He Stone Nandi

A Closed-Form GARCH Option
Valuation Model
Steven L. Heston
Goldman Sachs & Company
Saikat Nandi
Research Department
Federal Reserve Bank of Atlanta
This paper develops a closed-form option valuation formula for a spot asset whose variance follows a GARCH(p, q) process that can be correlated with the returns of the
spot asset. It provides the rst readily computed option formula for a random volatility
model that can be estimated and implemented solely on the basis of observables. The
single lag version of this model contains Hestons (1993) stochastic volatility model
as a continuous-time limit. Empirical analysis on S&P500 index options shows that the
out-of-sample valuation errors from the single lag version of the GARCH model are substantially lower than the ad hoc BlackScholes model of Dumas, Fleming and Whaley
(1998) that uses a separate implied volatility for each option to t to the smirk/smile in
implied volatilties. The GARCH model remains superior even though the parameters of
the GARCH model are held constant and volatility is ltered from the history of asset
prices while the ad hoc BlackScholes model is updated every period. The improvement is largely due to the ability of the GARCH model to simultaneously capture the
correlation of volatility with spot returns and the path dependence in volatility.
Since Black and Scholes (1973, henceforth BS) and Merton (1973) originally
developed their option valuation formulas, researchers have developed option
valuation models that incorporate stochastic volatility (see Heston (1993)
and the references therein). The two types of volatility models have been
continuous-time stochastic volatility models and discrete-time Generalized
Autoregressive Conditional Heteroskedasticity (GARCH) models. A related
class of volatility models is the implied binomial tree or the deterministic
volatility models of Derman and Kani (1994), Dupire (1994), and Rubinstein
Opinions in this paper are not those of Goldman Sachs & Co., Federal Reserve Bank of Atlanta, or the Federal
Reserve System. The paper was written when the rst author was a faculty member at the John M. Olin School
of Business, Washington University in St. Louis. Special thanks to an anonymous referee whose comments
and suggestions have greatly improved the scope and contents of the paper. We also thank the editor (Bernard
Dumas), executive editor (Maureen O Hara), participants of the workshop at Atlanta Fed, 1999 American
Finance Association meetings, the 8th annual derivatives securities conference, 1998 Financial Management
Association meetings, 1st annual conference on quantitative and computational nance, Jin Duan, Jeff Fleming,
Jens Jackwerth, Peter Ritchken and Steve Smith for comments. We are immensely grateful to Daniel Waggoner
for superb assistance with processing the data sets obtained from the CBOE and the Futures Industry Institute
as well as many other issues. Address all correspondence to Saikat Nandi, Research Department, Federal
Reserve Bank of Atlanta, 104 Marietta Street, N.W., Atlanta, GA 30303, or e-mail: [email protected].
The Review of Financial Studies Fall 2000 Vol. 13, No. 3, pp. 585625
2000 The Society for Financial Studies
The Review of Financial Studies / v 13 n 3 2000
(1994) in which the spot volatility is a function of the current asset price and
time only.
This paper presents an option formula for a stochastic volatility model with
Generalized Autoregressive Conditional Heteroskedasticity (GARCH). The
new formula describes option values as functions of the current spot price
and the observed path of historical spot prices. It captures both the stochastic
nature of volatility and correlation between volatility and spot returns. On a
daily frequency the model is numerically close to the continuous-time stochastic volatility model of Heston (1993), but much easier to apply with available
data. Our empirical analysis on S&P 500 index options shows that the out-ofsample valuation errors from the GARCH model are much lower than those
from other models, including heuristic rules that are used by market makers
to t to the variations in implied volatilities across strike prices and maturities. The GARCH model successfully predicts out-of-sample option prices
because it exploits the correlation of volatility with the path of stock returns.
In addition to improving the prediction of volatility, the correlation parameter
induces strike price and maturity patterns across option values such as the
pronounced smirk in implied volatilities in the index options market.
Continuous-time stochastic volatility models [for example Heston (1993)]
are difcult to implement and test. Although these models assume that volatility is observable, it is impossible to exactly lter a volatility variable from
discrete observations of spot asset prices in a continuous-time stochastic
volatility model. Consequently it is not possible to compute out-of-sample
options valuation errors from the history of asset returns. Also the unobservability of volatility implies that one has to use implied volatilities computed
from option prices to value other options. Holding the model parameters constant through time [as in Bates (1996, 1999) and Nandi (1998)], this approach
requires estimating numerous implied volatilities from options records, one
for every date and is computationally very burdensome in a long time series
of options records. Another alternative is to estimate all parameters (including volatility) daily from the cross-section of observed option prices as in
Bakshi, Cao and Chen (1997) or directly using the BS implied volatility
from a particular option/options as a proxy for the unobserved spot volatility
as in Knoch (1992). However, using implied volatilities to value an option
requires the use of other contemporaneous options that may not always be
feasible if one does not have reliable option prices such as in cases of thinly
traded or illiquid markets.
In contrast to the continuous-time models, GARCH models have the inherent advantage that volatility is readily observable from the history of asset
prices. As a result, a GARCH option model allows one to value an option
using spot volatilities computed directly from the history of asset returns
without necessarily using the implied volatilities inferred from other contemporaneous options. Thus it is possible to value an option solely on the
basis of observables because the parameters of the valuation formula can be
586
A Closed-Form GARCH Option Valuation Model
readily estimated from the discrete observations of asset prices. If a closedform solution were available, a GARCH model would enable one to readily
combine the cross-sectional information in options with the information in
the time series of the underlying asset. Since volatility is a readily computed
function of the history of asset prices, only a nite number of parameters
need to be estimated irrespective of the length of the time series, thus considerably simplifying the estimation procedure.
Unfortunately, existing GARCH models do not have closed-form solutions for option values. These models are typically solved by simulation
[Engle and Mustafa (1992), Amin and Ng (1993), Duan (1995)] that can be
slow and computationally intensive for empirical work. More recently, Duan,
Gauthier and Simonato (1999) provide a series approximation and Ritchken
and Trevor (1999) provide a lattice approximation to value American options
for GARCH processes with single lags in the variance dynamics. In contrast,
this paper develops a closed-form solution for European option values (and
hedge ratios) in a GARCH model. The model allows for multiple lags in the
time series dynamics of the variance process and also allows for correlation
between returns of the spot asset and variance. The single lag version of the
model reconciles the discrete-time GARCH approach with the continuoustime stochastic volatility approach to option valuation by including Hestons
(1993) closed-form stochastic volatility model as a continuous-time limit.
This generalizes the BlackScholes and Merton approach to option valuation
because it is possible to value options by the absence of arbitrage only in
the continuous-time limit, even though volatility is path dependent. In the
BS model option values are functions of the current spot asset price, while
in the GARCH model option values are functions of current and lagged spot
prices. Except for this difference the models are operationally similar.
We test the empirical implications of our GARCH model in the S&P 500
index options market. As a benchmark model we choose the ad hoc BS
model of Dumas, Fleming and Whaley (1998, henceforth DFW) that has
the exibility of tting to the strike and term structure of observed implied
volatilities by using a separate implied volatility for each option. It is found
that the GARCH model has smaller valuation errors (out-of-sample) than the
ad hoc BS model even though the ad hoc model is updated every period.
In contrast, the parameters of the GARCH model are held constant over a
sample period and variance is ltered from the history of asset prices. When
we update the parameters of the GARCH model every period, the out-ofsample prediction errors decrease even further and substantially. Also the
out-of-sample results remain essentially unchanged if we use the S&P 500
futures to lter the spot variance for our GARCH model instead of the S&P
500 cash index.
Our out-of-sample valuation results stand in contrast to previous empirical tests of the implied binomial tree/deterministic volatility models. In these
587
tests DFW (1998) found that the ad hoc BS model dominated the deterministic volatility models in terms of out-of-sample options valuation errors in
the S&P 500 index options market. Most of the options valuation improvements by the GARCH model are seen to result from its ability to simultaneously capture the path dependence in volatility and the negative correlation
of volatility with index returns. This negative correlation allows the model
to quickly adapt to changes in volatility associated with changes in the market levels. Also the negative correlation generates a negative skewness in the
risk-neutral distribution of the S&P 500 index return. This is associated with
the strike price and maturity specic biases in the index options market.
Section 1 describes the GARCH process and presents the option formula. Section 2 applies it to the S&P500 index option data, Section 3
reports the in-sample and out-of-sample results, while Section 4 concludes.
Appendix A contains detailed calculations and derivations of the option formula while Appendix B contains the calculations regarding the convergence
of the GARCH model to its continuous-time limit.
1. The Model
The model has two basic assumptions. The rst assumption is that the logspot price follows a particular GARCH process.
Assumption 1. The spot asset price, S(t) (including accumulated interest
or dividends) follows the following process over time steps of length ,

(1a)
log(S(t)) = log(S(t )) + r + h(t) + h(t)z(t)
h(t) = +
p

i=1
i h(t i) +
q

i=1

i (z(t i) i h(t i))2 ,
(1b)
where r is the continuously compounded interest rate for the time interval
and z(t) is a standard normal disturbance. h(t) is the conditional variance of
the log return between t and t and is known from the information set at
time t . The conditional variance in equation (1b), although distinct from
the classic GARCH models of Bollerslev (1986) and Duan (1995), is quite
similar to the NGARCH and VGARCH models of Engle and Ng (1993).
The conditional variance h(t) appears in the mean as a return premium. This
allows the average spot return to depend on the level of risk.1 Equation (1a)
assumes that the expected spot return exceeds the riskless rate by an amount
proportional to the variance h(t). Since volatility equals the square root of
h(t), this implies the return premium per unit of risk is also proportional
1
We assume that is constant, but option prices are very insensitive to this parameter. The functional form
of this risk premium, h(t), prevents arbitrage by ensuring that the spot asset earns the riskless interest rate
when the variance equals zero.
588
to the square root of h(t), exactly as in the Cox, Ingersoll and Ross (1985)
model. In particular limiting cases the variance becomes constant. As the i
and i parameters approach zero, it is equivalent to the BlackScholes model
observed at discrete intervals.
This paper will focus on the rst-order case (p = q = 1) for pointing
out some of the properties of the particular GARCH process. The rst-order
process remains stationary with nite mean and variance if 1 + 1 12 < 1.2
In this model one can directly observe h(t + ), at time t, as a function of
the spot price as follows:
h(t + ) = + 1 h(t)
+ 1
(log(S(t)) log(S(t )) r h(t) 1 h(t))2

.
h(t)
(2)
1 determines the kurtosis of the distribution and 1 being zero implies a

deterministic time varying variance. The 1 parameter results in asymmetric
inuence of shocks; a large negative shock, z(t) raises the variance more
than a large positive z(t). In general the variance process h(t) and the spot
return are correlated as follows,
Covt [h(t + ), log(S(t))] = 21 1 h(t).
(3)
Given positive 1 , positive value for 1 results in negative correlation between

spot returns and variance. This is consistent with the postulate of Black
(1976) and the leverage effect documented by Christie (1982) and others.
Thus, 1 controls the skewness or the asymmetry of the distribution of the
log-returns [see Heston, (1993) for an illustration in a continuous time framework]; the distribution is symmetric if 1 (and ) is zero.
Although equations (1a) and (1b) refer to a stochastic process observed at
a prescribed time interval , they have an interesting continuous-time limit.
It can be shown following Foster and Nelson (1994) that as the observation
interval, shrinks, the variance process h(t) converges weakly to a variance
process, v(t) which is the square-root process of Feller (1951), Cox, Ingersoll
and Ross (1985), and Heston (1993)
dv = ( v)dt + vdz,
(4)
where z(t) is a Wiener process. The details are in Appendix B. Consequently

the option valuation model (1) contains Hestons (1993) continuous-time
2
In the multiple factor case one must add the additional condition that the polynomial roots of x p
i i2 )x pi lie inside the unit circle.
p
i=1 (i
589
stochastic volatility model (that also admits a closed-form solution for option
values) as a special case.3
At this point we cannot value options or other contingent claims because
we do not know the risk-neutral distribution of the spot price. Motivated by
previous lognormal option formulas, we rewrite equation (1) in the form

