A Study towards a Uni ed Approach to the Joint Estimation of Objective and Risk Neutral

Measures for the Purpose of Options Valuation1

Mikhail Chernov and Eric Ghysels

First Draft: October 1997

This Revision: April 22, 1999

1 The rst author is with the Department of Finance, The Pennsylvania State University, 609 Business Administration
Building 1, University Park, PA 16802-3006, phone: 814-863-0486, fax: 814-865-3362, e-mail: mchernov@psu.edu. The second
author is with the Departments of Economics and Finance, The Pennsylvania State University, 523 Kern Graduate Building,
University Park, PA 16802-3305, phone: 814-865-7924, fax: 814-863-4775, e-mail: eghysels@psu.edu. We would like to
thank Torben Andersen, Mark Broadie, Stephen Brown, Charles Cao, Jer^ome Detemple, Jens Jackwerth, Eric Jacquier, Frank
Hatheway, Roger Lee and especially Ron Gallant and George Tauchen for invaluable comments and insightful discussions. We
also bene ted greatly from the comments of Bill Schwert (the Editor) and the Referees who helped us improve the quality
of our paper. In addition, we thank the participants of the NBER/NSF Time Series Conference held at the University of
Chicago, the RISK Computational and Quantitative Finance conference, the Newton Institute Workshop on Econometrics and
Financial Time Series, the Third Annual New England Finance Doctoral Students Symposium and Computational Finance 99
conference both held at NYU, the 9th Derivative Securities Conference at Boston University and the 1999 Western Finance
Association Meetings as well as seminars at CIRANO, Michigan State, Penn State, Universite Libre de Bruxelles, the University
of Montreal, the University of Michigan and the University of Virginia. We are solely responsible for all remaining errors. The
material in this paper subsumes two earlier papers: \What Data Should be Used to Price Options?" and \Filtering Volatility
and Pricing Options: A Comparison of Implied Volatility, GARCH and Stochastic Volatility".
 Abstract
The purpose of this paper is to bridge two strands of the literature, one pertaining to the objective or physical
measure used to model the underlying asset and the other pertaining to the risk-neutral measure used to price
derivatives. We propose a generic procedure using simultaneously the fundamental price St and a set of option
contracts [(itI )i=1;m ] where m  1 and itI is the Black-Scholes implied volatility. We use Heston's (1993) model
as an example and appraise univariate and multivariate estimation of the model in terms of pricing and hedging
performance. Our results, based on the S&P 500 index contract, show that the univariate approach only involving
options by and large dominates. A by-product of this nding is that we uncover a remarkably simple volatility
extraction lter based on a polynomial lag structure of implied volatilities. The bivariate approach involving both
the fundamental and an option appears useful when the information from the cash market provides support via
the conditional kurtosis to price options. This is the case for some long term options.
JEL classi cation: G13; C14, C52, C53
Key Words: Derivative securities, Ecient Method of Moments, State Price Densities, Stochastic Volatil-
ity Models, Filtering
Introduction
According to modern asset pricing theory the value of any asset can be computed as the expectation under
the risk-neutral measure of the future cash ows discounted by the pricing kernel. The valuation of any
contingent claim, like a European style option for instance, consists of specifying the pricing kernel and
determining the appropriate risk-neutral measure transformation. These operations involve several critical
steps. First and foremost, it should be noted that there is no absolute \model-free" way to proceed.1 In
particular, the characterization of the risk-neutral measure is intimately related to the price of market risk
which in turn is determined by the model one adopts to describe the behavior of the fundamental asset
underlying the option contract. This step is also inextricably linked to the estimation of parameters which
select the data generating process among the class of models considered. There are di erent sources of data
one could use for the purpose of estimating or calibrating parameters and there is certainly an abundant
choice as there are many option contracts actively traded and long time series of the fundamental typically
available. To further complicate matters it should be noted that many models describing the behavior of the
fundamental process feature latent factors such as stochastic volatility. Therefore one faces also the task of
using observations to not only estimate parameters but also to lter or extract the unobservable factors.
In recent years we have made considerable progress on various aspects of this research program. We know
more about estimating di usions particularly those involving stochastic volatility or other latent factors.2
Parallel to this we witnessed the emergence of several studies suggesting schemes to extract risk-neutral
measures from option prices (see inter alia references in footnote 2). Considerable e ort was also devoted to
lters for extracting the latent factors like volatility. These lters either involve options data or underlying
fundamentals (but not both jointly).
The purpose of this paper is to bridge two strands of the literature, one pertaining to the objective or
physical measure used to model the underlying asset and the other to the risk-neutral measure used to price
derivatives. In fact we start rst with estimating both measures jointly. This poses several challenges as
building a bridge between the objective and risk-neutral world prompts us to think about many new issues
which need to be addressed. In addition, it also opens up new possibilities for comparing the information in
the underlying fundamental and options data, a theme which has been the subject of many previous studies.
1 Even so called nonparametric approaches (see for instance At-Sahalia and Lo (1998) and Broadie et al. (1998)) either im-
plicitly restrict the class of models by imposing regularity conditions to guarantee valid statistical inference (for more discussion
on this point see for instance Ghysels et al. (1998)) or assume an explicit class of models (see e.g. Rubinstein (1994)).
2 Numerous techniques have been proposed for the estimation of continuous time processes pertaining to the pricing of

derivative securities. The literature on the estimation of di usions with or without stochastic volatility and/or jumps, is
summarized in a number of surveys and textbooks, including Bates (1996), Campbell et al. (1996), Ghysels et al. (1996),
Melino (1994), Renault (1997) and Tauchen (1997).

1
Numerous papers have confronted empirical evidence obtained from derivative security markets with results
from the underlying and vice versa. In particular, issues related to the informational content of option prices
have been examined extensively in the literature (see for instance Christensen and Prabhala (1998) for the
most recent example). Our attempt to model the price behavior of fundamental and derivative securities
jointly is motivated by the very same issues hitherto raised in the literature. Namely, we want to learn more
about the informational content of option prices. We also want to know how we can improve the statistical
precision of di usion parameters by incorporating options.
Our goal is to investigate these questions in a unifying framework. While we use the Heston (1993) as
speci c example, it should be stressed at the outset that our analysis is not limited to any particular model.
The choice of Heston's model is motivated by two important factors: (i) it has closed-form option pricing
formula which represents some considerable computational advantages, and (ii) because of the existence of
analytic solutions it has received much attention in the literature (see for instance Bakshi et al. (1997a) for
references) which makes our analysis directly comparable with results previously reported in the literature.
Financial theory also suggests that for stochastic volatility models with two state variables (such as
the models of Hull and White (1987), Scott (1987), Wiggins (1987), Heston (1993) and many others) one
should consider the fundamental and its derivative contracts jointly to estimate di usion parameters and
price options simultaneously. There are indeed appealing theoretical reasons to pursue this approach, as in
a stochastic volatility economy we need to add for instance options to be able to complete the market.3 The
complete market setup guarantees the existence and uniqueness of the risk-neutral probability density used
to price the option contracts. If done judiciously this challenging task should dominate the use of a single
source, whether it is options or fundamental.
Although to the best of our knowledge no attempts were made to estimate and appraise stochastic
volatility models using the joint distribution of fundamentals and options, it is clear that much of the
evidence in the literature suggests that we should gain from addressing this issue.4 In this paper we propose
a generic procedure for estimating and pricing options using simultaneously the fundamental price St and a
set of option contracts [(itI )i=1;m ] where m  1 and itI is the Black-Scholes implied volatility. Please note
that we can in principle deal with a panel of options, i.e. a time series of cross-sections. The procedure we
propose consists of two steps, rst we t a seminonparametric (henceforth SNP) density of [St ; (itI )i=1;m ]
conditional on its own past [S ; (iI )i=1;m ] for  < t using market data. Next we simulate the fundamental
3 See for instance, Romano and Touzi (1993) who show that the combination of the fundamental and a European-type option
contract complete the market in a stochastic volatility type economy.
4 A recent paper by Gallant et al. (1998) adopts a strategy similar to ours though not involving options. They consider

the bivariate process of the fundamental and the daily high/low spread, which provides extra information about the course of
volatility.

2
price and option prices and calibrate the parameters of the di usion and its associated option pricing model
to t the conditional density of the market data dynamics. The procedure coined by Gallant and Tauchen
(1996) as Ecient Method of Moments (EMM), has been used primarily to estimate di usions using only
St : We extend it to handle option prices and fundamentals simultaneously. The EMM procedure, which
is a simulation-based estimation technique, allows estimating the model parameters under both objective
and risk-neutral probability measures if we use implied volatilities and the underlying asset data jointly.
Indeed, time series of the underlying asset provide estimators under the objective probability measure while
risk-neutral parameters can retrieve from options. Since the model we adopt has a closed-form option pricing
formula, we can obtain the BS implied volatilities from the simulated data and contrast them with their
counterparts from the real data via the EMM framework. This leads to the parameter estimates under
the risk-neutral measure. Having estimated separately the risk-neutral and objective measures allows us to
appraise the typical risk-neutral representations used in the literature. In particular, in order to obtain the
closed-form solutions, the standard approach assumes that the linearity of the volatility drift is preserved.
We are able to determine if this assumption is consistent with the data.
The task is challenging and besides knowing how to proceed with it we don't know a priori what gains can
be made in terms of better parameter estimates for di usions and in terms of pricing options and hedging
performances. We compare univariate and multivariate models in terms of pricing and hedging performance.
The univariate speci cations consist of models only using the fundamental (i.e. the usual setup) and models
using only options data. It should be noted, however, that the knowledge of the estimated model parameters
is not sucient to compute an option price or a hedge ratio. We have to know the latent spot volatility
as well. Previous studies treated spot volatility as a parameter and estimated it from the previous day
cross-section of options prices. This approach introduces inconsistencies with the model. A recent extension
of the SNP/EMM methodology introduced in Gallant and Tauchen (1998) allows us to address the problem.
We lter spot volatilities via reprojection, i.e. we compute the expected value of the latent volatility process
using an SNP density conditioned on the observable processes such as returns and/or options data.
Our results show that the univariate approach only involving options by and large dominates. A by-
product of this nding is that we uncover a remarkably simple volatility extraction lter based on a polyno-
mial lag structure of implied volatilities. The bivariate approach appears useful when the information from
the cash market provides support via the conditional kurtosis to price options. This is the case for some
long term options.
These ndings prompt us to consider alternative volatility lters, such as for instance the one obtained
from the GARCH class of models. We examine the quality of various lters through the window of the BS
option pricing model, which allows us to separate the e ects of a particular pricing kernel from the lters

3
contribution. Interestingly, we nd that the role of the pricing kernel is marginal compared to that of ltering
spot volatility.
The remainder of the paper is organized as follows. Section 1 sets the stage for the analysis of the joint
density function of fundamentals and options. We discuss rst the issues addressed so far in the literature
and present the model we estimate. In section 1 we also provide a brief review of the EMM estimation and
reprojection method. Section 2 reports the estimation results and examines the mapping from objective to
risk-neutral measures while section 3 evaluates the performance of the estimated models. Section 4 studies
the role of di erent volatility lters. The last section concludes. Technical material is covered in several
appendices to the paper.

1 Joint Estimation of the Fundamental and Option Pricing Pro-

cesses
It has long been recognized that there is no reason to focus on either cash or option market prices separately.
Numerous papers have confronted empirical evidence obtained from derivative security markets with results
from the underlying and vice versa. In particular, issues pertaining to the informational content of option
prices have been examined extensively in the literature (see Bates (1996), Canina and Figlewski (1993),
Christensen and Prabhala (1998), Day and Lewis (1992), Fleming (1994), Lamoureux and Lastrapes (1993),
among others). At-Sahalia et al. (1997) address essentially the same issue comparing state-price densities
(SPD) implied by times series of the S&P 500 index and the SPD implied by a cross-section of S&P 500
index options. They reject the hypothesis that the two SPD's are the same. As they examine a class of
models with volatility speci ed as a deterministic function of the stock price, one can view their result as a
rejection of such a class of models. Along the same lines, Dumas et al. (1998) examine the out of sample
pricing and hedging performance of the same class of volatility models using also the example of the S&P
500 options and nd that "simpler is better", i.e. this class of models performs no better than ordinary
implied volatility.
Our attempt to model the price behavior of fundamental and derivative securities jointly is motivated by
the very same issues hitherto raised in the literature. Namely, we want to learn more about the informational
content of option prices, like Canina and Figlewski (1993) and many others did. We also want to know how
we can improve the statistical precision of di usion parameters by incorporating options.5 Moreover, we
also want to assess the advantages of multivariate schemes using nancial criteria such as the out-of-sample
5In a similar vein, Pastorello et al. (1994) used at-the-money implied volatilities, replacing latent spot volatility, to estimate
the Hull and White (1987) model. However, they did not estimate the joint process as we propose to do in this paper.

4
pricing and hedging performance of models, like Bakshi et al. (1997a), Dumas et al. (1998) and Jacquier
and Jarrow (1998), among others.
We attempt to investigate these questions in a unifying framework. We use the SV model due to Heston
(1993) for that purpose, though it should be stressed at the outset that our analysis is not limited to this
particular model. The Heston model will be covered in a rst subsection.
The joint modeling of returns and derivative security prices will represent some new challenges which we
will discuss in this section. First we describe the data in subsection 1.2. Since the EMM procedure is widely
used and described elsewhere, notably in Gallant and Tauchen (1998), we will only summarize its major
steps in a third subsection. A fourth subsection deals with reprojection methods, an extension of EMM to
extract latent volatility processes. The empirical results are discussed in the last two subsections.

1.1 The Heston Model

Following Heston (1993) we can write the model as:

p
dS (t)=S (t) = Rdt + V (t)dWS (t) (1.1)
p 2p p
dV (t) = ( ,  V (t))dt + V 1 ,  V (t)dWV (t) + V  V (t)dWS (t); (1.2)
where the model is stated under the risk-neutral probability measure. Equation (1.1) implies that the
stock-price process S (t) follows a geometric Brownian motion with stochastic variance V (t). The second
equation (1.2) states that V (t) follows a square-root mean-reverting process with the long-run mean  = ,
speed of adjustment  and variation coecient V . The Brownian motions WS (t) and WV (t) are assumed
independent. Equations (1.1) and (1.2) imply, however, that Corrt (dS (t)=S (t); dV (t)) = dt. Parameters
with asterisks are those which change when the model is rewritten under the objective probability measure.
Under the change of measure the risk-free rate R is substituted by a drift parameter S and all asterisks
are removed. This model yields the following formula for a price of a call at time t, time to maturity  and
strike K :
C (t; ; K ) = S (t)1 (t; ; S; V ) , Ke,R 2 (t; ; S; V ); (1.3)
where the expressions for j ; j = 1; 2 are provided in Appendix A.
The common practice of estimating di usions using the underlying asset and then relying on option
pricing formula has a number of drawbacks. Standard complete market asset pricing theory determines that
one has to change measure, from the objective to the risk neutral.6 This transformation is often somewhat
6See Harrison and Kreps (1979) and Harrison and Pliska (1981) for further discussion. Arguments about completeness of
markets are typically imposed to guarantee the existence of a unique risk-neutral measure.

