Complete Market Models
Complete Market Models
Complete Market Models
Volatility
By Mark H.A. Davis
Department of Mathematics, Imperial College, London SW7 2AZ, UK
In the Black-Scholes option pricing theory, asset prices are modelled as geometric
Brownian motion with a fixed volatility parameter , and option prices are determined as functions of the underlying asset price. Options are in principle redundant
in that their exercise values can be replicated by trading in the underlying. However,
it is an empirical fact that the prices of exchange-traded options do not correspond
to a fixed value of as the theory requires. This paper proposes a modelling framework in which certain options are non-redundant: these options and the underlying
are modelled as autonomous financial assets, linked only by the boundary condition at exercise. A geometric condition is given, under which a complete market is
obtained in this way, giving a consistent theory under which traded options as well
as the underlying asset are used as hedging instruments.
Keywords: Financial options, Black-Scholes, volatility, vega hedging,
stochastic flows, Bismut formula
1. Introduction
The Black-Scholes theory is based on an asset price model which, in a risk-neutral
measure Q, takes the form
dSt = rSt dt + St dwt ,
(1.1)
where r is the riskless rate, wt is a Brownian motion and is the volatility. In this
paper we are not concerned with interest-rate volatility generally a minor factor
in equity option pricing so r will be taken as a constant (sometimes 0). We also
assume the asset pays no dividends. As is well known, the price model (1.1) leads
to a 5-parameter formula C(S, K, r, , T ) for the price of a call option. Of these
parameters, (K, T ) (strike and exercise time) define the option contract while (S, r)
are market data, leaving the formula essentially as a map 7 p = C(S, K, r, , T )
from volatility to price. Because the call option exercise function is convex, p is an
increasing function of , and we can compute the inverse map, the so-called implied
volatility.
Evidence from the traded option market shows that the model (1.1) is not an
accurate description of reality (see Ghysels et al. 1996 for a comprehensive survey
and Tompkins 2001 for empirical evidence). Figure 1 shows the implied volatility for
FTSE index options for a range of strike prices and maturity dates, while Figure 2
shows the evolution of at-the-money implied volatility over a 15-year period. Figure
1 shows that the log-normal distribution of ST implied by model (1.1) cannot be
correct, and Figure 2 shows that volatility is in some sense stochastic.
Article submitted to Royal Society
TEX Paper
Im p lie d V o l
27%
26%
25%
June
July
Sept
Dec
24%
23%
22%
21%
20%
5825
5925
6025
6125
6225
6325
6425
6525
Strike
Complete-market Models
40%
35%
30%
25%
20%
15%
10%
5%
Jan-88
Jan-89
Jan-90
Jan-91
Jan-92
Jan-93
Jan-94
Jan-95
Jan-96
Jan-97
or the implied tree models of Derman and Kani (1998) or Dupire (1994). By suitably choosing () one can obtain price distributions that match observed volatility
smiles. However, these models are somewhat restricted and the implications for
hedging are not at all clear. In particular, since options are redundant these models
say nothing about vega hedging (see below). They also contradict the empirical fact (Tompkins, 2001) that price and volatility are not perfectly correlated, as
implied by (2.1).
A more interesting class of models arises from the following observation: it is
easily checked that the 2-vector random variable
w
Rt t
w ds
0 s
t
1 2
2t
1 2
2t
1 3
3t
dx1
1
1/2
1/2
(x1 + x1 /2)dt + x1 x0 dwt
2
( x1 )dt + dwt
x
)x
x0
1 1
1
2
4
2
The H
ormander condition is satisfied if rank(A1 , [A0 , A1 ]) = 2 and in fact the
determinant is
1
1
1 1/2
1
3/2
1/2
.
x1
1 x1
x0
2 x1
2
8
2
2
The determinant is non-zero for all x1 > 0 for reasonable ranges of the coefficients
, , , for instance the representative values = .0025, = .05, = .01.
The above calculations show that we can produce complete-market models
where the volatility is stochastic in the sense that it is not just a function of
current price. While this is satisfactory from an econometric standpoint, the trading strategies that these models imply delta-hedging in the underlying asset
are completely unrealistic. The obvious way to hedge volatility is to use traded options, and these models give us no clue how to do so, since all options are in theory
redundant.