1
(5a)
log(S(t)) = log(S(t )) + r h(t) + h(t)z (t)
2
p
q

h(t) = +
i h(t i) +
i (z(t i) i h(t i))2
i=1
i=2
+ 1 (z (t ) 1 h(t ))2 ,
(5b)
where,

1
z (t) = z(t) + +
h(t),
2
1
1 = 1 + + .
2
Equations (5a) and (5b) appear to be risk-neutral versions of equations (1a)
and (1b), though at this point, equations (5a) and (5b) are merely algebraic
rearrangements of equation (1a) and (1b); there is no reason for the riskneutral distribution of z (t) to be normal. In order for z (t) to have a standard
normal risk-neutral distribution, we introduce Assumption 2.
Assumption 2. The value of a call option with one period to expiration
obeys the BlackScholes-Rubinstein formula.
This assumption is equivalent to Duans (1995) valuation assumption.
The BlackScholes-Rubinstein formula is natural to use here because the
spot price has a conditionally lognormal distribution over a single period.
However, BS prices do not follow from absence of arbitrage with discretetime trading. Instead, one must appeal to other arguments such as those of
Rubinstein (1976) and Brennan (1979). If the BS formula holds for a single
period, then the risk-neutral distribution of the asset price is lognormal with
mean, S(t )er . This implies that one can nd a random variable, z (t)
which has a standard normal distribution under the risk-neutral probabilities.4
We formalize this property as the following proposition.
3
We can value options by the absence of arbitrage alone in the continuous-time model as the asset returns and
variance are instantaneously perfectly correlated. Note however that returns and volatility are not perfectly
correlated over any discrete interval of time.
For details of the preference assumptions that give rise to risk neutralization in a discrete-time model with
continuously distributed returns, one can refer to Duan (1995) for a GARCH model and Rubinstein (1976)
and Brennan (1979) for the BlackScholes model.
590
Proposition 1. The risk-neutral process takes the same GARCH form as

equations (1a) and (1b) with replaced by 1/2 and 1 replaced by
1 = 1 + + 1/2.
The proof of this proposition is trivial by noting that z (t), 1 and as
dened above make the one period return from investing in the spot asset
equal to the risk free rate in equation (5a). Assumptions 1 and 2 allow us
to derive the values of all contingent claims that can be written as functions
of the spot asset price. Since long-term options are functions of S(t) and
h(t + ), and h(t + ) can be written as a function of S(t) in equation (2),
this includes options of all maturities. Also note that unlike continuous-time
stochastic volatility models, in the GARCH model, the parameter that governs the risk-neutral skewness, 1 and the parameter that governs the actual
skewness, 1 can differ due to the risk premium parameter, .
The risk premium parameter, has some useful interpretations. In practice, exerts a negligible inuence on the current ltered values of h(t). But
it does change the risk-neutral distribution of future variance. The variance,
h(t +) does not carry an independent risk premium in this model. But since
our empirical evidence shows that variance is highly negatively correlated
with spot S&P 500 index returns, variance should behave like a negative
beta asset.5 If the spot asset has a positive return premium ( > 1/2) and
variance is negatively correlated with spot returns (1 > 0) then Proposition 1
(and equation B1 of Appendix B) shows that the risk-neutral drift of variance will be higher than the true drift of variance. This means that implied
volatility will typically be higher than expected future volatility. For example,
the term structure of volatility inferred from option prices may be upward
sloping even when the spot volatility, h(t + ) is not expected to increase.
In this case a trader using the BlackScholes formula would incorrectly perceive an arbitrage opportunity from selling and hedging volatility-sensitive
options, e.g., long-term at-the-money straddles. However the BlackScholes
hedge ratio does not accommodate GARCH effects. Since volatility is very
negatively correlated with S&P 500 returns, a short option position with a
BlackScholes hedge would actually be bullish on the market. While the
option position would earn an expected return premium, the premium would
be no higher than the return on a comparable levered market position. And
in the event of a sharp market downturn, volatility would rise signicantly,
causing the short option position to lose considerable value. This is consistent
with losses at arbitrage hedge funds, such as Long Term Capital Management, which held short index option positions in some markets in the third
quarter of 1998 [Dunbar (1999)].
5
Using equations (3) and (B2) of Appendix B, with the parameter estimates in Table 1 from spot S&P 500
returns, gives a correlation of 0.96 when h(t) is at its long-run value. This is a reasonable approximation
to the limiting diffusion case of Appendix B with perfect negative correlation.
591
We proceed to solve for the generating function of the GARCH process

of (1a) and (1b) and use it to produce option values.
Let f () denote the conditional generating function of the asset price
f () = Et [S(T ) ].
(6)
This is also the moment generating function of the logarithm of S(T ). The
function f () depends the parameters and state variables of the model, but
these arguments are suppressed for notational convenience. We shall use the
notation f () to denote the generating function for the risk-neutral process
in (5a) and (5b).
Proposition 2. The generating function takes the log-linear form

p
Bi (t; T , )h(t + 2 i)

f () = S(t) exp A(t; T , ) +
i=1
q1

i=1

Ci (t; T , )(z(t + i) i h(t + i))2 , (7)
where,
A(t; T , ) = A(t + ; T , ) + r + B1 (t + ; T , )
1
ln(1 21 B1 (t + ; T , ))
2
1
B1 (t; T , ) = ( + 1 ) 12 + 1 B1 (t + ; T , )
2
1/2( 1 )2
+
,
1 21 B1 (t + ; T , )
(8a)
(8b)
for the single lag (p = q = 1) version and these coefcients can be calculated recursively from the terminal conditions:
A(T ; T , ) = 0.
(9a)
B1 (T ; T , ) = 0.
(9b)
The appendix derives the recursion formulas for the coefcients A(t; T , ),
Bi (t; T , ), and Ci (t; T , ) in the general case i.e., for any p and q.
Since the generating function of the spot price is the moment generating
function of the logarithm of the spot price, f (i) is the characteristic function
of the logarithm of the spot price. Note that to use the characteristic function,
in equations (8a) and (8b) must be replaced by i everywhere. One can
calculate probabilities and risk-neutral probabilities following Feller (1971)
or Kendall and Stuart (1977) by inverting the characteristic function.
592
Proposition 3. If the characteristic function of the log spot price is f (i)

then

i
1
1
K f (i + 1)
Et [Max(S(T ) K, 0)] = f (1)
+
d
Re
2 0
if (1)

i
1
1
K f (i)
K
+
d , (10)
Re
2 0
i
where Re[ ] denotes the real part of a complex number. Proposition 3 involves
a somewhat new inversion formula, different from that of Heston (1993) and
others. In particular, it enables us to calculate the expectation in (10) once
we just have the characteristic function of the logarithm of the spot price,
instead of calculating two separate integrals.6
An option value is simply the discounted expected value of the payoff,
Max(S(T ) K, 0) calculated using the risk-neutral probabilities, i.e., using
the characteristic function, f (i). In particular, a European option value is
given by the following corollary.
Corollary. At time t, a European call option with strike price K that expires
at time T is worth
1
C = er(T t) Et [Max(S(T ) K, 0)] = S(t)
2
i
er(T t)
K f (i + 1)
+
d
Re
i
0

i
1
K f (i)
r(T t) 1
Ke
+
d ,
Re
2 0
i
(11)
where Et [] denotes the expectation under the risk-neutral distribution. This

