econstor

www.econstor.eu

Der Open-Access-Publikationsserver der ZBW – Leibniz-Informationszentrum Wirtschaft

The Open Access Publication Server of the ZBW – Leibniz Information Centre for Economics

Stilger, Przemyslaw Stan; Poon, Ser-Huang

Working Paper
Multi-level Monte Carlo simulations with importance
sampling

Manchester Business School Working Paper, No. 637

Provided in Cooperation with:

Manchester Business School, The University of Manchester

Suggested Citation: Stilger, Przemyslaw Stan; Poon, Ser-Huang (2013) : Multi-level Monte
Carlo simulations with importance sampling, Manchester Business School Working Paper, No.
637

This Version is available at:

http://hdl.handle.net/10419/102376

Nutzungsbedingungen: Terms of use:

Die ZBW räumt Ihnen als Nutzerin/Nutzer das unentgeltliche, The ZBW grants you, the user, the non-exclusive right to use
räumlich unbeschränkte und zeitlich auf die Dauer des Schutzrechts the selected work free of charge, territorially unrestricted and
beschränkte einfache Recht ein, das ausgewählte Werk im Rahmen within the time limit of the term of the property rights according
der unter to the terms specified at
→ http://www.econstor.eu/dspace/Nutzungsbedingungen → http://www.econstor.eu/dspace/Nutzungsbedingungen
nachzulesenden vollständigen Nutzungsbedingungen zu By the first use of the selected work the user agrees and
vervielfältigen, mit denen die Nutzerin/der Nutzer sich durch die declares to comply with these terms of use.
erste Nutzung einverstanden erklärt.

zbw
Leibniz-Informationszentrum Wirtschaft
Leibniz Information Centre for Economics
 Working Paper Series
Multi-Level Monte Carlo Simulations with Importance
Sampling

Przemyslaw Stan Stilger

Ser-Huang Poon

Manchester Business School Working Paper No 637

Manchester Business School

Copyright © 2013, AUTHOR STILGER & POON All rights reserved.
Do not quote or cite without permission from the author.

Manchester Business School

The University of Manchester
Booth Street West
Manchester M15 6PB

+44(0)161 306 1320

http://www.mbs.ac.uk/cgi/apps/research/working-papers/

ISSN 0954-7401

The working papers are produced by The University of Manchester - Manchester Business School and
are to be circulated for discussion purposes only. Their contents should be considered to be preliminary.
The papers are expected to be published in due course, in a revised form and should not be quoted
without the authors’ permission.
 Multi-Level Monte Carlo Simulations with Importance

Sampling

Przemyslaw Stan Stilger and Ser-Huang Poon∗

June 2, 2013

Abstract

We present an application of importance sampling in a Monte Carlo simulation for multi-

asset options and in a Multi-Level Monte Carlo simulation. We demonstrate that applying im-

portance sampling only on the first level of the Multi-Level Monte Carlo significantly improves

its effective performance. We extend the Likelihood Ratio Method Based on Characteristic

Function to estimate the Greeks of multi-asset options and in a Multi-Level Monte Carlo in a

computationally efficient manner. Moreover, we combine it with the importance sampling to

reduce the variance of the Greeks. Finally, we study the impact of the skew on the effective

performance of importance sampling.

Keywords: Importance sampling; Simulation; Stochastic volatility

∗ Przemyslaw Stan Stilger ([email protected]) and Ser-Huang Poon (ser-

[email protected]) are both at Manchester Business School, Crawford House, University of Manchester,
Oxford Road, Manchester, M13 9PL, United Kingdom.

1
1 Introduction

In practice, the valuation of contingent claims is typically a multi-dimensional problem that involves

Monte Carlo simulation. The rate of convergence of a Monte Carlo simulation is n−1/2 , where n

is the number of sample paths. Hence, improving the accuracy by a factor of 10 requires 100

times as many sample paths. For this reason variance reduction techniques become essential.

Importance sampling reduces the variance by changing the drift of the simulated sample paths.

Variance reduction achieved through importance sampling very much depends on the change of

drift. Much of research effort focuses on how to change the drift to fully exploit the variance

reduction potential of importance sampling. Multi-Level Monte Carlo is a Monte Carlo simulation

performed on different levels. The main advantage of Multi-Level Monte Carlo is that it has lower

computational complexity due to reduced variance compared to the basic Monte Carlo. Moreover,

the variance of Multi-Level Monte Carlo can be further reduced by combining it with other variance

reduction technique such as importance sampling.

In this paper, we focus on importance sampling for multi-asset options and importance sampling

in a Multi-Level Monte Carlo simulation. Our contribution is as follows. First, we present an

application of importance sampling with a stochastic change of drift to multi-asset options. Next,

we provide an efficient importance sampling scheme in a Multi-Level Monte Carlo simulation.

Finally, we extend the Likelihood Ratio Method Based on Characteristic Function to estimate

the Greeks of multi-asset options and in a Multi-Level Monte Carlo and we combine it with the

importance sampling to reduce the variance of the Greeks.

There is relatively little work on variance reduction for multi-asset options. Barraquand (1995)

introduces quadratic resampling and combines it with the importance sampling. Avramidis (2002)

proposes an algorithm that selects the importance sampling density as a mixture of multivariate

Normal densities. Neddermeyer (2011) develops non-parametric importance sampling in conjunc-

tion with quasi-random numbers. Su and Fu (1999), Bouhari (2004) as well as Caprotti (2008)

use importance sampling with a deterministic change of drift. In Su and Fu (1999) the change of

drift is depends on stochastic optimization. In Bouhari (2004) the change of drift relies on the

Robbins-Monro algorithms whereas in Caprotti (2008) it is based on the least squares minimization.

Pellizzari (1998) suggests the use of control variate based on unconditional and conditional expec-

tations of asset prices as a variance reduction technique for multi-asset options in the Black-Scholes

model.

