QUANTI TATI VE FI NANCE VOLUME 2 (2002) 3144 RESEARCH PAPER

I NSTI TUTE O F PHYSI CS PUBLI SHI NG quant.iop.org

Deterministic implied volatility
models
P Balland
Merrill Lynch Financial Centre, 2 King Edward Street, London EC1, UK
E-mail: philippe balland@ml.com
Received 20 September 2001
Published 4 February 2002
Online at stacks.iop.org/Quant/2/31
Abstract
In this paper, we characterize two deterministic implied volatility models,
dened by assuming that either the per-delta or the per-strike implied
volatility surface has a deterministic evolution. Practitioners have recently
proposed these two models to describe two regimes of implied
volatility (see Derman (1999 Risk 4 559)). In an arbitrage-free sticky-delta
model, we show that the underlying asset price is the exponential of a process
with independent increments under the unique risk neutral measure and that
any square-integrable claim can be replicated up to a vanishing risk by
trading portfolios of vanilla options. This latter result is similar in nature to
the quasi-completeness result obtained by Bjork et al (1997 Finance
Stochastics 1 14174) for interest rate models driven by Levy processes.
Finally, we show that the only arbitrage-free sticky-strike model is the
standard BlackScholes model.
1. Introduction
In classic extensions of the BlackScholes (1973) model that
accounts for the smile effect, the underlying asset price S is
driven by one or two Brownian motions. These models are
referredas local volatilitymodels as theydiffer fromthe Black
Scholes model by simply allowing the local volatility
t
of the
underlying asset price to be stochastic:
dS
t
/S
t
=
t
dt +
t
dW
t
.
Two types of local volatility models have been proposed: the
so-called deterministic and stochastic local volatility models.
In the deterministic local volatility models, the local volatility
satises
t
= (t, S
t
) as in Dupire (1994) and Derman
and Kani (1994). Deterministic local volatility models
are complete and complex options are replicated using the
underlying S alone! In the stochastic local volatility models,

t
=
2
t
is a stochastic process characterized by its mean-
reversion rate, its volatility and its correlation with S
t
as in
Hull and White (1987) and Heston (1993):
dv
t
= (
t
v
t
) dt +

v
t
dB
t
(Heston)
dv
t
/v
t
=
t
dt + dB
t
(HullWhite)
The processes B and W are Brownian motions with correlation
and are dened under the risk-neutral measure. This risk-
neutral measure is unique and the stochastic local volatility
models are complete if call options are traded instruments.
The parameters
t
, , and are implied by calibration to
the initial smile surface. When inferred from historical data,
these parameters are typically much lower. The reason for
this can be understood if we observe that in stochastic models,
d ln S
t
has a normal conditional distributionwithvariance
2
t
dt
as in the BlackScholes model. Consequently, the calibration
of a stochastic local volatility model to a short-dated smile
curve, which typically has large variations near the at-the-
money strike, results in unrealistically large correlation and
volatility . To compensate for these large parameters, the
calibration to a long-dated smile curve, which is typically
atter, implies a large mean-reversion rate . Deterministic
local volatility models have similar problems despite accurate
calibrations. The local volatility function inferred from the
initial smile surface, has typically unrealistically large slope
and convexity for small maturity while it is almost at for large
maturity.
1469-7688/02/010031+14$30.00 2002 IOP Publishing Ltd PII: S1469-7688(02)32504-1 31
P Balland QUANTI TATI VE FI NANCE
It is important to understand that a complex option will
be hedged efciently using a smile model only if the model
implies a dynamic of the smile that is sufciently realistic and
stationary for the model to not require frequent re-calibration.
Local volatility models are, in this sense, unsatisfactory. The
dynamic of the implied volatility surface is entirely specied
once the model has been calibrated with the initial smile
surface, leaving thus no control on this dynamic.
To illustrate further this point, consider the foreign-
exchange market, where implied volatilities are quoted per
delta. The underlying S is typically traded more frequently
than vanilla options and thus, the implied volatilities do
not change as frequently as S does. In a deterministic
local volatility model such as the Dupire model, the implied
volatilities are however functions of S. Hence, over a short
time interval during which implied volatilities do not change,
the model will still need re-calibration every time S changes!
These frequent re-calibrations cause the hedging strategies
implied by the model to be inefcient as reported by Dumas
et al (1998). In a stochastic local volatility model, implied
volatilities depend on t , the local volatility
t
and on the
percentage-in-the-money S
t
/K. This last dependence implies
a greater stability of the corresponding hedging strategies.
However, to be consistent with market smiles, such a model
over-estimates, as previously explained, the volatility of
volatility and the correlation of volatility with S and, as a
consequence, the cost of hedging is not accurately accounted
for.
The need for smile models that are calibrated to an
initial smile surface and that imply a realistic evolution of
implied volatility explains the recent interest in the so-called
implied volatility models. These models are dened by direct
assumptions on the stochastic evolution of the smile surface
from an initial surface, as in Schonbucher (1999) and in Cont
and da Fonseca (2001). It is important to stress, however, that
these models need severe restrictions to be arbitrage-free.
In this paper, we characterize two recently proposed
deterministic implied volatility models, dened by assuming
that either the per-delta or the per-strike implied volatility
surface has a deterministic evolution. In arbitrage-free sticky-
delta models, we show that the underlying asset price is the
exponential of a process with independent increments under
the risk neutral measure and that any square-integrable claim
can be replicated up to a vanishing risk by trading portfolios of
vanilla options. Finally, we show that the only arbitrage-free
sticky-strike model is the BlackScholes model.
2. The option market and the implied
volatility models
We consider a continuous trading economy on a nite horizon
[0, T ] with traded assets at time t , a nancial asset S, the money
market account B, the call and the put options on S with strike
K > 0 and maturity in (t, t + x
m
). We also assume that static
portfolios consisting of a continuum of traded vanilla options
are traded instruments as in Breeden and Litzenberger (1978)
and Carr and Madan (1998).
We assume that only short-dated vanilla options are traded
assets because in typical option markets, only short-dated
options have liquid enough prices to be regarded as traded
instruments. In typical foreign exchange markets for example,
only the vanilla options with maturity less than a couple of
years have liquid prices.
We denote by S
t
> 0, C
t
(x, K) > 0 and P
t
(x, K) > 0
the price at time t of one unit of S, of one call and of one put
option on S with maturity t + x (t, t + x
m
) and with strike
K > 0.
We assume no transaction costs. We assume deterministic
interest rates and we denote by P
t x
the price at time t of a
discount bond with maturity t + x and by B
t
1/P
0t
the
money-market account. We denote by F
t x
the forward price at
time t of receiving one unit of S at time t + x t .
We assume that there exists a real bounded function (t )
such that
F
t x
/S
t
= exp
__
t +x
t
(s) ds
_
m
t x
.
We assume that there exists a stochastic basis (, ` =
`
T
, {`
t
: t [0, T ]], P) with a right-continuous, complete,
increasing ltration with respect to which S
t
and 1/S
t
are
cadlag quasi-left continuous square-integrable processes and
sup{E[S
2
t
+ S
2
t
] : 0 t T ] < ,
C
t
(x, K) = P
t x
E
t
[(S
t +x
K)
+
],
P
t
(x, K) = P
t x
E
t
[(K S
t +x
)
+
].
We assume that the map (x, K) . C
t
(x, K) from (0, x
m
)
[0, +) into (0, +) is of class C
1
(resp. C
2
) in the rst
(resp. second) variable for all t (0, T ). The market ltration
{`
t
: t [0, T ]] is the ltration generated by all primary traded
asset prices.
Denition 2.1. A family {C
t
(x, K), P
t
(x, K), S
t
: (x, K)
(0, x
m
) (0, +)} satisfying the above assumptions is
called an implied volatility model.
We use the terminology of implied volatility models
instead of option price models because implied volatilities
rather than option prices are the nancial observable (see
section 1). It is clear, however, that a process S
t
and a
two-parameter family of implied volatility processes dene
an implied volatility model providing that this specication is
arbitrage-free. Our denition of implied volatility models is
based on option prices to ensure our implied volatility models
are arbitrage-free.
We shall need the following technical restriction to ensure
that S has bounded local characteristics. This restriction is
fairly mild and met by most local volatility models.
Denition 2.2. An implied volatility model is regular if
(i) lim
x0
E[f (S
t +x
)[S
t
= S] = f (S), (f C
0
),
(ii) sup
x

x
E[(S
t +x
/F
t x
)
2
[S
t
] is locally bounded,
(iii) sup
t
E[sup
x
[
x
E
t
[(S
t +x
/F
t x
)
2
][] < .
We write C
0
for the space of continuous functions
vanishing at zero and innity. We refer to Derman and Kani
(1998), Schonbucher (1999) and Cont and da Fonseca (2001)
for examples of stochastic implied volatility models. We now
restrict our attention to two deterministic implied volatility
models.
32
QUANTI TATI VE FI NANCE Deterministic implied volatility models
3. The sticky-delta and the sticky-strike
implied volatility models
The BlackScholes function is denoted C
BS
(V, F, K) with
C
BS
(V, F, K) = F N(d) KN(d

