Introduction Local volatility models Stochastic volatility models

Local and Stochastic Volatility

Johannes Ruf

University College London and

Oxford-Man Institute of Quantitative Finance
[email protected]

Slides available at
http://www.oxford-man.ox.ac.uk/~jruf

April, 2016
Introduction Local volatility models Stochastic volatility models

Some first remarks

• This mini course only touches on a few themes in the world of

local and stochastic volatility models.
• The course puts more emphasis on models used for pricing
and hedging than on models used for estimation.
• This presentation is partially based on notes by Michael
Monoyios and Sergey Nadtochiy.
Introduction Local volatility models Stochastic volatility models

Definition: spot volatility

• In the Black-Scholes (BS) model,

dSt = µSt dt + σSt dBt ,

the constant σ is the (spot) volatility of S.

• Observe that
  
1 St+∆t d
lim Var log = [log(S)]t = σ 2 . (1)
∆t↓0 ∆t St dt
• We can use (1) as a definition for spot volatility in general
models of the form

dSt = µt St dt + σt St dBt

(in particular, without jumps) and have

d
[log(S)]t = σt2 .
dt
Introduction Local volatility models Stochastic volatility models

Estimation: spot volatility

• Usually, the volatility is easier to estimate than the drift.
• Large amount of current research on high-frequency / tick
data.
• In a simple setup: assume discrete price observations
si = Si∆t (ω) for i = 0, . . . , n with n∆t = T .
• Define logarithmic returns ri := log(si /si−1 ) for i = 1, . . . , n.
• Itô’s formula yields
 
1 2
d log(St ) = µt − σt dt + σt dBt .
2

• Thus,
  Z i∆t
  Z i∆t
Si∆t 1 2
Ri = log = µt − σt dt + σt dBt .
S(i−1)∆t (i−1)∆t 2 (i−1)∆t
Introduction Local volatility models Stochastic volatility models

Estimation: spot volatility II

• Assume, for a moment, that µt ≡ µ and σt ≡ σ.

• Then,
  
1 2 2
Ri ∼ N µ − σ ∆t, σ ∆t .
2

b2 of σ 2 is, with
• Thus, maximum likelihood estimator σ
Pn
r¯ = i=1 ri /n,

n n
1 1X 1 X
b2 =
σ · (ri − r¯)2 = (ri − r¯)2 .
∆t n T
i=1 i=1

• How can we find an estimator for σt2 if σt is not a constant?

Introduction Local volatility models Stochastic volatility models

Estimation: spot volatility III

• Remember:
Z T
σt2 dt = [log(S)]T .
0

• Thus,
Z T n
X
σt2 dt ≈ ri2 ,
0 i=1

where the right-hand side is the approximate quadratic

variation.
e2 of σ 2 :
• If σt ≡ σ, this then yields an estimator σ
n n
1 X 2 1 X
e2 =
σ ri ≈ (ri − r¯)2 = σ
b2 .
T T
i=1 i=1
Introduction Local volatility models Stochastic volatility models

Estimation: spot volatility IV

• General theory yields (remember, we did not allow for jumps)
that
n
X Z T
ri2 → σt2 dt
P

i=1 0

as ∆t ↓ 0.
RT
• Thus, we can estimate 0 σt2 dt consistently with
high-frequency data.
• However, keep in mind that model assumptions (diffusion) do
not describe well super-high frequency data (ticker size, ...).
RT
• If we know 0 σt2 dt for all T > 0, then we can determine σt2
Lebesgue-almost everywhere. (More cannot be expected.)
However the integrated version seems to be the more natural
quantity in any case.
Introduction Local volatility models Stochastic volatility models

Estimation: spot volatility V

• Consider again the case σt ≡ σ.

• One then needs to decide on a choice of T , often 30 or 180
days when using daily observations.
• Development of more sophisticated methods, e.g.
(exponentially weighted) moving averages (RiskMetrics)
• See notes for an example with Dow Jones data.
Introduction Local volatility models Stochastic volatility models

Realized versus implied volatility

• Realized volatility estimate: based on historical data (past
observations).
• Implied volatility: based on current market prices.
• Observe that BS-pricing formula (for calls and puts) implies,
as a function of σ, an inverse function. For each price (in its
range), there exists a unique σ, which, when put into the BS
formula, yields that price.
• Given a market price CtMKT (T , K ), the implied volatility
Σt (T , K ) is the unique volatility, that solves

CtBS (Σt (T , K )2 , St , T , K ) = CtMKT (T , K ).