2
 The rest of the paper is organized as follows. In Section 2 we present an application of impor-

tance sampling with a stochastic change of drift to multi-asset options in the Heston stochastic

volatility model and the Bates stochastic volatility model with jumps. We consider basket, best-of,

worst-of, spread, absolute and composite options. We also extend the Likelihood Ratio Method

Based on the Characteristic Function to multi-asset options. In Section 3 we discuss the appli-

cation of importance sampling in a Multi-Level Monte Carlo using the Heston-Hull-White model

and the Heston-Cox-Ingersoll-Ross model. We demonstrate that applying importance sampling

only on the first level can significantly improve the effective performance of the Multi-Level Monte

Carlo. We use the Likelihood Ratio Method Based on the Characteristic Function to estimate

the Greeks in a Multi-Level Monte Carlo and combine it with the importance sampling to reduce

the variance of the Greeks. We also study the impact of the skew on the effective performance

of importance sampling. What we mean by the skew is the correlation between asset returns and

their volatility. We study this feature because it is an important stylized fact of financial time

series. Finally, Section 4 concludes the paper.

2 Importance Sampling

Here, we present an application of importance sampling for multi-asset option in the Heston model.

We use a stochastic change of drift for a stochastic volatility model that was derived by Fouque

and Tullie in [7].

In the Heston model the stock price dynamics under the risk-neutral measure Q is

√ Si
dSi,t = ri Si,t dt + vi,t Si,t dWi,t (1)

whereas the variance dynamics is given by

√ vi
dvi,t = κi (θi − vi,t ) dt + ξi vi,t dWi,t (2)

where κi is the i-th mean-reversion rate, θi is the i-th long-term variance and ξi is the i-th volatility

of the volatility. The correlation matrix is

 
 C1 C2 
C=  (3)
C2> C3

3
where  
 ρ1,1 ··· ρ1,n 
 .. .. .. 
C1 = 
 . . . 
 (4)
 
ρn,1 ··· ρn,n

is the correlation corresponding to the asset price processes

 
 ρ1,n+1 ··· ρ1,2n 
 .. .. .. 
C2 = 
 . . . 
 (5)
 
ρn,n+1 ··· ρn,2n

is the correlation corresponding to the asset price processes and the variance processes and

 
 ρn+1,n+1 ··· ρn+1,2n 
 .. .. .. 
C3 = 
 . . . 
 (6)
 
ρ2n,n+1 ··· ρ2n,2n

is the correlation corresponding to the variance processes.

The difference between the Heston model and the Bates model is that in the Bates model the

stock price dynamics under the risk-neutral measure Q becomes

√ Si
dSi,t = Si,t ri − λi k̄i dt + Si,t vi,t dWi,t + Si,t dZi,t (7)

where Zi,t is a compound Poisson process with intensity λi and log-normal distribution of jump

sizes such that if ki is its jump size then ln (1 + ki ) ∼ N ln 1 + k̄i − 12 δi2 , δi2 .

To apply this result to multi-asset options we will express the Heston model in matrix notation.

Let

dXt = b (Xt ) dt + a (Xt ) dηt (8)

where C =ΣΣ> , η
t is a 2n-dimensional
  correlated
 Q-Brownian  motion and
S1
S W r1 s1
 1,t   1,t   
 .   .  ..
 ..   .. 
 
   

 . 

     
Sn
 Sn,t   Wn,t rn sn
     
  
dXt = 

 , ηt = 
   , b(x) = 
 
v1 
 
 v1,t   W1,t  κ1 (θ1 − v1 ) 
     
 .   .   .. 
 .   . 
 .   .  .
 
 
     
vn
vn,t Wn,t κn (θn − vn )

4
 √
 
v1 s1 0 ··· ··· ··· 0
 
 .. .. 

 0 . ··· ··· ··· . 

 .. √ .. 
. ··· vn sn ··· ··· .
 
 
a(x) =  .. ..

 √ 

 . ··· ··· ξ1 v1 ··· . 

 .. ..

. ··· ··· ··· . 0
 
 
 
√
0 ··· ··· ··· 0 ξ n vn
Let h (t, Xt ) be η-adapted process. We introduce the martingale

ˆ T ˆ T
!
−1 −1 1 −1 −1
Ht = exp Σ h(s, Xs ) · Σ dηt + Σ h(s, Xs ) · Σ h(s, Xs )ds (9)
0 2 0

and define a new probability measure Q̃ by the density

dQ̃ −1
= (HT ) (10)
dQ

By the Girsanov theorem for correlated Brownian motions, the process

ˆ t
η̃t = ηt + h (s, Xs ) dηs (11)
0

is a 2-dimensional correlated Q̃-Brownian motion. Thus under a new probability measure Q̃ the

model dynamics becomes

dXt = (b (Xt ) − a (Xt ) h (t, Xt )) dt + a (Xt ) dη̃t (12)

The optimal choice of h for which the variance of the Monte Carlo estimator under Q̃ is minimized

is
1
h(t, Xt ) = − a(t, Xt )> ∇P (t, Xt ) (13)
P (t, Xt )

This result
is also
 valid for the  Bates model with the difference that 
√ dZ1,t
v1 + S s1 0 ··· ··· ··· 0
 dW1,t1 
 
 .. .. 

 0 . ··· ··· ··· . 

. ..
 
√
 
 .. ···
dZ
vn + dWn,t sn ··· ··· .

 Sn 
a(x) =  n,t 
 .. √ .. 
··· ··· ξ1 v1 · · ·
 
 . . 
 
 .. .. 

 . ··· ··· ··· . 0 

√
 
0 ··· ··· ··· 0 ξn vn
We refer to [13] for details.

5
 For both models we approximate P (t, Xt ) using its fast mean-reversion expansion as in [7]

which is given by the price of a European option in the Black-Scholes model with volatility

v
uN N
uX X
σ̄ = t σi2 − 2 σi σj ρi,j (14)
i=1 1≤i<j≤N

PN
and initial stock price S̄ = i=1 Si . We use the fast mean-reversion expansion because it can

be used in conjunction with a look-up table introduced in [13] which dramatically reduces the

computational time.

2.1 Basket Option

The payout of a basket option depends on the performance of a basket of underlying assets, each

with its own corresponding weight. The weights wi must satisfy the constraints 0 ≤ wi ≤ 1 for all
Pn
i = 1, · · · , n and i=1 wi = 1. The payout of the basket call with maturity T is given by

max (w1 S1 (T ), · · · , wn Sn (T ) − K, 0) (15)

The main advantage of a basket option is that it offers great flexibility in the construction of the

underlying basket and it is usually cheaper than buying vanilla options on each of the underlying

assets.

