Professional Documents
Culture Documents
Article
Markovian Approximation of the Rough Bergomi Model for
Monte Carlo Option Pricing
Qinwen Zhu 1 , Grégoire Loeper 2 , Wen Chen 3 and Nicolas Langrené 3, *
Abstract: The recently developed rough Bergomi (rBergomi) model is a rough fractional stochastic
volatility (RFSV) model which can generate a more realistic term structure of at-the-money volatility
skews compared with other RFSV models. However, its non-Markovianity brings mathematical and
computational challenges for model calibration and simulation. To overcome these difficulties, we
show that the rBergomi model can be well-approximated by the forward-variance Bergomi model with
wisely chosen weights and mean-reversion speed parameters (aBergomi), which has the Markovian
property. We establish an explicit bound on the L2-error between the respective kernels of these two
models, which is explicitly controlled by the number of terms in the aBergomi model. We establish
and describe the affine structure of the rBergomi model, and show the convergence of the affine
structure of the aBergomi model to the one of the rBergomi model. We demonstrate the efficiency and
accuracy of our method by implementing a classical Markovian Monte Carlo simulation scheme for
Citation: Zhu, Q.; Loeper, G.; Chen,
W.; Langrené, N. Markovian
the aBergomi model, which we compare to the hybrid scheme of the rBergomi model.
Approximation of the Rough Bergomi
Model for Monte Carlo Option Keywords: rough fractional stochastic volatility; forward variance model; markovian representation;
Pricing. Mathematics 2021, 9, 528. volatility skew; Volterra integral; rough heston; hybrid scheme; sum of ornstein-uhlenbeck processes
https://doi.org/10.3390/math9050528
Inspired by the technique by Abi Jaber and El Euch [18], Gatheral and Keller-Ressel [19],
and Harms and Stefanovits [20], in which the authors designed a multi-factor stochastic
volatility model with Markovian structure to approximate the rough Heston model, we
establish an analogous multi-factor affine structure for the rBergomi model. Indeed, the
Volterra kernel of the rBergomi model corresponds to a superposition of infinitely many
Ornstein-Uhlenbeck (OU) processes with different speeds of mean reversion. Truncating
this infinite sum into a finite sum of OU processes yields an approximation of the rBergomi
model which is a classical Markovian multi-factor Bergomi model. We refer to this affine,
Markovian approximation of the rBergomi model as the aBergomi model. We prove the
existence and uniqueness of the solution to this aBergomi model, and show that its affine
structure converges to the one of the rBergomi model. Finally, we implement a Monte Carlo
scheme for the aBergomi model, and compare it to the hybrid scheme of the rBergomi
model (Bennedsen et al. [21]). Our numerical tests demonstrate that using 20 exponential
terms in the aBergomi kernel is sufficient to obtain accurate implied volatility curvatures
while remaining computationally efficient.
The idea to interpret the conventional two-factor Bergomi model as a Markovian
approximation of the rBergomi model was originally briefly suggested by Bayer et al. [1]
(p. 892). Our work explores and expands upon this intuition by testing the number
of factors to use in the Bergomi model and establishing their respective parameters for
best approximation of the rBergomi model. For comprehensiveness, one can mention the
alternative Markovian approximation proposed in Carr and Itkin [22] of the rough volatility
version of the mean-reverting lognormal volatility model of Sepp [23], Langrené et al. [24],
based on a closed-form vol-of-vol expansion for solving the pricing PDE arising from the
use of the Dobrić-Ojeda process (Dobrić and Ojeda [25]) to approximate the fractional
Brownian motion.
Compared to alternative pricing methods for the rBergomi model, the main advantage
of our proposed Markovian approximation approach is that it does not require pricing
methods specifically designed for rough volatility models; instead, classical Markovian
pricing methods can be used for both vanilla and exotic options. In practice, the Monte
Carlo pricing method is the method of choice for the aBergomi model in view of the number
of terms needed for good accuracy. The computational cost of simulating our proposed
aBergomi model is proportional to the number of time-steps N, which makes it an interest-
ing alternative to the approximate O( N log N ) hybrid scheme of Bennedsen et al. [21] and
the exact O( N 3 ) covariance-based scheme of Bayer et al. [1], Bayer et al. [10]. The main
downside is that the approximation of the rBergomi power kernel by a sum of exponential
terms introduces some error, for which we provide an explicit bound in the L2 sense. In
particular, as in the case of the Riemann-sum scheme of Bennedsen et al. [21], the Fourier-
based scheme of Benth et al. [26], or the approximation by a Dobrić-Ojeda process in Carr
and Itkin [22], a truncation of the power kernel singularity at s = t cannot be avoided.
The paper is organized as follows. In Section 2, we introduce the Bergomi and rBer-
gomi models and discuss their respective ATM volatility skews. The rBergomi model is
closely related to the RFSV model introduced in Alòs et al. [27], for which the ATM volatil-
1
ity is proved to be equivalent to the power T H − 2 for short maturity using the Malliavin
technique, a result confirmed in Fukasawa [3] using a martingale expansion approach.
We prove in this section that a similar result holds for the rBergomi model, while this
does not hold for the Bergomi model (Equation (9)). We also establish the quasi-affine
structure of the rough Bergomi model. Section 3 is dedicated to the approximation of the
rough Bergomi model by a multi-factor Bergomi model, both theoretically and numerically.
Finally, Section 4 compares numerical simulations of the rBergomi model with our approxi-
mated Bergomi (aBergomi) model with a finite number of terms, showing the effectiveness
of our approximation.
Mathematics 2021, 9, 528 3 of 21
1 √
dXt = − Vt dt + Vt dWt , (1)
2
where Vt ≥ 0 is the instantaneous spot variance process. Let ξ tu , u ≥ t be the instantaneous
forward variance for date u observed at time t; in particular, ξ tt = Vt corresponds to the
spot variance.
