0% found this document useful (0 votes)
2 views32 pages

WBS Risk RQMC PCA

The document discusses the effectiveness of Randomized Quasi-Monte Carlo (RQMC) methods in option pricing and risk analysis, highlighting its advantages over traditional Monte Carlo methods. RQMC combines the benefits of Quasi Monte Carlo's faster convergence with the ability to compute confidence intervals, making it a powerful tool in financial applications. The paper also examines various Sobol' sequence generators, discretization schemes, and the impact of randomization techniques on computational efficiency.

Uploaded by

Julien Hok
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
2 views32 pages

WBS Risk RQMC PCA

The document discusses the effectiveness of Randomized Quasi-Monte Carlo (RQMC) methods in option pricing and risk analysis, highlighting its advantages over traditional Monte Carlo methods. RQMC combines the benefits of Quasi Monte Carlo's faster convergence with the ability to compute confidence intervals, making it a powerful tool in financial applications. The paper also examines various Sobol' sequence generators, discretization schemes, and the impact of randomization techniques on computational efficiency.

Uploaded by

Julien Hok
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

The unreasonable effectiveness of

Randomized Quasi-Monte Carlo in option


pricing and risk analysis

J. Hok1 S. Kucherenko2,3

1
Investec Bank
2
Imperial College London, London, SW7 2AZ, UK
3
BRODA Ltd., UK

julienhok@[Link] [Link]@[Link]

October 6, 2025
Disclaimer

"The opinions expressed in this presentation and on the following slides are solely
those of the presenter and not necessary those of INVESTEC. INVESTEC does
not guarantee the accuracy or reliability of the information provided herein."

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Introduction and plan

Option pricing problem can be formulated as


Z
I[f ] = E[f (x)] = f (x)dx, (1)
Hd

E[.] is the mathematical expectation.

f the payoff function, integrable in the d-dimensional unit hypercube H d = [0, 1]d .

The standard MC estimator:


N
1 X
IN [f ] = f (xi ), (2)
N i=1

a.s. IN [f ] −→ I[f ]. CLT provides confidence intervals:


σ̂N
I[f ] ∈ [IN [f ] ± c √ ] (3)
N

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Introduction and plan

→ In many financial applications, Quasi Monte Carlo (QMC) outperforms Monte


Carlo
showing faster empirical convergence rate: IN [f ] −→ I[f ] as rate O(N −α )
with 0.5 < α ≤ 1

more stable convergence.

However, QMC lacks a practical error estimate.

→ Randomized QMC (RQMC) method by randomizing the LDS points {xi }


combines the best of two methods:
it allows to compute confidence intervals around the estimated value as in
MC;

It may further improve the convergence rate

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Outline

Randomized QMC;

Different Sobol’ sequence generators;

Hyperbolic local volatility model;

Standard, Brownian Bridge, PCA discretization schemes;

Results MC, QMC, RQMC pricing and Greeks computation of Asian options;

Global Sensitivity Analysis, Effective dimensions.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Randomised QMC

Generate a set of n LDS points {Qi },


Generate a set of K randomised replication of {Qi }: {Vi } = Vik , k = 1, ..., K.

Define µ̂kn - the k − th RQMC estimator for (1):

n
1X
µ̂kn = f (Vik ), (4)
n i=1

µ̂kn are i.i.d. random variables


The RQMC sample mean
K
1 X k
µ̄n = µ̂n . (5)
K
k=1

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Computation of MC and RQMC confidence intervals

Table: MC and RQMC sample standard deviations σ, RMSE errors ε and confidence
intervals. The total number of function evaluations N = nK.

q PN q PK
1 2 1 k
σM C = (N −1) i=1 (f (Xi ) − IN [f ]) σRQM C = (K−1) k=1 (µ̂n − µ̄n )2
σ√ σRQM C
εM C = MC
N
εRQM C = √
K
IN [f ] ± zδ/2 εM C µ̄n ± zδ/2 εRQM C

K is large enough - µ̄n ∼ N (I[f ], σRQM C ).


zδ is the 1 − δ quantile of the standard normal distribution: F (zδ ) = 1 − δ.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Different Sobol’ sequence generators

The efficiency of Sobol’ LDS generator depends on direction numbers.

Figure: Sobol’ sequences with badly initialized direction numbers.1

Joe&Kuo’s generator - ’optimized’ 2D projections (maximum dim. d =21201)2 .

