Sei sulla pagina 1di 31

Stochastic Volatility III:

Volatility Models and Volatility Surfaces


(Lecture script version June 2, 2010)

Master of Advanced Studies in Finance

ETH Zurich / University of Zurich

Spring Term 2010

Paolo Vanini

Zürcher Kantonalbank and Swiss Finance Institute


E-mail address: paolo.vanini@zkb.ch
We consider implied volatility, time-dependent volatility, local volatility and
stochastic volatility. That is, derive the relationships between the different con-
cepts. The relationships are of an exact analytical type if this possible, else we use
expansions to obtain approximate expressions. We close the section with a discus-
sion of the mixing theorem.

Keywords: Implied Volatility, Stochastic Volatility, Local Volatility, Volatility


Statics, Mixing Theorem, Asymptotic Expansions, Option Pricing.

These are rough Lecture Notes which we use in the MAS in Finance at the
University of Zurich and ETH Zurich. Feedback is welcome.
Contents

Chapter 1. Volatility Models and Volatility Surface 1


1.1. Implied volatility 1
1.2. Time-dependent volatility 2
1.3. Local volatility 2
1.4. Implied Volatility in Terms of Local Volatility 13
1.5. Local variance and instantaneous variance 16
1.6. Volatility statics under no arbitrage condition 19
1.7. Mixing Theorems 21
1.8. Appendix: Forward, backward equations 24
Index 27

iii
CHAPTER 1

Volatility Models and Volatility Surface

Reading for this chapter:


• S. Barle and N. Cakici. How to Grow a Smiling Tree The Journal of
Financial Engineering, 7: 127146, 1998.
• H. Berestycki, J. Busca, and I. Florent. Asymptotics and Calibration of
Local Volatility Models. Quantitative Finance, 2:6169, 2002.
• E. Derman and I. Kani. The Volatility Smile and its Implied Tree. Risk,
7:32-39, 1994. (available on-line)
• E. Derman, I. Kani and N. Chriss. Implied Trinomial Trees of the Volatil-
ity Smile. Goldman Sachs Quantitative Strategies Research Notes, 1996.
Risk, 7:32-39, 1996.
• W. Haerdle and A. Mysickova. Numerics of Implied Binomial Trees. SFB
649 Discussion Paper 2008-044. Humboldt-Universitaet zu Berlin. 2008.
• E. G. Haug, The Complete Guide to Option Pricing Formulas. Mc Graw
Hill. Second Edition, 2007.
• B. Dupire. Pricing with a Smile. Risk, 7:18-20, 1994.
• P.H. Labordère. Analysis, Geometry and Modeling in Finance. Advanced
Methods in Option Pricing. Chapman and Hall/CRC Financial Mathe-
matics Series, 2009.
• A.L. Lewis. The Mixing Approach To Stochastic Volatility and Jump
Models. Working Paper, 2002.
• R. Lee. Implied volatility: Statics, dynamics, and probabilistic interpre-
tation. In R. Baeza-Yates, J. Glaz, H. Gzyl, J. Hüsler, and J. Pala-
cios, editors, Recent Advances in Applied Probability. Springer, 2004.
(available on-line)

1.1. Implied volatility


Take the usual notation of C for a call option paying off (ST − K)+ at expiry
T.
In the Black-Scholes framework, where S follows the geometric Brownian mo-
tion
dSt = µSt dt + σSt dWt ,
we denote the theoretical option price by
C = CBS (σ).
The implied Black-Scholes volatility, I(K, T ) = IBS , is the implicit solution to
C(K, T ) = CBS (K, T, I(K, T )) ,
1
2 1. VOLATILITY MODELS AND VOLATILITY SURFACE

where C(K, T ) is the market price of the option. Moreover, both existence and
uniqueness of I(K, T ) are guaranteed since C BS is strictly increasing in σ and hits
the call market price exactly once. Note that explicitly, I is a function of K, T, t, S.
How is the implied volatility of the Black-Scholes model related to the volatil-
ity in local volatility models, and what is the relationship to stochastic volatility
models? Why is this interesting? We know that there is volatility smile for exam-
ple which cannot be explained within the Black and Scholes model. Hence, model
extensions such as stochastic volatility or local volatility models should be able to
explain this and other stylized facts. But traders in the market like to think in Black
and Scholes implied volatility. Therefore, it is natural to consider the possible links
between the different volatility notions.

1.2. Time-dependent volatility


As an initial case, take volatility to be a deterministic function of time, i.e.
σ(t). The risk-neutral dynamics of the underlying price is
dSt = rSt dt + σ(t)St dWt .
Applying Itô’s lemma to log S gives
1 1 1
d log S = dS − dhSit
S 2 S2
1
= r dt + σ(t) dW − σ 2 (t) dt
2
Z Z t
1 t 2
⇒ log St = log S0 + rt − σ (s) ds + σ(s)dWs
2 0 0
so that   
σ̄ 2
log St /S0 ∼ N r− t, σ̄ 2 t .
2
Therefore the price CDT of an option in the time-dependent case is related to
the Black and Scholes prices as follows:
CDT = CBS (σ̄),
where Z
1 t 2
σ̄ 2 = σ (s) ds
t 0
is the quadratic mean value from 0 to t.

1.3. Local volatility


Local volatility models assume the following underlying price dynamics:
(1.1) dS(t) = rS(t) dt + σ(St , t)S(t) dWS (t),
where σ is now a deterministic function of both time and the underlying’s price.
This is opposite to stochastic volatility models. Hence, the same risk source drives
both the underlying and the volatility.

The motivation for this model class is as follows. For vanilla options, implied
volatility I(K, T ) = IBS is a function of strike and maturity. Hence, the fair vanilla
option price is a function of K and T or in other words, I(K, T ) determines the
risk neutral probability for vanilla option’s underlying value S(T ) at maturity. The
1.3. LOCAL VOLATILITY 3

option value at maturity is independent of the underlying’s path. If we consider


a barrier option, not only the terminal value S(T ) matters for option pricing, i.e.
pricing becomes path dependent. The implied volatility surface for vanilla options
will not tell the right answer, since the volatility figure is of a global type, i.e. path
independent. The path dependency requires a volatility figure which is also path
dependent, i.e. σ(St , t) as for the local volatility model. Since this volatility is not
a global figure but is only valid for given St and t, the expression ”local volatility”
is used.

Dupire (94) shows that, given a complete set of European option prices for
all strikes and maturities, local volatilities are uniquely determined by the vanilla
option prices and their sensitivities, see below. Therefore, local volatility models
need no extra calibration.

1.3.1. Implied Trees. Before we consider first the discrete setup in this sec-
tion. The model is discrete both in time and state space, i.e. we consider trees.
The goal is to use information of liquid options with different strikes and maturi-
ties to construct an arbitrage-free model. Hence, all relevant market information is
included in the model - hence we speak about an implied model. Consider a stock
dynamics with a local volatility structure σ(S, t). The functional form can either
be assumed or estimated from market data. Implied trees estimate numerically
this function using the liquid option prices. The trees can then be used to price
other options such as exotics for example. We consider implied trees in this section.

The standard approach is to start with a recombining binomial tree much in


the same spirit as the Cox-Ross-Rubinstein (CCR) model for the constant volatil-
ity case. Although the implied tree is still recombining and with equidistant time
spacing (for the moment being) there will be a distortion in the state space which
reflects the approximation to the function σ(S, t).