completes the option valuation formula.7 As in the BlackScholes formula,
(11) can be written as the asset price multiplied by a probability, P1 () and
the discounted strike price multiplied by a probability, P2 (). P2 () is the risk
neutral probability of the asset price being greater than K at maturity and the
delta of the call value is simply P1 (). The other hedge ratios like the vega and
the gamma can be calculated by straight differentiation in equation (11) and
the expression for P1 () respectively. Put option values can be calculated using
the put-call parity. In contrast to the BlackScholes formula, this formula
is a function of the current asset price, S(t), and the conditional variance,
6
The new inversion formula exploits the inherent relationship between the two probabilities, P1 () and P2 () of a
European option valuation model, thus requiring the calculation of only one integral, instead of two separate
integrals as in Heston (1993).
The integrands converge very rapidly and the integration can be very efciently performed in fractions of
a second using a numerical integration routine such as Rombergs method on an open interval [Press et al.
(1992)] or quadrature based integration routines (end of Appendix A has some sample option values).
593
h(t + ). Since h(t + ) is a function of the observed path of the asset
price, the option formula is effectively a function of current and lagged asset
prices. In contrast to continuous-time models, volatility is a readily observable function of historical asset prices and need not be estimated with other
procedures.
The next section describes the empirical performance of the single lag
(p = q = 1) version of the GARCH model in the S&P 500 index options
market.
2. Empirical Analysis
The empirical analysis starts with a description of the options data. It proceeds to estimate the GARCH model with time series data on index returns
and with options data.
2.1 Description of data
Intra-day data on S&P 500 index options traded on the Chicago Board
Options Exchange (CBOE) are used to test the model. The raw data set is
obtained directly from the exchange. The market for S&P 500 index options
is the second most active index options market in the United States and, in
terms of open interest in options, it is the largest. Unlike options on the S&P
100 index, there are no wild card features [see Fleming and Whaley (1994),
French and Maberly (1992)] that can complicate valuation. Also it is easier
to hedge S&P 500 index options because there is a very active market for
the S&P 500 futures. In fact, according to Rubinstein (1994) it is one of the
best markets for testing a European option valuation model.
As many of the stocks in the S&P 500 index pay dividends, one needs a
time series of dividends for the index. We use the daily cash dividends for the
S&P 500 index collected from the S&P 500 information bulletin.8 We arrive
at the present value of the dividends and subtract it from the current index
level. For the risk free rate, the continuously compounded Treasury bill rates
(from the average of the bid and ask discounts reported in the Wall Street
Journal), interpolated to match the maturity of the option is used. Also the
change in the expiration time of the SPX options on August 24, 1992 from
close to open (see DFW) resulting in the reduction of the time to expiration
by one day was taken into account.
The intra-day data set is sampled every Wednesday (or the next trading
day if Wednesday is a holiday)9 between 2:30 P.M. and 3:15 P.M. (central standard time, CST) for the years 1992, 1993 and 1994 to create the
8
We thank the referee for suggesting the use of this dividend series and Jeff Fleming and the referee for making
the dividend series available to us.
In our options sample, all but one day are Wednesdays.
594
data that we work on. We follow DFW (1998) in ltering the intra-day data
to create weekly data and use the mid-point of the bid-ask quote as the
option price. As in DFW only options with absolute moneyness, |K/F 1|,
(K is the strike and F is the forward price) less than or equal to ten percent
are included. In terms of maturity, options with time to maturity less than six
days or greater than one hundred days are excluded.10 However, unlike DFW,
we do not infer the index level simultaneously with the other parameters in
the estimation procedure. Instead, the level of the S&P 500 index reported
for a particular record is used. The S&P 500 is a value-weighted index and
the bigger stocks that trade more frequently constitute the bulk of the index
level. Since intra-day data and not the end-of-the-day option prices is used,
the problem with the index level being somewhat stale may not be severe
enough to undermine an estimation procedure. In theory, one could possibly overcome this problem by using implied index levels from the put-call
parity equation. However, this is conditional on put-call parity holding as an
equality and in the presence of transactions costs (bid-ask spreads that are
non-negligible), the equality becomes an inequality. Thus the implied index
levels from the put-call parity equation may not equal the true index level.
Also, even if one assumes away transactions costs, it is very difcult to create a sample of sufcient size by creating matched pairs of puts and calls
because the level of the S&P 500 index changes quite frequently through
the day. Another alternative is to use the implied index levels from S&P
500 futures prices. However, one must then assume that the futures and the
options market are closely integrated.11 The following criteria are also used
as lters.
1) An option of a particular moneyness and maturity is represented only
once in the sample on any particular day. In other words, although the
same option may be quoted again in our time window (with same or
different index levels) on a given day, only the rst record of that option
is included in our sample for that day.
2) A transaction must satisfy the no-arbitrage relationship (Merton, 1973)
in that the call price has to be greater than or equal to the spot price
minus the present value of the remaining dividends minus the discounted
strike price. Similarly, the put price has to be greater than or equal to
the present value of the remaining dividends plus the discounted strike
price minus the spot price.
The data set consists of 10,100 records/observations. The average number
of options per day is 65 with a minimum of 24 and a maximum of 106.
The average bid-ask spread is $0.481. The number of options of distinct
maturities for various days were: 30 daystwo maturities, 115 daysthree
maturities and 11 daysfour maturities.
10
See DFW for justication of the exclusionary criteria about moneyness and maturity.
11
As will be discussed later, we do, however, use the S&P 500 futures data to address potential problems in
ltering volatility from the history of asset prices in the GARCH model.
595
2.2 Estimation
The empirical analysis focuses mainly on the single lag version of the GARCH
model. We set = 1 and use daily index returns to model the evolution of
volatility. Unlike continuous-time stochastic volatility models in which the
volatility process is unobservable, all the parameters in our valuation formula can be easily estimated directly from the history of asset prices. We
do this with the maximum likelihood estimation (MLE) used by Bollerslev
(1986) and many others.12 To illustrate the importance of the skewness parameter, 1 , we performed this estimation with an unrestricted model and with a
restricted GARCH model in which 1 was constrained to equal zero (symmetric GARCH). Table 1 shows the maximum likelihood estimates of the
GARCH model, both when 1 is non-zero and when it is restricted to zero,
on the daily S&P 500 levels, closest to 2:30 P.M. (CSTCentral Standard
Time) from 01/08/92 to 12/30/94. The skewness parameter, 1 is substantially positive indicating that shocks to returns and volatility are strongly
negatively correlated. Using a likelihood ratio test, the symmetric version
is easily rejected implying that the negative correlation between returns and
volatility is a signicant feature of the S&P 500 time series. The daily volatility series (annualized) are shown in Figure 1A and 1B for the unrestricted
and restricted/symmetric versions of the model.13 These gures show that the
skewness parameter 1 has an important effect on the qualitative behavior
of the variance process. Including this parameter makes the ltered variance
more volatile, and produces sudden drops in volatility in addition to sudden
increases.
We also investigate how different our maximum likelihood estimation
results would be if we use S&P 500 futures prices to imply out the S&P 500
index levels. Towards this purpose we use the closest to 2:30 P.M. (CST)
lead/nearest maturity S&P 500 futures prices from 01/08/92 to 12/30/94 to
get the implied S&P 500 index levels. These futures prices are created from
tick-by-tick S&P 500 futures data sets that are obtained from the Futures
Industry Institute.14 Given a discrete dividend series, we use the following
equation [see Hull (1998)] to get the implied spot price (i.e. S&P 500 index
level)
F (t) = (S(t) PVDIV)er(t)(T t)
(12)
where F (t) denotes the futures price, PVDIV denotes the present value of
dividends to be paid from time t until the maturity of the futures contract
12
The procedure sets h(0) equal to the sample variance of the changes in the logarithm of S(t). Due to the
strong mean reversion of volatility, all results were insensitive to the starting value of h(0).
13
Unreported results show that various symmetric and asymmetric GARCH specications of Engle and Ng
(1993) produce similar results to our symmetric and asymmetric models, respectively on the same data
set. The values of the likelihood function are very similar and the time series graphs of ltered volatility
(Figures 1A and 1B) lie virtually on the top of each other.
14
We thank the referee for pointing us towards this source for obtaining tick-by-tick S&P 500 futures data.
596
1
1.32e6
(0.03e6)
1.07e6
(0.04e6)
1.33e6
(0.03e6)
1.71e6
(0.04e6)
1
0.589
(0.007)
0.922
(0.013)
0.586
(0.006)
0.859
(0.02)
424.69
(9.2)
421.39
(11.01)
5.02e6
(0.19e6)
1.63e6
(0.43e6)
4.96e6
(0.13e6)
3.57e6
(0.43e6)
0.205
(0.228)
0.732
(0.22)
0.335
(0.228)
0.671
(0.22)
9.71%
9.54%
9.33%
9.51%
0.859
0.826
0.922
0.823
1 + 1 12
3467.6
3482.8
3492.4
3503.7
Log-Likelihood

h(t)z(t),
252( + 1 )/(1 1 1 12 ) is the annualized (252
days) long run volatility (standard deviation) implied by the parameter estimates. 1 + 1 12 measures the degree of mean reversion in that 1 + 1 12 = 1 implies that
the variance process is integrated.
time). Number of Observations = 755. Asymptotic standard errors appear in parentheses. dened to be equal to

The log-likelihood function is Tt=1 0.5 log(h(t)) + z(t)2 , where T is the number of days in the sample. The daily cash index levels closest to (before) 2:30 P.M.
(central standard time) from 01/08/9212/30/94 are used. The futures prices are those of the shortest/lead maturity
contracts closest to (before) 2:30 P.M. (central standard

2
h(t) = + 1 h(t ) + 1 z(t ) 1 h(t )
log(S(t)) = log(S(t )) + r + h(t) +
Maximum Likelihood Estimates of the GARCH model with p = q = 1 and = 1 (day) using the spot/cash S&P 500 levels and the S&P 500 futures prices for the
unrestricted (1 = 0) and restricted (1 = 0) model.
GARCH, 1 = 0 (futures)
GARCH (futures)
GARCH, 1 = 0 (spot)
GARCH (spot)
Maximum likelihood estimation
Table 1
597
Figure 1A
This gure shows the daily annualized spot volatility from the unrestricted/asymmetric GARCH model from
January 09, 1992 to December 30, 1994 using daily S&P 500 index levels (closest to and before 2:30 P.M.,
central standard time).
at time, T and r(t) is the continuously compounded Treasury bill rate (from
the average of the bid and ask discounts reported in the Wall Street Journal),
interpolated to match the maturity of the futures contract.
Table 1 also reports the various parameter estimates and the value of
the log-likelihood function using the futures data set. The parameters are
very similar across the analysis of cash/spot and futures data for the unrestricted/asymmetric model. As with the cash/spot data, shocks to returns and
volatility are negatively correlated (i.e. 1 > 0) using the S&P 500 futures
data. Other features of the time series dynamics of volatility are quite similar across the cash/spot and futures data sets. For example, the parameter that measures the degree of mean reversion (as given by 1 + 1 12 ) is
0.823 from the cash/spot data and 0.826 from the futures data. Similarly, the
volatility of volatility, as measured by 1 , is 1.32e6 from the cash/spot
data and 1.33e6 from the futures data. The
annualized long-run mean
of volatility/standard deviation as given by 252( + 1 )/(1 1 1 12 )
(assuming a year with 252 trading days) is 9.51% from the cash/spot data
and 9.54%
for the futures data. This is reected in the actual ltered time
the average
series of h(t + 1) (i.e. volatility, not variance). For example,
(across our sample of 755 days) annualized level of h(t + 1) is 9.34%

from the cash/spot data and 9.46% from the futures data.15 The mean
15
For each day in our sample, the annualized level of volatility is
598
252 h(t + 1).
Figure 1B
This gure shows the daily annualized spot volatility from the restricted/symmetric GARCH model from
January 09, 1992 to December 30, 1994 using daily S&P 500 index levels (closest to and before 2:30 P.M.,
central standard time).
difference (futures spot) in the annualized volatility between the two time
series is 0.12% with a standard deviation of 0.249% and a maximum difference of 1.4%.
In the restricted/symmetric model (i.e. 1 = 0), the differences between the
parameter estimates from the futures and the cash/spot are somewhat greater
than in the unrestricted model. For example, the annualized long-run mean
of volatility is 9.33% from the cash/spot data and 9.71% from the futures
data. Similarly, the average annualized level of h(t + 1) is 9.39% from