BRODA’s SobolSeq - additional uniformity properties3 :


1) Property A for all dimensions (maximum dim. d =131072)
2) Property A′ for adjacent dimensions.

1 P. Jackel, Monte Carlo Methods in Finance, John Wiley&Sons, 2002


2 [Link], [Link]. SIAM J. Scientific Comp., 30, 2635-2654, 2008
3 I. Sobol’, D. Asotsky, A. Kreinin, S. Kucherenko. 2011, Wilmott Journal, Nov, 64-79

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Spurious variance component

Pd
Consider z = (z1 , . . . , zd ) ∼ N (0, I), and define z̄d = √1
d i=1 zi .
N
zi = F −1 (xi ) , xi ′ s− are Sobol’ points: {xi }i=1 ⊂ (0, 1)d

Example: A terminal asset value S(T ) in the case of d time steps


1 2
 √ 
S(T ) =S0 exp r− σ T + σ ∆t (z1 + z2 + · · · + zd ) = (6)
2

  
1 2
=S0 exp r − σ T + σ T z̄d (7)
2

Here the Wiener path is sampled using the Standard (incremental) discretization
scheme.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Spurious variance component

Consider variance V (z̄d ), assuming that E (z̄d ) = 0 :

 
d d d d Xd
1 XX 1 X X
V (z̄d ) = ρij = 1+ ρij  = 1 + ρ̄d . (8)
d i=1 j=1 d i=1 i=1 j:j̸=i
d d
1XX
ρ̄d = ρij is an average correlation, ρij = E [zi zj ] . (9)
d i=1
j:j̸=i

Theoretically ρij = 0, i ̸= j, hence V (z̄d ) = 1.

Numerically ρ̄d ̸= 0 due to the presence of spurious correlations between different


dimensions of actual LDS sequences. We call ρ̄d - a "spurious variance
component"

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Comparison of Joe-Kuo and BRODA SobolSeq generators
European call:
R S0 = 100, K= 100, r 2= 0.0, σp = 0.2, T = 1y. CBS = 7.966.
Pd
C = e−rT H d max[0, (S0 exp[(r − σ2 )T + σ Td j=1 Φ−1 (uj )] − K)]du1 ...dud

BRODA - SobolSeq generator; BRODA Scrambled - Owen’s scrambling with additional


permutations; Joe-Kuo - Direction numbers of Joe&Kuo
Joe-Kuo’s generator has unacceptably high spurious variance components.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Randomization with Digital Shift

A set d-dimensional Sobol’ points {Qi } in base b = 2


m
X
Qji = j
qi,p 2−p , (10)
p=1
Pm
Generate r.n. U ∼ U [0, 1]d , U j = p=1 ujp 2−p
Randomised version with Digital shift (DS):
m
X
Vij = j
vi,p 2−p : (11)
p=1

j j
vi,p = (qi,p ⊕ ujp ) (12)

⊕ - binary addition modulo 2 (a bitwise XOR operator):


0 ⊕ 0 = 0; 1 ⊕ 1 = 0; 0 ⊕ 1 = 1; 1 ⊕ 0 = 1

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Comparison of Randomization Methods

Method Pros Cons


High CPU/memory;
Higher rate: εRQM C ∼ O(1/n(3/2−α) )
Owen’s Scrambling4 √  Permutation tree size
n faster than QMC: O 1/n(1−α) ;
Π ∼ d(bM − 1)/(b − 1)
Simple to implement No increased rate
Digital Shift
No extra memory of convegence

Owen’s scrambling is the most efficient sampling method. Accuracy is improved by


random error cancellations.
Note: BRODA’s modification of Owen’s scrambling5 reduces memory and CPU
demands.

4 A. Owen, Ann. Stat., 25(4):1541, 1997


5 E. Atanassov, S. Kucherenko, BRODA Ltd., UK, 2021
J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo
Comparison of Sobol’, Digital Shift and Owen scrambling in
2D

Fig. First 128 points of 2D Sobol sequence6 . First 16 points colored red.

Left: Sobol points: stratified 1D spacing. Aligned on 2D diagonals.


Middle: Digit scrambling: offsets 1D. 2D structure unchanged.
Right: Owen scrambling: jitters 1D/2D, preserves stratification.