Bayer et al. [1] proposed the so-called rough Bergomi model where the forward
variance follows √
dξ tu = ξ tu η 2α + 1(u − t)α dBt , u ≥ t, (2)
where W and B have correlation ρ, α , H − 12 ∈ (− 12 , 0) is a negative exponent depending
on the Hurst exponent H ∈ (0, 12 ) of the underlying fractional Brownian motion, and
η is a positive parameter depending on H. The definition of the rBergomi model is
summarized below:
dX = − 1 V dt + √V dW ,
t t t t
2√ (3)
u
dξ t = ξ tu η 2α + 1(u − t)α dBt ,
1
where α = H − 2 ∈ (− 12 , 0), and dhW, Bit = ρdt.
Definition 2. The two-factor Bergomi model (Bergomi [28], Bergomi [29]) is defined by:
1 √
dXt = − Vt dt + Vt dWtS ,
2 (4)
dξ u = ξ u αθ ω (1 − θ )e−κX (u−t) dW X + θe−κY (u−t) dW Y ,
t t t t
with
dhW S , W X it = ρSX dt,
dhW S , W Y it = ρSY dt,
dhW X , W Y it = ρ XY dt,
where ξ tt = Vt = ω is the lognormal volatility of the instantaneous variance under the normalizing
1
factor αθ = (1 − θ )2 + 2ρ XY θ (1 − θ ) + θ 2 − 2 and θ is a mixing parameter of the short-term
factor driven by W X and the long-term factor driven by W Y (κ X > κY ).
Assumption 1. Without loss of generality, we assume throughout the paper that the initial forward
variance curve ξ 0u , u ≥ 0 is flat. This simplification is common in the rBergomi literature; see, for
example, Bayer et al. [1], Bayer et al. [6], and Bayer et al. [15]. We henceforth use the notation ξ 0
for the constant initial forward variance curve.
Mathematics 2021, 9, 528 4 of 21
dX = − 1 V dt + √V dW ,
t t t t
2 (5)
u u
dξ t = λ(t, u, ξ t )dBt ,
where Xt = ln(St ) is the log-spot, Vt is the instantaneous spot variance, ξ tu is the instanta-
neous forward variance for date u observed at time t, and λ = (λ1 , · · · , λd ) is the volatility
of forward instantaneous variances which takes values in Rd where d is the dimension of
the Brownian motion B. Note that in this formulation, the covariance between spot and
variance is modelled through the first component of λ, see Bergomi and Guyon [30] for
more details.
One can derive the following second-order expression (w.r.t. volatility of volatility)
for the Black-Scholes implied volatility:
ε2 Xξ 2
ε
σ̂TATM = σ̂TVS 1 + C Xξ + 12 C − v ( v + 4 ) C ξξ
+ 4v ( v − 4 ) C µ
,
4v 32v3
ε2 µ
ε Xξ
ST = σ̂TVS C + 4C v − 3 ( C Xξ 2
) ,
2v2 8v3
ε2 h i
CT = σ̂TVS 4 4C µ v + C ξξ v − 6 C Xξ 2 ,
8v
r
q RT s
0 ξ 0 ds
RT s VS v
where v = 0 ξ 0 ds is the total variance to expiration T, σ̂T = T = T is the
effective volatility. Here, u
ξ 0 = ξ 0 for any u ≥ 0 under Assumption 1, which means that
√
v = ξ 0 T and σ̂TVS = ξ 0 .
From Bergomi and Guyon [30], we can derive the following second-order expansion
for the autocorrelations C Xξ , C ξξ , C µ :
Xξ RT RT R T R T E[dXs dξ su ]
• Ct (ξ ) = t ds s duµ(s, u, ξ ) = t ds s du ds is the doubly integrated spot-
variance covariance function,
Xξ R T R T E[dXs dξ 0u ]
• C Xξ = C0 (ξ 0 ) = 0 ds s du ds .
0
h i
T T T T T T E dξ su dξ su
Ct (ξ ) = t ds s du s du0 ν(s, u, u0 , ξ ) = t ds s du s u0
ξξ R R R R R R
• ds is the triply
integrated variance/variance covariance function,
0
h i
T T T E dξ 0u dξ 0u
ξξ 0
ξξ
R R R
• C = C0 (ξ 0 ) = 0 dt s du s du ds .
RT RT
µ Xξ
• Ct (ξ ) = t ds s duµ(s, u, ξ )∂ξ 0u Cs (ξ ) is the double time-integral of the instance
Xξ
spot variance covariance function times the sensitivity of Ct (ξ ) with respect to
instantaneous forward variances,
µ R T R T E[dXs dξ 0u ] Xξ
• C µ = C0 (ξ 0 ) = 0 ds s du ds ∂ξ 0u Cs (ξ ) ,
Mathematics 2021, 9, 528 5 of 21
∂ 1
ψ( T ) , σ (k, T ) ∼ T H− 2 . (8)
∂k BS k =0
Proof. We first explicit the autocorrelationq functional in the rBergomi model. Using the
E[dXt dξ tu ] √
fact that dt = ρη 2α + 1(u − t) ξ tt ξ tu , the autocorrelation functionals C Xξ and C ξξ
α
are given by
E[dXs dξ 0u ]
Z T Z T
Xξ
C = ds du
0 s ds
√ Z Tq Z T
= ρη 2α + 1 ξ 0s ds ξ 0u (u − s)α du + O ε3 ,
0 s
Z T Z T Z T u u0
ξξ 0 E[ dξ 0 dξ 0 ]
C = ds du du
0 s s ds
Z T Z T Z T
0
= ds du0 η 2 (2α + 1)(u − s)α (u0 − s)α ξ 0u ξ 0u
du
0 s s
Z T Z T 2
= η 2 (2α + 1) ds ξ 0u (u − s)α du + O ε4 .
0 0
we obtain
E[dXs dξ 0u ] Xξ
Z T Z T
Cµ = ds du ∂ξ 0u Cs (ξ )
0 s dt
Z Tq Z T
=ρ2 η 2 (2α + 1) ξ 0s ds (u − s)α du
0 s
"Z p uZ #
uq ξ0 T t
u
× t
ξ 0 ξ 0 (u − t) dt +
α
ξ 0 (t − u) dt + O ε4 .
α
s 2 u
1 xξ VS 1
σBS (k, T ) = σ̂TVS + C σ̂T ε + 2 C Xξ σ̂TVS εk
4v
2v 3
VS 1 k 3
= σ̂T + + CH ρξ 02 T H + 2 σ̂TVS ε
4v 2v2
ξ0 T k 1
CH ρT H − 2 ε.
p
= ξ0 + +
4 2
Thus, the ATM volatility skew generated by the rBergomi model satisfies (8), which is
consistent with empirical evidence (see for example, Gatheral et al. [4]).