(b) Multi-factor models
The standard way to hedge against volatility risk is vega hedging. The vega
of an option C is v = C/, the sensitivity of the Black-Scholes value to changes
Article submitted to Royal Society
Complete-market Models
S(t)dt + (t)S(t)dwt
a(S(t), (t))dt + b(S(t), (t))dwt
where a, b define the volatility model and wt , wt are Brownian motions with constant correlation Edwt dwt = dt, so movements of volatility are possibly correlated
with movements of underlying asset price. Well known models of this type are those
0 0
of Hull and White (1987) and Heston (1993). We can write
p wt = wt + wt where
wt0 is a Brownian motion independent of wt and 0 = 1 2 . Measures Q equivalent to P then have densities of the form
!
Z
Z
Z T
Z T
1 T 2
1 T 2
dQ
0
= exp
s dws
s ds +
s dws
s ds
(2.5)
dP
2 0
2 0
0
0
for some integrands , . Taking = (r )/ and = (S, ) we find that the
equations for S, under measure Q are
dS(t) = rS(t)dt + S(t)dw
et
d(t) = e
a(S(t), (t))dt + b(S(t), (t))dw
et
where w,
e w
e are Q-Brownian motions with Edwd
e w
e = dt and e
a(S, ) = a +
0
b + b . Then S(t) has the riskless growth rate r, but is not a traded asset so
arbitrage considerations do not determine the drift of , leaving as an arbitrary
choice. Suppose we now have an option written on S(t) with exercise value g(S(T ))
at time T . We define its value at t < T to be
h
i
C(t, S(t), (t)) = EQ er(T t) g(S(T )) S(t), (t) .
C
2C
1
2C
1 2C
C
C
+ rs
+e
a
+ 2 s2 2 + b2 2 + sb
rC = 0
t
s
2
s
2
s
and we find that the process Y (t) := C(t, S(t), (t)) satisfies
dY (t) = rY (t)dt +
C
C
Sdw
e+
bdw
e .
s
(2.6)
(2.7)
where w
t is another Brownian motion, again correlated with w
et . S(t) and Y (t)
are linked by the fact that, at time T , Y (T ) = g(S(T )). We have now created
a complete market model with traded assets S(t), Y (t) for which Q is the unique
EMM. By trading these assets we can perfectly replicate any other contingent claim
in the market. We have however created a whole range of such models, one for
each choice of the integrand in (2.5). The choice of ultimately determines the
volatility structure F of Y (t) in (2.7), which is all that is relevant for hedging. This
choice is an empirical question. The relationship with implied volatility is clear:
if BS(t, S, ) denotes the Black-Scholes price at time t with volatility parameter
, then the implied volatility
b(t) must satisfy Y (t) = BS(t, S(t),
b(t)), so each
stochastic volatility model implicitly specifies a model for implied volatility.
=
=
=
E[er(T2 t) Y |Ft ],
t [0, T2 ]
E[e
[Y K2 ]+ |Ft ], t [0, T2 ]
E[er(T1 t) [S(T1 ) K1 ]+ |Ft ], t [0, T1 ].
r(T2 t)
(3.1)
(3.2)
(3.3)
As in the previous section, such a model automatically specifies a model for implied
volatilities
b1 ,
b2 of the two options, which (with obvious notation) must satisfy
Oi (t) = BS(Ti t, S(t),
bi (t), Ki ) i = 1, 2.
Complete-market Models
Associated with this is the backward equation, for a function v(t, x):
v
+ Av rv
t
v(T, x)
(t, x) [0, T ] R3
0,
h(x).
(3.4)
(3.5)
In this equation, h is given boundary data at some terminal time T and A is the
generator of t , i.e.
Af (x) = f (x) m(x) +
1X
2f
((x)T (x))i,j
(x).
2 i,j
xi xj
=
=
PT t h(x)
h
i
Et,x er(T t) h(T ) .
i = 1, 2,
(3.6)
where h2 (x) = [h0 (x) K]+ and h1 (x) = [PT2 T1 h0 (x) K1 ]+ . By the Ito formula,
the discounted asset price satisfies
d ert St = ert v0 dw,
(3.7)
(3.8)
The first three terms on the right are the increments in value of the holdings in
underlying asset and options, and the fourth term indicates that all residual value
Article submitted to Royal Society
is held in the riskless account, where it earns interest at rate r. Let tilde denote
t = ert Xt etc. Applying the Ito formula to (3.8) and using
discounted quantities: X
(3.7) we obtain
t
dX
=
=
1 (t) + 2 (t)dO
2 (t)
0 (t)dSt + 1 (t)dO
!
2
X
ert
i (t)vi dw.