the cash/spot data and 9.72% from the futures data. The mean difference
(futures spot) between the two time series of annualized volatility is 0.3%
with a standard deviation of 0.31% and a maximum difference of 2.3%.
Note however that as with the spot/cash data, the futures data also rejects
the restricted model in favor of the unrestricted model (using likelihood ratio
test).
One could plug the parameter estimates obtained from the above MLEs
(using historical asset prices) into the options valuation formula to compute
option values. However, the information that one would be using is only historical and could be different from the expectations about the future evolution
of the asset price that are embedded in option prices. In contrast, the information in option prices is forward looking. Since our model has a closed-form
solution for option values, a natural candidate for parameter estimation is a
non-linear least squares (NLS) procedure that tries to match model option
values to observed option prices as closely as possible.
599
The option value at time t is not only a function of the current level of
variance, h(t + 1) but also of the parameters that drive the variance process,
namely, 1 , 1 , 1 , and . One could do a cross-sectional tting every
week (i.e. each Wednesday) to imply out all these parameters including the
variance, h(t + 1). But purely cross-sectional estimation from a single dayss
data suffers from two problems if the model has more parameters to be
estimated than a single implied volatility as in the BS model. First of all,
because of the limited sample size, there is a problem of overtting as noted
in DFW (1998). Furthermore, in the context of our model, this procedure
does not use the information in the evolution of the index or equivalently
the time series of historical volatilities. It is quite possible that the history
of the index provides some information about the future over and above the
information contained in the option prices. Our model can readily exploit the
combined information in the history of asset prices (as variance is observable)
and the cross-section of option prices (due to a closed-form solution). In order
to take the implications of the model seriously, we hold the time invariant
parameters constant over the rst six-month period of each year. Later on
we relax this restriction and allow the parameters of the GARCH model to
be updated every week (in the process of computing out-of-sample valuation
errors in the second half of each year). However, we always compute the
variance, h(t + 1) from the history of asset prices.
As mentioned previously, for each year we choose the option prices in the
interval, 2:303:15 P.M. (CST) for the rst twenty six Wednesdays (or the
next trading day) to create our sample. Let e(i, t) denote the model error in
valuing option i at time t, i.e., e(i, t) is the difference between the model
value of option and
price of that option at time t. Then our crite themarket
Nt
e(i, t)2 , where T denotes the number of weeks in
rion function is, Tt=1 i=1
the sample and Nt is the number of options traded on the Wednesday (or the
next trading day) of week t. The criterion function needs to be minimized
over 1 , 1 , 1 , and . Note that in order to minimize the above criterion
function we need h(t + 1) for each t. However, at each t, h(t + 1) is a function of 1 , 1 , 1 , , and and the history of asset prices [see equation (2)].
Therefore h(t + 1) is known at time t given these parameters.16 This feature
considerably simplies our estimation procedure in contrast to continuoustime stochastic volatility models where the daily volatility is not known as
a function of the history of asset prices. Consequently, one has to estimate
the daily volatility separately. However, if the sample consists of a long time
series of option prices, then the number of parameters increases proportionately with the number of days in the sample. In contrast, as noted previously,
in our GARCH setting, only a few time-invariant parameters need to be estimated irrespective of the sample size.
16
At each iteration of an optimization routine, h(t + 1) is needed to compute option prices. But given the
parameters that are in use for that iteration, h(t + 1) is known from the history of asset prices.
600
For the BS model, a single implied volatility can be estimated for each
day in the sample by minimizing the above criterion function. However, a
BS model with a single implied volatility across all strikes and maturities,
although consistent theoretically is perhaps too restrictive in practice. Since
the GARCH model has four more parameters than the BS model, it may
have an unfair advantage over the BS model. Therefore, we follow DFW
and construct an ad hoc BS model in which each option has its own implied
volatility depending on the strike price and time to maturity. Specically, the
spot volatility of the asset that enters the BS option formula is a function of
the strike price and the time to maturity or combinations thereof (see DFW
for details). For example, one functional form is,
= a0 + a1 K + a2 K 2 + a3 + a4 2 + a5 K,
(13)
where is the implied volatility (using BS) for an option of strike K and
time to maturity . There are simpler parameterizations (subsets of the above,
see DFW for details) that one can use when data are limited. As per DFW the
particular functional form of we select on a given day depends on the number of distinct option maturities in the sample on that day. The coefcients
of the ad hoc model are estimated every week via ordinary least squares,
minimizing the sum of squared errors between the BS implied volatilities
across different strikes (and maturities) and the models functional form of
the implied volatility.
The ad hoc BS model, although theoretically inconsistent is denitely a
more challenging benchmark than the simple BS model for any competing
option valuation model. Furthermore, DFW shows that the implied binomial
tree or the deterministic volatility models of Derman and Kani (1994), Dupire
(1994), and Rubinstein (1994) underperform the ad hoc BS model in out-ofsample options valuation errors in the S&P 500 index options market. Thus
comparing the GARCH to the ad hoc BS model should also yield insights
on the relative efcacies of path-independent volatility models such as the
implied binomial tree models and path-dependent volatility models such as
the GARCH in terms of valuing options.
The NLS estimation procedure was carried out to estimate the parameters
for the rst six months of 1992, 1993 and 1994.17 Although the options data
are weekly, the conditional variance, h(t +1), that is relevant for option values
at time t is drawn from the daily evolution of index returns and not from
the weekly evolution. Specically we use the levels of the index closest to
(before than) 2:30 P.M. (CST). The starting variance, h(0), is kept xed at the
in-sample estimate of the variance i.e., the variance for the rst six months of
1992, 1993 and 1994 respectively, computed from daily logarithmic returns.
17
The NLS procedure that we use is the Levenberg-Marquardt method in Press et al. (1992).
601
3. Model Comparisons
This section describes the parameter estimates and in-sample and out-ofsample comparisons of the GARCH model with the BS model and the ad hoc
BS model of DFW (1998).
3.1 In-sample model comparison
The parameter estimates from the NLS estimation for the three years, 1992,
1993, and 1994 and the average in-sample valuation errors appear in Table 2.
The actual and risk-neutral skewness parameter, 1 and 1 , respectively are
always positive. This indicates that variance tends to rise when the index falls,
and vice versa under both the actual and risk-neutral distribution. Because
shocks to variance and index returns are negatively correlated, there is negative skewness in the distributions of multiperiod index returns. The long-run
annualized volatility (standard deviation) implied by the options data for the
three years, 1992, 1993 and 1994 are 13.1%, 10.6% and 10.6% respectively.18
The average option price in the sample is somewhat more than thirteen
dollars, ranging from $13.37 in 1993 to $13.70 in 1994. The root mean
squared error (RMSE) of the simple BS model is around one dollar insample, ranging from $0.95 in 1992 to $1.14 in 1994.19 These are economically signicant errors in matching market option prices. Nevertheless the
BS model ts better than the symmetric GARCH (1 = 0) model. The root
mean squared valuation errors (averaged across all options) for the symmetric
GARCH model are $1.06, $0.986 and $1.29 respectively, for the years 1992,
1993 and 1994. The symmetric GARCH model has worse in-sample t than
the BS model in every year. This version of the model allows conditional
heteroskedasticity (as 1 is non-zero) and it also allows a term structure of
volatility. Since the BS model is a special case of the symmetric GARCH
model, it may seem puzzling that the BS model ts better. However, the BS
model has the advantage of using a different volatility to t option prices each
day. In contrast the symmetric GARCH model uses the actual time series of
index returns to generate volatility. The symmetric GARCH model has more
exibility to t the term structure of the option prices across maturity. But
apparently this does not adequately compensate for the more accurate calibration of the simple BS model period by period.
The results improve dramatically with the full GARCH model that allows
pronounced skewness/correlation effects with a nonzero 1 parameter. This
parameter allows better time-series tracking of volatility by accounting for
changes in volatility that accompany upward or downward movements in the
index. The skewness parameter also gives the model an important dimension
of exibility in valuing options across different strike prices. The root mean
18
Since we only use options of up to 100 days in maturity, these numbers should be treated with caution.
19
RMSE is the square root of the average squared valuation error.
602
0.616
(0.128)
0.286
(0.038)
0.504
(0.08)
0.105
(0.021)
0.783
(0.068)
0.214
(0.005)
1.16e5
(5.23e6)
3.42e6
(0.52e6)
1.08e5
(4.6e7)
4.48e6
(0.12e6)
9.39e6
(1.04e7)
7.98e7
(0.45e7)
969.24
(82.15)
122.95
(108.7)
412.31
(30.21)
196.57
(138.6)
426.99
(68.32)
172.53
(151.1)
8.22e7
(0.01e6)
8.0e18
(3.2e16)
1.48e6
(0.11e6)
9.09e6
(1.21e8)
2.76e6
(0.6e6)
1.18e5
(5.9e6)
0.48
(1.06)
122.45
(158.7)
1.03
(2.05)
196.07
(138.6)
0.32
(0.49)
173.03
(151.1)
0.106
0.176
0.106
0.252
0.131
0.39
0.964
0.924
0.867
0.921
0.909
0.961
1 + 1 12
13.70
13.70
13.70
13.70
0.644
0.688
13.37
13.37
0.586
0.615
1.14
1.29
13.37
13.37
13.61
13.61
0.526
0.686
0.964
0.986
13.61
13.61
Average price
0.95
1.06
RMSE
1662
1662
1662
1662
1750
1750
1750
1750
1744
1744
1744
1744
Observations
Reports the parameter estimates and in-sample valuation errors from minimizing the sum of squared errors between model option values and market option prices in the rst half of each year.
Asymptotic standard errors appear in parentheses. Two versions of the GARCH model are used, one in which 1 is unrestricted and another in which 1 = 0. BS is the BlackScholes model
in which a single implied volatility is estimated across all strikes and maturities on a given day while ad hoc BS is an ad hoc version of the BlackScholes model with strike and maturity
specic implied volatilities. Both BS and the ad hoc BS are calibrated every week while the parameters of the GARCH model are held constant over the estimation period. is the annualized
long run standard deviation under the GARCH (dened in Table 1) while 1 + 1 12 measures the degree of mean reversion in that 1 + 1 12 + 1 implies that the variance process is integrated.
1 = 1 + + 1/2, measures the skewness of the risk neutral distribution. If 1 = 0, then 1 = ( + 1/2). RMSE is the root mean squared pricing error (in $). Average price is the average
option price in the sample (in $).
1994
BS
GARCH-non-updated
(1 = 0)
Ad hoc BS
GARCH-non-updated
1993
BS
GARCH-non-updated
(1 = 0)
Ad hoc BS
GARCH-non-updated
1992
BS
GARCH-non-updated
(1 = 0)
Ad hoc BS
GARCH-non-updated
Non-linear least squares estimation
Table 2
603
squared valuation errors for the GARCH model in 1992, 1993 and 1994 are
68.6 cents, 61.5 cents and 68.8 cents respectively. Recall that the corresponding errors in the BS model are 95 cents, 96.4 cents, and $1.14 respectively.
Hence the GARCH model provides a substantially better in-sample t. Thus
although not updated and constrained to use the variance from the history
of asset prices, the GARCH model ts better because it matches the shape
of the option prices better. And the skewness parameter, 1 is essential for
this purpose. Bates (1999) and Nandi (1998) have previously emphasized the
importance of skewness effects when applying continuous-time stochastic
volatility models to explain index option prices.20
We should be able to get the best in-sample t to options prices by using
a weekly implied volatility along with a exible tting option formula. The
ad hoc BS model has more exibility than the GARCH model because it is
designed to t both the volatility smile in strike prices and the term structure
of implied volatilities. Also it is updated every period. Table 2 shows that the
ad hoc version of the BS model dominates the GARCH model in-sample in
each of the three years. The in-sample RMSE over the rst six months are
52.6 cents, 58.6 cents and 64.4 cents for 1992, 1993 and 1994 respectively.
Thus a exible but theoretically inconsistent model may appear to be better
than our theoretically consistent GARCH model in terms of in-sample t.
Of course the above comparison of the GARCH model with the ad hoc
BS model is somewhat unfair, because the implementation of the ad hoc
BS model allows the model to be updated every week. It is not clear whether
the improved in-sample t of the ad hoc model stems from a more exible
functional form or from the instability of the functional form of the GARCH
process over a long enough time period. In order to determine this we estimated an updated GARCH model by minimizing the sum of squared errors
between model option values and market option prices, allowing the parameters to change every week. Although the parameters change each week, the
variance, h(t + 1) is still drawn from the history of asset prices at time t.
At each time we used the time-series of returns from the previous 252 days
to lter the variance.21 Since our ultimate goal is to compare out-of-sample
valuation errors in the second half of each year, this updating is done only in
the second half of each year. Table 3 presents a fair in-sample comparison
of the ad hoc BS model with the updated GARCH model.22 In every year the
updated GARCH model ts options prices better than the ad hoc BS model
in terms of root mean squared error, ranging from around 3 cents in 1994
20
As mentioned earlier, the true time series skewness parameter is slightly different from the risk-neutral skewness parameter that affects option prices. However, in our model both of these skewness parameters are
estimated to have the same sign.
21
We have also used longer time intervals, such as two or three years for drawing the variance. However, the
results are essentially the same due to the strong mean reversion in variance.
22
To facilitate comparison with subsequent tables these results are reported for the second half of each year,
but results are similar in the rst half.
604
Table 3
In-sample comparison of the ad hoc BS model and the updated GARCH model
RMSE
Average option price
Number of observations
1992
Ad hoc BS
GARCH (updated)
0.416
0.366
13.09
13.09
1550
1550
1993
Ad hoc BS
GARCH (updated)
0.553
0.483
13.24
13.24
1511
1511
1994
Ad hoc BS
GARCH-updated
0.489
0.459
13.12
13.12
1881
1881
In-sample valuation errors (in $) from the weekly (every week) estimation using option prices in the second half of each year
(1992, 1993 and 1994) for the updated GARCH and the ad hoc BS model. Note however that the last Wednesday of the rst
half of each year appears in this sample. Both the ad hoc BS and the GARCH model are estimated each week using ordinary
least squares and non-linear least squares respectively. For the GARCH model, variance, h(t +1) is drawn from the daily history
(last 252 days) of S&P 500 levels (closest to and before 2:30 P.M., central standard time).
to around 5 cents in 1992 and 1993. Thus the exibility of updating appears
to make a difference in terms of its ability to t options prices in-sample.
Table 4 reports the mean and standard deviations of the updated GARCH
coefcients from this estimation. As we can see, the parameter that is least
stable is , followed by 1 . The parameters, 1 and 1 are relatively more
stable. This should not be too surprising because option values are more
sensitive to 1 (that measures the volatility of volatility), and 1 (that controls
the skewness of index returns) than they are to the other parameters. This
stability is important for the GARCH model to t the data reasonably well
even with constant parameters.
Although both the ad hoc BS and the updated GARCH formulas appear
quite exible, there is an important distinction between the two functional
forms. The ad hoc BS formula is a function of the current index level. But
the updated GARCH model is a function of the historical path of the index.
Both models are consistent with the skew in implied volatilities and a term
structure of volatility. However, the shape of the ad hoc model is constant
by denition, whereas the shape of the GARCH option valuation function
changes depending on the past sequence of spot returns. Of course exibility is not the only criterion of an option valuation formula. DFW (1998)
Table 4
Mean estimates from the updated GARCH model using non-linear least squares
Parameter
1
1
1
1
Mean
4.06e 6
0.159
430.7
1.10e6
Standard Deviation
7.52e7
0.084
45.31
1.24e6
This table reports the mean and standard deviation of the parameter estimates from the weekly (every week) estimation of the
GARCH model (using option prices in the second half of each year, 1992, 1993 and 1994) using non-linear least squares. Note
however that the last Wednesday of the rst half of each year appears in this sample. Variance, h(t + 1) is drawn from the daily
history (last 252 days) of S&P 500 levels (closest to and before 2:30 P.M., central standard time).
605
have shown that a more exible model may dominate in-sample but have
much less predictive power for out-of-sample options prices. This occurs
when a misspecied model achieves good in-sample results by overtting
the data. We examine this issue in the next section by comparing all the
models out-of-sample.
3.2 Out-of-sample model comparison
Having estimated the parameters in-sample from the rst six months of
each year, we turn to out-of-sample valuation performance of the unrestricted/asymmetric GARCH model for the next six months of each the
three years under consideration. In computing out-of-sample option values
for the second half of a particular year for the non-updated GARCH model,
we keep the parameters xed at their in-sample estimates for the particular
year and obtain the conditional variance, h(t + 1), from the dynamics of
daily asset returns. In updating the variance we use the same starting variance, h(0) as in the in-sample estimation and for any given time, t, in the
out-of-sample period, obtain h(t + 1) from the entire daily history of asset
prices for that year23 (given the in-sample estimates). Therefore the entire
out-of-sample options calculations for the non-updated GARCH model are
based on options prices from the rst 6 months for that year. The BS, ad hoc
BS, and updated GARCH models are also estimated every week in the second half of each year. For the BS, the estimated implied volatility from the
current week is used to value options in the next week. For the ad hoc
BS and updated GARCH models, the estimated parameters from the current
week are used to value options in the next week. For the updated GARCH
model, at each time, t, in the out-of-sample period, the variance, h(t + 1)
is drawn from the daily history of S&P 500 levels including and up to the
rst day in the history of S&P 500 levels that was used for getting the insample estimates in the previous week.24 The important distinction between
the out-of-sample implementations is that the non-updated GARCH model
predicts options values up to 26 weeks ahead, whereas the BS, ad hoc BS,
and updated GARCH models only predict one week ahead.
Table 5 (Panel A) reports the out-of-sample valuation errors for the various models aggregated across all three out-of-sample periods. The aggregate
root mean squared valuation errors are $1.14, $0.771, $0.737 and $0.58,
respectively for the BS, ad hoc BS, non-updated GARCH and the updated
GARCH respectively. Recall the BS formula ts worst in-sample and it is
still the worst performer out-of-sample. The biggest deterioration occurs with
23
Daily S&P 500 levels (closest to and before 2:30 P.M, CST). Note that the sample of option prices for each
day start at 2:30 P.M (CST).
24
For example, in computing out-of-sample option values on a given Wednesday, we would use the daily
history of S&P 500 levels (closest to and before 2:30 P.M, CST) that go up to 252 days prior to the previous
Wednesday.
606
Table 5
Out-of-sample valuation errors
Panel A: Aggregate valuation errors across all years
RMSE
MAE
MOE
Average
option price
Number of
observations
1.14
0.771
0.737
0.58
0.807
0.47
0.492
0.373
0.125
0.011
0.057
0.031
13.1
13.1
13.1
13.1
4944
4944
4944
4944
RMSE
MAE
MOE
Average
option price
Number of
observations
1992
BS
Ad hoc BS
GARCH (non-updated)
GARCH (updated)
1.058
0.689
0.833
0.479
0.76
0.418
0.571
0.303
0.09
0.022
0.366
0.027
13.01
13.01
13.01
13.01
1545
1545
1545
1545
1993
BS
Ad hoc BS
GARCH (non-updated)
GARCH (updated)
1.138
0.899
0.689
0.624
0.789
0.544
0.454
0.401
0.161
0.008
0.029
0.008
13.20
13.20
13.20
13.20
1500
1500
1500
1500
1994
BS
Ad hoc BS
GARCH (non-updated)
GARCH (updated)
1.229
0.725
0.691
0.637
0.894
0.448
0.452
0.417
0.135
0.002
0.28
0.128
13.09
13.09
13.09
13.09
1899
1899
1899
1899
BS
Ad hoc BS
GARCH (non-updated)
GARCH (updated)
Panel B: Valuation errors by years
Panel A reports the aggregate (across the three years, 1992, 1993 and 1994) out-of-sample valuation errors (in $) for all options
by various models. Panel B reports the out-of-sample valuation errors (in $) by each year (1992, 1993 and 1994). Option values
are computed every Wednesday (or the next trading day) in the second half of each year. For the GARCH (non-updated) model,
option values are computed by holding the parameters at their in-sample estimates from Table 2 and updating the variance from
the daily S&P 500 levels (closest to and before 2:30 P.M., central standard time). For the GARCH (updated) model, parameters
are estimated in the previous week and variance is computed from the history of the daily S&P 500 levels (closest to and
before 2:30 P.M., central standard time). BS is the BlackScholes model in which a single implied volatility is estimated across
all strikes and maturities on a given day while ad hoc BS is an ad hoc version of the BlackScholes model with strike and
maturity specic implied volatilities; both the BS and ad hoc BS are estimated every week and then used to value options in
the following week. RMSE is the root mean squared out-of-sample/prediction valuation error in dollars. MOE (in $) or the
mean outside error measures the mean valuation error outside the bid-ask spread (difference between the model value and the
ask price if the model value exceeds the ask price or the difference between the model value and the bid price if the bid price
exceeds the model value). MAE (in $) or the mean absolute error is the average absolute value of the valuation errors outside
the bid-ask spread.
the ad hoc BS model. Although it has a competitive t in-sample, it underperforms both the non-updated GARCH and the updated GARCH one week
out-of-sample. Table 5 (Panel B) reports the out-of-sample RMSE for each
out-of-sample period. We nd that the ad hoc BS is superior to the nonupdated GARCH in 1992, but is outperformed enough by the non-updated
GARCH in 1993 and 1994 so that the aggregate errors are lower under the
non-updated GARCH. Note however that the updated GARCH substantially
outperforms the ad hoc BS in all three out-of-sample periods.
Although the ad hoc BS model is exible, it achieves a tight in-sample t
only by overtting the data. In contrast the GARCH model holds up surprisingly well. It continues to predict options values in the second half of the
year using parameter estimates from only the rst half. Furthermore, despite
607
using a variance h(t + 1) ltered from the time series of index returns, it outperforms the BS and ad hoc BS formulas calibrated to the previous weeks
options. The updated GARCH model, however, provides the best predictive
t. While it is only slightly better than the ad hoc BS model in-sample, it
does not suffer as much deterioration in out-of-sample t. Consequently it is
the best model both in-sample and out-of-sample.
Note that the updated GARCH model ts the dynamics of index returns
jointly with the pattern of option prices across strike price and maturity on a
given day. Thus it is somewhat restricted in that the spot variance, h(t + 1)
is not a free parameter. However, this restriction improves the predictions. In
unreported diagnostics we have also estimated the updated GARCH model
with h(t + 1) as a tted parameter, instead of ltering h(t + 1) from the historical time-series of index returns. This resembles Bakshi, Cao and Chens
(1997) estimation of continuous-time stochastic volatility models.25 Naturally
this improves the in-sample t due to increased degrees of freedom. But it
does not improve out-of-sample predictions (the variance is updated from
the asset prices to compute out-of-sample values). Compared with the earlier
updated GARCH, the retting of all the parameters increases the out-ofsample root mean squared error by an average of 5 cents (across all years).
Note that even with this increased error the updated GARCH model still
produces better out-of-sample predictions than other models. But it produces
the best predictions by simultaneously using the information in option prices
and the history of the index.
Table 5 (Panel A and Panel B) also reports out-of-sample mean absolute
error (MAE) that measures the absolute values of the valuation errors outside the bid-ask spread for all options.26 The aggregate (i.e. across the three
years) out-of-sample MAEs are $0.807, $0.47, $0.492 and $0.373 for the BS,
ad hoc BS, the non-updated GARCH and the updated GARCH respectively.
Table 6 and Table 7 report the valuation errors (both the RMSE in dollars and
the percentage valuation error) by different option moneyness and maturity
categories for puts and calls respectively. They also report the average error
outside the bid-ask spread for each category. Looking at the valuation errors
by moneyness, maturity and option type (i.e. a call or a put), we nd that the
GARCH model (both updated and non-updated) is able to value deep out-ofthe-money options (K/F < 0.95 for puts and K/F > 1.05 for calls) better
for all maturities than the ad hoc BS. For example, the RMSE for deep outof-the-money puts (K/F < 0.95) that have less than forty days to mature
is 28.2 cents for the non-updated GARCH and 22.7 cents for the updated
GARCH versus 42.3 cents for the ad hoc BS and 55.4 cents for the BS
25
The parameter turned out to be negative in three (out of 68 cases) using this kind of estimation. A negative
does not guarantee the long-run variance to be positive, although the updated, one week ahead variance
was positive due to the relatively higher value of the estimated h(t + 1) with respect to .
26
Note that the MAE is not directly comparable to the RMSE because MAE is dened only when the valuation
error does not fall within the bid-ask spread.
608
0.554
0.572
0.873
0.953
0.418
0.423
0.466
0.595
0.553
0.366
0.282
0.581
0.773
0.522
0.311
0.227
0.539
0.649
0.459
0.315
<0.95
[0.950.99)
[0.99,1.01)
[1.01,1.05]
>1.05
<0.95
[0.950.99)
[0.991.01)
[1.011.05]
>1.05
<0.95
[0.95,0.99)
[0.991.01)
[1.011.05]
>1.05
RMSE
<0.95
[0.950.99)
[0.99,1.01)
[1.01,1.05]
>1.05
Moneyness
38.08
29.43
12.89
3.44
1.06
47.14
31.70
15.35
3.91
1.05
70.94
25.45
11.81
4.14
1.23
92.79
31.22
17.34
7.13
1.41
% Error
<40
0.00
0.291
0.275
0.005
0.220
0.102
0.142
0.245
0.018
0.176
0.127
0.086
0.254
0.241
0.016
0.406
0.277
0.522
0.646
0.194
MOE
0.361
0.627
0.683
0.526
0.404
0.514
0.787
0.940
0.844
0.415
0.545
0.697
0.762
0.802
0.421
1.186
0.954
0.965
1.331
1.017
RMSE
19.22
14.44
8.49
3.33
1.36
27.38
18.13
11.68
5.35
1.40
29.03
16.06
9.47
5.08
1.42
63.18
21.98
12.00
8.43
3.43
% Error
[4070]
MOE
0.022
0.342
0.417
0.079
0.124
0.176
0.181
0.489
0.223
0.029
0.160
0.192
0.206
0.002
0.024
1.011
0.566
0.513
0.935
0.802
Days to Expiration
0.507
0.698
0.843
0.763
0.516
0.612
0.847
1.118
1.149
0.718
1.035
0.709
0.795
1.087
1.018
1.645
1.322
0.965
1.589
1.475
RMSE
>70
16.80
11.57
8.25
4.39
1.75
20.31
14.03
10.94
6.60
2.44
34.31
11.76
7.78
6.25
3.45
54.52
21.90
9.44
9.13
5.01
% Error
0.101
0.164
0.477
0.395
0.090
0.155
0.233
0.634
0.619
0.297
0.427
0.224
0.059
0.446
0.593
1.452
0.953
0.340
1.129
1.104
MOE
Reports out-of-sample valuation errors by moneyness and maturity for put options. Moneyness is dened to be K/F where K is the strike, F is the forward price. RMSE and
MOE, both in dollars are as dened in Table 5. % error is the ratio of the RMSE to the average option price for that option category.
GARCH (updated)
GARCH (non-updated)
Ad hoc BS
BS
Model
Out-of-sample valuation errors for put options
Table 6
609
610
RMSE
0.937
0.826
0.701
0.785
0.337
0.628
0.640
0.512
0.405
0.163
0.672
0.687
0.685
0.427
0.161
0.607
0.621
0.625
0.402
0.154
Moneyness
<0.95
[0.95, 0.99)
[0.99,1.01)
[1.01, 1.05]
>1.05
<0.95
[0.95, 0.99)
[0.99, 1.01)
[1.01, 1.05]
>1.05
<0.95
[0.95, 0.99)
[0.99, 1.01)
[1.01, 1.05]
>1.05
<0.95
[0.95, 0.99)
[0.99, 1.01)
[1.01, 1.05]
>1.05
1.80
4.13
12.64
35.15
87.30
2.00
4.57
13.85
37.36
91.44
1.86
4.26
10.35
35.42
92.32
2.78
5.49
14.17
68.63
191.82
% Error
<40
0.310
0.015
0.060
0.192
0.093
0.441
0.140
0.024
0.141
0.097
0.199
0.243
0.064
0.083
0.016
0.606
0.470
0.333
0.588
0.232
MOE
0.593
0.609
0.590
0.455
0.295
0.730
0.822
0.805
0.595
0.265
0.687
0.898
0.826
0.659
0.842
1.410
1.217
0.763
1.151
0.919
RMSE
1.73
3.39
7.20
15.35
50.51
2.12
4.58
9.82
20.08
45.32
2.00
5.00
10.08
22.24
144.13
4.10
6.78
9.30
38.80
157.19
% Error
[4070]
MOE
0.160
0.050
0.252
0.051
0.185
0.360
0.059
0.269
0.073
0.156
0.039
0.522
0.393
0.097
0.235
0.946
0.768
0.328
0.916
0.732
Days to Expiration
0.860
0.720
0.644
0.526
0.341
0.927
0.899
0.953
0.826
0.391
1.186
0.947
0.941
1.189
1.581
2.157
1.663
0.747
1.155
1.189
RMSE
>70
2.40
3.55
6.08
11.52
28.25
2.59
4.43
9.00
18.09
32.34
3.31
4.67
8.88
26.04
130.96
6.03
8.19
7.05
25.32
98.46
% Error
0.502
0.283
0.100
0.089
0.135
0.523
0.212
0.270
0.269
0.020
0.174
0.571
0.357
0.168
0.692
1.657
1.226
0.096
0.817
1.005
MOE
Out-of-sample valuation errors by moneyness and maturity for call options. Moneyness is dened to be K/F where K is the strike and F is the forward price. RMSE and MOE,
both in dollars are as dened in Table 5. % error is the ratio of the RMSE to the average option price for that option category.
GARCH (updated)
GARCH (non-updated)
Ad hoc BS
Model
BS
Out-of-sample valuation errors for call options
Table 7
(see Table 6). For near-the-money options, the results are mixed. We nd that for
short-term (<40 days to expiration) near-the-money (0.99 <= K/F < 1.01)
call options, the ad hoc BS model has lower valuation errors than both versions of the GARCH. However, for medium-term (4070 days) and long-term
(> 90 days) near-the-money calls, the updated GARCH has lower valuation errors than the ad hoc BS while for medium-term near-the-money call
options, the non-updated GARCH dominates the ad hoc BS. In contrast, for
near-the-money put options, irrespective of maturity, the ad hoc BS is better
than the non-updated GARCH, but not necessarily better than the updated
GARCH. In terms of maturity only, the percentage valuation errors under
the GARCH tend to decrease with an increase in maturity, especially for
out-of-the-money options. Short-term (<40 days to expire) out-of-the-money
options often tend to be the most difcult to value (in terms of percentage valuation error) under both GARCH and the ad hoc BS, although the magnitude
of valuation errors under the GARCH is substantially lower. Figures 2A, 2B
and 2C show the absolute percentage valuation errors for put options of the
three different maturities by moneyness while gures 3A, 3B and 3C show
the same for call options.
The superior out-of-sample valuation performance of the GARCH model
is especially encouraging in the context of results reported by DFW (1998).
Figure 2A
This gure shows the percentage out-of-sample valuation errors (i.e. 100RMSE/Option Price) for put options
(less than 40 days to mature) by various models in the second half of each year. Out-of-sample option values
are computed every Wednesday. BS is the BlackScholes model and ad hoc BS is the ad hoc BlackScholes
model. BS and ad hoc BS are estimated from option prices every week. GARCH-non-updated is the GARCH
model in which parameters have been estimated in the rst half of each year, while GARCH-updated is the
GARCH model in which parameters have been estimated in the previous week.
611
Figure 2B
(between 40 and 70 days to mature) by various models in the second half of each year. K is the strike and F
is the forward price. Out-of-sample option values are computed every Wednesday. BS is the BlackScholes
model and ad hoc BS is the ad hoc BlackScholes model. BS and ad hoc BS are estimated from option prices
every week. GARCH-non-updated is the GARCH model in which parameters have been estimated in the rst
half of each year, while GARCH-updated is the GARCH model in which parameters have been estimated in
the previous week.
Figure 2C
(more than 70 days to mature) by various models in the second half of each year. K is the strike and F is the
forward price. Out-of-sample option values are computed every Wednesday. BS is the BlackScholes model
and ad hoc BS is the ad hoc BlackScholes model. BS and ad hoc BS are estimated from option prices every
week. GARCH-non-updated is the GARCH model in which parameters have been estimated in the rst half
of each year, while GARCH-updated is the GARCH model in which parameters have been estimated in the
previous week.
612
Figure 3A
This gure shows the percentage out-of-sample valuation errors (i.e. 100 RMSE/Option Price) for call
options (less than 40 days to mature) by various models in the second half of each year. K is the strike and
F is the forward price. Out-of-sample option values are computed every Wednesday. BS is the BlackScholes
the previous week.
Figure 3B
options (between 40 and 70 days to mature) by various models in the second half of each year. K is the
strike and F is the forward price. Out-of-sample option values are computed every Wednesday. BS is the
BlackScholes model and ad hoc BS is the ad hoc BlackScholes model. BS and ad hoc BS are estimated
from option prices every week. GARCH-non-updated is the GARCH model in which parameters have been
estimated in the rst half of each year, while GARCH-updated is the GARCH model in which parameters
have been estimated in the previous week.
613
Figure 3C
options (more than 70 days to mature) by various models in the second half of each year. K is the strike and
F is the forward price. Out-of-sample option values are computed every Wednesday. BS is the BlackScholes
the previous week.
They found that the deterministic volatility models inspired by the implied
binomial trees of Derman and Kani (1994), Dupire (1994), and Rubinstein
(1994) [see Brown and Toft (1999) for an extension of Rubinstein (1994)]
underperform the ad hoc BS model in out-of-sample valuation tests in the
S&P 500 index options market. While the deterministic volatility models
can generate negative skewness in the distribution of asset returns, they are
Markovian and assume that volatility is a function of the index level and time.
The results of DFW show that this path-independence assumption is poorly
specied for options prices. Path-independent volatility dynamics also contradict the extensively documented empirical literature on ARCH-GARCH
effects or path dependence in volatility. The deterministic volatility models
are quite capable of tting the volatility smile of option prices in-sample.
But they do not properly relate this shape to the path-dependent dynamics
of volatility. Consequently these models do not properly update the volatility
using subsequent index returns, leading to poor out-of-sample option predictions. In contrast the GARCH model is able to both t the shape of options
prices and use this information to predict volatility on the basis of subsequent
index returns. The negative correlation between returns and volatility is an
important element of this relationship.
614
3.3 Using S&P 500 futures to lter volatility

