Volatility Models in Option Pricing

Miguel Ribeiro
[email protected]

Instituto Superior Técnico, Lisboa, Portugal

Month 2018

Abstract
Volatility is one of the most important subjects in all of quantitative finance, due not only to
its impact on the prices of options but also to its elusiveness. In this thesis we study some of the
models most used to forecast this variable, namely Dupire’s local volatility as well as Heston and
Static/Dynamic SABR stochastic volatility models. We train these models with some options’ implied
volatility data, making them able to replicate real market behavior. We find that, when dealing
with options with a single maturity, the Static SABR model is the one that best fits the data, while
with multiple maturities, the Heston model outperforms Dynamic SABR. All these models vastly
outperform the constant volatility model, assumed in Black-Scholes. We then use these trained models
to price European and Barrier options with the Monte Carlo numerical pricing method, which is able
to accurately predict implied volatilities for near-the-money options, failing for deep in-the-money
European call options.
Keywords: Volatility, Option pricing, Dupire, Heston, Static SABR, Dynamic SABR

1. Introduction lowing the option to expire without further losses.

Derivatives are currently one of the most studied Due to their importance, finding the ideal price
subjects in all of mathematical finance. In finance, a of an option is a major concern to investors, though
derivative is simply a contract whose value depends this can be very difficult for some option types.
on other simpler financial instruments, known as
1.2. Volatility
underlying assets, such as stock prices or interest
Volatility is a parameter that is used in all pricing
rates. These contracts are currently responsible for
models and that greatly affects option prices. De-
over $542 trillion worth of trades, in the Over-the-
spite its importance and our great efforts to study
Counter (OTC) market alone [Bank for Interna-
it, this parameter remains as one of the most elusive
tional Settlements, 2018]. We can thus see that
phenomena in all of quantitative finance, mainly
fully understanding the behavior of derivatives is
due to our inability not only to forecast its behav-
crucial to investors.
ior but even actually to accurately measure it.
1.1. Options Many models have been developed throughout
the years in an attempt to model this parameter,
Of all classes of derivatives, in this master thesis we
with varying degrees of success. We will analyze
focus particularly on the most traded type [Hull,
in detail some of the most famous ones, comparing
2009]: options. A (European) option contract
them with one another.
grants its buyer the option to buy (in the case of
a call type option) or sell (for put options) its un- 2. Background
derlying asset, referred to as stock, at a future date, We begin by providing some of the financial back-
known as the maturity, for a fixed price, known as ground required to understand the later results.
the strike price. There are other option types, such
as Barrier options, that behave similarly to Euro- 2.1. Option payoffs and prices
pean options with the condition that they only be- From the definition of a European option we can
come valid if the stock price increases past a given deduce the payoff function of this contract as
threshold at any point until the maturity. PayoffEuro, call (K, T ) = max (S(T ) − K, 0) ; (1a)
It’s important to emphasize the fact that an op- PayoffEuro, put (K, T ) = max (K − S(T ), 0) , (1b)
tion grants its buyer the right to do something. If
exercising the option would lead to losses, the buyer where K is the option’s strike price and S(T ) is the
can simply decide to let the maturity date pass, al- asset’s price, S(t), at the maturity, T .

1
 An up-and-in Barrier option’s payoff function is To price options, we need to solve the PDE in
exactly as shown in eq.(1) conditioned on a given eq.(3) as we would for the diffusion equation’s initial
threshold B being surpassed at any point until the value problem [Dilão et al., 2009], resulting in
maturity (i.e. ∃ t < T : S(t) > B). It should be
C(K, T ) = N (d1 )S0 − N (d2 )Ke−rT ; (5a)
zero otherwise. The influence of this threshold B
−rT
on the option price is trivial: because it is linked to P (K, T ) = −N (−d1 )S0 + N (−d2 )Ke , (5b)
the probability of the stock price reaching B, the where N (·) is the cumulative distribution function
higher the barrier B, the lower the corresponding of the standard normal distribution and where d1 ,
option price. d2 are given by
We can price options by setting their expected
σ2
     
profit to be the same as a risk-neutral investment, 1 S0
d1 = √ log + r+ T ; (6a)
(e.g. bank deposit). The price of an option can σ T K 2
√
thus be deduced as it’s expected future payoff, dis- d2 = d1 − σ T . (6b)
counted back to the present
In Figure 2 we represent the prices of call and
Price(K, T ) = e−rT E [Payoff(K, T )] , (2) put options at both the inception (i.e. t = 0) and
where r is the risk-free interest rate, defined as the maturity as a function of the ratio K/S0 .
interest an investor would receive from any risk-free 0.6
investment (e.g. treasury bills). In general, this rate Call (at Inception)
0.5 Put (at Inception)
changes slightly with time and is unknown. Call (at Maturity)
Put (at Maturity)
0.4
2.2. Black-Scholes Formulae
Fischer Black and Myron Scholes developed a math- 0.3

ematical model to price European options - the fa-

0.2
mous Black-Scholes (BS) model [Black and Scholes,
1973] - still in use in present days [Wilmott, 2006]. 0.1

