Chapter 2

Local Volatility Models

2.1 Local volatility models and Dupires equation
The local volatility models are based on the following prescription of asset price dynamic
dS
t
S
t
=
t
dt +(S
t
, t; S
0
)dW
t
,
where (S
t
, t; S
0
) is a deterministic function of S
t
and t, and S
0
is a xed parameter. Under
the risk-neutral measure,
t
= r
t
q
t
. Hence, the model is specied once (S
t
, t; S
0
) is
specied.
Dupire (1994) showed that, given the call price C(S
0
, 0, K, T) as a bivariate function of
K and T , the local volatility (S
t
, t; S
0
) can be uniquely determined. The derivation will
need the forward Fokker-Plank equation for the transition density function.
Let (S
t
, t; Y, T) be the transitional density function of S
t
that is dened by
(S
t
, t; Y, T) = E
Q
t
[(S
T
, T; Y, T)]
= E
Q
t
[(S
T
Y )]
where Q stands for the risk-neutral measure. It is well know that (S
t
, t; Y, T) satises the
backward Fokker-Planck equation:
_
_
_
_

t
+L
t,S
_
(S
t
, t; Y, T) = 0,
(S
T
, T; Y, T) = (S
T
Y ),
and
L
t,S
=
1
2

2
t
S
2

2
S
2
+
t
S

S
.
Next, we will show that also satises the forward Fokker-Planck equation.
1
2 CHAPTER 2. LOCAL VOLATILITY MODELS
Proposition 2.1.1 The probability density function (S
t
, t; Y, T) satises the following equation
_

T
L

T,Y
_
(S
t
, t; Y, T) = 0, T > t,
where
L

T,Y
(S
t
, t; Y, T) =
1
2

2
Y
2
_

2
T
Y
2

Y
(
T
Y ) ,
and
(S
t
, t; Y, T) = (Y S).
Proof: Consider an intermediate time such that t < < T. Then
(S, t; Y, T) =
_
R
(S, t; Z, )(Z, ; Y, T)dZ
The above equation expresses that the diffusion process starts from S and ends at Y must
be at some point Z at time . Differentiating both sides of the last equality with respect to
and use integration by parts repeatedly, we obtain
0 =
_
R
(S, t; Z, )

(Z, ; Y, T)dZ +
_
R
(S, t; Z, )
(Z, ; Y, T)

dZ
=
_
R
(S, t; Z, )

(Z, ; Y, T)dZ
_
R
(S, t; Z, )L
,Z
(Z, ; Y, T)dZ
=
_
R
(S, t; Z, )

(Z, ; Y, T)dZ
_
R
_
L

,Z
(S, t; Z, )
_
(Z, ; Y, T)dZ
Now let T, we then have
0 =
_
R
__

T
L

T,Z
_
(S, t; Z, T)
_
(Z Y )dZ
It follows that
_

T
L

T,Y
_
(S, t; Y, T) = 0.
The proposition is thus proved
Now we are ready to derive the Dupires equation. The value of call option satises
C(S
0
, K, T) = e
rT
_

K
(S
0
, 0; S
T
, T)(S
T
K)dS
T
, (2.1.1)
where (S
0
, 0; S
T
, T) is the risk-neutral density function of S
T
. Differentiating (2.1.1) with
respect to K once and twice, we obtain (Breeden and Litzenberger, 1978)
C
K
= e
rT
_

K
(S
0
, 0; S
T
, T)dS
T
,

2
C
K
2
= e
rT
(S
0
, 0; K, T).
(2.1.2)
2.1. LOCAL VOLATILITY MODELS AND DUPIRES EQUATION 3
We have already known that satises the forward Fokker-Plank equation:

T
=
1
2

2
S
2
T
_

2
T
S
2
T

S
T
(
T
S
T
) ,
Differentiating (2.1.1) with respect to T yields
C(S
0
, K, T)
T
= e
rT
_

K
dS
T
_
(S
0
, 0; S
T
, T)
T
_
(S
T
K) rC
= e
rT
_

K
dS
T
_
1
2

2
S
2
T
_

2
T
S
2
T

S
T
(
T
S
T
)
_
(S
T
K) rC
= e
rT
_
1
2

S
T
_

2
T
S
2
T

S
T
(
T
S
T
)
_
(S
T
K)

