Deep Learning Volatility

A deep neural network perspective on pricing and calibration in

(rough) volatility models∗
arXiv:1901.09647v2 [[Link]] 22 Aug 2019

Blanka Horvath
Department of Mathematics, King’s College London
[Link]@[Link], bhorvath@[Link]
Aitor Muguruza
Department of Mathematics, Imperial College London & NATIXIS
[Link]-gonzalez15@[Link]
Mehdi Tomas
CMAP & LadHyx, École Polytechnique
[Link]@[Link]

August 26, 2019

Abstract
We present a neural network based calibration method that performs the calibration task within
a few milliseconds for the full implied volatility surface. The framework is consistently applica-
ble throughout a range of volatility models—including the rough volatility family—and a range
of derivative contracts. The aim of neural networks in this work is an off-line approximation
of complex pricing functions, which are difficult to represent or time-consuming to evaluate by
other means. We highlight how this perspective opens new horizons for quantitative modelling:
The calibration bottleneck posed by a slow pricing of derivative contracts is lifted. This brings
several numerical pricers and model families (such as rough volatility models) within the scope
of applicability in industry practice. The form in which information from available data is ex-
tracted and stored influences network performance. This approach is inspired by representing
the implied volatility and option prices as a collection of pixels. In a number of applications
we demonstrate the prowess of this modelling approach regarding accuracy, speed, robustness
and generality and also its potentials towards model recognition.

2010 Mathematics Subject Classification: 60G15, 60G22, 91G20, 91G60, 91B25

Keywords: Rough volatility, volatility modelling, Volterra process, machine learning, accurate
price approximation, calibration, model assessment, Monte Carlo
∗ The authors are grateful to Jim Gatheral, Ben Wood, Antoine Savine and Ryan McCrickerd for stimulating

discussions. MT conducted research within the Econophysique et Systèmes Complexes under the aegis of the
Fondation du Risque, a joint initiative by the Fondation de l’École Polytechnique, l’École Polytechnique, Capital
Fund Management. MT also gratefully acknowledges the financial support of the ERC 679836 Staqamof and the
Chair Analytics and Models for Regulation.

1
Contents
1 Introduction 3

2 A neural network perspective on model calibration 5

2.1 A brief reminder of some (rough) models considered . . . . . . . . . . . . . . . . . . 5
2.2 Calibration bottlenecks in volatility modelling and deep calibration . . . . . . . . . . 7
2.3 Challenges in neural network approximations of pricing functionals . . . . . . . . . . 8
2.4 Motivations for our choice of training setup and features of neural networks as ap-
proximators of pricing functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.1 Reasons for the choice of grid-based implicit training . . . . . . . . . . . . . . 10
2.4.2 Some relevant properties of deep neural networks as functional approximators 10

3 Pricing and calibration with neural networks: Optimising network and training 12
3.1 The objective function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.1 For vanillas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.2 Some exotic payoffs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Network architecture and training . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.1 Network architecture of the implied volatility map approximation . . . . . . . 15
3.2.2 Training of the approximation network . . . . . . . . . . . . . . . . . . . . . . 16
3.3 The calibration step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

4 Numerical experiments 17
4.1 Numerical accuracy and speed of the price approximation for vanillas . . . . . . . . . 18
4.1.1 Neural network price approximation in (rough) Bergomi models with piece-
wise constant forward variance curve . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Calibration speed and accuracy for implied volatility surfaces . . . . . . . . . . . . . 21
4.2.1 A calibration experiment with simulated data in (rough) Bergomi models
with piecewise constant forward variances . . . . . . . . . . . . . . . . . . . . 22
4.2.2 Calibration in the rough Bergomi model with historical data . . . . . . . . . 23
4.3 Numerical experiments with barrier options in the rough Bergomi model . . . . . . . 26

5 Conclusions and outlook: “best-fit” models 26

2
1 Introduction
Approximation methods for option prices came in all shapes and forms in the past decades and
they have been extensively studied in the literature and well-understood by risk managers. Clearly,
the applicability of any given option pricing method (Fourier pricing, PDE methods, asymptotic
methods, Monte Carlo, . . . etc.) depends on the regularity properties of the particular stochastic
model at hand. Therefore, tractability of stochastic models has been one of the most decisive quali-
ties in determining their popularity. In fact it is often a more important quality than the modelling
accuracy itself: It was the (almost instantaneous) SABR asymptotic formula that helped SABR
become the benchmark model in fixed income desks, and similarly the convenience of Fourier pric-
ing is largely responsible for the popularity of the Heston model, despite the well-known hiccups of
these models. Needless to say that it is the very same reason (the concise Black Scholes formula)
that still makes the Black-Scholes model attractive for calculations even after many generations of
more realistic and more accurate stochastic market models have been developed. On the other end
of the spectrum are rough volatility models, for which (despite a plethora of modelling advantages,
see [7, 23, 28] to name a few) the necessity to rely on relatively slow Monte Carlo based pricing
methods creates a major bottleneck in calibration, which has proven to be a main limiting factor
with respect to industrial applications. This dichotomy can become a headache in situations when
we have to weigh up the objectives of accurate pricing vs. fast calibration against one another in
the choice of our pricing model. In this work we explore the possibilities provided by the availability
of an algorithm that –for a choice of model parameters– directly outputs the corresponding vanilla
option prices (as the Black-Scholes formula does) for a large range of maturities and strikes of a
given model.

In fact, the idea of mapping model parameters to shapes of the implied volatility surface directly is
not new. The stochastic volatility inspired SSVI, eSSVI surfaces (see [27, 29, 35]) do just that: A
given set of parameters is translated directly to different shapes of (arbitrage-free) implied volatil-
ity surfaces, bypassing the step of specifying any stochastic dynamics for the underlying asset. For
stochastic models that admit asymptotic expansions, such direct mappings from model parameters
to (approximations of) the implied volatility surface in certain asymptotic regimes can be obtained
(one example is the famous SABR formula). Such asymptotic formulae are typically limited to
certain asymptotic regimes along the surface by their very nature. Complementary to asymptotic
expansions we explore here a direct (approximative) mapping from different parameter combinations
of stochastic models to different shapes of implied volatility surface for intermediate regimes. It’s
appeal is that it combines the advantages of direct parametric volatility surfaces (of the SSVI fam-
ily) with the possibility to link volatility surfaces to the stochastic dynamics of the underlying asset.

In this paper we apply deep neural networks (merely) as powerful high-dimensional functional
approximators to approximate the multidimensional pricing functionals from model parameters to
option prices. The advantage of doing so via deep neural networks over standard (fixed-basis) func-
tional approximations is that deep neural networks are agnostic to the approximation basis [32].
This makes them robustly applicable to several stochastic models consistently. Our objective in
doing so is to move the (often time-consuming) numerical approximation of the pricing functional
into an off-line preprocessing step. This preprocessing amounts to storing the approximative di-
rect pricing functional in form of the network weights after a supervised training procedure: Using
available numerical approximations of option prices as ground truth (in a stochastic model of our

3
choice), we train a neural network to learn an accurate approximation of the pricing functional.
After training, the network outputs–for any choice of model parameters–the corresponding implied
volatilities within milliseconds for a large range of maturities and strikes along the whole surface.
Furthermore, we show that this procedure generalises well for unseen parameter combinations: the
accuracy of price approximation of our neural network pricing functional on out-of-sample data is
within the same range as the accuracy of the original numerical approximation used for training.
The accuracy of this direct pricing map is demonstrated in our numerical experiments.

One of the striking advantages of this approach is that it speeds up the (on-line) calibration
Rough Volatility models to the realm of just a few milliseconds. There have been several recent
contributions on neural network calibrations of stochastic models [9, 13, 36, 49, 17]. Clearly, much
depends on the finesse of the particular network design with respect to the performance of these
networks. One contribution of this paper is to achieve a fast and accurate calibration of the rough
Bergomi model of [7] with a general forward variance curve (approximated by piecewise constant
function). To demonstrate this, we first perform calibration experiments on simulated data and
show calibration accuracy in controlled experiments. To demonstrate the speed and prowess of the
approach we then calibrate the rough Bergomi model to historical data and display the evolution
of parameters on a dataset consisting of 10 years of SPX data.
Another advantage of our modelling choice is that by its very design it can be applied to portfolios
including multiple strikes and maturities at the same time which is the first step towards their
application as hedging instruments. See for example Buehler et al. [13] a motivation.