5
ad hoc. Although continuous time general equilibrium preference-based asset pricing models readily yield
the mapping from the objective to risk-neutral measure, they often result in rather complex di usion models
for the underlying asset. This is because the equilibrium asset price process is derived endogenously, based
on the discounted ow of dividends using an endogenously determined risk-neutral rate.7 It is therefore
common to use a simple di usion for the asset return and volatility dynamics and assume that the change of
drift (which by Girsanov's Theorem amounts to changing the measure) maintains the same type of processes.
In the case of the Heston (1993) model, this means that the drift change for the volatility process is an ane
function of volatility.8

1.2 The data

Like many previous studies we examine the S&P 500 index and the SPX European option contract traded
on the index. Our analysis obviously requires both options and returns data. The source of our series is the
Chicago Board of Options Exchange (CBOE). The data consist of daily last sale prices of options written
on the S&P 500 index as well as the closing price of the index.9 Speci cally, the dataset contains the date,
maturity month, option price, call/put ag, strike, open interest, volume and, nally, the underlying index.
The sample covers the time period from November 1985 until October 1994. We set aside the last year of data
(i.e. 11:93 to 10:94) for the out-of-sample tests, and use the rest for estimation purpose. Plot (a) of Figure
1 displays the S&P 500 series and shows the familiar pattern, including the negative return corresponding
to the crash of October 1987. The dashed vertical line represents the end of the estimation sample and the
beginning of the data used for the purpose of out-of-sample appraisals of the models.
Various schemes to extract implied volatilities from options have been suggested in the literature (see for
instance Bates (1996) for a survey). We concentrate our attention on the at-the-money (henceforth ATM)
calls, where we de ne at-the-money as S=K 2 [0:97; 1:03], with short maturities as these are the most liquid
instruments. Because of the active trading the implied volatilities of these contracts should convey the most
precise information. Moreover, Harvey and Whaley (1991) note that the ATM volatilities are the most
sensitive to changes in the spot volatility rate since an option's vega is maximized when moneyness is close
7 See for instance, Broadie et al. (1997) for such a derivation.
8 Namely, both under the objective and risk-neutral measures the drift in volatility is linear. Any nonlinear functional drift
transformation would yield a risk-neutral process not within the same class as the objective measure process.
9 As is well known, there is a 15 minute di erence between the close of the AMEX, NASDAQ and NYSE stock markets where

the 500 stocks included in the index are traded and the Chicago options market. This leads to non-synchronicity biases. Harvey
and Whaley (1991), Bakshi et al. (1997), among others, suggest various schemes based on option price quotes or transactions
around the 3 PM market close. We control for the possibility of such biases via our simulation procedure, which will be discussed
later and also in Appendix C.

6
to 1. To be more precise we select the calls with the shortest maturities and, in addition to being ATM, we
require the option to have the strike as close to the index level as possible: K  = arg minK
j S=K , 1 j :10
In order to make the observed S&P 500 index and the simulated underlying fundamental data comparable,
we adjusted the S&P 500 index for dividends. We took a constant continuously compounded dividend rate
of 2% (this is consistent with for instance Broadie et al. (1997)). Therefore, we ignore the lumpiness of
the dividend payments and also conveniently avoid using the historical observed dividends which would
considerably complicate our simulation design. Moreover, since we are dealing with European index options,
this matters little as we are only interested in the total ow of dividends paid over the life of the contract.
Finally, we used the monthly 3-month T-bill yield from CITIBASE as a proxy for the short-term interest
rate. Since the stochastic volatility models we consider assume a constant interest rate, we take the average
yield, which is equal to 5.81396%.
Finally, the estimation of SNP density requires the use of stationary and ergodic data. As noted before
we will estimate SV models, using three types of data: (i) time series data on the S&P 500 index, (ii) BS
volatilities implied by the closest to maturity and at the money calls on the index, and (iii) both jointly.
Therefore, the data entries to the SNP estimation routine are the log-returns on the index. Likewise, rather
than using the implied volatilities which are nonnegative, we will work with the log-volatilities. Despite the
transformations we will refer to these data series as S&P500 and BS volatilities for convenience. Figure 1
plot (c) displays the BS implied volatilities. We note that the highest volatility is, not surprisingly, observed
at the time of the crash. Moreover, it is also important to note the downward trend in volatility beginning
around 1991 and the reversal of this trend in the post estimation part of our data set. This obviously will
make the out-of-sample exercise particularly interesting.11

1.3 The Ecient Method of Moments Estimation Procedure

Several methods have been proposed to estimate the parameters of stochastic volatility models. These
methods range from GMM, Quasi-Maximum likelihood and various simulation based procedures including
Bayesian methods and Ecient Method of Moments (see Ghysels et al. (1996) for a literature review). In
10 Apart from the moneyness lter, several other lters were applied to the data. In particular, observations with a call price
missing or equal to zero were obviously dropped as well. The various lters applied to the data leave us with 1978 observations,
which roughly corresponds to 247 observations per year, i.e. we have a week per year of missing observations.
11 In addition, the more recent years (which are not covered by our sample), especially the summer/fall of 1998, indicate an

upward trending volatility. In other words, the volatility process seems to mean revert as assumed in the Heston (1993) model
we consider here. It should also be noted that the options contracts on S&P 500 were of American type during a very short
period of our sample, namely prior to April 1986. Because we select short maturity contracts where the exercise premium is
very close to zero this should not a ect much of our results. Figure 1 (c) also shows that the 'American option' part of our
sample, prior to April 1986, does not introduce any abnormal patterns.

7
this paper we use the Ecient Method of Moments (henceforth EMM) procedure of Gallant and Tauchen
(1996) which has already found many applications in the estimation of both continuous time and discrete
time stochastic volatility models. Examples include Andersen and Lund (1997), Andersen et al. (1997),
Gallant et al. (1997, 1998), Gallant and Tauchen (1998), Ghysels and Jasiak (1996), Jiang and van der Sluis
(1998) and Tauchen (1997). So far these applications only involve a single data series, either a short rate
process or a stock price (index).12 EMM can be divided into two main parts: (1) the estimation of the so
called score generator, which will be discussed rst, and (2) the estimation of the di usion parameters, which
will conclude this section.
Suppose the process of interest is denoted t . In our application this process can be univariate, involving
either asset returns or BS implied volatilities, it can be bivariate involving both returns and implied volatil-
ities, or in general a panel data series of returns and M option contracts with di erent moneyness and/or
maturities represented by their BS implied volatilities, namely of [(itI )i=1;m ]: In a generic context we assume
that t is a vector with L elements. It has a conditional distribution p(t j It ; ), where It is the information
set and  are the parameters of the stochastic volatility model for t . The asymptotically ecient method
to estimate  is maximum likelihood (MLE), which involves maximizing the function:
1X T
T t=1 log p(t j It ; ) (1.4)

Maximization of (1.4) is equivalent to solving

1X T @
^
T t=1 @  log p(t j It ; ) = 0 (1.5)

where @ log(p(
j
))=@  is the score function. The above expression is the sample equivalent of:
 
E @@ log p(t j It ; ) = 0: (1.6)

Unfortunately, it is very dicult to obtain the likelihood function for stochastic volatility models and hence
it is impossible to compute the score generator @ log(p(
j
))=@ . Gallant and Tauchen suggest to compute
instead a SNP density fk (t j Xt ; ), where Xt is the vector of M lags of t and  is the vector of parameters
of the SNP model. The index k relates to the dimension of  and should expand at the appropriate rate
as the sample size grows to accomplish MLE asymptotic eciency, see Gallant and Long (1997) for further
discussion. We provide some speci c details regarding the SNP density in Appendix B.13
12 There are exceptions, notably Ghysels and Jasiak (1996) who consider the joint process of stock returns and trading volume
and Gallant et al. (1998), who as noted before, consider the bivariate process of the fundamental and the daily high/low spread.
13 It should be noted that EMM can use any score generator which represents the data well. In this paper we will use the

SNP score generator, which is also required for the reprojection procedure described later.

8
 We noted that EMM has two parts. The rst part is the estimation of the auxiliary score generator model.
The estimated SNP density provides the input to the second stage of the EMM estimation procedure. More
precisely, the SNP score function provides the underpinnings for constructing a set of moment conditions.
In particular,  can be estimated through the moment conditions (score function) similar to (1.6), which in
this case will be:
 Z
m(; ^ ) = E @@ log fk (t j Xt ; ^ ) = @@ log fk (t j Xt ; ^ )dP (; X; ): (1.7)

Since these moment conditions should have mean zero they can be used as the basis for a GMM-type
estimation procedure which yields the desired estimate for the parameter vector . The moment conditions
are easier to compute by simulation instead of computing numerically the integral in (1.7). Hence, we
compute sample moments by simulating N observations of t from the SV model.14 With simulated time
series of length N for t with candidate parameters  the left hand side of (1.5) translates into
X
N @
mN (; ^ ) = N1 ^
log fk (t () j Xt (); ): (1.8)
t=1 @ 
Then we can formulate the EMM estimator for  using the following quadratic minimization criterion:

^ = arg min ^ ^
 mN (; ) WT mN (; ): (1.9)
0

Because of the properties of the SNP model,

XT @ log f ( j X ; ^ ) @ log f ( j X ; ^ )
WT = T1 k t t
@ 
k t t :
@ 0 (1.10)
t=1
Asymptotically, the EMM estimator is consistent and normal and has, under suitable regularity conditions,
the same eciency as MLE. See Gallant and Tauchen (1997a) and Gallant and Long (1997) for further
discussion.
We will focus primarily on the case where t represents a bivariate process of a stock return and a BS
implied volatility. The distribution p(t j It ; ) in this context is implicitly de ned by equations (1.1), (1.2)
and (1.3). Suppose now that the parameter vector  can be written as (S ; c ; o; n ), where S is the rate
of return on the underlying asset under the objective probability measure, c contains all the parameters
common to the objective and risk-neutral measure, whereas o represents the objective probability measure
volatility drift parameters and n the risk-neutral ones. In equations (1.1) and (1.2) these correspond to c
14 Since the SV model is formulated in continuous time we need to discretize the process to generate simulated paths. We use
the explicit order 2.0 weak SDE discretization scheme to simulate the processes (1.1) and (1.2) (Kloeden and Platen (1995) pp.
486-487). In Appendix C we provide further details of the discretization scheme. The empirical results reported in the next
section are based on simulated samples of size N = 10; 000:

9
= (V ; ), o = (; ) and n = ( ;  ): The rst step consists of simulating the underlying processes (1.1)
and (1.2) under the objective probability measure for a given set of parameters values for (S ; c; o): It
should be noted that we simulate the latent volatility process V (t), though it is not part of t since we need
to use V (t) to compute S (t). Using the risk-neutral measure parameters (c ; n ); where the values of c
remain the same, we compute the options price according to (1.3) and calculate the BS implied volatilities
from these prices. It is important to note here that since we estimate o and n separately we can test
certain hypotheses about the transformation from objective to risk-neutral measures.
To obtain the BS volatilities, we need to apply the option pricing formula to the simulated data. This
requires not only the parameters, but also choosing time to maturity and strike features of the contract.
Obviously, time to maturity is not available in the simulated data. To make the simulated option prices
comparable with the observed ones in the actual data, we replicate the maturities from the observed data.
Since the sample size of the simulated data is much larger (in order to decrease the simulation error), we
cycle through the sequence of observed maturities in the actual data to cover the entire simulated dataset. In
particular, for the simulated ith observation we use the maturity from the mod(i; T ) where T is the sample
size of the observed data. If the length of the simulated sample is N and a multiple of T , say N = lT , then
this scheme amounts to replicating l times each maturity appearing in the observed data.
We apply a similar strategy for the strike features of the contracts. In particular, we simulate moneyness
instead of strikes, because the simulated sample path of S (t) can be quite di erent from the observed one.
Matching the moneyness with the real data implies strikes not always observed in the real data sample.
This strategy preserves, however, the crucial properties of options. Since we can rewrite the option pricing
formula in terms of moneyness, the dependence on strike will be eliminated. We rotate moneyness from the
observed data in exactly the same way as maturities to simulate BS implied volatilities.
Finally, it should be noted that the simulation approach described in this section also extends to situations
only involving options data or the more commonly used univariate setup based on returns series. In neither
case it is possible to estimate the entire parameter vector  which we wrote as (S ; c; o; n ). With
options data we cannot estimate the drift parameter S under the objective measure. Hence with options
we can estimate the parameter vector (c; o ; n).15 When returns series are used we cannot recover the
risk neutral volatility parameters, i.e. n : Hence, only cases with both fundamentals and derivatives will
involve the full parameter vector (S ; c; o ; n).
15 Please note that options data allow us to estimate the volatility parameter vector under both measures, i.e. both o and
n :

10
1.4 Reprojection
Having obtained the EMM estimates of the model parameters ^ we would like to extract the unobserved
spot volatility V (t) in order to price options according to (1.3). Several lters have been proposed in the
literature, all involving exclusively the return process. Harvey et al. (1994) suggested to make use of the
approximate Kalman lter based on a discrete time SV model. The exact lter was derived by Jacquier et
al. (1994) in the context of a Bayesian analysis of the same model. Nelson and Foster (1994) showed how
di usion limit arguments applied to the class of EGARCH models and provided a justi cation for EGARCH
models as lters of the instantaneous volatility. Some attempts were made to extend these lters to a
multivariate context, see in particular Harvey et al. (1994), Jacquier et al. (1995) and Nelson (1996). These
multivariate extensions all involve exclusively return series and cannot accommodate derivative security
market information. We propose a ltering method, based on the reprojection procedure introduced by
Gallant and Tauchen (1998). We brie y describe rst the method intuitively in a generic context. Then we
focus on the speci c applications we will consider.
Suppose we have a vector process consisting of observable and unobservable time series. For example,
observables could be returns, or BS implied volatilities, while the latent spot volatility would be unobserv-
able. Let us denote the vector of contemporaneous and lagged observable variables by xt and the vector of
contemporaneous unobservable variables by yt ,  is the parameter vector. The ltering problem is equivalent
to computing the following conditional expectation:
Z
y~t = E (yt jxt ) = yt p(yt jxt ; )dyt (1.11)
This expectation involves the conditional probability density of yt given xt . If we knew the density we could
estimate it by p^(yt jxt ) = p(yt jxt ; ^ ). Unfortunately, for SV models there is no analytical expression for the
conditional density available. Therefore, we need to estimate this density as p^(yt jxt ) = fk (^yt jx^t ), where
y^t ; x^t are simulated from the SV model with parameters set equal to ^ and where fk is again a SNP density.
Gallant and Long (1997) show that:
^
lim f (^y jx^ ) = p(yt jxt ; ) (1.12)
k!1 k t t
Hence, the SNP density converges (in terms of the Sobolev norm speci ed by Gallant and Long) asymptot-
ically to the true conditional probability density. Hence, reprojection provides an unbiased estimate of the
latent process, say spot volatility.
Speci cally in the context of our paper the reprojection ltering method, in its generic form, can be
multivariate (i) only involving a vector of returns (multiple series, something not considered in this paper
but feasible), (ii) only involving a vector of options (more on this later), (iii) a mixture of the previous two.