Remark 1. Besides the rBergomi model, there exist other fractional volatility models which also
satisfy Equation (8); see, for example, Fukasawa [31] (subsection 3.3).
Theorem 2. In the two-factor Bergomi model, the ATM volatility skew satisfies
C1 κ X T − 1 + e−κX T C2 κY T − 1 + e−κY T
ψ( T ) ∼ + . (9)
T2 T2
W S = W1,
q
W X = ρSX W 1 + 1 − ρ2SX W 2 ,
q q
W Y = ρSY W 1 + χ 1 − ρ2SY W 2 + (1 − χ2 )(1 − ρ2SY )W 3 ,
ρ −ρSX ρSY
where W 1 , W 2 , W 3 are three independent Brownian motions and χ , √ XY 2 √ 2
. Thus,
1−ρSX 1−ρSY
the volatilities of variance λ = (λ1 , λ2 , λ3 ) in the general formulation (5) can be written as:
h i
λ1 (t, u, ξ ) = αθ ωξ 0u (1 − θ )ρSX e−κX (u−t) + θρSY e−κY (u−t) ,
q q
u 2 −κ X ( u − t ) 2 −κY ( u − t )
λ2 (t, u, ξ ) = αθ ωξ 0 (1 − θ ) 1 − ρSX e + θχ 1 − ρSY e ,
q
λ3 (t, u, ξ ) = αθ ωξ 0u θ (1 − χ2 ) 1 − ρ2SY e−κY (u−t) ,
or equivalently:
λi (t, u, ξ ) = αθ ωξ 0u ωiX e−κX (u−t) + ωiY e−κY (u−t) ,
where q
>
(ωiX )i=1,2,3 , (1 − θ )ρSX , (1 − θ ) 1 − ρ2SX , 0 ,
q q
>
(ωiY )i=1,2,3 , θρSY , θχ 1 − ρ2SY , θ (1 − χ2 )(1 − ρ2SY ) .
Mathematics 2021, 9, 528 7 of 21
Once again using Assumption 1 and the autocorrelations provided by Bergomi and
Guyon [30], we obtain
3
C Xξ =αθ ωξ 02 T 2 (ω1X J (κ X T ) + ω1Y J (κY T )),
C ξξ =α2θ ωξ 02 T 3 (ω0 + ωX I(κ X T ) + ωY I(κY T ) + ωXX I(2κ X T ) + ωYY I(2κY T ) + ωXY I((κ X + κY ) T )),
where
3 2 3 3
ωiX ω ωiX ωiX ω ω ωiX ω
ω0 = ∑ κX T
+ iY
κY T
, ω X = −2 ∑
κ T κX T
+ iY , ωY = −2 ∑ iY
κY T κ T κX T
+ iY ,
κY T
i =1 i =1 X i =1 Y
3 ω2 3 ω2 3
ωiX ωiY
ωXX = ∑ κ2 iXT2 , ωYY = ∑ κ2 iYT2 , ωXY = 2 ∑
κ κ T2
,
i =1 X i =1 Y i =1 X Y
and
1 − e−z z − 1 + e−z 1 − e−z − ze−z J (z) − K(z)
I(z) = , J (z) = , K( z ) = , H(z) = .
z z2 z2 z
µ µ
Similarly, we have C µ = α2θ ω 2 ξ 02 T 3 C1 + C2 , with the coefficients
µ 1 2 1 2 J (κY T ) − J (κ X T )
C1 = ω1X H(κ X T ) + ω1Y H(κY T ) − ω1X ω1Y ,
2 2 (κ X + κY ) T
00
J (κ X T ) + ωY00 J (κY T ) + ωXX
00 00 00
µ
C2 = ωX J (2κ X T ) + ωYY J (2κY T ) + ωXY J ((κ X + κY ) T ),
and
2
ω1X ω ω ω2
00
ωX = + 1X 1Y , ωY00 = κY1YT + ω1X ω1Y
κY T ,
κX T κY T
ω2 2 ω ω ω ω
ωXX00
= − 1X , 00 = − ω1Y ,
ωYY
YT
ωXY00
= − 1X 1Y − 1X 1Y .
κX T κ κX T κY T
−κ X T − κY T
Since C Xξ ∼ T 2 C1 · κX T(−κ1+Te)2 + C2 · κY T(−κ1+Te)2 and C1 , C2 are constants, we
X Y
can derive the term structure of the ATM volatility skew as in Equation (9) with the first
order in ε.
However, this result derived for the Bergomi model by the Bergomi-Guyon expan-
sion [30] is inconsistent with empirical evidence; see, for example, Bayer et al. [1]. This
suggests that the power-law kernel of the forward variance curve in the rBergomi model
Mathematics 2021, 9, 528 8 of 21
will lead to more realistic and accurate pricing and hedging results than the exponential
kernel of the forward variance curve in the Bergomi model.
Definition 3. An Ornstein-Uhlenbeck (OU) process Ytx is the solution of the following stochastic
differential equation (SDE):
dYtx = x ( a − Ytx )dt + σdBt , (10)
where x > 0 is the mean-reversion speed, a > 0 is the mean-reversion level, and Bs is a standard
Brownian motion. Its strong solution is explicitly given by
Z t
Ytx = Y0 + σ e− x(t−s) dBs . (11)
0
a , Y0 , (12)
√
σ , η 2α + 1, (13)
where η and α come from Definition 1 of the rBergomi model (see Bayer et al. [1]).
Definition 4. Without loss of generality, we define, for H < 21 , the sigma-finite measure µ(dx ) on
(0, ∞) as
dx
µ(dx ) = 1 .
x 2 Γ( 21 − H )
+ H
which can be recognised as the power-law kernel in the Volterra-type integral. Con-
RtR∞
sequently, we have σ X̃t = 0 0 σe− x(t−s) µ(dx )dBs , and using Fubini’s stochastic the-
R∞Rt
orem, see Protter [32], we obtain σ X̃t = 0 0 σe− x(t−s) dBs µ(dx ). From Definition 3,
R t − x (t−s)
where 0 σe dBs = Ytx − Y0 , we obtain the Markovian representation given by
Equation (14).