The GARCH model uses the history of asset prices (S&P 500 levels) to lter
the variance, h(t + 1), that is needed to value options on day t, while the
ad hoc BS and the simple BS model (with a single implied volatility) do not
use the history of asset prices. The in-sample and out-of-sample valuation
errors reported in the previous sections are computed using volatility ltered
from the history of S&P 500 cash/spot levels. If stale price quotations affect
the measured S&P 500 cash index levels then the volatilities are contaminated
by the accumulated history of these quotation errors. Table 1 reports the
slightly different volatility dynamics computed (using maximum likelihood
estimation) from the history of S&P 500 cash/spot data and the history of
S&P 500 futures data. A related question is how robust our options valuation
results are if we use the S&P 500 futures to lter volatility. In other words,
one needs to explore how minor changes in data affect our option valuation
results. In order to address this issue, we rst create a daily time series of
the prices of lead/closest maturity S&P 500 futures contracts from the intraday data on S&P 500 futures. Specically as with the cash/spot, we use the
closest to (before) 2:30 P.M. (CST) futures prices to lter volatility. However,
we retain the same parameter estimates that were obtained from the previous
in-sample optimizations (on the options data) using the S&P 500 cash/spot
index. Table 8 reports the in-sample and out-of-sample RMSE in the various
periods for the two versions of the GARCH model i.e. the non-updated and
the updated GARCH model.
The out-of-sample RMSE for the non-updated GARCH model using S&P
500 futures (to lter variance) are 85 cents, 69 cents and 69.6 cents for 1992,
1993 and 1994 respectively. Recall that the corresponding out-of-sample
RMSE using variance ltered from the S&P 500 cash index for the nonupdated GARCH model are 83.3 cents, 68.9 cents and 69.1 cents respectively.
Table 8
In-sample and out-of-sample valuation errors using S&P 500 futures
RMSE(in-sample)
RMSE (out-of-sample)
1992
GARCH (non-updated)
GARCH (updated)
0.711
0.40
0.85
0.498
1993
GARCH (non-updated)
GARCH (updated)
0.723
0.505
0.69
0.619
1994
GARCH (non-updated)
GARCH (updated)
0.689
0.527
0.696
0.617
Reports the dollar root mean squared valuation errors (RMSE) for the in-sample and out-of-sample estimation (for 1992, 1993
and 1994) for the two versions of the GARCH model if the variance, h(t + 1) is ltered from the S&P 500 levels that are
implied from the S&P 500 futures. The nearest maturity futures with prices closest to (before) 2:30 P.M. (central standard time)
are used. Parameter estimates are xed at the ones obtained from the in-sample estimations of the respective models using the
S&P 500 cash/spot levels.
615
This shows that there is a maximum 1.7 cents difference for out-of-sample
valuation errors if we change the data used to lter variance. For the updated
GARCH model, the out-of-sample RMSE (using S&P 500 futures) are 49.8
cents, 61.9 cents and 61.7 cents for 1992, 1993 and 1994 respectively. Recall
that the corresponding RMSE using variance ltered from the S&P 500 cash
index for the updated GARCH model are 47.9 cents, 62.4 cents and 63.7
cents respectively. Thus at most, there is a 2 cent difference in terms of outof-sample valuation errors for the updated GARCH model. Therefore using
futures data to lter volatility does not change our primary result that the
GARCH model outperforms the BS and the ad hoc BS out-of-sample and
the quantitative implications of using index returns implied from the S&P
500 futures prices are quite trivial.
For in-sample valuation errors, the RMSE from the updated GARCH
model using the S&P 500 futures are 40 cents, 50.5 cents and 52.7 cents for
1992, 1993 and 1994 respectively. Recall that the corresponding in-sample
RMSE using the S&P 500 cash index are 36.6 cents, 48.3 cents and 45.9 cents
respectively. Thus the differences (between futures and spot) in in-sample
valuation errors are higher than the differences in out-of-sample valuation
errors and the same is true of the differences in in-sample valuation errors if
we use the non-updated GARCH model. These in-sample differences would
undoubtedly be smaller if we re-estimated the parameters of the GARCH
model from the options data based on the volatility ltered using the futures
data. We do not pursue this because it would confound the effects of changing parameters with the effects of changing data. The empirical goal of our
paper is to compare models in terms of out-of-sample performance, and the
results demonstrate that the out-of-sample superiority of the GARCH model
is insensitive to the data used to lter volatility.
4. Conclusion
This paper presents a closed-form solution for option values (and hedge
ratios) when the variance of the spot asset follows a GARCH(p, q) process
and is correlated with asset returns. The discrete-time GARCH model with
single lag converges to Hestons (1993) continuous-time stochastic volatility
model as the observation interval shrinks. In this limiting case the formula
gives a unique option value that is based on the absence of arbitrage only
(the option can be replicated by trading in the underlying asset and the risk
free asset). This limit is a path-dependent continuous-time model of the type
suggested by DFW (1998) to overcome the limitations of the path independent implied binomial tree models. In practice the numerical results of the
discrete-time GARCH model with daily increments are quite close to those of
the continuous-time model. But unlike continuous-time stochastic volatility
models, the GARCH model can be easily estimated from observing only the
history of asset prices. As the model has closed-form solutions for option values, one can easily combine the information in the cross-section of options
616
with the time-series information in the history of asset prices to estimate