To construct an implied tree assume that there are n time steps. We denote
by (n, j) the node j at time step n. We determine the tree quantities using forward
induction: Suppose we are at time n and we know all parameters at this time point.
Then we want to find the unknown parameters at time n + 1. Suppose that n = 1.
Then at time n = 2 we have to find
• 3 asset prices S2,j (n + 2 for general n),
• 2 risk-neutral transition probabilities p1,j from the nodes (1, j) to the
nodes (2, j) (n + 1 for general n).
What do we know to find this unknown parameters?
• There are 2 forward prices F1,j = S1,j er∆t with r the risk free rate (not
necessary time independent) (n + 1 for general n),
• There are 2 option prices O1,j expiring at time T1 (n + 1 options expiring
at Tn+1 for general n)
In summary, there are 2n + 3 unknown parameters and 2n + 2 known parameters.
Hence, there is a single extra freedom. Typically, the extra freedom is chosen to
center the implied tree around a standard CCR: At each odd time step, the middle
node of all possible nodes is set equal to the initial node value at time 0.
4 1. VOLATILITY MODELS AND VOLATILITY SURFACE

The unknown parameters are found from the known one using risk-neutrality
both for the forward and the option prices. At time n, the known forward value at
time n + 1 at any node j must be equal to the risk-neutral expected value of the
stock prices, i.e. for all j = 1, 2, . . . , n
(1.2) Fn+1,j = pn+1,j Sn+1,j+1 + (1 − pn+1,j )Sn+1,j .
For the option prices we set C(sn,j , Tn+1 ) for a liquid European call option with
maturity Tn+1 and struck at today’s n-time value sn,j of the stock. Similar,
P (sn,j , Tn+1 ) reflects the set of put options. The theoretical binomial value of
a call struck at K and expiring at Tn+1 is given by the sum over all nodes j at
the (n + 1)th level of the discounted probability of reaching each node (n + 1, j)
multiplied by the call payoff there:
n
X
C(K, Tn+1 ) = e−r∆t (λn+1,j pn+1,j + λn+1,j+1 (1 − pn+1,j+1 )) max(Sn+1,j+1 −K, 0)
j=1

where λ are the Arrow-Debreu state prices. That is, λn,j is the price today of a
security that has a cash flow of unity in state j at time n and zero elsewhere, i.e.
λn,j = E[e−r∆t χS∆t =Sn,j |S0 =S] .

The state prices for the next time step n + 1 are given by

 pn,n λn,n , for i = n + 1;
er∆t λn+1,j = pn,i−1 λn,i−1 + (1 − pn,i )λn,i , 1 ≤ i ≤ n + 1;

(1 − pn,0 )λn,0 , i = 0.
The option price formula follows from the continuous time formula as follows. In
continuous time we have
Z ∞
−rτ
C(K, τ ) = e max(ST − K)q(ST |St , r, τ )dST .
0
In the implied binomial model the option price is calculated using the Arrow-Debreu
prices, i.e.
Xn
C(K, n∆t) = λn+1,j+1 max(Sn+1,j+1 − K, 0) .
j=0

When the strike K equals the known stock prices sn,k , the contribution from the
transition to the first in-the-money up node can be separated from the other con-
tributions. Using Equation (1.2) this contribution can be written in terms of the
known Arrow-Debreu prices, stock prices and forwards, i.e.
(1.3)
n
X
C(sn,k , Tn+1 ) = e−r∆t [λn,k pn+1,k (Sn+1,k+1 − sn,k ) + λn,j (Fn,j − sn,k )] .
| {z } j=k+1
S, p unknown | {z }
Known Expression

Since the forward and call prices are known from the smile, the equations (1.2) and
(1.3) can be solved simultaneously for the unknown stock prices and probabilities.
Solving the system, one gets the solution for the asset prices above the central node:

Sn+1,j er∆t C(Sn,j , Tn+1 ) − S − λn,j Sn,j (Fn,j − Sn+1,j )
(1.4) Sn+1,j+1 =
er∆t C(Sn,j , Tn+1 ) − S − λn,j (Fn,j − Sn+1,j )
1.3. LOCAL VOLATILITY 5

with
n
X
S= λn,k (Fn,k − Sn,j ) .
k=j+1
Below the central node, we have

Sn+1,j+1 er∆t P (Sn,j , Tn+1 ) − T − λn,j Sn,j (Fn,j − Sn+1,j+1 )
(1.5) Sn+1,j =
er∆t C(Sn,j , Tn+1 ) − T − λn,j (Fn,j − Sn+1,j+1 )
with
j−1
X
T = λn,k (Fn,k − Sn,j ) .
k=0
The transition probabilities are given by
Fn,j − Sn+1,j
(1.6) pn,j = .
Sn+1,j+1 − Sn+1,j
Given the solution algorithm how do we start to build a new level of states of the
tree? If the number of time steps n already solved for is odd, the initial central
node Sn+1,j is equal to the central node of the CRR tree. If the number of solved
steps n is even, we use the logarithmic centering condition
S2
Sn+1,j =
Sn+1,j+1
with S today’s asset price. Inserting this in (1.5) leads to Sn+1mj+1 . The implied
local volatilities at each node are finally given by
q  
Sn+1,j+1
(1.7) σn,j = pn,j (1 − pn,j ) ln .
Sn+1,j
Example 1.1. We construct a three-year implied binomial tree with annual
time steps. The risk free rate is 10%, the at-the-money volatility is 15%, today’s
stock price is 100. Volatility increases (decreases) by 0.5% with every 10-point
decrease (increase) in the strike price.
The node in the up state S1,1 after one period n = 1 is first calculated as
100(e0.05×1 C(S, 1) + 1 × 100
S1,1 = .
1 × F0,0 − e0.05×1 C(S, 1)
The expression S is zero since there are no nodes above this stock price, F0,0 =
100e0.05×1 = 105.13 and the call ATM has a price C(100, 1, σ = 15%) = 9.74.
Therefore,
S1,1 = 116.18 .
For the down node at n = 1 we get
S2 1002
S1,0 = = = 86.07 .
S1,1 116.16
The up transition probabilities are
105.13 − 86.07
p0,0 = = 0.633
116.18 − 86.07
and finally, the Arrow-Debreu prices follow:
λ1,1 = 0.633 × 1 × e−0.05×1 = 0.602 .
6 1. VOLATILITY MODELS AND VOLATILITY SURFACE


The CRR stock binomial tree follows from u = e0.15× 1 = 1.162 and the normaliza-
e0.05 −0.861
tion d = u1 = 0.861. The constant up probability p reads p = 1.162−0.861 = 0.633.
See Figure 1.1 for the illustration of the different trees.

CRR Tree Implied Binomial


Arrow-Debreu
Stock Tree Stock

116.18 116.18 0.602

100 100 1

86.07 86.07 0.349

Figure 1.1. Trees in the first time step.

We consider the next period n = 2. Since the already solved time step n = 1 is
odd, we set for the stock price of the central node S2,1 = 100. The node S2,2 above
this center node is calculated as:
100[e0.05×1 C(S1,2 , T2 ) − S] − 0.602 × 116.18(F1,1 − 100)
S2,2 = .
e0.05×1 C(S1,2 , T2 ) − S − −0.602 × (F1,1 − 100)
We have for the option and forward the following values:
C(S1,2 , T2 ) = C(116.18, 2, 14.19%) = 6.25 , F1,1 = 116.18 × e0.05×1 = 122.14 .
Hence,
S2,2 = 131.94 .
In the same way we obtain for the node below the central node
S2,0 = 70.49 .
Since there are no nodes below, the S-term is zero (as for the node above the central
node). The transition probabilities leading to these nodes are given by, see Figure
1.2 for an illustration:
122.14 − 100
p1,1 = = 0.693 , p1,0 = 90.48 − 70.49100 − 70.49 = 0.678 .
131.94 − 100
The Arrow-Debreu prices and the transition probabilities are shown in Figure 1.3.
The Arrow-Debreu price λ2,2 is calculated as follows:
λ2,2 = 0.693 × 0.602 × e−0.05×1 = 0.397 .
The local volatilities are given by:
p 116.18
σ0,0 = 0.633(1 − 0.633) ln = 14.458% ∼ 14.5% .
86.07
1.3. LOCAL VOLATILITY 7

Implied Binomial
CRR Tree 134.99 Tree Stock 131.94
Stock
116.18 116.18

100 100 100 100

86.07 86.07

74.08 70.49

Figure 1.2. Stock trees in the CRR and implied Binomial model
for n = 2.

