Mixing Models To Capture Stock Price Volatility Markov Lecture INFORMS Annual Meeting October 12-15, 2008 Washington, DC

Download as pdf or txt
Download as pdf or txt
You are on page 1of 35

MIXING MODELS TO CAPTURE

STOCK PRICE VOLATILITY

Markov Lecture
INFORMS Annual Meeting
October 12–15, 2008
Washington, DC

Steven E. Shreve
Department of Mathematical Sciences
Carnegie Mellon University

Slides available at
www.math.cmu.edu/users/shreve

1 / 35
Outline

1. The Black-Scholes-(Merton) model


1.1 Derivation of the Black-Scholes-(Merton) formula
1.2 Risk-neutral pricing
1.3 The role of volatility
2. Non-constant volatility
2.1 Mixture of Black-Scholes-(Merton) models
2.2 Stochastic volatility models
2.3 Implied volatility models
2.4 Local volatility model
3. New results on mixture models and local volatility

2 / 35
Black-Scholes-(Merton)
European call: Right to buy one unit of a risky asset at an
expiration time T at a strike price K agreed upon today (time
zero). The call pays (ST − K )+ , max{ST − K , 0} at expiration.

The price today of the call should be

S0 N(d+ ) − e −rT KN(d− ),

where
I S0 is today’s price of the risky asset,
I r is the continuously compounding rate of interest,
I N is the standard cumulative normal distribution function,
I    
1 S0 1
d± = √ log − r ± σ2 T .
σ T K 2
I σ, which is positive, is the risky asset price “volatility.”
3 / 35
What Black-Scholes-(Merton) teaches us

Preposterous result: The price of the European call does not


directly depend on the expected rate of growth of the underlying
asset price.

Insightful result: The volatility σ of the risky asset is the key


parameter.

4 / 35
Derivation of Black-Scholes-(Merton)
Replicate the call payoff by trading.

The two assets for trading are:


I a money market account with constant rate of interest r (we
assume r = 0),
I the risky asset with price satisfying

dSt = αt St dt + σSt dWt ,

where
I W is a Brownian motion on (Ω, F, P) relative to a filtration
{Ft }0≤t≤T ,
I αt is an adapted process,
I σ is a positive constant.
Because αt is not required to be constant, nor even
deterministic, St is not assumed to be a geometric Brownian
motion.
5 / 35
Portfolio

Create a portfolio of cash and the risky asset whose value at each
time t is the value of the call at that time.
I Xt – Value of the portfolio at time t.
I ∆t – Number of units of the risky asset held by the portfolio
at time t.
I Xt − ∆t St – Amount of cash held by the portfolio at time t.
I Evolution of portfolio value –

dXt = ∆t dSt , 0 ≤ t ≤ T.

6 / 35
Matching evolutions
Let c(t, St ) denote the price of the call at time t when the
underlying asset price is S(t). Note: c(T , s) = (s − K )+ for all
s > 0.

Itô’s formula:
   1 
dc t, St = ct t, St dt + cs t, St dSt + css t, St dSt dSt
2 | {z }
σ 2 St2 dt
?
= dXt
= ∆t dSt .

Need ct (t, s) + 21 σ 2 s 2 css (t, s) = 0 for 0 ≤ t < T and s > 0.


Black-Scholes-(Merton) formula provides the solution to this PDE.

Replication:
Start with X0 = c(0, S0 ) and take ∆t = cs (t, St ), 0 ≤ t ≤ T .
Then XT = c(T , ST ) = (ST − K )+ .
7 / 35
Risk-neutral (martingale) probability measure1,2

We write
hα i
t
dSt = αt St dt + σSt dWt = σSt dt + dWt = σSt d W
ft ,
σ
where Z t
αu
W
ft = du + Wt .
0 σ
Using Girsanov’s Theorem, we construct a new probability measure
e under which W
P f is a Brownian motion. Under P,
e the risky asset
price is a geometric Brownian motion and a martingale. P
e is called
the risk-neutral measure.

1
Harrison, J. M. & Kreps, D. M. (1979) Martingales and arbitrage in
multiperiod security markets, J. Econom. Theory 20, 381–408
2
Harrison, J. M. & Pliska, S. R. (1981) Martingales and stochastic
integrals in the theory of continuous trading, Stoch. Proc. Appl. 11, 215–260.
8 / 35
Risk-neutral pricing

dSt = σSt d W
ft .
For 0 ≤ t ≤ T and s > 0, define
e ST − K + St = s .
  
f (t, s) = E

Then
f (T , s) = (s − K )+ = c(T , s), s > 0.
The Kolmogorov backward equation for f (t, s) is
1
ft (t, s) + σ 2 s 2 fss (t, s) = 0.
2
In other words, f (t, s) = c(t, s).