As an example we will consider basket call on three underlying assets. We assume that the

time to maturity is 1 year. Other parameters are

i w S r v0 ξ κ θ
1 50% 70 0.05 0.04 0.4 3 0.09
2 30% 35 0.05 0.09 0.3 0.5 0.25
3 20% 40 0.05 0.25 0.2 5 0.04

Table 1: Model parameters

and the correlation matrix is

 
 1 0.4 0.2 −0.6 −0.28 −0.1 
 
 0.4
 1 0.5 −0.24 −0.7 −0.25 
 
 0.2 0.5 1 0.0282 −0.35 −0.5 
 



 (16)
 −0.6 −0.24 0.0282 1 0.168 0.0294 
 
 
 −0.28 −0.7 −0.35 0.168 1 0.175 
 
 
−0.1 −0.25 −0.5 0.0294 0.175 1

6
 We note that the Feller condition is satisfied for each underlying asset.

In a Monte Carlo simulation we use 10 000 sample paths and a time increment of 0.001. In

Table 2 we report the results for basic Monte Carlo and importance sampling for the Heston model

and the Bates model. For the Bates model, we assume that the jump intensity is 1 per year,

standard deviation of the jumps is 2% and the mean jump size is -5%.
K
S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Heston
MC 23.0003 18.0939 13.4929 9.4321 6.1308 3.6583 2.0060 0.9916
price
IS 23.0432 18.1301 13.5388 9.4302 6.0855 3.6288 1.9740 0.9759
MC 139.1319 131.7882 116.7452 93.7079 66.5853 41.7007 22.8492 11.1104
variance
IS 23.9918 22.8241 27.9146 15.9959 11.1225 10.7502 5.6949 1.8306
Bates
MC 22.9964 18.0972 13.5079 9.4722 6.1935 3.7289 2.0761 1.0521
price
IS 22.9257 18.0389 13.4853 9.4724 6.2015 3.7453 2.0880 1.0719
MC 143.5208 135.8888 120.4728 96.8471 69.1762 43.8538 24.5189 12.2881
variance
IS 32.9773 31.0928 26.9779 21.0307 14.3609 8.3187 4.1190 1.7322
K
Table 2: Price and variance of price for basket call as a function of S0 .

On average importance sampling reduces the variance 5 times compared to the basic Monte

Carlo.

2.2 Best-of Option

Best-of option depends on the performance of the best performing asset in a basket. The payout

of the best-of call with maturity T is given by

max (max (S1 (T ), · · · , Sn (T )) − K, 0) (17)

Best-of call has a higher upside potential compared to a call option on the same a basket of

underlying assets.

As an example we will consider best-of three underlying assets call option. We assume that the

time to maturity is 1 year. Other parameters are

i S r v0 ξ κ θ
1 30 0.05 0.04 0.4 3 0.09
2 35 0.05 0.09 0.3 0.5 0.25
3 40 0.05 0.25 0.2 5 0.04

Table 3: Model parameters

and the correlation matrix is given by (16). We also note that the Feller condition is satisfied

for each underlying asset.

7
 In a Monte Carlo simulation we use 10 000 sample paths and a time increment of 0.001. In

Table 4 we report the results for basic Monte Carlo and importance sampling for the Heston model

and the Bates model. For the Bates model, we assume that the jump intensity is 1 per year,

standard deviation of the jumps is 2% and the mean jump size is -5%.

K
S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Heston
MC 23.3939 20.0806 16.8117 13.6660 10.7524 8.1722 5.9940 4.2450
price
IS 23.3965 20.0837 16.8275 13.6950 10.7276 8.1459 5.9882 4.2398
MC 114.9330 114.2491 112.0612 106.5591 96.5682 82.5957 66.5238 50.5769
variance
IS 58.0854 49.6236 40.2004 31.9864 24.6266 19.9598 15.4075 11.2659
Bates
MC 23.5387 20.2257 16.9608 13.8165 10.8969 8.3088 6.1219 4.3675
price
IS 23.4815 20.1805 16.9499 13.8360 10.8988 8.3327 6.1667 4.4003
MC 118.2053 117.5027 115.1506 109.5496 99.5909 85.5577 69.3061 53.0277
variance
IS 57.7069 49.5691 40.9534 33.3357 26.6619 21.6881 16.7958 12.4169
K
Table 4: Price and variance of price for best-of call as a function of S0 .

On average importance sampling reduces the variance 3 times compared to the basic Monte

Carlo.

2.3 Worst-of Option

Worst-of option depends on the performance of the worst performing asset in a basket. The payout

of the worst-of call with maturity T is given by

max (min (S1 (T ), · · · , Sn (T )) − K, 0) (18)

Worst-of call has a lower upside potential compared to a call option on the same a basket of

underlying assets.

As an example we will consider worst-of three underlying assets call option. We use the same

parameters as in the previous section. In a Monte Carlo simulation we use 10 000 sample paths and

a time increment of 0.001. In Table 5 we report the results for basic Monte Carlo and importance

sampling for the Heston model and the Bates model. For the Bates model, we assume that the

jump intensity is 1 per year, standard deviation of the jumps is 2% and the mean jump size is -5%.

8
 K
S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Heston
MC 7.5960 5.0870 3.1251 1.7429 0.8760 0.3895 0.1544 0.0590
price
IS 7.6359 5.1117 3.1181 1.7279 0.8628 0.3849 0.1524 0.0533
MC 42.3979 32.2608 21.3585 12.1857 5.9854 2.5606 0.9838 0.3423
variance
IS 19.3028 13.8514 8.7862 4.7116 2.1009 0.7885 0.2509 0.0701
Bates
MC 7.4994 5.0152 3.0794 1.7277 0.8753 0.3964 0.1649 0.0615
price
IS 7.5106 5.0174 3.0583 1.6931 0.8489 0.3835 0.1561 0.0580
MC 42.7988 32.4943 21.5437 12.3510 6.1418 2.6990 1.0523 0.3754
variance
IS 20.1260 14.3138 8.9745 4.7930 2.1609 0.8268 0.2723 0.0798
K
Table 5: Price and variance of price for worst-of call as a function of S0 .

On average importance sampling reduces the variance 3 times compared to the basic Monte

Carlo.

2.4 Spread option

Spread option depends on the difference between two underlying assets. Seller of such an option

is long correlation which differentiates it from the majority of multi-asset options that leave the

seller short correlation. The payout of the spread call with maturity T is given by

max (S1 (T ) − S2 (T ) − K, 0) (19)

As an example we will consider spread call. We assume that the time to maturity is 1 year.