6 B. Burley, Practical Hash-based Owen Scrambling. J. Comp. Graphics Tech., 9 (4) 2020
J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo
Comparison of Standard and Scrambled Sobol Sequences

(a) SobolSeq (b) SobolOwen

Video: Standard and Scrambled Sobol Sequences

Comparison of Random numbers and Sobol Sequences – YouTube, BRODA


Difference between Standard and Scrambled Sobol Sequences – YouTube, BRODA

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Time homogeneous HLV model and time discretisation
Asset prices follow:
dSt = rSt dt + σ̃(St )dWt , S0 = 1, (13)

with the risk free interest rate r and time-homogeneous hyperbolic local volatility
(HLV) model7 :
n (1 − β + β 2 ) (β − 1) p 2 o
σ̃(S) = ν St + St + β 2 (1 − St )2 − β , (14)
β β
Here ν > 0 - the level of volatility; β ∈ (0, 1] - the skew parameter.
Euler time discretization of the SDE
1
dY (t) = [r − σ 2 (Y (t))]dt + σ(Y (t))dWt , Y (0) = log(S(0)), (15)
2
σ̃(eY )
with Y (t) = ln(S(t)) and σ(Y ) = eY
.

1 2 n
Y n (ti+1 ) = Y n (ti ) + [r − σ (Y (ti ))](ti+1 − ti ) + σ(Y n (ti ))(W (ti+1 ) − W (ti )) (16)
2
with Y n (0) = log(S(0)), ∆t = T
d
, ti = i∆t, i = 0, .., d.
7 Jaeckel, P. 2008. Hyperbolic local volatility. [Link]
J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo
Time discretization schemes
T
Brownian path W = (W1 , W2 , . . . , Wd ) is normally distributed: E(W ) = 0 and
d
cov. matrix Σ = [min (ti , tj )]i,j=1
The value of options can be written as

EQ ((f (S, K)) =EQ ((f (W, K)) (17)


exp − 12 W T Σ−1 W
Z 
= P (W ) p dW (18)
Rd (2π)d det(Σ)
exp − 12 z T z
Z 
= P (Lz) p dz (19)
Rd (2π)d
Z
P LΦ−1 (x) dx,
 
= (20)
[0,1]d
Z
= f (x)dx (21)
[0,1]d

where we use the change of variable W = Lz in the second equality and for the
last equality, the mapping z = Φ−1 (x) with Φ−1 (x) the inverse of the standard
normal cumulative distribution function (applied elementwise). Here Σ = LLT .
J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo
Discretization of the Wiener process

We will consider 3 different ways of Σ = LLT decomposition:


Cholesky factorization or standard discretization:

W (ti ) = W (ti−1 ) + ∆tZi 1 ≤ i ≤ d, (22)
Zi independent standard normal variates. Cholesky factorisation of Σ = LLT .

Brownian bridge (BB) algorithm:


1. First we generate the variable at the terminal point

W (T ) = T Z1 (23)
2. Then we fill other points using already found values of W (ti )
p
W (ti ) = (1 − γ)W (tl ) + γW (tm ) + γ(1 − γ)(m − l)∆tZi , (24)
i−l
where γ = m−l
with l ≤ i ≤ m. It can be seen from equation (24) that the
variance of the stochastic part of the BB formula decreases rapidly at the
successive levels of refinement and the first few points contain most of the
variance.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Discretization of the Wiener process

PCA  
1/2
λ1
L = U Λ1/2 = (U1 |U2 | . . .)  1/2
. (25)
 
λ2
...
where:
U is an orthogonal matrix whose columns are the corresponding unit
eigenvectors (i.e., U ⊤ U = Id ),
Λ = diag(λ1 , λ2 , . . . , λd ) the diagonal matrix of eigenvalues of Σ with
λ1 ≥ λ2 ≥ · · · ≥ λd ≥ 0.

PCA selects the transformation matrix L such that the first principal
component U1 captures the maximum variance of W ; the second component
U2 captures the maximum remaining variance, conditional on U1 ; and
subsequent components (U3 , . . . , Ud ) are determined iteratively.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


MC simulation of Asian option price and Greeks
Geometric average Asian call option payoff:

PA = max(S̄ − K, 0), (26)


Qn 1
with S̄ = ( i=1 Si ) d , where Si = S(ti ), ti = i Td , 1 ≤ i ≤ d.
Price:

" N
#
−rT Q −rT 1 X (l)
AC(T, K) = e E [PA ] ≈ ACN (T, K) = e max(S̄ − K, 0) ,
N
l=1
(27)
where S̄ (l) is S̄ at the simulated price paths l.
Sensitivities ∆ and Γ defined and approximated as :