11
The latter strategy will be our prime focus here. It should be noted that one can also consider univariate
schemes which involve either the return series or the BS implied volatilities. The former univariate scheme
would be comparable to the ltering methods of Harvey et al. (1994), Jacquier et al. (1994) and Nelson
and Foster (1994). In section 5 we will digress on such univariate volatility lters based on returns data.
Univariate schemes only involving options have been proposed mostly informally and often, though not
exclusively, rely on the Black-Scholes model where a cross-section of options is used and today's volatility
is treated as a parameter. The volatility is then backed out from the cross-section via the minimization of
the pricing error. Such a method appears in Bakshi et al. (1997a) for instance. The reprojection approach
applied to a vector of options (i.e. reprojection scheme (ii)) is more general, however, as it truly takes
advantage of the panel structure (i.e. cross-section of time series). Even the univariate reprojection method
only involving for instance an ATM option, will result in a time series lter of implied volatilities, which to
the best of our knowledge is a ltering scheme for instantaneous volatility not much exploited so far (and as
we will see in the next section, remarkably good).
In the remainder of this section we will focus exclusively on the applications of ltering via reprojection
which will contain novel features. The speci c applications can readily be extended, however, to all the
aforementioned generic speci cations. There are two novel applications, one deals with a bivariate model of
returns and an ATM option and the other deals with a univariate lter based on options only.16 The repro-
jection scheme relies on a one-step ahead forecast, which is an expectation computed from the distribution of
volatility conditional on the contemporaneous and lagged returns, denoted rt ; and lagged implied volatilities:
^ : Hence, xt = (rt ; tI,1 ; rt,1 ; :::; tI,M ; rt,M ) and yt = V (t) in (1.11).
p(V (t)jrt ; tI,1 ; rt,1 ; :::; tI,M ; rt,M ; )
The computation of the SNP density for V (t) proceeds in several steps. First, using the estimates ^ one
simulates the processes (1.1) and (1.2) which produces the series f^tI,1 ; r^t ; V^t gNt=1 in case of bivariate data
and f^tI,1 ; V^t gNt=1 in the univariate case.17 Let us again denote the vector (^rt ; ^tI,1 ; r^t,1 ; :::; ^tI,M ; r^t,M ) 0

(or (^tI,1 ; :::; ^tI,M ) in the univariate case) by x^t .

0

Second, since the SNP density has a Gaussian lead term we need to transform the simulated volatility
since it is a process which only takes positive values. It would be natural to consider the log-transformation
of V^t , as was done with the implied volatilities data (see section 1.2). However, in this case we need an
estimate of the untransformed spot volatility to substitute into the option pricing formula. We conclude this
section with the description of a piecewise linear approximation to the log transformation which will allow
us to recover spot volatility without any biases.
16 In section 4 we will also consider ltering based exclusively on returns series. Since this is more standard we will not digress
here on the details of this procedure.
17 To streamline the notation we use V ^t for the simulated V (t). Furthermore, since k grows with the sample size N we select
a simulation size of 10,000 observations.

12
 Consider a rst order Taylor expansion of the logarithm of volatility, which we denote by L(V (t)). We
nd the SNP density fk (L(V^t )jx^t ).18 >From Jensen's inequality we have that E (C (V (t)))  C (E (V (t))),
for any concave function C . This inequality becomes an equality only if C is linear, hence the use of L(
).
Hence, the SNP density tted to the simulated data therefore produces a mean of V^t conditional on the
observable vector x^t , which in turn yields the desired ltered values. The rst order Taylor expansion is an
approximation and a priori can be centered anywhere. One obvious starting point is to compute the Taylor
expansion around the mean of simulated V^t 's. This approximation may be quite inaccurate in the tails,
however. This is particularly important as the inverse transform may easily result in negative volatilities.19
Indeed, experiments showed that the rst order Taylor expansion centered around the mean of simulated
V^t resulted in 32% negative volatilities. We therefore took the rst order Taylor expansion at di erent
points yielding a piecewise linear approximation similar to a spline transformation. Further experiments
showed that a linear approximation center around two points yielded roughly 3.6% negative volatilities and
around four points yielded roughly 1.6%. We therefore took a four-point linear spline and centered the rst
order Taylor expansions around the 8th, 24th, 50th, 76th and 92nd percentiles of the simulated V^t marginal
distribution with breakpoints at the 16th, 32nd, 68th and 84th percentiles. The remaining few negative
values were replaced by a small positive number, namely by 0.0001.

2 Empirical Results and an Examination of the Mapping between

Objective and Risk-Neutral Measures
The discussion of the empirical results is divided in three subsections. First we cover the two stages of
the EMM procedure. The rst stage, which pertains to the SNP density estimation is covered in a rst
subsection. The parameter estimates of Heston's model are discussed in a second subsection while the nal
subsection discusses the lessons we can draw from having estimated jointly the risk-neutral and objective
probability measures.
18 Fitting SNP densities in a reprojection exercise requires a slightly di erent setup since the explanatory variables are
exogenous to the dependent variable, hence we set MR equal to 0 (see Appendix B). The SNP code also requires other slight
modi cations, so that the lags of dependent variables would not be included in the conditioning set as in the standard SNP
density speci cation.
19 This happens because the linear transformation imposes lower and upper limits on the range of a volatility (as opposed to

the log-transformation which can potentially take any value on the real line).

13
2.1 SNP density estimation results
The SNP density estimation results are reported in Panel A of Table 1. We estimate the SNP density for
the three types of data: (i) the log-returns on the S&P500; (ii) the log of the BS implied volatilities of
the closest to maturity and ATM call options; (iii) both series jointly. Rather than report the parameter
estimates we focus instead on the density structures as characterized by the tuning parameters Kz , KX ,
M , M and MR (see Appendix B for further details). To facilitate the interpretation of the results we
supplement Panel A of Table 1 with a second Panel B which describes the generic features of SNP densities
for di erent combinations of the tuning parameters. Panel A reports in addition to the structure of the
estimated densities, the values of the objective function sn based on the density in (B.1) (in Appendix B),
the values of the Akaike information criterion (AIC), the Hannan and Quinn criterion (HQ) and the Schwarz
Bayes information criterion (BIC). Each line reports the best BIC model within each class as outlined in
Panel B. The overall best model selected by BIC is in boldface as it is the model selection criterion which
is predominantly used for model selection. Comparing panels A and B, we note that all series require
models with Kz > 0; meaning non-Gaussian innovations. The fact that MR > 0 means that we also nd
ARCH e ects, even with BS implied volatilities, in the SNP Hermite polynomial expansions. We also need
autoregressive terms in the mean, since all expansions have M > 0: For BS implied volatilities we need
the longest lags in the mean, namely M equals 9. Obviously the mean equation for BS volatilities is like a
conditional second moment equation. This is comparable to the returns series where we need 8 lags in the
ARCH expansion (i.e. MR = 8 for S&P 500 return series).20
It is worthwhile to examine the estimated density plots. In addition to the raw data Figure 1 also
features plots of the estimated SNP densities for the univariate cases. Figure 2 does the same for the
bivariate case involving the joint process of returns and BS implied volatilities. Plots (a) and (c) in Figure
1 were already discussed in the data section. They display the S&P 500 series and the corresponding BS
implied volatilities. Plots (b) and (d) show the corresponding SNP densities with a standard normal p.d.f.
superimposed (represented by the dashed line). The estimated densities show the familiar patterns, namely
they are peaked, leptokurtic and weakly skewed. Figure 2 reports the joint density in (a) and the contour
plot in (b). The marginal densities appear in (c) and (d). The contour plot suggests the presence of slight
negative correlation between returns and volatility, which supports the presence of a leverage e ect. Hence,
20An alternative SNP density speci cation involving an AR-GARCH lead term has been suggested by Andersen et al. (1997).
This alternative speci cation would be appropriate for discrete-time analogs to the continuous time di usions and potentially
represent statistical eciency gains. While these arguments clearly apply to univariate return series, it is not clear they apply
to BS volatilities and the bivariate models. We therefore used the original speci cation suggested by Gallant and Tauchen
(1989).

14
in the estimation we imposed the restriction that   0: In addition, we also imposed a restriction to prevent
volatility reaching zero. Since we have a square-root model, we use the results of Cox et al. (1985), namely if
V2  2; then V (t) in (1.2) does not reach 0 with probability 1. This condition is also important because it
excludes arbitrage opportunities which occur in Heston's model when V (t) = 0. In particular, in the absence
of uncertainty (i.e. when V (t) equals zero) the drift is not equal to the risk-free rate in Heston's model, an
issue we will further elaborate in the next section.

2.2 SV model parameter estimates

We turn now to parameter estimates of Heston's SV model. Table 2 reports all the estimation results.
Traditionally this model is estimated with returns only, so we report this con guration as a benchmark. Then
we proceed to the estimation results exclusively relying on options, in the spirit of Pastorello et al. (1994).
Next, we consider the structural parameters obtained from matching the moments dictated by the bivariate
SNP score. The rst observation is that Heston's SV model is mostly rejected no matter what the data
con rmation is (all z -statistics reported in the brackets are large).21 The rejections vary dramatically across
the rows in Table 2 depending on which data is used. In particular, the returns data allow to approximate
the process reasonably well, however the precision of the estimates is very poor (for example the standard
error for  is 0.29). This result is consistent with previous ndings (see, for example, Gallant et al. (1997)).
It is worth noting here that the estimated parameters are annualized as is typically done in option pricing
models. Hence the reported values are quite di erent from their GARCH counterparts, for instance, which
are typically characterized on a daily basis. It is easy to provide the link between them, however, since the SV
model allows for temporal aggregation. For example the value of   0:93 corresponds to 0:93=252  0:004
on a daily scale or roughly 0:996 as the corresponding persistence parameter in a GARCH(1,1) model.
Of course our major interest is not to appraise the model exclusively on the grounds of its statistical
properties. Its pricing and hedging features, discussed in the next section, will be of prime interest.
Using the BS volatilities data improves the precision of the estimated parameters substantially (the
standard errors are of order 10,2 or less). There are ve parameters which overlap between the rst and
second row of Table 2. Recall from the discussion in section 1.3 that we can estimate with options data the
parameter vector (c ; o; n ) but not the drift S under the objective measure. Therefore, the parameter
vector (c ; o) is common to the rst and second row in Table 2. For all the parameters common to both
data sets, we obtain roughly the same point estimates, yet the precision of the estimates are dramatically
21 The standard Normal distribution is used to evaluate the z -statistics which are an approximation to asymptotically 2
distributed GMM-type overidentifying restrictions test statistic (see Gallant and Tauchen (1997) for further discussion). We
report the statistics, not their p-values as they are all basically equal to zero.

15
improved with the second data set. This is not so surprising as almost all estimated parameters are related
to the volatility process (1.2). Therefore, looking at the process through the observed implied volatilities
should give us more precision, an observation also made by Pastorello et al. (1994).22 The z -statistic is
huge (247.447), however, indicating serious inconsistencies between the model and the data. In particular,
it means that our model does not explain very well the information extracted from the options prices.
Last but not least in Table 2 is the model estimated with the joint data. Here, we can t both the
objective and risk-neutral density parameters and therefore make direct comparisons with the previous two
univariate setups. The bivariate model shows a great improvement in t compared to the options-based
approach, the z -statistic is greatly reduced although the model is still strongly rejected. The precision of
the estimates is the best of all across the di erent con gurations, while the point estimates roughly remain
the same. The latter observation is important as the bivariate estimation approach is a ected by the non-
synchronicity in options and returns data. This con rms the observation made in Appendix C that matching
moments instead of sample paths appears to render any spurious non-synchronicity e ects insigni cant.
Since we are interested which type of data we should use to price and hedge options, we can conclude
from Table 2, that the returns series only should not be used as we cannot directly infer the parameters
under the risk-neutral probability measure, unless auxiliary assumptions are made. The two competing data
sets which allow us to identify directly the necessary parameters are: (i) BS implied volatilities, and (ii)
joint returns and implied volatilities series. We will therefore focus exclusively on these two to appraise how
options are priced and how well they perform for the purpose of hedging.

2.3 An Examination of the Mapping between Objective and Risk-Neutral Mea-

sures
Having estimated the parameters of the risk neutral and objective measures simultaneously allows us to
examine whether the change of measure restrictions imposed in Heston's model are supported by the data.
The bivariate setup indeed yields estimates of the entire parameter vector  = (S ; V ; ; ; ;  ;  ). Obvi-
ously, the mapping between the two measures also pertains to the general equilibrium underpinnings of the
model. The joint estimation of the objective and risk-neutral measures introduces a new set of issues hitherto
not addressed in the literature of equity option pricing which focused exclusively on univariate estimation
schemes. It should be noted that similar issues have been examined in the literature on ane models for the
term structure of interest rates. For instance, de Jong (1997) estimated the market price of risk using time
series and cross-section data for the class of models introduced by Due and Kan (1996) where the assumed
22 Pastorello et al. (1994) actually consider the model of Hull and White (1987) and conduct a Monte Carlo study showing
the dramatic improvement of parameter estimates when options data are used.

16
price of risk is proportional to the underlying factors.
The market price of volatility risk in Heston (1993) has the same structure, namely it is assumed to be
proportional to volatility. In particular, Heston assumes (1) an equilibrium process for the stock price, (2) a
speci c choice of the representative agents' utility and (3) an equilibrium consumption process. These three,
when combined, may result in overidentifying assumptions regarding the underlying economy as it suces
to specify any two attributes of the economy to derive the third. More speci cally, the analysis in Broadie et
al. (1997) shows that particular assumptions in Heston (1993) are indeed overidentifying. This leads us to
consider a more general form of the volatility risk premium. However, since it is crucial to preserve the linear
structure of the volatility drift we can only augment the speci cation by adding a constant term. Therefore,
we consider the volatility risk premium is of the form k0 + k1 V (t), where k0 =  ,  and k1 =  ,  .
Given the complete parameter vector ^ , we can evaluate the prices of risk. Namely, the Radon-Nikodym
derivative of the objective probability measure with respect to the risk-neutral one is computed as follows:
 1 Z t+ Z t+ Z t+ 
t; = exp , 2 2 2
(1 (u) + 2 (u))du , 1 (u)dWS (u) , 2 (u)dWV (u) ; (2.13)
t t t
where (t) = (1 (t); 2 (t)) is the market price of risk. Since we know the parameter values under both
0

measures we can infer (t). By Girsanov's theorem we have:

p
S S (t) , 1 (t) V (t)S (t) = RS (t); (2.14)
p 2p p
 , V (t) , 1 (t)V 1 ,  V (t) , 2 (t)V  V (t) =  ,  V (t): (2.15)

Therefore,

1 (t) = pS , R ; (2.16)

V (t)
p
2 (t) = pC1 , C2 V (t); (2.17)
V (t)
where
 p 2
C1 =  ,  , (S , R )V 1 ,  ; (2.18)
V

C2 = ,  : (2.19)
V
The estimated parameters also allow us to estimate the volatility V (t) via the reprojection procedure (in
the next section we will give the details of the actual implementation) and therefore compute the sample
paths for prices of risk appearing in (2.16) and (2.17). They are reported in Figure 3 which has three panels.
The top panel plots the values of 1 (t) computed with the reprojected volatilities, while the second does
the same for 2 (t): The lower panel reproduced the sample path of the actual reprojected volatilities V^ (t):

17
The rst two panel display the time series processes which represent the risk adjustments to the Brownian
motions WS and WV : Realizing that for instance dWV (t) = dWV (t) , 2 (t)dt it is clear that the risk adjusted
densities may di er substantially from the objective Brownian motions. Therefore one has to be careful when
simulating fundamental processes, as it is not possible to directly simulate processes under the risk-neutral
density.
It appears from Figure 3 that there is a strong symmetry between 1 and 2 , i.e. the return and volatility
risk adjustment. In particular, the main observation we can make from Figure 3 is that the market price of
the asset risk decreases as volatility increases. This represents a violation of the no-arbitrage principle, since
one can increase returns on a portfolio without bounds by increasing its risk. On the other hand, the market
price of the volatility risk is consistent with the pattern of volatility displayed in the lower panel of Figure
3 , i.e. it has positive comovement with the volatility. If we consider the theoretical expressions for i (t) in
(2.16) and (2.17), we can note that while the behavior of 2 (t) is data dependent, i.e. di erent parameters
values may in principle yield a di erent pattern, the dependence of 1 (t) on V (t) is purely model driven.
The fact that the di erence between the drifts of the process under the two measures is constant regardless
of the uncertainty (volatility) in the economy imposes unrealistic properties on the price of risk. As Garman
(1976) noted, the assumption of the constant risk-free rate imposes a severe restriction on the market prices
of risk and admissible di usion processes. If taken to the extreme, as V (t) approaches 0, the market price
goes to the in nity and creates potential arbitrage opportunities. The arbitrage violation inherent in the
model speci cation may manifest itself as a violation of the Novikov condition or inconsistencies in the risk-
neutral measure.23 The conclusion we can make is that the Heston model may be improved by considering a
stochastic drift in at least one of the probability measures. This conclusion is also related to the equilibrium
foundation of the model.