Other parameters are

i S r v0 ξ κ θ
1 30 0.05 0.04 0.4 3 0.09
2 5 0.05 0.09 0.3 0.5 0.25

Table 6: Model parameters

and the correlation matrix is

 
1 0.4 −0.6 −0.28
 
 
 0.4 1 −0.24 −0.7 
(20)
 
 
 −0.6 −0.24 1 0.168 
 
 
−0.28 −0.7 0.168 1

We note that the Feller condition is satisfied for each underlying asset.

In a Monte Carlo simulation we use 10 000 sample paths and a time increment of 0.001. In

Table 7 we report the results for basic Monte Carlo and importance sampling for the Heston model

9
and the Bates model. For the Bates model, we assume that the jump intensity is 1 per year,

standard deviation of the jumps is 2% and the mean jump size is -5%.

K
S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Heston
MC 10.9116 8.7664 6.8078 5.0821 3.6278 2.4616 1.5897 0.9777
price
IS 10.8962 8.7438 6.7701 5.0316 3.5795 2.4298 1.5730 0.9676
MC 51.5956 47.0730 40.6259 32.9627 25.0319 17.7671 11.7748 7.3314
variance
IS 3.4245 3.7729 2.9216 1.7977 1.3673 1.0649 0.7246 0.3987
Bates
MC 10.9352 8.8015 6.8525 5.1433 3.7056 2.5509 1.6764 1.0552
price
IS 10.9055 8.7716 6.8252 5.1218 3.6918 2.5521 1.6944 1.0786
MC 54.2643 49.5032 42.8706 34.9473 26.7347 19.1973 12.9460 8.2356
variance
IS 5.4703 4.7225 4.1768 3.4911 2.5868 1.9515 1.4214 0.9444
K
Table 7: Price and variance of price for spread call as a function of S0 .

On average importance sampling reduces the variance 13 times compared to the basic Monte

Carlo.

2.5 Absolute option

Absolute option is an option written on the absolute value of the difference between the two

underlying assets at maturity. The payout of the spread call with maturity T is given by

max (max (S1 (T ), S2 (T )) − min (S1 (T ), S2 (T )) − K, 0) (21)

As an example we will consider absolute call on two underlying assets. We assume that the

time to maturity is 1 year. Other parameters are

i S r v0 ξ κ θ
1 30 0.05 0.04 0.4 3 0.09
2 35 0.05 0.09 0.3 0.5 0.25

Table 8: Model parameters

and the correlation matrix is given by (20). We also note that the Feller condition is satisfied

for each underlying asset.

In a Monte Carlo simulation we use 10 000 sample paths and a time increment of 0.001. In

Table 9 we report the results for basic Monte Carlo and importance sampling for the Heston model

and the Bates model. For the Bates model, we assume that the jump intensity is 1 per year,

standard deviation of the jumps is 2% and the mean jump size is -5%.

10
 K
S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Heston
MC 0.8110 0.4772 0.2775 0.1561 0.0860 0.0470 0.0260 0.0137
price
IS 0.8157 0.4738 0.2703 0.1517 0.0837 0.0457 0.0241 0.0124
MC 8.4177 4.9509 2.8182 1.5635 0.8538 0.4603 0.2410 0.1229
variance
IS 4.3064 1.9271 0.9153 0.4107 0.1753 0.0765 0.0256 0.0071
Bates
MC 0.8988 0.5351 0.3091 0.1733 0.0947 0.0509 0.0271 0.0143
price
IS 0.8731 0.5180 0.3015 0.1726 0.0973 0.0530 0.0289 0.0158
MC 9.2814 5.4544 3.0895 1.6995 0.9159 0.4850 0.2541 0.1307
variance
IS 3.9397 1.8787 0.8458 0.3651 0.1514 0.0554 0.0207 0.0082
K
Table 9: Price and variance of price for absolute call as a function of S0 .

On average importance sampling reduces the variance 7 times compared to the basic Monte

Carlo.

2.6 Composite

Composite option is an option on a foreign underlying asset with a strike denominated in the

domestic currency. The holder of a composite option faces foreign exchange risk, but benefits from

having fixed the strike in the domestic currency. The payout of the composite call with maturity

T is given by

max (S(T )F X(T ) − K, 0) (22)

where FX is the foreign exchange rate.

As an example we will consider composite call. We assume that the time to maturity is 1 year.

Other parameters are

i S r v0 ξ κ θ
1 30 0.05 0.04 0.4 3 0.09
2 2 0.05 0.09 0.3 0.5 0.25

Table 10: Model parameters

and the correlation matrix is given by (20). We also note that the Feller condition is satisfied

for each underlying asset.

In a Monte Carlo simulation we use 10 000 sample paths and a time increment of 0.001. In

Table 11 we report the results for basic Monte Carlo and importance sampling for the Heston

model and the Bates model. For the Bates model, we assume that the jump intensity is 1 per year,

standard deviation of the jumps is 2% and the mean jump size is -5%.

11
 K
S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Heston
MC 14.7229 11.8809 9.2202 6.8346 4.8012 3.1750 1.9658 1.1369
price
IS 14.7025 11.8549 9.1903 6.8037 4.7791 3.1696 1.9776 1.1546
MC 132.9938 113.6243 93.6035 73.2430 53.7984 36.7750 23.3859 13.9662
variance
IS 65.9776 54.7920 44.3035 34.0895 24.4774 16.1298 9.6617 5.2622
Bates
MC 14.7446 11.9122 9.2596 6.8880 4.8743 3.2598 2.0555 1.2193
price
IS 14.7270 11.8902 9.2423 6.8705 4.8617 3.2663 2.0775 1.2507
MC 138.4657 118.5183 98.0459 77.1913 57.2165 39.7261 25.8310 15.8694
variance
IS 70.2021 58.6839 47.6199 36.9331 26.9349 18.1927 11.4240 6.8354
K
Table 11: Price and variance of price for composite call as a function of S0 .

On average importance sampling reduces the variance 2 times compared to the basic Monte

Carlo.