(3.9)
i=0
v0 (t, x)
G(t, x) = v1 (t, x)
v2 (t, x)
(3.10)
is nonsingular for all (t, x) [0, T ] R3 . Then we have a complete market model
and the hedge ratios for hedging any other contingent claim are given by (3.12)
below.
Proof. Let H be the exercise value of a contingent claim exercised at time T < T1 ,
i.e. H is an FT -measurable random variable with E[H 2 ] < . By the martingale
representation theorem for Brownian motion (Rogers & Williams 2000, Theorem
RT
IV.36.1), there is an integrand t such that E 0 |t |2 dt < and
erT H = E[erT H] +
t dwt .
(3.11)
(3.12)
Then ert G = , and we see from (3.9) and (3.11) that H = XT a.s. if X0 =
E[erT H]. Thus arbitrary contingent claims can be replicated and the market is
complete.
As a special case, suppose that H = h(T ) and define
v(t, x) = PT t h(x),
t T.
(3.13)
(3.14)
Complete-market Models
Underlying spot
Option strike K1
Option maturity T1
Implied vol
Option price
Delta
Vega
Riskless rate r
Dividend yield
100
96
1
20.085%
10.00
0.6193
38.10
0
0
next section. This question has been considered by Bajeux-Besnainou and Rochet
(1996).
The advantage of our approach is that there is no calibration, since market
option prices are inputs to the model, and no complicated conditions to avoid
arbitrage (as in Sch
onbucher, 1999), as the model is automatically arbitrage-free.
On the other hand the implied model for implied volatility is rather indirect, leaving
us with the problem of determining good classes of factor processes t to capture
the volatility structures we need. In the following section we present some quick
calculations using the Hull-White volatility model (Hull & White, 1987).
(b) An example
Here we will assume there is just one exchange-traded option, maturing at time
T1 . The parameter values are as shown in Table 3; in particular, purely for ease of
exposition, the riskless rate r is zero. The factor process (t) = (1 (t), 2 (t)) is
d1 (t)
=
=
d2 (t)
d2
1
1
log(S/K) + a
a
2
d1 a.
10
=
=
=
(3.15)
(Of course, there is positive probability that A < 0; we take BS(S, K, a) to be equal
to the intrinsic value when a 0.) A short computation shows that
Z
where
b(T, ) =
1
1 eT
c(T, , 0 ) = 02 (T b(t, ))
and Z is a zero-mean gaussian random variable with standard deviation
p
T 2b(T, ) + b(T, 2).
(T, , ) =
In view of these expressions and the fact that the t equations are time-invariant,
we can express the option value (3.15) as (with = T t)
r
Z
1 z2 /2
e
O1 (t) =
BS[1 (t), K1 , b(, )2 (t) + c(, , 0 ) + (, , )z]
dz.
2
(3.16)
This is the convenient representation noted by Hull and White (1987). In this
case the inverse problem is readily solved: given St = s1 and O1 (t) = o1 , the
corresponding values of the factor process are 1 (t) = s1 and 2 (t) = x0 , where
x0 is the value of 2 (t) such that (3.16) is satisfied when the left-hand side is
equal to o1 . Since the right-hand side is monotone increasing in 2 (t), thispvalue
is easily found by one-dimensional search. Figure 4 shows the values of 2 (0)
corresponding to different values of , 0 . The Black-Scholes implied volatility is
shown for comparison in the right-hand column.
The option O we wish to hedge has strike K = 110 and matures at T = 0.5.
Taking the volatility as the implied volatility of O1 , the Black-Scholes value of O is
2.231 and the delta and vega are = 0.2742, = 23.56. Thus referring to Figure
3 the standard vega hedge at time 0 has 1 = 23.56/38.10 = 0.618 units of O1 ,
leaving a residual delta of -0.109.
To calculate the hedge corresponding to our stochastic volatility model we simply apply the 2 2 version of formula (3.14), computing the gradients by finite
Article submitted to Royal Society
Complete-market Models
11
25%
s q rt(x 2 (0 ))
20%
15%
sig0=15%
sig0=25%
10%
5%
0%
0
0.05
Figure 4. Values of
0.1
0.15
0.2
0.25
BS
vol of vol
0.8
0.6
0.4
a0,
a1,
a0,
a1,
0.2
s0=15%
s0=15%
s0=25%
s0=25%
0.0
0
0.05
0.1
0.15
0.2
0.25
BS
-0.2
-0.4
vol of vol
differences. The resulting initial hedge parameters 0 (0), 1 (0) are shown in Figure
5 for the same range of , 0 as in Figure 4. Recall that in this model these are
the true hedge parameters for a perfectly-replicating portfolio. The standard vega
hedge is shown on the right, for comparison.