This model states that the price of a European 0

option follows the partial differential equation 0.4 0.6 0.8 1 1.2 1.4 1.6
K/S 0
2
∂V 1 ∂ V ∂V
+ σ 2 S 2 2 + rS − rV = 0, (3)
∂t 2 ∂S ∂S Figure 2: Call and Put option values at inception
where V is the price of the option and σ is the and maturity
stock price volatility, which affects how erratically
the stock price moves. The interest rate, r, is as- We can thus use eqs.(5) to precisely price Eu-
sumed to be a known function of time, and the ropean options, so long as all the parameters are
volatility, σ, is assumed to be a known constant. exactly known and remain constant throughout the
Under this model, we assume that stock prices option’s duration (which is never true).
follow a Geometric Brownian Motion (GBM), de-
fined as 3. Volatility
Volatility is a measure of the uncertainty in future
dS(t) = rS(t)dt + σS(t)dW (t), (4) stock price movements - a higher volatility will lead
with {W (t), t > 0} defining a one-dimensional to greater future fluctuations in the stock price,
Brownian motion. An example of such processes whereas a stock with lower volatility is more sta-
is represented in Figure 1. ble. This influence is represented in Figure 3.
1.15
1.2
1.1

1.1
1.05

1 1

0.95 0.9
=0.05 yr -0.5

0.9 =0.1 yr -0.5

0.8 =0.2 yr -0.5
0.85
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time(yr) Time(yr)

Figure 1: Example of three realizations of a Geo- Figure 3: Example of three realizations of GBM
metric Brownian Motion process. processes with different volatilities

2
 Of all the parameters in the BS model, volatility It can be shown that the relation between im-
is the only one we can’t easily estimate or predict, plied volatility and option price is a monotonous
even though it has a major impact on the prices of increasing function (i.e. the relationship is bijec-
options. tive), which means that we can obtain the im-
It is usually estimated empirically from the plied volatility of an option from its price and vice
standard deviation of the historical rate of log- versa [Wilmott, 2006]. Thus, we can train our mod-
returns [Hull, 2009], defined as els on implied volatility data instead of using option
v prices, since they are equivalent.
n
One important property of implied volatility is
u
u 1 X
s=t (ui − u)2 , (7) that, when observing real market data, we can
n − 1 i=1
clearly see that it depends on the strike price, which
where the log-return rate, ui , is given by is incompatible with the BS view (which assumes it
  is independent). This phenomenon is known as the
Si implied volatility smile and is represented in Fig-
ui = log , (8)
Si−1 ure 4 (a skew can also be observed for some types
of underlying assets).
with u defining its average and Si corresponding
to the stock price at the ith measurement of some
given set of past observations.
Using this result, we are able to estimate the
volatility of any given asset at the present moment
Implied Volatility
and use it in the BS model, if we assume it remains
constant in the future. The clear problem with this
approach is that, when observing market data, we
can see that volatilities change over time [Chour-
dakis, 2008]. If we try to price options assuming
a constant volatility, our options will become mis-
priced, causing potential losses. 1.0

Our goal is to model the instantaneous volatil- K/S 0

ity (i.e. the volatility at a given point in time)