R∞
Proof. From Fubini’s stochastic theorem, 0 Ytx µ(dx ) is Gaussian under the filtration Fs
for 0 ≤ s ≤ t, with mean
Z ∞ Z ∞
E x
Yt µ(dx ) Fs = Ysx e−(t−s) x µ(dx ).
0 0
Thus,
∞
Z Z ∞ Z ∞
x 1 x x
E exp Yt µ(dx ) Fs = exp Var Yt µ(dx ) Fs + E Yt µ(dx ) Fs
0 2 0 0
Z ∞ 2 Z ∞
!
σ2 t−s
Z
−sx x −(t−s) x
= exp e µ(dx ) ds + Ys e µ(dx ) .
2 0 0 0
where Xt is the log stock price, ξ 0 is the initial flat forward variance curve, and W, B are two
Brownian motions with correlation dhW, Bit = ρdt and ρ ∈ [−1, 1]. Our aim is now to write
the log stock price Xt in a quasi-affine form as the first coordinate of an infinite-dimensional
affine process. To do so, we introduce the following symmetric non-negative tensor:
n o
L 1 ( µ ) ⊗ s L 1 ( µ ) = y ⊗2 : y ∈ L 1 ( µ ) ⊂ L 1 ( µ ) ⊗2 ⊂ L 1 ( µ ⊗2 ) ,
where we used the notation y⊗2 , y ⊗ y. Let Πt = (i ⊗ 1)(Ytx )⊗2 ∈ iL1 (µ) ⊗s L1 (µ), where
R∞ 2 R ∞
i is the imaginary unit (i ×i = −1). The relation 0 Ytx µ(dx ) = 0 (i ⊗ 1)(Ytx )⊗2 µ⊗2 (dx )
holds. Therefore, the log stock price dynamics can be written as
R∞ ⊗2
0 Πt µ (dx )
p 1 R ∞ Y x µ(dx)
dXt = ξ 0 · E dWt − E
4 0 t
2
R∞ ⊗ 2
√ R
0 Πt µ (dx ) − η t2α+1
2
p ξ 0 ∞ Y x µ(dx) − η2 t2α+1
= ξ0 e 4 e 4 dWt − e 0 t e 2 dt,
2
where E is the Doléans-Dade stochastic exponential.
where
Z t−s Z ∞ 2 !
1 −ux
Φ1 , − log 1 − 2 e µ(dx ) du , (16)
2 0 0
R ∞ −(t−s)x ⊗2 ⊗2
0 Πs e µ (dx )
Φ2 , R t−s R ∞ . (17)
−ux µ ( dx ) 2 du
σ2 − 2σ2 0 0 e
R∞ x
0 Yt µ ( dx )
Proof. From Fubini’s stochastic theorem, qR
t−s R ∞ −ux
is Gaussian under the
σ 0 ( 0 e µ(dx ))2 du
filtration Fs for 0 ≤ s ≤ t, with conditional mean
R∞ x R ∞ x −(t−s)x
0 Yt µ ( dx ) 0 Ys e µ(dx )
E q R t−s R ∞ 2 Fs =
q R t−s R ∞
−ux µ ( dx ) du −ux µ ( dx ) 2 du
σ 0 0 e σ 0 0 e
Thus, the Formulas (16) and (17) for Φ1 and Φ2 follow from the characteristic function
of the noncentral χ2 distribution, which concludes the proof.
Proof. This corollary follows from Theorem 5 which exhibits that the rBergomi model has an
exponential-affine dependence on x; hence, the model is Markovian in each dimension.
and the n-term Bergomi model with the same Brownian motion in the variance process can
be represented by √
dSt = St Vt dWt ,
!
Z t n
V (18)
log ξ 0 = 0 ∑ αi e
t −κ i ( t − s )
dBs ,
i =1
we can view the rBergomi model as a continuous infinite-term Bergomi model under
the measure µ(·), in which the mean-reversion speed x has been integrated from 0 to
∞, with respect to the Brownian motion Bs . We can therefore approximate the rBergomi
model by an n-term exponential kernel Kexp = ∑in=1 αi e−κi (t−s) instead of the power kernel
√
Kpow = 2α + 1(t − s)α of the Volterra process in the rBergomi model. R∞
Following Equation (18), after approximating the exponential kernel K (τ ) = 0 e− xτ µ(dx )
n
by the kernel K n (τ ) = ∑in=1 αin e−τxi , we can rewrite the aBergomi model (18) as follows:
q
dStn = Stn Vtn dW,
n n
Vt
log
ξ 0
= ∑ αin Vtn,i , (19)
i =1
√
dVtn,i = − xin ( a − Vtn )dt + σdBt a = Y0 , σ = η 2α + 1,
where (αin )1≤i≤n are positive weights, ( xin )1≤i≤n are mean-reverting speeds, and hW, Bit =
ρdt, with initial conditions S0n = S0 = 1 and V0n,i = V0 = 0.
Theorem 6. Under the conditions of the model (19), there exists a unique, strong, non-negative
solution V n to Equation (20).
Proof. Øksendal and Zhang [33] imply that there exists a unique, strong, non-negative
solution V n to Equation (20) under the conditions of the model (19).
Then, the strong existence and uniqueness of (Sn , V n ) follows, along with its Marko-
vianity w.r.t. the spot price Sn and the factors V n,i for i ∈ {1, · · · , n}.
Theorem 7. There exist weights (αin )1≤i≤n > 0, mean reversion speeds ( xin )1≤i≤n > 0, and a
constant C depending on H and T only such that
−4H
kK n − K k2,T ≤ Cn 5 ,
Theorem 8. The process V n (Equation (20)) has the following affine structure
n
( ! )
σ2 n n 1 e−(t−s) xi n
2 i∑
n
E[Vt | Fs ] = ξ 0 exp αi − n,i n −(t−s) xin
+ ∑ Vs αi e .
=1
xin xin i =1
2 i∑
= ξ 0 exp αin
− + ∑ Vs αi en,i n −(t−s) xin
.