The paper is organised as follows: In Section 2 we present a neural network perspective on model
calibration and recall stochastic models that are considered in later sections. In this section we also
formalise our objectives about the accuracy and speed of neural network approximation of pricing
functionals and the basic ideas of our training. Section 2.4.2 recalls some background on neural
networks as functional approximators and some aspects of neural network training that influenced
the setup of our network architecture and training design. In Section 3.2 we describe the network
architecture and training of the price approximation network such as the calibration methods we
consider. In Section 4 we present numerical experiments of price approximations of vanilla and
some exotic options, calibration to synthetic data and to historical data. Section 5 points to further
potential applications and outlook to future work.
Numerical experiments and codes are provided on GitHub: NN-StochVol-Calibrations , where an
accessible code demo of our results can be downloaded. We also created a library of stochastic
models where this approach is demonstrated to work well.

4
2 A neural network perspective on model calibration
In plain words, any calibration procedure is meant to fix the model parameters such that the model
is as close as possible to the observed reality. In a financial context, our model represents the un-
derlying (stocks, indices, volatility, etc.) and we are interested in calibrating the model to available
market prices of financial contracts based on this underlying.
Let us first formalise this by setting the notation M := M(θ)θ∈Θ which represents an abstract
model with parameters θ in the set Θ ⊂ Rn , for some n ∈ N. Thus the model M(θ) (stochastic or
parametric) and the corresponding prices of financial contracts are fully specified by the choice of
the parameter combination θ ∈ Θ. Furthermore, we introduce a pricing map P : M(θ, ζ) → Rm ,
where ζ : (C(R) → Rm ), m ∈ N denote the financial products we aim to price, such as vanilla
options for (a set of) given maturities and strikes. Let us denote the observed market data corre-
sponding to the contracts, by P M KT (ζ) ∈ Rm , m ∈ N.

Parameter Calibration: The parameter configuration θ̂ solves an (ideal) δ-calibration problem

for a model M(Θ) for the conditions P M KT (ζ) if

θ̂ = argmin δ(P (M(θ), ζ), P M KT (ζ)) (1)

θ∈Θ

where δ(·, ·) is a suitable choice of metric for the financial contract ζ at hand.
For most financial models however (1) represents an idealised form of the calibration problem as
in practice there rarely exists an analytical formula for the option price P (M(θ), ζ) and for the
vast majority of financial models it needs to be computed by some numerical approximation scheme.

Approximate Parameter Calibration We say that the parameter configuration θ̂ ∈ Θ solves

an approximate δ-calibration problem for the model M(Θ) for the conditions P M KT (ζ) if

θ̂ = argmin δ(Pe(M(θ), ζ), P M KT (ζ)) (2)

θ∈Θ

where δ(·, ·) is a suitably chosen metric and Pe is a numerical approximation of the pricing map P .

In the remainder of this paper it is this second type of calibration problem that we will be concerned
with: In our numerical experiments (Section 4) we consider the numerical approximation Pe of the
pricing map P as the benchmark (available truth) for generating synthetic training samples in the
training a neural network to approximate pricing maps. Clearly, the better the original numerical
approximations, the better the network approximation will be. In a separate work we will illuminate
this perspective with a Bayesian analysis of the calibration procedure.

2.1 A brief reminder of some (rough) models considered

We would like to emphasize that our methodology can in principle be applied to any (classical or
rough) volatility model. From the classical Black Scholes or Heston models to the rough Bergomi
model of [7], also to large class of rough volatility models (see Horvath, Jacquier and Muguruza [41]
for a general setup). In fact the methodology is not limited to stochastic models, also parametric
models of implied volatility could be used for generating training samples of abstract models, but
we have not pursued this direction further.

5
The Rough Bergomi model
In the abstract model framework, the rough Bergomi model is represented by MrBergomi (ΘrBergomi ),
with parameters θ = (ξ0 , ν, ρ, H) ∈ ΘrBergomi . On a given filtered probability space (Ω, F, (Ft )t≥0 , P)
the model corresponds to the following system
1 p
dXt = − Vt dt + Vt dWt , for t > 0, X0 = 0,
2

√ Z t  (3)
H−1/2
Vt = ξ0 (t)E 2Hν (t − s) dZs , for t > 0, V0 = v0 > 0
0

where H ∈ (0, 1) denotes the Hurst parameter, ν > 0 , E(·) the stochastic exponential [21], and
ξ0 (·) > 0 denotes the initial forward variance curve (see [11, Section 6]), and W and Z are correlated
standard Brownian motions with correlation parameter ρ ∈ [−1, 1]. To fit the model parameters
into our abstract model framework ΘrBergomi ⊂ Rn for some n ∈ N, the initial forward variance
curve ξ0 (·) > 0 is approximated by a piecewise constant function in our numerical experiments
in Sections 4.1.1, and 4.2.1. We refer the reader to Horvath, Jacquier and Muguruza [41] for one
general setting of rough volatility models and their numerical simulation.

The Heston model

The Heston model, appearing in our numerical experiments of Section 5 is described by the system
p
dSt = Vt St dWt for t > 0, S0 = s0
p (4)
dVt = a(b − Vt )dt + v Vt dZt for t > 0, V0 = v0

with W and Z Brownian motions with correlation parameter ρ ∈ [−1, 1], a, b, v > 0 and 2ab > v 2 .
In our framework it is denoted by MHeston (θ) with θ = (a, b, v, ρ) ∈ ΘHeston ⊂ R4 . The Heston
model is considered in our numerical experiments in Section 5. It was also considered by [9, 18] in
different neural network contexts.

The Bergomi model

In the general n-factor Bergomi model, the volatility is expressed as
n Z t
!
X
i


Vt = ξ0 (t)E ηi exp −κi (t − s)dWs for t > 0, V0 = v0 > 0, (5)
i=1 0

where η1 , . . . , ηn > 0 and (W 1 , . . . , W n ) is an n-dimensional correlated Brownian motion, E(·) the

stochastic exponential [21], and ξ0 (·) > 0 denotes the initial forward variance curve, see [11, Section
6] for details. In this work we consider the Bergomi model for n = 1, 2 in Section 4. Henceforth,
M1F Bergomi (ξ0 , β, η, ρ) represents the 1 Factor Bergomi model, corresponding to the following
dynamics:
1 p
dXt = − Vt dt + Vt dWt for t > 0, X0 = 0
2
 Z t  (6)
Vt = ξ0 (t)E η exp(−β(t − s))dZs for t > 0, V0 = v0 > 0,
0

6
where ν > 0, and W and Z are correlated standard Brownian motions with correlation parameter
ρ ∈ [−1, 1]. To fit the model parameters into our abstract model framework Θ1F Bergomi ⊂ Rn ,
for some n ∈ N, the initial forward variance curve ξ0 (·) > 0 is approximated in our numerical
experiments by a piecewise constant function in Sections 4.1.1, and 4.2.1.

The SABR model

The stochastic alpha beta rho model of Hagan et al. [33, 34] is denoted in our setting as MSABR (α, β, ρ)
and is defined as
dSt = Vt Stβ dWt for t > 0, S0 = s0 .
(7)
dVt = αVt dZt for t > 0, V0 = v0

where v0 , s0 , α > 0 and β ∈ [0, 1]. The SABR model is considered by McGhee in [55] in a neural
network context (see also Section 2.3).

2.2 Calibration bottlenecks in volatility modelling and deep calibration

Whenever for a stochastic volatility model the numerical approximate calibration procedures (2) are
computationally slow, a bottleneck in calibration time can deem the model of limited applicability
for industrial production irrespective of other desirable features the model might have. This is the
case in particular for the family rough volatility models, where the rough fractional Brownian mo-
tion in the volatility dynamics rules out usual Markovian pricing methods such as finite differences.
So far such calibration bottlenecks have been a major limiting factor for the class of rough volatility
models, whose overwhelming modelling advantages have been explored and highlighted in rapidly
expanding number of academic articles [1, 2, 7, 6, 8, 10, 22, 26, 28, 44, 40, 45] in the past years.
Other examples include models with delicate degeneracies (such as the SABR model around zero
forward) which for a precise computation of arbitrage-free prices require time consuming numerical
pricing methods such as Finite Element Methods [42], Monte Carlo [15, 52] or the evaluation of
multiple integrals [3].

Contrary to Hernandez’s [36] pioneering work, where he develops a direct calibation via NN, we set
up and advocate a two setp calibration approach.