• The function Σt is called volatility surface.

• If BS model was correct, Σt (·, ·) would be constant.
Introduction Local volatility models Stochastic volatility models

Implied volatility of SP500 index options

Introduction Local volatility models Stochastic volatility models

Black-Scholes and implied volatility

• BS assumption of constant implied volatility clearly does not

hold in markets where calls and puts are liquidly traded
(otherwise, implied volatilities cannot be observed).
• Graphs illustrate that implied volatilities today change as
maturity and strike changes.
• Moreover, implied volatilities for fixed maturities and strikes
also change over time; that is, Σt (T , K ) as a function of t is
not constant.
Introduction Local volatility models Stochastic volatility models

VIX
Introduction Local volatility models Stochastic volatility models

VIX II
Introduction Local volatility models Stochastic volatility models

Modeling of implied volatilities

• A word of warning: In this course, we shall not model implied

volatilities;
• instead, we shall model the process σt .
• Implied volatilities correspond, in some sense, to averages of
future realizations of paths of σt .
• The direct modeling of implied volatilities is highly complex;
in particular to check whether these models satisfy standard
no-arbitrage conditions, e.g.
• call prices are convex functions in strike K ,
• call prices are increasing functions in maturity T .
Introduction Local volatility models Stochastic volatility models

Stylized facts

• Volatility clustering and persistence: small price moves follow

small moves, large moves follow large moves (high
autocorrelation of volatility measures).
• Thick tails: distribution of asset returns have heavier tails
than normal distribution (leptokurtic distribution).
• Negative correlation between prices and volatility: when prices
go down, volatility tends to rise (leverage effect).
• Mean reversion: volatility tends to revert to some long-run
level.
BS model does not capture these stylized facts.
Introduction Local volatility models Stochastic volatility models

Autocorrelation of squared returns

From Lai and Xing, Statistical Models and Methods for Financial
Markets (Springer).
Introduction Local volatility models Stochastic volatility models

Various volatility models

• Pricing and hedging models. In increasing generality:

• deterministic models: σt = σ(t) only function of t,
• local volatility models: σt = σ(t, St ) function of t and St ,
• stochastic volatility models: additional stochastic factors,
e.g. SABR, Heston.
• Econometric models (mainly for estimation / forecasting):
ARCH, GARCH, EGARCH, IGARCH, ARMA-EGARCH, ...
Introduction Local volatility models Stochastic volatility models

Econometric models
• Usually formulated in discrete time, suitable for statistical
estimation (“time series” analysis).
• Motivated by an attempt to model volatility clustering.
• Heteroskedastic = variance / volatility can change.
• Autoregressive Conditional Heteroskedastic (ARCH) models:
With Pi denoting the log-price at times i∆t,
Pi = Pi−1 + α + ηi εi ,
where α represents the trend, {εi }i∈N is a family of
i.i.d. standard normally distributed random variables and
k
X
ηi2 = β0 + βj (Pi−j − Pi−j−1 − α)2 .
j=1

Thus, large observed recent price changes increase volatility of

next price change.
Introduction Local volatility models Stochastic volatility models

Econometric models II

• Generalized ARCH (GARCH) models depend also on past

values of ηi .
• Usually a GARCH model needs less parameters. (ARCH
models often need large k to get good fit.)
• Analytic expressions for maximum-likelihood estimators or
forecasted volatilities are available.
• In EGARCH models, one distinguishes between positive and
negative returns.
• For details, see for example books by Lai and Xing, Statistical
Models and Methods for Financial Markets (Springer) or Tsay,
Analysis of Financial Time Series (Wiley).
Introduction Local volatility models Stochastic volatility models

Outline from now on

Pricing and hedging models:

1. deterministic models: σt = σ(t) only function of t
2. local volatility models: σt = σ(t, St ) function of t and St
3. stochastic volatility models: additional stochastic factors
Introduction Local volatility models Stochastic volatility models

Deterministic volatility models

• Simplest generalization of BS:

dSt = µSt dt + σ(t)St dBt ,

where σ(·) is a deterministic function of t.