+
S
T
=K
e
rT
_

K
dS
T
_
1
2

S
T
_

2
T
S
2
T

_
(
T
S
T
)
_
rC
= e
rT
1
2
_

2
T
S
2
T

+
S
T
=K
+e
rT
_

K
dS
T
(
T
S
T
) rC
= e
rT
1
2

2
T
K
2
(S
0
, 0; K, T) +
T
e
rT
_

K
dS
T
S
T
rC
=
1
2

2
K
2

2
C
K
2
+e
rT

T
_

K
dS
T
(S
T
K +K) rC
=
1
2

2
K
2

2
C
K
2
+
T
_
C +e
rT
K
_

K
dS
T

_
rC
=
1
2

2
K
2

2
C
K
2
+
T
_
C K
C
K
_
rC
(2.1.3)
Here we have assumed
T
to be independent of S
T
, which is non-restrictive at all. Thus,
we arrive at the Dupires equation
C
T
=
1
2

2
T
K
2

2
C
K
2
+
T
_
C K
C
K
_
rC. (2.1.4)
The major merit of the Dupires equation is that it allows us to express the local volatility
function in terms of the values of options and their differentiations, that

2
(K, T, S
0
) =
C
T

T
_
C K
C
K
_
+rC
1
2
K
2

2
C
K
2
. (2.1.5)
Note that the above result extends to time-dependent interest rate, such that we only need
to replace r by r
T
in the above equation.
4 CHAPTER 2. LOCAL VOLATILITY MODELS
2.2 Local Volatility in Terms of Implied Volatility
The markets quote options in terms of Black-Scholes implied volatility,
BS
(K, T, S
0
).
Thus, we can write
C(S
0
, K, T) = C
BS
(S
0
, K, T,
BS
(K, T, S
0
)).
For convenience, we write
w(K, T, S
0
) =
BS
(K, T, S
0
)T.
and the log-moneyness by
y = ln
F
T
K
,
where F
T
= S
0
exp{
_
T
0

t
dt} is the forward price. In terms of these variables, we can write
C
BS
(F
T
, y, w) = F
T
{(d
1
) e
y
(d
2
)}
= F
T
_
(
y

w
+

w
2
) e
y
(
y

w
2
)
_
and (2.1.4), the Dupires equation, into
C
T
=
v
L
2
_

2
C
y
2

C
y
_
+
T
C.
where v
L
stands for the local volatility,
v
L
=
2
(S
0
, K, T).
After some derivations, we will arrive at
v
L
=
w
T
1
y
w
w
y
+
1
4
_

1
4

1
w
+
y
2
w
2
__
w
y
_
2
+
1
2

2
w
y
2
.
(2.2.6)
When there is no skew,
w
y
= 0, we have
v
L
=
w
T
,
and the implied Black-Scholes variance is obtained from the local variance, through
w(T) =
_
T
0
v
L
(t)dt.
2.3. CALIBRATION OF THE LOCAL VOLATILITY MODELS 5
2.3 Calibration of the Local Volatility Models
Although elegant, the Dupires result has very limited application in reality. The reasons
are
1. There are only a small set of with which we have reliable quotes.
2. Numerical differentiation is an unstable procedure. To do 2nd -order differentiation
is even more unreliable.
There is, however, a numerical procedure to determine the function using
1. bivariate spline interpolation, and
2. calibration of the model.
This result was developed by Coleman, Li and Verma (1997). They rst interpolate the lo-
cal volatility function by bivariate cubic spline, and then determine the spline function by
matching to the mid prices of the input call or put options. Conventional nite difference
method was adopted to solve for the option prices. From the numerical point of view, this
procedure is feasible. Figure 2.1 shows a recovered local volatility surface.
Figure 2.1 Original and the reconstructed local volatility surfaces
When taking this approach, we should avoid using a big number of spline knots. In
fact, when the number of spline knots is much greater than the number of observations,
the calibration problem becomes severely underdetermined and the solution can be very
different from the true one and unreliable, affecting the quality of the Greeks so calcu-
lated. Figure 2.2 gives one of such examples, where the reconstructed local volatility curve
6 CHAPTER 2. LOCAL VOLATILITY MODELS
Figure 2.2 Original and the reconstructed local volatility curves
2.3. CALIBRATION OF THE LOCAL VOLATILITY MODELS 7
Figure 2.3 True and the reconstructed price and greeks
8 CHAPTER 2. LOCAL VOLATILITY MODELS
(for a single maturity) is signicantly different from the original one, and is sensitive to
the use of spline knots. Although the prices are correctly calibrated, the Greeks calculated
using the calibrated local-vol surface can be very different from the one calculated using
the true local-vol surface. In application the true local volatility surface may not exist,
but the Greeks so obtained denitely carry the so-called model risk.
2.4 Local Variance as a Conditional Expectation of Instan-
taneous Variance
Consider the general model of stochastic volatility of the form
_
dS
t
= S
t
(
t
dt +
t
dW
t
),
d
t
= (S
t
,
t
, t)dt +(S
t
,
t
, t)dZ
t
,
(2.4.7)
with
dW
t
dZ
t
=
t
dt.
In terms of the forward price,
F
T
t
= S
t
e

