A Benchmark Approach

to Filtering in Finance

Eckhard Platen1 and Wolfgang J. Runggaldier2

March 11, 2002

Abstract. The paper proposes the use of the growth optimal

portfolio for the construction of
nancial market models with
unobserved factors that have to be
ltered. This benchmark
approach avoids any measure transformation for the pricing of
derivatives. The suggested framework allows to measure the
reduction of the variance of derivative prices for increasing degrees
of available information.

1991 Mathematics Subject Classi
cation: primary 90A09 secondary 60G99, 62P20.

JEL Classi
cation: G10, G13
Key words and phrase: Financial modelling,
lter methods, benchmark approach,
growth optimal portfolio.

1 University of Technology Sydney, School of Finance & Economics and Department of

Mathematical Sciences, PO Box 123, Broadway, NSW, 2007, Australia
2 Universit
a degli Studi di Padova, Dipartimento di Matematica, Pura ed Applicata
Via Belzoni, 7 I - 35131 Padova, Italy
1 Introduction
In nancial modelling it is sometimes the case that not all quantities, which
determine the dynamics of security prices, can be fully observed. Some of the
factors that characterize the evolution of the market are hidden. However, these
unobserved factors may be essential to reect in a market model the type of
dynamics that one empirically observes. This leads naturally to lter methods.
These methods determine the distribution, called lter distribution, of the unob-
served factors, given the available information. This distribution allows then to
compute the expectation of quantities that are dependent on unobserved factors,
for instance, derivative prices.
There is a growing literature in the area of ltering in nance. To mention a few
recent publications let us list Elliott & van der Hoek (1997), Fischer, Platen &
Runggaldier (1999), Elliott, Fischer & Platen (1999), Fischer & Platen (1999),
Landen (2000), Gombani & Runggaldier (2001), Frey & Runggaldier (2001),
Elliott & Platen (2001), Bhar, Chiarella & Runggaldier (2001a, 2001b) and
Chiarella, Pasquali & Runggaldier (2001). All these papers provide examples,
where lter methods have been applied in the area of nance. Such applications
involve optimal asset allocation, interest rate term structure modelling, estima-
tion of risk premia, volatility estimation and hedging under partial observation.
A key problem that arises in most ltering applications in nance is the deter-
mination of a suitable risk neutral equivalent martingale measure for the pricing
of derivatives. The resulting derivative prices and hedging strategies depend of-
ten signicantly on the chosen measure. On the other hand it is obvious that
in ltering one has to deal with the real world probability measure. It is there-
fore important to explore alternative methods that are based on the real world
measure and allow consistent derivative pricing.
In this paper we suggest a benchmark approach to ltering, where the bench-
mark portfolio is chosen as the growth optimal portfolio (GOP), see Long (1990)
and Platen (2002). The GOP has the important economic interpretation of be-
ing the portfolio that maximizes expected logarithmic utility. The dynamics
of the growth optimal portfolio depends on the degree of available information.
Given a certain information structure, one naturally obtains in this approach a
fair price system, where benchmarked prices equal their expected future bench-
marked prices. This avoids the involvement of a risk neutral equivalent martingale
measure. All resulting prices, when expressed in units of the GOP, turn out to
be local martingales under the given real world measure. In cases when bench-
marked prices are strict local martingales the benchmark approach generalizes
the standard risk neutral approach.
The paper is structured in the following way. It summarizes in Section 2 the
general ltering methodology for multi-factor jump diusion models with unob-
served factors. Section 3 describes the proposed ltered benchmark model. The

2
fair pricing and hedging of derivatives is then studied in Section 4. This section
also demonstrates how to quantify the reduction of the variance for derivative
prices using more information.

2 Filtered Multi-Factor Models

2.1 Factor Model
To build a nancial market model with a suciently rich structure and high com-
putational tractability we introduce a multi-factor model. This model provides
the basis for the dynamics of nancial quantities.
We consider a multi-factor model with n
2 factors z1
z2
: : :
zn , forming the
vector process
n ;  o
z = zt = zt1
: : :
ztk
ztk+1
: : :
ztn >
t 2 0
T ] : (2.1)
We shall assume that not all of the factors are observed. More precisely, only the
rst k factors are directly observed, while the remaining n ; k are not. Here k is
an integer with 1  k < n that we shall suppose to be xed during most of this
paper. However, in Section 4.4 we shall discuss the implications of a varying k.
For xed k we shall consider the following subvectors of zt
yt = (yt1
: : :
ytk )> = (zt1
: : :
ztk )> and xt = (x1t
: : :
xnt ;k )> = (ztk+1
: : :
ztn)>
(2.2)
with yt representing the observed and xt the unobserved factors. For instance,
yt may represent the vector of logarithms of the continuous and jump parts of
observed risky security prices.
Let there be given a ltered probability space (
AT
A
P ), where A = (At)t20
T ]
is a given ltration to which all the processes will be adapted. We assume that
the observed and unobserved factors satisfy the system of stochastic di
erential
equations (SDEs)
dxt = at (zt ) dt + bt (zt ) dwt + gt; (zt;) dmt
dyt = At (zt ) dt + Bt(yt) dvt + Gt;(yt;) dNt (2.3)
for t 2 0
T ] with given initial value z0 . Here
n ; o

wtn >


w = wt = wt
: : :1

wtk
wtk+1
: : : t 2 0
T ] (2.4)
is an n-dimensional (A
P )-Wiener process and
; >
vt = wt1
: : :
wtk (2.5)