• Pricing and hedging of contingent claims is basically the same
as in BS.
• Market is complete (all contingent claims can be replicated by
trading in the underlying).
• Contingent claim price v (t, St ) corresponding to terminal
payoff h(ST ) usually satisfies BS PDE (assume no dividends)

1
vt (t, s) + rsvs (t, s) + σ(t)2 s 2 vs,s (t, s) − rv (t, s) = 0,
2
v (T , s) = h(s).
Introduction Local volatility models Stochastic volatility models

Deterministic volatility models II

• By the Feynman-Kac theorem,
h i
v (t, s) = E Q e −r (T −t) h(ST )|St = s ,
where S has Q-dynamics
dSt = rSt dt + σ(t)St dBtQ .
• Now, solve SDE to obtain
1 T
Z Z T
2
log(ST ) = log(St ) + r (T − t) − σ(u) du + σ(u)dBuQ
2 t t
to observe that distribution of ST , given St = s, is normal:
   
1 2 2
log(ST ) ∼ N log(s) + r − σ̄t (T − t), σ̄t (T − t)
2
where
Z T
1
σ̄t2 = σ(u)2 du.
T −t t
Introduction Local volatility models Stochastic volatility models

Deterministic volatility models III

• Thus, in all BS pricing formulas for European,

path-independent contingent claims, just replace σ by σ̄t .
• E.g., price of a call option at time t if St = s is given by

CtBS (σ̄t2 , s, T , K ) = sN (d1 ) − e −r (T −t) K N (d2 ),

where
log(s/K ) + (r + σ̄t2 )(T − t)
d1 = √ ,
σ̄t T − t
√
d2 = d1 − σ̄t T − t,
Z T
1
σ̄t2 = σ(u)2 du.
T −t t
Introduction Local volatility models Stochastic volatility models

Definition: local volatility model

• Further generalization of BS:

dSt = µSt dt + σ(t, St )St dBt .

• The deterministic function

(t, s) → σ(t, s)

is called local volatility.

• Model is complete.
• Option price v (t, St ) for terminal payoff h(ST ) usually
satisfies BS PDE
1
vt (t, s) + rsvs (t, s) + σ(t, s)2 s 2 vs,s (t, s) − rv (t, s) = 0,
2
v (T , s) = h(s).
Introduction Local volatility models Stochastic volatility models

Example: Constant Elasticity of Variance (CEV) model

• Important example:

σ(t, s) = δs β

for δ > 0 and usually β ≤ 0.

• Case β > 0 needs care (strict local martingality).
• β = −1/2: Cox-Ingersoll-Ross / square-root process
(well-known from modeling interest rates).
• β < 0 yields leverage effect: spot volatility increases as asset
price declines.
• Generally, be cautious concerning possibly positive probability
of hitting zero.
• Process is analytically tractable (including analytic formulas
for barrier and lookback options).
• Can be extended (by making δ stochastic) to SABR model.
Introduction Local volatility models Stochastic volatility models

Another example: Quadratic Normal Volatility model

• Risk-neutral dynamics described by

dSt = (aSt2 + bSt + c)dBt .

• Here,
c
σ(t, s) = as + b + .
s
• Used to price Foreign Exchange options.
• Strict local martingality is again an issue,
• but process is analytically tractable.
Introduction Local volatility models Stochastic volatility models

Calibration

• We shall assume that there is a liquid market for “vanilla”

calls / puts.
• Then, these contingent claims are marked to market,
• and could be used as hedging instruments to price exotic
contingent claims.
• E.g., vanillas written on major indices (SP500, SP100, DJ,
DAX, FTSE, ...), large stocks, currencies.
• Throughout this section, we shall try to find a local volatility
function σ, such that the model prices (expectations under
the risk-neutral measure) agree with the observed market
prices (calibration).
• In other words, we are trying to choose a certain distribution
among a class of distributions.
Introduction Local volatility models Stochastic volatility models

Digression: Breeden-Litzenberger formula

• Assume that S is a Markov process with a density p(t, s, T , ·)
for ST , conditioned on St = s.
• Then,
Z ∞
−r (T −t)
Ct (s, T , K ) = e p(t, s, T , y )(y − K )+ dy .
0