and, with this model, predict its future behav- Figure 4: Implied volatility smile.
ior, using this knowledge to better price options.
Many models have been developed in the past, so It can also be shown that the implied volatility
we will only study some of the most used ones, decreases with maturity, though this relationship is
namely Dupire’s formula, where we assume that harder to define.
the volatility depends on the stock price and time
3.2. Option Price Sensitivity and Vega
(i.e. σ = σ(t, S(t))), and three other models, Hes-
When studying volatilities, it’s very important to
ton and Static/Dynamic SABR, where the volatil-
consider the sensitivity of the option price to the
ity itself is assumed to be a stochastic process (i.e.
volatility, i.e. how a small variation in the volatility
σ = σ(t, S(t), W (t))).
of a given option affects its price. This analysis is
3.1. Implied Volatility particularly important for volatilities because there
To fully understand the results shown next we need is a high uncertainty associated with this parameter
the concept of implied volatility. Implied volatility and a high sensitivity might lead to severe mispric-
can be defined as the value of stock price volatility ing errors.
that, when input into the BS pricer in eq.(5), out- This sensitivity is usually called Vega, or V, and
puts a price equal to the market price of a given is defined as
∂V
option. V= , (10)
Because eq.(5) is not explicitly invertible w.r.t. σ, ∂σ
we need to use some numerical method (e.g. New- where V denotes the option price. In particular for
ton’s method) to find the value of implied volatil- European calls, this value can be shown to be [Hull,
ity that matches market with model prices, i.e. we 2009] √
must find, numerically, the solution to the equation V = S0 T N 0 (d1 ), (11)

C(σimp ) = Cmkt , (9) where d1 is given in eqs.(6) and N 0 (·) is the prob-
ability density function for a standard normal dis-
where C(σimp ) is the BS option price using σimp tribution.
as volatility and Cmkt is the price observed in the Despite its usefulness, the Vega doesn’t grasp the
market. whole picture e.g. a Vega of 4 means that an abso-

3
lute change of 2 in the volatility produces an abso- the volatility). The opposite effect is observed for
lute change of 8 in the option price - we don’t have call options with lower strikes, since we can see that
any information regarding the relative change of the the relative variation of the option price w.r.t. the
price (i.e. if it changed by 1% or 50%). As we will volatility is extremely small in these cases, meaning
see shortly, this information is indeed important. not only that the option price is extremely robust
We thus define the relative change as to the volatility (i.e. a change in the volatility will
∂V σ barely affect the option price), but also implying
Relative Change = , (12) that the volatility is extremely sensitive to the op-
∂σ V
tion price (i.e. a very slight (relative) change in the
where now a Relative Change of 4 implies that a price of a call option will dramatically change its
change of 2% in the volatility produces a variation volatility). This phenomenon will become impor-
of 8% in the option price. This relationship is plot- tant when we examine our results.
ted in Figure 5, for call options, against their strike
price. 3.3. Volatility Models
As previously stated volatility not only changes
35
T=63 days
with time but is also dependent on the strike price.
30 T=42 days The (constant volatility) BS model is therefore
T=21 days
25 clearly insufficient to completely grasp real-world
Relative Change

trading and we should try to find some more ap-

20
propriate volatility models.
15
3.3.1. Dupire’s Local Volatility
10 One of the most used volatility models was devel-
5 oped by Dupire [Dupire, 1994] and uses the concept
0
of local volatility, where we assume that volatility is
0.4 0.6 0.8 1 1.2 1.4 1.6 a function of both time and stock price: σ(S(t), t).
K/S 0
The stock price process now follows the diffusion
process
Figure 5: Relationship between the (call) option
price’s relative change (w.r.t. volatility) and the dS(t) = rS(t)dt + σ(S(t), t)S(t)dW (t), (13)
respective strike prices, for different maturities.
where the local volatility, σ(S(t), t), is a nonlinear
As we can see in Figure 5, the relative change deterministic function of S(t) and t.
of the option price w.r.t. volatility is very large for To estimate σ(S(t), t), we must have some im-
call options with high strikes, which means that the plied volatility data for options with multiple strikes
prices of such options are very sensitive to volatility over multiple maturities. From this data we extract
(i.e. a very slight (relative) change in the volatility the implied volatility surface, σimp (S(t), t), and its
will produce a very large (relative) change in the op- respective gradients w.r.t K and T , which we can
tion price). It also means that the volatility is very use to generate σ(S(t), t) using
robust w.r.t. the option price (i.e. a very large (rel-
ative) variation in the option price will barely affect
v
∂σimp ∂σimp
u
2
σimp + 2tσimp + 2r(S(t))tσimp
u
u
σ(S(t), t) = u ∂T ∂K
2 ! , (14)
√ ∂σimp 2 √
u  2

u 2 ∂ σ imp ∂σimp
t 1 + (S(t))d1 t + (S(t)) tσimp − d1 t
∂K ∂K 2 ∂K

with d1 given by price options, as we will see shortly.