3
is the subvector of its rst k components. The process m = fmt = (m1t
: : :
mkt ,
mkt +1
: : :
mnt)>, t 2 0
T ]g is an n-dimensional (A
P )-jump martingale dened
as follows: Consider n counting processes N 1
: : :
N n having no common jumps.
These are at time t 2 0
T ] characterized by the corresponding vector of intensi-
ties t (zt ) = (1t (zt )
: : :
nt(zt ))>, where for i 2 f1
2
: : :
kg
it(zt ) = ~it (yt): (2.6)
This means, we assume without loss of generality that the jump intensities of the
rst k counting processes are observed. Dene the ith jump martingale by
dmit = dNti ; it (zt ) dt (2.7)
for t 2 0
T ] and i 2 f1
2
: : :
ng. Let
; >
Nt = Nt1
: : :
Ntk (2.8)
be the vector of the rst k counting processes at time t 2 0
T ].
Concerning the coecients in the SDE (2.3), we assume that the vectors at (zt ),
At (zt), t (zt ) and the matrices bt (zt ), Bt(yt), gt(zt ) and Gt(yt) are such that a
unique strong solution of (2.3) exists that does not explode until time T , see
Protter (1990). We shall also assume that the k  k-matrix Bt (yt) is invertible for
all t 2 0
T ]. Finally, gt(zt ) may be any bounded function and the k  k-matrix
Gt(yt ) is assumed to be a given function of yt, invertible for each t 2 0
T ].
This latter assumption implies that, since there are no common jumps among
the components of Nt, by observing a jump of yt we can establish which of the
processes N i
i 2 f1
2
: : :
kg, has jumped.
In addition to the ltration A, which represents the complete information, we
shall also consider the subltration
A~k = (A~kt )t20
T ]  A
(2.9)
where A~kt = fys = (zs1
: : :
zsk )>
s  tg represents the observed information
at time t 2 0
T ]. Thus A~k provides the structure of the actually available
information in the market, which depends on the specication of the degree of
available information k.
We shall be interested in the conditional distribution of xt , given A~kt , that, ac-
cording to standard terminology we call the lter distribution at time t 2 0
T ].
There exist general lter equations for the dynamics described by the SDEs in
(2.3). It turns out that these are SDEs for the conditional expectations of inte-
grable functions of the unobserved factors xt , given A~kt . Notice that, in particular,
exp{ xt ] is, for given  2 <k and with { denoting the imaginary unit, a bounded
and thus integrable function of xt . Its conditional expectation leads therefore to
the conditional characteristic function of the distribution of xt , given A~kt . The

4
latter characterizes completely the entire lter distribution. Considering condi-
tional expectations of integrable functions of xt is thus not a restriction for the
identication of lter equations.
The general case of lter equations is beyond the scope of this paper. These are,
for instance, considered in Liptser & Shiryaev (1977). We assume that the SDEs
(2.3) are such that the corresponding lter distributions admit a representation
of the form
   ;  
P ztk+1  zk+1
: : :
ztn  zn  A~kt = Fz +1
:::
z zk+1
: : :
zn  t1
: : :
tq
k
t
n
t

(2.10)
for all t 2 0
T ]. This means, that we have a nite-dimensional lter, character-
ized by the lter state process
n ;  o
 = t = t1
: : :
tq >
t 2 0
T ]
(2.11)
which is an A~kt -adapted process with a certain dimension q
1. We shall denote
by z~tk the resulting (k + q)-vector of observables
; 
z~tk = yt1
: : :
ytk
t1
: : :
tq >
(2.12)
which consists of the observed factors and the components of the lter state
process. Furthermore, we assume that the lter state t satises an SDE of the
form
dt = Ct (~ztk ) dt + Dt;(~ztk; ) dyt (2.13)
with Ct() denoting a q-vector valued function and Dt () a (q  k)-matrix valued
function, t 2 0
T ].
There are various models of the type (2.3) that admit a nite-dimensional lter
with t satisfying an equation of the form (2.13). In the following two subsections
we recall two classical such models. These are the conditionally Gaussian model,
which leads to a generalized Kalman-lter and the nite-state jump model for x,
which is related to hidden Markov chain lters. Various combinations of these
models have nite-dimensional lters and can be readily applied in nance, as
demonstrated in the literature that we mentioned in the introduction.

Example 2.1 : Conditionally Gaussian Filter Model

Assume that in the system of SDEs (2.3) the functions at () and At () are linear
in the factors and that bt (zt ) bt is a deterministic function, while gt(zt )
Gt(yt ) 0. This means the model (2.3) takes the form
 
dxt = a0t + a1t xt + a2t yt dt + bt dwt
 
dyt = A0t + A1t xt + A2t yt dt + Bt (yt) dvt
(2.14)
5
for t 2 0
T ] with given deterministic initial values x0 and y0. Here a0t and A0t are
column vectors of dimensions (n ; k) and k respectively, and a1t , a2t , bt , A1t , A2t ,
Bt (yt) are matrices of appropriate dimensions. Recall that w is an n-dimensional
(A
P )-Wiener process and v the vector of its rst k components.
In this case the lter distribution is a Gaussian distribution with vector mean
t = ( 1t
: : :
(tn;k))>, where
  
it =E xit  A~kt (2.15)
and covariance matrix ct = c`
i
t ]`
i2f1
2
:::
n;kg, where
 