• Thus,
Z ∞
∂
Ct (s, T , K ) = −e −r (T −t) p(t, s, T , y )dy ,
∂K K
∂2
Ct (s, T , K ) = e −r (T −t) p(t, s, T , K ),
∂K 2
• implying that observing all call prices yields the (marginal)
density of ST .
Introduction Local volatility models Stochastic volatility models

Digression: Breeden-Litzenberger formula II

Several practical issues:

• Not all call prices can be observed (as the corresponding calls
are not traded).
• Call prices are only observed with “some error”
(bid-ask-spread).
• Differentiation is numerically highly unstable.
• Interpolating observed call prices with a smooth function is
very difficult (interpolation scheme might not be
arbitrage-free [call prices monotone and convex in K ]).
Introduction Local volatility models Stochastic volatility models

Kolmogorov equations
• Assume from now on that S is given by

dSt = rSt dt + σ(t, St )St dBt ,

in particular, S is Markovian, and (assume that it) has a

density p(t, s, T , ·) for ST , conditioned on St = s.
• Then p(t, s, T , y ) satisfies “BS PDE”

1
−pt (t, s, T , y ) = rsps (t, s, T , y ) + σ(t, s)2 s 2 ps,s (t, s, T , y ),
2
p(T , s, T , y ) = δ(s − y )

where δ(·) is Dirac delta function, satisfying

Z ∞
h(y )δ(x − y )dy = h(x).
−∞

• This PDE is also called Kolmogorov backward equation.

Introduction Local volatility models Stochastic volatility models

Kolmogorov equations II

• Observed that T , y are constant in Kolmogorov backward

equation.
• By a simple argument based on integrating by parts (see
lecture notes), one can derive the following PDE for
p(t, s, T , y ):
∂ 1 ∂2  2 2

pT (t, s, T , y ) = −r (yp(t, s, T , y )) + σ(T , y ) y p(t, s, T , y ) ,
∂y 2 ∂y 2
p(t, s, t, y ) = δ(s − y )

• This PDE is called Kolmogorov forward equation or

Fokker-Planck equation.
• Now, t, s are constant.
• This PDE is quite useful as its solution corresponds to the
whole surface p(t, s, ·, ·).
Introduction Local volatility models Stochastic volatility models

Dupire’s formula
• Multiply forward Fokker-Planck equation by (y − K )+ and
integrate over y to obtain
Z ∞
∂
p(t, s, T , y )(y − K )dy =
∂T K
∂2 
Z ∞ Z ∞
∂ 1 2 2

− r (yp(t, s, T , y ))(y − K )dy + σ(T , y ) y p(t, s, T , y ) (y − K )dy .
K ∂y 2 K ∂y 2

• Integrating by parts (under sufficient regularity assumptions),

multiplying by e −r (T −t) , and applying the
Breeden-Litzenberger formula yield
Z ∞ Z ∞
∂ 1

−r (T −t) −r (T −t) 2 2
e p(t, s, T , y )(y − K )dy = e ryp(t, s, T , y )dy + σ(T , K ) K p(t, s, T , K )
∂T K K 2
∂ 1 2 2 ∂2
= rCt (s, T , K ) − rK Ct (s, T , K ) + σ(T , K ) K Ct (s, T , K ).
∂K 2 ∂K 2

• We obtain Dupire’s formula

∂ ∂
Ct (s, T , K ) + rK ∂K Ct (s, T , K )
σ(T , K )2 = 2 ∂T ∂ 2 .
K 2 ∂K 2 Ct (s, T , K )
Introduction Local volatility models Stochastic volatility models

Dupire’s formula II

• Do not confuse Dupire’s equation

∂ ∂ 1 ∂2
C + ry C − σ(T , y )2 y 2 2 C = 0
∂T ∂y 2 ∂y
with the BS equation

∂ ∂ 1 ∂2
C + rs C + σ(t, s)2 s 2 2 C − rC = 0.
∂t ∂s 2 ∂s
• Main advantage of Dupire’s equation is that it treats call price
as a function of strike and maturity.
• Dupire’s formula can be used to calibrate a local volatility
model to call prices.
Introduction Local volatility models Stochastic volatility models

Dupire’s formula III

Uniqueness: Given a continuum of arbitrage-free call prices there

exists at most one local vol surface which calibrates them:
v
u ∂ ∂
u Ct (s, T , K ) + rK ∂K Ct (s, T , K )
σ(T , K ) = t2 ∂T ∂ 2 .
K 2 ∂K 2 Ct (s, T , K )

Existence: Not obvious as the SDE

dSt = rSt dt + σ(t, St )St dBt

needs to have a solution.