2.7 Greeks

Price of a multi-asset option can be expressed as expectation under the risk-neural measure Q

h i
P = E e−r(T −t) φ (S1 (T ), · · · , Sn (T )) (23)

where φ (S1 (T ), · · · , Sn (T )) is the payout function. Using the definition of expectation, this can

be written as

ˆ ∞ ˆ ∞
P = ··· e−r(T −t) φ (S1 (T ), · · · , Sn (T )) f (x1 , · · · , xn )dx1 · · · dxn (24)
0 0

where f (x1 , · · · , xn ) is the joint density.

Let us consider the first derivative of the option price with respect to S1 (0)

ˆ ∞ ˆ ∞
∂
∆ = ··· e−r(T −t) φ (S1 (T ), · · · , Sn (T )) f (x1 , · · · , xn )dx1 · · · dxn
∂S1 (0) 0 0
ˆ ∞ ˆ ∞ ∂
∂S (0) f (x1 , · · · , xn )
= ··· e−r(T −t) φ (S1 (T ), · · · , Sn (T )) 1 f (x1 , · · · , xn )dx1 · · · dxn
0 0 f (x1 , · · · , xn )

∂
∂S1 (0)
f (x1 ,··· ,xn )
where f (x1 ,··· ,xn ) is the likelihood ratio. By Sklar’s Theorem there exists a copula C such

that

F (x1 , · · · , xn ) = C (F1 (x1 ) , · · · , Fn (xn )) = C (u1 , · · · , un ) (25)

In [13] we showed that the CDF and PDF for both the Heston model and the Bates model can

12
be obtained as

ˆ ∞  
1 1 exp (−iω ln (x1 )) φT (ω)
F1 (x1 ) = P r (S1 (T ) ≤ x1 ) = − Re dω
2 π 0 iω
F1 (x1 + ∆x) − F1 (x1 )
f1 (x1 ) ≈
∆x

where φ is the characteristic function. nth order differentiation of (25) gives an expression for the

joint density.
n
Y
f (x1 , · · · , xn ) = fi (xi ) c (u1 , · · · , un ) (26)
i=1

∂ n C(u1 ,··· ,un )

where c (u1 , · · · , un ) = ∂u1 ···∂un . Other Greeks can be computed in a similar manner. In order

to estimate the Greeks with respect to correlation between assets we will use an analytical copula

as an approximation of copula in (25).

As an example let us again consider basket call. As an approximation of the joint PDF for the

Heston model we will use t copula with 62 degrees of freedom and correlation matrix

 
 1 0.4 0.2 
 
 0.4 1
 0.5 
 (27)
 
0.2 0.5 1

In Figure 1 we display the joint PDF for the first and second underlying asset in Table 1. The

joint PDF in the upper panel was obtained using the Monte Carlo simulation. The joint PDF in

the lower panel was obtained using t copula with 62 degrees of freedom and marginal PDFs of

the first and second underlying asset in Table 1. Correlation matrix and the number of degrees of

freedom were estimated using maximum likelihood.

13
Figure 1: Joint PDF for the first and second underlying asset in Table 1 for the Heston model.
 In Tables 12 and 13 we report delta and gamma of each underlying asset computed by the

Likelihood Ratio Method Based on the Characteristic Function for the Heston model.
K
i S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
MC 0.4973 0.4850 0.4572 0.4070 0.3346 0.2500 0.1678 0.1005
delta
IS 0.4978 0.4832 0.4528 0.4005 0.3290 0.2499 0.1702 0.1045
1
MC 3.9044 2.9080 2.0993 1.4687 0.9938 0.6445 0.3937 0.2223
variance
IS 2.9194 1.9176 1.1636 0.6470 0.3326 0.2544 0.1250 0.0439
MC 0.2968 0.2873 0.2671 0.2329 0.1870 0.1366 0.0906 0.0524
delta
IS 0.2961 0.2865 0.2673 0.2365 0.1910 0.1352 0.0874 0.0501
2
MC 10.9800 7.8665 5.3811 3.4908 2.1292 1.2105 0.6341 0.3048
variance
IS 9.8116 6.5965 4.1291 2.3213 1.1894 0.6826 0.2735 0.0868
MC 0.1991 0.1886 0.1728 0.1506 0.1219 0.0897 0.0583 0.0340
delta
IS 0.2025 0.1932 0.1789 0.1593 0.1315 0.0995 0.0689 0.0430
3
MC 7.7937 5.4444 3.6204 2.2803 1.3568 0.7634 0.4083 0.2102
variance
IS 6.2314 4.0029 2.3606 1.2534 0.6082 0.2816 0.1127 0.0388
K
Table 12: Delta and variance of delta for basket call as a function of S0 .

On average importance sampling reduces the variance of delta 2 times.

K
i S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
MC 0.0001 0.0009 0.0025 0.0048 0.0071 0.0084 0.0083 0.0067
gamma
IS 0.0020 0.0026 0.0039 0.0060 0.0079 0.0089 0.0086 0.0071
1
MC 0.0479 0.0378 0.0291 0.0217 0.0157 0.0108 0.0070 0.0042
variance
IS 0.0372 0.0262 0.0174 0.0107 0.0061 0.0033 0.0018 0.0009
MC 0.0142 0.0115 0.0093 0.0077 0.0065 0.0054 0.0042 0.0027
gamma
IS 0.0197 0.0169 0.0144 0.0120 0.0100 0.0079 0.0055 0.0036
2
MC 0.4378 0.3224 0.2279 0.1536 0.0977 0.0581 0.0322 0.0164
variance
IS 0.4136 0.2865 0.1852 0.1098 0.0587 0.0298 0.0127 0.0046
MC 0.0104 0.0082 0.0063 0.0047 0.0036 0.0029 0.0024 0.0020
gamma
IS 0.0127 0.0103 0.0081 0.0062 0.0047 0.0034 0.0025 0.0017
3
MC 0.1819 0.1294 0.0883 0.0576 0.0360 0.0218 0.0130 0.0078
variance
IS 0.1433 0.0928 0.0554 0.0301 0.0150 0.0079 0.0032 0.0011
K
Table 13: Gamma and variance of gamma for basket call as a function of S0 .

On average importance sampling reduces the variance of gamma 2 times.

As an approximation of the joint PDF for the Bates model we will use t copula with 30 degrees

of freedom and the correlation matrix given by (27).