; ` ` ; i i  k
t = E xt ; t xt ; t  A~t :
c`
i (2.16)
The dependence of t and ct on k is for simplicity suppressed in our notation.
The above lter can be obtained from a generalization of the well-known Kalman
lter, see Chapter 10 in Liptser & Shiryaev (1977), namely
   
d t = a0t + a1t t + a2t yt dt + bt Bt (yt)> + ct (A1t )> (Bt(yt) Bt (yt)>);1
 ;  
 dyt ; A0t + A1t t + A2t yt dt

dct = a1t ct + ct (a1t )> + (bt b>t )
 ; ;1   o
1 > >

; bt Bt (yt) + ct (At ) Bt(yt ) Bt(yt)
> 1 > >  >
bt Bt (yt) + ct (At ) dt

(2.17)
where bt is the k-dimensional vector obtained from the rst k components of bt ,
t 2 0
T ]. We recall that Bt (yt) is assumed to be invertible.
Although for t 2 0
T ], ct is dened as a conditional expectation, it follows from
(2.17) that if Bt (yt) does not depend on the observable factors yt, then ct can
be computed o-line. Notice that the computation of ct is contingent upon the
knowledge of the coecients in the second equation of (2.17). These coecients
are given deterministic functions of time, except for Bt (yt) that depends also on
observed factors. The value of Bt(yt) becomes known only at time t, however, this
is sucient to determine the solution of (2.17) at time t. Model (2.14) is in fact
of the type of a conditionally Gaussian lter model, where the lter process  is
given by the vector process = f t
t 2 0
T ]g and the upper triangular array of
the elements of the matrix process c = fct
t 2 0
T ]g with q = (n ; k) 3+(n2;k)] .
Note by (2.16) that the matrix ct is symmetric. Obviously, in the case when
Bt (yt) does not depend on yt for all t 2 0
T ], then we have a Gaussian lter
model.

6
Example 2.2 : Finite-State Jump Model
Here we assume that the unobserved factors form a continuous time, (n ; k)-
dimensional jump process x = fxt = (x1t
: : :
xnt ;k )>
t 2 0
T ]g, which can take
a nite number M of values. More precisely, given an appropriate time t and>
zt -dependent matrix gt (zt ), and an intensity vector t(zt ) = (1t (zt )
: : :
nt(zt ))
at time t 2 0
T ] for the vector counting process N = fNt = (Nt1
: : :
Ntn)>,
t 2 0
T ]g, we consider the particular case of model (2.3), where in the xt -
dynamics we have at (zt ) = gt(zt )t (zt ) and bt (zt ) 0. Thus, by (2.3) and (2.7)
we have
dxt = gt;(zt; ) dNt (2.18)
for t 2 0
T ]. Notice that the process x of unobserved factors has here only
jumps and is therefore piecewise constant. On the other hand, for the vector yt
of observed factors we assume that it satises the second equation in (2.3) with
Gt(yt ) 0. This means that the process of observed factors y is only perturbed
by continuous noise and does not jump.
In this example, the lter distribution is completely characterized by the vector of
conditional probabilities pt = (p1t
: : :
pMt )>, where M is the number of possible
states
1
: : :

M of the vector xt and
  
pjt = P xt =
j  A~kt
(2.19)
j 2 f1
2
: : :
M g. Let ai
j h
t (y

) denote the transition kernel for x at time t to
jump from state i into state j given yt = y and xt =
h. The vector pt satises
the following dynamics
; j h i; 
dpt = a~t(yt
pt) pt dt + pt At (yt

) ; At(yt
pt ) Bt (yt) Bt(yt)> ;1
j > j j ~
h i
 dyt ; A~t(yt
pt) dt
(2.20)
see, Liptser & Shiryaev (1977), Chapter 9, where
M M !
;  X X
a~t (yt
pt )> pt j = ai
j h h
t (yt

) pt pit
i=1 h=1
At (yt

j ) = At (yt
xt )  x = t
j

M
X
A~t (yt
pt) = At(yt

j ) pjt (2.21)
j =1

7
for t 2 0
T ], j 2 f1
2
: : :
M g. The lter state process  = ft = (t1 : : :
tq )>,
t 2 0
T ]g for the model of this example is thus given by the vector process
p = fpt = (p1t
: : :
pqt )>, t 2 0
T ]g with q = M ; 1. Since the probabilities add
to one, we need only M ; 1 probabilities to compute.

2.2 Markovian Representation

As in the two previous examples we have, in general, in our lter setup to deal
with the quantity E (At (zt ) j A~kt) assuming that it exists. This is the conditional
expectation of At (zt ) = At (yt1
: : :
ytk
x1t
: : :
xtn;k ), given in (2.3), with respect
to the lter distribution at time t for the unobserved factors xt . Since the lter
is characterized by the lter state process  , we obtain the representation
  
A~t (~ztk ) = E At(zt )  A~kt
(2.22)
where the vector z~tk is as dened in (2.12).
Notice that, in the case of Example 2.1, namely the conditionally Gaussian model,
the expression A~t(~ztk ) takes the particular form
A~t (~ztk ) = A0t + A1t t + A2t yt: (2.23)
Furthermore, for Example 2.2, namely the nite-state jump model, A~t(~ztk ) can
be represented as
M
X
A~t (~ztk ) = A~t(yt
pt) = At(yt

j ) pjt (2.24)
j =1
for t 2 0
T ], see (2.21).
We have now the following generalization of Theorem 7.12 in Liptser & Shiryaev
(1977), which provides an important representation of the SDE for the observed
factors.
Proposition 2.3 Let At (zt ) and the invertible matrix Bt(yt ) in (2.3) be such
that
Z T Z T
E (jAt (zt )j) dt < 1 and Bt (yt) Bt(yt )> dt < 1 (2.25)
0 0

P -a.s. Then there exists a k-dimensional A~k -adapted Wiener process v~ = fv~t
t 2
0
T ]g such that the process y = fyt
t 2 0
T ]g of observed factors in (2.3)
satises the SDE
dyt = A~t (~ztk ) dt + Bt (yt) dv~t + Gt; (yt;) dNt (2.26)
with A~t (~ztk ) as in (2.22).
8
The proof of Proposition 2.3 is given in Appendix A.