In the same way we obtain the local volatilities at the future nodes:
σ1,1 = 0.128 , σ1,0 = 0.163 , σ2,2 = 0.110 , σ2,1 = 0.146 , σ2,0 = 0.174 .

Transition
Arrow Debreu 0.397 Probabilities 0.706

0.602 0.693

1 0.401 0.633 0.631

0.349 0.678

0.107 0.532

Figure 1.3. Arrow-Debreu prices and the transition probabilities.

We considered Binomial trees due to their simplicity. More accurate is the use
of Trinomial trees. The main reason to use them is their greater flexibility in the
description of the implied stock market process. More precisely, in a binomial tree
arbitrage violation are possible. Consider for example a node where the value of the
Forward is above the up-value of the next time step implied stock value. This is an
arbitrage condition which can lead to negative values for the implied probabilities.
Using a trinomial tree there is more flexibility to model the implied stock tree such
8 1. VOLATILITY MODELS AND VOLATILITY SURFACE

that such arbitrage situations can be avoided. Derman et al. (96) explain in detail
the construction of the Trinomial tree.

A detail analysis of the implied Binomial tree is given by Haerdle and Mysickova
(08). They also compare the Derman and Kani approach with the alternative
method of Barle and Cakici (98).
1.3.2. Derivation of the Dupire equation. The Dupire equation describes
the relationship between implied and local volatility. The non-discounted risk-
neutral value C = C(S0 , K, T ) of a European call option is given by
Z ∞
(1.8) C= ϕ(ST , T )(ST − K) dST
K
where ϕ is the unknown probability density of the final spot price at maturity which
satisfies the Fokker-Planck equation (see the Appendix for details)
1 ∂2  ∂ ∂ϕ
2 σ 2 ST2 ϕ − (rST ϕ) = .
2 ∂ST ∂ST ∂T
Differentiating (1.8) w.r.t. K twice gives
Z ∞
∂C ∂
= ϕ(ST − K) dST
∂K ∂K K
Z ∞
= ϕ(K − K) − ϕ dST
K
Z ∞
= − ϕ dST
K
and
Z ∞
∂2C ∂
= − ϕ dST
∂K 2 ∂K K
= ϕ(K, T ).
Hence we can recover the risk neutral density ϕ from option data. Differentiating
(1.8) w.r.t. T gives
Z ∞ 
∂C ∂
= ϕ(ST , T ) (ST − K) dST
∂T K ∂T
Z ∞ 
1 ∂2 2 2
 ∂
= σ S T ϕ − (rS T ϕ) (ST − K) dST
K 2 ∂ST2 ∂S
where we used the backward Fokker-Planck equation for the density function. In-
tegrating by parts twice gives
∂C σ2 K 2 ∂ 2 C ∂C
= − rK ,
∂T 2 ∂K 2 ∂K
which is the Dupire equation with initial condition
C(K, 0) = (S(0) − K)+ .
Implied local volatility σLoc (K, T ) is then defined as:
∂C ∂C
2 ∂T + rK ∂K
(1.9) σLoc (K, T ) = K2 ∂2C
.
2 ∂K 2
1.3. LOCAL VOLATILITY 9

This is the local volatility function consistent with the given prices of options and
their sensitivities whereas the unknown density function φ has been eliminated.
Equation (1.9) holds for non-dividend paying stocks. With a continuous dividend
stream d, Dupire’s equation reads:
∂C ∂C
2 ∂T + (r − d)K ∂K + dC
(1.10) σLoc (K, T ) = K2 ∂2C
.
2 ∂K 2
.

Dupire equation is obtained by switching from the PDE in the variables S, t to a


PDE in the variables K, T .

There is a simple interpretation of Dupire equation in terms of static option


strategies. That for, we set the interest rate equal to zero. Then,
∂C
2 ∂T
σLoc (K, T ) = K2∂2C
.
2 ∂K 2
∂C
But ∂T is the infinitesimal version of
C(S, t, K, T + ∆T ) − C(S, t, K, T )
∆T
i.e. a long call position with maturity T + ∆T and a short call with maturity T .
∂2C
In other words, a calendar spread with strike K. Similar, ∂K 2 is the infinitesimal

version of
C(S, t, K + ∆K, T ) − 2C(S, t, K, T ) + C(S, t, K − ∆K, T )
.
(∆K)2
But this is a butterfly spread with strike K.

Local variance is proportional to the ratio of a calendar and a butterfly spread.

Dupire equation looks much like the Black-Scholes equation with t replaced
by T and S replaced by K. But whereas the Black-Scholes equation holds for any
contingent claim on S, Dupire equation holds only for standard calls and puts. This
equation tells you how to find σLoc (K, T ) and hence build an implied local volatility
tree from options prices and their derivatives. You can then use that implied tree
to value exotic options and to hedge standard options, knowing that you have one
consistent model that values all standard options correctly rather than having to
use several different inconsistent Black-Scholes models with different underlying
volatilities.
Dupire’s approach requires a continuous set of options data for all K and T .
Since data are only available for a discrete set and options out- and in- the money
are suffering from illiquidity several problems arise in the implementation of the
approach. First, an interpolation between the discrete data points is needed. It
turns out that different interpolation methods have a strong impact on the outcome.
Besides this instability due to the interpolation, the calibration of the local volatility
surface is also instable over time.
If we re-express Dupire equation in term of the original variable, by applying
the chain rule and using the formula for the Greeks in Black and Scholes we get:
10 1. VOLATILITY MODELS AND VOLATILITY SURFACE

Proposition 1.2 (Local variance in terms of Black-Scholes implied variance).


For zero dividends and zero interest rates, implied local variance reads in term of
Black-Scholes implied variance:
(1.11)
2
2 ∂I∂T
BS
+ TIBS
−t
σLoc (K, T ) =   2 
∂ 2I √ ∂I
 2 1 1 ∂IBS
K 2 ∂KBS 2 − d1 T − t ∂K
BS
+ IBS

K T −t
+ d1 ∂K

with
ln(S/K) + (r + 12 IBS
2
)(T − t)
d1 = √ .
IBS T − t
Using the forward price and the transformation
 
K 2
x = ln , y(T, x) = IBS (K, T |S, t)(T − t),
F0,T
the result (1.11) reads:
∂y
2 ∂T
σLoc (x, T ) =   2 .
x ∂y 1 x2 ∂y 1 ∂2y
1− y ∂x + 4 − 14 − 1
y + y2 ∂x + 2 ∂x2

This proposition relates the two concepts: Implied local volatility and implied
Black and Scholes volatility.
Proof. Starting with the Black-Scholes implied volatility definition
C(S0 , K, T ) = CBS (S0 , K, I(K, T ), T ),
and using the definitions x, y, the Black and Scholes formula becomes:
C(F0,T , x, y) = F0,T (Φ(d1 ) − ex Φ(d2 )) ,
with
x 1√ x 1√
d1 = − √ + y, d2 = − √ − y.
y 2 y 2
The Dupire equation (1.9) becomes
 2 
∂C σ2 ∂ C ∂C
(1.12) = Loc − + rC,
∂T 2 ∂x2 ∂x
2 2
where σLoc = σLoc (S0 , K, T ) denotes the local variance. The derivatives of the
Black-Scholes formula are
 