In Figure 2 we display the joint PDF for the first and second underlying asset in Table 1. The

joint PDF in the upper panel was obtained using the Monte Carlo simulation. The joint PDF in

the lower panel was obtained using t copula with 30 degrees of freedom and marginal PDFs of

the first and second underlying asset in Table 1. Correlation matrix and the number of degrees of

freedom were estimated using maximum likelihood.

15
Figure 2: Joint PDF for the first and second underlying asset in Table 1 for the Bates model.
 In Tables 14 and 15 we report delta and gamma of each underlying asset computed by the

Likelihood Ratio Method Based on the Characteristic Function for the Bates model.
K
i S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
MC 0.4966 0.4845 0.4567 0.4066 0.3347 0.2512 0.1706 0.1044
delta
IS 0.4912 0.4769 0.4489 0.3976 0.3292 0.2488 0.1722 0.1092
1
MC 3.8657 2.8944 2.1049 1.4874 1.0193 0.6713 0.4182 0.2421
variance
IS 2.8849 1.9074 1.1758 0.6710 0.3565 0.1826 0.0906 0.0463
MC 0.3056 0.2933 0.2708 0.2342 0.1862 0.1364 0.0906 0.0529
delta
IS 0.2895 0.2812 0.2598 0.2284 0.1803 0.1296 0.0833 0.0477
2
MC 10.8908 7.8325 5.3818 3.5098 2.1576 1.2400 0.6602 0.3245
variance
IS 9.6904 6.5213 4.1040 2.3413 1.2146 0.5691 0.2377 0.0889
MC 0.1803 0.1722 0.1588 0.1390 0.1136 0.0831 0.0543 0.0317
delta
IS 0.1766 0.1701 0.1611 0.1442 0.1206 0.0939 0.0655 0.0404
3
MC 7.5413 5.2603 3.4916 2.1943 1.3011 0.7289 0.3872 0.1981
variance
IS 6.0178 3.8632 2.2887 1.2276 0.6000 0.2648 0.1058 0.0389
K
Table 14: Delta and variance of delta for basket call as a function of S0 .

On average importance sampling reduces the variance of delta 2 times.

K
i S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
MC 0.0023 0.0028 0.0040 0.0060 0.0080 0.0091 0.0087 0.0071
gamma
IS 0.0037 0.0040 0.0055 0.0071 0.0089 0.0100 0.0094 0.0078
1
MC 0.0477 0.0379 0.0295 0.0223 0.0163 0.0114 0.0076 0.0047
variance
IS 0.0342 0.0242 0.0168 0.0105 0.0063 0.0036 0.0020 0.0011
MC 0.0210 0.0174 0.0142 0.0115 0.0093 0.0072 0.0053 0.0033
gamma
IS 0.0267 0.0227 0.0199 0.0163 0.0132 0.0098 0.0072 0.0048
2
MC 0.4307 0.3207 0.2294 0.1565 0.1009 0.0608 0.0340 0.0175
variance
IS 0.3641 0.2523 0.1679 0.1049 0.0561 0.0279 0.0123 0.0051
MC 0.0123 0.0096 0.0071 0.0050 0.0035 0.0025 0.0019 0.0014
gamma
IS 0.0117 0.0094 0.0075 0.0057 0.0041 0.0032 0.0021 0.0013
3
MC 0.1674 0.1184 0.0801 0.0515 0.0315 0.0185 0.0105 0.0059
variance
IS 0.1210 0.0781 0.0472 0.0258 0.0129 0.0061 0.0026 0.0009
K
Table 15: Gamma and variance of gamma for basket call as a function of S0 .

On average importance sampling reduces the variance of gamma 2 times.

3 Multi-Level Monte Carlo with Importance Sampling

Multi-Level Monte Carlo is a Monte Carlo simulation with different number of time steps hl = 2−l T

on each level l = 0, 1, · · · , L. Let P denote the payout and Pl denote its approximation on level l.

The Multi-Level Monte Carlo estimator is given by

L
X
Yb = E [PL ] = E [P0 ] + E [Pl − Pl−1 ] (28)
l=1

Let Ŷ0 denote an estimator for E [P0 ] using N0 sample paths and let Ŷl denote an estimator for

17
E [Pl − Pl−1 ] using Nl sample paths. Ŷl is calculated as a mean of Nl independent samples

Nl
1 X
Pli − Pl−1
i


Ŷl = (29)
Nl i=1

It is important to note that Pl and Pl−1 are obtained using the same Brownian path. This

is done by first constructing Nl Brownian increments to evaluate Pli and then summing them in
Nl i
pairs, yielding 2 Brownian increments, to evaluate Pl−1 .

The variance of the Multi-Level Monte Carlo estimator is

L L
h i X h i X 1
V ar Ŷ = V ar Ŷl = Vl (30)
Nl
l=0 l=0

where Vl = V ar [Pl − Pl−1 ].

The mean square error is given by

 2  h i  h i 2
M SE = E Ŷ − E [P ] = V ar Ŷ + E Ŷ − E [P ] (31)

 h i   h i 2
where E Ŷ − E [PL ] is the discretization error. By Theorem 3.1 in [8] E Ŷ − E [PL ] ≤
h i
2 2

2 and V ar Ŷ ≤ 2 so the upper bound on the mean square error is 2 .

PL
We wish to minimize the variance of Y for a given computational cost C = l=0 N
hl . It can be
l

shown that the optimal number of sample paths for level l is

& L r !'
−2
p X Vl
Nl = 2 Vl hl (32)
hl
l=0

Overall, Monte Carlo for a given  has computational complexity proportional to −3 whereas
2
that of the Multi-Level Monte Carlo is proportional to −2 (log ) due to reduced variance.

To illustrate the performance of Multi-Level Monte Carlo with the importance sampling we

will use the Heston-Hull-White model and the Heston-Cox-Ingersoll-Ross model. The dynamics of

these models under the risk neutral measure Q is given by

√
dSt = rt St dt + vt St dWtS (33)
√
dvt = κ (v̄ − vt ) dt + γ vt dWtv (34)

drt = λ (θ − rt ) dt + ηrtp dWtr (35)

where dWtS dWtv = ρS,v dt, dWtS dWtr = ρS,r dt and hdWtr dWtv i = 0. If p = 0 we have the

Heston-Hull-White model and if it is 0.5 we have the Heston-Cox-Ingersoll-Ross model.