=1
xin xin i =1
Then, we describe the so-called hybrid scheme and introduce an algorithm to approxi-
mate the rBergomi model by the aBergomi model.
√ Z t
η 2 2α+1
Vt = ξ 0 exp η 2α + 1 (t − s) dBs − t
α
.
0 2
√ Rt
To simulate the Volterra-type integral X̃ = 2α + 1 0 (t − s)α dBs , we apply the hybrid
scheme proposed in Bennedsen et al. [21], which approximates the kernel function of the
Brownian semi-stationary processes by a Wiener integral of the power function at t = s
and a Riemann sum elsewhere. Let Ω, F , (Ft )t∈R , P be a filtered probability space
which supports a standard Brownian motion W = (Wt )t∈R . We consider a Brownian
semi-stationary process (Bss):
Z t
X̄t = g(t − s)σs dWs t ∈ R , (22)
−∞
where σ = (σt )t∈R is an (Ft )t∈R -predictable process which captures the stochastic volatility
of X̄ and g : (0, ∞) → [0, ∞) is a Borel-measurable kernel function. We assume that
E σt2 < ∞ for all t ∈ R and the process is covariance-stationary, namely,
E[σs ] = E[σt ],
cov(σs , σt ) = cov σ0 , σ|s−t| , s, t ∈ R.
Mathematics 2021, 9, 528 13 of 21
g( x ) = x α L g ( x ), x ∈ (0, 1],
In order to implement the hybrid scheme to the rBergomi model, we need to introduce
a particular class of non-stationary processes, namely, truncated Brownian semi-stationary
(tBss) processes,
Z t
X̃t = g(t − s)σs dWs t ≥ 0, (23)
0
where the kernel function g(t), the volatility process σs , and the driving Brownian motion
Ws are as defined in the definition of Bss processes. X̃t can also be seen as the truncated
stochastic integral at 0 of the Bss process X̄t . Equation (23) is integrable since g(t) is
differentiable on (0, ∞).
Now, we can discretise Equation (23) in time. Let N be the total number of time-steps,
∆t = T/N be the time-step size, and t0 = 0 ≤ . . . ≤ t j = j∆t ≤ . . . ≤ t N = T be a time
grid on the interval [0, T ].
According to Bennedsen et al. [21], the observations X̃tNj , j = 0, 1, · · · , N can be
computed via (κ = 1 case)
j
X̃tNj = L g (∆t)σjN−1 WjN−1,1 + ∑ g(bk∗ ∆t)σjN−k W̄jN−k (24)
k =1
Then,
∆t
Z t α
j
WjN−1,1
α
= t j − s dWs ≈ (Wt j − Wt j−1 )
t j −1 2
Z t
j +1
W̄jN = dWs = Wt j+1 − Wt j
tj
σjN = σt j .
Mathematics 2021, 9, 528 14 of 21
In the rBergomi model, σti = σ is a constant for i = 1, 2, ..., N defined in Equation (13).
When simulating X̃ti , we need to perform a matrix multiplication, the computational
complexity of which is of order O( N 2 ) when using the conventional matrix multiplication
algorithm. However, multiplying a lower triangular Toeplitz matrix can be regarded as a
discrete convolution which can be evaluated efficiently by fast Fourier transform. Therefore,
the computational complexity can be reduced to O( N log N ). The algorithm to simulate
the Volterra process X̃ is described in Algorithm 1 below. Then, we can use a standard
Euler scheme to simulate the price (St1 , St2 , · · · , St N ), as shown in Algorithm 2.
Table 1 reports the parameters used for our numerical experiments, which are the
same as in Bayer et al. [1] and Bennedsen et al. [21]. Recall from the definition of α that the
chosen value α = −0.43 corresponds to the Hurst exponent H = 0.07. Such small values of
H are indeed consistent with empirical experiments, and one can refer to the recent works
Forde et al. [36] and Gerhold [37] about the behaviour of the rBergomi model for small H.
ξ0 η α
0.026 1.9 –0.43
number of terms is greater than two. The two-term Bergomi model (4) that we used to
approximate the rBergomi model is given by
√
dSt = St Vt dWt ,
(26)
dξ st = ηξ st α1 e−κ1 (t−s) + α2 e−κ2 (t−s) dBs ,
where the two parameters κ1 and κ2 come from the exponential kernel Kexp , and Ys1 and Ys2
are two OU processes. Hence, the process yts can be written as a driftless Gaussian process
as follows:
dyts = α1 e−κ1 (t−s) dBs + α2 e−κ2 (t−s) dBs ,
and its qquadratic variation is given by hdyt , dyt is = ς2 (t − s)ds where
ς(u) = α21 e−2κ1 u + α22 e−2κ2 u + 2α1 α2 e−(κ1 +κ2 )u . The forward variation process ξ st can
be written as dξ st = ηst dyts . Thus, the solution
t
of the forward variation process is ξ s =
2
ξ 0 f t (s, yts ), where f t (s, y) = exp ηy − 2 χ(s, t) and
η
Z t
χ(s, t) = ς2 (u)du
t−s
Z t
= α21 e−2κ1 u + α22 e−2κ2 u + 2α1 α2 e−(κ1 +κ2 )u du
t−s
1 − e−2κ1 s 1 − e−2κ2 s 1 − e−(κ1 +κ2 )s
= α21 e−κ1 (t−s) + α22 e−2κ2 (t−s) + 2α1 α2 e−(κ1 +κ2 )(t−s) . (28)
2κ1 2κ2 κ1 + κ2
η2
Recall that Vt = ξ tt = ξ 0 exp ηytt − 2 χ(t, t) and χ(t, t) ' t2α+1 when s → t when
s→t
the number of terms n is large enough.
Using the approximation by the Bergomi model, we consider the parameters
{ i i }(i=1,2,··· ,n) in the exponential kernel Kexp = ∑in=1 αi e−κi (t−s) on s ∈ [0, t). Note
α , κ
that when s → t, the power kernel Kpow → ∞ while Kexp is finite. To compute the ap-
proximation numerically, we need to truncate the kernel Kexp . To do so, we can use the
scipy.optimize module in Python or the nlinfit function in MATLAB for the nonlinear re-
gression of the parameters {αi , κi }(i=1,2,··· ,n) and the simulated price {St }. We exemplify the
truncation of Kexp by letting s ∈ [0, T − ∆t], the truncated parameter θ = T − N
T
= T − ∆t,
and let T = 1.