Two Step Approach (i) Learn a model and (ii) Calibrate to data: One separates the
calibration procedure described in (2) (resp. (2)) into two parts: (i) We first learn (approximate)
the pricing map by a neural network that maps parameters of a stochastic model to pricing func-
tions (or implied volatilities (cf. section (2.1) and we store this map during an off-line training
procedure. In a second step (ii) we calibrate (on-line) the now deterministic approximative learned
price map, which speeds up the on-line calibration by orders of magnitude. To formalise the two
step approach, we write for a payoff ζ and a model M with parameters θ ∈ Θ

(i) Learn: Fe(Θ, ζ) = Pe(M(Θ, ζ)) (ii) Calibrate: θ̂ = argmin δ(Fe(θ, ζ), P M KT (ζ)). (8)
θ∈Θ

Note that in part (ii) of (8) we essentially replaced Pe(M(Θ, ζ)) in equation (2) by its learned
(deterministic) counterpart Fe(Θ, ζ) (which will be a Neural Network see Section 3.2) from (i).
Therefore, this second calibration is–by its deterministic nature–considerably faster than calibra-
tion of all those traditional stochastic models, which involve numerical simulation of the expected

7
payoff P (M(θ, ζ)) = E[ζ(X(θ))] for some underlying stochastic process X θ . The first part (i) in
(8) denotes an approximation of the pricing map through a neural network, which is calibrated in a
supervised training procedure using the original (possibly slow) numerical pricing maps for training
(see sections 3.2 and 4 for details in specific examples).

In the following sections we elaborate on the objectives and advantages of this two step calibra-
tion approach and present examples of neural network architectures, precise numerical recipes and
training procedures to apply the two step calibration approach to a family of stochastic volatility
models. We also present some numerical experiments (corresponding codes are available on GitHub:
NN-StochVol-Calibrations ) and report on learning errors and on calibration times.

2.3 Challenges in neural network approximations of pricing functionals

In general problem (1) and henceforth (2) is solved using suitable numerical optimisation techniques
such as gradient descent [32], specific methods for certain metrics (such as Lavenberg-Marquadnt
[53] for L2 ), neural networks, or tailor-made methods to the complexity of the optimisation prob-
lem and objective function at hand1 . But irrespective of their level of sophistication all optimis-
ers for calibration share a common property: repeated (iterative) evaluation of the pricing map
θ 7→ P (M(θ), ζ) (resp. an approximation Pe thereof) on each instance θ of consecutive parame-
ter combinations until a sufficiently small distance δ(Pe(M(θ), ζ), P M KT (ζ) between model prices
and observed prices is achieved. Consequently, the pricing map is arguably the computational
cornerstone of a calibration algorithm. Main differences between specific calibration algorithms ef-
fectively lie in the way the specific choice of evaluated parameter combinations {θ1 , θ2 . . .} are deter-
mined, which
hence determines the total number N of functional evaluations of the pricing function
P (M(θi ), ζ) i=1...N used in the calibration until the desired precision δ(Pe(M(θ̂), ζ), P M KT (ζ)) is
achieved. In case the pricing map

P (M(·), ζ) : Θ −→ P (M)
θ 7→ P (M(θ), ζ)

involved in (1) is available in closed form, and can be evaluated instantaneously, the calibration (2)
is fast even if a high number N of functional evaluations is used. If the pricing map is approxi-
mated numerically, calibration time depends strongly on the time needed to generate a functional
evaluation of the numerical approximation

θi 7→ Pe(M(θi ), ζ), θi ∈ {θ1 , . . . θN } (9)

at each iteration i = 1, . . . , N of the calibration procedure. Slow functional evaluations potentially

cause substantial bottlenecks in calibration time. This is where we see the most powerful use of the
prowess of neural network approximation:

A neural network is constructed to replace in (i) of (8) the pricing map, that is to approximate (for
a given financial contract ζ) the pricing map from the full set2 of model parameters Θ of the model
to the corresponding prices P (M(θ, ζ)). The first challenge for the neural network approximator of
pricing functionals is to speed up this process and enable us to obtain faster functional evaluations
1 For details and an overview on calibration methods see [32].
2 Note that the set θ1 , . . . , θN in (9) is extended to the full set of possible parameter combinations Θ in (10).

8
and thereby lift the bottleneck of calibration. The second challenge is to do so with an accuracy
that remains within the error bounds of the original numerical pricing discretisation:

Fe : Θ −→ Pe(M)
(10)
θ 7→ Fe(θ, ζ)
More precisely (motivated by (2)), for any parameter combination θ ∈ Θ we aim to approximate
the numerical approximation Pe of the true option price P with the neural network Fe up to the
same order of precision  > 0 up to which Pe approximates P . That is, for any θ ∈ Θ
Fe(θ) = P (M(θ), ζ) + O() whenever Pe(M(θ), ζ) = P (M(θ), ζ) + O().
Therefore, our training objective is
Fe(θ) = Pe(M(θ), ζ) + O(). (11)

where Pe is the available numerical approximation of the pricing function, which is considered as
ground truth. In our numerical experiments in Section 4 we demonstrate that our approximation
network achieves this approximation accuracy and yields a substantial speedup in terms of functional
evaluations.

2.4 Motivations for our choice of training setup and features of neural
networks as approximators of pricing functionals
There are several advantages of separating the tasks of pricing and calibration which we address in
full detail in a separate work. Here we recall some of the most convincing reasons to do so. Above
all, the most appealing reason is that it allows us to build upon the knowledge we have gained about
the models in the past decades, which is of crucial importance from a risk management perspective.
By its very design, deep learning the price approximation (i) combined with (ii) deterministic
calibration does not cause more headache to risk managers and regulators than the corresponding
stochastic models do. Designing the training as described above demonstrates how deep learning
techniques can successfully extend the toolbox of financial engineering, without making compromises
on any of our objectives.
1. The knowledge gathered in many years of experience with traditional models remains useful
and risk management libraries of models remain valid. The neural network is only used as a
computational enhancement of models.
2. The availability of training data for training the deep neural network does not cause any
constraints as it is synthetically generated by traditional numerical methods.
3. This can be extended beyond the models presented in this work: Whenever a consistent
numerical pricer exists for a model, it can be approximated and replaced by a deep neural
network that provides fast numerical evaluations of the pricing map.
Here, we identify the grid-based apporach as our choice of training. Though a thorough analysis of
the best training approaches is subject to further research, we have good reason to believe that the
grid-based approach provides a powerful and robust methodology for training:

9
2.4.1 Reasons for the choice of grid-based implicit training
In the grid-based approach we evaluate the values of implied volatility surface along 8×11 gridpoints
with 80, 000 different parameter combinations we effectively evaluate the ”fit” of the surface to
numerically generated ones across the same number of points. By moving the evaluation of the
implied volatilities into the objective function we improve the learning in many aspects:
• The first advantage of implicit training is that it efficiently exploits the structure of the data.
Updates in neighbouring volatility points σn−1 and σn can be incorporated in the learning
process. If the output is a full grid as in (16) this effect is further enhanced. Updates of
the network on each gridpoint also imply additional information for updates of the network
on neighbouring gridpoints. One can say that we regard the implied volatility surface as an
image with a given number of pixels.
• A further advantage of the image based implicit training is, that by evaluating the objective
function on a larger set of (grid) points, injectivity of the mapping can be more easily guar-
anteed than in the pointwise training: Two distinct parameter combinations are less likely to
yield the same value across a set of gridpoints, then if evaluated only on a single point.
• We do not limit ourselves to one specific grid on the implied volatility surface. We store the
generated 60, 000 sample paths for the training data and chose a set of maturities (here 8)
and strikes (here 11) to evaluate prices corresponding to these paths. But we can easily add
and evaluate additional maturities and strikes to the same set of paths. Note in particular
that in this training design we can refine the grid on the implied volatility surface without
increasing the number of training samples needed and without significantly increasing the
computational time for training as the portfolio of vanilla options on the same underlying
grows with different strikes and maturities.

2.4.2 Some relevant properties of deep neural networks as functional approximators

Deep feed forward3 neural networks are the most basic deep neural networks, originally designed to
approximate some function F ∗ , which is not available in closed form but only through sample pairs
of given input data x and output data y = F ∗ (x). In a nutshell, a feed forward network defines
a mapping y = F (x, w) and the training determines (calibrates) the optimal values of network
parameters w b that result in the best function approximation4 F ∗ (·) ≈ F (·, w)b of the unknown
∗
function F (·) for the given pairs of input and output data (x, y), cf. [32, Chapter 6].
To formalise this, we introduce some notation and recall some basic definitions and principles of
function approximation via (feedforward) neural networks:
Definition 1 (Neural network). Let L ∈ N and the tuple (N1 , N2 · · · , NL ) ∈ NL denote the number
of layers (depth) and the number of nodes (neurons) on each layer respectively. Furthermore, we
introduce the affine functions
wl : RNl −→ RNl+1 for 1 ≤ l ≤ L − 1
(12)
x 7→ Al+1 x + bl+1
3 The network is called feed forward if there are no feedback connections in which outputs of the model are fed

back into itself.

4 In our case y is a 8 × 11-point grid on the implied volatility surface and x are model parameters θ ∈ Θ, for

details see Section 3.

