Sie sind auf Seite 1von 12

A Short Note on Volatility Models

May 20, 2013

Didier Kouokap Youmbi

(didier.kouokap@gmail.com)

This document is a short summary regarding the evolution of the volatility


models from Black and Scholes to the Local-Stochastic Volatility models. We
show advantages as drawbacks linked to each model, and how the community
has moved from one model to another in order to overcome drawbacks.

1 The Model: Dynamics of the underlying asset


The volatility here is the instantaneous standard deviation of the underlyings
return. The equity underlying is modelled as follow (without jump):
dSt
= t dt + t dWt
St
(Wt )t>0 is a brownian motion;
t is the drift term (the instantaneous expectation of the underlyings
return);

t is the volatility.

Up to now and as far as we know, the industry has experimented 5 classes of


volatility models: the constant volatility model also called Black and Scholes;
the local volatility or Dupire model; the uncertain volatility or Avellaneda model;
the stochastic volatility models; and the mixed local-stochastic volatility (LSV)
model.

2 What are we expecting from the volatility mod-


els?
From the pricing theory we get that a fair price for a derivative product only
exists if the market is complete (i.e there is a strategy for replicating the product,

1
using liquid assets as the equity underlying). This is equivalent to say that there
is a risk-neutral measure (probability), under which the drift of any tradable or
replicable asset or strategy is the risk-free rate: t = rt .
Pricing is done under the risk-neutral probability, this implies that the drift
term is the risk-free short interest rate and is known. Therefore, the only degree
of freedom we have to drive the equity underlying is the volatility. Thats why
the volatility modelling is so crucial.
When modelling the volatility, our first requirement will be to match the so
called vanilla options (call and put in equity world). Matching the call and put
prices for all the maturities will allow us to replicate the forward on the equity
underlying, as the call and put prices are linked to the equity forward through
the so called call-put parity formula:

Call(t, T, K) P ut(t, T, K) = DF (t, T ) (Ft,T K)


K is the options strike price (payoff is (ST K)+ for call option and
(K ST )+ for put option);
t is the calculation date;
T is the options maturity (contracts end date);
DF (t, T ) is the discount factor of maturity T ;
Ft,T is the forward of maturity T .
Matching the call and put prices for all the strike prices K [0; +[ will allow
us to replicate payoffs that are twice differentiable, as any such payoff can be
written as the infinite sum of weighted vanilla call and put prices through the
following formula:

x0 +
0 00 00
g(x) = g(x0 )+(x x0 ) g (x0 )+ (K x)+ g (K)dK+ (x K)+ g (K)dK
0 x0
(1)
This formula, known as Carr formula, also gives a static replication strat-
egy for the payoff. Of course one could never succeed to perfectly exploit this
formula, as the market prices available are not continuous (in strike) and there
is no market liquidity for small strikes (below 60%) and for high strikes (above
200%). There are some numerical techniques for minimizing the effect of the
non continuity in strike. These techniques consist in smartly computing the
calls and puts weights on some specific strikes; depending on the payoff. The
use of the above formula has been largely documented in the pricing of Variance
Swaps. The reader can refer to [6] for further details.
Some exotic products as forward start options (all the forward start options
depend on volatility of the volatility), clicquets or napoleon strongly depends
on the volatility of volatility, and the forward skew (forward prices of the call
and put options).
Finally we are expecting the volatility model to behave such that we match:

2
1. The market prices of vanilla put and call options for all the strikes and all
the maturities;
2. The volatility of volatility;
3. The forward implied volatilitys skew (skew at the money forward). We
will clearly define the skew in next section, but before that we need to
define the implied volatility.

3 The Black and Scholes model (1973) and the


implied volatility surface
The Black and Scholes is a basic model where the volatility and the drift are
supposed to be constant and fixed: t = r and t = . If this model were
realistic, then the same calibrated constant volatility would allow to replicate
vanilla call and put prices for all the strikes and all the maturities. Unfortunately
it is not the case. We observe that for each set (T, K) represesenting a vanilla
option of maturity T and strike K, we need a constant volatility different from
any other set. In other words the constant volatility to be used depends on the
maturity and the strike of the option we are trying to replicate. If we use the
same constant volatility for a different (strike or maturity) vanilla option, we
wont be able to replicate it using this model. For a given maturity, the function
that represents the constant volatility with respect to the different strike prices
is called the implied volatility smile. It is called smile because (in equity world
at least) it has the shape of a smile. The skew is the skew at the money forward
of the implied volatility smile represented as a function of the log strike:

Imp
Skew(t, T ) =
ln(K) |K=Ft,T

3.1 Black and Scholes Model advantages


It is easy to calibrate, as there are closed form formulae for vanilla prices:

Call(t, T, K) = DF (t, T ) (Ft,T N (d+ ) KN (d ))

P ut(t, T, K) = DF (t, T ) (Ft,T N (d+ ) KN (d ))


 
Ft,T
ln K + 12 2 (T t)
d+ =
T t
 
Ft,T
ln K 12 2 (T t)
d =
T t
x
1 t2
N (x) = e 2 dt
2

3
Can help having an idea how the volatility can impact some complicated
products (as it is easy to handle);
it is a good base for building and testing more complicated volatility mod-
els;

can be used as an input for more complicated models, if we succeed to


parametrize the implied volatility surface such that it matches the market
implied surface

3.2 Black and Scholes Model drawbacks


This model is based on a number of assumptions (log normal underlying
process, no jump, no dividend, market liquidity, no arbitrage opportu-
nities, no transaction cost). Unfortunately almost all these assumptions
doesnt meet in the real world;
This model is static, there is no dynamics as the market moves;

It is not possible to diffuse the underlying equity to price a product other


than the vanilla put or call from which the constant volatility used is
implied. In particular it is not possible to compute a Monte Carlo fair
price for a non vanilla call and put option with this model.

From the last point we may want to derive a model, a volatility function that
allows to replicate the whole market vanilla grid (strikes and maturities) within
a single Monte Carlo pricing for instance. This has been achieved by Bruno
Dupire in 1994 when he discovered the local volatility function.

4 The local volatility or Dupire volatility func-


tion (1994)
The question behind the local volatility is the following: given a market grid
(strikes and maturities) of vanilla put and call options, is there a volatility
function depending on the time and the underlying equity price, such that one
can replicate the whole market grid 1 in a single pricing? The answer is yes,
and this volatility function is known as the Dupire local volatility function, from
Bruno Dupire:
C C
+ (rT qT )K K + qT C
2 (T, K) = T
1 2 2C
2 K K 2
With
1 We remind that this is not possible with the implied volatility, as each implied volatility
is only computed to replicate the vanilla options price from which it is computed.

4
C = Call(T, K) is the price of a vanilla call option for maturity T and
strike K;
r is the the short interest rate;
q is the proportional dividends rate.
There is a second interpretation of the local volatility function: the square of the
local volatility function is the projection of the unknown real stochastic variance
in the space (information) defined by the equity underlying:

2 (t, S) = E0 [Vt |St = S]


Where V is the unknown and stochastic variance of the equity underlying
process. This second interpretation justifies the use of the local volatility within
the Monte Carlo simulation.

4.1 Local volatility advantages


It is easy to compute, with a very negligible cost. In practice one computes
only once a grid of local volatility points (K, T ), and this grid is stored.
these points are then interpolate and extrapolate within the pricing;
As it matches the market vanilla grid of prices by construction, any payoff
that is twice differentiable with respect to the equity underlying variable
can be fairly priced with the local volatility function. The justification is
equation 1.

4.2 Local volatility drawbacks


As in Black And Scholes model, there is no dynamics with local volatility
function, as the market prices move;
Numerical tests show that the local volatility collapses the skew of a given
maturity. To be more precised if we consider the skew of a 1 year maturity
implied volatility smile seen as of today, then if we diffuse the same skew
for 6 months for instance (1 year maturity skew as seen in six months)
then it will have collapsed in absolute value. And this is not an accurate
intuitive behaviour. We are expecting the skew of a given maturity to
keep constant in the time;
The same behaviour can be observed regarding the volatility of volatility.
It is not kept constant.
While matching the market vanilla prices, local volatility function has a wrong
dynamics, and is not suitable for pricing forward start products, forward skew
sensitive products (as clicquets) and volatility of volatility depending products
(Napoleon). The next step will be to find a volatility model that can be cal-
ibrated on market vanilla prices with a cheap cost, and be able to properly
handle forward skew and volatility of volatility.