We define the integral Itrunc on the truncated region [0, θt) and apply the scaling
property of Brownian motion as follows:
r
n Z θt n Z t
θ
∑ αi ∑ αi
T −κ i ( t − s )
e−κi (1− T )s dBs .
θ
Itrunc = e dBs =
i =1 0 i =1
T 0
After scaling Bs , the process ys has to remain driftless Gaussian and satisfy
ys = ∑in=1 αi e−κi (1− T )s Ysi , where dYsi = κi (1 − Tθ )Ysi ds + dBs , Y0i = 0. Then, the pro-
θ
In practice, the truncation of the rBergomi power kernel means that, as is the case for the
Riemann-sum scheme of Bennedsen et al. [21], this scheme is able to capture the shape of
the implied volatility smile, but not its level. A multiplication factor is used in Algorithm 3
for each time-step to correct for this phenomenon. In practice, these factors can be estimated
using another calibrated scheme, or more simply, from quoted option prices.
4. Simulation Results
In this section, we compare the simulated volatilities of the rBergomi and aBergomi
models. To demonstrate the approximation’s accuracy and efficiency, we investigate the
Mean Absolute Error (MAE) of simulated results for different number of terms and number
of time-steps in numerical tests.
Figure 1 displays the power kernel Kpow in the rBergomi model and the Kexp kernel of
the 20-term aBergomi model with T = 1 and N = 100. This figure suggests that this Kexp
obtained by nonlinear regression is sufficiently accurate, with a MAE of 4.05806 × 10−6 .
Figure 1. The power kernel Kpow in the rBergomi model and the exponential Kexp in the 20-term
aBergomi model when T = 1 and N = 100.
The volatility smiles in Figure 2 are obtained by simulating the rBergomi model as
described in Section 3.3, and the aBergomi model as described in Section 3.4 using the
multiplication factors reported in Table 2. From Figure 2, we note that the at-the-money
Mathematics 2021, 9, 528 17 of 21
Figure 2. Volatility smiles for rBergomi and 20-term aBergomi models with T = 1 using 20,000 Monte Carlo paths.
We compute the MAE of the implied volatility approximation with different numbers
of terms in the aBergomi model and different time-steps in Figure 3, and compare the
pricing speed in Table 3. As expected, the higher the number of terms in the aBergomi
model, the lower the MAE for all time-steps, but the difference between the models
decreases when the number of time-steps decreases. Another expected result is that the
computational time increases with both the number of terms and time-steps. The number
of terms and time-step combinations provide a good trade-off between speed and accuracy,
such as the 20-term aBergomi model with 100 time-steps and 20,000 Monte Carlo paths.
Figure 3. MAE of the implied volatility smiles of the aBergomi model with respect to the number of
time-steps, for four different numbers of terms (10, 15, 20, and 25), using 20,000 Monte Carlo paths.
Mathematics 2021, 9, 528 18 of 21
Table 3. Runtime (in s) of the rBergomi model and the aBergomi model for different time-steps with T = 1 and 20,000 Monte Carlo paths.
Time-Steps rBergomi 10-Term aBergomi 15-Term aBergomi 20-Term aBergomi 25-Term aBergomi
50 0.4081 0.0994 0.1392 0.1721 0.2164
100 0.5001 0.2369 0.3130 0.4024 0.4543
150 0.5602 0.3650 0.4499 0.5589 0.6869
200 0.5861 0.4257 0.5908 0.7258 0.8727
5. Conclusions
In this paper, we proved the power-law behavior of the ATM volatility skew as
time to maturity goes to zero of the rough Bergomi model (rBergomi), and proposed an
approximate Bergomi (aBergomi) model with a finite number of forward variance terms
to approximate the rBergomi model. The approximation enables the adoption of classical
pricing methods, while keeping the fractional feature of the model. We theoretically prove
the convergence of the aBergomi model towards the rBergomi model when the number of
terms is large enough, and verify this convergence numerically. We numerically compared
the fast hybrid scheme for the rBergomi model to the Euler scheme for the aBergomi model.
The numerical simulation results illustrate the accuracy and efficiency of the approximation.
The parameters of the aBergomi model are numerically obtained by nonlinear regression on
the power-law kernel of the rBergomi model. Other alternative calibration and truncation
methods are worth investigating for future research, as well as further comparisons on
more complex options.
Author Contributions: Conceptualization, Q.Z. and G.L.; Formal analysis, Q.Z. and W.C.; Inves-
tigation, Q.Z.; Methodology, Q.Z. and G.L.; Software, Q.Z. and N.L.; Supervision, G.L. and W.C.;
Validation, Q.Z., G.L., W.C. and N.L.; Visualization, Q.Z.; Writing—original draft, Q.Z.; Writing—
review & editing, Q.Z., G.L., W.C. and N.L. All authors have read and agreed to the published version
of the manuscript.
Funding: The Centre for Quantitative Finance and Investment Strategies has been supported by
BNP Paribas.
Acknowledgments: The authors thank the three anonymous reviewers for their useful comments
which helped us to improve the article significantly.
Conflicts of Interest: The authors declare no conflict of interest.
Appendix A
Appendix A.1. Proof of Theorem 7
This subsection is devoted to the proof of Theorem 7.
Proof. Let pin 0≤i≤n be auxiliary mean reversion speeds such that pin−1 ≤ xin ≤ pin for
R∞
i ≤ {1, · · · , n} and p0n = 0. Recall that K (τ ) = 0 e− xτ µ(dx ). We have
n Z ∞
∑ αin e−xi τ −
n
kK n − K k2,T = e− xτ µ(dx )
i =1 0
2,T
Z ∞ Z pn (A1)
n
µ(dx ) + ∑
− x (·) n i
≤ e αin e− xi (·) − e − x (·)
µ(dx ) .