Risk-neutral pricing formula:


e ST − K + St = s .
  
c(t, s) = E
9 / 35
Criticism of Black-Scholes-(Merton)
Fix an underlying asset S. There is no single volatility parameter σ
that causes the Black-Scholes-(Merton) formula to give the correct
(market) price of options written on that asset with a variety of
different strikes and expiration times.

Implied Implied volatility varies with strike and expiration time.


Vol
1.00 Call options on MSFT expiring September 19, 2008.
September 4, 2008, closing prices.
0.75

0.50

0.25
Stock
price
22 23 24 25 26 27 28 29 30 31 32
Strike
10 / 35
Why does this matter?
Pricing – We want a model that we can calibrate to options that
are traded and then use to price options that are not traded.

Vanilla options by interpolation – If we know the prices (and hence


the implied volatilites) for calls with strikes K0 and K2 and want to
price a call with strike K1 ∈ (K0 , K2 ), we can interpolate the
implied volatilities and use Black-Scholes-(Merton) with the
interpolated volatility parameter.

Exotic options – Suppose we want to price a knock-out barrier


option, which pays
(ST − K )+ lI{MT ≤B} ,
where
MT , max St .
0≤t≤T

Now what volatility should we use?


11 / 35
Why does this matter?

Hedging
I Hedge a book of options. Negative of replication.
I Want total portfolio (book plus hedge) to be insensitive to
price changes.
I Requires a single model for the underlying asset price.

12 / 35
Four ways to proceed

1. Mixture of Black-Scholes-(Merton) models


2. Stochastic volatility models
3. Implied volatility models
4. Local volatility model

13 / 35
Mixture of Black-Scholes-(Merton) models

Recall dSt = σSt d W


ft , so
 
fT − 1 σ 2 T
ST = S0 exp σ W .
2

Let σ be a random variable, taking either the positive value c1 or


the positive value c2 . Then the distribution of ST is a mixture of
log-normals.
I For fixed T , mixture model fits well to call and put prices.
I Easy to implement.
I Not a reasonable dynamic model. Once σ has been realized, it
will immediately be observed, and the mixture property is lost.
I No good for hedging.

14 / 35
Four ways to proceed

1. Mixture of Black-Scholes-(Merton) models


2. Stochastic volatility models
3. Implied volatility models
4. Local volatility model

15 / 35
Stochastic volatility models

Example (Heston3 )


dSt = αt St dt + vt St dWt1 ,

= λ v − vt dt + β vt dWt2 ,

dvt
hW , W 2 it
1
= ρt.

Closed-form formulas for Fourier transforms of option prices.

3
S. Heston (1993) A closed-form solution for options with stochastic
volatility, with applications to bond and currency options, Review Financial
Studies 6, 327–343.
16 / 35
Stochastic volatility models (continued)

Example (Fouque, Papanicolaou, Sircar4 )

dSt = αt St dt + e Yt St dWt1 ,
λ β
= 2 Y − Yt dt + dWt2 ,

dYt
ε ε
hW 1 , W 2 it = ρt.

Asymptotic expansion of call prices in powers of ε, with zero-order


term given by Black-Scholes-(Merton) formula.

4
J.-P. Fouque, G. Papanicolaou and R. Sircar (2000) Derivatives in
Financial Markets with Stochastic Volatility, Cambridge University Press.
17 / 35
Comments on stochastic volatility models

Calibration:
I First choose parameters, and then “run” the model to output
option prices.
I Compare model outputs with market prices. Try to choose
better parameters, typically to reduce the sum of squared
errors. This is a nonlinear optimization problem in which each
function evaluation is expensive.
I Subsequent recalibrations result in paper profits or losses and
upset hedges.

18 / 35
Four ways to proceed

1. Mixture of Black-Scholes-(Merton) models


2. Stochastic volatility models
3. Implied volatility models
4. Local volatility model

19 / 35
Implied volatility models

Start with c 0, S0 ; t, K for 0 < t ≤ T and K > 0. For each t and
K , define the implied volatility σ(t, K ) to be the solution of the
equation
 + 
 
  1 2  
c 0, S0 ; t, K = E S0 exp σ W (t) − σ t −K  .
e  f  
 2  
| {z }
St

Obtain thereby a implied volatility surface

σ(t, K ), 0 < t ≤ T , K > 0.