3 Assessing Pricing and Hedging Performance

Statistical criteria for model selection are one thing, nancial criteria such as pricing and hedging performance
out-of-sample are obviously the ultimate scope of model selection. Therefore, we investigate the performance
23The Novikov condition is imposed to suce that the Radon-Nikodym derivative is a martingale (see Karatzas and Shreve
(1991)). Even if the Novikov condition is satis ed we may still not nd the volatility drift assumption acceptable. Therefore
the use of another test may be warranted. Breeden and Litzenberger (1978) show that the underlying asset state-price density
(SPD) is a second derivative of a call option price with respect to strike evaluated at the assets price and normalized to integrate
to one. This approach yields a formula for the fundamental SPD at time t and time to maturity  as a function of the risk-
neutral parameters only. An alternative characterization of the SPD is also feasible since we have the parameters estimated
under both measures. Chernov and Ghysels (1998) suggest tests which examine whether two SPD's are identical. Appraising
the similarity of two probability densities is rather non-trivial, however. These tests are therefore tedious to implement.

18
of the alternative model speci cations obtained from the previous section. In order to assess the magnitude
of the forecast errors we use the Black-Scholes valuation model as a benchmark. All the evaluations reported
in this section are out-of-sample. Namely the estimation sample used in the preceding section covered
November 1985 to October 1993, whereas the sample used to appraise all the models runs from November
1993 to October 1994.24 Therefore, since we use out-of-sample data, our results will not be contaminated by
in-sample data mining. We noted before, particularly when discussing Figure 1 plot (c) which displays the
BS implied volatilities, that a downward trend in volatilities which started around 1991 reversed in the post
estimation part of our data set. This obviously will make the out-of-sample exercise particularly interesting.
A rst subsection is devoted to pricing of options. A second looks at hedging and last but not least we
devote a separate section to a remarkable simple volatility lter for option pricing which emerges from our
analysis. The examination of this lter is indeed important as all the model con gurations yield roughly the
same parameter estimates (recall the results in Table 2). Hence, whenever the same option pricing formula
is used, the di erences in hedging and pricing are mostly due to ltering.

3.1 Pricing reliability

The model parameter estimates, the reprojection lters to extract the latent volatility process and Heston's
European call pricing formula give us all the ingredients to price any SPX contract with a particular time to
maturity, strike and given cash price for the underlying index. Equipped with these tools we can compare the
call prices predicted by the models with observed prices. In our appraisal we do not consider in-the-money
calls with moneyness greater than 1.03, as there is a very thin market for such options. Furthermore, we
separate the calls into twelve groups by moneyness and time to maturity. Since the number of contracts
varies through time we assume that each group contains nt options at time t.
Two measures of pricing errors are considered. The rst is an absolute measure, denoted Dap , and has a
dollar value scale. The second is a relative measure, denoted Drp , and re ects percentage deviations. The
two measures are de ned as follows:
v
u
u
t PT1 X Xt (Citobserved , Citmodel )2;
T n
Dap = (3.1)
t=1 nt t=1 i=1
v
u nt  C observed , C model 2
u
t PT1 X T X
Drp = it
Citobserved
it : (3.2)
n
t=1 t t=1 i=1
The results, are divided in four di erent categories of moneyness, namely deep out-of-the-money (OTM with
S=K less than 0:94), OTM (with :94 < S=K < :97), slightly OTM (with :97 < S=K < 1) and slightly in-the-
24One could call this a genuine out-of-sample approach, as in Dumas et al. (1998) or Jacquier and Jarrow (1998), in
comparison to some performance evaluations which use in-sample one step ahead forecasts.

19
money (with 1 < S=K < 1:03). The last two categories are usually viewed as ATM options. Three maturity
horizons are considered for all the moneyness categories, short (less than 60 days), medium (between 60 and
180 days) and long (more than 180 days). The respective sample sizes for each of the twelve out-of-sample
cells are reported in Table 3. For deep out-of-the-money options we have roughly between 313 and 448
contracts to compute pricing errors and slightly less for computing hedging errors.25 For the remaining
three moneyness categories, there is a pronounced downward trend in the number of contracts as time-
to-maturity increases. The maximum number of observations is 1377 and smallest is 90: Table 4 contains
the pricing errors Dap and Drp computed for three di erent model speci cations. The rst is \BS ", which
involves pricing options with the Black-Scholes model and a volatility estimate based on the previous day
ATM option contract (see for instance Bates (1996) for further discussion on this approach).
A slight digression is in order here. The aim of the present exercise is to compare pricing performance of
di erent models giving them an equal chance to succeed. Since the implementation of BS is ad-hoc, while
we make every attempt to implement SV in a consistent way, this task is very dicult. The discussion
in Bates (1996) allows us to view the ATM implied volatility as a way to introduce some consistency into
the BS implementation. Furthermore, it is well known that if we price a deep OTM contract based on its
volatility implied from the previous day, we will obtain a smaller pricing error. However, it does not give
an equal chance to SV which could have been estimated based on OTM contracts as well. Since the focus
of the present study is the information which can be extracted from the ATM options and asset returns, we
believe that the proposed scheme allows for a balanced approach to the task.
Next is the speci cation denoted \V ol" in the table, which involves Heston's call price model esti-
mated using a univariate speci cation using options. The speci cation denoted \Joint" corresponds to the
bivariate model. Table 4 reports besides the absolute and relative pricing errors for each of the twelve mon-
eyness/maturity groups also statistics which test whether there are any statistically signi cant di erences
between the pricing errors computed from BS , V ol and Joint: We observe that in general pricing errors can
be large with dollar values (i.e. Dap ) ranging from 5:23 dollars to 42 cents. Not surprisingly, large errors
occur for long maturities. The relative errors (i.e. Drp ) range from 20:71 to 0:14 with the same pattern.
For our purpose, a comparison across di erent model speci cations is more important. As noted before,
to appraise the di erences we compute formal tests based on the following set of moment conditions, namely:

0 PT ,1 PT Pnt , observed vol:model observed
2 1 0 1

( n ) ( C
@ PT t=1 ,1 PT t=1Pnti=1 observed
t it , C it ) = ( C it )
 , m A @ 0A
joint:model 2 = ; (3.3)
( n)
t=1 t t=1 (Ci=1 ,C
it it)=(C observed ) , m
it 0
25 The computation of hedging errors will be discussed in the next subsection. Suce here to say that they involve less
observations as contracts are required to be traded at least two consecutive days.

20
where m is the common mean relative error under the null.26 The common parameter m in (3.3) implies
that the two pricing errors have the same mean, and hence the two models being compared are equivalent in
pricing (at least as far as the mean is concerned). Rejecting the null of a common mean implies one model
is superior in pricing. We can proceed in several ways to test the null hypothesis. We should note that
the errors form a panel data set, because we have a time series of cross-sectional observations since within
each cell a total of nt contracts are traded. The sample data are therefore correlated which precludes us
from considering a simple t-statistic. Instead, we use a GMM-based procedure involving a Newey and West
estimator for the covariance matrix. The common mean m entails one overidentifying restriction on the
moment conditions in (3.3). Therefore the overidentifying restrictions test statistic is distributed 2 with
one degree of freedom under the null hypothesis (see Hansen (1982)). The test statistics are reported in the
left panel of Table 4. We report overidentifying restrictions test statistics for pairwise comparisons between
V ol and BS , V ol and Joint, and nally Joint versus BS . These tests are reported for the three maturity
horizons, i.e. short (less than 60 days), medium (between 60 and 180 days) and long (more than 180 days)
separately for each of the moneyness categories. In addition we also report one joint statistic per moneyness
category for pricing errors over all three maturity horizons combined (denoted as group tests).
The most striking result which emerges from the table is dominance of the V ol approach. The comparisons
of V ol and BS , V ol and Joint are always signi cant at least at the 1 % level for all moneyness categories,
with the three maturities combined. In each of the individual maturity/moneyness cells we also observe the
statistical signi cance of V ol, though there are exceptions and most tests are only at the 5 % or 10 % level.
Moreover, in each of the cases where the overidenti cation test involving pairwise comparisons of V ol with
either BS or Joint are signi cant we nd that the former has the smallest pricing error. Hence, the estimation
of SV models only involving ATM options outperforms BS and the bivariate approach denoted Joint: Besides
yielding statistically signi cant superior pricing performance, we observe from the results reported in Table
4 that the improvements over BS range from a 1:21 dollar reduction (the relative error dropping from 0:45
with BS to 0:33) for OTM long maturity, to a small eight cents reduction with relative error dropping from
5:78 to 4:46 for slightly OTM short maturity. The V ol speci cation outperforms the Joint with the largest
improvement for the deep OTM long maturity category, though this time the improvement is only 48 cents.
The smallest one is also in the same moneyness category but has short maturity and is equal to only 2 cents.
It is worth re-emphasizing here that the comparisons between Joint and V ol involved the same call price
formula of Heston and basically the same parameters (see Table 2). Hence the di erence are very much due
to the ltering procedures, i.e. the procedure to extract the latent volatility process from the data.
Since the outperformance of the Joint by the V ol model speci cation is somewhat surprising we conducted
26 Computations with the absolute errors could be performed as well but are not reported here.

21
some further tests which are not reported in the table. These additional tests required a departure from
the reprojection procedure described in section 1.4. The details of the reprojection lter underlying V ol will
be discussed in the next section. We therefore focus only on the augmentation from V ol to Joint ltering,
the latter involving the bivariate lter and therefore supplementing the options data with returns series.
It is not clear from the tted SNP density in the reprojection routine whether certain individual moments
of returns may provide information that is enhancing the pricing of options. To separate the potential
contribution of individual moments we computed reprojection lters only involving a particular moment
of returns in addition to implied volatilities. For instance, to determine whether the mean return has any
pricing information, we consider a bivariate lter with implied volatilities and past and concurrent returns.27
This speci c lter therefore investigates how much of the leverage e ect in the stock index which is not
already incorporated in the implied volatilities improves the pricing of European-type contracts. The same
strategy can be applied to isolate the informational content of the second, third and fourth power of returns.
These computations reinforced our ndings with the SNP reprojection lter, with one potentially important
exception. Indeed, we found that the conditional kurtosis, i.e. a lter using a lag operator in the fourth
power of returns in addition to implied volatilities, helps to improve the pricing of long maturity slightly
in-the-money option contracts. The statistically signi cant improvement decreased the pricing error from
3:85 to 3:03 dollars.
The general conclusion we can draw so far is that using Heston's call price model with a relatively simple
volatility ltering scheme using past options data yields the most desirable outcome across all the three
model speci cations considered. By and large, all the models perform relatively well at short maturities, of
course. The discrepancies really play out in the cells involving long maturities and OTM contracts. It should
be noted that the estimation of the SV models was con ned to ATM options. Therefore, one would expect
that incorporating in the estimation sample contracts similar in nature to the ones priced out-of-sample, like
OTM contracts, would certainly improve the record across the V ol and Joint speci cations. We will further
elaborate on this in the concluding section and delegate this to future research. In section 4 we will also
revisit some of these questions when we disentangle ltering and pricing kernel e ects.

3.2 Hedging performance

It is common to hedge the market risk via a combination of stocks and options. There are numerous hedging
techniques available in theory, some are quite sophisticated and involve several instruments simultaneously.
However, since options are in zero net supply (i.e. there always has to be a party holding an opposite position
27 Hence, we ignore the Hermite polynomial expansion terms in the SNP density and isolate the Gaussian kernel.

22
in an option contract), many of the sophisticated hedging strategies are hard to implement in practice (see
Figlewski (1998) for further discussion). We will therefore concentrate on a simple minimum variance hedging
which uses just one option. In particular, if we want to hedge our position in one call with NS (t) shares of
stock, we choose the number of shares in such a way, that the remaining cash position
CP (t) = C (t; ; K ) , NS (t)S (t) (3.4)
has minimum variance. This is achieved by taking
NS (t) =
S (t; ; K ) + S(tV)
V (t; ; K ); (3.5)
where

S (t; ; K ) = @C (t;@S; K ) = 1 (3.6)
Note that this simple formula can be obtained because the SV option price (1.3) is homogeneous of degree
one in S (t) and K .28 Finally, we also have that

(t; ; K ) = @C (t; ; K ) = S (t) @ 1 , Ke,R @ 2 ;
V @V @V @V (3.7)
Since the SV model does not account for all sources of risk (for example, the interest rate is assumed to
be constant), we will not end up with zero, if we try to unwind our position the next day. The hedging error
will therefore be:
H (t +
t) = NS (t)S (t +
t) + CP (t)eR
t , C (t +
t;  ,
t; K ); (3.8)
where
t is equal to one day.
We use summary statistics similar to those in the previous subsection to report the model hedging
performance. In particular, we construct absolute (Dah ) de ned as:
v
u
u X
T X
nt
Da = t PT1
h Hi2 (t): (3.9)
n t
t=1 t=1 i=1
and comparable relative (Drh ) measures. To obtain the latter, let us rewrite (3.8) as:
H (t +
t) = CP (t)eR
t , CP (t +
t) = Cashmodel observed
t+
t , Casht+
t ; (3.10)
where CP is di erent from CP because the number of shares was computed in the previous period, not as
de ned in (3.4). This yields the relative measure of the hedging error which is de ned as:
v
u !2 vu
u XT Xn t observed , model u XT X nt  H (t) 2
Dr = PT
h t 1 Cash it Cash
observed
it t
= PT 1 i (3.11)
t=1 nt t=1 i=1 Cashit t=1 nt t=1 i=1 CP i (t)
28 For further details see Nandi (1998).

23
 The hedging performance comparisons for the BS , V ol and Joint are reported in Table 5, again using the
classi cation of twelve moneyness/time-to-maturity cells. The absolute and relative errors are complemented
with statistical tests built on moment conditions similar to (3.3) which yield an overidentifying restrictions
test. We compute the overidentifying restrictions tests for all individual moneyness/maturity cells as in Table
4, including the grouped maturity tests in the last column of Table 5. We observe that the V ol speci cation is
again the dominant one, yet not signi cantly di erent from both alternatives. In fact all three speci cations
perform rather well and hardly any of the overidentifying restrictions tests are signi cant, except in four
individual cells suggesting some very weak improvements due to the use of the V ol speci cation. Overall we
should conclude from the results in Table 5 that hedging strategies, unlike the pricing errors, appear to be
rather insensitive to model speci cation.