Instead of the original factors zt = (yt1
: : :
ytk
x1t
: : :
xnt ;k )> = (zt1
: : :
ztn )>,
where xt = (x1t
: : :
xnt ;k )> is unobserved, we may now base our analysis on the
components of the vector z~tk = (yt1
: : :
ytk
t1
: : :
tq )>, see (2.12), that are all
observed. Just as was the case with z = fzt
t 2 0
T ]g, also the vector process
z~k = fz~tk
t 2 0
T ]g has a Markovian dynamics. In fact, replacing dyt in (2.13)
by its expression resulting from (2.26), we obtain
h i
dt = Ct(~ztk ) + Dt(~ztk ) A~t(~ztk ) dt + Dt (~ztk ) Bt(yt) dv~t + Dt;(~ztk; ) Gt;(yt;) dNt
= C~t (~ztk ) dt + D~ t(~ztk ) dv~t + G~ t;(~ztk; ) dNt
(2.27)
whereby we implicitly dene the vector C~t(~ztk ) and the matrices D~ t(~ztk ) and G~ t (~ztk )
for compact notation.
From equations (2.26) and (2.27) we immediately obtain the following result.
Corollary 2.4 The dynamics of the vector z~tk = (yt
t ) can be expressed by
the system of SDEs
dyt = A~t (~ztk ) dt + Bt(yt) dv~t + Gt;(yt;) dNt
dt = C~t(~ztk ) dt + D~ t (~ztk ) dv~t + G~ t; (~ztk; ) dNt: (2.28)
From Corollary 2.4 it follows that the process z~k = fz~tk
t 2 0
T ]g is Markov.
Due to the existence of a Markovian lter dynamics we have our original Marko-
vian factor model, given by (2.3), projected into a Markovian model for the
observed quantities. Here the driving observable noise v~ is an (A~k
P )-Wiener
process and the observable counting process N is generated by the rst k com-
ponents N 1
N 2
: : :
N k of the n counting processes.
For ecient notation we write for the vector of observables z~tk = zt = (zt1
zt2

: : :
ztk+q )> the corresponding system of SDEs in the form
Xk
dzt` = `(t
zt1
zt2
: : : k +q

zt ) dt +  `
r (t
zt1
zt2
: : :
ztk+q ) dv~tr
r=1
k
X  
+ `
r t;
zt;
zt;
: : :
1 2

ztk;+q dNtr (2.29)
r=1
for t 2 0
T ] and ` 2 f1
2
: : :
k + qg. The functions, `,  `
r and `
r follow
directly from A~, B , G, C~ , D~ and G~ appearing in (2.28).
We also have as an immediate consequence of the Markovianity of z~k = z, as well
as property (2.10), the following result.
9
Corollary 2.5 Any expectation of the form E (u(t
zt ) j A~kt ) < 1 for a given
function u : 0
T ]  <n ! < and given k 2 f1
2
: : :
n ; 1g can be expressed as
  
E u(t
zt )  A~kt = u~k (t
z~tk ) (2.30)
with a suitable function u~k : 0
T ]  <k+q ! <.

Relation (2.30) in Corollary 2.5 will be of importance for contingent claim pricing
as we shall see later on.

3 Filtered Benchmark Model

On the basis of the Markovian dynamics for the prices, generated by the observed
factors introduced above, we formulate a ltered benchmark model. As described
in Platen (2002), we model the dierent denominations of the growth optimal
portfolio (GOP), see Long (1990). We only use the observed factors to model
the GOP. However, these factors evolve in conjunction with unobserved factors
that inuence the observed ones. The resulting ltered benchmark model has the
key advantage that a consistent price system is automatically established without
using any measure transformation.

3.1 Primary Security Accounts

We assume that there are d+1 primary assets in the market, where d = 2 k. These
are, for instance, currencies or shares. For the domestic currency as primary
asset we express the time evolution of its value by the savings account process
B 0 = fB 0(t)
t 2 0
T ]g. We call B 0 also the 0th primary security account
process.
For the modelling of the time value of the j th primary asset, j 2 f1
2
: : :
dg,
we introduce the j th primary security account process S j = fS j (t)
t 2 0
T ]g.
For instance, in the case of currencies, S j (t) is the value of the savings account of
the j th foreign currency, expressed in units of the domestic currency. If the j th
asset is a share, then S j (t) is the cum-dividend share price, where all dividend
payments are reinvested. We then denote by Sj (t) the discounted value at time
t of the j th primary security account, that is
Sj (t) = BS 0((tt))
j
(3.1)

for t 2 0
T ] and j 2 f0
1
: : :
dg. We assume that Sj is A~k -adapted and the

10
unique strong solution of the stochastic dierential equation (SDE)
k 
X
dSj (t) = Sj (t;) (0
r (t) ; j
r (t)) (0
r (t) dt + dv~tr )
r=1
 j
r  
' (t ; ) ~
' (t;) ; 1 ;' (t) t (yt) dt + dNt
0
r r r
+ 0
r (3.2)

for t 2 0
T ] with Sj (0) > 0, j 2 f1
2
: : :
dg.

We assume that j
r and 'j
r are A~k -predictable with 'j
r (t)
0, '0
r (t) > 0 a.s.
for t 2 0
T ] and
Z T 
(j
r (s))2 + ~rs(ys) ds < 1
0

for j 2 f1
2
: : :
dg and r 2 f1
2
: : :
kg. The given parameterization of the

above SDE (3.2) does not restrict its generality but is convenient for the bench-
mark approach.