1 2


log(S0 /S(t)) + r + 2 σimp t
d1 = √ . (15) 3.3.2. Heston’s Stochastic Volatility
σimp t
Volatility is not constant, is not directly observable
where we define σimp = σimp (K, T ) as the implied and is unpredictable. This seems to suggest that
volatilities of options with maturity T , and strike volatility is itself also a stochastic process [Rebon-
K. Furthermore, σimp and all its derivatives are ato, 2004].
evaluated at K = S(t) and T = t.
With eq.(14) we generate a local volatility sur- The Heston model, developed by Steven Hes-
face, which we can use in eq.(13) to create the re- ton [Heston, 1993], is one of the most used stochas-
spective stock price path. From this, we are able to tic volatility models, and it states that stock prices

4
satisfy the relations (i.e. the mean value of variance) and the volatil-
p ity of the variance (i.e. how erratic is the variance
dS(t) = rS(t)dt + ν(t)S(t)dW1 (t), (16) process).
To appropriately use the model, we have to find
the values for the parameter set θ = {κ, ν, η, ν0 , ρ}
p
dν(t) = κ(ν − ν(t))dt + η ν(t)dW2 (t), (17)
that best fit market data. For many models, this
with ν(t) corresponding to the stock price variance calibration process requires lengthy simulations and
(i.e. the square of the volatility, ν(t) = (σ(t))2 ) and is impractically slow to converge. One of the rea-
where we define ν0 as the initial variance. Further- sons why the Heston model is so popular is the
more, the one-dimensional Brownian motion pro- fact that there exists a closed-form solution that we
cesses W1 (t) and W2 (t) have a constant correlation can use to directly obtain the option prices under
ρ, typically negative, which can be justified from this model with any given parameter set θ. This
market behavior [Chourdakis, 2008]. closed form solution (with modifications made by
The parameters κ, ν and η are, respectively, the Schoutens [Schoutens et al., 2004], Rollin et al. [del
mean-reversion rate (i.e. how fast the variance con- Baño Rollin et al., 2010] and Cui et al. [Cui et al.,
verges to its mean value), the long-term variance 2017]) is given by

CH (K, T ; θ) = e−rT E (S(T ) − K) 1{S(T )>K} | θ

 

= e−rT E S(T )1{S(T )>K} | θ − KE 1{S(T )>K} | θ

   

(18)
−rT
= S0 P1 (K, T ; θ) − e KP0 (K, T ; θ),
∞
e−iu log K
Z  
1 1
P1 (K, T ; θ) = + Re φ(u − i, T ; θ) du, (19)
2 π 0 iuS0 erT
1 ∞
Z  −iu log K 
1 e
P0 (K, T ; θ) = + Re φ(u, T ; θ) du, (20)
2 π 0 iu
 
tκνρiu 2κν
φ(u, t; θ) = exp iu (log S0 + rt) − − ν0 A + 2 D , (21)
η η
 
(κ − α)t α+ξ α − ξ −αt
D = log α + − log + e , (22)
2 2 2

to distinguish it from the Dynamic SABR model

A1 shown next). Under this model we assume that the
A= , (23)
A2 stock price and volatility processes follow [Vlaming,
ξ = κ − ηρiu, (24) 2011]
p
α = ξ 2 + η 2 (u2 + iu), (25) dS(t) = rS(t)dt + e−r(T −t)(1−β) σ(t)(S(t))β dW1 (t),
αt (28)
A1 = (u2 + iu) sinh , (26)
2 dσ(t) = νσ(t)dW2 (t), (29)
αt αt
A2 = α cosh + ξ sinh . (27) where α = σ(0) and, as before, the two Brownian
2 2 motion processes W1 (t) and W2 (t) have a constant
where CH (K, T ; θ) corresponds to the model’s Eu- correlation of ρ.
ropean call option price, assuming a parameter set The parameters β and ν correspond, respectively
θ, i is the imaginary unit and where φ(u, t; θ) is to the skewness (i.e. how the volatility smile moves
the characteristic function of the logarithm of the when the stock price changes) and the volatility of
stock price process (the characteristic function cor- volatility (i.e. how erratic is the volatility process).
responds to the Fourier transform of the probability As for the Heston model, in Static SABR we also
density function of a random variable). With this have a (quasi-)closed form solution from which we
result we can easily calibrate the model by minimiz- can extract the option’s implied volatility for any
ing the difference between model and market prices. given parameter set (the respective price can be ob-
3.3.3. Static SABR Stochastic Volatility tained with eq.(9)). This formula is given by (with
One other very commonly used stochastic volatility modifications made by Oblój [Obloj, 2008])
model was developed by Hagan et al. [Hagan et al.,
2002] and is known as SABR (short for stochastic-
αβρ) (we henceforth refer to it as Static SABR