model parameters.
Empirical analysis of the single lag version of the GARCH model on the
S&P 500 index options data shows out-of-sample valuation improvements
over a exible ad hoc version of the BS model. The ad hoc BS model uses
a separate implied volatility for each option (specic to its strike and time
to maturity) extracted from market prices and is designed to produce a very
close t to the shape of the implied volatilities across strike prices and maturities; also it is updated every period. In contrast, the GARCH model lters
the volatility from the history of asset prices and its parameters are held
constant over the sample periods. If the parameters of the GARCH model
are updated every period, the out-of-sample valuation improvements over the
ad hoc model are more substantial. However, the GARCH model underperforms even the simple BS model when it does not allow correlation between
index returns and volatility. DFW (1998) have shown that models which capture this correlation, but do not allow for path dependence in volatility are
dominated by the ad hoc BS model in out-of-sample valuation errors in the
S&P 500 index options market. We conclude that the out-of-sample valuation improvements of the GARCH model depends on the models ability
to simultaneously capture the path-dependence in volatility and the correlation of volatility with asset returns. Also our results are robust to using the
history of S&P 500 futures to lter the variance instead of the S&P 500
cash index.
Although the empirical analysis in this paper has focused on the single lag
version of the GARCH model, one could estimate the model with multiple
lags. Since Assumption 2 states that the GARCH model is identical to the
BS model for one-period options, additional lags will probably not improve
valuation errors of short-term options. One might also want to incorporate
a fat-tailed distribution of the one step ahead index returns. Also instead of
ltering the variance from daily returns, one could potentially use intra-day
levels of S&P 500 to update the variance intra-day. Attention could also be
given to how the existence of huge open interest in deep-out-of-the-money
puts could potentially impact the prices of these puts, a subject that does not
lie within the realm of this paper, but may be important in itself. Although
this paper has focused on equity/index options, the model can also be easily
applied to currency options and quantos. One can also extend the model
to bond options by assuming that (continuously compounded) interest rates
follow an autoregressive moving-average process with the GARCH effects of
equation (1b). This leads to a family of log-linear (afne) bond models that
in continuous-time limit nest the diffusion models of Vasicek (1977), Heston
(1990) and others.
617
Appendix A. Derivation of the Generating Function and Option Formula