3.2 Portfolios
Let us now form portfolios of primary security accounts. We say that an A~k -
predictable stochastic process  = f(t) = (0 (t)
: : :
d(t))>, t 2 0
T ]g is a
self-nancing strategy, if  is S-integrable, see Protter (1990), the corresponding
portfolio V0(t) has at time t the discounted value
d
V (t) = j (t) Sj (t)
V0(t) = B
0 X
(3.3)
0 (t)
j =0
and it is
d
X
dV0 (t) = j (t;) dSj (t) (3.4)
j =0

for all t 2 0
T ]. The j th component j (t), j 2 f0
1
: : :
dg, of the self-nancing
strategy  expresses the number of units of the j th primary security account
held at time t in the corresponding portfolio. Under a self-nancing strategy no
outow or inow of funds occurs for the corresponding portfolio. All changes in
the value of the portfolio are due to gains from trade in the primary security
accounts.
We assume that no primary security account is redundant. That means, no pri-
mary security account can be expressed as a self-nancing portfolio of other pri-

11
mary security accounts. Let us set
8
>
< 0
r (t) ; j
r (t) for r 2 f1
2
: : :
kg
bj
r (t) =  q (3.5)
>
: ' ; (t)
j
r k
;1 ~rt ;k (yt) for r 2 fk + 1
: : :
dg
'0 ; (t)

r k

for t 2 0
T ] and j 2 f1
2
: : :
dg. We then dene the matrix b(t) = bj
r (t)]dj
r=1
for t 2 0
T ] and assume that b(t) is for Lebesgue-almost-every t 2 0
T ] invert-
ible. Note that the observed market is complete, which means that the observed
primary security accounts securitize the uncertainty generated by the Wiener
processes v~1
: : :
v~k and the counting processes N 1
: : :
N k .

3.3 Growth Optimal Portfolio

The GOP is the self-nancing portfolio that achieves maximum expected loga-
rithmic utility. We denote by Vi (t) the value of the GOP when it is expressed at
time t in units of the ith primary security account. For the diusion case without
jumps the corresponding SDE is well known, see, for instance, Long (1990) or
Karatzas & Shreve (1998). In the case with jumps the derivation of the SDE for
the GOP is more involved and described in Platen (2002). It has the form
k   
dVi(t) = Vi (t;)
X
i
r (t) (i
r (t) + dv~r (t)) + 1 ;1
r=1 'i
r (t;)
 o
 ;'i
r (t) ~ rt (yt) dt + dNtr (3.6)
for t 2 0
T ] and i 2 f0
1
: : :
dg.
To make the above framework computationally tractable, given our factor model,
we specify Vi(t) as a function of time t and the vector of observables z~tk , that is
Vi(t) = Vi (t
z~tk ) = Vi(t
zt1
: : :
ztk+q ) (3.7)
for t 2 0
T ] and i 2 f0
1
: : :
dg. Assuming sucient smoothness of Vi(
), by
application of the It^o formula and using (2.29), we then obtain
Z t k Z t
X
Vi(t
z~tk ) = Vi(0
z~0k ) + L Vi(s
z~sk ) ds +
0
Lr Vi(s
z~sk ) dv~sr
0 r=1 0

k Z t
X
+ rV (s;
z~sk;) dNsr
i (3.8)
r=1

0

12
for t 2 0
T ] and i 2 f0
1
: : :
dg. Here we use the operators
k+q 
@ +X
L0 = @t ` t
zt1
: : :
ztk+q @@z`


`=1
k+q X
k
 `
r t
zt1
: : :
ztk+q  p
r t
zt1
: : :
ztk+q @ z@` @ zp
(3.9)
   
+ 12
X 2

`
p=1 r=1

k+q
 `
r t
zt1
: : :
ztk+q @@z`
X  
Lr = (3.10)
`=1
and
rF (t;
z~tk; ) =
    
F t
zt1; + 1
r t;
zt1;
: : :
ztk;+q
: : :
ztk;+q + k+q
r t;
zt1;
: : :
ztk;+q
 
; F t;
zt1;
: : :
ztk;+q (3.11)
for t 2 0
T ], r 2 f1
2
: : :
kg and F : 0
T ]  <k+q ! < being any given
function of time and vector of observables, where z~tk = zt = (zt1
: : :
ztk+q )> as
introduced before (2.29).
By comparison of (3.6) and (3.8) it follows that the i
`th GOP-volatility i
`(t)
has the form
L` V i(t
z~k )
i
`(t) =  i  k t (3.12)
V (t
z~t )
and the inverted i
`th GOP-jump ratio 'i
`(t), see (3.6) is given by the expression
Vi(t;
z~tk; )
' (t;) = ` (t;
z~k ) + V i(t;
z~k )
i
` (3.13)
V t; 
i t;

for t 2 0
T ], i 2 f0
1
: : :
dg, ` 2 f1
2
: : :
kg.

4 Fair Pricing and Hedging of Derivatives

4.1 Benchmarked Prices
In what follows we call prices that are expressed in units of the GOP, benchmarked
prices. This means for j 2 f0
1
: : :
dg that the j th benchmarked primary
13
security account S^j = fS^j (t)
t 2 0
T ]g with
j
S^j (t) = VS 0((tt)) = S 0(t)
j
(4.1)
 V (t)
satises by (3.2), (3.5), (3.6) and application of the It^o formula the SDE
k 
X 
dS^j (t) = S^j (t;) ;j
r (t) dv~tr + ('j
r (t;) ; 1) dmrt (4.2)
r=1
for t 2 0
T ] and j 2 f0
1
: : :
dg, see Platen (2002). Here mrt denotes the rth
component of the jump martingale m dened in (2.7). Note that the j th bench-
marked primary security account is an (A~k
P )-local martingale. Moreover, as
shown in Platen (2002), for any self-nancing portfolio V0 it 0follows 0by applica-
tion of the It^o formula that its benchmarked value V^ (t) = VV 0((tt)) = VV 0 ((tt)) satises

 

the SDE
 

k ( d
X X
dV^ (t) = V^ (t;) ; j (t) j
r (t) dv~r (t)
r=1 j =0
! )
Xd
+ j (t;) 'j
r (t;) ; 1 dmrt (4.3)
j =0

for t 2 0
T ]. This shows that V^ is an (A~k
P )-local martingale too. Note that
these processes are, in general, not (A~k
P )-martingales. Since a nonnegative
benchmarked portfolio process is here an (A~k
P )-supermartingale, the result-
ing ltered benchmark model can be shown to exclude standard arbitrage. This
means, it is impossible to generate, with strictly positive probability, strictly
positive wealth from zero initial capital.