Build a stochastic model to evolve the implied volatility surface


forward.

In contrast to stochastic volatility models, the initial market call


prices are an input, not an output, of implied volatility models.
20 / 35
Arbitrage problem for implied volatility models

I Assume that today σ(t, K ) = 0.20 for all t and K . This is a


flat implied volatility surface. (Black-Scholes-(Merton)
assumes the implied volatility surface is flat.)
I Suppose that tomorrow we will have another flat implied
volatility surface, at either 0.25 or 0.15. Today we do not
know which scenario will occur.
I At time zero, buy ∆i calls with strike Ki , i = 1, 2, 3.
I Choose non-zero ∆1 , ∆2 and ∆3 so that the portfolio has
value zero tomorrow, regardless of which scenario occurs.
This requires that we solve two homogeneous equations in
three unknowns.
I Price of portfolio must be zero today, or else there is an
arbitrage.

21 / 35
Arbitrage problem for implied volatility models

I A necessary and sufficient condition to rule out arbitrage is


known.5,6
I It is so complex that it has thus far prevented the
implementation of such a model.

5
P. Schönbucher, A market model for stochastic implied volatility, Working
paper, May 1999.
6
M. Schweizer and J. Wissel (2008), Term structure of implied volatilities:
absence of arbitrage and existence results, Math. Finance 18, 77-114.
22 / 35
Four ways to proceed

1. Mixture of Black-Scholes-(Merton) models


2. Stochastic volatility models
3. Implied volatility models
4. Local volatility model

23 / 35
Local volatility model – Implying risk-neutral densities from
call prices.

Suppose we are given c(0, S0 ; T , K ) for all K > 0. We can imply


the risk-neutral densities p(T , s) = p(0, S0 ; T , s) from the formula
e ST − K + S0
   
c 0, S0 ; T , K = E
Z ∞
= (s − K )p(T , s) ds.
K
Differentiate once:
Z ∞
∂ 
c 0, S0 ; T , K = −(K − K )p(T , K ) − p(T , s) ds
∂K K
Z ∞
= − p(T , s) ds.
K
Differentiate again:
∂2 
c 0, S(0); T , K = p(T , K ).
∂K 2
24 / 35
Implying the local volatility surface.7,8
Assume the evolution of the risky asset price:
 
dSt = σ t, St St d W ft = γ t, St d W ft , 0 ≤ t ≤ T.

Forward equation satisfied by transition density:

∂ 1 ∂2  2 
p(0, S0 ; T , s) = γ (T , s)p(0, S0 ; T , s) .
∂T 2 ∂s 2
Again we set p(T , s) = p(0, S0 ; T , s). Then
Z ∞
∂  ∂
c 0, S(0); T , K = (s − K ) p(T , s) du
∂T K ∂T
1 ∞ ∂2 
Z 
= (s − K ) 2 γ 2 (T , s)p(T , s) ds.
2 K ∂s

7
B. Dupire (1994) Pricing with a smile, Risk 7, 18–20.
8
E. Derman and I. Kani (1994) Riding on a smile, Risk 7, 32–39
25 / 35
Implying the local volatility surface.

Integrate by parts:

∂2  2
Z
∂  1 
c 0, S(0); T , K = (s − K ) γ (T , s)p(T , s) ds
∂T 2K ∂s 2
1 ∞
Z
∂ 2 
= − γ (T , s)p(T , s) ds
2 K ∂s
1 2
= γ (T , K )p(T , K ).
2
Recall
∂2 
p(T , K ) = 2
c 0, S(0); T , K .
∂K
Solve for the function γ 2 (T , K ) for T in some range and K > 0.
Recall σ(T , K )K = γ(t, K ).

26 / 35
Local volatility model – Practical issues

I Local volatility model takes call prices as inputs.


I There is no freedom to calibrate to exotic options.
I Volatility surface not stable over time.

27 / 35
Local volatility model – Theoretical issues
Theorem (First Fundamental Theorem of Asset Pricing9 )
Assume
I underlying asset is a semimartingale,
I call price processes are semimartingales,
I there is no free lunch with vanishing risk (essentially, no
arbitrage).
Then there exists a risk-neutral probability measure P
e under which
all these price processes are local martingales.
Under mild additional conditions, then S has the representation

dSt = γt · d W
ft ,

where γt is a vector-valued adapted process and W


ft is a vector of
independent Brownian motions under P. e
9
F. Delbaen and W. Schachermayer (1994) A general version of the
fundamental theorem of asset pricing, Math. Annalen 300, 463–520.
28 / 35
Local volatility model – Theoretical issues
Assume
dSt = γt · d W
ft .