t
0
(rsqs)ds
,
the asset price process becomes
dF
T
t
= F
T
t

t
dW
t
.
Since F
T
T
= S
T
, the T-forward price of the European call option is
C(F
T
0
, K, T) = E
_
(F
T
T
K)
+

.
Differentiating the above equation with respect to K we have
C
K
= E
_
H(F
T
T
K)

2
C
K
2
= E
_
(F
T
T
K)

,
where H and are Heaviside function and Dirac delta function, respectively. Applying
the Itos lemma to the terminal payoff and then letting t T we have
d(F
T
T
K)
+
= H(F
T
T
K)dF
T
T
+
1
2

2
T
(F
T
T
)
2
(F
T
T
K)dT.
Take expectation on both sides conditional on F
T
0
, we then have
dC = dE
_
(F
T
T
K)
+

=
1
2
E
_

2
T
(F
T
T
)
2
(F
T
T
K)

dT.
2.5. PROS AND CONS OF THE LOCAL VOLATILITY MODELS 9
where we have used the martingale property of F
T
t
. Next, we write
E
_

2
T
(F
T
T
)(F
T
T
K)

= E
_

2
T
(F
T
T
)|F
T
T
= K

E
_
(F
T
T
K)

= K
2
E
_

2
T
|F
T
T
= K

2
C
K
2
.
Here, the second equality above is a slight generalization of the formula for conditional
expectation:
E[X|A] =
E[X 1
A
]
E[1
A
]
,
or
E[X 1
A
] = E[X|A]E[1
A
].
Since
dC =
C
T
dT,
it follows that
C
T
=
1
2
K
2
E
_

2
T
|F
T
T
= K

2
C
K
2
.
By comparison, we conclude that the corresponding local volatility function of a local-
volatility model is

2
loc
(K, T, S
0
) = E
_

2
T
|F
T
T
= K

= E
_

2
T
|F
T
0
e

T
0

1
2

2
t
dt+tdWt
= K
_
That is, local variance is the risk-neutral expectation of the instantaneous variance condi-
tional on the nal stock price S
T
= F
T
T
being equal to strike K. This result can be regarded
as the projection of the stochastic volatility model to local volatility model. No one, how-
ever, has ever obtained the local volatility variance through this result.
2.5 Pros and Cons of the Local Volatility Models
As the rst generation of smile models, the local volatility models are appreciated by its
intuitive extension to the Black-Scholes model. Under the LV models, the market remains
complete, and consistent pricing of exotic derivatives is achieved using the calibrated
models. Soon, however, the LV models are found to suffer serious drawbacks, includ-
ing
1. They tend to produce at future short-term skews.
2. It requires frequent calibrations.
10 CHAPTER 2. LOCAL VOLATILITY MODELS
3. Implied volatility curves moves in the opposite direction to that of the underlying
security.
Among the three drawbacks listed above, the third one is the most serious one, which can
cause hedging errors bigger that that of the Black-Scholes model, and thus almost rejects
the approach of smile modeling.