The hedge parameters vary surprisingly little over the different model parameters. The truth is that this class of models doesnt have much pizazz. With only
one traded option we are not capturing any smile effect, and the independence
of the two underlying Brownian motions is computationally convenient but hardly
realistic. The example however gives us a proof of principle: the method has been
completely implemented and the hedge parameters computed. With more realistic
models the procedure would be exactly the same, but with efficient PDE solvers
replacing the one-dimensional integration (3.16).
Article submitted to Royal Society
12
v1 (t, x)
..
G(t, x) =
,
.
vn (t, x)
(3.18)
where denotes the gradient in the x variables. Consider first the Brownian case.
Proposition 3.2. Suppose t is Brownian motion in Rn . Then G(t, x) defined by
(3.18) is non-singular if and only if there exists no non-zero vector Rn such
that
H L{T1 , . . . , Tn }.
(3.19)
Here L{ } denotes the linear subspace spanned by the indicated random variables
in L2 (, FT , Pt,x ), and
n
X
k hk (T ).
H =
k=1
hi (y)e|yx|
/2
dy,
Complete-market Models
so that
vi
(t, x)
xj
=
=
=
13
Z
2
1
1
hi (y)(yj xj )e|yx| /2 dy
(2 )n/2
1
Et,x [hi (T )(Tj xj )]
1
covt,x (hi (T ), Tj ).
Thus
1
cov(Hi , T )
(3.20)
G =
n
X
i cov(Hi , T ) = cov(H , T ),
i=1
k > 1.
n
X
i (s,t (x))dwti ,
s,s = x
(3.21)
i=1
dDs,t =
X i
m
Ds,t dwti ,
Ds,t dt +
x
x
1
Ds,s = I.
14
( is the matrix whose ith column is i .) From (3.22) we immediately obtain the folRT
lowing integration by parts formula: for a vector process u such that Es,x s |ut |2 dt <
we have
"
# Z
Z T
n
T
X
1
i
i
E h(s,T (x))
ut dwt =
E h(s,T (x))Ds,T Ds,t
(s,t (x))ut dt .
s
(3.23)
If Ut is an n n matrix-valued process each of whose components satisfies the
integrability condition then
"
# Z
Z T
T
1
T
E h(s,T (x))
dwt Ut =
E h(s,T (x))Ds,T Ds,t
(s,t (x))Ut dt . (3.24)
s
Indeed, (3.24) is an equality between row vectors in which the jth column is (3.23)
with ut equal to the jth column of Ut .
Define v(s, x) = E[h(s,T (x))]; then v(s, x) = E[h(s,T (x))Ds,T ], and taking
1
Ut = 1 Ds,t , so that Ds,t
U = I, we obtain from (3.24) the following version of
the Bismut formula (Bismut 1984):
#
"
Z T
1
dwtT 1 (s,t (x))Ds,t
(3.25)
v(s, x) =
E h(s,T (x))
T s
s
We can now state the generalization of Proposition 3.2.
Proposition 3.3. Suppose s,t satisfies (3.21) and let vi and G be defined by (3.17)
and (3.18) respectively. For fixed (s, x) define random variables Y1 , . . . , Yn by
n Z T
X
Yj =
1 (s,t (x))Ds,t ij dwti .
i=1
H =
n
X
k hk (s,T (x)).
k=1
n
X
i=1
"
i E h(s,T (x))
T
s
Complete-market Models
15
Elworthy & Li (1994) give the Bismut formula in a geometric setting. Singularity
of G can then be interpreted in terms of vanishing of the first component of a certain
Wiener chaos expansion of H . The implications of this will be explored in later
work.
Thanks to Robert Tompkins for discussions on stochastic volatility, to participants at the
Financial Options Research Centre conference, Warwick University, September 2002, for
helpful comments and suggestions, and to David Elworthy and Alexander Grigoryan for
pointing out interesting directions in connection with the non-singularity problem.
Appendix A. The H
ormander theorem
Full details of the following can be found in section IV.38 of Rogers and Williams
(2000), or section 2.3 of Nualart (1995).