0 2,T
i =1 pin−1
2,T
The first term on the RHS of the inequality (A1) can be estimated as below:
Z ∞ Z ∞r
− x (·) 1 − e−2xT ( pnn )− H
e µ(dx ) = µ(dx ) ≤ √ .
pnn 2,T pnn 2x 2HΓ 12 − H
Mathematics 2021, 9, 528 19 of 21
For the second term, applying a second-order Taylor expansion of the exponential
function
x2 ( x − u )3
Z x
ex = 1 + x + + du
2 0 6
R pn 14
i x 4 µ ( dx )
p n p n
for t ∈ [0, T ], choosing αin = pni µ(dx ) and xin = Ri−pn1
R
, yields
i −1 i µ(dx )
pn
i −1
! Z n !
(− xt)2
Z pn
n i (− xin t)2 pi
αin e− xi t − e − xt
µ(dx ) = αin 1 + (− xin t) + − n 1 + (− xt) + µ(dx )
pin−1 2 p i −1 2
3 ! Z n Z
Z xn t
i xin t − u pi xt ( xt − u )3
n
+ αi du − duµ(dx )
0 6 pin−1 0 6
2
Z pn
i n − xin t − (− xt)2
= n ( xt − xi t) + µ(dx )
p i −1 2
Z pn
t2 i
≤ ( x − xin )2 µ(dx )
2 pin−1
since
Z pn
(Z )
i xin t
( xin t − u)3 Z xt
( xt − u)3
du − du µ(dx )
pin−1 0 6 0 6
Z pn
( Z 1 n
)
i n ( xi t − xin ts)3 Z 1
( xt − xts)3 u
= n xi t ds − xt ds µ(dx ) , s =
p i −1 0 6 0 6 xt
Z pn Z 1 3 Z 1 3
(1 − s ) (1 − s )
i
= n ( xin t)4 ds − ( xt)4 ds µ(dx )
p i −1 0 6 0 6
Z pn n
(1 − s )3
Z 1 o
i
= t 4
ds ( xin )4 − ( x )4 µ(dx )
0 6 pin−1
4
R p n
i
xµ(dx )
Z pn
(1 − s )3
Z 1
pin−1
i
4 4
= t ds − ( x ) µ(dx )
n
6 R pi µ(dx )
0 pin−1
n
p i −1
=0.
Hence,
n Z pn 5 Z pn
T2 n
∑
n
≤ √ ∑
i i
αin e− xi (·) − e− x(·) µ(dx ) ( x − xin )2 µ(dx ).
i =1 pin−1 2 5 i =1 pin−1
2,T
Thus, the convergence of K n depends on the weights αi and mean reversions xi . Let
pin = iπn for each i ∈ {1, · · · , n} and πn > 0. We have
5 −H 1
n Z pn Z pn
n πn2 n 2 −H
∑
i
(x − xin )2 µ(dx ) ≤ πn2 µ(dx ) =
n
i =1 p i −1 0 1
2 − H Γ 12 − H
We can also proceed to get the explicit expressions of αin and xin as follows:
3 3
1 1
(iπn ) 2 − H −[(i −1)πn ] 2 − H 1 − 2H (iπn ) 2 − H − [(i − 1)πn ] 2 − H
αin = , xin = · .
( 12 − H )Γ( 21 − H ) 3 − 2H (iπ ) 21 − H − [(i − 1)π ] 12 − H
n n
Mathematics 2021, 9, 528 20 of 21
5 −H 1 1
Since pnn = nπn → ∞ , we have πn2 n 2 − H → 0 as n → +∞ when πn < n− 6 ,
5
1 T 2H 1
kK n − K k2,T ≤ √ ( pnn )− H + √ ( pnn ) 2 − H πn2
2HΓ 12 − H 10 12 − H
5
1 T 2H 1 5 −H
= √ n− H πn− H + √ n 2 − H πn2
1 1
2HΓ 2 − H 10 2 − H
5
= ax − H + bx 2 − H (A2)
5 0 3
Let x = πn , y = ax − H + bx 2 − H and y = − aHx − H −1 + b 5
2 − H x 2 − H = 0; solving
5
2 1
for x, we obtain x 5 = aH
5 , where a = n− H and b = √ T 21 H n 2 −H
b( 2 −H ) 10( 2 − H )
√ 2
5
1√ 2
5
√ 2
5
n− H H 10 12 − H n− 2 10 12 − H n − 1
5 10 1
2 − H
x = πn = 5 1 = 5
=
5
T 2 Hn 2 − H 25 − H T 2 25 − H T 2 −H
√ 2
− 51 10( 21 − H ) 5
n
When πn = T 5 , the RHS of Equation (A2) attains its minimum and
( 2 −H )
√ − 5 H
−4H 1 10( 21 − H ) 2 5
kK n − K k2,T ≤ Cn 5 where C = √ TH 5 −H
2
5 −H is a constant.
2HΓ( 12 − H ) 2 2
References
1. Bayer, C.; Friz, P.; Gatheral, J. Pricing under rough volatility. Quant. Financ. 2016, 16, 887–904. [CrossRef]
2. Forde, M.; Zhang, H. Asymptotics for rough stochastic volatility models. SIAM J. Financ. Math. 1993, 8, 114–145. [CrossRef]
3. Fukasawa, M. Short-time at-the-money skew and rough fractional volatility. Quant. Financ. 2017, 17, 189–198. [CrossRef]
4. Gatheral, J.; Jaisson, T.; Rosenbaum, M. Volatility is rough. Quant. Financ. 2018, 18, 933–949. [CrossRef]
5. Jusselin, P.; Rosenbaum, M. No-arbitrage implies power-law market impact and rough volatility. Math. Financ. 2020, 30, 1309–1336.
[CrossRef]
6. Bayer, C.; Ben Hammouda, C.; Tempone, R. Hierarchical adaptive sparse grids and quasi-Monte Carlo for option pricing under
the rough Bergomi model. Quant. Financ. 2020, 20, 1457–1473. [CrossRef]
7. Jacquier, A.; Martini, C.; Muguruza, A. On VIX futures in the rough Bergomi model. Quant. Financ. 2018, 18, 45–61. [CrossRef]
8. McCrickerd, R.; Pakkanen, M. Turbocharging Monte Carlo pricing for the rough Bergomi model. Quant. Financ. 2018, 18,
1877–1886. [CrossRef]
9. El Euch, O.; Fukasawa, M.; Gatheral, J.; Rosenbaum, M. Short-term at-the-money asymptotics under stochastic volatility models.