V),
d = ln(F/K)/

V +
1
2

V.
We denote

BS
(V, F, K) =
F
C
BS
(V, F, K),

BS
(V, F, K) =
2
F
C
BS
(V, F, K).
We recall that

V
C
BS
(V, F, K) =
1
2
F
2

BS
(V, F, K).
Denition 3.1. Let (x, I) (0, x
m
) (0, +). The
per-delta implied volatility denoted
t
(x, I) where I stands
for the forward moneyness ratio, is the unique positive
solution of the equation
P
t x
C
BS
(x
t
(x, I)
2
, F
t x
, F
t x
I) = C
t
(x, F
t x
I).
The per-strike implied volatility denoted
t
(x, K) is the
unique positive solution of the equation
P
t x
C
BS
(x
t
(x, K)
2
, F
t x
, K) = C
t
(x, K).
Since the function C
BS
(V, F, K) is strictly increasing
with respect to V, we conclude that

t
(x, K) =
t
(x, K/F
t x
),
lim
x0
x
t
(x, I)
2
= lim
x0
x
t
(x, K)
2
= 0.
At a given time t , the volatility functions (x, I) .

t
(x, I) and (x, K) .
t
(x, K) are of class C
1
in the
rst variable and C
2
in the second variable in (0, x
m
)
(0, +). For (x, I) (0, x
m
) (0, +), the implied
volatility processes
t
(x, I) and
t
(x, I) are cadlag quasi-left
continuous processes adapted to the market ltration.
Following Derman (1999) and Reiner (1999), we propose
the following denition of sticky-delta and sticky-strike
implied volatility models.
Denition 3.2. An implied volatility model is sticky-delta if
the per-delta volatility process
t
(x, I) is deterministic on
[0, T ] for all (x, I) in (0, x
m
) (0, +).
An implied volatility model is sticky-strike if the per-strike
volatility process
t
(x, K) is deterministic on [0, T ] for all
(x, K) in (0, x
m
) (0, +).
We observe that the BlackScholes model is an implied
volatility model dened by
t
(x, I) =
t
(x, K) =
t
(x, 1).
4. Characterization of the sticky-delta
implied volatility models
For a sticky-delta implied volatility model, we dene the
following deterministic function:
c(t, x, I) = C
BS
(x
t
(x, I)
2
, 1, I). (4.1)
Since y
2
< 2(e
y
+ e
y
) for all real y and S
t
, 1/S
t
are
square-integrable, ln S
t
is square-integrable. Finally, we have
the following result.
Theorem 4.1. In a sticky-delta implied volatility model
dened on [0, T ], the market ltration {`
t
: 0 t T ]
coincides with the ltration generated by S, all risk-neutral
probability measures are equal to the probability measure P
on the -algebra `
T
and the stochastic process ln S
t
has
independent increments under P.
Proof. We note that P is a risk-neutral measure. Consider a
forward-start call option with xing date t and maturity t + x
with x < x
m
. This option pays at maturity the quantity
(S
t +x
kF
t x
)
+
/F
t x
.
At time t , this forward-start option is a regular call option
on S with strike kF
t x
, time-to-maturity x and notional 1/F
t x
.
At time t , the cost of replicating this option is P
t x
c(t, x, k).
This value is deterministic. Hence at time s t , the forward
value of this option is c(t, x, k) and thus independent of s.
Finally, we have proved that for any risk-neutral measure Q
and any s t , we have
E
Q
s
[(S
t +x
kF
t x
)
+
/F
t x
] = c(t, x, k). (4.2)
Next, we note that
0
1
F
t x

k
(S
t +x
kF
t x
)
+
1. (4.3)
We can thus permute differentiation with respect to k and
Q-expectation in (4.2):
Pr
Q
{S
t +x
< kF
t x
[ `
s
] = 1 +
k
c(t, x, k). (4.4)
The cumulative distribution of ln S
t +x
ln S
t
conditional
on `
s
is thus independent of s t , S
s
and Q.
Let u
1
= ln S
r
ln S
s
and u
2
= ln S
u
ln S
t
with
s < r < t < u < T , 0 < r u < x
m
and u t < x
m
.
Equation (4.4) implies that for any risk-neutral measure Q, we
have
E
Q
[e
iu
1
+iu
2
] = E
Q
[e
iu
1
] E
Q
[e
iu
2
]. (4.5)
The variables u
1
and u
2
are thus independent and have
the same joint distribution under any risk-neutral measure Q.
This result is extended by convolution to n arbitrary non-
overlapping increments of ln S
t
in [0, T ]. It thus follows
that ln S
t
has independent increment in [0, T ] under any risk-
neutral measure.
33
P Balland QUANTI TATI VE FI NANCE
Finally, we show by induction that for any integer n, any
sequence {T
i
] [0, T ]
n
and any real U
i
:
Q{S
T
1
< U
1
, . . . , S
T
n
< U
n
] = P{S
T
1
< U
1
, . . . , S
T
n
< U
n
].
All risk-neutral probability measures Q coincide with P
on the algebra A of cylindrical sets { : S(T
1
, ) <
U
1
, . . . , S(T
n
, ) < U
n
]. Adirect application of the monotone
class theorem as in Jacod and Protter (1991, p 32) implies that
these probabilitymeasures coincide on(A) andthus on`
T
. .
We note that the uniqueness of a risk-neutral measure
does not, however, imply completeness of the model since
the number of traded assets is innite (see Jarrow and Madan
1999). We have proved that in a sticky-delta implied volatility
model, there exists a cadlag square-integrable P-martingale X
with P-independent increments such that
S
t
= S
0
m
0t
exp(X
t
)/E[exp(X
t
)].
Examples of processes with independent increments are
Brownian motions, Poisson processes, Levy processes and
jump-diffusion processes. We refer to Protter (1995) for
the construction of stochastic integrals with respect to cadlag
square-integrable martingales. In this paper, the stochastic
integrals froma 0 to b are integrals on (a, b]. We recall that
the quadratic variationof a cadlagsquare-integrable martingale
Z is an increasing cadlag adapted process dened by
[Z, Z]
t
= Z
2
t
2
_
t
0
Z
u
dZ
u
.
To ease notation, this process will also be denoted [Z]
t
or [Z
t
].
Restricting our attention to regular sticky-delta models
will ensure that X has nite moments and satisfy the
representation property of Nualart and Schoutens. This will
allowus toprove the quasi-completeness of regular sticky-delta
models. But rst, we needsome notationandsome preliminary
results.
Since S is assumed quasi-left continuous, the martingale
X has no xed points of discontinuity and can be decomposed
as in Jacod and Shiryaev (1987, p 77)
X
t
= ( W)
t
+ (x (N n))
t
,

0<st
X
t
= (x N)
t
=
_
t
0
_
xN(dx ds),
where X
t
= X
t
X
t
is the jump of Xat time t , is a square-
integrable deterministic function, W is a Brownian motion,
N is a Poisson measure independent of W with deterministic
compensator n such that n
s
(dx) ds = E[N(dx ds)] where
N(dx ds) is the number of non-zero jumps of X in (s, s +
ds) (x, x + dx).
Since ln S
t
and thus X
t
are square-integrable, we conclude
that for t in [0, T ]
E
_

0<ut
X
2
u
_
=
_
t
0
_
+

x
2
n
u
(dx) du < ,
_
t
0

2
u
du < .
We recall the following notation:
(H W)
t
=
_
t
0
H
u
dW
u
,
( (N n))
t
=
_
t
0
_

u
(x){N(dx du) n
u
(dx) du].
By applying Itos formula (see Jacod and Shiryaev
1987) to S
t
. E[(S
s
K)
+
[S
t
], we derive the following
representation for the short-dated option prices:

C
t,st,K
= C
0sK
+ (H
C
sK
W)
t
+ (
C
sK
(N n))
t
(4.6)

P
t,st,K
= P
0sK
+ (H
P
sK
W)
t
+ (
P
sK
(N n))
t
. (4.7)
where

X
t
= P
0t
X
t
. Jensen inequality implies that the
martingales P
0t
C
t
(s t, K) and P
0t
P
t
(s t, K) are square-
integrable. It follows that H
C
(s, K), H
P
(s, K) are in L
2
W
and

C
(s, K),
P
(s, K) are in L
2
n
with
L
2
W
=
_
H : E
__
T
0
H
2
u
du
_
<
_
,
L
2
n
=
_
B(R) : E
__
T
0
_

2
u
(x) n
u
(dx) du
_
<
_
,
where is the predictable -algebra on [0, T ] , that is the
smallest -algebra making all adapted processes that are left
continuous with right limits, measurable (see Protter 1995).
Let V
u
(t, K) = P
0u
P
u
(t u, K) and recall that the
quadratic variation of this process satises with equation (4.7)
(see Protter 1995)
[V(t, K), V(t, K)]
u
= V
0
(t, K)
2
+
_
u
0
H
P
s
(t, K)
2
ds
+
_
u
0
_