For two continuous semimartingales X, Y the Stratonovich integral is defined as
Z
t
0
Y dX =
Y dX +
0
1
< Y, X >t ,
2
d
X
1
Ai f (t ) dwi (t)
n
X
mk (x)
f
,
xk
Ai f (x) =
n
X
1
ik (x)
f
, i = 1 . . . d.
xk
The Lie bracket of two vector fields Ai , Aj is the vector field [Ai , Aj ] = Ai Aj Aj Ai .
Theorem 3.4. Suppose that the coefficients of the SDE (A 1) are infinitely differentiable with bounded derivatives of all orders, and that the vector space spanned
by the vector fields
A1 , . . . , Ad , [Ai , Aj ], 0 i, j, d, [Ai , [Aj , Ak ]], 0 i, j, k d, . . .
at the initial point x0 is equal to Rn . Then for any t > 0 the random vector t has
a density that is absolutely continuous with respect to Lebesgue measure.
Article submitted to Royal Society
16
References
Babbar, K. 2001 PhD Thesis, Imperial College London.
Bajeux-Besnainou, I. and Rochet, J.-C. 1996 Dynamic spanning: are options an appropriate instrument?, Mathematical Finance 6, 116
Barndorff-Nielsen, O.E and Shephard, N. 2001 Non-gaussian Ornstein-Uhlenbeck-based
models and some of their uses in financial economics (with discussion), JRSS(B) 63,
167241.
Bergman, Y.Z., Grundy, R.D. and Wiener, Z. 1996 General properties of option prices,
Journal of Finance 51, 15731610.
Bismut, J.-M. 1984 Large deviations and the Malliavin calculus, Boston: Birkh
auser.
Black, F. and Scholes, M. 1973 The pricing of options and corporate liabilities, J. Political
Economy 81, 637654.
Davis, M.H.A. 1980 Functionals of diffusion processes as stochastic integrals, Math. Proc.
Camb. Phil. Soc. 87, 157166
Davis, M. 2001 Mathematics of financial markets. In Mathematics Unlimited: 2001 and
Beyond (eds B. Engquist and W. Schmid), Heidelberg: Springer-Verlag.
Derman, E and Kani, I. 1998 Stochastic implied trees: arbitrage pricing with stochastic
term and strike stucture of volatility, Int. J. Theor. and Appl. Finance 1 61110.
Dowd, K. 1998 Beyond value at risk: the new science of risk management, Chichester:
Wiley.
Dupire, B. 1994 Pricing with a smile, Risk 7, 1820.
El Karoui, N., Jeanblanc-Picque, M and Shreve, S.E. 1998 Robustness of the Black and
Scholes formula, Mathematical Finance 8, 93126.
Elliott, R.J. and Kohlmann, M. 1989 Integration by parts, homogeneous chaos expansions
and smooth densities, Ann. Prob. 17, 194207.
Elworthy, K.D. and Li, X.-M. 1994 Formulae for the derivatives of heat semigroups, J.
Functional Anal. 125, 252286.
Fouque, J.-P., Papanicolaou, G. and Sircar, K.R 2000 Derivatives in financial markets
with stochastic volatility. Cambridge University Press.
Ghysels, E., Harvey, A.C. and Renault, E. 1996 Stochastic volatility. In Handbook of
statistics Vol. 14 (eds G.S. Maddala and C.R. Rao), Amsterdam: Elsevier.
Heston, S.L. 1993 A closed-form solution for options with stochastic volatility with applications to bond and currency options, Rev. Fin. Studies 6, 327343.
Hobson, D.G. and Rogers, L.C.G. 1998 Complete models with stochastic volatility, Mathematical Finance 8, 2748.
Hull, J.C. 2000 Options, Futures and Other Derivatives, 4th ed., Upper Saddle River NJ:
Prentice Hall.
Hull, J.C. and White, A. 1987 The pricing of options on assets with stochastic volatility,
J. Finance 42, 281300.
Nualart, D. 1995 The Malliavin calculus and related topics, Heidelberg: Springer-Verlag.
Rogers, L.C.G. and Williams, D. 2000 Diffusions, Markov processes and martingales, Cambridge University Press.
Romano, M. and Touzi, N. 1997 Contingent claims and market completeness in a stochastic
volatility model, Mathematical Finance 7, 399410.
Sch
onbucher, P.J. 1999 A market model for stochastic implied volatility, Phil. Trans. R.
Soc. Lond. A 357, 20712092.
R. Tompkins 2001 Stock index futures markets: stochastic volatility models and smiles,
Journal of Futures Markets 21, 4378.