∂AC(T, K) ACN (T, K, S(0) + ϵs ) − ACN (T, K, S(0) − ϵs )


∆ := ≈ (28)
∂S(0) 2ϵs
∂ 2 AC(T, K) ACN (T, K, S(0) + ϵs ) + ACN (T, K, S(0) − ϵs ) − 2ACN (T, K, S(0))
Γ := ≈ (29)
∂S02 ϵ2s

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Numerical results:Confidence intervals for prices and
Greeks

Recall:

Table: MC and RQMC sample standard deviations σ, RMSE errors ε and confidence
intervals. The total number of function evaluations N = nK.
q PN q PK
1 2 1 k
σM C = (N −1) i=1 (f (xi ) − µ̂N ) σRQM C = (K−1) k=1 (µ̂n − µ̄n )2
σ√ σRQM C
εM C = MC
N
εRQM C = √
K

Table: εM C , εRQM C−BB and εRQM C−P CA of price estimations.

ITM ATM OTM


εM C 3.26 10−2 2.2 10−2 1.24 10−3
εRQM C−BB 3.67 10−4 6.09 10−4 4.04 10−4
εRQM C−P CA 3.2 10−4 3.11 10−3 3.05 10−4
εM C /εRQM C−BB 89 36 3
εM C /εRQM C−P CA 102 71 4

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Numerical results:Confidence intervals for prices and
Greeks

Table: εM C , εRQM C−BB and εRQM C−P CA of Deltas estimations.

ITM ATM OTM


εM C 3.29 10−4 2.43 10−3 4.09 10−4
εRQM C−BB 3.23 10−5 5.98 10−4 1.41 10−4
εRQM C−P CA 3.25 10−5 4.91 10−4 4.05 10−4
εM C /εRQM C−BB 10 4 3
εM C /εRQM C−P CA 10 49 10

Table: εM C , εRQM C−BB and εRQM C−P CA of Gammas estimations.

ITM ATM OTM


εM C 4.41 10−5 9.08 10−4 2.24 10−4
εRQM C−BB 3.72 10−5 6.79 10−4 1.82 10−4
εRQM C−P CA 2.20 10−5 1.33 10−4 4.62 10−5
εM C /εRQM C−BB 1 1 1
εM C /εRQM C−P CA 2 7 5

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Numerical results: performance analysis
Power law integration error approximation:
C
εn ∼ . (30)

Use the RMSE below to approximate the rate of convergence


v
u K  2
u1 X (k)
εn = t V − Vn , (31)
K
k=1

V is the reference Price or Greek values.


ITM ATM OTM
RQMC BB (Price) 1.0 0.8 0.77
RQMC PCA (Price) 1.0 0.91 0.97
RQMC BB (Delta) 0.7 0.64 0.58
RQMC PCA (Delta) 0.88 0.73 0.87
RQMC BB (Gamma) 0.58 0.6 0.55
RQMC PCA (Gamma) 0.54 0.86 0.71

Table: Extracted α in RQMC BB and RQMC PCA

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


performance analysis for ATM price (a), delta (b) and
gamma (c)

(c) (d)

(e)
J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo
ANOVA decomposition and Sobol’ Sensitivity Indices

Let’s consider a square integrable f (x) in Hd = [0, 1]d . f has an unique ANOVA
decomposition:

n
X XX
f (x) = f0 + fi (xi ) + fij (xi , xj ) + . . . + f1,2,...d (x1 , x2 , . . . , xd ) (32)
i=1 i j>i

if
Z
f0 = f (x) dx, (33)
[0,1]d

and for all k = 1, 2, . . . , s


Z 1
fi1 ,...,is (xi1 , ....., xis )dxik = 0. (34)
0

Each ANOVA term fi1 ,...,is (xi1 , ....., xis ) is a function of a unique subset of
variables from x and the terms are orthogonal.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


ANOVA decomposition and Sobol’ Sensitivity Indices

Variance decomposition:
d
X d X
X d
2
σ = σi2 + 2
σij 2
+ .... + σ12...d . (35)
i=1 i=1 i<j

Here σ 2 is the total variance,


Z 1
2
σi1 ,...,is = fi21 ,...,is (xi1 , ....., xis )dxi1 ...dxis (36)
0

are called partial variances.