P
s
(t, K)(x)
2
N(dx ds). (4.8)
Let (K) be a square-integrable function. The Lebesgue
Fubini theorem for positive integrand (Malliavin and Airault
1994, p 46) implies that the following Lebesgue integrals, if
nite, satisfy
E
_ _
+
0
_
T
0
(K)
2
d[V
t,K
]
u
dK
_
= E
__
T
0
__
+
0
(
K
H
P
ut K
)
2
dK
_
du
_
+E
_ _
T
0
_ __
+
0
(
K

P
ut K
(x))
2
dK
_
n
u
(dx) du
_
.
Since S
t
and 1/S
t
are assumed to be square-integrable, we
obtain that for all t in [0, T ] and all i 0
E[[ ln S
t
[
i
exp([ ln S
t
[)] < i! E[S
2
t
+ S
2
t
]. (4.9)
It follows that for all i 0:
sup
t [0,T ]
{E[S
t
[ ln S
t
[
i
] + E[S
1
t
[ ln S
t
[
i
]] < . (4.10)
With equation (4.10), we prove the following lemma that
will be useful in establishing our quasi-completeness result.
34
QUANTI TATI VE FI NANCE Deterministic implied volatility models
Lemma 4.1. For any integers p 0 and any real > 0, we
have
E
_ _

0
_
t
0
(ln K)
2p
K
4
d
u
[

P
u,t u,K
]dK
_
< , (4.11)
E
_ _
+

_
t
0
(ln K)
2p
K
4
d
u
[

C
u,t u,K
]dK
_
< , (4.12)
where [M] [M, M],

M M/B for a cadlag process M.
Proof. Since

P
u
(t u, K) is a P-martingale, we observe that:
E
_ _
t
0
d
u
[

P
u
(t u, K)]
_
= E[

P
t
(0+, K)
2
]

P
0
(t, K)
2
.
Therefore, we have
E
__
t
0
(ln K)
2p
K
4
d
u
[

P
u,t u,K
, ]
_
(ln K)
2p
K
4
E[(K S
t
)
+2
].
Using the LebesgueFubini theorem for a positive
integrand, we obtain
_

0
(ln K)
2p
K
4
E[(KS
t
)
+2
]dK E[[ ln S
t
[[ ln [
2p
S
1
t
]
With (4.10), we conclude that
_

0
E
_ _
t
0
(ln K)
2p
K
4
d
u
[

P
u,t u,K
]
_
<
By LebesgueFubini, we obtain (4.11). Similarly, we
derive inequality (4.12). .
Since X has independent increments and E[exp(2X
t
)] +
E[exp(2X
t
)] < , the Laplace transform of X
t
is dened
for all [[ 2, t [0, T ] and satises the time-dependent
LevyKhintchine formula for square-integrable processes:
E[exp((X
t +x
X
t
))] = exp
__
t +x
t

s
() ds
_
,

s
() =
1
2

2
s

2
+
_
+

(e
z
1 z) n
s
(dz).
Using the above formula, we show that the martingale X
for a regular sticky-delta model as in denition 2.2 is regular
in the following sense.
Denition 4.1. The martingale X is regular if there exists
2:
sup
s(0,T )
_

2
s
+
_
(1,1)
c
exp([x[)n
s
(dx)
_
< , (4.13)
Thanks to (4.13), we derive for i 2:
_
+

[x[
i
n
s
(dx) <
i!

i
_
(1,1)
c
exp([x[)n
s
(dx)+
_
1
1
x
2
n
s
(dx)
Therefore for k, i 2, m
i
(t )
_
+

x
i
n
t
(dx) and
M
k
(t ) E
_
0<st
(X
s
)
k
_
=
_
t
0
m
k
(s) ds are uniformly
bounded in [0, T ].
In the case where the increments of X have stationary
distributions, we note that X is a Levy martingale and the
condition (4.13) is similar to the condition introduced by
Nualart and Schoutens (2000). With that restriction, we will
extend in section 6, the polynomial representation obtained by
Nualart and Schoutens for regular Levy martingales, to regular
martingales. With this representation property, we will derive
our quasi-completeness result in section 7. But rst we need
to dene precisely our admissible trading strategies and what
we mean by quasi-completeness, as this type of completeness
is not standard.
5. Trading strategies, attainable claims
and quasi-completeness
In this section, we dene the trading strategies that will be used
to replicate contingent claims in a regular sticky-delta model.
We rst dene the static trading strategies, which are static
portfolios having a continuumof traded vanilla options. These
static trading strategies are traded instruments by assumption
(see section 2). As in the BlackScholes theory, we then
dene tradingstrategies as dynamic portfolios involvinga nite
number of traded instruments.
We highlight that in this paper, static portfolios with a
continuum of traded options, i.e. static trading strategies, are
assumedtobe tradedinstruments, as inCarr andMadan(1998).
We note that this assumption is equivalent to assuming that at
time t , all European claims, with maturity u (t, t + x
m
) and
payoffs f (S
u
), that can be decomposed in a continuum of call
and put option payoffs as in Breeden and Litzenberger (1978),
are traded instruments.
We dene a static trading strategy to be a portfolio of
short-dated options with maturity t
1
and of the money market
account such that the portfolio holdings are constant over the
trading interval (t
0
, t
1
] (t
0
, t
0
+ x
m
) and there are no ows
coming in or out of the strategy up to time t
1
.
A static strategy is characterized by an initial date t
0
, an
end date t
1
(t
0
, t
0
+ x
m
), an initial endowment C
0
`
t
0
,
two locally bounded functions
C
(K) and
P
(K) dened in
[, +) and in (0, ] respectively, and two positive -nite
measures c

(dK) and p

(dK) dened on the positive half line.

The static trading strategy = (t
0
, t
1
,
P
,
C
, p, c, ) is
a portfolio that has constant holdings in (t
0
, t
1
] and that is
instantiated at zero cost using the initial endowment at time t
0
to purchase
C
(K)c

(dK) unit(s) of the call option with strike

K (, +) and maturity t
1
and
P
(K)p

(dK) unit(s) of the

put option with strike K (0, ) and maturity t
1
.
We assume that the functions
C
(K) and
P
(K) satisfy
the following conditions:
E
_ _

0
[
P
K
[

P
0t
1
K
p

(dK) +
_
+

[
C
K
[

C
0t
1
K
c

(dK)
_
< ,
(5.1)
E
_ _

0
_
t
1
t
0

P
(K)
2
d
s
[

P
s,t
1
s,K
]p

(dK)
_
< , (5.2)
E
_ _
+

_
t
1
t
0

C
(K)
2
d
s
[

C
s,t
1
s,K
]c

(dK)
_
< (5.3)
35
P Balland QUANTI TATI VE FI NANCE
where we recall [X] = [X, X]. Using the CauchySchwartz
inequality, we show as in the proof of lemma 4.1, that
equations (5.1)(5.3) are implied by the single equation
E
_ _

S
t
1

(
P
K
K)
2
p

(dK) +
_
S
t
1

(
C
K
S
t
1
)
2
c

(dK)
_
< .
We impose the holding functions to satisfy (5.2) and (5.3)
in order to apply the stochastic Fubini theorem as in Protter
(1995, p 160), so as to permute strike and time integrations.
As we will see, this will allow us to dene dynamic portfolios
based on static trading strategies.
The value of the static trading strategy at time t (t
0
, t
1
]
is obtained by adding the values of its constituents:
V
t
() =
_

0

P
(K)P
t
(t
1
t, K)p

(dK)
+
_
+

C
(K)C
t
(t
1
t, K)c

(dK) + C
0
B
t
/B
t
0
.
The above two integrals are guaranteed to exist thanks
to (5.1). Since the static portfolio does not involve any cost at
initiation, we derive for t in (t
0
, t
1
):

V
t
() =
_

0

P
(K)
__
t
t
0
d
s

P
s
(t
1
s, K)
_
p

(dK)
+
_
+

C
(K)
__
t
t
0
d
s

C
s
(t
1
s, K)
_
c

(dK).
Using (5.2) and (5.3) together with (4.6)(4.9), we show
that we can permute strike and time integration, by application
of the stochastic Fubini theoremas formulatedinProtter (1995)
for martingales and in Lebedev (1995) for random measures:

V
t
()=
___

0

P
K
H
P
t
1
K
p

(dK) +
_
+

C
K
H
C
t
1
K
c

(dK)
_
W
_
t
t
0
+
___

0

P
K

P
t
1
K
p

(dK) +
_
+

C
K

C
t
1
K
c

(dK)
_
(N n)
_
t
t
0
where we have used the notation [X]
b
a
= X
b
X
a
. In short
and with abuse of notations, we simply write