4.2 Derivative Prices

To provide an intuitive link between the benchmark framework and the standard
risk neutral approach, let us discuss a situation where we assume for the moment
that the following steps can be made and a standard equivalent risk neutral
probability measure P k exists. We underline that such assumptions will not
be needed for our results. Then all prices, discounted by the domestic savings
account B 0 would be (A~k
P k)-martingales. Denoting by E the expectation with
respect to P and by E k that with respect to P k , we would have, taking into

14
account (4.1), that
  
S j (t) = Ek B 0
(t) S j (T )  A~k
B 0 (T )  t
 k  
 B 0
( t)
= E k 0 S (T )  A~kt
T j 
t B (T )
 !
= V0 (t) E VS 0((TT ))  A~kt
j 
(4.4)


for t 2 0
T ] and j 2 f1
2
: : :
dg. Here the Radon-Nikodym derivative kT = dPdP k

would satisfy the expression

k V0(0) B 0 (t) S^0(t)
t = V 0(t) B 0(0) = ^0 (4.5)
 S (0)
for t 2 0
T ]. Furthermore, by (4.5), the price of a self-nancing portfolio V0
would satisfy the relation
 
B 0
(t ) 
 ~k
V (t) = Ek V (  )  At
0 0
B ( )
0 
 k 0  
 B ( t )
= E k B 0 ( ) V ( )  At
 0
 k
 ~
t
 !
V 0
= V0(t) E 0  A~kt (  ) 
(4.6)
V ( )
for t 2 0
 ] and any A~k -stopping time 0 . Thus, under the above assumptions all
benchmarked portfolio prices V^ (t) = VV 0 ((tt)) would be (A~k
P )-martingales, that is


  
V^ (t) = E V^ ( )  A~kt (4.7)
for all t 2 0
 ].
In the benchmark framework we avoid the above steps and the assumption on
the existence of an equivalent risk neutral measure by introducing the concept
of a fair price. A price process is called fair, if its benchmarked values form an
(A~k
P )-martingale, see Platen (2002).
At a given maturity date  , which is assumed to be an A~k -stopping time, we
consider a benchmarked contingent claim U (
y ) as a function of  and the
corresponding values of observed factors y , where we assume that

E (jU (
y )j  A~kt ) < 1 (4.8)

15
a.s. for all t 2 0
 ]. There is no point to let the payo function depend on
any other than observed factors, otherwise the payo would not be veriable at
time  . The benchmarked fair price process u~k = fu~k (t
z~tk )
t 2 0
 ]g for the
benchmarked contingent claim U (
y ) is then the (A~k
P )-martingale, obtained
by the conditional expectation
  
u~k (t
z~tk ) = E U (
y )  A~kt (4.9)
for t 2 0
 ]. This means, we form directly the conditional expectation (4.7)
without using any measure transformation. The corresponding fair price at time
t for this contingent claim, when expressed in units of the domestic currency, is
then
u~0
k (t
z~tk ) = V0 (t) u~k(t
z~tk ) (4.10)
for t 2 0
 ). The above concept of fair pricing generalizes the well-known concept
of risk neutral pricing and avoids not only the assumption on the existence of an
equivalent risk neutral measure but also some issues that arise from measure
changes under dierent ltrations.
The vector of observables y is a subvector, not only of z~k but also of z . This al-
lows us to dene the (A
P )-martingale u = fu(t
zt )
t 2 0
 ]g by the conditional
expectation
;  
u(t
zt) = E U (
y )  At (4.11)
for t 2 0
 ], which at time t exploits the complete information characterized
by the -algebra At. The above derivation can be summarized in the following
result.

Corollary 4.1 The benchmarked fair price u~k (t
z~tk ) for the benchmarked con-
tingent claim U (
y ) can be expressed as the conditional expectation
  
u~k (t
z~tk ) = E u(t
zt)  A~kt (4.12)
for t 2 0
 ].

We recall that A~kt denotes in Corollary 4.1 the information, which is available
at time t, whereas At is the complete information at time t that determines the
original model dynamics including also the unobserved factors.
Note that the benchmarked fair price, given in Corollary 4.1, ts perfectly the
expression of our result for the ltered factor model given in (2.30). The advan-
tage of the representation (4.12) is that it allows us to express the benchmarked
fair price u~k (t
z~tk ) as conditional expectation with respect to A~kt . The actual
computation of the conditional expectation in (4.12) is equivalent to the solution
of the ltering problem for the unobserved factors.
16
4.3 Hedging Strategy
Assume that the above benchmarked pricing function u~k (
) in (4.9) and (4.12)
is dierentiable with respect to time and twice dierentiable with respect to the
observables. Then we obtain by the It^o formula the representation
U (
y ) = u~k (
z~k )
k Z 
u~k (s
z~sk ) Lu~ku~(s
(s
z~kz~)s ) dv~s`
X ` k k
= u~k (t
z~tk ) +
`=1 t s
k Z 
X ` (s;
z~k )
+ u~k (s;
z~sk;) u~u~k (s;
z~ks;) dm`s (4.13) k

`=1 t s ;
for t 2 0
 ]. Let us search for a fair benchmarked price process V^ , with self-
nancing hedging strategy U , that possibly matches u~k . This means, we consider
U

V^ (t) with

U

d
X
dV^ (t) =
U
Uj (t;) dS^j (t) (4.14)
j =0
for t 2 0
 ]. By (4.3) we then have
k Z 
X d
X
V^ ( ) = V^ (t) ;
U U
V^ (s)
U
j (s) j
`(s) dv~s`
`=1 t
U
j =0
k Z  d !
X X
+ V^ (s;)
U
j (s;)'j
r (s;) ; 1 dm`s: (4.15)
`=1 t j =0
U