Sobol’ Sensitivity Indices (SI):
n
X X X
1= Si + Sij + Sijl + . . . + S1,2,...,d (37)
i=1 i<j i<j<l

σi21 ,...,is
where Si1 ,...,is := σ2 .

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Back to the option pricing problem

EQ [f (S, K)] = EQ [f (W, K)] (38)


exp − 12 W T Σ−1 W
Z 
= P (W ) p dW (39)
Rd (2π)d det(Σ)
exp − 1 z T z
Z 
= P (Lz) p 2 dz (40)
Rd (2π)d
Z
P L Φ−1 (x) dx
 
= (41)
[0,1] d
Z
= f (x) dx. (42)
[0,1]d

We transformed the option price into an expectation over independent uniforms.


Which dimensions of x impact the price most?

The key is to analyze the integrand f (x):


1 By computing its Sobol’ indices, we identify the most influential drivers.
2 By calculating its mean effective dimension, we gain insight into the overall
complexity of the problem.
J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo
Mean effective dimension

Definition: The mean dimension of f is


X
dˆ = |u|Su . (43)
u⊆1:d

with |u| the cardinality of a set of variables u


Pd
Theorem (Owen 2011): dˆ = i=1 Sitot
with the total Sobol’ index for input variable xi defined as
X Dytot
Sitot := Su := (44)
σ2
u⊆{1,...,d}
i∈u

It satisfies 1 ⩽ dˆ ⩽ d. Various examples show that QMC outperforms MC


integration if the integrand f (x) has dˆ < 3

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Mean effective dimension
By denoting x = (y, z), we have
Theorem (Sobol 2001): Subset’s total variance Dytot is equal to
Z
1
Dytot = [f (y, z) − f (y ′ , z)]2 dxdy ′ (45)
2

Let ξj = (ηj , ζj ), ξj′ = (ηj′ , ζj′ ) independent random points uniformly ditributed
in H d with j = 1, ..., N . Then
N
1 X 2 P
f (ξj ) − f (ηj′ , ζj ) − → Dytot . (46)
2N i=1

Table: Mean effective dimension dˆ for Asian option price and Greeks, OTM

BB PCA Standard
Price 2.21 1.12 5.8
Delta 2.10 1.1 4.28
Gamma 4.10 1.62 26.43

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Sobol SI for Asian option price

Figure: Total Sobol SI w.r.t variable xi for Standard (SD), BB and PCA schemes

The initial coordinates of Sobol’ sequences are much better distributed than the later
high dimensional coordinates. The Brownian bridge and PCA discretizations use low well
distributed coordinates from each d-dimensional LDS vector point to determine most of
the structure of a path and reserves the later coordinates to fill in fine details.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


Conclusions

RQMC with Owen’s scrambling outperforms MC and QMC, enabling reliable


confidence intervals and faster convergence.

Effective dimension reduction (PCA, Brownian Bridge) is the key driver of


efficiency, yielding substantial accuracy gains in both pricing and Greeks.

PCA-based RQMC consistently delivers the lowest RMSE, outperforming all


other methods and making difficult sensitivities (especially Gamma) feasible
and reliable.

Effective dimension, not nominal dimension, determines success for


high-dimensional, path-dependent financial simulations.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo


References

Hok, J and Kuchenrenko, S. Pricing and Risk Analysis in Hyperbolic Local


Volatility Model with Quasi Monte Carlo. Wilmott, 2021(113):62-9
Kuchenrenko, S and Hok, J. The importance of being scrambled:
supercharged Quasi Monte Carlo. Journal of Risk, 26(1):1-20,2023.
Scoleri S, Bianchetti M, Kucherenko S. Application of Quasi Monte Carol and
Global Sensitivity Analysis to Option Pricing and Greeks: Finite Differences
vs. AAD. Wilmott. 2021(116):66-83
Sobol I., Kucherenko S. Global Sensitivity Indices for Nonlinear Mathematical
Models. Review, Wilmott, 2005(1):56-61.
Sobol, I.M., Asotsky, D., Kreinin, A. and Kucherenko, S. Construction and
Comparison of High-Dimensional Sobol Generators, Wilmott, 2011(56),
64-79.
Sobol, I.M. Global sensitivity indices for nonlinear mathematical models and
their Monte Carlo estimates, Mathematics and Computers in Simulation,
2001(55), 271-280.

J. Hok and S. Kucherenko The unreasonable effectiveness of Randomized Quasi-Monte Carlo

You might also like