V
t
() =
_
t
t
0
_

0

P
(K)p

(dK) d
s

P(t
1
s, K)
+
_
t
t
0
_
+

C
(K)c

(dK) d
s

C(t
1
s, K).
We set

V
t
() =

V
(t t
0
)t
1
() for t [0, T ]. This
discounted value process is a square-integrable cadlag
martingale which is zero before t
0
and constant after t
1
.
These static strategies which are portfolios with a
continuum of traded options, are traded instruments by
assumption.
A trading strategy is dened as a portfolio of n static
trading strategies and of the money market account. A trading
strategy is thus characterized by a nite sequence {
i
: 1
i n] of static trading strategies and by some predictable
processes {
i
t
: 1 i n] and by a progressively measurable
adapted process
B
t
such that
sup
t (0,T )
E
__
n

i=1

i
t

V(
i
) +
B
t
_
2
_
< , (5.4)
n

i,j=1
E
__
T
0

i
s

j
s
d[

V(
i
),

V(
j
)]
s
_
< . (5.5)
At time t , the strategy is the portfolio consisting of
i
t
unit(s) of each of the static portfolios
i
and of
B
t
unit(s) of
the money market account.
The value of such a strategy is given at time t by the sum
of the values of its constituents:
V
t
() =
n

i=1

i
t
V
t
(
i
) +
B
t
B
t
.
Equation (5.4) guarantees that this process is a square-
integrable semimartingale. The gain cumulated up to time t
by the trading strategy is dened as usual by
G
t
()
n

i=1
_
t
0

i
s
dV
s
(
i
) +
_
t
0

B
s
dB
s
.
Equation (5.5) guarantees the existence of the above
integrals. A trading strategy is self-nancing if and only if
the value process satises as usual
V
t
() = V
0
() + G
t
().
Hence the discounted value process of a self-nancing
trading strategy is a square-integrable martingale that satises

V
t
() = V
0
() +
n

i=1
_
t
0

i
s
d

V
s
(
i
). (5.6)
A self-nancing trading strategy is thus entirely
characterized by = (
i
t
,
i
) where
i
t
is a predictable process
satisfying (5.4) and (5.5). We denote by the space of all self-
nancing trading strategies. We dene the following linear
subspace of L
2
(, `), V() = {V
t
() : t [0, T ], ].
As explained in Jarrow and Madan (1999), an arbitrage
strategy is a self-nancing strategy such that
V
0
() = 0, P(V
T
() 0) = 1, P(V
T
() > 0) > 0.
Proposition 5.1. Implied volatility models are arbitrage-free.
Proof. Suppose that is an arbitrage strategy then
E[1{V
T
() < 0]] = 0 and thus, we have E[V
T
() 1{V
T
() <
0]] = 0. It follows that E[

V
T
()] = E[[

V
T
()[] > 0 =

V
0
().
This contradicts the fact that

V
t
() is a P-martingale. .
Remark 5.1. This result is not surprising since we have
assumed the existence of at least one risk-neutral measure P.
We can now dene attainability and attainability up to a
vanishing risk.
Denition 5.1. A claim L
2
(, `) with maturity t T is
attainable or can be replicated if there is a self-nancing
trading strategy such that = V
t
(), i.e. V().
A claim L
2
(, `) with maturity t T is attainable up
to a vanishing risk or can be replicated up to a vanishing risk
if there is a sequence of self-nancing strategies {
n
] such
that lim
n
E[(V
t
(
n
))
2
] = 0, i.e.

V().
36
QUANTI TATI VE FI NANCE Deterministic implied volatility models
A claim L
2
(, `) with maturity t is thus
attainable if there are some static trading strategies {
i
=
(t
0i
, t
1i
,
P
i
,
C
i
, p
i
, c
i
,
i
) : 0 i n], some predictable
processes {
i
t
: 0 i n] satisfying (5.4) and (5.5) and a real
V
0
= E[P
0t
] which is the cost of replication, such that
P
0t
= V
0
+
n

i=0
_
t
1i
t
0i
_

i
0

i
s
g
i
(K)p
i
(dK) d
s

P
s
(t
1i
s, K)
+
n

i=0
_
t
1i
t
0i
_
+

i
s
f
i
(K)c
i
(dK) d
s

C
s
(t
1i
s, K).
We say that a claim L
2
(, `) with maturity t is
attainable at time s if there exists a self-nancing strategy
and C
s
`
s
such that P
0t
=

V
t
()

V
s
() + C
s
. Similarly,
we dene attainable up to a vanishing risk, at time s.
With the above notations, we dene quasi-completeness
as in Jarrow and Madan (1999) and in Bjork et al (1997).
Denition 5.2. An implied volatility model is complete up to
a vanishing risk or quasi-complete if any L
2
(, `) is
attainable up to a vanishing risk i.e. L
2
(, `)

V().
We now prove the following classic result to be used later.
Proposition 5.2. Let a (0, x
m
), b = t + a [0, T ] and let
f be a C
2
(0, +) function with f (S
b
) L
2
(, `) such that
there is 0 < < satisfying
E
__

S
b

f
//
(K)
2
K
2
dK +
_
S
b

f
//
(K)
2
S
2
b
dK
_
< .
(5.7)
Then the claim with payoff f (S
b
) at time b can be replicated
at time t at a cost P
t b
E
t
[f (S
b
)].
Proof. Following Carr and Madan (1998), we decompose the
payoff into four components
f (S
b
) =
_

0
(K S
b
)
+
f
//
(K) dK
+
_
+

(S
b
K)
+
f
//
(K) dK + f ()
+f
/
()(S
b
) A + B + C + D. (5.8)
Thanks to our assumptions on f
//
, we derive by Cauchy
Schwartz:
_

0
P
0bK
[f
//
(K)[dK +
_
+

C
0bK
[f
//
(K)[dK < . (5.9)
[f
//
(K)[(K S)
+
1{K < ] + [f
//
(K)[(S K)
+
1{K > ] is
thus P{S
b
dS]dK integrable on (0, +)
2
and the Fubini
Lebesgue theoremimplies (Malliavin and Airault (1993, p 46))
P
t b
E
t
[A+B] =
_

0
f
//
(z)P
t
(a, z) dz+
_
+

f
//
(z)C
t
(a, z) dz.
Thanks to our assumption on f
//
, we conclude that the
claim A + B is square-integrable and attainable at t by the
static trading strategy (t, b, f
//
, f
//
, dK, dK, ):
P
t b
(A + B) = P
t b
E
t
[A + B] +
_
b
t
_

0
f
//
(z) dz d
s

P
s
(b s, z)
+
_
b
t
_
+

f
//
(z) dz d
s

C
s
(b s, z).
The third component is trivially replicated by taking
position in the money market account. The fourth component
is replicated by buying a call option and selling a put option
since it is the terminal value of a forward contract. .
Corollary 5.1. The above result still holds if we replace in
(5.7) the squares by absolute values and replicated by
replicating up to a vanishing risk.
Proof. Consider the sequence of smooth functions f
n
dened
by replacing in (5.8) f
//
(K) by f
//
(K)1{[f
//
(K)[ < n]. By
applying the Lebesgue dominated convergence theorem, we
obtain lim
n
E[(f
n
f )
2
] = 0.
By application of proposition 5.2 to f
n
(S
b
), we conclude
that f
n
(S
b
) is attainable and thus at time t , the claim f (S
b
)
L
2
(, `) can be replicated up to a vanishing risk at a cost
P
t b
E
t
[f (S
b
)]. .
Proposition 5.3. Let U
t
be a square-integrable martingale
such that the claim U
T
with maturity T is attainable and let

t
be a predictable bounded process. Then the claim with
payoff at time a, Z
a
=
_
a
0

t
dU
t
L
2
(, `), can be
replicated at zero cost.
Proof. The claim with payoff U
T
/P
0T
is attained by a
self-nancing trading strategy = (
i
s
,
i
). We note that

V
t
() = E
t
[U
T
] = U
t
satises
U
t
= U
0
+
n

i=1
_
t
0

i
s
d

V
s
(
i
).
Since
t
is a predictable bounded process, (
s

i
s
,
i
)
denes a self-nancing trading strategy and we have
_
a
0

t
dU
t
=
n

i=1
_
a
0

i
t
d

V
t
(
i
).
Therefore, the claim Z
a
with maturity a is attainable at
zero cost. .
We conclude this section with a rst illustration of
how hedging in a sticky-delta model works in practice (see
remark 7.2).
Proposition 5.4. All call and put options with maturity in
[0, T ] are attainable.
Proof. Consider the long-dated call option with strike K > 0
and maturity t = kx
m
/2 + x in [0, T ] where k is an integer
and 0 < x x
m
/2. Given the set of dates {t
i
= (ix
m
/2) t :
1 i k + 1], we shall construct a sequence of static trading
strategies
i
having trading intervals (t
i
, t
i+1
] such that (1,
i
)
replicates the long-dated call option.
At time t
k
= kx
m
/2, the call option is a short-dated
option and thus can be replicated by purchasing the call option
itself. The discounted value process associated with this static
37
P Balland QUANTI TATI VE FI NANCE
strategy is

V
u
=

C
k
(t u, K) for u [t
k
, t ]. Given the sticky-
delta assumption, we obtain

V
t
k
= P
0t
E[(S
t
K)
+
[S
t
k
] = P
0t
S
t
k
c(t t
k
, K/S
t
k
),
where c(x, y) = E[(S
t
/S
t x
y)
+
] is a C
2
function which is
decreasing and convex with respect to the second argument.
The density of S
t
/S
t x
is
2
y
c(x, y) and y
2