Introduction Local volatility models Stochastic volatility models

Dupire’s formula IV

• Neither Dupire nor Derman-Kani (who developed a

discrete-time version) thought of local volatility as a realistic
model for the evolution of actual volatility.
• Local volatility can be interpreted as a “code-book”, a
translation of a call price surface, due to the one-to-one
mapping of a (arbitrage-free) call price surface and a local
volatility surface.
• Why do we need more complicated models? Local volatility
models basically capture all marginal distributions.
Introduction Local volatility models Stochastic volatility models

Dupire’s formula and implied volatility

• Remember: Implied volatility Σt (T , K ) defined via

CtBS (Σt (T , K )2 , s, T , K ) = Ct (s, T , K ).

• Reparameterize (dimensionless variables):

w (T , x) = Σt (T , se r (T −t) e x )2 (T − t).

• Then,
wT
σ(T , se r (T −t) e x )2 =   .
x 1 x2
1− w wx + 4 − 14 − 1
w + w2
wx2 + 12 wx,x

• For details, see for example Gatheral, The Volatility Surface

(Wiley).
Introduction Local volatility models Stochastic volatility models

Dupire’s formula: drawbacks

• Requires continuum of observed call prices.

• Prices are not exactly observed due to bid-ask spreads.
• Differentiation is numerically unstable.
• Interpolations are difficult as no-arbitrage conditions have to
be guaranteed.
• More generally, local volatility models do not capture the
correct dynamics. (E.g., calibrate local volatility model at
times t0 and t1 and observe that parameters usually change
completely.)
Introduction Local volatility models Stochastic volatility models

Outline for the rest of this section

Calibration of local volatility models

1. Inverse problems
2. Examples
3. Regularization
4. Application to calibration of local volatility models
Introduction Local volatility models Stochastic volatility models

Inverse problems

• A problem is called an inverse problem if it is defined as a

inverse of some other, “more explicitly” stated problem.
• E.g., instead of going from a model (here described through
the local volatility σ(·, ·)) to option prices, going the other
way around, from option prices to model parameters.
• A problem is called well-posed if
1. a solution exists,
2. the solution is unique, and
3. the solution depends “continuously” on the data (input).
• Otherwise, the problem is called ill-posed.
Introduction Local volatility models Stochastic volatility models

Examples: well- and ill-posed problems

Rx
• Integration: given f , find F (x) = 0 f (z)dz. Consider an
erroneous observation f˜ := f + g with supx |g (x)| ≤ . Then,
Z x Z x  Z x
f˜(z)dz −
 

 f (z)dz  ≤ |g (z)|dz ≤ x.
0 0 0

Thus, integration represents a well-posed problem.

∂
• Differentiation: given F , find f (x) = ∂x F (x). Consider an
erroneous observation F̃ := F +  sin(x/2 ). Then
|F̃ (x) − F (x)| ≤ , but

cos x2 
  

∂
 F̃ (x) − ∂ F (x) =

 ∂x →∞
∂x  

as  ↓ 0. Thus, differentiation represents an ill-posed problem.

Introduction Local volatility models Stochastic volatility models

Implied volatility of SP500 index options

From Cont, Encyclopedia of Quantitative Finance (Wiley).

Introduction Local volatility models Stochastic volatility models

Tikhonov regularization

• Consider an abstract ill-posed inverse problem

F (x) = y , x ∈ X , y ∈ Y .

• Solution may not exist, may not be unique, or may be

unstable.
• Remedy: Solve a regularized optimization problem instead:

min kF (x) − y k2Y + αG (x),

x∈X

where G is some (convex) penalty function, e.g.,

G (x) = kx − x0 k2X with x0 a “prior” guess.
• Regularization factor α needs to be determined by
cross-validation.
Introduction Local volatility models Stochastic volatility models