Note that the volatilities and jump ratios in (4.15) are those identied in (3.12)
and (3.13). Above we used the j th proportion
 j (t) S^j (t)
j
 (t) = ^ U (4.16)
U
V (t) U

of the value of the corresponding hedging portfolio that has to be invested into
the j th primary security account at time t 2 0
 ]. To replicate the benchmarked
contingent claim U (
y ) we can start at a given time t 2 0
 ] by forming a
portfolio with fair benchmarked price
  
V^ (t) = u~k (t
z~tk ) = E U (
y )  A~kt :
U
(4.17)
By comparison of (4.13) and (4.15) the proportions must satisfy the system of
linear equations
` u~k (t
z~k ) X d
L
; u~k (t
z~k ) = j (t) j
`(t)
t (4.18)
t
U

j =0

17
and
`u~ (t;
z~tk; ) X d
k
+ 1 = j (t;) 'j
`(t;) (4.19)
u~ (t;
z~t; )
k k
j =0
for ` 2 f1
2
: : :
kg and t 2 0
T ]. Let us use the d-dimensional vector c(t;) =
(c1(t;)
c2 (t;)
: : :
cd(t;))> with components
8
Lr u~k (t;
z~tk; )
>
>
<
0
r
u~k (t;
z~tk; ) +  (t) for r 2 f1
2
: : :
kg
cru~k (t;) =   r;k
k (t;
z~tk; )
 q
>
: '0
` (t;) u~u~k (t;
z~k ) + 1
> 1
t;
; 1 ~rt ;k (yt;) for r 2 fk + 1
: : :
dg

(4.20)
t 2 0
 ). By involving the matrix b(t) given in (3.5), we can then rewrite the
system of equations (4.18) - (4.19) in the form
cu~ (t;)> =  (t;)> b(t;)
k
U
(4.21)
for t 2 0
T ]. Now, we obtain the following result.

Proposition 4.2 For a given benchmarked contingent claim U (
y ) with cor-
responding vector cu~ (t) given in (4.20) the proportions of the corresponding hedg-
k

ing portfolio are of the form

; >
 (t;) = cu~ (t;)> b;1 (t;)
U
k (4.22)
for t 2 0
 ).

Note that the invertibility of the matrix b(t) is not linked to a specic contingent
claim. Thus, one can form a perfectly replicating hedging portfolio for all bench-
marked contingent claims U (
y ). The introduced ltered benchmark model
forms a complete market despite the fact that the original model involves unob-
served factors. The benchmarked pricing functions can always be obtained from
the conditional expectation (4.12) on the basis of the lter distribution.
We did not consider the case d < 2k. In such a case the market is incomplete.
Incomplete markets of this type can be handled by a generalization of the above
described ltered benchmark approach.

4.4 Variance of Benchmarked Prices

Let us now investigate the impact of varying degrees of information concerning
the factors zt = (zt1
: : :
ztn)> that underly our model dynamics, see (2.2) - (2.3).
As already mentioned in Section 2.1, the degree of available information is indexed
18
by the parameter k. A larger value of k means that more factors are observed,
providing thus more information in A~k . Again we use the notation z~tk for the
vector of observables dened in (2.12), where we stress its dependence on k and
recall that, by (2.28), the process z~k is Markovian.
Consider from now on a benchmarked contingent claim
U (
y ) = U (
y1
y2
: : :
yr ) (4.23)
for some xed r 2 f1
2
: : :
n ; 1g, where we assume that the number of ob-
served factors that inuence the claim equals r. For k 2 fr
r + 1
: : :
n ; 1g let
u~k (t
z~tk ) be the corresponding benchmarked fair price under the information A~kt ,
as given by (4.12). Recall by (2.30) that u~k (t
z~tk ) can be computed as conditional
expectation via the lter distribution. Then
; 2  k 
Vart (u) = E u(t
zt ) ; u~ (t
z~t )  A~t
k k k (4.24)
is the corresponding conditional variance at time t 2 0
 ). Note that for larger k
we have more information available, which naturally should reduce the conditional
variance.
For each degree of available information one obtains, in general, dierent equiva-
lent risk neutral probability measures. The complexity of working with dierent
pricing measures can be signicant. This is avoided by using the suggested l-
tered benchmark model. All conditional expectations can be taken under the
real world probability measure P . Furthermore, it is clear that ltering itself is
always performed under the real world measure.
We can prove the following proposition, which expresses the reduction in condi-
tional variance and can also be seen as a generalization of the celebrated Rao-
Blackwell theorem towards ltering.
Proposition 4.3 For m 2 f0
1
: : :
n ; kg and k 2 fr
r + 1
: : :
n ; 1g we
have
  
E Varkt +m(u)  A~kt = Varkt (u) ; Rtk+m
(4.25)
where
; 2  
Rtk+m = E u~k+m(t
z~tk+m ) ; u~k (t
z~tk )  A~kt (4.26)
for t 2 0
 ).
Proof: For t 2 0
 ) and k 2 fr
r + 1
: : :
n ; 1g we have
;  ;  ; 
u(t
zt ) ; u~k (t
z~tk ) 2 = u(t
zt) ; u~k+m(t
z~tk+m ) 2 + u~k+m(t
z~tk+m ) ; u~k (t
z~tk ) 2
; ; 
+ 2 u(t
zt) ; u~k+m(t
z~tk+m ) u~k+m(t
z~tk+m ) ; u~k (t
z~tk ) :
(4.27)
19
By taking conditional expectations with respect to A~kt on both sides of the above
equation it follows that

   ;  
Varkt (u) = E Varkt +m(u)  A~kt + Rtk+m + 2 E u~k+m(t
z~tk+m ) ; u~k (t
z~tk )
;   k+m   k 
E u(t
zt) ; u~ (t
z~t )  A~t
k +m k +m
 A~t : (4.28)
Since the last term on the right hand side is equal to zero by denition, we obtain
(4.25).