5
  
1 ν log (f /K)
σStatSABR (K, f, T ) ≈  .
(1 − β)2 (1 − β)4
   
f f x(z)
1+ log2 + log4
24 K 1920 K (30)
(1 − β)2 α2 2 − 3ρ2 2
  
1 ρβνα
. 1+T + + ν ,
24 (Kf )1−β 4 (Kf )(1−β)/2 24

with z and x(z) defined as the Static SABR model but where we replace the

parameters ν and ρ with some functions of time,
ν f 1−β − K 1−β
ν(t) and ρ(t), chosen empirically.
z= , (31)
α(1 − β) Fernandez et al. [Fernández et al., 2013] claimed
(p ) that ν(t) and ρ(t) should tend to zero with time.
1 − 2ρz + z 2 + z − ρ In this work, we will use the functions suggested by
x(z) = log , (32) those authors
1−ρ
ρ(t) = ρ0 e−at , (33)
using f = S0 erT .
ν(t) = ν0 e−bt , (34)
We again need to find the optimal values for the
parameters that best fit the market data. with ρ0 ∈ [−1, 1], ν0 > 0, a > 0 and b > 0.
3.3.4. Dynamic SABR Stochastic Volatility For this particular choice of ν(t) and ρ(t), the Dy-
One of the main setbacks of the Static SABR model namic SABR model also has a (quasi-)closed form
is the fact that it behaves badly when we try to solution, which we can use to directly obtain the op-
fit options with different maturities [Hagan et al., tion’s implied volatility with any given parameter
2002]. To solve this, the authors developed an alter- set. This formula (with some modifications made
native model, known as Dynamic SABR, similar to by Osajima [Osajima, 2007]) is given by

     
1 K K
σDynSABR (K, f, T ) = 1 + A1 (T ) log + A2 (T ) log2 + B(T )T , (35)
ω f f
β − 1 η1 (T )ω
A1 (T ) = + , (36)
2 2
(1 − β)2 1 − β − η1 (T )ω 4ν12 (T ) + 3(η22 (T ) − 3η12 (T )) 2
A2 (T ) = + + ω , (37)
12 4 24
1 (1 − β)2 ωβη1 (T ) 2ν22 (T ) − 3η22 (T ) 2
 
B(T ) = 2 + + ω , (38)
ω 24 4 24
2 2
  
6ν0 (2bT ) −2bT
ν12 (T ) = − 2bT + 1 − e , (39)
(2bT )3 2
12ν02  −2bT
ν22 (T ) =

3
e (1 + bT ) + bT − 1 , (40)
(2bT )
2ν0 ρ0 h −(a+b)T
i
η1 (T ) = 2 (a + b)T + e − 1 , (41)
T (a + b)2
3ν 2 ρ2 h i
η22 (T ) = 4 0 0 4 e−2(a+b)T − 8e−(a+b)T + (7 + 2T (a + b)(−3 + (a + b)T )) . (42)
T (a + b)

where f = S0 erT and ω = f 1−β /α. Starting with Dupire’s local volatility, we applied
a Delaunay triangulation on the data to produce the
4. Implementation implied volatility surface, from which we extracted
4.1. Model Training
the gradients required for eq.(14). Having obtained
To use the models for predicting future option prices
the local volatility surface, we can obtain the op-
we first need to train them on some real market
tion’s implied volatility with a numerical method,
data. We had access to some implied volatility data
as we will explain shortly.
for options with an index as underlying asset, with
7 different strike prices over maturities of 1, 2, 3 Regarding the stochastic volatility models (He-
and 6 months. ston and Static/Dynamic SABR), we trained our

6
models on the aforementioned data using their re- 5. Results
spective closed form solutions. This calibration was We are now able to train the models on the afore-
done by minimizing the distance between the data mentioned data using the implementation methods
and the model predictions, measuring it with the described before.
cost function
n X m We begin by showing in Figure 6 the local volatil-
X
Cost(θ) = wi,j (σimp,mkt (Ti , Kj )− ity surface resulting from the Dupire’s local volatil-
i=1 j=1 (43) ity model, which we obtained by interpolating the
implied volatility data and applying Dupire’s for-
−σimp,mdl (Ti , Kj ; θ))2 ,
mula. This surface can then be used in a Monte
where σimp,mkt (·) and σimp,mdl (·) correspond to the Carlo pricer, as we explained before.
real-market and the model’s implied volatilities, re-
spectively, for maturities {Ti , i = 1, . . . , n} and
strikes {Kj , j = 1, . . . , m} and where we defined
the weight function, wi,j , as (assuming that the
strikes are restricted to K < 2S0 )
 2 1
Kj
wi,j = 1 − 1 − , (44)
S0