18
 We compare the performance of the Multi-Level Monte Carlo with the importance sampling and

without importance sampling for European call option. We use the following parameters: κ = 2,

γ = 0.06, v0 = v̄ = 0.04, λ = 0.05, r0 = θ = 0.07, η = 0.01, S0 = 100, ρS,v = −0.3, ρS,r = 0.2,

T = 1. We set  = 0.01 and L = 8. We consider five strikes: 60, 80, 100, 120 and 140. In Table 16

we report European option prices and relative errors with respect to the semi-analytical solution

for different strikes under the Heston-Hull-White model and the Heston-Cox-Ingersoll-Ross model.

Strike 60 80 100 120 140

Heston-Hull-White
Semi closed-form (price) 44.0682 26.0077 11.5943 3.7583 0.9221
MLMC (price) 44.0918 26.0258 11.6100 3.7636 0.9219
MLMC (relative error) 0.05% 0.07% 0.13% 0.14% 0.02%
MLMC+IS (price) 44.0879 26.0256 11.6085 3.7648 0.9229
MLMC+IS (relative error) 0.04% 0.07% 0.12% 0.17% 0.08%
Heston-Cox-Ingersoll-Ross
Semi closed-form (price) 44.0686 25.9996 11.5668 3.7296 0.9071
MLMC (price) 44.0724 26.0067 11.5696 3.7320 0.9037
MLMC (relative error) 0.01% 0.03% 0.02% 0.06% 0.37%
MLMC+IS (price) 44.0769 26.0045 11.5693 3.7326 0.9041
MLMC+IS (relative error) 0.02% 0.02% 0.02% 0.08% 0.33%

Table 16: Option price and relative error with respect to the semi-analytical solution.

Given the nature of the Multi-Level Monte Carlo, we observe that it is possible to use impor-

tance sampling on all levels or some levels. We will refer to the former case as full importance

sampling. We note that the first level, where l = 0, is the coarsest, because the time step at this

level is T . The point is that variance at level l decreases as l increases because both Pl−1 and Pl

accurately approximate P as they are obtained using the same Brownian path. Therefore, as an

alternative to the full importance sampling we will consider importance sampling on the first level

only. In Figure 3 we illustrate the effective performance defined as the ratio of variance reduction

from importance sampling to speed for different strikes. Speed is defined as the ratio of computa-

tional time of the Multi-Level Monte Carlo with the importance sampling to computational time of

the Multi-Level Monte Carlo without importance sampling. We compare Multi-Level Monte Carlo

with full importance sampling, Multi-Level Monte Carlo with the importance sampling on the first

level only and Multi-Level Monte Carlo without importance sampling. The results indicate that the

Multi-Level Monte Carlo with the importance sampling is more efficient than Multi-Level Monte

Carlo without importance sampling and Multi-Level Monte Carlo with the importance sampling

on the first level only is much more efficient than both Multi-Level Monte Carlo without impor-

tance sampling and Multi-Level Monte Carlo with full importance sampling. The performance

improvement compared to the Multi-Level Monte Carlo full importance sampling comes from two

19
sources. The first one is variance reduction from importance sampling. The second one is reduced

computational time. This is due to the fact that the number of sample paths at level l which is

given by (32) depends on the variance at level l. Since importance sampling reduces the variance at

the first level, the required number of sample paths at this level is less compared to the Multi-Level

Monte Carlo without importance sampling.

Figure 3: Effective performance for different strikes.

It is also possible to use the Likelihood Ratio Method Based on Characteristic Function to

estimate the Greeks in a Multi-Level Monte Carlo. Let us consider the first derivative of the

Multi-Level Monte Carlo estimator (28) with respect to S0 .

∂
∆ = E [PL ]
∂S0
L
∂ X ∂
= E [P0 ] + E [Pl − Pl−1 ]
∂S0 ∂S0
l=1
ˆ ∞ ∂ L ˆ ∞ ∂
f (x) X ∂S0 f (x)
= e−r(T −t) P0 ∂S0 f (x)dx + e−r(T −t) (Pl − Pl−1 ) f (x)dx
0 f (x) 0 f (x)
l=1

∂
∂S0 f (x)
where f (x) is the likelihood ratio which can be obtained from the characteristic function as

in [13].

In Tables 17 and 18 we report delta and gamma computed by the Likelihood Ratio Method

Based on the Characteristic Function for European option under the Heston-Hull-White model

and the Heston-Cox-Ingersoll-Ross model.

20
 Strike 60 80 100 120 140
Heston-Hull-White
MLMC (delta) 0.9595 0.9187 0.6691 0.3227 0.1068
MLMC (variance) 8.5492 4.7461 2.5166 1.1986 0.4324
MLMC+IS (delta) 0.9621 0.9172 0.6690 0.3235 0.1058
MLMC+IS (variance) 4.3355 1.1883 0.2358 0.0869 0.0379
Heston-Cox-Ingersoll-Ross
MLMC (delta) 0.9570 0.9179 0.6698 0.3231 0.1062
MLMC (variance) 8.6794 4.8099 2.5586 1.2152 0.4403
MLMC+IS (delta) 0.9577 0.9195 0.6686 0.3224 0.1057
MLMC+IS (variance) 4.4050 1.2070 0.2374 0.0861 0.0350

Table 17: Delta and variance of delta.

On average importance sampling reduces the variance of delta 8 times.

Strike 60 80 100 120 140

Heston-Hull-White
MLMC (gamma) 0.0001 0.0042 0.0165 0.0175 0.0092
MLMC (variance) 0.0569 0.0370 0.0219 0.0119 0.0055
MLMC+IS (gamma) 0.0001 0.0042 0.0164 0.0174 0.0092
MLMC+IS (variance) 0.0264 0.0108 0.0035 0.0011 0.0005
Heston-Cox-Ingersoll-Ross
MLMC (gamma) 0.0001 0.0041 0.0165 0.0176 0.0092
MLMC (variance) 0.0591 0.0383 0.0227 0.0123 0.0057
MLMC+IS (gamma) 0.0002 0.0042 0.0164 0.0175 0.0091
MLMC+IS (variance) 0.0271 0.0112 0.0036 0.0011 0.0005

Table 18: Gamma and variance of gamma.

On average importance sampling reduces the variance of gamma 7 times.