10
acting between layers for some Al+1 ∈ RNl+1 ×Nl . The vector bl+1 ∈ RNl+1 denotes the bias term
and each entry Al+1
(i,j) denotes the weight connecting node i ∈ Nl of layer l with node j ∈ Nl+1
of layer l + 1. For the the collection of affine functions of the form (12) on each layer we fix the
notation w = (w1 , . . . , wL ). We call the tuple w the network weights for any such collection of affine
functions. Then a Neural Network F (w, ·) : RN0 → RNL is defined as the composition:
F := FL ◦ · · · ◦ F1 (13)

where each component is of the form Fl := σl ◦ W l . The function σl : R → R is referred to as

the activation function. It is typically nonlinear and applied component wise on the outputs of the
affine function W l . The first and last layers, F1 and FL , are the input and output layers. Layers in
between, F2 · · · FL−1 , are called hidden layers.
The following central result of Hornik justifies the use of neural networks as approximators for
multivariate functions and their derivatives.
Theorem 1 (Universal approximation theorem (Hornik, Stinchcombe and White [38])). Let N N σd0 ,d1
be the set of neural networks with activation function σ : R 7→ R, input dimension d0 ∈ N and out-
put dimension d1 ∈ N. Then, if σ is continuous and non-constant, N N σd0 ,1 is dense in Lp (µ) for
all finite measures µ.
There is a rapidly growing literature on approximation results with neural networks, see [37, 39,
56, 65] and the references therein. Among these we would like to single out one particular result:
Theorem 2 (Universal approximation theorem for derivatives (Hornik, Stinchcombe and White
[39])). Let F ∗ ∈ C n and F : Rd0 → R and N N σd0 ,1 be the set of single-layer neural networks with
activation function σ : R 7→ R, input dimension d0 ∈ N and output dimension 1. Then, if the
(non-constant) activation function is σ ∈ C n (R), then N N σd0 ,1 arbitrarily approximates f and all
its derivatives up to order n.
Remark 1. Theorem 2 highlights that the smoothness properties of the activation function are of
significant importance in the approximation of derivatives of the target function F ∗ . In particular,
to guarantee the convergence of l-th order derivatives of the target function, we choose an activation
function σ ∈ C l (R). Note that the ReLu activation function , σReLu (x) = (x)+ is not in C l (R) for
any l > 0, while σElu (x) = α(ex − 1) is smooth.

Bias
b
Weights
x1 ai,1
Activation
3
X function Output
Node inputs x2 ai,2 bi + ai,j xj σELU y
j=1

x3 ai,3

Figure 1: In detail neuron behaviour

11
 The following Theorem provides theoretical bounds for the above rule of thumb and establishes
a connection between the number of nodes in a network and the number of training samples needed
to train it.
Theorem 3 (Estimation bounds for Neural Networks (Barron [5])). Let N N σd0 ,d1 be the set of
x
single-layer neural networks with Sigmoid activation function σ(x) = exe+1 , input dimension d0 ∈ N
and output dimension d1 ∈ N. Then:
!
2
C
 
∗ 2 f nd0
EkF − F̂ k2 ≤ O +O log N
n N

where n is the number of nodes, N is the training set size and CF ∗ is the first absolut moment of
the Fourier magnitude distribution of F ∗ .

Remark 2. Barron’s [5] insightful result gives a rather explicit decomposition of the error in terms
of bias (model complexity) and variance:
 2 
C
• O nF ∗ represents the model complexity, i.e. the larger n (number of nodes) the smaller
the error

• O nd


N log N represents the variance, i.e. a large n must be compensated with a large training
0

set N in order to avoid overfitting.

Finally, we motivate the use of multi layer networks and the choice of network depth. Even though a
single layer might theoretically suffice to arbitrarily approximate any continuous function,in practice
the use of multiple layers dramatically improves the approximation capacities of network. We
informally recall the following Theorem due to Eldan and Shamir [20] and refer the reader to the
original paper for details.
Theorem 4 (Power of depth of Neural Networks (Eldan and Shamir [20])). There exists a simple
(approximately radial) function on Rd , expressible by a small 3-layer feedforward neural networks,
which cannot be approximated by any 2-layer network, to more than a certain constant accuracy,
unless its width is exponential in the dimension.
Remark 3. In spite of the specific framework by Eldan and Shamir [20] being restrictive, it provides a
theoretical justification to the power of “deep” neural networks (multiple layers) against “shallower”
networks (i.e. few layers) as in [55] with a larger number of neurons. On the other hand, multiple
findings indicate [9, 47] that adding hidden layers beyond 4 hidden layers does not significantly
improve network performance.

3 Pricing and calibration with neural networks: Optimising

network and training
In this section we compare different objective functions (direct calibration to data to an image-based
implicit learning approach) and motivate our choice of image-based objective function. We give
details about network architectures for the approximation network and compare different optimisers
for the calibration step.

12
3.1 The objective function
1. Learn the map F ∗ (θ) = {P M(θ) (ζi )}ni=1 via neral network, where {ζi }i=1,...,n represents the
exotic product attributes (such as maturity, strike or barrier level) on a prespecified grid with
size n.
NX n
T rain X
2
ŵ = argmin (F (θu , w)i − F ∗ (θu )i ) .
w∈Rn u=1 i=1

2. Solve
n
X
θ̂ := argmin (Fe(θ)i − P M KT (ζi ))2 . (14)
θ∈Θ i=1

3.1.1 For vanillas

As in many academic and industry research papers, we pursue the calibration of vanilla contracts
via approximation of the implied volatility surface5 .

We take this idea further and design an implicit form of the pricing map that is based on storing the
implied volatility surface as an image given by a grid of ”pixels”. This image-based representation
has a formative contribution in the performance of the network we present in Section 4. We present
our contribution here; Let us denote by ∆ := {ki , Tj }n, m
i=1, j=1 a fixed grid of strikes and maturities,
then we propose the following two step approach:
M(θ)
1. Learn the map F ∗ (θ) = {σBS (Ti , kj )}n, m
i=1, j=1 via neural network F (θ) := F (θ, ŵ) where
e

F ∗ : Θ −→ Rn×m (16)
∗
θ 7→ F (θ)

where the input is a parameter combination θ ∈ Θ of the stochastic model M(Θ) and the
M(θ)
output is a n × m grid on the implied volatility surface {σBS (Ti , kj )}n, m
i=1, j=1 where n, m ∈ N
are chosen appropriately (see Section 3.2). Then,
NX n X
T rain X m
2
ŵ = argmin (F (θu , w)ij − F ∗ (θu )ij ) .
w∈Rn u=1 i=1 j=1
5 Forsake of completeness we introduce the Black-Scholes Call pricing function in terms of log-strike k, initial spot
S0 , maturity T and volatility σ:
√
log(S0 ) − k Tσ
BS(σ, S0 , k, T ) := S0 N (d+ ) − KN (d− ), d± := √ ± ,
Tσ 2
where N (·) denotes the Gaussian cumulative distribution function. The implied volatility induced by a Call option
pricing function P (K, T ) is then given by the unique solution σBS (k, T ) of the following equation
BS(σBS (k, T ), S0 , k, T ) = P (k, T ).
Precisely, we seek to solve the following calibration problem
M(θ)
θ̂ := argmin d(ΣBS , ΣM
BS
KT
) (15)
θ∈Θ
M(θ) M(θ)
where ΣBS := {σBS (ki , Tj )}i=1,..,n, j=1,...,m represents the set of implied volatilities generated by the model
pricing function P (M(θ), k, T ) and ΣM
BS
KT := {σ M KT (k , T )}
BS i j i=1,..,n, j=1,...,m are the corresponding market implied
volatilities, for some metric d : Rn×m × Rn×m → R+ .

13
 2. Solve
n X
X m
M KT
θ̂ := argmin (Fe(θ)ij − σBS (Ti , kj ))2 .
θ∈Θ i=1 j=1

Remark 4. Notice that ŵ(∆) depends on ∆ implicitly, consequently so does Fe(θ) = F (θ, ŵ(∆))
(hence the name implicit learning). This setting is similar to that of image recognition and exploits
the structure of the data to reduce the complexity of the Network (see Section 4 for details).
Remark 5. In our experiments we chose n = 8 and m = 11. At first, a criticism of mapping
(16) might be the inability to extrapolate/interpolate between maturities/strikes outside the grid
∆. However, one is free to choose the grids ∆ as fine as needed. In addition, one may use
standard (arbitrage free) uni/bi-variate splines techniques to extrapolate/interpolate across strikes
and maturities, as with traditional market data observable only at discrete points.

Figure 2: Volatility surface generated by the neural network approximator and the corresponding
original counterpart on a grid given by 8 maturities and 11 strikes.