5 Conclusions
We constructed a ltered benchmark model by specifying the growth optimal
portfolio for a given degree of available information. A consistent price system has
been established. Benchmarked fair derivative prices are obtained as martingales
under the real world probability measure. In general, benchmarked security prices
are not forced to be martingales. They may be just local martingales. The
reduction of the conditional variance of fair derivative prices under increased
information is quantied via a generalization of the Rao-Blackwell theorem.

A Appendix
Proof of Proposition 2.3
Denote by yc the continuous part of the observation process y, that is
X
ytc = yt ; G ;(y ;)  N

j j j
(A.1)
 t
j

where the j denote the jump times of N = fNt
t 2 0
T ]g and  N = N ;

N ; is the vector ( N1 ;
: : :
 Nk ;)>. Let us now dene the k-dimensional
j j

A~k -adapted process v~ = fv~t
t 2 0
T ]g by

j j

Bt (yt) dv~t = dytc ; A~t (~ztk ) dt: (A.2)

From (2.3), (A.1) and (A.2) it follows that
h i
dv~t = dvt + Bt(yt );1 At (zt ) ; A~t (~ztk ) dt: (A.3)

20
From this we nd, by the multi-variate It^o formula with  2 <k a row vector and
{ the imaginary unit, that
Z t
exp { (~vt ; v~s)] = 1 + { exp { (~vu ; v~s)] dvu
s
Z t  
+ { exp { (~vu ; v~s)] Bu;1(yu) Au(zu) ; A~u(~zuk ) du
s
  > Z t
; 2 exp { (~vu ; v~s)] du: (A.4)
s
Recalling that v is an A~k -measurable Wiener process, notice that
Z t 

E exp { (~vu ; v~s)] dvu  A~ks =0 (A.5)
s
and that, by our assumptions and by the boundedness of exp { (~vu ; v~s)],
Z t   

E exp { (~vu ; v~ )] B ;1 (y
s u u) Au(zu ) ; A~u(~zuk ) du  A~ks =
s
Z t     
E exp { (~vu ; v~ )] B ;1(y
s u u)E Au(zu) ; A~u(~zuk )  A~u du  A~ks = 0:
s
(A.6)
Taking conditional expectations on the left and the right hand sides of (A.4) we
end up with the equation
  k   0Z t   k
~
E exp ({ (~vt ; v~s)]) As = 1 ; 2
 E exp { (~vu ; v~s)]  A~s du
(A.7)
s
which has the solution
 > 
  k  
E exp { (~vt ; v~s)]  A~s = exp ; 2 (t ; s) (A.8)
for 0  s  t  T . We can conclude that (~vt ; v~s) is a k-dimensional vector
of independent A~kt -measurable Gaussian random variables, each with variance
(t ; s) and independent of A~ks . By Levy's theorem, v~ is thus a k-dimensional
A~k -adapted standard Wiener process.

References
Bhar, R., C. Chiarella, & W. Runggaldier (2001a). Estimation in models of the
instantaneous short term interest rate by use of a dynamic Baysian algo-
rithm. Technical report, University of Technology, Sydney. QFRG Research
Paper 68.
21
Bhar, R., C. Chiarella, & W. Runggaldier (2001b). Filtering equity risk premia
from derivative prices. Technical report, University of Technology, Sydney.
QFRG Research Paper 69.
Chiarella, C., S. Pasquali, & W. Runggaldier (2001). On ltering in Marko-
vian term structure models (an approximation approach). Adv. in Appl.
Probab. 33(4), 794{809.
Elliott, R. J., P. Fischer, & E. Platen (1999). Hidden Markov ltering for a
mean reverting interest rate model. In 38th IEEE CDC meeting in Phoenix,
pp. 2782{2787. IEEE Control Systems Society.
Elliott, R. J. & E. Platen (2001). Hidden Markov chain ltering for generalised
Bessel processes. In Stochastics in Finite and Innite Dimensions, pp. 123{
143. Birkh"auser.
Elliott, R. J. & J. van der Hoek (1997). An application of hidden Markov
models to asset allocation problems. Finance and Stochastics 1, 229{238.
Fischer, P. & E. Platen (1999). Applications of the balanced method to stochas-
tic dierential equations in ltering. Monte Carlo Methods Appl. 5(1), 19{
38.
Fischer, P., E. Platen, & W. Runggaldier (1999). Risk minimizing hedging
strategies under partial observation. In Seminar on Stochastic Analysis,
Random Fields and Applications, Volume 45 of Progr. Probab., pp. 175{
188. Birkh"auser.
Frey, R. & W. J. Runggaldier (2001). A nonlinear ltering approach to volatility
estimation with a view towards high frequency data. Int. J. Theor. Appl.
Finance 4(2), 199{210.
Gombani, A. & W. J. Runggaldier (2001). A ltering approach to pricing in
multifactor term structure models. Int. J. Theor. Appl. Finance 4(2), 303{
320.
Karatzas, I. & S. E. Shreve (1998). Methods of Mathematical Finance, Vol-
ume 39 of Appl. Math. Springer.
Landen, C. (2000). Bond pricing in a hidden Markov model of the short rate.
Finance and Stochastics 4, 371{385.
Liptser, R. & A. Shiryaev (1977). Statistics of Random Processes: I. General
Theory, Volume 5 of Appl. Math. Springer.
Long, J. B. (1990). The numeraire portfolio. J. Financial Economics 26, 29{69.
Platen, E. (2002). Benchmark models with intensity based jumps. University
of Technology Sydney, (working paper).
Protter, P. (1990). Stochastic Integration and Di
erential Equations. Springer.

22