2
y
c(x, y) is
uniformly bounded in (0, +). We note that

V L
2
(, `)
and
2
S
t
k

V
t
k
= K
2

2
y
c(t
k
, K/S
t
k
)/S
3
t
k
0. We derive
E
__
1
S
k
1

2
y
c(t
k
, K/S
k
)
2
K
6
/S
6
k
dK
_
< ,
E
__
S
k
1
1

2
y
c(t
k
, K/S
k
)
2
K
4
/S
4
k
dK
_
< .
We can thus apply proposition 5.2 and replicate the claim

V
t
k
between time t
k1
and t
k
using a static portfolio with
positive holdings in short-dated call and put options having
maturity t
k
, at a discounted cost

V
t
k1
= P
0t
E[(S
t
K)
+
[S
t
k1
] = P
0t
S
t
k1
c(t t
k1
, K/S
t
k1
).
By repeating the above argument, we obtain a set of static
trading strategy {
i
: 0 i k] with positive holdings such
that the self-nancing trading strategy (1,
i
) replicates the
long-dated call option at the following initial cost:

V
0
= P
0t
E[(S
t
K)
+
].
We obtain the result for the long-dated put options by put
call parity. .
In order to construct similar hedging strategies for square-
integrable claims, we need a representation property for X.
6. Representation property for regular
martingales with independent increments
Inthis sectionandinthe appendix, we extendthe representation
property obtained by Nualart and Schoutens (2000) for Levy
processes, to regular martingales with independent increments.
We adapt the notation and the approach taken by Nualart and
Schoutens to our purpose.
We dene the Teugels martingales on [0, T ] as in Nualart
and Schoutens (2000):
Y
(1)
t
= X
t
,
Y
(k)
t

0<st
(X
s
)
k

_
t
0
m
k
(s) ds, (k 2).
It is clear that these martingales are square-integrable
with independent increments and nite moments of all orders.
In the appendix and by [ ]-orthogonalization of the Teugels
martingales, we construct the following family of pairwise [ ]-
orthogonal martingales having independent increments:
H
(1)
t
= Y
(1)
t
,
H
(i)
t
=
_
t
0
dY
(i)
s
+ a
i,i1
(s) dY
(i1)
s
+ . . . + a
i,1
(s) dY
(1)
s
.
The deterministic functions a
ij
(s) are bounded in [0, T ]
and are such that the martingales H
(i)
t
are square-integrable
and pairwise strongly orthogonal.
We show in the appendix that the [ ]-orthogonal family
{H
(i)
: i 1] forms a complete basis of L
2
(, `). More
precisely, we have the following representation property.
Proposition 6.1. Let F L
2
(, `) then there are
predictable processes {
(i)
t
: i 1] such that
F = E[F] +
+

i=1
_
T
0

(i)
s
dH
(i)
s
,
where {
(i)
t
: i 1] belongs to
+
i=1
L
2
H
(i)
.
Proof. See appendix. .
Using this representation property, we prove in the next
section that in regular sticky-delta models, all square integrable
claims can be replicated up to a vanishing residual risk by
trading portfolios of vanilla options. We cannot typically
replicate, in the classic sense, contingent claims because the
family of martingales underlying the representation property
of proposition 6.1 has in general an innite dimension and thus
exact replication would be possible only if strategies based on
an innite number of traded instruments were admissible.
7. Quasi-completeness of regular
sticky-delta implied volatility models
In this section, we prove that all regular sticky-delta implied
volatility models are quasi-complete. But rst, we show that
the claims H
(i)
T
are attainable by application of Itos lemma to
polynomial functions. We dene f
0b
F
0b
/E[exp(X
b
)] and
observe that X
b
= ln(S
b
/f
0b
).
Lemma 7.1. Let i 2. The claim with payoff (X
b
)
i
at time b
is attainable at time a > b x
m
using the static strategy:
X
i
b
= E
a
[(X
b
)
i
] +
_
b
a
_
f
0b
0
x
(i)
(K)K
2
dKd
u

P
u
(b u, K)
+
_
b
a
_
+
f
0b
x
(i)
(K)K
2
dK d
u

C
u
(b u, K),
x
(i)
(K) = {i(i 1)(ln{K/f
0b
])
i2
i(ln{K/f
0b
])
i1
]/P
0b
.
Proof. Using equation (4.10) and proposition 5.2 with f (S) =
ln(S/f
0b
)
i
, we derive the result. .
38
QUANTI TATI VE FI NANCE Deterministic implied volatility models
Proposition 7.1. For i 1 and b (a, a + x
m
), H
(i)
b
H
(i)
a
satises
H
(i)
b
H
(i)
a
=
_
b
a
_
f
0b
0
h
(i)
u
(K)K
2
dKd
u

P
u
(b u, K)
+
_
b
a
_
+
f
0b
h
(i)
u
(K)K
2
dK d
u

C
u
(b u, K)
+
_
b
a
g
(i)
u
d
u

C
u
(b u, f
0b
)
_
b
a
g
(i)
u
d
u

P
u
(b u, f
0b
),
h
(i)
u
(K) =

0jki1
h
jk
u,i
X
j
u
(ln(K/f
0b
))
k
,
g
(i)
u
=

0ki1
g
k
u,i
X
k
u
.
The coefcients h
jk
u,i
and g
k
u,i
are deterministic and uniformly
bounded on [0, T ]. For all i 1, the claim H
(i)
T
is attainable,
i.e. H
(i)
T
V().
Proof. Y
(1)
t
= ln(S
t
/f
0t
) is a square-integrable martingale. By
proposition 5.2, we conclude that Y
(1)
b
is attainable using the
static trading strategy dened by
P
0b
Y
(1)
t
=
_
f
0b
0

P
t
(b t, K)K
2
dK

_
+
f
0b

C
t
(b t, K)K
2
dK
+(

C
t
(b t, f
0b
)

P
t
(b t, f
0b
))/f
0b
.
This equation implies for t [a, b]
Y
(1)
t
Y
(1)
a
=
_
t
a
_
f
0b
0
y
(1)
u
(K)K
2
dK d
u

P
u
(b u, K)
+
_
t
a
_
+
f
0b
y
(1)
u
(K)K
2
dK d
u

C
u
(b u, K)
+
_
t
a
f
(1)
u
d
u

C
u
(b u, f
0b
)
_
t
a
f
(1)
u
d
u

P
u
(b u, f
0b
),
where y
(1)
u
(K) =
1
P
0b
and f
(1)
u
=
1
P
0b
f
0b
.
With Itos formula for a real function f of class C
2
:

0<ut
{f (X
u
) f
/
(X
u
)X
u
] = f (X
t
) f (X
0
)

_
t
0
f
/
(X
u
) dX
u

1
2
_
t
0
f
//
(X
u
)
2
u
du.
Taking f (x) = x
2
, we derive:
Y
(2)
b
= (X
b
)
2
E[(X
b
)
2
] 2
_
b
0
X
u
dY
(1)
u
.
Finally, the above equation and lemma 7.1 imply that
Y
(2)
b
Y
(2)
a
is attainable:
Y
(2)
b
Y
(2)
a
=
_
b
a
_
f
0b
0
y
(2)
u
(K)K
2
dK d
u

P
u
(b u, K)
+
_
b
a
_
+
f
0b
y
(2)
u
(K)K
2
dK d
u

C
u
(b u, K)
+
_
b
a
f
(2)
u
d

C
u
(b u, f
0b
)
_
b
a
f
(2)
u
d

P
u
(b u, f
0b
)
y
(2)
u
(K) = 2{X
u
ln(K/f
0b
) + 1]/P
0b
f
(2)
u
= 2X
u

f
(1)
u
Similarly, we obtain by applying Itos lemma to
f (x) = x
3
:

0<ub
(X
u
)
3

_
b
0
m
3
(s) ds = (X
b
)
3
E[(X
b
)
3
]
3
_
b
0
X
u
dY
(2)
u
3
_
b
0
(X
2
u
+ V
b
V
u
+ M
(2)
b
M
(2)
u
) dY
(1)
u
where V
t
=
_
t
0

2
s
ds and M
(2)
t
=
_
t
0
m
2
(s) ds.
It follows that:
Y
(3)
b
Y
(3)
a
=
_
b
a
_
f
0b
0
y
(3)
u
(K)K
2
dK d
u

P
u
(b u, K)
+
_
b
a
_
+
f
0b
y
(3)
u
(K)K
2
dK d
u

C
u
(b u, K)
+
_
b
a
f
(3)
u
d

C
u
(b u, f
0b
)
_
b
a
f
(3)
u
d

P
u
(b u, f
0b
)
where y
(3)
u
(K) = Q
(2)
u
(X
u
, ln(K/f
0b
)), f
(3)
u
= R
(2)
u
(X
u