4 Disentangling Filtering and Pricing Kernel E ects

To disentangle the advantages of various lters we need to be able to construct a situation where we separate
the e ect of ltering from that of the pricing formula. In Table 4, for instance, the comparison of BS with
V ol involves two components, one due to the ltering and the other due to the di erent pricing formula, i.e.
Black-Scholes versus Heston's pricing model. In the rst subsection, we examine all models rst through the
window of the BS option pricing model, which allows us to separate and appraise the e ect of the alternative
volatility lters on the pricing of options. In the second subsection we compare pricing performance of SV,
BS and GARCH option pricing models with their proper pricing formula. For GARCH models we follow
Duan (1995) who developed a framework for option pricing under a local risk-neutral probability measure.
This comparison will allow us to see whether pricing formula more advanced than BS add any improvement
in pricing performance. Moreover, this two-tier comparison clearly allows us to separate the role of volatility
ltering from that of the pricing kernel in pricing derivative contracts.

4.1 Filtering latent spot volatility

We propose to use BS, GARCH and SV models rst as volatility lters and evaluate them via the BS formula.
To facilitate the discussion we will introduce the following notation for the ltering schemes (estimators of
p
the V (t)):

 tI is the estimate of today's volatility by the previous day's BS implied ATM option
 tG is the estimate of instantaneous volatility using a univariate GARCH(1,1) model (see later for
details)

24
  rtR is the estimate of instantaneous volatility via reprojection using a univariate scheme based on
returns.
 vtR is the estimate of instantaneous volatility via reprojection using a univariate scheme based on BS
implieds, i.e. the scheme underlying V ol in Table 4.

Each of these ltering schemes warrant some further discussion. Since the univariate V ol speci cation of
Table 4 involving options is the most successful and novel, we digress rst on the practical implementation
of vtR . Constructing a reprojection lter involves a model selection procedure for the tuning parameters
Kz , KX , M , M and MR of the SNP density, except that the density is estimated this time with simulated
data. To facilitate comparison with the sample data SNP ts we report the reprojection model selection
results in Panel C of Table 1 right side by side with the empirical sample results which appear in Panel A.
It is important to note here that the SNP densities applied to the sample data are not comparable to those
used in the reprojection. Recall from (1.11) that we are computing a conditional distribution of, say latent
spot volatility given BS implieds. Hence, in the case of V ol the densities reported in Panel A are univariate
conditional densities of BS implieds given their own past, whereas in Panel C we t a conditional reprojection
density of latent conditional on implied volatility. Therefore, we do not expect the SNP densities to coincide.
Moreover, the densities in Panel C cannot be autoregressive (i.e. involving past latent volatilities). Hence,
no AR or ARCH parts appear in the tted reprojection densities. The tuning parameters are again selected
via the BIC criterion. For the reprojection density involving BS implied volatilities the model speci cation
which minimizes the BIC criterion is a simple twenty two lags linear operator. In Panel C of table 1 we
note that M equals 22 and besides a constant no other tuning parameters are set to values greater than
zero. Hence the ATM implied volatilities are combined in a weighted historical moving average with a 22
day window.
The lter is indeed remarkably simple given that we started out with a general speci cation of a SNP
conditional distribution. Table 6 lists the lter weights which extract spot volatility from BS implieds. The
weights range from 0:24514 to 0:03277: The intercept is negative and equal to ,0:16320: The largest weights
in the lter appear at the shorter lags and the decrease in weights is roughly linear in the lags. Hence, the
extraction scheme puts less weight on observations from the options markets which date back about one
month or 22 trading days.
The simplicity of this scheme is rather surprising if one thinks of the complexity of the task. Indeed,
alternative lters, such as those proposed by Harvey et al. (1994), Jacquier et al. (1994) and Nelson and
Foster (1994), involve highly nonlinear functions of returns. Hence, the virtue of using volatility data to
predict future spot volatility is that one can limit the lter to a linear structure. It should be noted, however,

25
that the construction of the lter is unfortunately not as simple as running a linear regression model. Since
spot volatility is a latent process one can only recover the lter weights via the reprojection procedure. To
clarify this, suppose we would consider the much simpler task of regressing implied rather than spot volatilities
on the same window of past implieds. Such a linear regression model is also reported in Table 6, with its lag
coecients appearing side-by-side with the reprojection lter. The OLS parameter estimates show absolutely
no resemblance, neither numerically nor statistically, to extraction lter weights. The di erence between
the two lag polynomials is rather easy to explain when one realizes that implied volatility is not like spot
volatility but rather relates to the expected volatility over the remaining time to maturity of a contract (see
for instance Hull and White (1987)). Therefore, the calculation of the lter weights remains a nontrivial task
which cannot be performed by simple regression methods. However, once di usion parameters are available
the task is relatively straightforward.
Furthermore, this volatility ltering scheme provides insights into the literature which evaluates implied
volatilities as estimates of the actual volatility. The motivation for the work in this area is the implication
of the Hull and White (1997) model that the BS volatility is equal to the average integrated volatility over
the remaining life of an option. The conclusions of the studies are somewhat mixed (see Bates (1996) for a
review), but the major ndings are that the implieds are positively biased estimates of the actual volatility,
but still contain information about the future volatility value.
We noted before that to construct rtR ; Harvey et al. (1994) suggested to make use of the approximate
Kalman lter based on a discrete time SV model. The exact lter was derived by Jacquier et al. (1994) in
the context of a Bayesian analysis of the same model. Along the same lines one can apply the reprojection
approach using EMM applied to returns data. The parameter estimates are reported in Table 2, rst line.
This will be the rst of two return-based lters we will consider. The reprojection SNP tuning parameters
for returns are again reported in Panel C of Table 1 ( rst block).29
The next lter, namely tG ; is based on the work of Nelson and Foster (1994) who showed how di usion
limit arguments applied to the class of ARCH models provides a justi cation of ARCH models as lters of
the instantaneous volatility. To implement this lter let us consider the GARCH option pricing model due to
Duan (1995) who proposed a GARCH(1,1) model which is described under the local risk-neutral probability
measure by:
29We also report in Panel C of Table 1 the reprojection density of the joint speci cation. We did not include this lter in
the analysis of this section because the SV model using options data was shown to dominate by and large the joint model in
terms of pricing. However, we still report this reprojection density speci cation as it was used in the analysis of the objective
and risk-neutral measures in section 2.3.

26
 p
log St =St,1 = R , 21 V (t) + V (t)t ; (4.1)
V (t) = 0 + 1 V (t , 1)(t,1 , c , )2 + 1 V (t , 1) (4.2)

where t  N (0; 1). Volatility in this case follows a NGARCH(1,1), which is the conventional GARCH(1,1)
process of Bollerslev (1986) adjusted by the unit risk premium . The leverage e ect is represented by c,
which has an e ect similar to the one of  in (1.2). However, since we have only one source of uncertainty
in GARCH, c is not the correlation. To streamline the presentation and for the sake of simplicity we will
focus on a GARCH model without leverage e ects, i.e. set c = 0. However, we did consider both GARCH
with and without leverage in our empirical work. Finally, for the GARCH model, equations (4.1) and (4.2)
suggest to lter volatility as tG , which is determined by the following recursive relationship
 S 1 2 2
1=2
G t , 1 G ^ G 2
t = ^0 + ^1 (log S , R + 2 t,1 , t,1 ) + 1 t,1 ^ G (4.3)
t,2
where the maximum likelihood estimates are: ^ = :0599, ^0 = 0:556
10,5, ^1 = 0:127, ^1 = 0:828: As is
typically the case for these models: the parameters are statistically signi cant and the models are not rejected
by the data using standard diagnostics. The residuals and the squared residuals of the GARCH model are
uncorrelated and Portmanteau tests with up to 24 lags con rm the absence of residual autocorrelation.30
Figure 4 presents the plots of the ltered volatilities. Panel (a) contrasts the BS implied volatilities and
volatilities obtained from the GARCH model according to (4.3). Since the GARCH model is speci ed for
daily observations in our case, all the parameters are daily. Hence, the ltered volatility has daily units of
measurements. To make it comparable to the BS volatilities and the SV-based volatilities, which have yearly
p
units of measurements, we have multiplied the GARCH volatility by 250.31 We see that the two lters of t
(tI and tG ) provide very di erent forecasts. Panel (b) allows us to evaluate the volatility reprojected from
the SV model estimated based on the returns data, rtR . In order to have a common benchmark of comparison
with tG , we also plot tI in this panel. We note that rtR has a much smaller amplitude in comparison to both
tI and tG . Finally, panel (b) of Figure 4 also reports the reprojection lter from the SV model estimated
with implied volatilities data, vtR : It is worth noting that, though the lters in panel (b) are obtained from
the same model speci cation with similar parameters estimates , rtR and vtR are very di erent indicating
that the key feature is not the model itself but the volatility lter it o ers. This observation motivates us to
30 In particular, the p-values for residuals are equal to 0.212 (up to 6 lags), 0.574 (up to 12 lags), 0.743 (up to 18 lags) and
0.776 (up to 24 lags). The p-values for the squared residuals are equal to 0.928, 0.991, 0.998 and 0.999 respectively.
31 We know that GARCH models do not temporally aggregate. Therefore, this operation gives us only a rough idea about

the yearly data which is adequate for the visual comparison.

27
consider all of the obtained lters in the framework of a single simple model, namely BS. The next section
proceeds with the evaluation of the option pricing quality of the lters.
Table 7, Panel A, contains the pricing errors Dap and Drp . Obviously, GARCH and SV lters are a
suboptimal schemes here as they are not used with their corresponding option pricing formula. We consider
primarily the relative errors Drp to appraise the di erences, though Table 7 reports both. We observe that the
GARCH lter tG is dominated by tI everywhere except for the two long maturity ATM (0:97 < S=K < 1:03)
groups. The best improvement of the GARCH lter over the implied volatilities is 1:79 dollars for the
slightly OTM group with relative error decreasing from 0:31 to 0:19. The best case of the implied volatility
domination over GARCH is OTM medium maturity where the price improvement is 95 cents with relative
error decreasing from 19.62 to 12.35. It should parenthetically be noted that we also studied the performance
of the GARCH model with leverage e ects. The results are not reported in the table as they are qualitatively
similar to the ones without leverage, though slightly and uniformly worse than the results obtained from
GARCH without leverage.32 The reprojection lter based on the model estimated with returns data rtR
outperforms tI only for short maturity deep OTM and for long maturity OTM and ATM calls. The best
improvement occurs again for the slightly OTM group and is equal to 2:90 dollars with the relative error
decreasing from 0:31 to 0:14. The best case where BS implied volatility dominates the returns reprojected
volatility is the slightly OTM short maturity group with a 44 cents gain (relative error changes from 12:29
to 5:78). It is easy to see that the BS { rtR speci cation uniformly dominates BS { tG with highest gain for
the long maturity deep OTM calls equal to 1:48 dollars with the relative error dropping from 16:49 to 13:63.
The lowest gain is 21 cents in the short maturity slightly ITM category. BS { vtR dominates BS { tI almost
uniformly. The exceptions are slightly ITM short and medium maturities, and all long maturities except
for deep OTM. However, in these cases di erences are not statistically signi cant. The best improvement,
which is equal to 55 cents occurs in the medium maturity deep OTM group. These observations allow us to
make some additional conclusions regarding the relative pricing performance of the four speci cations. In
particular, BS { tG (and, therefore BS { rtR ) clearly outperforms BS { vtR for long maturity ATM calls. BS
{ rtR also does better for the long maturity OTM group.
We can again conclude that the best volatility lter is vtR with the exception of pricing long maturity
and 0:94 < S=K < 1:03 calls where rtR is the best. Both are SV lters, the former, namely vtR , is relatively
straightforward. We can provide several explanations for this result. The rst is that long maturity contracts
32 This is yet another example of con icts between statistical and nancial criteria. The leverage parameter in the GARCH
model is statistically signi cant (though marginally). Yet, the resulting lter through the eyes of the BS formula worsens the
pricing performance. We already noted before, that the same observation also applies to SV models, which are rejected on
statistical grounds while GARCH models are not rejected.

28
may re ect long memory, which is not taken into an account by the models we evaluate in this paper. One
would have to consider the long memory models (for SV see e.g. Comte and Renault (1995) and Harvey
(1993) and for GARCH models see e.g. Baillie et al. (1995)). Nevertheless, one could speculate that the
SV lters considered here have greater exibility and capture some of the long memory. The fact that
rtR outperforms vtR at long maturities may be exploited by relying on Backus et al. (1997) and Das and
Sundaram (1997) who show that in the SV model kurtosis of returns increases with the time interval (or
time to maturity in the option pricing context). Therefore, the volatilities implied from the long maturity
contracts have a more pronounced smile e ect. It means that, for shorter maturities, volatilities implied from
the ATM options (the ones we used in our computation of vtR ) are quite a good proxy for all BS implied
volatilities. However, this is not the case for longer maturities, as the ATM implied volatilities may be a good
re ection of the returns kurtosis for short maturities, but they are not adequate for long maturities. Hence
we have to use the fourth power of returns to be able to adequately re ect the kurtosis in pricin g long term
options. Our results support this theoretical conjecture, especially because the SNP density involved in the
returns based reprojection required all the moments of returns up to the fourth. This nding also con rms
the results reported in the previous section which showed that bivariate returns/options lters dominate vtR
only at long maturities due to the kurtosis e ect captured by the fourth moment of returns and not present
in the ATM options at short maturities underlying vtR .

4.2 The contribution of the pricing formula

We want to investigate now whether a model generating a particular volatility lter is important for option
pricing. In particular, we address the question whether there is any pricing improvement when we use the
original, i.e. SV and GARCH pricing formula, rather than the BS one. Panel B of Table 7 reports the results.
Surprisingly, we observe that GARCH option pricing errors are dramatically and uniformly dominated by
their BS { tG counterparts. The GARCH call option value is determined as the expected value of normalized
payo at maturity under the local risk-neutral measure and can be computed via Monte-Carlo simulations.
In particular, we adopt in this paper the empirical martingale simulation strategy proposed by Duan and
Simonato (1998). The GARCH with leverage model (not reported in the table) performs roughly the same
way. Overall, we therefore need to conclude that the GARCH pricing formula does not perform very well
in terms of option pricing. The rtR lter combined with the SV option pricing formulas shows a very
di erent and opposite picture: the SV formula uniformly improves upon its BS counterpart. Contrary to
the rtR case, the volatilities-based SV lter substituted into the SV pricing formula shows a mixed picture.
In particular, the BS { vtR combination is signi cantly better in three groups: OTM short and medium
maturities and slightly OTM short maturity. These ndings prompt the question whether the suboptimal

29
BS { tG dominates GARCH option pricing and why this also happens in some of the BS { vtR cases. One
possible interpretation is that the parsimony of BS comes into play. The Black and Scholes model appears to
be so robust that, when used for instance with the GARCH volatility lter tG , it smoothes out the GARCH
de ciencies and prices considerably better.
The general conclusion we can draw from Table 7 is that the Heston SV model almost always provides
both superior volatility ltering and option pricing. Within the Heston model, the implied volatility based
lter speci cation is the best except for the long maturity options, which are better priced with rtR . On
the other hand, one can notice that the volatility ltering is much more important than the option pricing
formula the volatiltiy is subsequently substituted in. For example, the di erences in pricing errors between
BS - vtR and SV - vtR are minimal if compared with pricing errors obtained from other volatility ltering
schemes for short and medium maturity options.