5
5 Stochastic volatility models
With stochastic volatility models we are trying to handle forward skew and
volatility of volatility while being able to be calibrated on market vanilla prices
with a cheap cost.
Most of the stochastic volatility models used can be ranged in a same family
models called Constant Elasticity Variance (CEV) models. The generic dynam-
ics within this models is the following:
dSt p
= t dt + Vt dWtS
St

dVt = (V Vt ) dt + Vt dWtV

d W S ; W V t = dt

V > 0 is the level of the variance towards which the process converges
after un very long time;
0 is the mean reverting parameter. It is the intensity of calling back
the variance process towards the long term variance level;
> 0 is the volatility of volatility parameter;
]0; 1] is the elasticity parameter:

The case = 1 is known as the Gatherals model;


The case = 21 is known as the Hestons (1993) model. This case,
which is an affine model provides a closed form formula (through
Fourrier transform) for vanilla options. It has been widely used by
practitionners until Lorenzo Bergomi showed in (smile dynamics I,
2004) that this model is structurally wrong, as the skew and the
volatility of volatility provided by this model are inconsistent with
the market observations.

In addition to the CEV models, one can add the Lorenzo Bergomis variance
model, the Ballands log normal model and the Karasinski and Sepp two factors
model.
Ballands model: p 1
t = Vt = 0 eZt 2 V ar(Zt )

Z is a normal Ornstein-Uhlenbeck process: dZt = (Z Zt ) dt +


dWtV ;
The model is built such that the volatility process is log normal and
local martingale.

Bergomis Forward Variance model (Bergomi 2005, Smile dynamics II [9]):

6
Begomi chooses to model the forward instantaneous variance: tT =
Et [VT ]; with Vt = tt
The main idea behind this choice is that the forward instantaneous
variance can be initialised using the term structure of variance swap
prices, as we have that

tT = ((T t)V arSwap(t, T ))
T
The model is built such that there is a separation between the skew
and the volatility of volatility (which is not the case with classical
models above, as shown by Lorenzo Bergomi in [8])
The 1-factor version of the model is the following
(T t)
Xt 21 w2 e2(T t) V ar(Xt )}
tT = 0T e{we
dXt = Xt dt + dWtV

Karasinski and Sepp model (Karasinski and Sepp 2012, The beta stochas-
tic volatility model [12]):

As suggested by Lorenzo Bergomi in [8], by adding a second factor to


the volatilitys diffusion, Karasinski and Sepp propose a model where
it is possible to drive the volatility of volatility without altering the
skew:
dSt
= t dt + t dWt
St

dt = ( t ) dt + t dWt + dWt

d W ; W t = 0

(Wt )t0 and Wt



t0
are two independant brownian motions;
The parameter drives the skew and also appears in the volatility
of volatility expression;
The parameter drives the volatility of volatility only.

5.1 Stochastic volatility models advantages


Most of the stochastic volatility models handle volatility of volatility better than
the local volatility does, as Stoch-Vols models have a volatility parameter with
actually drives the volatility of volatility. smart calibration of the correlation
(between the underlying equity and the stochastic volatility) parameter, or more
recently some models built to handle separately the skew and the volatility of
volatility (Karasinski and Sepp 2012 [12]; Bergomi 2005 [9]) allow to drive both
the skew and the volatility of volatility. Therefore stochastic volatility models
are suitable for pricing forward skew or volatility of volatility sensitive products.

7
5.2 Stochastic volatility models drawbacks
Can be unstable and difficult to implement. We need to find or adapt exist-
ing discretization schemes such that the variance process remains positive
while diffusing. We should avoid flooring the variance process at zero, as
doing this brings a distorsion in the process distribution, and may lead to
wrong prices;
Not all the parameters sets are acceptable;
Except the affine models (Heston) for which there are closed form for-
mula for vanilla call and prices, there are no closed form formula for most
stochastic volatility models. Therefore the only way to calibrate the model
parameters on vanilla prices is through a Monte Carlo simulation (or PDE
pricing). The PDE is unstable due to the dimension, and the Monte Carlo
is time consuming (more than five hours for performing one calibration).

