretization approximates the integrals using the left-point rule
Z t+dtt
( vu) du ( vt) dtZ t+dtt
p
vudW2;u pvt(Wt+dt Wt)
=
pvtdtZv
where Zv is a standard normal random variable. The right hand side involves( vt)rather than( vt+dt)since at time t we dont know the value ofvt+dt.This leaves us with
vt+dt= vt+ ( vt) dt + p
vtdtZv:
To avoid negative variances, we can replace vt with v+t = max(0; vt). This is
the full truncation scheme. Thereectionscheme replaces vt with its absolutevaluejvtj.
1.2.2 Process for St
In a similar fashion, the SDE for St in (8) is written in integral form as
St+dt = St+ r
Z t+dtt
Sudu +
Z t+dtt
pvuSudWu:
Euler discretization approximates the integrals with the left-point rule
Z t+dtt
Sudu StdtZ t+dtt
pvuSudW1;u
pvtSt(Wt+dt Wt)
=p
vtdtStZs
where Zs is a standard normal random variable that has correlation with Zv.We end up with
St+dt= St+ rStdt +p
vtdtStZs:
3
8/10/2019 Euler and Milstein Discretization
4/8
1.3 Process for lnSt
By Itos lemma ln St follows the diusion
d ln St =
r 1
2vt
dt +
pvtdW1;t
or in integral form
ln St+dt = ln St+
Z t0
r 1
2vu
du +
Z t0
pvudW1;u:
Euler discretization of the process forln St is thus
ln St+dt = ln St+ r 1
2vt
dt +p
vt(W1;t+dt W1;t) (10)
= ln St+
r 1
2vt
dt +
pvtdtZs:
Hence the Euler discretization ofSt is
St+dt = Stexp
r 1
2vt
dt +
pvtdtZs
:
Again, to avoid negative variances we must apply the full truncation or reectionscheme by replacing vt everywhere with v
+t or withjvtj.
1.3.1 Process for (St; vt) or (ln St; vt)
Start with the initial valuesS0for the stock price andv0 for the variance. Givena value for vt at time t, we rst obtain vt+dt from
vt+dt= vt+ ( vt) dt + p
vtdtZv
and we obtain St+dt from
St+dt = St+ rStdt +p
vtdtStZs
or from
St+dt = Stexp
r 1
2vt
dt +
pvtdtZs
:
To generate Zv and Zs with correlation , we rst generate two independent
standard normal variable Z1 and Z2, and we set Zv = Z1 and Zs = Z1 +p
1 2Z2.
4
8/10/2019 Euler and Milstein Discretization
5/8
2 Milstein Scheme
This scheme is described in Glasserman [2] and in Kloeden and Platen [4] forgeneral processes, and in Kahl and Jackel [3] for stochastic volatility models.The scheme works for SDEs for which the coecients (St) and (St)dependonly on S, and do not depend on t directly. Hence we assume that the stockprice St is driven by the SDE
dSt = (St) dt + (St) dWt (11)
= tdt + tdWt:
In integral form
St+dt = St+ Z t+dt
t
sds + Z t+dt
t
sdWs: (12)
The key to the Milstein scheme is that the accuracy of the discretization isincreased by considering expansions of the coecients t = (St) and t = (St) via Itos lemma. This is sensible since the coecients are functions ofS. Indeed, we can apply Itos Lemma to the functions t and t as we wouldfor any dierentiable function of S. By Itos lemma, then, the SDEs for thecoecients are
dt =
0tt+
1
200t
2t
dt + (0tt) dWt
dt =
0tt+
1
200t
2t
dt + (0tt) dWt
where the prime refers to dierentiation in Sand where the derivatives in t arezero because we assume that t and t have no direct dependence on t. Theintegral form of the coecients at time s (witht < s < t + dt)
s = t+
Z st
0uu+
1
200u
2u
du +
Z st
(0uu) dWu
s = t+
Z st
0uu+
1
200u
2u
du +
Z st
(0uu) dWu:
Substitute for s and s in (12) to produce
St+dt = St+ Z t+dt
t t+ Z
s
t 0uu+
1
200u
2u
du +
Z s
t
(0uu) dWuds
+
Z t+dt
t
t+
Z s
t
0uu+
1
200u
2u
du +
Z s
t
(0uu) dWu
dWs
The terms higher than order one are dsdu= O
(dt)2
and dsdWu = O
(dt)3=2
and are ignored. The term involving dWudWs is retained since dWudWs =
5
8/10/2019 Euler and Milstein Discretization
6/8
O (dt) is of order one. This leaves us with
St+dt = St+ t
Z t+dt
t
ds + tZ t+dt
t
dWs+Z t+dt
t
Z s
t
(0uu) dWudWs: (13)
Apply Euler discretization to the last term to obtain
Z t+dtt
Z st
0uudWudWs 0ttZ t+dtt
Z st
dWudWs (14)
= 0tt
Z t+dtt
(Ws Wt) dWs
= 0tt
Z t+dtt
WsdWs WtWt+dt+ W2t!