3.1.2 Some exotic payoffs

Our framework extends to a number of exotic products such as: Digital barriers, no-touch (or dou-
ble no-touch) barrier, cliquets or autocallables.

We present some numerical experiments in Section 4.3, to demonstrate the pricing of digital barrier
options. More precisely, in Section 4.3 we consider down-and-in such as down-and-out digital barrier
options, the main building blocks of many Autocallable products. For a barrier level B < S0 and
maturity T the payoff is given by:

P Down−and−In (B, T ) = E 1{τB ≤T }

 
(17)
P Down−and−Out (B, T ) = E 1{τB ≥T }
 
(18)

where τB = inf {St = B}. In this setting, we may easily generate a grid for barrier levels and
t
maturities ∆Barrier := {Bi , Tj }n, m
i=1, j=1 that we can fit in the objective function specified in (14)

3.2 Network architecture and training

Motivated by the above analysis, we choose to set up the calibration in the implicit two-step
approach. This involves a separation of the calibration procedure into (i) “Deep approximation” an
approximation network with an implicit training and (ii)“Calibration” a calibration layer on top.
We first start by describing the approximation network in the implicit image-based training and
discuss the calibration in Section 3.3 below. In addition, we will highlight specific techniques that
contribute to the robustness and efficiency of our design.

14
 3.2.1 Network architecture of the implied volatility map approximation
Here we motivate our choice of network architecture for the following numerical experiments which
were inspired by the analysis in the previous sections. Our network architecture is summarised in
the graph 3.2.1 below.

1. A fully connected feed forward neural network with 4 hidden layers (due to Theorem 4) and
30 nodes on each layers (see Figure 3.2.1 for a detailed representation)
2. Input dimension = n, number of model parameters
3. Output dimension = 11 strikes× 8 maturities for this experiment, but this choice of grid can
be enriched or modified.
4. The four inner layers have 30 nodes each, which adding the corresponding biases results on a
number
(n + 1) × 30 + 4 × (1 + 30) × 30 + (30 + 1) × 88 = 30n + 6478
of network parameters to calibrate (see Section 2.4.2 for details).
5. Motivated by Theorem 2 we choose the Elu σElu = α(ex − 1) activation function for the
network.

1st 2nd 3rd 4th

Input Output
Hidden Hidden Hidden Hidden
layer layer
layer layer layer layer
Output 1

..
1 1 1 1 .

Input 1 .. .. .. .. ..
θ1 . . . . .

..
. .. .. .. .. .. ..
. . . . . .

Input n .. .. .. .. ..
θn . . . . .

..
30 30 30 30 .

Output 88

Figure 3: Our neural network architecture with 4 hidden layers and 30 neurons on each hidden
layer, with the model parameters of the respective model on the input layer and with the 8 × 11
implied volatility grid on the output layer.

15
3.2.2 Training of the approximation network
We follow the common features of optimization techniques and choose mini-batches, as described in
Goodfellow, Bengio and Courville [32]. Typical batch size values range from around 10 to 100. In
our case we started with small batch sizes and increased the batch size until training performance
consistently reached a plateau. Finally, we chose batch sizes of 32, as performance is similar for
batch sizes above this level, and larger batch sizes increase computation time by computing a larger
number of gradients at a time.

In our training design, we use a number of regularisation techniques to speed up convergence of the
training, to avoid overfitting and improve the network performance.

1) Early stopping: We choose the number of epochs as 200 and stop updating network pa-
rameters if the error has not improved in the test set for 25 steps.

2) Normalisation of model parameters: Usually, model parameters are restricted to a given

domain i.e. θ ∈ [θmin , θmax ]. Then, we perform the following normalisation transform:
2θ − (θmax + θmin )
∈ [−1, 1].
θmax − θmin
3) Normalisation of implied volatilities: The normalisation of implied volatilities is a more
delicate matter, since σBS (T, k, θtrain ) ∈ [0, ∞), for each T and k. Therefore, we choose to normalise
the surface subtracting the sample empirical mean and dividing by the sample standard deviation.

3.3 The calibration step

Once the pricing map approximation Fe for the implied volatility is found, only the calibration step
in (2) is left to solve. In general, for financial models the pricing map F ∗ is assumed to be smooth
(at least C 1 differentiable) with respect to all its input parameters θ.

Gradient-based optimizers

A standard necessary first order condition for optimality in (2) is that

 
∇θ δ Fe(M(θ), ζ), P M KT (ζ) = 0, (19)

provided that the objective function is smooth. Then, a natural update rule is to move along the
gradient via Gradient Descent i.e.
 
θi+1 = θi − λ∇θ δ Fe(M(θi ), ζ), P M KT (ζ) , λ > 0. (20)

A common feature
 of gradient based optimization methods building on (20) is the use of the
gradient ∇ δ Fe(M(θ), ζ), P M KT (ζ) , hence its correct and precise computation is crucial for
θ

subsequent success. Examples of such algorithms, are Levenberg-Marquardt [53, 54], Broyden-
Fletcher-Goldfarb-Shanno (BFGS) algorithm [58], L-BFGS-B [70] and SLSQP [50]. The main

16
advantage of the aforementioned methods is the quick convergence towards condition (19). However,
(19) only gives necessary and not sufficient conditions for optimality, hence special care must be
taken with non-convex problems.
Remark 6. Notably, making use of Theorem 2 we use a smooth activation functions in order to
guarantee ∇θ Pe ≈ ∇θ Fe

Gradient-free optimizers

Gradient-free optimization algorithms are gaining popularity due to the increasing number of high
dimensional nonlinear, non-differentiable and/or non-convex problems flourishing in many scientific
fields such as biology, physics or engineering. As the name suggests, gradient-free algorithms make
no C 1 assumption on the objective function. Perhaps, the most well known example is the Simplex
based Nelder-Mead [57] algorithm. However, there are many other methods such as COBYLA [60]
or Differential Evolution [68] and we refer the reader to [61] for an excellent review on gradient-
free methods. The main advantage of these methods is the ability to find global solutions in (2)
regardless of the objective function. In contrast, the main drawback is a higher computational cost
compared to gradient methods.

To conclude, we summarise the advantages of each approach in Table 1.

Gradient-based Gradient-free
Convergence Speed Very Fast Slow
Global Solution Depends on problem Always
Smooth activation
Yes to apply Theorem 2 No
function needed
Accurate gradient
Yes No
approximation needed

Table 1: Comparison of Gradient vs. Gradient-free methods.

4 Numerical experiments
In our numerical experiments we demonstrate that the accuracy of the approximation network
indeed remains within the accuracy of the Monte Carlo error bounds and proclaimed in the in-
troductory sections’ objectives. For this we first compute the benchmark Monte Carlo errors in
Figures 4-5 and compare this with the neural network approximation errors in Figures 6 and 7.
For this separation into steps (i) and (ii) to be computationally meaningful, the neural network
approximation has to be a reasonably accurate approximation of the true pricing functionals and
each functional evaluation (i.e. evaluation an option price for a given price and maturity) should
have a considerable speed-up in comparison to the original numerical method. In this section we
demonstrate that our network achieves both of these goals.

17
 4.1 Numerical accuracy and speed of the price approximation for vanillas
As mentioned in Section 2 one crucial difference that sets apart this work from direct neural net-
work approaches, as pioneered by Hernandez [36], is the separation of (i) the implied volatility
approximation function, mapping from parameters of the stochastic volatility model to the implied
volatility surface–thereby bypassing the need for expensive Monte-Carlo simulations—and (ii) the
calibration procedure, which (after this separation) becomes a simple deterministic optimisation
problem. As outlined in Section 2.3 our aim for the Step (i) in the two-step training approach is
to achieve a considerable speedup per functional evaluation of option prices while maintaining the
numerical accuracy of the original pricer. Here we demonstrate how our NN training for Step (i)
achieves these goals outlined in Section 2.3:

1. Approximation accuracy: here we compare the error of the approximation network error to
the error of Monte Carlo evaluations. We compute Monte Carlo prices with 60, 000 paths as
reference at the nodes where we compute the implied volatility grid using Algorithm 3.5 in
Horvath, Jacquier and Muguruza [41]. In Figures 4 and 5 the approximation accuracy of the
Monte Carlo method for the full implied volatility surface is computed using pointwise relative
error with respect to the 95% Monte Carlo confidence interval. Figures 6 and 7 demonstrate
that the same approximation accuracy for the neural network is achieved as for the Monte
Carlo approximation (i.e. within a few basis points). For reference, the spread on options
is around 0.2% in implied volatility terms for the most liquid and those below a year. This
translates into 1% relative error for a implied volatility of 20%.
2. Approximation speed: Table 2 shows the CPU computation time per functional evaluation of
a full surface under two different models; rBergomi 3 and 1 Factor Bergomi 6 (for a reminder
see Section 4.1.1 for details).