Tikhonov regularization applied to local vol. calibration

• Market prices V mkt (Ti , Ki ) and model prices V σ (Ti , Ki )
(produced in a model with local volatility σ(·, ·))
• V : usually out-of-the money calls / puts.
• Consider, for some weights wi ,
X n  2
min wi V mkt (Ti , Ki ) − V σ (Ti , Ki ) + αkσk2 .
σ∈S
i=1
• For example,
T ∞ 2 2
∂2σ
Z Z 
∂σ
kσk2 = + dK dt
0 0 ∂K 2 ∂T
(“Flatter” volatility surfaces are preferred)
• Choice of weights wi should depend on liquidity of
corresponding options.
• Choice of S: often, spline-based representation (local
volatility is parameterized by finitely many values, thus
optimization problem is finite-dimensional problem).
Introduction Local volatility models Stochastic volatility models

Tikhonov reg. applied to local vol. calibration II

Advantages:
• No need to have continuum of observed strikes and maturities.
• No need to interpolate market prices.
• Convex penalization leads to numerical stability.
• Calibrated surface is smooth due to choice of penalization
norm.
Disadvantages:
• Computationally demanding.
• Penalization criterion does not include weights to take into
account distribution of St ; thus, criterion overweights values
with small probability of occurrence.
Introduction Local volatility models Stochastic volatility models

Problems of local volatility models

• Local volatility models can perfectly fit marginals

(European-style path-independent options),
• but have problems with pricing path-dependent options, and
• their dynamics are not realistic.
Introduction Local volatility models Stochastic volatility models

Formal description: stochastic volatility models

• Allow now σt to be a stochastic process:

dSt = µ(t, St , Yt )St dt + σ(t, St , Yt )St dBt , (2)

dYt = a(t, St , Yt )dt + b(t, St , Yt )dWt , (3)
dhB, W it = ρdt.
p
• We can write Wt = ρBt + 1 − ρ2 Bbt , where Bb is a BM
independent of B.
• If the underlying asset (with price process St ) is the only
hedging instrument, then the market is incomplete since, in
general, contingent claims written on Y cannot be replicated
by trading in S only.
• ρ < 0 corresponds to the leverage effect.
Introduction Local volatility models Stochastic volatility models

Risk-neutral measures

• Denote by P the historical (physical) probability measure

under which S and Y have the dynamics of (2) and (3).
• If we assume no-arbitrage (NFLVR to be precise), then there
exists a measure Q, equivalent to P, such that e −rt St is a
local martingale under Q. (1st Fundamental Theorem of
Asset Pricing)
• However, Q might not need to be unique. Indeed, it is unique
if and only if the model is complete (each contingent claim
can be replicated perfectly by trading in S alone). (2nd
Fundamental Theorem of Asset Pricing)
Introduction Local volatility models Stochastic volatility models

Risk-neutral measures II
• Denote the class of equivalent local martingale measures by
M.
• Any Q ∈ M is characterized by its stochastic discount factor
/ Radon-Nikodym derivative with respect to P:

Q dQ 
Zt := = E(−λ · B − ψ · B)b t,
dP Ft

where E is the Doléans-Dade (stochastic) exponential

1
E(X )t := e Xt −X0 − 2 [X ]t ,

and
µ(t, St , Yt ) − r
λt := ,
σ(t, St , Yt )
and ψ is progressively measurable.
Introduction Local volatility models Stochastic volatility models

Risk-neutral measures III

• As we have seen, λ is fixed by the model (representing the

market price of risk), but ψ is an (almost) arbitrary process.
• For any Q ∈ M, we have E[ZTQ ] = 1.
• Thus, this yields a necessary condition on ψ. For this
condition to hold, Novikov’s condition is sufficient:
h 1 RT 2 2 i
E e 2 0 (λt +ψt )dt < ∞.
Introduction Local volatility models Stochastic volatility models

Risk-neutral measures IV

• Under Q, we obtain that B Q , B

bQ are independent BM with
Z t
BtQ = Bt + λ(u, Su , Yu )du,
Z0 t
BbtQ = Bbt + ψu du.
0

• Furthermore, we obtain the dynamics

dSt = rSt dt + σ(t, St , Yt )St dBtQ ,

  p 
dYt = a(t, St , Yt ) − b(t, St , Yt ) ρλ(t, St , Yt ) + 1 − ρ2 ψt dt
+ b(t, St , Yt )dWtQ ,
dhB Q , W Q it = ρdt.
Introduction Local volatility models Stochastic volatility models

Arbitrage-free prices
• Consider a claim that pays h(ST , YT ) at time T .
• Denote
h i
CtQ = EQ e −r (T −t) h(ST , YT )|Ft

• If (S, Y ) is Markovian under Q, e.g. if ψt = ψ(t, St , Yt ), we

have CtQ = v Q (t, St , Yt ) for
h i
v Q (t, s, y ) = EQ e −r (T −t) h(ST , YT )|St = s, Yt = y .