∂ 2 CBS 1 1 x2 ∂CBS
= − − + ,
∂y 2 8 2y 2y 2 ∂y
 
∂ 2 CBS 1 x ∂CBS
= − ,
∂x∂y 2 y ∂y
and
∂ 2 CBS ∂CBS ∂CBS
2
− =2 .
∂x ∂x ∂y
We may rewrite (1.12) in terms of implied variance by making the substitutions
∂C ∂CBS ∂CBS ∂y
= + ,
∂x ∂x ∂y ∂x
1.3. LOCAL VOLATILITY 11

 2
∂2C ∂ 2 CBS ∂ 2 CBS ∂y ∂ 2 CBS ∂y ∂CBS ∂ 2 y
= + 2 + + ,
∂x2 ∂x2 ∂x∂y ∂x ∂y 2 ∂x ∂y ∂x2
and
∂C ∂CBS ∂CBS ∂y ∂CBS ∂y
= + = + rCBS .
∂T ∂T ∂y ∂T ∂y ∂T
Now (1.12) becomes
2

∂CBS ∂y σLoc (K, T ) ∂CBS ∂ 2 CBS ∂CBS ∂y ∂ 2 CBS ∂y
= − + 2
− +2
∂y ∂T 2 ∂x ∂x ∂y ∂x ∂x∂y ∂x
 2 #
2 2
∂ CBS ∂y ∂CBS ∂ y
+ 2
+
∂y ∂x ∂y ∂x2
2
  
σLoc ∂CBS ∂y 1 x ∂y
= 2− +2 −
2 ∂y ∂x 2 y ∂x
   2 #
1 1 x2 ∂y ∂2y
+ − − + + .
8 2y 2y 2 ∂x ∂x2
Simplifying,
"    2 #
2 2
∂y 2 x ∂y 1 1 1 x ∂y 1 ∂ y
= σLoc 1− + − − + 2 + .
∂T y ∂x 4 4 y y ∂x 2 ∂x2
Finally, inverting this expresses the local variance as a function of the Black-Scholes
2
implied variance y = IBS , i.e.
∂y
2 ∂T
σLoc =   2 .
x ∂y 1 x2 ∂y 1 ∂2y
1− y ∂x + 4 − 14 − 1
y + y2 ∂x + 2 ∂x2

Inserting the original variables and after some algebra the proof follows. 

The next Proposition summarizes some important facts.


Proposition 1.3. We assume that r = 0.
(1) Implied variance is an average of local variance if there is no skew, i.e.
implied volatility does not depend on strike.
(2) Assume that implied volatility does not depend on τ = T − t but only
weakly on strike K, i.e. higher order sensitivities than ∂I∂KBS
are negligible.
Furthermore, we consider the case where we are close to at-the-money.
Then the slope of the skew w.r.t strike is 1/2 the slope of the local volatility
w.r.t. spot.
(3) Implied volatility is an harmonic average over local volatility at short ex-
pirations.
This Proposition relates implied variance in the Black and Scholes model to
local variance or volatility. We observe, that the relationships are not holding in
general but only in certain regimes, for example no-skew regime or only for short
times to maturity. The harmonic average contrasts usual or expected arithmetic
or quadratic averages. Probabilistic considerations rule out the arithmetic and
quadratic means. Berestycki, Busca, and Florent (2002) provide the following ar-
gument: Consider a local volatility diffusion in which there exists a price level
12 1. VOLATILITY MODELS AND VOLATILITY SURFACE

between initial stock price S(0) and strike K above which the local volatility van-
ishes, but below which it is positive. Then the option must have zero premium,
hence zero implied volatility. This is inconsistent with taking a spatial mean of
volatility arithmetically or quadratically, but is consistent with taking a spatial
mean of volatility harmonically.
Proof. If there is no skew, i.e. implied volatility is independent of strike K,
then all derivatives w.r.t. to K in (1.11) vanish. Hence, we have
2 ∂y
(1.13) σLoc (x, T ) = .
∂T
Using the definition y 2 = IBS 2
(T − t), the last equality can be rewritten using
τ = T − t as
2 ∂ 2

σLoc (K, T ) = τ IBS .
∂τ
Integrating this equation and setting the integration constant to zero, we get
Z
2 1 τ 2
IBS = σ (K, u)du
τ 0 Loc
which proves the first claim.

For the second claim by assumption all τ -derivatives and all derivatives of
higher order than first order in K are set equal to zero in (1.11). This implies
2
2 IBS
σLoc (K, T ) = √ BS 2 .
1 + d1 K τ ∂I∂K
Taking the positive square root - the negative beeing meaningless - we get
IBS
σLoc (K, T ) = √ BS .
1 + d1 K τ ∂I∂K
To continue with the approximations we consider that we are close at-the-money,
1
i.e. we write K = S + ∆K. Using the approximations log(1 + x) ∼ −x, 1−x ∼ 1+x
for small x, we get
S S
log( K ) 1 √ log( K ) log(1 + ∆K ) ∆K 1
d1 = √ + IBS τ ∼ √ = √S ∼− √ .
IBS τ 2 IBS τ IBS τ S IBS τ
Replacing S by K − ∆K we get
∆K 1 ∆K 1
d1 ∼ − √ =− √
K − ∆K IBS τ K IBS τ
Therefore,
√ ∂IBS ∆K 1 √ ∂I 1 ∂IBS
d1 K τ ∼− √ K τ BS = −∆K
∂K K IBS τ ∂K IBS ∂K
and
 
IBS IBS 1 ∂IBS
σLoc (K, T ) = √ ∂IBS ∼ 1 ∂IBS
∼ IBS 1 + ∆K .
1 + d1 K τ ∂K 1 − ∆K IBS ∂K
IBS ∂K
In summary, we obtain
∂IBS (K)
σLoc (K, T ) ∼ IBS (K) + ∆K .
∂K
1.4. IMPLIED VOLATILITY IN TERMS OF LOCAL VOLATILITY 13

We next Taylor-expand σLoc (K, T ), i.e.


∂σLoc (S, T )
σLoc (K, T ) = σLoc (S + ∆K, T ) ∼ σLoc (S, T ) + ∆K
∂S
But a Taylor series expansion also implies:
∂IBS (S)
IBS (K) = IBS (S + ∆K) ∼ IBS (S) + ∆K .
∂K
Hence,
∂IBS (K) ∂IBS (S)
IBS (K) + ∆K ∼ IBS (S) + 2∆K |K=S .
∂K ∂K
In summary,
∂σLoc (S, T ) ∂IBS (S)
=2 |K=S
∂S ∂K
follows which proves the claim, i.e. local volatility grows twice as fast with stock
price as the implied volatility does with respect to the strike variable.

We prove the third claim and start with equation (1.11), i.e.

2 2τ ∂I∂τBS + IBS
(1.14) σLoc (K, T ) =  2 √ 2 √ BS 2  .
K 2 τ ∂∂K
IBS
2 − d1 τ τ ∂I∂K
BS 1
+ IBS 1
K + d1 τ ∂I∂K

We take the limit τ → 0. From the expression of d1 follows, that


τ d1 → ln(S/K)IBS , for τ → 0.
∂IBS
If we further assume, that the slope of ∂τ is bounded, then 1.14 becomes in the
limit τ → 0:
IBS
(1.15) σLoc (K, T ) = K ∂IBS
.
1+ IBS ∂K ln(S/K)
This is an ordinary differential equation. Using the moneyness variable x = ln(S/K),
we can write
IBS
(1.16) σLoc (K, T ) = x ∂IBS
.
1 − IBS ∂x

The solution of this ODE is


Z x
1 1 1
= dy
IBS (x) x 0 σLoc (y)
which follows by inserting the result in the ODE. In other words, at very short times
to expiration, the implied volatility is the harmonic mean of the local volatility as
a function of ln(S/K) between spot and strike. 