(yr -1/2 )
such that a higher weight is given to the prices close 0.5

loc
to the starting price, S0 .
126
Finally, to find the parameters that minimize the 105
0
aforementioned distance, we have to use an opti- 84
0.5
mization algorithm, able to deal with the nonlinear- 63
s)
ities of the cost function. We chose an algorithm de- 1
42 ay
(d
K/S 1.5
T
veloped by Hansen [Hansen, 2006] known as CMA- 0 21

ES. Other algorithms were tested, but CMA-ES

performed best. Figure 6: Local volatility surface obtained with
Dupire’s formula
4.2. Numerical Option Pricing
Having trained all the models on some real mar-
ket data, we should now be able to price any op-
Regarding now the stochastic volatility models,
tions, European or not. To achieve this we chose
we calibrated them with the cost function described
the Monte Carlo numerical pricing algorithm, a ver-
before using the CMA-ES optimization algorithm.
satile but powerful method.
The calibrated parameters for these models are rep-
The Monte Carlo algorithm consists of simulat-
resented in Tables 1-3, with the resulting costs in
ing a very large number of stock price paths, using
Table 4.
any of the mentioned models, and then calculate the
option’s payoff for each of the stock prices. Averag- We should note that the Static SABR model was
ing the payoffs and discounting them to the present trained on data for each maturity independently,
should provide a fairly good estimate of the option’s whereas the Heston and Dynamic SABR models
value. were trained on all maturities together dependently,
Because the Brownian motion is a self-similar as an ensemble. In other words, the Static SABR
process, to simulate the stock prices we need first model was trained 4 times (once for each maturity)
to discretize their diffusion process [Mikosch, 1998]. on a data set of 7 implied volatilities (with different
For the constant volatility model and Dupire’s local strikes), whereas Heston and Dynamic SABR were
volatility we used the Euler–Maruyama discretiza- trained on a data set of 28 (7×4) implied volatilities,
tion method, meaning that the stock price process disregarding maturity. For this reason, in Table 4,
follows for the Static SABR model we show the sum of the
costs of calibration on each maturity, ΣCosts .
S(t + ∆t) =S(t) + rS(t)∆t+
√ (45) To have a benchmark for how well each model
+ σ(S(t), t)S(t) ∆tZ(t),
performs, we also trained a constant volatility
where Z(t) ∼ N (0, 1) defines a normal distributed model twice, with the calibration done indepen-
random variable and ∆t a (small) subinterval of dently for each maturity, as for Static SABR, and
the whole time to maturity. As for the stochastic dependently, like Heston. Because the calibrated
volatility models we used the more robust Milstein constant volatilities are not relevant for our results,
discretization method for both the stock price and they are not shown here, though we represent their
the volatility processes. costs in Table 4.