SIAM J. Financ. Math. 2019, 10, 491–511. [CrossRef]
10. Bayer, C.; Friz, P.; Gulisashvili, A.; Horvath, B.; Stemper, B. Short-time near-the-money skew in rough fractional volatility models.
Quant. Financ. 2019, 19, 779–798. [CrossRef]
11. Friz, P.K.; Gassiat, P.; Pigato, P. Short dated smile under rough volatility: Asymptotics and numerics. arXiv 2020, arXiv:2009.08814.
12. Tomas, A. Pricing of Asian Options in the Rough Bergomi Model. Ph.D Thesis, Technische Universität Wien, Wien, Austria, 2018.
13. Bayer, C.; Tempone, R.; Wolfers, S. Pricing American options by exercise rate optimization. Quant. Financ. 2020, 20, 1749–1760.
[CrossRef]
14. Bayer, C.; Qiu, J.; Yao, Y. Pricing options under rough volatility with backward SPDEs. arXiv 2020, arXiv:2008.01241.
15. Bayer, C.; Horvath, B.; Muguruza, A.; Stemper, B.; Tomas, M. On deep calibration of (rough) stochastic volatility models. arXiv
2019, arXiv:1908.08806.
16. Zeron, M.; Ruiz, I. Tensoring volatility calibration Calibration of the rough Bergomi volatility model via Chebyshev Tensors.
arXiv 2020, arXiv:2012.07440.
17. Horvath, B.; Muguruza, A.; Tomas, M. Deep learning volatility: A deep neural network perspective on pricing and calibration in
(rough) volatility models. Quant. Financ. 2021, 21, 11–27. [CrossRef]
18. Abi Jaber, E.; El Euch, O. Multifactor approximation of rough volatility models. SIAM J. Financ. Math. 2019, 10, 309–349.
[CrossRef]
19. Gatheral, J.; Keller-Ressel, M. Affine forward variance models. Financ. Stochastics 2019, 23, 501–533. [CrossRef]
Mathematics 2021, 9, 528 21 of 21
20. Harms, P.; Stefanovits, D. Affine representations of fractional processes with applications in mathematical finance. Stoch. Process.
Their Appl. 2019, 129, 1185–1228. [CrossRef]
21. Bennedsen, M.; Lunde, A.; Pakkanen, M. Hybrid scheme for Brownian semistationary processes. Financ. Stochastics 2017, 21,
931–965. [CrossRef]
22. Carr, P.; Itkin, A. ADOL: Markovian approximation of a rough lognormal model. Risk Mag. 2019, 32. Available online:
https://www.risk.net/cutting-edge/banking/7209816/adol-markovian-approximation-of-a-rough-lognormal-model (accessed
on 21 February 2021).
23. Sepp, A. Log-Normal Stochastic Volatility Model: Affine Decomposition of Moment Generating Function and Pricing of Vanilla
Options. 2016. Available online: https://dx.doi.org/10.2139/ssrn.2522425 (accessed on 21 February 2021).
24. Langrené, N.; Lee, G.; Zhu, Z. Switching to nonaffine stochastic volatility: A closed-form expansion for the Inverse Gamma
model. Int. J. Theor. Appl. Financ. 2016, 19, 1–37. [CrossRef]
25. Dobrić, V.; Ojeda, F. Fractional Brownian fields, duality, and martingales. In High Dimensional Probability; Institute of Mathematical
Statistics Lecture Notes—Monograph Series; Institute of Mathematical Statistics: Beachwood, OH, USA, 2006; Volume 51,
pp. 77–95.
26. Benth, F.E.; Eyjolfsson, H.; Veraart, A. Approximating Lévy semistationary processes via Fourier methods in the context of power
markets. SIAM J. Financ. Math. 2014, 5, 71–98. [CrossRef]
27. Alòs, E.; León, J.; Vives, J. On the short-time behavior of the implied volatility for jump-diffusion models with stochastic volatility.
Financ. Stochastics 2007, 11, 571–589. [CrossRef]
28. Bergomi, L. Smile dynamics II. Risk Mag. 2005, 18. Available online: https://www.risk.net/derivatives/equity-derivatives/1500
225/smile-dynamics-ii (accessed on 21 February 2021). [CrossRef]
29. Bergomi, L. Smile dynamics IV. Risk Mag. 2009, 22. Available online: https://www.risk.net/derivatives/equity-derivatives/1564
129/smile-dynamics-iv (accessed on 21 February 2021). [CrossRef]
30. Bergomi, L.; Guyon, J. Stochastic volatility’s orderly smiles. Risk Mag. 2012, 25. Available online: https://www.risk.net/
derivatives/2171452/stochastic-volatilitys-orderly-smiles (accessed on 21 February 2021).
31. Fukasawa, M. Asymptotic analysis for stochastic volatility: Martingale expansion. Financ. Stochastics 2011, 15, 635–654. [CrossRef]
32. Protter, P. Stochastic differential equations. In Stochastic Integration and Differential Equations; Springer: Berlin/Heidelberg,
Germany, 2005; pp. 249–361.
33. Øksendal, B.; Zhang, T.-S. The stochastic Volterra equation. Barcelona Seminar on Stochastic Analysis; Birkhäuser: Basel, Switzerland,
1993; pp. 168–202.
34. Carmona, P.; Coutin, L.; Montseny, G. Approximation of some Gaussian processes. Stat. Inference Stoch. Process. 2000, 3, 161–171.
[CrossRef]
35. Muravlev, A. Representation of a fractional Brownian motion in terms of an infinite-dimensional Ornstein-Uhlenbeck process.
Russ. Math. Surv. 2011, 66, 439–441. [CrossRef]
36. Forde, M.; Fukasawa, M.; Gerhold, S.; Smith, B. The Rough Bergomi Model as H→0—Skew Flattening/Blow up and Non-Gaussian
Rough Volatility; King’s College London Working Paper; King’s College: London, UK, 2020.
37. Gerhold, S. Asymptotic analysis of a double integral occurring in the rough Bergomi model. Math. Commun. 2020, 25, 171–184.