1.4. Implied Volatility in Terms of Local Volatility


Equation (1.11) expresses local volatility in terms of implied volatility. Is it
possible to express implied volatility as a function of local volatility? Given the
complicated expression (1.11), only an approximative solution is obtainable. There
are different approaches to obtain the approximation.
14 1. VOLATILITY MODELS AND VOLATILITY SURFACE

We first consider a power series approach where both local volatility and implied
volatility are represented as power series in the time to maturity variable T , i.e. we
set

X ∞
X
i
(1.17) IBS (y, T ) = Ii (y)T , σLoc (y, T ) = σLoc,i (y)T i
i=0 i=0

with y = ln(K/F0,T ) the moneyness variable. Using the moneyness variable y, the
Dupire relationship between local and implied volatility reads:

2 2T IBS ∂I∂T
BS
+ IBS2
(1.18) σLoc (y, T ) =   2  2
2 I2 T 2 y ∂IBS
T IBS ∂ ∂yIBS
2 − BS4 ∂IBS
∂y + 1− IBS ∂y

The first order approximation T 0 implies that all terms which have a coefficient
proportional to T or T 2 do not contribute. Hence, inserting the power series (1.17)
into (1.18) implies

2 I02 (y) 0
(1.19) σLoc,0 (y) =  2 , up to order T .
y ∂I0 (y)
1− I0 (y) ∂y

1 ∂f (y) ∂ ln f (y)
Taking the positive square root of the last equation and using f (y) ∂y = ∂y
we obtain the ODE
 
1 ∂ y
(1.20) = , up to order T 0 .
σLoc,0 (y) ∂y I0 (y)
Integrating w.r.t. y we obtain the following representation of implied volatility in
terms of local volatility
y
I0 (y) = R y du
, up to order T 0 ,
0 σLoc,0 (u)
+ c
with c the integration constant. To fix the constant c, one imposes that for y → 0
Rb
implied volatility to zeroth order remains finite. Since a f (x)dx = f (w)(b − a) for
a given value w, in the limit y → 0 for c = 0 the right hand side behaves as
y
1 = σLoc,0 (0) , , up to order T 0 ,
σLoc,0 (0)y

i.e. is finite if local volatility is finite for y = 0. Therefore, c = 0 follows. If we


define by
∂IBS
S(T ) = K |K=S
∂K
the implied volatility skew and by
 
∂ σLoc (F0,t )
S(T ) = F0,T
∂F0,T F0,T
the local skew, then the relationship up to order T 0 given in (1.20) implies
S(T ) = 2S(T ) , up to order T 0 .
Hence, local skew is twice implied volatility skew. This confirms the result (2) ob-
tained in Proposition 1.2.
1.4. IMPLIED VOLATILITY IN TERMS OF LOCAL VOLATILITY 15

We consider next the approximation up to order T 1 of (1.18). First, the term


proportional to T 2 can be neglected. Second, each term with is proportional to T
can only contain I0 and σLoc,0 expressions of the power series to match the right
expansion order. Hence, we get:

2 2T I0 ∂I 2
∂T + IBS
0

(1.21) (σLoc,0 (y) + σLoc,1 T + . . .) =  2 .


2 y ∂IBS
T I0 ∂∂yI20 + 1 − IBS ∂y

Continuing, we get:
(1.22)
 2
2 ∂ 2 I0 y ∂IBS 2 ∂I0 2
σLoc,0 (y)T I0 2 + 1 − (σLoc,0 (y) + σLoc,1 T + . . .) = 2T I0 +IBS .
∂y IBS ∂y ∂T
2
The next steps in the calculation are: First, insert the power series for IBS and for
y ∂IBS
IBS ∂y . Second, consider the zeroth order equation, i.e. the zeroth order term
I1 (y)
cancel. Performing these two calculation steps we obtain with g(y) = I0 (y) :

∂ I0 1 ∂2 I0 σLoc,1
(1.23) y g(y) + 2 g(y) = σLoc,0 2 I0 (y) + , up to order T 1 .
∂y ILoc,0 2 ∂y σLoc,0

This is a linear, inhomogeneous first order ODE for the function g. Solving this
equation with the boundary condition that the function must be finite for y to
infinity, we get up to order T 1 :

 2
1 I0 (y)
g(y) = −
2 y
   Z y   2
I0 (y)2 σLoc,1 (y 0 ) ∂ y0
(1.24) × ln − dy 0 .
σLoc,0 (y)σLoc,0 (0) 0 σLoc,0 (y 0 ) ∂y 0 I0 (y 0 )

Using the definition of the function g it follows that implied volatility up to order
1 is expressed as a function of local volatility only if we too use the expression
of implied volatility up to order 0 in terms of local volatility. This result can be
further simplified by approximating the integral using the mean-value theorem and
by approximation the log-term, see Labordere (09) Chapter 5.2 for details.

A different approach to express implied volatility in terms of local volatility


is to start with the pricing PDE for a call option both for the Black and Scholes
model and and local volatility model. Let CBS be the price of an European call in
the Black and Scholes model and CLoc the price in the local volatility model. The
pricing equation in the variables time t and forward F = F (t, T ) reads for both
options:

∂CBS 1 2 ∂ 2 CBS
+ σBS F2 = 0
∂t 2 ∂F 2
2
∂CLoc 1 ∂ CLoc
(1.25) + σLoc (t, F )2 F 2 = 0
∂t 2 ∂F 2
16 1. VOLATILITY MODELS AND VOLATILITY SURFACE

∂ 2 CBS
If we set h(t, F ) = CBS − CLoc and define the Gamma ΓBS = ∂F 2 , taking the
difference of the two pricing equations leads to a PDE for h:

∂h ∂2h 1
(1.26) + σLoc (t, F )2 F 2 + F 2 ΓBS (σLoc
2 2
− σBS ) = 0
∂t ∂F 2 2
Since both options have the same payoff at maturity T , the terminal condition
h(T, F ) = 0 holds. Using the Feynman-Kac theorem for the inhomogeneous diffu-
sion equation, the solution reads

1  2 2 2

(1.27) g(0, F0 ) = E F ΓBS (σLoc − σBS ) .
2

By definition of implied volatility at initiation g(0, F0 ) = 0 holds true. But this


2
implies, that σBS becomes equal to implied volatility which depends on the variables
K, T since t, F are integrated out. But this expression can be taken out of the
integral in (1.27) leading to

RT R∞
  F 2 σLoc
2
(t, F )ΓBS (t, F )p(t, F |0, f0 )dtdF
2 E F 2 ΓBS σLoc
2
0 0
IBS (K, T ) = = .
E [F 2 ΓBS ] RT R∞
F 2Γ BS (t, F )p(t, F |0, f0 )dtdF
0 0
(1.28)

This representation can be used to obtain an approximation of implied volatility in


terms of local volatility as follows:

Proposition 1.4. Assume that the forward F (t, T ) satisfies under the forward
measure QT the driftless SDE

dF (t, T ) = F (t, T )(σ + (t, F (t, T ))dW (t) .