)
and the polynomials Q
(k)
u
(X, Y) =

0ijk
q
ij
u,k
X
i
Y
j
and R
(k)
u
(X) =

0ik
r
i
u,k
X
i
have coefcients that are
deterministic and uniformly bounded in [0, T ].
By induction, we extend the above formula to all Y
(k)
b

Y
(k)
a
. We recall that:
[H
(k)
]
b
a
=
_
b
a
dY
(k)
u
+ a
k,k1
(u) dY
(k1)
u
+ . . . + a
k,1
(u)dY
(1)
u
,
where the coefcients are deterministic and bounded in [0, T ].
We nally obtain the promised expression for H
(k)
b
H
(k)
a
by using the previous equation for Y
(k)
b
Y
(k)
a
. By adding
the above decompositions obtained for a = t
l1
, b = t
l
with
l = 1, . . . , [2T/x
m
] + 1, t
l
= (lx
m
/2) T , we obtain a self-
nancing trading strategy that replicates H
(k)
T
. .
Finally, we have the following quasi-completeness result.
Theorem 7.1. Any square-integrable claim L
2
(, `)
with maturity T can be replicated up to a vanishing risk at a
cost P
0T
E[] by trading the underlying, the money-market
account and some portfolios of traded call and put options. A
regular sticky-delta model is quasi-complete or complete up
to a vanishing risk.
Proof. According to proposition 6.1, the random variable
L
2
(, `) can be represented as follows:
= E[] +
+

i=1
_
T
0

(i)
s
dH
(i)
s
,
where
(i)
t
is in L
2
H
(i)
.
We dene
n
= E[] +

n
i=1
_
T
0

(i,n)
s
dH
(i)
s
V()
with
(i,n)
t
=
(i)
y
1{[
(i)
t
[ < n].
39
P Balland QUANTI TATI VE FI NANCE
Thanks to the strong orthogonality of H
(i)
t
, we obtain
E[(
n
)
2
] E
__
E[]
n

i=1
_
T
0

(i)
s
dH
(i)
s
_
2
_
+E
_
+

i=1
_
T
0
{
(i)
s

(i,n)
s
]
2
d[H
(i)
, H
(i)
]
s
_
.
The rst sequence on the RHS converges to zero since
{
(i)
t
: i 1] belongs to
+
j=1
L
2
H
(j)
. Observe next that:
lim
n
(
(i,n)
s

(i)
s
)
2
= 0, (
(i,n)
s

(i)
s
)
2
4(
(i)
s
)
2
.
By three applications of the dominated convergence
Lebesgue theorem, we derive
lim
n
E
_
+

i=1
_
T
0
(
(i)
s

(i,n)
s
)
2
d[H
(i)
, H
(i)
]
s
_
= 0.
Finally, we have lim
n
E[(
n
)
2
] = 0. Each claim

n
is attainable by application of propositions 7.1 and 5.3 at a
cost P
0T
E[]. .
Remark 7.1. The concept of quasi-completeness is due to
Jarrow and Madan (1999) and to Bjork et al (1997).
The next proposition shows that some square-integrable
claims can be replicated in the classical sense, by rolling a
portfolio of vanilla options.
Proposition 7.2. Consider a claim with square-integrable
payoff f (S
T
0
, . . . , S
T
n
) at time T
n
. We assume that this payoff
is such that S
T
i
. E[f (S
T
0
, . . . , S
T
n
)[`
T
i1
S
T
i
] satises,
after subtraction of a nite number of call and put payoffs,
equation (5.7) with b = T
i
. Then the claim can be replicated
by trading the underlying, the money-market account and
some portfolios of traded call and put options.
Proof. Consider a claim with maturity T
n
such as an Asian
option, a discrete barrier option, a Parisian option or a volatility
swap, that has a nite number of xing dates T
i
with 0 <
T
i
T
i1
< x
m
and a square-integrable payoff f (S
T
0
, . . . , S
T
n
)
as in proposition 7.2. At time T
n1
, the payoff can be replicated
using a portfolio of call and put options, a forward contract and
a zero coupon bond as in proposition 5.2. The value V
T
n1
of
the complex option at time T
n1
, is thus
V
T
n1
= P
T
n1
,T
n1
[
T
n1
,T
n1
f (S
T
0
, . . . , S
T
n1
, )](S
T
n1
),
where the linear operator
t x
is dened by
[
t x
f ](S)
_
+
0
f (I Sm
t x
)
2
I
C
BS
(x
t
(x, I)
2
, 1, I) dI.
The sticky-delta assumption implies that the operator

T
n1
,T
n1
is deterministic and thus V
T
n1
is a deterministic
function of S
T
0
, . . . , S
T
n1
. At time T
n2
, the complex option
can thus be replicated by the use of calls, puts, forwards and
zero coupon bonds with maturity T
n1
. Finally, we derive by
induction that the complex option can be replicated at a cost
V
0
= P
0T
n
_

0,T
0
n1

i=0

T
i
,T
i
_
(f )(S
0
) = P
0T
n
E[f ].
.
Remark 7.2. By modifying the above strategy, we obtain a
super-replication strategy for the complex option when there
are transaction costs on the implied volatility or when the
short-dated implied volatility smile is uncertain but
bounded. Observe that the transition operators
T
k
,T
k
are
nonlinear in this case.
The previous expectation can be estimated by using the
Monte Carlo method, a fast Fourier transform as in Carr and
Madan (1999) or by solving the integro-differential equation

t
V + A
t
V = r
t
V where A
t
is the generator dened by
A
t
= (
x

t x
)
x=0
+ .
8. Regular geometric Levy models
We consider a regular sticky-delta implied volatility model
and we suppose that the per-delta implied volatility processes
are independent of time. Our previous analysis in section 4
shows that the increments of ln S
t
are independent and have a
stationary distribution under the probability measure P, i.e.
S
t
= F
0t
exp(L
t
ln E[e
L
t
]),
where L is a regular P-Levy martingale (see Levy 1965).
We dene a regular geometric Levy under P by:
S
t
= F
0t
exp(L
t
ln E[e
L
t
]),
C
t
(x, K) = P
t x
E
t
[(S
t +x
K)
+
],
P
t
(x, K) = P
t x
E
t
[(K S
t +x
)
+
],
where x < x
m
and Lis an adapted regular P-Levy martingale.
Proposition 8.1. Regular geometric Levy models are
arbitrage-free and quasi-complete in the sense that all
square-integrable claims can be replicated up to a vanishing
risk by trading the underlying, the money-market account and
portfolios of short-dated call and put options.
Proof. By direct calculation, we show that regular geometric
Levy models are regular stationary sticky-delta implied
volatility models. Therefore, these models are arbitrage-free
and quasi-complete by application of theorem 7.1. .
There is a long list of geometric Levy models proposed
as alternatives to the BlackScholes model. We mention,
in particular, Mandelbrot (1963), Merton (1976), Madan and
Seneta (1990), the continuous-time formulation by Koponen
(1995) of the truncated Levy ight introduced by Mantegna
and Stanley (1994), Eberlein and Keller (1995), Barndorff-
Nielsen (1995), Bouchaud, Cont and Potters (1998), Bjork
et al (1997) and Schoutens (2001).
40
QUANTI TATI VE FI NANCE Deterministic implied volatility models
9. Characterization of sticky-strike
implied volatility models
As with sticky-delta models, we consider only regular sticky-
strike models. We denote v
t x
(K) x
t
(x, Km
t x
)
2
, the den-
sity of S
t
by p
t
and
t x
(S, K)
BS
(x
t
(x, Km
t x
)
2
, S, K).
Lemma 9.1. For each t < T , there exists a positive, locally
bounded, P-integrable function h
t
such that:
lim
k+

x
v
t x
k
(S
t
) = h
t
(S
t
) P a.s.,
where x
k
is a sequence with limit zero.
Proof. First, we note by Jensen inequality:
E
t
[(S
t +x
/m
t x
K)
+
] = E
t
[(S
t +x+y
/m
t,x+y
K)
+
].
Hence v
t x
(K) is increasing with x. To simplify notation,
we assume that p
t
> 0 in (0, +). The case where p
t
vanishes
is treated similarly since we have P(p
t
(S
t
) = 0) = 0.
Using the BlackScholes equation, we derive:

x
E
t
[S
2
t +x
/F
2
t x
] =
_
+
0

x
v
t x
(K)
t x
(S
t
, K)dK.
For

I (0, +), we dene the increasing function:
H
It x
: z .
_
I

x
v
t x
(K)E[
t x
(S
t
, K) 1{S
t
< z]dK.
We observe that for any positive z:
H
It x
(z) E[sup
x

x
E
t
[S
2
t +x
/F
2
t x
]].
Denition 2.2 implies that the increasing functions H
It x
are
uniformly bounded with respect to x and I. Hellys theorem
implies that there is a sequence x
I,n
with limit zero and an
increasing function H
I,t
such that (see Doob 1994):
lim
n+
H
It x
In
(z) = H
I,t
(z).
The increasing function H
I,t
satises for 0 a < b:
H
I,t
(b) H
I,t
(a)
_
b
a
sup
x