Conclusion
We considered a generic procedure for estimating and pricing options in the context of stochastic volatility
models using simultaneously the fundamental price St and a set of option contracts [(itI )i=1;K ] where K  1
and itI is the Black-Scholes implied volatility. The joint estimation enabled us to recover objective and
risk neutral measures simultaneously. This approach, bridging two strands of the literature, opens up new
possibilities for comparing the information in the underlying fundamental and options data, for conducting
tests examining the transformation between the two measures and investigating the general equilibrium
underpinnings of models. Moreover, the bivariate approach involving both the fundamental and an option
appears useful for pricing derivatives when the information from the cash market provides support via the
conditional kurtosis to price options. This is the case for some long term options but the e ect is only
marginal. However, our results based on the S&P 500 index contract, showed that the univariate approach
only involving options by and large dominated for the purpose of pricing. A by-product of this nding
is that we uncover a remarkably simple volatility extraction lter based on a polynomial lag structure of
implied volatilities. Since the competing model speci cations yielded roughly the same di usion parameter
estimates, we found that the ltering of spot volatility is the key ingredient in our procedure. In fact, by
comparing various volatility lters, we found that the implied volatility based lter performs very well even
if it is used in the simple setup of the Black-Scholes formula.
At the general level our results showed that volatility lters obtained from the same model, and the same
parameter estimates, can be very di erent as alternative lters exploit di erently the information in the
panel structure of options and returns. It is clear also from our analysis that nancial criteria, like pricing

30
and hedging performance, represent the most valuable selection methods. The results in this paper also
showed that models which are rejected using statistical goodness of t criteria, in our case Heston's model,
can be the basis for a relatively e ective lter, where e ectiveness is measured with respect to pricing. On
the contrary, models that t the data well like GARCH, might turn out to be quite ine ective when it comes
to valuing derivative securities.
Ultimately, nancial and statistical criteria need to be reconciled. Heston's model has its obvious limi-
tations, since it is assumed that interest and dividend rates are constant. Several extensions of the Heston
model have been suggested which include additional state variables which determine stochastic interest rates,
dividends and jump components. In particular Bakshi and Madan (1998) and Due et al. (1998) discuss a
general framework of jump-di usions of the ane class which yield analytical option pricing formula. The
generic framework imposes a certain uniformity on the empirical studies in option pricing because the models
are easier to compare as they are nested in the general model speci cation.
We also know that the methods proposed in this paper can be improved upon in several ways in addition
to the use of more elaborate model speci cations which would reconcile nancial and statistical criteria. First,
we estimated the continuous time processes using only ATM contracts. It would be worth investigating a
setup where more than one option is used, namely a set of option contracts [(itI )i=1;K ] where some are ATM
and others are OTM. This setup will surely improve the pricing of long term OTM options. It would also be
intriguing to nd out how multivariate lters involving exclusively options do in comparison to the univariate
option-based lter. Last but not least, it would be worth exploring the result suggesting that conditional
kurtosis information in the cash market improves pricing. One could consider using LEAPS option contracts
for such an exploration.

31
 Technical Appendices
A The call options pricing using SV model
Here we provide the details of the call options pricing formula under the SV model from Bakshi et al. (1997a).
The call price in (1.3) is expressed through j ; j = 1; 2, which are equal to:
Z 1 e,ilnK fj (t; ; S (t); V (t); )
j = 12 + 1 Re i d (A.1)
0
and
    ,
f1 (t; ; S (t); V (t); ) = exp iR , 2 2 ln 1 , [ ,  + (i + 21) V ](1 , e ) ,
V
 [ ,  + (i + 1) ] + i ln S (t) + i(i + 1)(1 , e, )V (t)

V2 V 2 , [ ,  + (1 + i)V ](1 , e, ) (A.2)

where:
q
 = [ , (1 + i)V ]2 , i(i + 1)V2
     V ](1 , e,  ) ,

f2(t; ; S (t); V (t); ) = exp iR , 2 2 ln 1 , [ ,  + i2

V
 [ ,  + i ] + i ln S (t) + i(i , 1)(1 , e,  )

V2 V 2 , [ ,  + iV ](1 , e,  ) V (t) (A.3)

where:
q
= [ , iV ]2 , i(i , 1)V2

32
B SNP density estimation
SNP is a method of nonparametric time series analysis. The code can be downloaded from George Tauchen's
website http://www.econ.duke.edu/~get/snp.html. The method employs a Hermite polynomial series
expansion to approximate the conditional density of a multivariate process. An appealing feature of this
expansion is that it is a nonlinear nonparametric model that directly nests the Gaussian VAR model, the
semiparametric VAR model, the Gaussian ARCH model, and the semiparametric ARCH model. The SNP
model is tted using conventional maximum likelihood together with a model selection strategy that deter-
mines the appropriate degree of the polynomial.
The SNP method is based on the notion that a Hermite expansion can be used as a general purpose
approximation to a density function. Here it is employed in the form fk (zt j Xt ; ) / [PK (zt ; Xt)2 ](zt ),
where PK (
) denotes a multivariate polynomial of degree Kz and (z ) denotes the standard normal (possibly
multivariate) p.d.f. The process Xt is the vector of M lags of the process of interest and the index k
denotes the dimension of  which expands as the sample size increases (see discussion in section 2.1). Since
fk (z ), as the distribution proxy, has to integrate to 1, the constant of proportionality is the inverse of
R
NK (Xt ) = [PK (s; Xt )2 ](s)ds. To achieve a unique representation, the constant term of the polynomial
part is put equal to one.
First of all, let us note that if the process of interest   N (;  = RR ), where R is an upper triangular
0

matrix, then  = Rz +  and

t ; Xt )2 ](zt ) ;
fk (t j Xt ; ) = N 1(X ) [PK (jzdet (B.1)
K t (R ) j
where  = (aij ; ; R) and is estimated by QML.
The mean  at time t depends on M lags of t : Xt = b0 + BXt . RtX also depends on MR lags of
the process of interest in case of ARCH leading term. A Hermite polynomial is used to expand the density
around the leading Gaussian density, namely:
0 KX 1
X
Kz X
Kz X
@
PK (zt ; Xt ) = ai (Xt )z i = aij Xtj A z i ; (B.2)
i=0 i=0 j =0
and, as discussed above, we set a00 equal to 1 for the uniqueness of the representation. The subscript K
denotes the vector (KX ; Kz ; M ). According to Andersen and Lund (1996), if the leading term is speci ed
carefully, M = 1 generally suces in the univariate SNP setting.
When Kz is put to zero, one obtains normal distributions. When Kz is positive and KX = 0, one obtains
a density which is conditionally homogeneous with respect to X . Finally, when both Kz and KX are positive,
we approximate over conditionally heterogeneous class of densities.

33
 The optimal values of K (as well as M ; MR ) are chosen based on standard criterions and tests, such
as AIC, BIC and HQ model selection criteria. Gallant and Tauchen (1993) provide a table which matches
values of the parameters K; M ; MR and popular time series models. We reproduce this table here (table
1B). Typically, BIC is used to select the model speci cation. We follow this strategy in Table 1 Panels A
and C.

34
C Details of the SDE Discretization Scheme
The discussion in this appendix is based on Kloeden and Platen (1995) to which we refer the reader for
additional details.
The EMM estimation procedure requires the process under study to be simulated. It is impossible to
simulate the actual process if it does not have an analytical solution, which is the case for Heston's model.
Therefore, to simulate we need to discretize the process in (1.1)-(1.2). The discretizations are based on
the truncation of the Ito-Taylor expansion formula. The more terms we keep, the more accurate is the
expansion. The Ito-Taylor formula also involves the derivatives of the drift and di usion coecients with
respect to the process of interest. Hence an approximation of these derivatives maybe advisable to improve
the computational eciency. The discretization schemes mainly di er in the number of terms that one keeps
and the way one approximate (or does not approximate) the above mentioned derivatives. We choose to
work with the explicit order 2.0 weak discretization scheme for the following reasons.
Each strong scheme of order p is a weak scheme of order 2p and schemes are called weak because they
approximate moments rather than sample paths of a di usion (which the strong schemes approximate).
Moreover, strong schemes take longer time to simulate and since we match moments instead of sample paths
in our estimation we use a weak scheme. The order of the scheme determines the quality of an approximation.
We say that a discrete approximation Y (
t) = (Sn ; Vn ) of the process X = (S (t); V (t)) converges weakly
0 0

with order p > 0 to X at time T as
t ! 0 if for each moment M there exists a positive constant C which
does not depend on
t, and a nite
0 t such that

jM (XT ) , M (YT (
t))j  C
tp (C.1)

for any
t 2 (0;
0 t). In our case p = 2, which yields a faster convergence than the commonly used Euler
discretization, a weak scheme of order 0:5.33 Kloeden and Platen show that under standard regularity
conditions the moments of the di usions exist provided the initial value is a constant (see Theorem 4.5.4).
The discretization scheme guarantees that the moments of the discretized processes also exist. Therefore, it
is legitimate to implement the moments matching procedure.
The term \explicit" in the discretization scheme refers to the derivatives approximation method which
works similary to the Runge-Kutta methods in case of deterministic di erential equations. Since this method
imposes less computational burdens than the implicit scheme, it is often preferable to use.
This approach also allows us to control for the non-synchronicity bias we discussed in the data section.
We can show that by matching moments instead of sample paths the error introduced by the di erential in
33 Note that, following the tradition in the option pricing literature, we estimate the annualized parmeters, i.e. take
t =
1=252.

35
stock prices over the 15 minute closing interval has a negligible impact on the estimation of the moments.
We simulate the fundamental process as being observed at 4:00 pm, or time T , and we use these simulated
values to substitute in the implied volatility function, I (St ; Vt ), implicitly based on the option price which
is observed at 4:15 pm, or time T + t. We will use the Ito-Taylor expansion formula to compute the error
introduced by such an operation. The EMM procedure matches the moments of the function I (
). Therefore,
we have to assess the size of the error for all possible moments E (I n (
)). Since it is a suciently smooth
function, we can approximate it with polynomials based on (St ; Vt ) to any degree of accuracy. Hence, in our
0

discussion, we will consider E (f (St ; Vt )) for any smooth function f . In particular, according to the theorem
5.5.1 of Kloeden and Platen (1995), we can write:
Z T +t Z T +t
f (ST +t ; VT +t ) = f (ST ; VT ) + f(0) (ST ; VT ) du + f(1)(ST ; VT ) dWS (u)
T T
Z T +t Z T +t Z T +t
+ f(2) (ST ; VT ) dWV (u) + f(1;1)(ST ; VT ) dWS (u)dWS (v)
T T T
Z T +t Z T +t
+ f(2;2) (ST ; VT ) dWV (u)dWV (v)
T T
Z T +t Z T +t
+ f(1;2) (ST ; VT ) dWS (u)dWV (v)
ZTT +t ZTT +t
+ f(2;1) (ST ; VT ) dWS (u)dWV (v) + o(t) (C.2)
T T
where f(
) are the Ito coecient functions (Kloeden and Platen (1995), 5.3). As we already mentioned, we
are interested in the accuracy of the moments, i.e. we have to compute E (f (ST +t ; VT +t )):

E (f (ST +t ; VT +t )) = E (f (ST ; VT )) + E (f(0) (ST ; VT ))t + o(t)

@ @ + ( , V ) @
= E (f (ST ; VT )) + E @t + S S @S @V
2 2 2 
@ +  V S @ + 1 2 V @ f (S ; V )t + o(t)
+ 12 V S 2 @S 2 V @S@V 2 V @V T T
= E (f (ST ; VT )) + o(t) (C.3)

The last equality follows from the fact that the expected value of the in nitesimal generator f(0) (ST ; VT ) of
the process (1.1)-(1.2) is equal to zero (see, for instance, Hansen and Scheinkman (1995)). Since the interval
t is relatively small (it is 1=96th of the simulation interval
t), substituting f (ST +t ; VT +t ) by f (ST ; VT )
in the implied volatility moments matching procedure introduces only a negligable error.

36
References
[1] At-Sahalia, Y. and A. Lo (1998): Nonparametric Estimation of State-Price Densities Implicit in Fi-
nancial Prices, Journal of Finance, 53, 499-548.
[2] At-Sahalia, Y., Y. Wang and F. Yared (1997): Do options markets correctly price the probabilities of
movements of the underlying asset?, Discussion paper, University of Chicago.
[3] Andersen, T.G., H.-J. Chung and B.E. Srensen (1997): Ecient Method of Moments Estimation of a
Stochastic Volatility Model: A Monte Carlo Study, Discussion Paper, Northwestern University.
[4] Andersen, T. and J. Lund (1996): The Short Rate Di usion Revisited: An Investigation Guided by the
Ecient Method of Moments. Discussion Paper, Northwestern University.
[5] Andersen, T.G. and J. Lund (1997): Estimating continuous-time stochastic volatility models of the
short-term interest rate. Journal of Econometrics, 77, 343-377.
[6] Bakshi, G., C. Cao and Z. Chen (1997a): Empirical performance of alternative option pricing models,
Journal of Finance, 52, 2003-2049.
[7] Bakshi, G., C. Cao and Z. Chen (1997b): Can Markovian models explain option price dynamics? Lessons
from high-frequency option data, Discussion paper, Penn State University.
[8] Bakshi, G. and D. Madan (1998): Spanning and Derivative-Security Valuation, Journal of Financial
Economics, (forthcoming)
[9] Bates, D.S. (1996): Testing Option Pricing Models. In Maddala, G.S. and C.R. Rao (eds.), Handbook
of Statistics, Vol. 14, Elsevier Science B.V.
[10] Black, F. (1976): Studies in Stock Price Volatility Changes, Proceedings of the 1976 Business Meeting
of the Business and Economic Statistics Section, American Statistical Association, 177-181.
[11] Black, F. and M.S. Scholes (1973): The pricing of options and corporate liabilities, Journal of Political
Economy, 81, 637-659.
[12] Breeden, D. and R. Litzenberger (1978): Prices of State Contingent Claims Implicit in Option Prices,
Journal of Business, 51, 621-652.
[13] Broadie, M., J. Detemple, E. Ghysels and O. Torres (1997): American Options with Stochastic Volatility
and Dividends: A Nonparametric Approach, Journal of Econometrics, (forthcoming).