The Heston Model is the only stochastic model to provide closed form formula
for the vanilla call and put prices; and thus a fast calibration on vanilla prices.
However it has been shown that the Heston model generates a wrong skew
(bergomi 2004, Smile Dynamics [8]). The Others stochastic models with ac-
ceptable skew and volatily dynamics are long to be calibrated on market vanilla
prices. The idea has been made to combine stochastic volatility model with a
local adjustment factor. The stochastic volatility will be used to manage the
forward skew and volatility of volatility, while the local adjustment factor will
be used to match the market vanilla prices.

6 Local-Stochastic-Volatility (LSV) model


The Volatility model here is the combination of a stochastic volatility model
with a local adjustment parameter. The stochastic volatility will be used to
manage the forward skew and volatility of volatility, while the local adjustment
factor will be used to match the market vanilla prices. The forward of the equity
underlying dynamics is therefore
dft
= (t, ft ) Vt dWtf
p
ft
V is any of the stochastic volatility model described in the previous section;
(t, ft ) is an adjustment factor. It is not a local volatility. The proper
volatility here is stochastic. The adjustment factor is only there to allow
the whole model to match the market vanilla prices, by construction, as
in the local volatility framework.

The main challenge with this model is how do we compute the adjustment
factor in order to match the market vanilla prices? The solution starts with
a result known as Gyongys Theorem. This theorem suggests that the equity

8
underlying process with LSV model has the same terminal distribution as the
equity underlying with the market local volatility function if

2 (t, f )E0 [Vt |ft = f ] = mkt


2
(t, f )
We thus get that
2
mkt (t, f )
2 (t, f ) =
E0 [Vt |ft = f ]
Where mkt (t, f ) is the Dupire local volatility function computed on market
vanilla prices grid. The challenge now is how to compute E0 [Vt |ft = f ]? We
write that

 
E0 Vt 1{ft [f ;f +df ]}
E0 [Vt |ft = f ] =   (2)
E0 1{ft [f ;f +df ]}
+
vp(t, f, v)dv
= 0 +
0
p(t, f, v)dv

p(t, f, v) is the probability density function of the joint distribution of the


variance and the underlying at time t.

Many techniques have been proposed for computing 2.

6.1 Kolmogorov-forward approach (Alexander Lipton (2002),


[7])
As all the continuous time probability transition functions, p is the solution
of both Kolmogorov-backward and Kolmogorov-forward equations. Here we
will exploit the Kolmogorov-Forward equation, also known as Fokker-Planck
equation. With a CEV stochastic volatility model, the Kolmogorov-Forward
equation is the following:

p 1 2 2
 1 2 2 2
 2  + 12

(t, f, v) = (t, f )vp + v p + (t, f )v p
t 2 f 2 2 v 2 f v
(3)

lim p(t, f, v) = (f f0 ; v v0 )
t0+

However, the numerical resolution of the multidimensional Kolmogorov-


forward equation is very challenging, especially when the correlation between
the underlying and the variance is not null. Assuming that we can solve the
Kolmogorov-forward equation, the algorithm for solving the problem will be the
following:
1. Build a fixed grid (fi , vj )i,j>0 ;

9
2. solve equation 3 for each point (fi , vj ) of the grid;
3. Integrate along v and compute the adjustment factor using 2.

6.2 Markovian projection approach (Pierre Henry-Labordere


(2009), [10])
The core idea of this technique is to find the path (function) that leads from
the underlying process (f ) to the process (X)defined as
dXt
= Vt dWtf
p
Xt
Labordere denotes this function H(t, ft ) = Xt . He then writes itos lemma on
the process (Xt )t0 . He finds that the function H is the solution of 2 equations.
One of the 2 equations is a PDE that can be integrated to find the expression
of H. This expression depends on a function which should be computed by
solving the other equation. Unfortunatenaly the second equation depends on
the stochastic volatility and the adjustment factors derivative with respect to
the equity underlying. To overcome the complexity Labordere comnsiders the
expectation of the second equation. Due to this approximation it is difficult to
have this method converging in practise. Below are the results:

mkt (t, f )
(t, f ) =
locSto (t, H(t, f ))
f
!
dx
H(t, f ) = 1
t
f0 xmkt (t, x)
x
dy
t (x) =
(t) ylocSto (t, y)

  T
(T ) 1 2
 0

ln = locSto (t, f0 ) 1 (t, f0 ) f0 (t, f0 ) dt (4)
f0 2 0

2
locSto (t, x) = E0 [Vt |Xt = x]
 