• Clearly, C Q and v Q depend on the choice of risk-neutral

measure Q.
• All possible, arbitrage-free prices of the claim with payoff
h(ST , YT ), are given by the interval
" #
inf CtQ , sup CtQ .
Q∈M Q∈M
Introduction Local volatility models Stochastic volatility models

Choice of risk-free measure

• If inf Q∈M CtQ = supQ∈M CtQ , then there is a unique price and
the claim can be perfectly replicated, by the general theory.
(e.g., usually [but not always] h(ST , YT ) = ST ).
• Otherwise, distinguish two cases:
1. Only S is a liquidly traded asset, and there are no other
hedging instruments available.
2. There is another hedging instrument available (e.g., a call).
• In case 1, we have a problem. Possible approaches:
• Reconsider, whether the model should be changed.
• Find a hedging strategy that minimizes risk (e.g., VaR),
quadratic hedging, ...
• Take an ad-hoc measure: Minimal Martingale (Entropy)
Measure, ...
• Choose the superreplicating price: supQ∈M CtQ . This often is a
very conservative approach.
Introduction Local volatility models Stochastic volatility models

Portfolio dynamics and lack of replication

• If an investor holds ∆t units of S at each time t and keeps /
borrows the remainder in / from the bank account
(self-financing strategy), then her wealth process
X := {Xt }t∈[0,T ] satisfies
dXt = rXt dt + ∆t St σ(t, St , Yt ) (λ(t, St , Yt )dt + dBt )
= rXt dt + ∆t St σ(t, St , Yt )dBtQ .
• Price of a claim CtQ = v Q (t, St , Yt ) satisfies:
dCtQ = (vtQ + Av Q )dt + (σSt vsQ + ρbvyQ )dBt
p
+ 1 − ρ2 bvyQ dBbt ,
where
1 1
Av = σ 2 s 2 vs,s + b 2 vy ,y + ρσsbvs,y + µsvs + avy .
2 2
• It is simple to see that, usually, perfect hedging with S only is
impossible.
Introduction Local volatility models Stochastic volatility models

Market completion
• Introduce another traded asset, for example, a contingent
claim (that cannot be replicated by trading in S only) with
payoff g (STe , YTe ) at time Te ≥ T . Denote its price at time t
by Ot = u(t, St , Yt ).
• Now,

dOt = (ut + Au)dt + (σSt us + ρbuy )dBt

p
+ 1 − ρ2 buy dBbt ,
p
= (ut + Au − λ(σSt us + ρbuy ) − ψ 1 − ρ2 buy )dt
p
+ (σSt us + ρbuy )dBtQ + 1 − ρ2 buy dBbtQ .
• Thus, there exists exactly one ψ (assuming uy 6= 0) such that
the drift term equals ru under the corresponding Q.
• Thus, with S and O as traded assets, there exists only one
equivalent local martingale measure and thus, the market is
complete.
Introduction Local volatility models Stochastic volatility models

Market completion II

• An investor can hold ∆t units of S, Nt units of O, and the

bank account. Her wealth process X satisfies

dXt = ∆t dSt + Nt dOt + r (Xt − ∆t St − Nt Ot )dt

= (r (Xt − Nt u) + ∆t σλSt + Nt (ut + Au)) dt
p
+ (∆t σSt + Nt (σSt us + ρbuy ))dBt + Nt 1 − ρ2 buy dBbt .

• In order to hedge a contingent claim perfectly, we need a

wealth process with XT = h(ST , YT ).
• Recall the dynamics of C Q :

dCtQ = (vtQ + Av Q )dt + (σSt vsQ + ρbvyQ )dBt

p
+ 1 − ρ2 bvyQ dBbt ,
Introduction Local volatility models Stochastic volatility models

Market completion III

• Equating the dB-terms
b yields

Nt uy = vyQ .