37
[14] Campbell, J., A. Lo and C. MacKinlay (1997): The Econometrics of Financial Markets, Princeton
University Press.
[15] Canina, L. and S. Figlewski (1993): The informational content of implied volatility, Review of Financial
Studies, 6, 659-681.
[16] Chernov, M. and E. Ghysels (1998): What Data Should be Used to Price Options, Discussion Paper
CIRANO, available at http://ftp.cirano.umontreal.ca/pub/publication/98s-22.pdf.zip.
[17] Comte, F. and E. Renault (1995): Long Memory Continuous Time Stochastic Volatility Models, Paper
presented at the HFDF-I Conference, Zurich
[18] Cox, J., J. Ingersoll and S. Ross (1985): A theory of the term structure of interest rates, Econometrica,
53, 385-408.
[19] Christensen, B. J. and N. R. Prabhala (1998): The relation between implied and realized volatility,
Journal of Financial Economics, 50, 125-150.
[20] Das, S. and R. Sundaram (1997): Of Smiles and Smirks: A Term-Structure Perspective, Working Paper,
Harvard University
[21] Day, T. and C. Lewis (1992): Stock Market Volatility And the Information Content of Stock Index
Options, Journal of Econometrics, 52, 267-287
[22] de Jong, Frank (1997): Time-series and Cross-section Information in Ane Term Structure Models,
Working paper, Tilburg University
[23] Duan, J.C. (1995): The GARCH option pricing model, Mathematical Finance, 5, 13-32
[24] Duan, J.C. and J.G. Simonato (1998): Empirical Martingale Simulation for Asset Prices, Management
Science, 44, 1218-1233.
[25] Due, D. and R. Kan (1996): A Yield-Factor Model of Interest Rates, Mathematical Finance, 6, 379-406
[26] Due, D. and J. Pan and K. Singleton (1998): Transform Analysis and Option Pricing for Ane
Jump-Di usions, Working Paper, Stanford University.
[27] Dumas, B., J. Fleming and R.E. Whaley (1998): Implied volatility functions: Empirical Tests, Journal
of Finance, 53, 2059-2106.
[28] Fenton, V.M. and A.R. Gallant (1996): Qualitative and Asymptotic Performance of SNP Density Esti-
mators, Journal of Econometrics 74, 77-118.

38
[29] Fleming, J. (1994): The Quality of Market Volatility Forecasts Implied by S&P 100 Index Option Prices,
Discussion Paper, Rice University
[30] Figlewski, S. (1998): Derivatives Risks, Old and New, Discussion Paper, New York University.
[31] Gallant, A.R., D. Hsieh and G. Tauchen (1997): Estimation of stochastic volatility models with diag-
nostics, Journal of Econometrics, 81, 159-192.
[32] Gallant, A.R., C.-T. Hsu and G.Tauchen (1998): Calibrating Volatility Di usions and Extracting Inte-
grated Volatility, Discussion Paper, Duke University.
[33] Gallant, A.R. and Jonathan R.Long (1997): Estimating Stochastic Di erential Equations Eciently by
Minimum Chi-Square, Biometrika, 84, 125-141.
[34] Gallant, A.R. and G. Tauchen (1989): Seminonparametric Estimation of Conditionally Constrained
Heterogeneous Processes: Asset Pricing Applications, Econometrica, 57, 1091-1120.
[35] Gallant, A.R. and G. Tauchen (1993): SNP: A Program for Nonparametric Time Series Analysis, version
8.3, User's Guide, Discussion paper, University of North Carolina at Chapel Hill.
[36] Gallant, A.R. and G. Tauchen (1996): Which Moments to Match? Econometric Theory, 12, 657-681.
[37] Gallant, A.R. and G. Tauchen (1997): EMM: a program for ecient method of moments estimation,
Version 1.4, User's guide, Discussion paper, University of North Carolina at Chapel Hill.
[38] Gallant, A.R. and G. Tauchen (1998): Reprojecting Partially Observed Systems with Application to
Interest Rate Di usions, Journal of American Statistical Association, 93, 10-24.
[39] Garman, M. (1976): A General Theory of Asset valuation Under Di usion State Processes, Working
Paper, UC Berkeley
[40] Ghysels, E., A. Harvey and E. Renault (1996): Stochastic Volatility. In Maddala, G.S. and C.R. Rao
(eds.): Handbook of Statistics, Vol. 14, Elsevier Science B.V.
[41] Ghysels, E. and J. Jasiak (1996): Stochastic Volatility and Time Deformation, Discussion Paper
CIRANO.
[42] Ghysels, E., V. Patilea, E. Renault and O. Torres (1998), Non-parametric Methods and Option Pricing,
in D.J. Hand and S. D. Jacka (eds.): Statistics in Finance, John Wiley, 261-282.
[43] Hansen, L. (1982): Large Sample Properties of Generalized Method of Moments Estimators, Economet-
rica, 50, 1029-1054.

39
[44] Hansen, L. and J. Scheinkman (1995): Back to the Future: Generating Moment Implications for
Continuous-Time Markov Processes, Econometrica, 63, 767-804
[45] Harrison, M. and D. Kreps (1979): Martingales and Arbitrgae in Multiperiod Securities Markets, Journal
of Economic Theory, 20, 381-408.
[46] Harrison, M. and S. Pliska (1981): Martingales and Stochastic Integrals in the Theory of Continuous
Trading, Stochastic Processes and Their Applications, 11, 215-260.
[47] Harvey, A.C., E. Ruiz and N. Shephard (1994): Multivariate Stochastic Variance Models, Review of
Economic Studies 61, 247-264.
[48] Harvey, C.R. and R.E. Whaley (1991): S&P 100 Index Option Volatility, Journal of Finance, 46,
1551-1561
[49] Heston, S.L. (1993): A Closed-Form Solution for Options with Stochastic Volatility with Applications
to Bond and Currency Options, Review of Financial Studies, 6, 327-343.
[50] Hull, J. and A. White (1987): The pricing of options on assets with stochastic volatilities, Journal of
Finance, 42, 281-300.
[51] Jacquier, E. and R. Jarrow (1998): Model Error in Contingent Claim Models: Dynamic Evaluation,
Journal of Econometrics, (forthcoming).
[52] Jacquier, E., N.G. Polson and P.E. Rossi (1994): Bayesian Analysis of Stochastic Volatility Models
(with discussion): Journal of Business and Economic Statistics 12, 371-417.
[53] Jacquier, E., N.G. Polson and P.E. Rossi (1995): Stochastic Volatility: Univariate and Multivariate
Extensions, Discussion Paper, Boston College.
[54] Jiang, G. and P. van der Sluis, Pieter (1998): Pricing Stock Options under Stochastic Volatility and
Interest Rates with Ecient Method of Moments Estimation, Discussion Paper, University of Groningen.
[55] Ingersoll, J. (1987): Theory of Financial Decision Making, Rowman and Little eld
[56] Karatzas, I. and S. Shreve (1991): Brownian Motion and Stochastic Calculus, 2nd edition, Springer
[57] Kloeden, P.E. and E. Platen (1995): Numerical Solution of Stochastic Di erential Equations, Springer
Verlag.
[58] Lamoureux, C.G. and W.D. Lastrapes (1993): Forecasting stock-return variance: Toward an under-
standing of stochastic implied volatilities, Review of Financial Studies, 6, 293-326.

40
[59] Melino, A. (1994): Estimation of Continuous-time Models in Finance. In Sims, C.A. (eds.): Advances
in Econometrics, Cambridge University Press.
[60] Nandi, S. (1998): How important is the correlation between returns and volatility in a stochastic
volatility model? Emprical evidence from pricing and hedging in the S&P 500 index options market,
Journal of Banking and Finance, 22, 589-610
[61] Nelson, D.B. and D.P. Foster (1994): Asymptotic Filtering Theory for Univariate ARCH Models, Econo-
metrica 62, 1-41.
[62] Nelson, D.B. (1996): Asymptotic Filtering Theory for Multivariate ARCH Models, Journal of Econo-
metrics 62, 1-41.
[63] Pastorello, S., E. Renault and N. Touzi (1994): Statistical inference for random variance option pricing,
Jornal of Business and Economic Statistics, forthcoming.
[64] Renault, E. (1997): Econometric models of option pricing errors. In Kreps, D.M. and K.F. Wallis (eds.):
Advances in economics and econometrics: Theory and applications, Seventh World Congress, Vol. 3,
Cambridge University Press.
[65] Renault, E. and N. Touzi (1996): Option Hedging and Implied Volatilities in a Stochastic Volatility
Model, Mathematical Finance, 6, 279-302.
[66] Romano, M. and N. Touzi (1993): Contingent Claims and Market Completeness in a Stochastic Volatility
Model, Discussion paper, CREST, Paris.
[67] Rubinstein, M. (1994): Implied binomial trees, Journal of Finance, 49, 771-818.
[68] Schwarz, G. (1978): Estimating the Dimension of the Model, Annals of Statistics, 6, 461-464.
[69] Scott, L. (1987): Option Pricing when the Variance Changes Randomly: Theory, Estimation, and an
Application, Journal of Financial and Quantitative Analysis, 22, 419-438.
[70] Tauchen, G. (1997): New Minimum Chi-square methods in empirical nance. In Kreps, D.M. and
K.F. Wallis (eds.): Advances in economics and econometrics: Theory and applications, Seventh World
Congress, Vol. 3, Cambridge University Press.
[71] Wiggins, J. B. (1987): Option values under stochastic volatility: Theory and empirical estimates,
Journal of Financial Economics, 19, 351-372.

41
 Figures and Tables
Figure 1
The univariate series and SNP densities
We estimate the SNP density for the two univariate types of data: (i) the log-returns on the S&P500; (ii) the log of
the BS implied volatilities of the closest to maturity and the money call options. The data are daily and span the
period from No vember 1985 to October 1993. The plots to the left are the time series of the data, the plots to the
right are the estimated densities of the series. The solid line is a plot of an SNP t, the dashed line is normal density
with the same mean and variance .

42
 g

0.10
0.05

80
0.0

60
returns

-0.10

40
20
-0.20

0
1986 1988 1990 1992 1994 -0.02 0.0 0.02

returns
(a) (b)
-0.5

3
-1.0
-1.5

2
volatility

-2.0

1
-2.5
-3.0

1986 1988 1990 1992 1994 -3.0 -2.5 -2.0 -1.5 -1.0

volatility
(c) (d)
 Figure 2
The bivariate SNP density

We estimate the joint SNP density for the following series: (i) the log-returns on the S&P500; (ii)
the log of the BS implied volatilities of the closest to maturity and the money call options. The
data are daily and span the period from November 1985 to October 1993. (a) is the perspective
plot of the estimated bivariate density; (b) is the contour plot at quantiles 10%, 25%, 50%, 75%,
90%, and 95%; (c) and (d) are the marginal densities of (i) and (ii) correspondingly, the solid line
is a plot of an SNP t, the dashed line is normal density with the same mean and variance.

43
 (a)
50 100 150 200

0.02
returns

0.0
0

0.0
30
.02
0.0
1 -1
0 -1.5

-0.02
re
tur -0.0 -2
ns 1-0
.02 -2.5 tility
-0
.03 -3 vola

-3.0 -2.5 -2.0 -1.5 -1.0

volatility
(b)
2.5
60

2.0
1.5
40

1.0
20

0.5
0.0
0

-0.02 0.0 0.02 -3.0 -2.5 -2.0 -1.5 -1.0

returns volatility
(c) (d)
 Figure 3
The market price of risk
The estimation of the SV model based on bivariate data allows to compute the market price of risk (t). We plot these prices for
the entire estimation sample. Panel (a) reports 1 (t) - the market price of the asset risk, see (2.4), panel (b) reports 2 (t) - the
market price of the volatility risk, see (2.5) and panel (c) is the time series of the reprojected volatility V (t) which was used to
compute 1 (t) and 2 (t).
44
1.5
0.5

1986 1987 1988 1989 1990 1991 1992 1993 1994

(a)
0 50
-100
-200

1986 1987 1988 1989 1990 1991 1992 1993 1994

(b)
0.08
0.04
0.0

1986 1987 1988 1989 1990 1991 1992 1993 1994

(c)
 Figure 4
The ltered volatilities
We plot the volatilities ^ which were estimated from di erent models. Panel (a) shows the volatility obtained from the GARCH
t
model and the previous day Black-Scholes implied volatility. Panel (b) shows the volatilty reprojected from the SV model based
on the returns series and on the implied volatilities series. It also shows the implied volatilities to have a common benchmark
with the GARCH volatility plot in (a).
45
 0.20

GARCH
BS
0.15
volatilities

0.10
0.05
0.0

1993 1993 1993 1993 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994
Nov Nov Dec Dec Jan Feb Feb Mar Apr Apr May May Jun Jul Jul Aug Sep Sep Oct Oct

(a)

SV-ret
0.20

BS
SV-vol
0.15
volatilities

0.10
0.05
0.0

1993 1993 1993 1993 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994 1994
Nov Nov Dec Dec Jan Feb Feb Mar Apr Apr May May Jun Jul Jul Aug Sep Sep Oct Oct

(b)
 Table 1
The SNP density estimation
We estimate the SNP density for the three types of data: (i) the log-returns on the S&P500;
(ii) the log of the BS implied volatilities of the closest to maturity and the money call options; (iii)
both series. Panel A reports the structure of the estimated densities and the values of the objective
function sn based on the density in (B.1), the values of the Akaike information criterion (AIC), the
Hannan and Quinn criterion (HQ) and the Schwarz Bayes information criterion (BIC). Each line
report the best BIC model within each class of density structures. The overall best model selected
by BIC is in boldface. Panel B reports possible densities structures, as described in Gallant and
Tauchen (1993). Panel C reports the reprojection SNP densities for spot volatility estimated with
simulated data from Heston's model with the data of types (i)-(iii) in the conditioning set.
Panel A
Data type Kz KX M M MR (^ )
sn AIC HQ BIC
S&P 500 0 0 1 3 0 1.4159 1.4185 1.4211 1.4256
0 0 1 3 8 1.2140 1.2207 1.2275 1.2392
9 0 1 3 8 1.1421 1.1534 1.1649 1.1848
9 1 1 3 8 1.1371 1.1535 1.1703 1.1991
BS vol's 0 0 1 9 0 1.0836 1.0893 1.0950 1.1050
0 0 1 9 5 0.9510 0.9592 0.9676 0.9821
6 0 1 9 5 0.8183 0.8296 0.8411 0.8610
6 3 1 9 5 0.7889 0.8109 0.8335 0.8723
Joint 0 0 1 4 0 2.4969 2.5077 2.5187 2.5376
0 0 1 4 4 2.1626 2.1856 2.2092 2.2498
4 0 1 4 4 1.9942 2.0214 2.0492 2.0970

Panel B
Parameter K; M ; MR setting Characterization of t
Kz = 0; KX = 0; M  0; M = 0; MR = 0 iid Gaussian
Kz = 0; KX = 0; M  0; M > 0; MR = 0 Gaussian VAR
Kz > 0; KX = 0; M  0; M > 0; MR = 0 non-Gaussian VAR, homog. innov.
Kz = 0; KX = 0; M  0; M  0; MR > 0 Gaussian ARCH
Kz > 0; KX = 0; M  0; M  0; MR > 0 non-Gaussian ARCH, homog. innov.
Kz > 0; KX > 0; M > 0; M  0; MR  0 general non-linear process, heterog. innov.