Theorem (Krylov10 , Gyöngy11 )


Assume that the the largest eigenvalue of γt γttr is bounded and the
smallest eigenvalue is bounded away from zero, uniformly in t and
ω. Then there is a volatility surface σ(t, s) such that
I dStvs = σ(t, Stvs )Stvs dWtvs has a weak solution, unique in law;
I for each t ≥ 0, we have L(St ) = L(Stvs );
I prices of calls on S agree with prices of calls on S vs .

10
N. Krylov (1984) Once more about the connection between elliptic
operators and Itô’s stochastic equations, Statistics and Control of Stochastic
Processes, Steklov Seminar, 214–229.
11
I. Gyöngy (1986) Mimicking the one-dimensional marginal distributions of
processes having an Itô differential, Prob. Theory and Related Fields 71,
501–516.
29 / 35
Local volatility model – Mixture example
e 0 = c1 } = P{σ
Assume P{σ e 0 = c2 }, where c2 > c1 > 0. Assume

σt = σ0 so γt = σ0 St , 0 ≤ t ≤ T,

where
dSt = σ0 St d W
ft , 0 ≤ t ≤ T.
”Smallest eigenvalue” of γt2 = σ0 St2 is not bounded away from
zero.
Nonetheless, there is a local volatility surface σ(t, s), and the
stochastic differential equation

dStvs = σ(t, Stvs )Stvs dWtvs

has a unique strong solution.12


12
D. Brigo and F. Mercurio (2002) Lognormal-mixture dynamics and
calibration to market volatility smiles, Internat. J. Theoret. Appl. Finance 5,
427–446.
30 / 35
Local volatility model – Mixture example

I dSt = σ0 St d W
ft – choose a volatility at time zero and use it
throughout.
I dStvs = σ(t, Stvs )Stvs dWtvs – time-varying, state-dependent
volatility.
I If S0 = S0vs , then St and Stvs have the same one-dimensional
distributions for all t ≥ 0, so ...
I prices of calls on S agree with prices of calls on S vs .
I S vs makes sense as a dynamic model. S does not.

31 / 35
Dynamic mixture model - Outline of a general construction
Choose a sequence of partitions

Πn : 0 = T0 < T1 < T2 < · · · < Tn < ∞.

Flip a coin at time T0 to choose a volatility σ0 . If we use this


volatility throughout, we obtain the process S.

Use S to construct a second process S Πn as follows. At time Ti ,


e 0 = ci |ST = s}. Redraw the volatility with this
compute P{σ i
distribution and use it on [Ti , Ti+1 ).
I For each t, St and StΠn have the same distribution.
I Immediately after each Ti , observation of S Πn reveals the
volatility being used on [Ti , Ti+1 ), but not the volatility that
will be chosen at Ti+1 .
I As n → ∞, S Πn converges weakly to some S vs in C [0, ∞).

dS vs = σ(t, S vs )S vs dW vs , where (σ vs )2 (t, s) = E
e σ 2 St = s .

I t t t t 0
32 / 35
Extension to barrier options
Theorem (Brunick13 )
Assume
dSt = γt · d W
ft ,

and define
Mt , max Su .
0≤u≤t

Then there exists a function γ(t, s, m) and a weak solution to the


stochastic differential equation

dStvs = γ(t, Stvs , Mtvs ) dWtvs ,

where
Mtvs , max Suvs ,
0≤u≤t

such that for each t ≥ 0, we have L(St , Mt ) = L(Stvs , Mtvs ).


13
G. Brunick (2008) A weak existence result with application to the financial
engineer’s calibration problem, Ph.D. dissertation, Carnegie Mellon University.
33 / 35
Further discussion

Comments:

γ 2 (t, s, m) = E
e γ 2 St = s, Mt = m

I t
I The stated theorem is a corollary of a more general result.
I At the level of generality given here, there can be multiple
weak solutions to

dStvs = γ(t, Stvs , Mtvs ) dWtvs .

Some open problems:


I Obtain conditions under which uniqueness holds.
I Obtain a formula for γ(t, s, m) in terms of prices of calls and
barrier options.

34 / 35
Conclusions

Two observations about the development of a new generation of


models for option pricing and hedging.
I Critical importance for the practice of finance.
I Source of fascinating new problems in mathematics.

35 / 35

You might also like