E0 Vt 1{Xt [x;x+dx]}
=  
E0 1{Xt [x;x+dx]}
  
E0 Vt pX x|FtV dx
=   
E0 pX x|FtV dx
 
E0 Vt pX x|FtV
=  
E0 pX x|FtV

And since Xt |FtV is a log normal process, we get that

10
K 2
1
2(12 ) t Vs ds
pX x|FtV = q

t e 0

x (1 2 ) 0 Vs ds 2
With
  t tp
x 1
K := ln + Vs ds Vs dWsV
x0 2 0 0

After simplification we finally get that


K 2
" #

2(12 ) t Vs ds
e 0
E0 Vt t
V ds
0 s
2
locSto (t, x) = " K 2
#

2(12 ) t Vs ds
e 0
E0 t
V ds
0 s

2
The computation of locSto (t, x) can be done in Monte Carlo.
This technique is not accurate due to equation 4 which is an expectation of
the original equation to be solved.

6.3 Particle simulation approach (Julien Guyon and Pierre


Henry-Labordere (2011), [11])
This method is the most simple, accurate and powerful, as it overcomes the mul-
tidimensional issue, and there is no approximation as in the projection method.
The idea used here is also very basic: use Monte Carlo simulations for estimating
the expectations in equation 2:

 
E0 Vt 1{ft [f ;f +df ]}
E0 [Vt |ft = f ] =  
E0 1{ft [f ;f +df ]}
+
vp(t, f, v)dv
= 0 + (5)
0
p(t, f, v)dv
 
PN (i) (i)
i=1 t V f t f
=  
PN (i)
i=1 ft f

(i) (i)
ft and Vt are the forward and the variance for the ith Monte Carlo
path;
The Dirac function can be approximated with the kernel function as
explained in Guyon-Labordere (2012), [11];

The adjustment factor (t, f ) is computed step by step as the variance


and the forward processes are diffused:

11
Given the initial adjustment factor (t0 , f0 ), diffuse the variance Vt1 ,
and diffuse the forward ft1 ;
repeat the operation for N paths, and compute the adjustment factor
(t1 , f ), using equation 5.
The only point to be checkeck regarding this technique is how the choice of the
kernel function affects the results.

References
[1] Fischer Black and Myron Scholes (1973). The Pricing of Options and Cor-
porate Liabilities. The Journal Of Political Economy, Vol. 81, No. 3 (May-
Jun. 1973), pp 637-654. The University Of Chicago Press.
[2] Gyongy I. (1986). Mimicking the One-Dimensional Marginal Distributional
Marginal Distributions of Processes Having an Ito Differential. Probability
Theory and Related Fields, 71, pp 501-516.
[3] Heston S. (1993). A Closed-Form Solution for Options with Stochastic
Volatility with Applications to Bonds and Currency Options. Review of
Financial Studies, 6, pp 327-343.
[4] Dupire B. (1994). Pricing With a Smile. Risk Magazine July 2007, Cutting
Edge Classic from January 1994. Volatility, pp 126-129.
[5] Avellaneda M., Levy A. and Paras A. (1995). Pricing and Hedging Deriva-
tive Securities in Markets with Uncertain Volatilities. Applied Mathemati-
cal Finance, 2, pp 73-88.
[6] Demeterfi K., Derman E., Kamal M. and Zou J. (1999). More Than You
Ever Wanted To Know About Volatility Swaps. Quantitative Strategies
Research Notes. Goldman Sachs. March.
[7] Lipton A. (2002). The vol smile problem. Risk Magazine February 2002,
Cutting Edge. pp 61-65.
[8] Bergomi L. (2004). Smile Dynamics. Risk Magazine September 2004, Cut-
ting Edge. pp 117-123.
[9] Bergomi L. (2005). Smile Dynamics II. Risk Magazine October 2005, Cut-
ting Edge. pp 67-73.
[10] Labordere P.-H. (2009). Calibration of Local Stochastic Volatility Models
to Market Smiles. Social Science Research Network (SSRN). October.
[11] Guyon J. and Labordere P.-H. (2011). The Smile Calibration Problem
Solved. Social Science Research Network (SSRN). July.
[12] Karasinski P. and Sepp A. (2012). The Beta Stochastic Volatility Model.
Risk Magazine October 2012, Cutting Edge. pp 66-71.

12

Das könnte Ihnen auch gefallen