46
 Panel C
Data type Kz KX M M MR (^ )
sn AIC HQ BIC
S&P 500 0 0 1 1 0 0.4601 0.4604 0.4607 0.4615
4 0 1 1 0 0.4058 0.4065 0.4073 0.4090
4 1 1 1 0 0.3859 0.3871 0.3886 0.3915
BS vol's 0 0 1 22 0 -0.8367 -0.8343 -0.8313 -0.8256
1 0 1 22 0 -0.7889 -0.7864 -0.7833 -0.7773
Joint 0 0 1 7 0 -25.2651 -25.2618 -25.2578 -25.2498
8 0 1 7 0 -30.5975 -30.5925 -30.5865 -30.5747
8 6 1 7 0 -30.6031 -30.5775 -30.5464 -30.4855

47
 Table 2
EMM estimation results
We have estimated the stochastic volatility model
dS (t) q
S (t)
= Rdt + V (t)dWS (t);

q q q
dV (t) = ( ,  V (t))dt + V 1 , 2 V (t)dWV (t) + V  V (t)dWS (t);
where the model is stated under the risk-free probability measure. We have denoted by asterisks only those parameters, which change
when we make a transition to the objective probability measure. With such a change R would have to be substituted by S and all
asterisks removed. Three types of data were used in this estimation: (i) the log-returns on the S&P500; (ii) the log of the BS implied
volatilities of the closest to maturity and the money call options; (iii) both series. Standard errors are reported in the parentheses,
z -statisticstesting the model are in the brackets, { denotes unidenti able parameters. The reported parameter estimates are annualized.
Data type ^
 S ^ ^ ^
 ^
 ^V
 ^
S&P 500 0.12767392 0.01518484 { 0.92586310 { 0.06340828 -0.01787568
[6:863] (0.00000084) (0.00514937) (0.29340050) (0.00814696) (0.24603696)
Vol { 0.01546074 0.00856834 0.92803334 0.70200229 0.06459151 -0.01939315
[247:447] (0.00017247) (0.00001727) (0.00003542) (0.00019527) (0.00028325) (0.00032025)
Joint 0.10190398 0.01430366 0.00659546 0.93086335 0.69011763 0.06150001 -0.01830000
[90:328] (0.00007322) (0.00005969) (0.00007541) (0.00009669) (0.00026902) (0.00007698) (0.00009685)
47
 Table 3
The number of observations in the options groups used for pricing/hedging
performance evaluations
To assess pricing and hedging performance of alternative model speci cations we separate the calls into
twelve groups by moneyness and time to maturity. The table reports number of observations in each group.
The total number of observations is reported in the headers.
Pricing (7424) Hedging (7132)
Moneyness Days to expiration Days to expiration
<60 60-180 180 <60 60-180 180
<0.94 337 448 313 299 437 279
0.94-0.97 1080 828 168 1037 818 159
0.97-1.00 1377 938 195 1324 930 187
1.00-1.03 1038 612 90 975 604 83

48
 Table 4
The SV model pricing errors
We consider post estimation sample data from 11:93 to 10:94 to compute out-of-sample pricing errors. We evaluate three model speci cations by
moneyness and time to maturity. The rst is \BS ", which involves pricing options with the Black-Scholes model and a volatility estimate based on the
previous day at-the-money option contract. Next is the speci cation denoted \V ol" in the table, which involves Heston's call price model estimated
using a univariate speci cation involving options. The speci cation denoted \Joint" corresponds to the bivariate speci cation. The rst column is
the moneyness category. The next column lists the model specifcations. The third column reports the absolute standard error Dap (see (3.1)) and
the relative standard error Drp (see (3.2)) by each maturity group. The remaining
,
columns report the 12 -statistics


from the GMM overidentifying
restrictions tests. These tests are aimed to determine whether the means of (Citobserved , Citmodel )=(Citobserved ) 2 computed from a pair of model
estimates are signi cantly di erent. We report results of the tests within each maturity/moneyness group and overall across all maturities for each
moneyness.
Pricing Errors Tests All
Moneyness Model Days to expiration Models Days to expiration groups
<60 60-180 180 <60 60-180 180 test
Dap Drp Dap Drp Dap Drp
BS 0.54 4.62 1.56 20.71 3.12 4.93 Vol-BS 3:61 5:79 0.29 16:81
<0.94 Vol 0.42 2.01 0.92 12.58 2.26 4.06 Vol-Joint 2.69 10:40 1.32 24:01
Joint 0.44 2.42 1.23 17.35 2.74 4.80 Joint-BS 2:81 1.30 0.01 5:93
BS 0.82 12.12 1.88 12.35 4.62 0.45 Vol-BS 10:95 4:17 17:64 16:81
0.94-0.97 Vol 0.58 8.39 1.38 7.63 3.41 0.33 Vol-Joint 15:90 5:42 11:33 24:01
Joint 0.73 10.79 1.66 9.91 3.86 0.37 Joint-BS 2.10 2.27 8:20 5:93
BS 1.13 5.78 2.48 0.38 5.23 0.31 Vol-BS 5:72 19:63 12:28 16:81
0.97-1.00 Vol 1.05 4.46 2.20 0.30 4.24 0.24 Vol-Joint 4:93 38:20 18:07 24:01
Joint 1.22 4.93 2.48 0.36 4.71 0.27 Joint-BS 3:08 1.38 5:78 5:93
BS 1.25 0.14 2.31 0.17 4.63 0.18 Vol-BS 0.07 3:04 8:16 16:81
1.00-1.03 Vol 1.35 0.15 2.33 0.16 3.85 0.14 Vol-Joint 11:31 13:49 4:19 24:01
Joint 1.40 0.16 2.43 0.17 4.21 0.16 Joint-BS 4:42 0.07 3:24 5:93
 indicates signi cance at the 1% level;  indicates signi cance at the 5% level;  indicates signi cance at the 10% level.
49
 Table 5
The SV model hedging errors
We consider post estimation sample data from 11:93 to 10:94 to compute out-of-sample hedging errors. We evaluate three model speci cations by
moneyness and time to maturity. The rst is \BS ", which involves pricing options with the Black-Scholes model and a volatility estimate based on the
previous day at-the-money option contract. Next is the speci cation denoted \V ol" in the table, which involves Heston's call price model estimated
using a univariate speci cation involving options. The speci cation denoted \Joint" corresponds to the bivariate speci cation. The rst column is
the moneyness category. The next column lists the model specifcations. The third column reports the absolute standard error Dah (see (3.9)) and
the relative standard error Drh (see (3.11)) by each maturity group. The remaining

columns report the 12 -statistics from the GMM overidentifying
2
restrictions tests. These tests are aimed to determine whether the means of (Cashitobserved , Cashitmodel )=(Cashitobserved ) computed from a pair of
model estimates are signi cantly di erent. We report results of the tests within each each maturity/moneyness group and overall across all maturities
for each moneyness.
Hedging Errors Tests All
Moneyness Model Days to expiration Models Days to expiration groups
<60 60-180 180 <60 60-180 180 test
Dah Drh Dah Drh Dah Drh
BS 2.35 0.968 1.33 0.875 1.35 0.067 Vol-BS 0.34 0.01 3:44 0.89
<0.94 Vol 2.43 0.751 1.33 0.881 1.37 0.029 Vol-Joint 0.81 2.24 1.02 1.07
Joint 2.44 1.525 1.37 1.396 1.36 0.092 Joint-BS 0.47 2.18 0.30 1.10
BS 2.04 1.202 1.10 0.804 1.51 0.024 Vol-BS 0.05 0.03 1.76 0.89
0.94-0.97 Vol 2.06 1.064 1.07 0.718 1.62 0.009 Vol-Joint 1.24 0.99 3:75 1.07
Joint 2.07 0.470 1.07 0.053 1.61 0.010 Joint-BS 1.65 1.10 1.67 1.10
BS 2.26 5.368 1.30 0.013 2.37 0.012 Vol-BS 0.87 1.99 1.07 0.89
0.97-1.00 Vol 2.25 1.570 1.30 0.007 2.34 0.009 Vol-Joint 1.00 4:81 1.77 1.07
Joint 2.22 0.066 1.30 0.008 2.35 0.009 Joint-BS 1.04 1.69 1.12 1.10
BS 1.75 0.005 1.67 0.005 3.97 0.012 Vol-BS 0.01 2.38 1.74 0.89
1.00-1.03 Vol 1.77 0.005 1.66 0.005 3.38 0.011 Vol-Joint 1.01 7:20 1.45 1.07
Joint 1.77 0.005 1.67 0.005 3.54 0.011 Joint-BS 0.29 0.19 1.39 1.10
 indicates signi cance at the 1% level;  indicates signi cance at the 5% level;  indicates signi cance at the 10% level.
50
 Table 6
Reprojection Model with Implied Volatilities
The coecients of the reprojection model used to lter the latent volatility process from the observed
past implied volatilities are listed in the table. We also report parameters of AR(22) model for the logarithm
of implied volatilities.
Lags Reprojection AR(22) model
parameters std. err. parameters std.err.
0 -0.16320 0.01043 -1.00038 0.05046
1 0.24514 0.00149 0.45085 0.06300
2 0.20591 0.00144 0.15847 0.06863
3 0.18804 0.00147 -0.03892 0.06936
4 0.16293 0.00154 0.10117 0.06895
5 0.14055 0.00167 0.00499 0.06893
6 0.11489 0.00184 0.00697 0.06887
7 0.09978 0.00203 0.04363 0.06875
8 0.08792 0.00225 -0.04505 0.06827
9 0.07836 0.00253 -0.01029 0.06835
10 0.06675 0.00285 0.01011 0.06821
11 0.05675 0.00321 0.02365 0.06805
12 0.05069 0.00337 -0.08067 0.06818
13 0.04506 0.00365 0.07710 0.06840
14 0.04715 0.00384 0.01280 0.06865
15 0.04298 0.00405 -0.13689 0.06861
16 0.04179 0.00449 0.07321 0.06911
17 0.04882 0.00468 0.05775 0.06927
18 0.04386 0.00503 -0.10836 0.06954
19 0.03987 0.00529 0.12760 0.06964
20 0.03976 0.00554 0.01582 0.07011
21 0.03277 0.00587 -0.13449 0.06957
22 0.03280 0.00629 -0.00279 0.06478

51
 Table 7
Volatility lter and model contributions to option pricing
In this table the role of volatility lter and particular model speci cation in option pricing is unraveled and evaluated separately. In Panel A we consider the
pricing performance of various volatility lters by substituting them into the Black-Scholes option valuation formula. In panel B we assess whether the pricing
formula associated with a particular volatility lter improves option pricing as compared to the Black-Scholes formula reported in Panel A. We consider three
models. The rst is \BS ", which is the Black-Scholes model, the second is \GARCH " as speci ed in Duan (1995), and nally \SV " is the stochastic volatility
model of Heston (1993). The volatility lters considered are: tI { the estimate of today's volatility by the previous day's \BS " implied ATM option; tG { the
estimate of instantaneous volatility using the \GARCH " model; rtR { the estimate of instantaneous volatility via reprojection from \SV " using a univariate
scheme based on returns; vtR is the estimate of instantaneous volatility via reprojection from \SV " using a univariate scheme based on BS implieds. In
panels A nad B the rst column is the moneyness category. The next column in Panel A lists the lters which are used in conjuction with the BS formula,
while the second column in Panel B lists the lter/pricing formula combinations. The remaining columns are similar to the ones in table 4. Namely, they
report the absolute, Dap , and relative, Drp , errors by moneyness and maturity categories. GMM overidentifying restrictions tests, which determine whether
the means of the relative pricing errors computed from a pair of models are signi cantly di erent, are reported next. In Panel A, pairwise tests within each
moneyness-maturity group for all possible lters combinations (omitting repetitions) and overall groups tests are reported via the 12 - statistics. The tests in
Panel B compare the performance of the lter/pricing formula combinations and their BS formula counterparts from Panel A.
The percentiles of the 12 distribution are: 90th { 2.71, 95th { 3.84, 99th { 6.63.
52
 Panel A
Pricing Errors Pairwise Tests All
Moneyness ^t Days to expiration Days to expiration groups
<60 60-180 180 <60 60-180 180 tests
Dap Drp Dap Drp Dap Drp tI tG rtR tI tG rtR tI tG rtR tI tG rtR
tI 0.54 4.62 1.56 20.71 3.12 4.93
tG 0.67 6.71 2.35 33.15 3.77 16.49 3.64 10.29 4.86 42.84
<0.94 rtR 0.44 3.42 1.51 22.89 2.29 13.63 2.37 5.09 0.57 12.68 4.53 3.20 27.26 18.99
vtR 0.42 2.03 1.01 12.61 3.03 3.65 3.53 4.53 2.29 5.61 18.62 13.47 0.44 4.45 3.91 17.51 53.69 51.19
tI 0.82 12.12 1.88 12.35 4.62 0.45
tG 1.39 21.57 2.83 19.62 4.09 0.45 30.43 9.30 0.00 42.84
0.94-0.97 rtR 1.08 18.34 2.13 18.93 2.93 0.34 22.87 14.64 7.64 0.49 11.95 8.88 27.26 18.99
vtR 0.58 8.17 1.57 7.27 4.65 0.46 11.74 37.96 38.14 4.59 15.49 15.89 0.07 .01 11.05 17.51 53.69 51.19
tI 1.13 5.78 2.48 0.38 5.23 0.31
tG 1.84 12.78 2.66 0.53 3.44 0.19 22.01 20.35 18.09 42.84
0.97-1.00 rtR 1.57 12.29 2.19 0.51 2.33 0.14 18.47 1.75 8.19 1.33 28.24 10.06 27.26 18.99
vtR 1.10 4.37 2.60 0.35 5.79 0.33 6.09 22.10 18.59 2.91 20.41 9.03 0.66 19.45 26.37 17.51 53.69 51.19
tI 1.25 0.14 2.31 0.17 4.63 0.18
tG 1.56 0.23 2.29 0.18 2.87 0.11 29.24 1.87 11.07 42.84
1.00-1.03 rtR 1.35 0.21 1.95 0.16 2.18 0.09 19.29 9.88 0.79 8.12 16.83 2.74 27.26 18.99
vtR 1.39 0.15 2.61 0.18 4.72 0.18 0.76 22.51 13.27 1.16 0.24 2.24 0.02 11.36 16.88 17.51 53.69 51.19
53
 Panel B
Pricing Errors Tests: Model vs BS All
Moneyness Model { ^t Days to expiration Days to expiration groups
<60 60-180 180 <60 60-180 180 test
Dap Drp Dap Drp Dap Drp
BS { tI 0.54 4.62 1.56 20.71 3.12 4.93
GARCH { tG 0.98 9.68 5.20 65.58 11.11 47.85 11.02 24.67 5.78 38.75
<0.94 SV{ rtR 0.43 3.30 1.41 21.86 1.91 12.42 4.36 22.87 4.49 37.16
SV { vtR 0.42 2.01 0.92 12.58 2.26 4.06 0.35 0.01 1.61 3.92
BS { tI 0.82 12.12 1.88 12.35 4.62 0.45
GARCH { tG 2.36 36.34 6.28 39.39 12.88 1.37 42.03 17.57 41.35 38.75
0.94-0.97 SV{ rtR 1.06 17.99 2.01 18.43 2.54 0.29 39.78 17.23 28.57 37.16
SV { vtR 0.58 8.39 1.38 7.63 3.41 0.33 22.20 8.89 30.49 3.92
BS { tI 1.13 5.78 2.48 0.38 5.23 0.31
GARCH { tG 3.19 18.03 6.79 1.23 12.33 0.73 13.59 59.65 50.69 38.75
0.97-1.00 SV{ rtR 1.54 12.16 2.07 0.49 2.12 0.13 15.81 54.50 24.49 37.16
SV { vtR 1.05 4.46 2.20 0.30 4.24 0.24 4.88 85.57 27.24 3.92
BS { tI 1.25 0.14 2.31 0.17 4.63 0.18
GARCH { tG 2.93 0.37 6.45 0.48 11.59 0.47 110.64 104.59 30.67 38.75
1.00-1.03 SV{ rtR 1.33 0.21 1.87 0.15 2.11 0.08 18.68 69.35 3.86 37.16
SV { vtR 1.35 0.15 2.33 0.16 3.85 0.14 6.61 63.22 25.95 3.92
54