Proof of Proposition 2.
Derivation of the Generating Function:
Let x(t) = log(S(t)) and let f (t; T , ) be the conditional generating function of S(T ), or
equivalently the conditional moment generating function of x(T ) i.e.,
f (t; T , ) = Et [exp(x(T ))].
(A1)
We shall guess that the moment generating function takes the log-linear form

f (t; T , ) = exp x(t) + A(t; T , )
+
p

i=1
Bi (t; T , )h(t + 2 i) +
q1
i=1
Ci (t; T , )

(z(t + i) i h(t + i))2 ,
(A2)
and solve for the coefcients A(), Bi () and Ci () as in Ingersoll (1987, p. 397), utilizing the fact
that the conditional moment generating function is exponential afne in the state variables, x(t)
and the h(t)s. The fact that the conditional moment generating function is exponential afne
can be easily veried by calculating the moment generating functions for x(t + 1), x(t + 2) and
so on. Equation (A2) species the general form of this function for x(T ).
Since x(T ) is known at time T , equations (A1) and (A2) require the terminal condition
A(T ; T , ) = Bi (T ; T , ) = Ci (T ; T , ) = 0.
(A3)
Applying the law of iterated expectations to f (t; T , ), we get,

f (t; T , ) = Et f (t + ; T , ) = Et exp x(t + ) + A(t + ; T , )
+
p

i=1
Bi (t + ; T , )h(t + 3 i) +
q1
i=1
Ci (t + ; T , )
2

.
z(t + 2 i) i h(t + 2 i)
Substituting the dynamics of x(t) in equations (1a) and (1b) shows

f (t; T , ) = Et exp x(t) + r + h(t + ) + h(t + )z(t + )
+ A(t + ; T , ) + B1 (t + ; T , )

2

1 h(t + ) + 1 z(t + ) 1 h(t + )

p1

i+1 h(t + 2 i)
+ B1 (t + ; T , ) +
i=1
q1
i=1
618
2

i+1 z(t + 2 i) i+1 h(t + 2 i)
(A4)
p1
i=1
Bi+1 (t + ; T , )h(t + 2 i) + C1 (t + ; T , )

2 q2

Ci+1 (t + ; T , )
z(t + ) 1 h(t + ) +
i=1
2
z(t + i) i+1 h(t + i)

(A5)
Rearranging terms through completing squares and some algebra shows

f (t; T , ) = Et exp (x(t) + r) + A(t + ; T , ) + B1 (t + ; T , )

+ B1 (t + ; T , )1 + C1 (t + ; T , ) z(t + )

1
2(B1 (t + ; T , )1 + C1 (t + ; T , ))
2

h(t + ) + + B1 (t + ; T , )1 + 1

2
h(t + )
4(B1 (t + ; T , )1 + C1 (t + ; T , ))
+ B1 (t + ; T , )
p1
i=1
i+1 h(t + 2 i) +
p1
i=1
Bi+1
(t + ; T , )h(t + 2 i) + B1 (t + ; T , )

2
z(t + 2 i) i+1 h(t + 2 i)
+
q2
i=1
q1

i=1
i+1

Ci+1 (t + ; T , ) z(t + i)

2
.
i+1 h(t + i)
(A6)
A useful result is that for a standard normal variable z

1

ab2
.
E exp(a(z + b)2 ) = exp ln(1 2a) +
2
1 2a
(A7)
Substituting this result in (A6) and subsequently equating terms in both sides of (A6) shows
A(t;T ,) = A(t +;T ,)+r +B1 (t +;T ,)

1
ln 121 B1 (t +;T ,)2C1 (t +;T ,) .
2
B1 (t; T , ) = ( + 1 )
(A8)
1 2
+ 1 B1 (t + ; T , ) + B2 (t + ; T , )
2 1
619
1/2( 1 )2
1 21 B1 (t + ; T , ) 2C1 (t + ; T , )
(A9)
Bi (t; T , ) = i B1 (t + ; T , ) + Bi+1 (t + ; T , ), for 1 < i p,

Ci (t; T , ) = i+1 B1 (t + ; T , ) + Ci+1 (t + ; T , ), for 1 < i q 1.
One can use equations (A8) and (A9) to calculate the coefcients recursively starting with
equation (A3).
Proof of Proposition 3.
Let f () denote the moment generating function of the probability density, p(x), where x is
the logarithm of the terminal asset price.27 Let p (x) be an adjusted probability density dened
by p (x) = exp(x)p(x)/f (1).28 It is easy to see that it is a valid probability density because it
is non-negative and f (1) = Et [exp(x(T ))] from equation (6). The moment generating function
for p (x) is

exp(x)p (x)dx =
=
1
exp(( + 1)x)p(x)dx
f (1)
f ( + 1)
.
f (1)
(A10)
Since the spot asset price is exp(x), the expectation (at time t) of a call option payoff separates
into two terms with probability integrals.
E[Max(ex K, 0)] =
ln(K)
= f (1)
exp(x)p(x)dx K

ln(K)
p (x)dx K

ln(K)
ln(K)
p(x)dx
p(x)dx.
(A11)
Note that f (i) is the characteristic function corresponding to p(x) and f (i + 1)/f (1) is
the characteristic function corresponding to p (x). Feller [1971] and Kendall and Stuart [1977]
show how to recover the probabilities from the characteristic functions

ln(K)
p(x)dx =
i ln(K)
1
f (i)
1
e
+
d,
Re
2
0
i
(A12)
and similarly the other integral of p (x). Substituting equation (A12) into (A11) proves the
proposition and noting that under the risk neutral distribution, S(t) = er(T t) E[S(T )], demonstrates the corollary.
As mentioned before, the option values under our model can be gotten very easily with
a numerical integration routine as the integrand converges very fast. For example, using the
parameters from the MLE estimation (of the unrestricted/asymmetric GARCH model) on daily
spot index returns that appear in Table 1, with S = K = $100, h(t + 1) = (0.15 0.15)/252.0
and setting the interest rate to zero, call options with 50 and 100 days to expiration have values
of $1.817 and $2.481 respectively.29
27
Note that p(x) is basically the conditional (at time t) density function of x(T ).
28
Note that p (x) is not the density corresponding to f ().
29
These option values were generated by using the integration routine qromo() of Press et al. (1992) as
well as by using Gauss-Hermite quadrature (Press et al. (1992)) after normalizing the integrand by a factor
proportional to the square-root of expected future variance (until the option expires).
620
Appendix B. Convergence to Continuous Time

In the single factor/single lag model the conditional mean and variance of h(t) are
Et [h(t + )] = + 1 + (1 + 1 12 )h(t).
Var t [h(t + )] = 12 (2 + 412 h(t)).
(B1)
(B2)
There are various ways to approach a continuous-time limit as the time interval shrinks.
Since h(t) is the variance of the spot return over time interval , it should converge to zero.
To measure the variance per unit of time we dene v(t) = h(t)/ and v(t) has a well dened
continuous-time limit. The stochastic process, v(t) follows the dynamics

v(t + ) = v + v v(t) + v (z(t) v v(t))2 ,
(B3)
where,
v =
, v = 1 , v = 1 , v = 1 .