The the first order approximation of implied volatility in  reads:


(1.29)
Z 1 Z ∞  √ 
du 1 2 K
IBS (K, T ) ∼ σ + 2σ dt √ e− 2 u  tT, F (0, T )et ln F (0,T ) + t(1−t)σu
0 −∞ 2π

See Labordere (09) Chapter 5.3 for a proof of the Proposition which uses (1.28).
Note that the function σ + (t, F (t, T ) represents local volatility, i.e. σLoc = σ +
(t, F (t, T ), which is assumed to be of the form constant plus a perturbation given
by the function .

1.5. Local variance and instantaneous variance


We derive the relationship between local variance and instantaneous variance,
i.e. the variance which is for example given in a stochastic volatility model such as
Heston’s one. We only assume that the underlying process is a diffusion - but no
further specification is assumed.
1.5. LOCAL VARIANCE AND INSTANTANEOUS VARIANCE 17

Let f (S) = (S − K)+ and applying Tanaka’s Formula, i.e. the generalization
of Itô’s Lemma for non-differentiable functions f , we get1
 
dT C = dT e−rT E(S − K)+
 
= EdT e−rT f (S)
 
−rT 0 1 00 2
= e E f (ST ) dST + f (ST ) dST − rf (ST ) dT
2
 
−rT 0 1 00 2 2
= e E f (ST ) (rSt dT + σT ST dWT ) + f (ST )σT ST dT − rf (ST ) dT
2

What is f 0 (ST )?

d
f 0 (ST ) = (S − K)+
dST
(
1 if ST ≥ K
=
0 otherwise
=: H(ST − K),

where H denotes the heaviside function.


What is f 00 (ST )?

d
f 00 (ST ) = H(ST − K)
dS
( T
0 if ST 6= K
=
∞ if ST = K
=: δ(ST − K),

where δ denotes the Dirac delta function. Therefore,


 
−rT 1 2 2 +
dT C = e E H(ST − K)rST + δ(ST − K)σT ST − r(ST − K) dT.
2

Since
(
+ rST − r(ST − K) if ST ≥ K
H(ST − K)rST − r(ST − K) =
0 otherwise
(
rK if ST ≥ K
=
0 otherwise
= rKH(ST − K),

then
 
1
dT C = e−rT E rKH(ST − K) + δ(ST − K)σT2 ST2 dT.
2

1See the Notes Stochastic Volatility I, Section ”Generalized Functions 2.3” for some details.
18 1. VOLATILITY MODELS AND VOLATILITY SURFACE

So
 
∂C 1
= e−rT E rKH(ST − K) + δ(ST − K)σT2 ST2
∂T 2
e −rT  
= e−rT rKE [H(ST − K)] + E δ(ST − K)σT2 ST2
| {z } | 2 {z }
(A)
(B)

First consider term (A):


∂C ∂  
= e−rT E (ST − K)+
∂K ∂K 

= e−rT E (ST − K)+
∂K
= e−rT E [−H(ST − K)] ,
so that
∂C ∂C
= −rK + (B).
∂T ∂K
Consider now term (B) and let
• p(ST , VT ) denote the joint density of (ST , σT2 ) and
• p(ST ) denote the marginal density of ST ,
then
e−rT  
(B) = E δ(ST − K)σT2 ST2
2
Z Z
e−rT
= v(s)s2 δ(s − K)p(s, v) ds dv
2
Z
e−rT 2
= K v(K)p(K, v) dv.
2
Therefore, Z
∂C ∂C e−rT 2
= −rK + K vp(K, v) dv.
∂T ∂K 2
Using the formula for local implied volatility (1.9), we get
∂C ∂C
2 ∂T + rK ∂K
σLoc (K, T ) = 1 2∂ C2
2 K ∂K 2
e−rT
2
R
2 K vK p(K, v) dv
(1.30) = 1 2 ∂2C
.
2 K ∂K 2
We may then simplify numerator and denominator of (1.30) and derive
∂2C ∂
= −e−rT E [H(ST − K)]
∂K 2 ∂K 

= −e−rT E H(ST − K)
∂K
= e−rT E [δ(ST − K)]
Z
−rT
= e δ(ST − K)p(ST ) dST

= e−rT p(K),
1.6. VOLATILITY STATICS UNDER NO ARBITRAGE CONDITION 19

so
2 R
2 e−rT K2 vK p(K, v) dv
σLoc = K 2
e−rT 2 p(K)
R
vK p(K, v) dv
=
p(K)
Z
1
= vK p(K, v) dv
p(K)
Z
= vK p(ST , VT )(v|ST = K) dv
 
= E σT2 |ST = K .
We finally get that implied local volatility is equal to the risk neutral expectation
of the instantaneous variance conditional on the final stock price ST being equal to
the strike price K.

The result holds true for any diffusion model of the underlying value. No specifi-
cation of the volatility function is assumed.

1.6. Volatility statics under no arbitrage condition


In this final section, we analyze the behavior of implied volatility IBS (K, T ) for
fixed t. We examine here the implications of various assumptions on the shape of
the implied volatility surface with only minimal assumptions of no-arbitrage.
Consider the FX market for equally strong currencies: the empirically observed
volatility smile reveals that I(K) is greater for values of K further away from the
at-the-money value than for K near-the-money. In the equity market, the smile
is rather a skew: at-the-money implied volatility slopes downward and the skew is
more prominent for smaller rather than larger values of K.
As for the relationship between implied volatility and maturity, a market’s rule
of thumb states that the smile or skew flattens out as maturity T increases and
that the skew’s slope decays at a rate ∼ √1T .

Proposition 1.5. Under the (sole) assumption of absence of arbitrage if K1 <


K2 , then C(K1 ) ≥ C(K2 ) and P (K
K1
1)
≤ P (K2)
K2 .

Proof. The proof is left as an exercise for the reader. 

Proposition 1.6. Assume C and P are differentiable, then we get slope bounds
 
∂C ∂ P
≤ 0, ≥ 0.
∂K ∂K K
Also assume we are in the Black-Scholes framework, so that
N (−d1 ) ∂I N (d2 )
−√ ≤ ≤√ .
0
T N (d1 ) ∂x T N 0 (d2 )
Proof. This proof, too, is left as an exercise for the reader. 
20 1. VOLATILITY MODELS AND VOLATILITY SURFACE

1−N (d)
Using Mill’s ratio, defined by R(d) := N 0 (d) , we get

R(d1 ) ∂I R(−d2 )
− √ ≤ ≤ √ .
T ∂x T
K
Remark 1.7. Behavior at-the-money, i.e. x = log S er(T −t)
= 0. Then
√ √
x σ T −t σ T −t
d1,2 = − √ ± =±
σ T −t 2 2

I(0, T ) T
d1,2 (x = 0, t = 0) = ±
2
So that d1,2 (0, 0) → 0 as T → 0, and d1,2 (0, 0) → ∞ as T → ∞.
Let us consider the behavior of the implied volatility surface for both very short
and very long maturities:
• Take T → 0. Since R(0) > 0 constant, ∂I(0,T ∂x
)
must have the short-dated
behavior
 
∂I(0, T )
= O √1 ,
∂x T
∂I
i.e. ∂x grows no faster than √1 and, thus, the rule of thumb holds for
T
short-dated options.
• Take T → ∞. We have seen above that d1,2 (0, 0) → ∞, but what about
R(d → ∞)?
2
N (d) → 1, N 0 (d) ∼ e−d → 0,

then
1 − N (d) 0
R(d) = → .
N 0 (d) 0
Applying de l’Hôpital gives
2
−N 0 (d) e−d 1
R(d) = ≈ − 2 ∼ ;
N 00 (d) −2d e−d d

therefore,
1
R(d → ∞) → .
d

∂I(0, T ) 1 1

∂x ∼ d = T , T → ∞.