MC Pricing MC Pricing
NN Pricing NN Gradient Speed up
1F Bergomi rBergomi
Full Surface Full Surface NN vs. MC
Full Surface Full Surface
Piecewise constant
300, 000 µs 500, 000 µs 30.9 µs 113 µs 9, 000 − 16, 000
forward variance

Table 2: Computational time of pricing map (entire implied volatility surface) and gradients via
Neural Network approximation and Monte Carlo (MC). If the forward variance curve is a constant
value, then the speed-up is even more pronounced

18
Figure 4: As benchmark we recall average relative errors of Monte Carlo prices computed across
80, 000 random parameter combinations of the Rough Bergomi model. Relative errors are given in
terms of Average-Standard Deviation-Maximum (Left-Middle-Right) on implied volatility surfaces
in the Rough Bergomi model, computed using 95% confidence intervals.

Figure 5: As benchmark we recall average relative errors of Monte Carlo prices computed across
80, 000 random parameter combinations of the 1 Factor Bergomi model. Relative errors are given in
terms of Average-Standard Deviation-Maximum (Left-Middle-Right) on implied volatility surfaces
in the 1 Factor Bergomi model, computed using 95% confidence intervals.

4.1.1 Neural network price approximation in (rough) Bergomi models with piecewise
constant forward variance curve
Pn
We consider a piecewise constant forward variance curve ξ0 (t) = i−1 ξi1{ti−1 <t<ti } where t0 =
0 < t1 < ... < tn and {ti }i = 1, .., n are the option maturity dates (n = 8 in our case). This is the
modelling approach suggested by Bergomi [11]. We will consider again the rough Bergomi 3 and 1

19
Factor Bergomi models 6

• Normalized parameters as input and normalised implied volatilities as output

• 4 hidden layers with 30 neurons and Elu activation function

• Output layer with Linear activation function
• Total number of parameters: 6808

• Train Set: 68, 000 and Test Set: 12, 000

• Rough Bergomi sample:(ξ0 , ν, ρ, H) ∈ U[0.01, 0.16]8 ×U[0.5, 4.0]×U[−0.95, −0.1]×U[0.025, 0.5]
• 1 Factor Bergomi sample:(ξ0 , ν, ρ, β) ∈ U[0.01, 0.16]8 × U[0.5, 4.0] × U[−0.95, −0.1] × U[0, 10]
• strikes={0.5, 0.6, 0.7, 0.8, 0.9, 1, 1.1, 1.2, 1.3, 1.4, 1.5}

• maturities={0.1, 0.3, 0.6, 0.9, 1.2, 1.5, 1.8, 2.0}

• Training data samples of Input-Output pares are computed using Algorithm 3.5 in Horvath,
Jacquier and Muguruza [41] with 60, 000 sample paths and the spot martingale condition i.e.
E[St ] = S0 , t ≥ 0 as control variate.

Figure 6: We compare surface relative errors of the neural network approximator against the
Monte Carlo benchmark across all training data (68, 000 random parameter combinations)in the
rough Bergomi model. Relative errors are given in terms of Average-Standard Deviation-Maximum
(Left-Middle-Right).

20
Figure 7: We compare surface relative errors of the neural network approximator against the Monte
Carlo benchmark across all training data (68, 000 random parameter combinations)in the 1 Factor
Bergomi model. Relative errors are given in terms of Average-Standard Deviation-Maximum (Left-
Middle-Right).

Figures 6 and 7 show that the average (across all parameter combinations) relative error between
neural network and Monte Carlo approximations is far less than 0.5% consistently (left image in
Figures 6 and 7) with a standard deviation of less than 1% (middle image in Figures 6 and 7).
The maximum relative error goes as far as 25%. We conclude that the methodology generalises
adequately to the case of non-constant forward variances, by showing the same error behaviour.

4.2 Calibration speed and accuracy for implied volatility surfaces

Figure 8 reports average calibration times on test set for different parameter combinations on each
of the models analysed. We conclude that gradient-based optimizers outperform (in terms of speed)
gradient-free ones. Moreover, in Figure 8 one observes that computational times in gradient-free
methods are heavily affected by the dimension of the parameter space, i.e. flat forward variances
are much quicker to calibrate than piecewise constant ones. We find that Lavenberg-Marquardt
is the most balanced optimizer in terms of speed/convergence and we choose to perform further
experiments with this optimizer. The reader is encouraged to keep in mind that a wide range of
optimizers is available for the calibration and the optimal selection of one is left for future research.

21
 Figure 8: Average calibrations times for all models using a range of optimizers.

In order to assess the accuracy, we report the calibrated model parameters θb compared to the
synthetically generated data with the set of parameters θ that was chosen for the generation of our
synthetic data. We measure the accuracy of the calibration via parameter relative error i.e.

b = |θ − θ|
b
ER (θ)
|θ|

as well as the root mean square error (RMSE) with respect to the original surface i.e.
v
u n X
m
uX
RMSE(θ)b =t b ij − σ M KT (Ti , kj ))2 .
(Fe(θ) BS
i=1 j=1

Therefore, on one hand a measure of good calibration is a small RMSE. On the other hand, a
measure of parameter sensitivity on a given model is the combined result of RMSE and parameter
relative error.

4.2.1 A calibration experiment with simulated data in (rough) Bergomi models with
piecewise constant forward variances
We consider the rough Bergomi model (3) and the Bergomi model (6) with a piecewise constant
term-structure of forward variances. Figures 9 and 10 show that the 99% quantile of the RMSE
is below 1% and shows that the Neural Network approach generalises properly to the piecewise
constant forward variance. Again, we find that the largest relative errors per parameter are con-
centrated around 0, consequence of using the relative error as measure. This suggests a successful
generalisation to general forward variances, which to our knowledge has not been addressed before
by means of neural networks or machine learning techniques.

22
Figure 9: Cumulative Distribution Function (CDF) of Rough Bergomi parameter relative errors
(left) and RMSE (right) after Levengerg-Marquardt calibration across test set random parameter
combinations.

Figure 10: Cumulative Distribution Function (CDF) of 1 Factor Bergomi parameter relative errors
(left) and RMSE (right) after Levengerg-Marquardt calibration across test set random parameter
combinations.

4.2.2 Calibration in the rough Bergomi model with historical data

As previously mentioned, the natural use of neural network approximators is the model calibration
to historical data. We discussed that as along as the approximation is accurate, the calibration task
should be performed within the given tolerance. Furthermore, one should expect such tolerance to
be aligned with the neural network accuracy obtained in both training and test sets.

In this section we will perform a historical calibration using the neural network approximation
and compare it with that of the brute force monte carlo calibration. Precisely we seek to solve the

23
following optimisation problem for the rough Bergomi model
5 X
9
ˆ
X
M KT
θrBergomi := argmin (Fe(θ)ij − σBS (Ti , kj ))2 .
θ rBergomi ∈ΘrBergomi i=1 j=1

where θrBergomi = (ξ1 , ξ2 , ξ3 , ξ4 , ξ5 , ν, ρ, H) and ΘrBergomi = [0.01, 0.25]5 × [0.5, 4] × [−1, 0] ×

[0.025, 0.5]. As for the time grid we choose
1
(T1 , T2 , T3 , T4 , T5 ) := × (1, 3, 6, 9, 12)
12
and for the strike grid
ki := 0.85 + (i − 1) × 0.05 for i = 1, ..., 9.
We consider SPX market smiles between 01/01/2010 and 18/03/2019 on the pre-specified time and
strike grid. Figure 11 shows the historical evolution of rough Bergomi parameters calibrated to SPX
using the neural network price. In particular we note that H < 21 as previously discussed in many
academic papers [1, 2, 7, 6, 8, 10, 22, 26, 28, 44, 40, 45], moreover we may confirm that under Q,
H ∈ [0.1, 0.15] as found in Gatheral, Jaisson and Rosenbaum [28] under P. Figure 12, benchmarks
the NN optimal fit using Levenberg-Marquardt and Differential Evolution against a brute force MC
calibration via Levenberg-Marquardt. Again, we find that the discrepancy between both is below
0.2% most of the time and conclude that the Differential Evolution algorithm does outperform the
Levenberg-Marquardt. This in turn, suggests that the neural network might not be precise enough
on first order derivatives. This observation, is left as an open question for further research. Perhaps
surprisingly, we sometimes obtain a better fit using the neural network than the MC pricing itself.
This could be caused by the fact that gradients in the neural network are exact, whereas when
using MC brute force calibration we resort to finite differences to approximate gradients.