Another interesting question is whether the skew has an impact on the effective performance

of importance sampling. To answer this question we consider two scenarios. In the first scenario

ρS,v = −0.3, and in the second scenario ρS,v = −0.7. In Figure 4 we illustrate the effective

performance in each scenario for different strikes. The results indicate that the skew has no

significant impact on the effective performance of importance sampling.

21
 Figure 4: Effective performance for different strikes.

Finally, we will use the Multi-Level Monte Carlo to price basket call on three underlying assets.

We will use the same parameters as in Section 2.1 and λ = 0.05, r0 = θ = 0.05, η = 0.01,

ρS,v = −0.3, ρS,r = 0.2, ρr,v = 0. We set  = 0.05 and L = 8. We note that a combination

of Multi-Level Monte Carlo and hybrid stochastic volatility model such as Heston-Hull-White

or Heston-Cox-Ingersoll-Ross is particularly suitable for pricing variable annuities which are in

principle long-dated basket put options. In Table 19 we report the results for basket call on three

underlying assets and in Figure 5 we illustrate the effective performance.

K
S0 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
Heston-Hull-White
MLMC 23.0163 18.1578 13.5499 9.4680 6.1793 3.6732 2.0032 1.0110
price
IS 23.0566 18.1401 13.5361 9.5040 6.1853 3.6966 2.0461 1.0250
MLMC 167.2807 163.5697 151.1191 128.0302 98.3946 68.8361 45.2237 28.9471
variance
IS 82.6510 83.0224 79.2609 68.8662 53.9843 39.5120 27.5075 18.7768
Heston-Cox-Ingersoll-Ross
MLMC 23.0473 18.1467 13.5498 9.4906 6.1786 3.7034 2.0303 1.0340
price
IS 23.0272 18.1399 13.5098 9.4610 6.1423 3.6764 2.0344 0.9912
MLMC 167.5398 163.5396 151.2621 128.0371 98.0752 69.0209 45.4317 28.6720
variance
IS 82.6687 82.9464 79.3410 68.9031 53.9411 39.3100 27.4990 18.6551
K
Table 19: Price and variance of price for basket call as a function of S0 .

On average importance sampling reduces the variance 2 times compared to the Multi-Level

Monte Carlo without importance sampling.

22
 K
Figure 5: Effective performance as a function of S0 .

4 Conclusion

We have presented an application of importance sampling with stochastic change of drift to multi-

asset options. We have illustrated the use of importance sampling with basket, best-of, worst-of,

spread, absolute and composite options as examples. Based on our results, importance sampling

reduces variance of multi-asset options by a factor of 3-13 on average across the strikes.

We have provided an extension of the Likelihood Ratio Method Based on the Characteristic

Function to multi-asset options and combined it with the importance sampling to reduce the

variance of the Greeks. Based on our results, importance sampling reduces variance of Greeks of

multi-asset options by a factor of 2 on average across the strikes.

We applied importance sampling in a Multi-Level Monte Carlo and demonstrated that applying

importance sampling on the first level significantly improves its effective performance. For Euro-

pean option the Multi-Level Monte Carlo with full importance sampling the effective performance

is on average almost 3 times better than that of the Multi-Level Monte Carlo without importance

sampling. The Multi-Level Monte Carlo with the importance sampling on the first level only the

effective performance is on average almost 19 times better than that of the Multi-Level Monte

Carlo without importance sampling. For basket option the Multi-Level Monte Carlo with the

importance sampling on the first level only the effective performance is on average almost 2 times

better than that of the Multi-Level Monte Carlo without importance sampling. We have also used

the Likelihood Ratio Method Based on the Characteristic Function to estimate the Greeks in a

23
Multi-Level Monte Carlo and combined it with the importance sampling to reduce the variance of

the Greeks. Based on our results, importance sampling reduces variance of Greeks of by a factor

of 6-8 on average across the strikes.

Finally, we have studied the impact of the skew on the effective performance of importance

sampling. Our results suggest that the skew has no significant impact on the effective performance

of importance sampling.

24
References

[1] Avramidis, A. (2002) Importance Sampling for Multimodal Functions and Applica-

tions to Pricing Exotic Options. Proceedings of the 2002 Winter Simulation Confer-

ence, pp. 1493-1501.

[2] Barraquand, J. (1995) Numerical Valuation of High Dimensional Multivariate Euro-

pean Securities. Management Science, 41(12):1882-1891.

[3] Bates D. (1996) Jumps and Stochastic Volatility: The Exchange Rate Processes Im-

plicit in Deutschemark Options. Review of Financial Studies, 9(1):69-107.

[4] Bouhari, A. (2004) Robbins-Monro Algorithms and Variance Reduction in Finance.

Journal of Computational Finance, 7(2):35-62.

[5] Bouzoubaa M. and A. Osseiran (2010) Exotic Options and Hybrids: A Guide to

Structuring, Pricing and Trading. John Wiley & Sons.

[6] Capriotti, L. (2008) Least Squares Importance Sampling for Monte Carlo Security

Pricing. Quantitative Finance, 8(5):485-497.

[7] Fouque, J.P. and T. Tullie (2002) Variance Reduction for Monte Carlo Simulation in

a Stochastic Volatility Environment. Quantitative Finance, 2(1):24-30.

[8] Giles, M.B. (2008) Multi-Level Monte Carlo Path Simulation. Operations Research,

56(3):607-617.

[9] Grzelak, L. and C. Oosterlee (2011) On the Heston Model with Stochastic Interest

Rates. SIAM Journal on Financial Mathematics, 2(1):255-286.

[10] Heston H. L. (1993) A Closed-form Solution for Options with Stochastic Volatil-

ity with Applications to Bond and Currency Options. Review of Financial Studies,

6(2):327-343.

[11] Neddermeyer, J.C. (2011) Non-Parametric Partial Importance Sampling for Financial

Derivative Pricing. Quantitative Finance, 11(8):1193-1206.

[12] Pellizzari, P. (1998) Efficient Monte Carlo Pricing of Basket Options. Working Paper.

[13] Stilger, P.S., S. Acomb and S.-H. Poon (2012) Pricing and Risk Management with

Stochastic Volatility Using Importance Sampling. Working Paper.

25
[14] Su, Y. and M.C. Fu (2002) Optimal Importance Sampling in Securities Pricing. Jour-

nal of Computational Finance, 5(4):27-50.

26