x
E[S
2
t +x
/F
2
t x
[S
t
]p
t
(S
t
)dS
t
.
The RadonNikodym theorem implies that there is a positive,
locally bounded, P-integrable function h
I,t
such that:
H
I,t
(b) H
I,t
(a) =
_
b
a
h
I,t
(z)p
t
(z)dz (a, b > 0).
0 h
I,t
(z) sup
0<x<T t

x
E[S
2
t +x
/F
2
t x
[S
t
= z] P a.s.
Finally, we conclude that for any compact A (0, +):
lim
n+
_
I

x
v
t x
I,n
E[
t x
I,n
(S
t
, K)1
A
(S
t
)]dK =
_
A
h
I,t
p
t
dz.
Since E[
t,x
I,n
1
A
]/p
t
converges uniformly to 1
A
on I:
lim
n+
_
I

x
v
t x
I,n
1
A
(K)p
t
(K)dK =
_
A
h
I,t
p
t
dz.
Hence
x
v
t x
I,n
(S
t
) converges in probability to h
I,t
(S
t
) in I and
there exists x
I,(n)
with limit zero, such that (see Doob 1994):
lim
n+

x
v
t x
I,(n)
(S
t
) = h
I,t
(S
t
) (S
t
I, P a.s.).
Application of Fatous theorem gives (see Doob 1994):
lim
n+
_
I
E[
t x
I,n
(S
t
, K)1
A
(S
t
)] [
x
v
t x
I,n
h
I,t
[dK = 0.
(9.1)
We dene I
k
=
_
1
k+1
,
1
k
_
[k, k + 1) and construct, as
previously, a family of positive P-integrable functions h

I
k
,t
and sub-sequences x

k
(n)
x

k1
(n)
converging to zero such
that:
lim
n+

x
v
t x

k
(n)
(S
t
) = h

I
k
,t
(S
t
) (k > 0, S
t
I
k
, P a.s.).
Using the diagonal procedure, we dene the sequence x
(n)
=
x

n
(n)
with limit zero, and the positive, locally bounded, P-
integrable function:
h
t
(S) =
+

k=1
1
I
k
(S)h

I
k
,t
(S) sup
x

x
E[S
2
t +x
/F
2
t x
[S
t
= S].
Using equation (9.1), we obtain for any compact sets A, C:
lim
n+
_
C
E[
t x
(n)
(S
t
, K)1
A
(S
t
)][
x
v
t x
(n)
h
t
[dK = 0.
(9.2)
Since

k
I
k
= (0, +), we nally conclude that:
lim
n+

x
v
t x
(n)
(S
t
) = h
t
(S
t
) P a.s.
.
In fact, we have a stronger result.
Theorem 9.1. In a regular sticky-strike implied volatility
model, the per-strike implied volatility
t
(x, K) is
independent of K. Hence the BlackScholes model is the only
arbitrage-free regular sticky-strike model.
Proof. The implied volatility model with zero-drift underlying
S
t
/m
0t
and implied volatility
t
(x, Km
0t
) is regular and
sticky-strike. Therefore, there is no loss of generality in
assuming zero drift i.e. m
0t
= 1.
For any bounded, Borel measurable function f , we have:
P
t x
E
P
t
[f (S
t +x
)] =
_
+
0

2
K
C
t
(x, K)f (K)dK.
With our assumptions, the asset price process S
t
is a Markov
process entirely characterized by the transition function

t x
(see Revuz and Yor 1991):
(

t x
f )(S
t
) = E[f (S
t +x
)[S
t
].
41
P Balland QUANTI TATI VE FI NANCE
The transition function is Feller and it is thus associated with
an innitesimal generator

A
t
dened on D
A
t
satisfying:
[

A
t
f ](S
t
) = lim
x0

x
[

t x
f ](S
t
).
Let f D
A
t
C
2
with compact support . We obtain:

x
[

t x
f ](S
t
) =
_

x
C
BS
(v
t x
(K), S
t
, K)f
//
(K)dK. (9.3)
The BlackScholes function satises:

V
C
BS
(V, F, K) =
1
2
F
2

BS
(V, F, K).
Hence, equation (9.3) can be written as follows:

x
[

t x
f ](S
t
) =
1
2
_

S
2
t

t x
(S
t
, K)
x
v
t x
(K)f
//
(K)dK.
According to lemma 9.1, there is a positive, locally bounded
P-integrable function h
t
such that:
lim
n

x
v
t x
n
(S
t
) = h
t
(S
t
) P a.s.
By taking rst the expectation of equation (9.3) on a compact
C and then the limit as x
n
tends to zero, we obtain with (9.2):
_
C
[

A
t
f ](K)p
t
(K)dK =
1
2
_
C
h
t
(K)K
2
f
//
(K)p
t
(K)dK.
We nally obtain the following expression for the generator:
[

A
t
f ](S
t
) =
1
2
h
t
(S
t
)S
2
t
f
//
(S
t
) P a.s. (9.4)
The forward price V(t, S
t
) of a call option with maturity a and
strike K is a P-martingale. Therefore, V satises the backward
equation (see Revuz and Yor 1991):

t
V(t, S
t
) + [

A
t
V](S
t
) = 0. (9.5)
On the other hand, V satises:
V(t, S
t
) = C
BS
((a t )
t
(a t, K)
2
, S
t
, K).
Since V has the same second-order derivatives with respect to
S as the BlackScholes function, it follows that:

t
V(t, S
t
) =
1
2

t
[(a t )
t
(a t, K)
2
]S
2

2
S
V(t, S
t
). (9.6)
Finally, equations (9.4)(9.6) imply that we have for all
K, t, a (t, t + x
m
):

t
(a t, K)
2
=
1
a t
_
a
t
E[h
u
(S
u
)]du.
Hence, the implied volatility
t
(x, K) is independent of K
and the regular sticky-strike model coincides with the Black
Scholes model! .
Acknowledgments
I would like to thank L P Hughston, D B Madan, T Bjork,
W Schoutens, D Nualart, R Cont, an anonymous referee, the
participants of the EURANDOM seminar on Levy processes
(2001), the participants of the AMS Meeting (2001) and
my colleagues at Merrill Lynch for stimulating and fruitful
discussions.
Appendix. Representation property for
regular martingales
In this appendix, we extend the SchoutensNualart
representation property obtained for regular Levy martingales
to regular martingales with independent increments. We use
the notations of section6andwe followcloselythe presentation
of Nualart and Schoutens (2000).
We recall that two square-integrable martingales M and N
are said to be strongly orthogonal or [ ]-orthogonal if [M, N]
t
is a martingale (see Protter 1995).
Proposition A.1. There exists a family of pairwise strongly
orthogonal square-integrable martingales {H
(i)
: i 1].
Proof. We dene the following inner product acting on the
space of real polynomials with time-dependent coefcients in
L
2
(0, T ):
Q, R) =
_
T
0
_
+

Q(x, s)R(x, s)(x

2
n
s
(dx)
2
s
(x)dx)ds.
By )-orthogonalization of the total family {1{s < t ]x
i
:
i 0, t (0, T )], we nd a family of pairwise )-orthogonal
polynomials {R
i
(x, s)] with bounded coefcients such that:
R
i
(x, s) = a
i+1,1
(s) + a
i+1,2
(s)x + . . . + a
i+1,i
(s)x
i1
+ x
i
,
R
i
, 1{s < t ]x
j
) = 0, (t < T, 1 j i 1).
The coefcients of the above polynomial are bounded
because the measure n(s, dx) has nite moments of order
i 2, uniformly bounded in [0, T ].
Dene Q
i+1
(x, s) x(R
i
(x, s) a
i+1,1
(s)), a
i+1,i+1
1
and the following square integrable martingale:
H
(i+1)
t

_
t
0
a
i+1,1
(s)dY
(1)
s
+ . . . + a
i+1,i+1
(s) dY
(i+1)
s
.
By direct calculation, we derive:
H
(i+1)
t
=
_
t
0
a
i+1,1
(s)dX
s
+