7
 T (days) α(yr−0.5 ) β ρ ν(yr−0.5 ) model (with dependent fits), which suggests that
21 0.24 0.38 -0.38 2.10 this model fits the data very well, without the over-
42 0.24 0.74 -0.37 1.45 fitting problem from Static SABR. The parameter κ
63 0.24 0.78 -0.31 1.14 is very large, which means that the variance process
126 0.23 0.88 -0.24 0.82 (ν(t) = (σ(t))2 ) tends to its mean value, ν, very fast
and that ν0 has almost no influence. Furthermore ρ
Table 1: Fitted parameters for each maturity (fitted is negative, as expected, and η is very large meaning
independently) under Static SABR model. that the variance process will be quite erratic.
Considering the Dynamic SABR model, the im-
provement of the cost value over the constant
κ(yr−1 ) ν(yr−1 ) ν0 (yr−1 ) ρ η(yr−1 ) volatility model (with dependent fits) is now (only)
53.44 0.07 0.10 -0.41 6.26 91.3%. The parameter a being 0 implies that the
function ρ(t) is stuck at ρ0 , and the parameter b be-
Table 2: Fitted parameters for all maturities (fitted ing very large means that ν(t) goes to 0 extremely
simultaneously) under the Heston model. fast. These results are very inconsistent with what
is observed in the Static SABR model, where the ρ
and ν tend (slowly) to zero with time. This seems
to suggest that the functions chosen to model these
α(yr−0.5 ) β ρ0 a(yr−1 ) ν0 (yr−0.5 ) b(yr−1 ) two parameters were not appropriate.
0.25 0.63 -0.42 0 1.87 41.69 Finally, comparing all the stochastic volatility
models, we can say that even though the Static
Table 3: Fitted parameters for all maturities (fitted SABR model presented the lowest cost, because of
simultaneously) under the Dynamic SABR model. the mentioned overfitting, the Heston model might
be the one that performs best among the three.
Having trained the models we are now able to
Model Cost ΣCosts simulate the results to produce forecasts and com-
Constant Vol. (indep.) - 0.1150 pare the models’ simulations with one another. The
Constant Vol. (dep.) 0.1248 - simulations were performed for all four maturities
Dupire - - mentioned before, but, due to redundancy, we only
Static SABR - 0.0008 show the plots for the first maturity of 1 month.
Heston 0.0025 -
Dynamic SABR 0.0108 Regarding Dupire’s local volatility model, having
obtained the local volatility surface we are able to
Table 4: Comparison between the costs from the simulate the stock price paths with the discretiza-
calibrated stochastic volatility models. tion method described before. From these, we can
extract the implied volatilities for many different
strikes (converting them from option prices). These
We now analyze the calibration results presented
simulations are performed 100 times for each strike
before for the stochastic volatility models.
and then averaged to produce the simulated func-
Starting with Static SABR, we first note that the
tion, as shown in Figure 7(a). The simulations don’t
cost after calibration is extremely low, showing an
always produce the exact same results and some
improvement of 99.3% over the cost of the constant
variation might occur. To represent this variation
volatility model (with independent fits). This seems
we also show the 95% confidence bands of the sim-
to suggest that the model fits the data almost per-
ulations in the plot (i.e. 95% of all observations are
fectly, though this conclusion should be taken care-
contained within these bands). In the same plot
fully: we have an extremely small amount of data
we also represent the implied volatility data for the
points (7 for each maturity) for the comparatively
first maturity.
large number of parameters used (4 parameters in
Considering the stochastic volatility models, we
total). This will cause our model to overfit the data,
show the same information as for Dupire’s model
explaining our very low costs. To further corrobo-
but we also include the calibrated closed form solu-
rate this hypothesis, we note that the parameter β,
tions, which we call theoretical functions.
which is expected to remain constant, varies wildly
To run the Monte Carlo pricers we used an initial
between maturities. Finally, we note that, as ex-
stock price of S0 = 1e, a risk-free interest rate r =
pected, the parameter ρ is always negative and that
0, a time step size ∆t = 0.5 days and simulated a
both ρ and ν seem to tend to zero with time, vali-
total of 100 000 paths.
dating the hypothesis from Fernandez et al.
Regarding now the Heston model, we first em-
phasize the low cost value after calibration, with an
improvement of 98.0% over the constant volatility

8
 We now analyze the results shown in the previous
plots.
We begin by noting that in the regions with
strikes around S0 all simulations follow the closed
form solutions very closely, with almost no varia-
tion, which seems to suggest that the implementa-
tion was done correctly for all models. Furthermore,
all models match the implied volatility data quite
well in this region.
Regarding the simulated functions, we can see
that they decrease significantly for large strikes,
presenting also a large variation (i.e. wide confi-
(a) Dupire’s model
dence bands). This can be explained by the low
number of stock price paths that are able to reach
such high strikes in such a short maturity, which
causes the resulting option prices to be too depen-
dent on only a few simulations, which explains the
variation. If we increase the maturity or the number
of simulations this phenomenon disappears, since
more paths are able to reach the high strikes and
contribute to the option price.
Finally, to explain the wide confidence bands for
the low strikes on all models we require the concept
of relative change introduced before. We noted pre-
viously that for low strikes the implied volatility is
extremely sensitive to the option price. Thus, be-
(b) Static SABR cause we are simulating option prices and only then
converting them to implied volatilities, even a very
slight variation in the simulations will produce a
a slightly different option price which is converted
into an entirely different implied volatility, justify-
ing the wide confidence bands.
With the Monte Carlo pricers working properly,
we should be able to price any options, European or
not, by adapting our algorithm. We priced several
Barrier options with different barrier levels under
each of the models. Due to redundancy, in Fig-
ure 8 we only represent Barrier option prices with
different barrier levels using the Heston model. For
comparison, we also show the prices of the corre-
(c) Heston model sponding European option.