Now dene dYt = WtdWt. Using Itos Lemma, it is easy to show1 that Yt hassolutionYt=
1
2W2t 12 t so thatZ t+dt
t
WsdWs = Yt+dt Yt = 12
W2t+dt 1
2W2t
1
2dt: (15)
Substitute back into (14) to obtain
Z t+dtt
Z st
0uudWudWs 1
20uu
h(Wt+dt Wt)2 dt
i= :
1
20uu
h(Wt)
2 dti
:
where Wt = Wt+dt Wt, which is equal in distribution to pdtZ with Zdistributed as standard normal. Combining Equations (13) and (15) the generalform of Milstein discretization is therefore
St+dt= St+ tdt + tp
dtZ+1
20ttdt
Z2 1 : (16)
2.1 Milstein Scheme for the Black-Scholes Model
In the Black-Scholes model Equation (4) we have (St) = rSt and (St) = Stso the Milstein scheme (16) is
St+dt= St+ rStdt + Stp
dtZ+1
22dt Z
2 1which adds the correction term 1
22dt
Z2 1 to the Euler scheme in (5). In
the Black-Scholes model for the log-stock price, Equation (6), we have (St) =
1 Indeed, @Y@t
= 12; @Y@W
= W, and @2Y
@W2 = 1, so that dYt =
1
2+ 0 + 1
2 11
dt+
(Wt 1) dWt = WtdWt.
6
8/10/2019 Euler and Milstein Discretization
7/8
r 1
22
and (St) = so that
0
t = 0
t = 0: The Milstein scheme (16) istherefore
ln St+dt= ln St+
r 12
2
dt + pdtZ
which is identical to the Euler scheme in (7). Hence, while the Milstein schemeimproves the discretization ofStin the Black-Scholes model, it does not improvethe discretization ofln St.
2.2 Milstein Scheme for the Heston Model
Recall that this model is given in Equation (8) as
dSt = rStdt +p
vtStdW1;t
dvt = ( vt) dt + pvtdW2;t
2.3 Process for vt
The coecients of the variance process are (vt) = ( vt)and (vt) = pvtso an application of Equation (16) for vt produces
vt+dt= vt+ ( vt) dt + p
vtdtZv+1
42dt
Z2v 1
(17)
which can be written
vt+dt =
pvt+
1
2p
dtZv
2+ ( vt) dt 1
42dt:
This last equation is also Equation (2.18) of Gatheral [1]. Milstein discretizationof the variance process produces far fewer negative values for the variance thanEuler discretization. Nevertheless, the full truncation scheme or the reectionscheme must be applied to (17) as well.
2.3.1 Process for St and ln St
The coecients of the stock price process are (St) =rSt and (St) =p
vtStso Equation (16) becomes
St+dt = St+ rStdt +p
vtdtStZs+1
4S2t dt
Z2s 1
: (18)
We can also discretize the log-stock process, which by Itos lemma follows the
processd ln St =
r 1
2vt
dt +
pvtdW1;t:
The coecients are (St) =
r 12
vt
and (St) =p
vt so that 0
t = 0
t = 0.Sincevt is known at time t, we can treat it as a constant in the coecients. An
7
8/10/2019 Euler and Milstein Discretization
8/8
application of (16) produces
ln St+dt = ln St+
r 12
vt
dt +p
vtdtdZs
which is identical to Equation (10). Hence, as in the Black-Scholes model, thediscretization ofln St rather than St means that there are no higher correctionsto be brought to the Euler discretization. The discretization of the stock priceis
St+dt = Stexp
r 1
2vt
dt +
pvtdtZs
: (19)
Again, it is necessary to apply the full truncation or reections schemes inEquations (18) and (19).
2.4 Process for(St; vt)
or(lnSt; vt)
Given a value for vt at time t, we rst update to vt+dt using (17)
vt+dt= vt+ ( vt) dt + p
vtdtZv+1
42dt
Z2v 1
and we obtain St+dt using
St+dt = St+ rStdt +p
vtdtStZs+1
4S2t dt
Z2s 1
:
or from
St+dt = Stexp
r 1
2vt
dt +
pvtdtZs
:
To generate Zv and Zs with correlation , we rst generate two independentstandard normal variable Z1 and Z2, and we set Zv = Z1 and Zs = Z1 +p
1 2Z2.
References
[1] Gatheral, J. (2006) The Volatility Surface: A Practitioners Guide. NewYork, NY: John Wiley & Sons.
[2] Glasserman, P. (2003). Monte Carlo Methods in Financial Engineering.New York, NY: Springer.
[3] Kahl, C., and P. Jackel (2006). Fast Strong Approximation Monte-Carlo
Schemes for Stochastic Volatility Models. Quantitative Finance, Vol. 6, No.6.
[4] Kloeden, P.E., and E. Platen (1992). Numerical Solution of Stochastic Dif-ferential Equations. New York, NY: Springer.
8