 
∂I(0, T )
=O 1 , T → ∞,
∂x T
which shows that the rule of thumb stated above does not hold for long-
dated options.
1.7. MIXING THEOREMS 21

(1.31) CT D (S(0), σ(0), T ) = CBS (S(0), σ eff , T )

1.7. Mixing Theorems


What mixing theorems do is express the option prices in the more complicated
models as a weighted sum of the option prices in simpler base models. The simpler
base model we prefer is the Black and Scholes model.
To derive the Mixing Theorem, we consider the time-dependent volatility in
Section 1.2. If CT D represents an option price in a a time dependent model with
maturity T , we then know that
where the effective variance
ZT
2,eff 2 1
σ = σ̄ = σ 2 (s) ds
T
0
is the mean variance over a time interval. The reason for this relationship between
CT D and the Black and Scholes formula is due to the fact that ln ST /S0 is normal
distributed with mean rT − 12 σ̄ 2 T and variance σ̄ 2 T , i.e. the distribution which
leads to Black and Scholes formula.

What happens if volatility is stochastic? Assume a model, where the two state
variables S and v = σ 2 are both driven by diffusion processes such as the Heston
model but with the sources of risk being independent: The two Brownian motions
which drive the two state variables have correlation zero. Then the price of the
option in such a stochastic volatility setup is as usual given by the discounted
expected terminal payoff, i.e.:
Z
CSV (St , vt , t) = e−r(T −t) f (ST )p(ST |St , vt )dST

with f the payoff at maturity and p the conditional distribution function determined
by the two diffusions for the two state variables. The basic observation then is that
for any density functions p, g, h and any random variables x, y, z the relationship
Z
p(x|y) = g(x|z)h(z, y)dz

holds. We apply this fact to our option pricing problem where we set z = σ̄ 2 , i.e.
mean variance over the time from 0 to T . We then have
Z Z
−r(T −t)
CSV (St , vt , t) = e f (ST )g(ST |St , σ̄ 2 )h(σ̄ 2 |St , vt )dST dσ̄ 2 .

This double integral can be written as


Z  Z 
CSV (St , vt , t) = e−r(T −t) f (ST )g(ST |St , σ̄ 2 )dST h(σ̄ 2 |St , vt )dσ̄ 2 .

If the correlation is zero, we claim that the inner integral is given by the Black and
Scholes price with a mean variance σ̄ 2 , i.e.
Z
(1.32) CSV (St , vt , t) = C(S(T ), T, σ̄ 2 )h(σ̄ 2 |St , vt )dσ̄ 2 .

This is the content of the Mixing theorem which was first derived by Hull and White:
This theorem is in particular useful if we simulate option prices. Since there is no
22 1. VOLATILITY MODELS AND VOLATILITY SURFACE

The price of a an option under stochastic volatility is equal to the weighted


sum/integral of Black and Scholes option prices if correlation between the stock
and volatility is zero. In the Black and Scholes formula, the mean path volatility
enters.

closed form solution for options in say the Heston model, we have to simulate the
prices. Suppose that we do not consider the simplification using the Fourier the-
ory but directly simulate the expected payoff formula using Monte Carlo. Then,
since there are two state variables we have to perform a two-dimensional simula-
tion. The Mixing Theorem reduces this, since one dimension - the stock price - can
be integrated, i.e. one applies Black and Scholes formula with the mean variance
replacement. We consider this issue in detail below. We next justify the theorem.

First, if volatility is constant the outer integral is superfluous and we are back
in the time-dependent case. Second, assume the stochastic case. If σ 2 is stochastic,
there is no longer a single path as in the deterministic case which leads to the mean
variance σ̄ 2 , but there is a continuum of such paths. But all these paths lead to the
same terminal distribution of the stock price. Therefore also for stochastic volatil-
ity the terminal distribution of the stock price conditional on the mean variance
is lognormal. This proves the mixing theorem. Third, if correlation is different
from zero, then again log S(T )/S(0) conditional on σ̄ 2 has a normal distribution.
But unlike in the case with zero correlation, both the mean and variance of the
distribution do not only depend on time, the risk free rate and the mean variance
but also on other parameters - in particular on the paths which the volatility takes
up to maturity.

Does the mixing theorem holds for non-zero correlation? To give an answer,
we consider Heston’s model:
p
dSt = rSt dt + Vt St dZ1
and p
dVt = −λ(Vt − V̄ ) dt + η Vt dZ2 .
We perform two transformations on the dynamics. First, the two Brownian motion
are correlated. Writing
p
dZ1 = ρdZ2 + 1 − ρ2 dZ
where dZ is another standard Brownian motion that is independent of Z2 , we can
insert this in the dynamics where the two risk sources are then independent. Second,
use a discrete time version of the dynamics. This has two advantages. First, we
do not need to consider advanced mathematics such as path integrals. Second, the
formula in discrete time can be used for Monte Carlo simulations of option prices.
The discrete version of the Heston dynamics reads:
p p √
∆St = rSt ∆t + Vt St (ρZ̃2,t + 1 − ρ2 Z̃t ) ∆t
and p √
∆Vt = −λ(Vt − V̄ )∆t + η Vt Z̃2,t ∆t,
where Z̃t is the result of drawing a standard normal at the discrete time step t. The
same applies to the other random variable. The price of a call in this stochastic
1.7. MIXING THEOREMS 23

volatility setup is given by the risk-neutral formula


CSV (S(0), V (0)) = e−rt E[(S(T ) − K)+ ] .
In discrete form, the formula becomes
 
CSV (S(0), V (0)) = e−rt E (S(T ) − K)+
Z Z TY−∆t
2 2 dZ̃2,t dZ̃t
−rT
(1.33) = lim e . . . (S(T ) − K)+ e−1/2(Z̃2,t +Z̃t ) .
∆t→0 R R t=0

The formula reflects that the expected value is approximated by a sequence of
discrete time integration. Each integration has a two-dimensional density of two
uncorrelated standard normal Gaussian. The exact value follows if the limit spacing
goes to zero. The advantage of the discretization is due to the following. Suppose
2
that we first made a complete sequence of drawing of the random variable Z2,t for
all t = 0, ∆t, 2∆t, . . . , T . Since we know this random variable, the solution of the
difference equation for the volatility is a deterministic function. We then solve the
S-equation, i.e. one gets

(1.34) !
TX
−∆t TX
−∆t p √
S(T ) = S(0)eY (T ) exp rt − 1/2 (1 − ρ2 )Vt ∆t + 1/2 (1 − ρ2 )1/2 Vt Z̃t ∆t
t=0 t=0

with
TX
−∆t TX
−∆t
2

Y (T ) = −1/2 ρ Vt ∆t + ρZ̃2,t ∆t.
t=0 t=0
The crucial point is, that (1.34) is the solution to a deterministic volatility problem-
a Black-Scholes type problem. The solution starts with the p value S(0) eY (T ) instead

of S(0) and the volatility is altered to the effective volatility Vteff = (1−ρ2 )1/2 Vt .
This observation implies that we can interpret the entire set of integrations in
(1.34) over the Z̃t variables, conditional on holding the Z̃2,t fixed, as a deterministic
volatility problem with the two modifications just explained. Since the variance is
replaced by an effective variance, we define:
RT RT √
ρ2 Vt dt+
Steff = lim S(0) eY (T ) = S(0) e−1/2 0 0
ρ Vt dZ2,t
∆t→0
T −∆t Z
1 X 1 T
(1.35) Vteff = lim (1 − ρ2 )Vt ∆t = (1 − ρ2 )Vt dt.
∆t→0 T T 0
t=0

If we denote the remaining integrations over the volatility process in (1.36) over
the variables dZ2,t by EZ2,t 2 [. . .], then the above shows that the call option price is

given by
Z Z TY−∆t
2 dZ̃2,t
eff eff eff eff
EZ2,t
2 [CBS (S
T , V T )] = lim . . . cBS (S T , V T ) e−1/2Z2,t √ .
∆t→0 R R t=0

We have so obtained the general form of the mixing theorem:

 
CSV (S(0), V (0)) = EZ2,t
2 CBS (STeff , VTeff ) .
24 1. VOLATILITY MODELS AND VOLATILITY SURFACE

This theorem was proved by Romano and Touzi, 1997. In words again, the
option value under stochastic volatility is a weighted sum or mixture of the Black-
Scholes values with an effective stock price and effective volatility. The effective
variables depend only upon the volatility process. Hence, the problem reduces to
a pricing expectation over the risk-adjusted volatility process alone. Note that for
zero correlation S eff = S(0). In general, one can not find closed form solutions for
the mixing problems - else the whole lecture so far would have been superfluous.
There are some specific cases where a closed-form solution exists. The value of the
Theorem is on the conceptual side, Monte Carlo simulations and approximative
solution methods.