24
Figure 11: Historical Evolution of parameters in the rough Bergomi model with a piecewise constant
forward variance term structure calibrated on SPX

Figure 12: The image above compares historical RMSE obtained by the neural network best fit via
Levenberg-Marquardt (dashed orange line) and 25Differential Evolution (dotted purple line) against
the brute force MC calibration (green line) via Levenberg-Marquardt. Picture below shows the
difference against MC brute force calibration.
4.3 Numerical experiments with barrier options in the rough Bergomi
model
In this section we show that our methodology can be easily extended to exotic options. To do so
we test our image-based approach on digital barrier options. We follow the same architecture and
experimental design described in Section 4.1 for the rough Bergomi model. As described in Section
3.1.2 we adapt the objective function to the payoffs given in (17) and (18) and replace the strike
grid by a barrier level grid. Figure 13 confirm the accuracy of the neural network approximation
with average absolute errors of less than 10bps with standard deviation of 10bps.

Figure 13: Picute above: Down-and-Out neural network absolute error analysis on test set. Picture
below: Down-and-In neural network absolute error analysis on test set

5 Conclusions and outlook: “best-fit” models

To sum up, neural networks have the potential to efficiently approximate complex functions, which
are difficult to represent and time-consuming to evaluate by other means. Using deep neural net-
works, as we will do here, to approximate the pricing map (or equivalently the implied volatility
mapping) from parameters of traditional models to shapes of the implied volatility surface repre-
sented by grid of implied volatility values speeds up each functional evaluation, while maintaining

26
control over reliability and interpretability of network outputs. The implicit grid based approach
that we advocate here, also allows further applications that opens up further landscapes for finan-
cial modelling.

Potential applications and outlook towards mixture of “expert” models: In the previous
sections we set up a powerful approximation method to closely approximate implied volatilities un-
der different stochastic models and highlighted that the choice of the objective function (evaluation
of the surface on a grid, inspired by pixels of an image) was crucial for the performance of the
network. Now we are interested in the inverse task and ask whether a neural network—trained by
this objective function to multiple stochastic models simultaneously—can identify which stochastic
model a given set of data comes from. By doing so, potential applications we have in mind are
twofold:
(1) Ultimately we are interested in which model (or what mixture of existing stochastic models)
best describes the market.
(2) From a more academic and less practical perspective, we are interested whether and to what
extent is it possible to“translate” parameters of one stochastic model to parameters of another.

We conduct a further, preliminary experiment as a proof of concept in the classification setting. We

train a further neural network to identify which of three given stochastic volatility model generated
a given implied volatility surface.

Training procedure: Implied volatility surfaces in this experiment were generated by the
Heston, Bergomi and rough Bergomi models (see Section 2.1 for a reminder). For each volatility
surface, a “flag” was assigned corresponding to the model (eg: 1 for Heston, 2 for Bergomi and 3
M(θ)
for rough Bergomi). The training set thus consists of surfaces of the form: (ΣBS , I), where M is
Heston Bergomi rBergomi
one of the three models M ,M ,M , θ an admissible combination of parameters
for that model (thus in ΘHeston , ΘBergomi or ΘrBergomi ) and I the flag identifying the model which
generated the surface (I = 1 if M = MHeston , I = 2 if M = MBergomi and I = 3 if M = MrBergomi ).
Mixture((a,b,c)) Heston Bergomi rough Bergomi
We define a mixture of these surfaces as ΣM := aΣM +bΣM +cΣM ,
where a, b, c ≥ 0 and a + b + c = 1. So far the training is suitable for recognition of a single model
surface (either a = 0, b = 0, c = 1, a = 0, b = 1, c = 0 or a = 1, b = 0, c = 0). To generalise this
to mixtures, we randomly select surfaces (one from each model) and compute the mixture surface
Mixture((a,b,c)) Heston Bergomi rough Bergomi
ΣM = aΣM + bΣM + cΣM . The corresponding probabilities are
(a, b, c = 1 − a − b).

Network Architecture: The classifier is a small, fully connected feedforward network for the
same reasons as those outlined in section 3.2. The network is composed of 2 hidden layers (of 100
and 50 output nodes respectively) with exponentially linear activation functions and an output
layer with a softmax activation function. Thus, the output of the network represents the probabil-
ities of a given surface belonging to a particular model. We used stochastic gradient descent with
20 epochs to minimize cross-entropy (theP cross-entropy of two discrete distributions (p, q) with K
possible distinct values is H(p, q) := − 1≤i≤K pi log qi .).

A numerical experiment on model recognition: We report on one of many experiments

here as a proof of concept: We test the method on mixtures of rough Bergomi and Heston surfaces

27
(hence setting b = 0 in the training). To vary the type of mixtures generated, we chose a ∈
{0, 0.1, · · · , 0.9, 1}. For each a, the mixture surface is computed as the convex combination of a
randomly chosen surface from the rough Bergomi and the Heston model, and repeated 20 times.
The training set has 320,000 surfaces. To further test the robustness of the model, validation
surfaces were generated using a finer grid of mixture parameters: a ∈ {0, 0.05, · · · , 0.95, 1}. In
total, the validation set is made up of 105,000 surfaces.
We report the classifiers’ effectiveness and comment on the results in Figure 5.

Figure 14: Error of the neural network classifier depending on the mixture coefficient a. Each point
of the plot corresponds to the average estimated coefficient by the neural network for all mixture
surfaces with a given a. For example, for a = 0, the surfaces are generated from the rough Bergomi
model. For each parameter combination from those surfaces, we compute the predicted mixture
coefficient and average all of them over the validation set to report â. The network never sets the
mixture coefficient very close to 1 or 0, attributing the surface to one specific model. This may be
explained using Bayesian reasoning.

References
[1] E. Alòs, D. Garcı́a-Lorite and A. Muguruza. On smile properties of volatility derivatives and
exotic products: understanding the VIX skew. arXiv:1808.03610, 2018.
[2] E. Alòs, J. León and J. Vives. On the short-time behavior of the implied volatility for jump-
diffusion models with stochastic volatility. Finance and Stochastics, 11(4), 571-589, 2007.

[3] A, Antonov, M. Konikov and M. Spector. SABR spreads its wings. Risk (August issue), pp.
58-63, 2013.
[4] M. Avellaneda, A. Carelli and F. Stella. Following the Bayes path to option pricing. Journal
of Computational Intelligence in Finance, 1998.

28
 [5] A.R. Barron. Approximation and estimation bounds for artificial neural networks. Machine
Learning. Vol.14:1,1994 Pages 115-133 .
[6] C. Bayer, P. Friz, P. Gassiat, J. Martin and B. Stemper. A regularity structure for rough
volatility. arXiv:1710.07481, 2017.
[7] C. Bayer, P. Friz and J. Gatheral. Pricing under rough volatility. Quantitative Finance, 16(6):
1-18, 2015.
[8] C. Bayer, P. Friz, A. Gulisashvili, B. Horvath and B. Stemper. Short-time near the money
skew in rough fractional stochastic volatility models. arXiv:1703.05132, 2017.
[9] C. Bayer and B. Stemper. Deep calibration of rough stochastic volatility models. Preprint,
arXiv:1810.03399
[10] M. Bennedsen, A. Lunde and M.S. Pakkanen. Hybrid scheme for Brownian semistationary
processes. Finance and Stochastics, 21(4): 931-965, 2017.
[11] L. Bergomi. Stochastic Volatility Modeling. Chapman & Hall/CRC financial mathematical
series. Chapman & Hall/CRC, 2015.
[12] A. Broström and R. Kristiansson. Exotic Derivatives and Deep Learning. Unpublished Thesis,
KTH Royal Institute of Technology, School of Engineering Sciences, Stockholm, Sweden, 2018.
[13] H. Buehler, L. Gonon, J. Teichmann and B. Wood. Deep Hedging. Preprint, arXiv:1802.03042,
2018.
[14] J. Cao, J. Chen and J.C. Hull. A Neural Network Approach to Understanding Implied Volatility
Movements. SSRN:3288067, 2018.
[15] B. Chen, C. W. osterlee and H. Van Der Weide. Efficient unbiased simulation scheme for the
SABR stochastic volatility model, 2011.
[16] D. Clevert, T. Unterthiner, S. Hochreiter. Fast and accurate deep network learning by expo-
nential linear units (ELUs). Preprint, arXiv:1511.07289, 2015.
[17] J. De Spiegeleer, D. Madan, S. Reyners and W. Schoutens. Machine learning for quantitative
finance: Fast derivative pricing, hedging and [Link], 2018.
[18] G. Dimitroff, D. Röder and C. P. Fries. Volatility model calibration with convolutional neural
networks. Preprint, SSRN:3252432, 2018.
[19] M. Duembgen and L.C.G. Rogers. Estimate Nothing. Quantitative Finance, 14(12), pp.2065-
2072, 2014.
[20] R. Eldan and O. Shamir. The power of depth for feedforward neural neworks. JMLR: Workshop
and Conference Proceedings Vol 49:1-34, 2016.
[21] C. Doléans-Dade. Quelques applications de la formule de changement de variables pour les
semimartingales. Z. Wahrscheinlichkeitstheorie verwandte Gebiete, Vol 16: 181-194, 1970.
[22] O. El Euch and M. Rosenbaum. The characteristic function of rough Heston models. To appear
in Mathematical Finance.