This component of the hedging strategy is called “Vega

hedging”.
• Now, equating the dB-terms yields

vyQ
∆t = vsQ − us .
uy

This component of the hedging strategy is called “Delta

hedging”.
• Thus, the “volatility risk” of the claim is offset by the vega
hedge and the delta hedge gets adjusted by the delta provided
through the vega hedge.
Introduction Local volatility models Stochastic volatility models

Market completion IV

• Equating now the dt-terms yields

ut + Au − σλSt us − ru v Q + Av Q − σλSt vsQ − rv Q

= t .
uy vyQ

• Observe that the left-hand side equals

p
(ρλ + 1 − ρ2 ψ)b.

• This comes from setting the drift equal to ru in

p
dOt = (ut + Au − λ(σSt us + ρbuy ) − ψ 1 − ρ2 buy )dt
p
+ (σSt us + ρbuy )dBtQ + 1 − ρ2 buy dBbtQ .
Introduction Local volatility models Stochastic volatility models

Market completion V
• Thus, both u and v Q satisfy the PDE

ft + AQ f − rf = 0,
where
1 2 2 1 2
  q 
Q
A f = σ s fs,s + b fy ,y + ρσsbfs,y + rsfs + a−b ρλ + 1 − ρ2 ψ fy .
2 2

• Observe that AQ is the generator of

dSt = rSt dt + σ(t, St , Yt )St dBtQ ,

  p 
dYt = a(t, St , Yt ) − b(t, St , Yt ) ρλ(t, St , Yt ) + 1 − ρ2 ψt dt
+ b(t, St , Yt )dWtQ ,
dhB Q , W Q it = ρdt.
• Thus, the Feynman-Kac formula yields
h i
v (t, s, y ) = EQ e −r (T −t) h(ST , YT )|St = s, Yt = y .
Introduction Local volatility models Stochastic volatility models

Special case: zero correlation

• Assume for the moment that ρ = 0.
• Assume moreover that
dSt = rSt dt + σ(t, Yt )St dBtQ ,
dYt = (a(t, Yt ) − b(t, Yt )ψt ) dt + b(t, Yt )dWtQ ,
dhB Q , W Q it = 0 × dt.
• A call with payoff h(ST ) = (ST − K )+ has a price
h i
v (t, s, y ) = EQ e −r (T −t) h(ST )|St = s, Yt = y
h i
= EQ EQ [e −r (T −t) h(ST )|Ft , {Yt }t≥0 ]|St = s, Yt = y
h i
= EQ CtBS (σ̄t2 , s, T , K )|Yt = y ,
where
Z T
1
σ̄t2 := σ 2 (s, Ys )ds.
T −t t
Introduction Local volatility models Stochastic volatility models

Examples of stochastic volatility models

√
• Hull-White model: ρ = 0, µ(·, ·, ·) = µ, σ(·, ·, y ) = y and Y
GBM:
p
dSt = µSt dt + Yt St dBt ,
dYt = aYt dt + bYt dWt ,
dhB, W it = 0 × dt.
√
• Heston model: µ(·, ·, ·) = µ, σ(·, ·, y ) = y and Y a
square-root / Cox-Ingersoll-Ross process (mean-reverting!):
p
dSt = µSt dt + Yt St dBt ,
p
dYt = a(b − Yt )dt + b Yt dWt ,
dhB, W it = ρdt.

• Standard models are mainly chosen due to their analytic

tractability.
Introduction Local volatility models Stochastic volatility models

Another class of stochastic volatility models

• Sometimes, stochastic volatility models are specified through
R t
a time-change Vt = 0 vs ds.
• The time change Vt has an interpretation as business time (in
contrast to versus calendar time).
• E.g.,

St = e rt+BVt .

• If B and V are independent, this is just another

representation of the stochastic volatility model above, due to
d √
the self-similarity of BM (Bct = cBt ).
• Advantage: Interpretation as business time.
• Disadvantage: It is not clear, a priori, how to model the
leverage effect, for example.
Introduction Local volatility models Stochastic volatility models

Hedging: stochastic volatility model vs. BS

• In theory, stochastic volatility models can help in hedging the

“volatility risk”, which is an improvement in comparison with
the BS model.
• However, hedging ratios depend strongly upon the
parameters, and are sensitive with respect to changes in
parameters (recalibration).
• This is another example of an ill-posed problem.
• Often, a simple model does better since its parameters can be
calibrated more efficiently, are more robust, and hedging errors
can get “averaged out”.