0<st
Q
i+1
(S
s
, s)

i+1

k=2
_
t
0
a
i+1,k
(s)m
k
(s)ds.
We note that the processes [Y
(k)
, Y
(j)
]
t
and [H
(i+1)
, Y
(j)
]
t
have independent increments. Furthermore, we have for
t [0, T ] and j = 1, . . . , i:
E[H
(i+1)
, Y
(j)
]
t
= E
__
t
0
i

k=1
a
i+1,k
(s) d[Y
(k)
, Y
(j)
]
s
_
= R
i
(x, s), 1{s < t ]x
j1
) = 0.
42
QUANTI TATI VE FI NANCE Deterministic implied volatility models
Therefore H
(i+1)
is strongly orthogonal to all Y
(j)
for
j = 1, . . . , i. The martingales H
(j)
are consequently pairwise
strongly orthogonal. .
Remark A.1. We observe that if the polynomial R
i
is such
that R
i
, R
i
) = 0 then the martingale H
(i+1)
satises
E[H
(i+1)
, H
(i+1)
]
T
= 0 and thus H
(i+1)
= 0 almost
everywhere. If the discontinuous component of the regular
martingale is the sum of a nite number of Poisson processes
with deterministic intensity then only a nite number of the
martingales H
(k)
will be non-zero.
We dene the following families of L
2
(, `) variables:

t
1
...t
n
= {X
k
1
t
1
(X
t
2
X
t
1
)
k
2
. . . (X
t
n
X
t
n1
)
k
n
: k
i
0],
= {X
k
1
t
1
. . . (X
t
n
X
t
n1
)
k
n
: 0 t
i
< t
i+1
T, k
i
0].
Lemma A.2. The family
t
1
...t
n
is total in the space
L
2
(, (X
t
1
, X
t
2
X
t
1
, . . . , X
t
n
X
t
n1
)).
Proof. Let us prove the result for the family
t
. The case with
n non-overlapping independent increments is treated similarly.
Since X is regular, the linear hull

t
of
t
is in L
2
(P{X
t

dx]). The space C of functions in L
2
(P{X
t
dx]) having
compact support is dense in L
2
(P{X
t
dx]). Let F C
with support in [a, a] and weakly orthogonal to

t
, i.e.
E[(X
t
)
k
F(X
t
)] = 0, (k 0). For any real Z, we have:
+

k=0
[Z[
k
k!
E[[X
t
[
k
[F(X
t
)[] < E[F(X
t
)
2
]
1/2
exp([Za[).
By the dominated convergence theorem, we derive
E[exp(iZX
t
)F(X
t
)] = 0 for all Z and thus F(X
t
) = 0
(Malliavin and Airault (1994, p 110)). Therefore

t
is dense
in C and thus in L
2
(P{X
t
dx]). .
Proposition A.2. The family is total in L
2
(, `).
Proof. A variable Z in L
2
(, `) can be approximated
arbitrarily closely by an element Y of L
2
(, (X
t
1
, X
t
2

X
t
1
, . . . , X
t
n
X
t
n1
)) for some sequence {t
i
]. Following
lemma A.2
t
1
...t
n
is total in L
2
(, (X
t
1
, . . . , X
t
n
X
t
n1
)).
Therefore, Y and thus Z can be approximated arbitrarily
closely by an element of

. .
We have the following representation property for
elements of .
Lemma A.3. For any integer k and any power-increment
(X
s
1
X
s
0
)
k
, there is a sequence of predictable processes
{
(i)
s
0
s
1
k
(s) : i 1] such that:
(X
s
1
X
s
0
)
k
= E[(X
s
1
X
s
0
)
k
] +
+

i=1
_
s
1
s
0

(i)
s
0
s
1
k
(s) dH
(i)
s
.
Proof. We prove the above equation by induction on k and by
application of Itos formula to power functions as in Nualart
and Schoutens (2000). .
Proposition A.3. Let R ; then there are predictable
processes {
(i)
t
: i 1] such that:
R = E[R] +
+

i=1
_
T
0

(i)
s
dH
(i)
s
.
Proof. With lemma A.3, we derive that the product of non-
overlapping power-increments Y
kl
([X]
s
1
s
0
)
k
([X]
s
3
s
2
)
l
with
s
0
< s
1
s
2
< s
3
can be represented as follows:
Y
kl
= E[Y
kl
] +
+

i=1
_
T
0

(i)
s
dH
(i)
s
.
The process
(i)
s
is a predictable process dened by:

(i)
s
= 1{s (s
2
, s
3
)]
(i)
s
2
s
3
l
(s)
+

j=1
_
s
1
s
0

(j)
s
0
s
1
k
(u) dH
(j)
u
.
Hence, we have proved the result for the product of two
non-overlapping power-increments. By induction, we prove
the result for the product of an arbitrary number of non-
overlapping power-increments. .
Since the linear space

spanned by is dense in
L
2
(, `), we deduce the following representation property
of square-integrable variables.
Proposition 6.1. Let F L
2
(, `) then there is a family of
predictable processes {
(i)
t
: i 1] such that:
F = E[F] +
+

i=1
_
T
0

(i)
s
dH
(i)
s
,
where
(i)
t
belongs to L
2
H
(i)
.
Proof. For a square-integrable adapted martingale m, we recall
that:
L
2
m
=
_
H :
_
T
0
H
s
dm
s
L
2
(, `)
_
,
where is the predictable -algebra on [0, T ] . We need
to prove that:
L
2
(, `) =
+

i=1
L
2
H
(i)
,
+

i=1
L
2
H
(i)

_
X L
2
(, `) : X=
+

i=1

(i)
H
(i)
,
(i)
L
2
H
(i)
_
.
We observe that
+
i=1
L
2
H
(i)
is closed in L
2
(, `) since the
martingales H
(i)
are pairwise strongly orthogonal. Thanks to
proposition A.3, we have:

i=1
L
2
H
(i)
L
2
(, `).
Since

= L
2
(, `) by proposition A.2, we derive the
result by closure of the above inclusion. .
43
P Balland QUANTI TATI VE FI NANCE
References
Balland P and Hughston L P 1999 Sticky-delta model Derivatives
Week
Barndorff-Nielsen O E 1995 Normal inverse Gaussian distributions
and the modeling of stock returns Research Report no 300
Department of Theoretical Statistics, Aarhus University
Bjork T, Masi D G, Kabanov Y and Runggaldier W 1997 Towards a
general theory of bond markets Finance Stochastics 1 14174
Black F and Scholes M 1973 The pricing of options and corporate
liabilities J. Political Economy 81 63759
Bouchaud J P, Cont R and Potters M 1998 Financial markets as
adaptive systems Europhys. Lett. 41 3
Breeden D and Litzenberger R 1978 Prices of state-contingent
claims implicit in option prices J. Business 51 62151
Brezis H 1992 Analyse Functionelle. Theorie et Applications (Paris:
Masson)
Carr P and Madan D 1998 Towards a theory of volatility trading
Volatility ed R A Jarrow (London: Risk) pp 41727
Carr P and Madan D 1999 Option valuation using the fast Fourier
transform J. Comput. Finance 6173
Cont R and da Fonseca J 2001 Deformation of implied volatility
surfaces: an empirical analysis Empirical Approaches to
Financial Fluctuations ed H Takayasu (Tokyo: Springer)
Derman E 1999 Regimes of volatility Risk 4 559
Derman E and Kani I 1994 Riding on a smile Risk 7 329
Derman E and Kani I 1998 Stochastic implied trees: Arbitrage
pricing with stochastic term and strike structure of volatility
Int. J. Theor. Appl. Finance 1 61110
Doob J L 1994 Measure Theory (New York: Springer)
Dumas B, Fleming J and Whaley R 1998 Implied volatility
functions: empirical tests J. Finance 53 2059
Dupire B 1994 Pricing with a smile Risk 7 1820
Eberlein E and Keller U 1995 Hyperbolic distributions in nance
Bernoulli 1 28199
Heston S 1993 A closed form solution for options with stochastic
volatility with applications to bond and currency options Rev.
Financial Studies 6 32743
Hull J C and White A 1987 The pricing of options with stochastic
volatilities J. Finance 42 281300
Jacod J and Protter P 1991 Probability Essentials (Berlin: Springer)
Jacod J and Shiryaev A N 1987 Limit Theorems for Stochastic
Processes (Berlin: Springer)
Jarrow R A and Madan D 1999 Valuing and hedging contingent
claims on semimartingales Finance Stochastics 3 11134
Koponen I 1995 Phys. Rev. E 52 1197
Lebedev V A 1995 Fubini-theorem for parameter-dependent
stochastic integrals with respect to L
0
-valued random measures
Probab. Theory Appl. 40 31323
Levy P 1965 Processus Stochastiques et Mouvement Brownien
(Paris: Gauthier-Villars)
Madan D B and Seneta E 1990 The variance gamma model for share
market returns J. Business 63 51124
Malliavin P and Airault H 1994 Integration et Analyse de Fourier
(Paris: Masson)
Mandelbrot B B 1963 The variation of certain speculative prices J.
Business 36 394419
Mantegna R N and Stanley H E 1994 The truncated levy ights
Phys. Rev. Lett. 73 2946
Merton R 1976 Option pricing when underlying stocks returns are
discontinuous J. Financial Economics 3 12544
Nualart D and Schoutens W 2000 Chaotic and predictable
representations for Levy processes Stochastic Processes
Applications 90 10922
Protter P 1995 Stochastic Integration and Differential Equations
(Berlin: Springer)
Reiner E 1999 Volatility rules and implied processes Global
Derivative Conf.
Revuz D and Yor M 1991 Continuous Martingales and Brownian
Motion (Berlin: Springer)
Schonbucher P J 1999 A market model for stochastic implied
volatility Phil. Trans. R. Soc. A 357 207192
Schoutens W 2001 Meixner processes in nance EURANDOM
Report 2001-002
Schoutens W 2000 Stochastic Processes and Orthogonal
Polynomials (Lecture Notes in Statistics) vol 146 (Berlin:
Springer)
44