29
[23] O. El Euch and M. Rosenbaum. Perfect hedging in rough Heston models, to appear in The
Annals of Applied Probability, 2018.
[24] J. Friedman, R. Tibshiran and T. Hastie. The Elements of Statistical Learning. Springer New
York Inc, 2001.

[25] R. Ferguson and A. Green. Deeply learning derivatives. Preprint arXiv:1809.02233, 2018.
[26] M. Fukasawa. Asymptotic analysis for stochastic volatility: martingale expansion. Finance and
Stochastics, 15: 635-654, 2011.
[27] J. Gatheral. A parsimonious arbitrage-free implied volatility parameterization with application
to the valuation of volatility derivatives, Presentation at Global Derivatives, 2004.

[28] J. Gatheral, T. Jaisson and M. Rosenbaum. Volatility is rough. Quantitative Finance, 18(6):
933-949, 2018.
[29] J. Gatheral, A. Jacquier. Arbitrage-free SVI volatility surfaces. Quantitative Finance, 14(1):
59-71, 2014

[30] A. Gulisashvili, [Link] and A. Jacquier. Mass at zero in the uncorrelated SABR model.
Quantitative Finance, 18(10): 1753-1765, 2018.
[31] A. Gulisashvili, [Link] and A. Jacquier. On the probability of hitting the boundary for
Brownian motions on the SABR plane. Electronic Communications in Probability, 21(75): 1-13,
2016.

[32] I. Goodfellow, Y. Bengio and A. Courville. Deep Learning. MIT Press, 2016.
[33] P. Hagan, D. Kumar, A. Lesniewski, and D. Woodward. Managing smile risk. Wilmott Maga-
zine, September issue: 84-108, 2002.
[34] P. Hagan, A. Lesniewski, and D. Woodward. Probability distribution in the SABR model of
stochastic volatility. Large Deviations and Asymptotic Methods in Finance, Springer Proceed-
ings in Mathematics and Statistics, 110, 2015.
[35] S. Hendriks, C. Martini The Extended SSVI Volatility Surface. Preprint, SSRN:2971502, 2017.
[36] A. Hernandez. Model calibration with neural networks. Risk, 2017.

[37] [Link]. Approximation capabilities of multilayer feedforward [Link],

4(2):251-257, 1991.
[38] K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward networks are universal
approximators. Neural Networks, 2(5):359-366, 1989.

[39] K. Hornik. M. Stinchcombe and H. White. Universal approximation of an unknown mapping

and its derivatives using multilayer feedforward networks. Neural Networks. Vol. 3:11, 1990.
[40] B. Horvath, A. Jacquier and C. Lacombe. Asymptotic behaviour of randomised fractional
volatility models. arXiv:1708.01121, 2017.

30
[41] B. Horvath, A. Jacquier and A. Muguruza. Functional central limit theorems for rough volatil-
ity. arXiv:1711.03078, 2017.
[42] B. Horvath, O. Reichmann. Dirichlet Forms and Finite Element Methods for the SABR Model.
SIAM Journal on Financial Mathematics, p. 716-754(2), May 2018.
Hutchinson, James M., Andrew W. Lo, and Tomaso Poggio. ”A nonparametric approach to
pricing and hedging derivative securities via learning networks.” The Journal of Finance 49.3
(1994): 851-889.
[43] J.M. Hutchinson, A.W. Lo, T. Poggio. A nonparametric approach to pricing and hedging
derivative securities via learning networks. The Journal of Finance, 49 p. 851-889(3), 1994.

[44] A. Jacquier, C. Martini and A. Muguruza. On VIX futures in the rough Bergomi model.
Quantitative Finance, 18(1): 45-61, 2018.
[45] A. Jacquier, M. Pakkanen and H. Stone. Pathwise large deviations for the rough Bergomi
model. Journal of Applied Probability, 55(4): pp.1078-1092, 2018.
[46] A. Jentzen, B. Kuckuck, A. Neufeld, P. von Wurstemberger. Strong error analysis for stochastic
gradient descent optimization algorithms, Preprint arXiv:1801.09324 ,2018.
[47] S. Ioffe and C. Szegedy. Batch normalisation: Accelerating deep network training by reducing
internal covariate shift. Preprint, arXiv:1502.03167, 2015.
[48] D.P. Kingman and J. Ba, Adam: A Method for Stochastic Optimization. Conference paper,
3rd International Conference for Learning Representations, 2015.
[49] A. Kondratyev. Learning curve dynamics with artificial neural networks. Preprint,
SSRN:3041232, 2018.
[50] D. Kraft. A Software Package for Sequential Quadratic Programming. DFVLR-FB pp.88-28,
1988.

[51] G. Kutyniok, H. Bölcskei, P. Grohs and P. Petersen, Optimal approximation with sparsely
connected deep neural networks, Preprint arXiv:1705.01714, 2017.
[52] A. Leitao Rodriguez, L.A. Grzelak and C.W. Oosterlee. On an efficient multiple time step
Monte Carlo simulation of the SABR model. Quantitative Finance, 17(10), pp.1549-1565, 2017.

[53] K. Levenberg. A Method for the Solution of Certain Non-Linear Problems in Least Squares.
Quarterly of Applied Mathematics. 2: pp. 164-168, 1944.
[54] D. Marquardt. An Algorithm for Least-Squares Estimation of Nonlinear Parameters. SIAM
Journal on Applied Mathematics. 11 (2): pp. 431-441,1963.‘

[55] W. A. McGhee. An artificial neural network representation of the SABR stochastic volatility
model. Preprint, SSRN:3288882, 2018.
[56] H.N. Mhaskar. Approximation properties of a multilayered feedforward artificial neural net-
work. Advances in Computational Mathematics, 1(1): 61-80, 1993.

31
[57] J. A. Nelder and R. Mead. A simplex method for function minimization. Computer Journal.
7: pp. 308-313, 1965.
[58] J. Nocedal and S. Wright. Numerical Optimization. Springer Series in Operations Research
and Financial Engineering. Springer-Verlag New York, 2006.

[59] D. Pedamonti. Comparison of non-linear activation functions for deep neural networks on
MNIST classification task. Preprint, arXiv:1804.02763
[60] M. J. D. Powell. A direct search optimization method that models the objective and constraint
functions by linear interpolation. Advances in Optimization and Numerical Analysis, eds. S.
Gomez and J-P. Hennart, Kluwer Academic (Dordrecht), pp. 51-67, 1994.

[61] L. M. Rios and N. V. Sahinidis. Derivative-free optimization: a review of algorithms and

comparison of software implementations. Journal of Global Optimization. Vol. 56: 3 pp. 1247-
1293, 2013.
[62] M. Sabate Vidales, D. Siska and L. Szpruch. Martingale Functional Control variates via Deep
Learning Preprint 1810.05094, 2018.

[63] N. Ruiz, S. Schulter and M. Chandraker, Learning to simulate, Preprint arXiv:1810.02513,

2018.
[64] L. Setayeshgar and H. Wang. Large deviations for a feed-forward network, Advances in Applied
Probability, 43: 2, pp. 545571, 2011.

[65] U. Shaham, A. Cloninger, and R. R. Coifman. Provable approximation properties for deep
neural networks. Appl. Comput. Harmon. Anal., 44(3): 537-557, 2018.
[66] J. Sirignano and K. Spiliopoulos. Stochastic Gradient Descent in Continuous Time: A Central
Limit Theorem. SIAM J. Finan. Math., 8(1), pp. 933-961, 2017.

[67] H. Stone. Calibrating rough volatility models: a convolutional neural network approach.
Preprint, arXiv:1812.05315, 2018.
[68] R. Storn and K. Price. A Simple and Efficient Heuristic for global Optimization over Continuous
Spaces. Journal of Global Optimization. Vol.11:4, pp341-359, 1997.
[69] V. N. Vapnik. Statistical Learning Theory. Wiley-Interscience, 1998.

[70] C. Zhu, R. H. Byrd and J. Nocedal. L-BFGS-B: Algorithm 778: L-BFGS-B, FORTRAN rou-
tines for large scale bound constrained optimization. ACM Transactions on Mathematical Soft-
ware, 23: 4, pp. 550-560, 1997.

32