1.8. Appendix: Forward, backward equations


An equation
Z t Z t
(1.36) S(t) = S(0) + µ(s, S(s))ds + σ(s, S(s))dW (s) ,
0 0
is called a stochastic differential equation and a solution of such an equation is a
diffusion. By a solution we mean:
Definition 1.8. Consider a stochastic base (Ω, A, F = (Ft )t , P ) where the
Brownian motion in (1.36) is adapted to the filtration. A solution to the SDE
(1.36) is an Ft -adapted process S(t) such that:
Rt Rt
(1) For any t ≥ 0,the integrals 0 µ(s)ds, 0 σ(s)dW (s) exist, i.e.
Z t Z t
|µ(s)|ds < ∞, |σ(s)|2 ds < ∞ .
0 0
(2) (S(t)) satisfies (1.36).
Sufficiency conditions which ensure that the coefficient functions µ, σ possess
the desired integrability conditions are similar to the conditions for ordinary differ-
ential equations. We state them without proof.
Theorem 1.9. If µ, σ are continuous functions and if there exists a constant
k such that:
|µ(t, x) − µ(t, y)| + |σ(t, x) − σ(t, y)| ≤ K|x − y|
(1.37) |µ(t, x)| + |σ(t, x)| ≤ K(1 + |x|) ,
then for any T ≥ 0 the SDE (1.36) admits a unique solution.
The first condition in the theorem is called a locally Lipschitz condition in x
uniformly in t and the second condition is a linear growth condition in x.
Using the differential notation and the properties of the Itô integral, we get
E[dS(t)| S(t) = s] = µ(S(t), t)dt + o(t)
2
E[(dS(t)) | S(t) = s] = σ 2 (S(t), t)dt + o(t)
n
(1.38) E[(dS(t)) | S(t) = s] = o(t) , n ≥ 3 .
This conditions not only holds for the differential increments but also for finite time
steps. For example
E[S(t + s) − S(s)| S(s) = s] = µ(S(t), t)t + o(t)
1.8. APPENDIX: FORWARD, BACKWARD EQUATIONS 25

and similar for the higher moments. A solution of the SDE (1.36) satisfies the
important Markov property:
Definition 1.10. A Ft -adapted process S(t) satisfies the Markov property, if
for any s and t with s ≤ t and any bounded, measurable function f
(1.39) E[f (S(t))|Fs ] = E[f (S(t))|S(s)]
holds.
A diffusion S(t), i.e. a solution of the SDE (1.36), satisfies the Markov property.
Let S(t) be a solution of the SDE (1.36). We define the expected value u of some
payoff f at maturity time T > t given that S(t) = s by
(1.40) u(t, s) = E[f (S(T ))|S(t) = s] := E S(t)=s [f (S(T ))] .
We define the infinitesimal generator A of the diffusion S by:
∂ 1 ∂2
+ σ 2 (t, s) 2 .
A := µ(t, s)
∂s 2 ∂s
The next theorem of Feynman-Kac relates PDEs to SDEs.
Theorem 1.11 (Feynman-Kac). Let S(t) be a solution of the SDE (1.36),
u(t, s) ∈ C 2,1 and µ, σ satisfy regularity conditions. Then u solves the backward
Kolmogorov PDE

(1.41) ( + A) u(t, s) = r(t, s) u(t, s) , u(T, s) = f (s) .
| ∂t {z } | {z } | {z }
(risky asset dyn.) (int. rate) (Claim)

The unique continuous solution of the PDE is given by the Feynman-Kac formula
RT
(1.42) u(t, s) = E S(t)=s [e− t
r(u,S(u))du
f (S(T ))] .
Basically, the above theory states that given a SDE for assets, then a final payoff
of a claim is the solution of the associated backward Kolmogorov PDE. Since the
solution of the SDE is a Markov process, it has a well defined transition probability
p(u, z|t, s) which describes the probability of being at z at time u given at time t the
process started in s. This transition probability satisfies also a PDE, the forward
Kolmogorov equation, i.e.

(1.43) − p + A∗ p = 0 , u > t
∂u
with the initial condition
(1.44) p(t, z|t, s) = δ(z − s)
where δ(z − s) is Dirac’s delta function. We consider the forward equation and the
construction of the adjoint operator A∗ . That for, we introduce the real Hilbert
space H of square integrable functions whose elements are at least twice continu-
ously differentiable with inner product
Z
(f, g) = f (x)g(x)dx , f, g ∈ H .
R
The adjoint operator A∗ to the infinitesimal operator A is defined similar than in
the finite dimensional case by
(Af, g) = (f, A∗ g) .
26 1. VOLATILITY MODELS AND VOLATILITY SURFACE

This definition allows us to calculate the adjoint using integration by parts. We do


the calculation for the drift term µ in the generator. The volatility part is done in
the same way. We write A|µ for the drift part of the generator.
Z
(A|µ f, g) = (A|µ f (x))g(x)dx
R
Z
∂f
= µ g(x)dx
R ∂x
Z

= µf g(x)|R − f (x) (µg(x))dx
R ∂x
Z

= − f (x) (µg(x))dx
∂x
ZR
= − f A∗ |µ g(x)dx
R
= −(f, A∗ |µ g(x)) .
where we used the decay properties of f, g at infinity. Summarizing,
∂ ∂
µ f (x) → − (µ(x)g(x))
∂x ∂x
under adjointness. With the similar calculation for the volatility part we get for
the adjoint operator:
∂ ∂2
(1.45) A∗ p(z) = − (µ(z)p(z)) + 2 (σ(z)p(z)) .
∂z ∂z
We finally discuss the initial condition (1.44). We denote by q(u, z) the probability
density of z at time u. Using the transition probability, we have the transition law
Z
q(u, z) = p(u, z|t, s)q(t, s)ds .
R
If t → s, the integral should become
Z
p(u, z|t, s)q(t, s)ds → q(t, s) .
R
The function which annihilates the integral in this sense is by definition the Dirac
delta function.
Index

Dirac Delta function, 12

Effective volatility, 17

Fokker-Planck equation, 3

Heaviside function, 12
Heston model, 17

Implied volatility skew, 9


Implied volatility surface asymptotic
behavior, 15
Interpretation of Dupire’s equation, 4

Local Skew, 10
Local variance
and instantaneous variance, 11
local variance, 2
and implied variance asymptotic
properties, 6
Dupire equation for, 3
in terms of Black-Scholes implied
variance, 5

Mill’s ratio, 15
Mixing Theorem, 16
for correlated securities, 17

27

Potrebbero piacerti anche