0.6
Heston
0.5 European

0.4

0.3

0.2

0.1

0
0.4 0.6 0.8 1 1.2 1.4 1.6
K/S 0

(d) Dynamic SABR Figure 8: Barrier option prices with different barrier
levels and corresponding European option prices
Figure 7: Simulations and closed-form solutions un- under the Heston model.
der each model
9
 As we can see, the prices of Barrier options are Y. Cui et al. Full and fast calibration of the heston
heavily dependent on their barrier levels, decreas- stochastic volatility model. European Journal of
ing when we increase this threshold. They are also Operational Research, 263(2):625 – 638, 2017.
strictly lower than the corresponding European op-
tion price, which is unsurprising. We can thus con- S. del Baño Rollin et al. On the density of log-
clude that the results are what would be expected. spot in the heston volatility model. Stochastic
Processes and their Applications, 120(10):2037 –
6. Conclusions 2063, 2010.
Volatility is one of the most important subjects in
all of quantitative finance, due not only to its im- R. Dilão, J. A. de Matos, and B. Ferreira. On the
pact on the prices of options but also to its elusive- value of european options on a stock paying a
ness. In this thesis we studied some of the models discrete dividend. Journal of Modelling in Man-
most used to forecast this variable. agement, 4(3):235–248, 2009.
We began by studying Dupire’s local volatility B. Dupire. Pricing with a smile. Risk Magazine,
model as well as Heston and Static/Dynamic SABR pages 18–20, 1994.
stochastic volatility models. We used some real im-
plied volatility data to train the models: we gener- J. Fernández et al. Static and dynamic sabr stochas-
ated the local volatility surface for Dupire’s model tic volatility models: Calibration and option pric-
and calibrated all the parameters for the stochastic ing using gpus. Mathematics and Computers in
volatility models using their closed form solutions. Simulation, 94:55 – 75, 2013.
From this calibration we concluded that the Static
P. Hagan et al. Managing smile risk. Wilmott Mag-
SABR model best fit the data, though some over-
azine, 1:84–108, 01 2002.
fitting is expected to have occurred, for which rea-
son we consider the Heston model to perform best. N. Hansen. The cma evolution strategy: a com-
All models vastly outperform the constant volatility paring review. In Towards a new evolutionary
model. computation, pages 75–102. Springer, 2006.
Having trained all models, we input them into a
numerical pricer, using the Monte Carlo method to S. Heston. A closed-form solution for options with
estimate the option prices under each model. All stochastic volatility with applications to bond
simulations followed the data very closely for op- and currency options. 6:327–43, 02 1993.
tions with strikes around S0 , though some signifi-
J. Hull. Options, Futures, and Other Derivatives.
cant variation was found in the results for very low
Prentice Hall, 2009. ISBN 9780136015864.
strikes and for high strikes with early maturities.
In these cases great care should be employed when T. Mikosch. Elementary Stochastic Calculus with
using these models. Finance in View. Advanced Series on Sta-
Finally, we adapted our Monte Carlo pricers to tistical Science and Applied Probability. World
price Barrier options and observed that the option Scientific Publishing Company, 1998. ISBN
prices decreased when we increased the barrier lev- 9789813105294.
els and remained strictly below the corresponding
European option prices. Both observations were ex- J. Obloj. Fine-tune your smile: Correction to hagan
pected. et al. Mar 2008.
Regarding future work, one clear possibility is to Y. Osajima. The asymptotic expansion formula of
improve the Monte Carlo pricer, using, for exam- implied volatility for dynamic sabr model and fx
ple, importance sampling or the antithetic variates hybrid model. 2007.
method. Some different functions for ρ(t) and ν(t)
could also be considered, besides the ones we used. R. Rebonato. Volatility and correlation: the per-
fect hedger and the fox. Wiley, 2004. ISBN
References 9780470091395.
Bank for International Settlements. Semiannual otc
derivatives statistics. http://stats.bis.org/ W. Schoutens, E. Simons, and J. Tistaert. A perfect
statx/srs/table/d5.1, 2018. (Accessed: 2018- calibration! now what? Wilmott Magazine, pages
05-20). 66–78, 2004.
F. Black and M. Scholes. The pricing of options and G. Vlaming. Pricing options with the sabr model.
corporate liabilities. Journal of political economy, Master’s thesis, 2011.
81(3):637–654, 1973.
P. Wilmott. Paul Wilmott on Quantitative Finance.
K. Chourdakis. Financial engineering: A brief in- The Wiley Finance Series. Wiley, 2006. ISBN
troduction using the matlab system, 2008. 9780470065372.

10