Let 1 () = 41 2 2 , 1 () = 0, () = ( 41 2 )2 , 1 () = 2/( ) / , and

() = . Noting that v(t + ) is observable at time t and taking conditional expectation at
time t ,
Et [v(t + ) v(t)] = ( v(t)) +

Var t [v(t + )] = 2 v(t) +
1 2
v(t)2 .
4

4
22
2 v(t) +
v(t) 2 .
8
4
(B4)
(B5)
(Note that 1 , 1 , , 1 as dened above are not v , v , v , and v corresponding to the v(t)
process). The correlation between the variance process and the continuously compounded stock
return is

sign(v ) 2v2 v(t)
.
(B6)
Corr t [v(t + ), log(S(t))] =

1 + 2v2 v(t)
As the time interval shrinks the skewness parameter, v () approaches positive or negative
innity. Consequently the correlation in equation (B6) approaches 1 (or negative 1) in the limit.
The variance process, v(t) has a continuous-time diffusion limit following Foster and Nelson
(1994). As the observation interval shrinks, v(t) converges weakly to the square-root process
of Feller (1951), Cox, Ingersoll Ross (1985), and Heston (1993)
d log(S) = (r + v)dt + vdz
dv = ( v)dt + vdz,
(B7a)
(B7b)
where z(t) is a Wiener process. Note that the same Wiener process drives both the spot asset and
the variance. The limiting behavior of this GARCH process is very different from those of other
GARCH processes such as GARCH 1-1 [Bollerslev (1986)] or many other asymmetric GARCH
processes in which two different Wiener processes drive the spot assets and the variance. Also,
while the above shows that the asset returns and variance processes under the data generating
measure converge to well-dened continuous-time limits, one still needs to verify that the discrete risk-neutral processes converge to appropriate continuous-time limits if the discrete-time
GARCH option values are to converge to their continuous-time limits.
621
As shown in Proposition 1, in the risk-neutral distribution, 1 equals 1 + + 1/2. Therefore,

the risk-neutral parameter for the v(t) process is,
v () = 1 () =

1
.
(B8)
Consequently the risk-neutral process has a different mean,

1
v(t)
[v(t + ) v(t)] = ( v(t)) + +
Et
2

2

1
1
+
+ +
v(t)2 .
4
2
(B9)
Again following Foster and Nelson (1994), it follows that the continuous-time risk-neutral processes are
v
(B10a)
dt + vdz
d log(S) = r
2

1
dv = ( v) + +
(B10b)
v dt + vdz ,
2
where z(t) is a Wiener process under the risk-neutral measure. As with the data generating
measure, the same Wiener process drives both asset returns and variance under the risk-neutral
measure. The above risk-neutral processes for the stock price and the variance are equivalent
to the risk-neutral processes of Heston (1993) with the two Wiener processes therein being
perfectly correlated. Consequently, the discrete-time GARCH option values converge to the
continuous-time option values of Heston (1993) as shrinks; such convergence has been veried
numerically. Figure 4 shows how the discrete time GARCH model converges to the continuoustime model as decreases (i.e. as the number of trading periods increases). The parameters
Figure 4
This gure shows how the discrete-time GARCH option values converge to the continuous-time option value
with an increase in the number of trading intervals.
622
used for an at-the-money option with a spot asset price, S = $100, strike price, K = $100
with 0.5 years to maturity are, ( + 1/2) = 2, = 0.02, = 1, = 0.1, v = 0.01.
Given these parameters, one can directly use Hestons (1993) formula to compute the value of
an option.
As the two Wiener processes are perfectly correlated, one can value options solely by the
absence of arbitrage only using the hedging arguments of Black and Scholes (1973) and Merton (1973) or equivalently by showing the existence of a unique risk-neutral distribution as
per Cox and Ross (1976), Harrison and Kreps (1979), Harrison and Pliska (1981). In this case
Assumption 2 is superuous and merely states that options are properly priced at maturity.
Note that although returns and volatility are perfectly correlated instantaneously in the continuous time model, they are imperfectly correlated over any discrete time interval. Also note
that the parameter related to the asset risk premium, can inuence option values unlike in
the BlackScholes-Merton setup due to the fact that the asset price process is non-Markovian.
Since variance is a function of only the uncertainty emanating from changes in the asset prices
and not driven by a separate Wiener process, the parameter related to the asset risk premium,
appears in the drift of the risk-neutral variance instead of a volatility risk premium as in Heston
(1993). This type of result has also been noted in Kallsen and Taqqu (1998).
References
Amin, K., and V. Ng, 1993, ARCH Processes and Option Valuation, working paper, University of Michigan.
Bakshi, G., C. Cao, and Z. Chen, 1997, Empirical Performance of Alternative Option Pricing Models,
Journal of Finance, 52, 20032049.
Bakshi, G., C. Cao, and Z. Chen, 1998, Pricing and Hedging Long-Term Options, Working Paper, University
of Maryland; forthcoming in Journal of Econometrics.
Bates, D., 1996, Jumps & Stochastic Volatility: Exchange Rate Processes Implicit in Deutschemark Options,
Review of Financial Studies, 9, 69107.
Bates, D., 1999, Post-87 Crash Fears in the S&P 500 Futures Options, Working Paper, University of Iowa;
forthcoming in Journal of Econometrics.
Black, F., and M. Scholes, 1973, The Pricing of Options and Corporate Liabilities, Journal of Political
Economy, 81, 637659.
Black, F., 1976, Studies of Stock Price Volatility Changes, Proceedings of the 1976 Meetings of the
American Statistical Association, Business and Economic Statistic Section, 177181.
Bollerslev, T., 1986, Generalized Autoregressive Conditional Heteroskedasticity, Journal of Econometrics,
31, 307327.
Brennan, M., 1979, The Pricing of Contingent Claims in Discrete Time Models, Journal of Finance, 34,
5368.
Brown, G. W., and K. Toft, 1999, Constructing Binomial Trees from Multiple Implied Probability Distributions, Working Paper, University of North Carolina.
Christie, A., 1982, The Stochastic Behavior of Common Stock Variances: Value, Leverage and Interest Rate
Effects, Journal of Financial Economics, 10, 407432.
Cox, J. and S. Ross, 1976, The Valuation of Options for Alternative Stochastic Processes, Journal of
Financial Economics, 3, 145166.
Cox, J., J. Ingersoll, and S. Ross, 1985, A Theory of the Term Structure of Interest Rates, Econometrica,
53, 385407.
Derman, E., and I. Kani, 1994, Riding on the Smile, Risk, 7, 3239.
Duan, J., 1995, The GARCH Option Pricing Model, Mathematical Finance, 5, 1332.
623
Duan, J., G. Gauthier, and J. Simonato, 1999, An Analytical Approximation for the GARCH Option Pricing
Model, Journal of Computational Finance, 2, 75116.
Dumas, B., J. Fleming, and R. Whaley, 1998, Implied Volatility Functions: Empirical Tests, Journal of
Finance, 53, 20592106.
Dunbar, N., 1999, LTCM and the Dangers of Marking to Market, Risk, 12, 67.
Dupire, B., 1994, Pricing with a Smile, Risk, 7, 1820.
Engle, R., and C. Mustafa, 1992, Implied ARCH Models from Options Prices, Journal of Econometrics,
52, 289311.
Engle, R., and V. Ng, 1993, Measuring and Testing the Impact of News on Volatility, Journal of Finance,
43, 17491778.
Feller, W., 1951, Two Singular Diffusion Problems, Annals of Mathematics, 54, 173182.
Feller, W., 1971, An Introduction to Probability Theory and Its Applications: Volume 2, Wiley and Sons, New
York.
Fleming, J., and R. Whaley, 1994, The Value of Wildcard Options, Journal of Finance, 49, 251236.
Foster, Douglas, and Daniel Nelson, 1994, Asymptotic Filtering Theory for Univariate ARCH Models,
Econometrica, 62, 141.
French, D., and E. Maberly, 1992, Early Exercise of American Index Options, Journal of Financial Research,
15, 127137.
Harrison, J. M., and D. Kreps, 1979, Martingales and Arbitrage in Multiperiod Securities Markets, Journal
Of Economic Theory, 20, 381408.
Harrison, J. M., and S. Pliska, 1981, Martingales and Stochastic Integrals in The Theory of Continuous
Trading, Stochastic Processes And Their Applications, 11, 215260.
Heston, S., 1990, Testing Continuous Time Models of the Term Structure of Interest Rates, working paper,
Yale University.
Heston, S., 1993, A Closed-Form Solution for Options with Stochastic Volatility, with Applications to Bond
and Currency Options, Review of Financial Studies, 6, 327343.
Hull, J., 1998, Options, Futures and Other Derivatives (3rd. ed.), Prentice-Hall, New Jersey.
Ingersoll, J., 1987, Theory of Financial Decision Making, Rowman and Littleeld, Totowa, N.J.
Kallsen, J., and M. Taqqu, 1998, Option Pricing in ARCH Type Models, Mathematical Finance, 8, 1326.
Kendall, M., and A. Stuart, 1977, The Advanced Theory of Statistics: Volume 1, Macmillan Publishing Co.,
New York.
Knoch, H., 1992, The Pricing of Foreign Currency Options With Stochastic Volatility, Ph.D. Dissertation,
Yale School of Organization and Management.
Merton, R., 1973. The Theory of Rational Option Pricing, Bell Journal of Economics and Management
Science, 4, 141183.
Nandi, Saikat, 1998, How Important is the Correlation Between Returns and Volatility in a Stochastic Volatility Model? Empirical Evidence from Pricing and Hedging in the S&P 500 Index Options Market, Journal
of Banking and Finance, 22, 589610.
Press, W., S. Teukolsky, W. Vetterling, and B. Flannery, 1992, Numerical Recipes in C (2d. ed.), Cambridge
University Press, New York.
Ritchken, P., and R. Trevor, 1999, Pricing Options Under Generalized GARCH and Stochastic Volatility
Processes, Journal of Finance, 54, 377402.
624
Rubinstein, M., 1976, The Valuation of Uncertain Income Streams and the Pricing of Options, Bell Journal
of Economics and Management Science, 7, 407425.
Rubinstein, M., 1994, Implied Binomial Trees, Journal of Finance, 69, 771818.
Vasicek, O., 1977, An Equilibrium Characterization of the Term Structure, Journal of Financial Economics,
5, 177188.
625

He Stone Nandi

Uploaded by

Document Informationclick to expand document informationResearch Paper on options Pricing

Document Informationclick to expand document information

Copyright:

Available Formats

He Stone Nandi

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

He Stone Nandi

Uploaded by

Copyright:

Available Formats

A Closed-Form GARCH Option

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

(log(S(t)) log(S(t )) r h(t) 1 h(t))2

1 determines the kurtosis of the distribution and 1 being zero implies a

Given positive 1 , positive value for 1 results in negative correlation between

where z(t) is a Wiener process. The details are in Appendix B. Consequently

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

Proposition 1. The risk-neutral process takes the same GARCH form as

The Review of Financial Studies / v 13 n 3 2000

We proceed to solve for the generating function of the GARCH process

Bi (t; T , )h(t + 2 i)

A Closed-Form GARCH Option Valuation Model

Proposition 3. If the characteristic function of the log spot price is f (i)

where Et [] denotes the expectation under the risk-neutral distribution. This

The Review of Financial Studies / v 13 n 3 2000

In our options sample, all but one day are Wednesdays.

A Closed-Form GARCH Option Valuation Model

The Review of Financial Studies / v 13 n 3 2000

252( + 1 )/(1 1 1 12 ) is the annualized (252

log(S(t)) = log(S(t )) + r + h(t) +

Maximum likelihood estimation

A Closed-Form GARCH Option Valuation Model

The Review of Financial Studies / v 13 n 3 2000

(across our sample of 755 days) annualized level of h(t + 1) is 9.34%

For each day in our sample, the annualized level of volatility is

252 h(t + 1).

A Closed-Form GARCH Option Valuation Model

data. Similarly, the average annualized level of h(t + 1) is 9.39% from

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

The Review of Financial Studies / v 13 n 3 2000

RMSE is the square root of the average squared valuation error.

Non-linear least squares estimation

A Closed-Form GARCH Option Valuation Model

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

Average option price

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

Panel B: Valuation errors by years

The Review of Financial Studies / v 13 n 3 2000

Out-of-sample valuation errors for put options

A Closed-Form GARCH Option Valuation Model

Out-of-sample valuation errors for call options

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

3.3 Using S&P 500 futures to lter volatility

The Review of Financial Studies / v 13 n 3 2000

A Closed-Form GARCH Option Valuation Model

with the time-series information in the history of asset prices to estimate

(log(S(t)) log(S(t )) r h(t) 1 h(t))2

Bi (t; T , )h(t + 2 i)

log(S(t)) = log(S(t )) + r + h(t) +

Bi (t; T , )h(t + 2 i) +

Bi+1 (t + ; T , )h(t + 2 i) + C1 (t + ; T , )

i+1 h(t + 2 i) +

Bi (t; T , ) = i B1 (t + ; T , ) + Bi+1 (t + ; T , ), for 1 < i p,

Let 1 () = 41 2 2 , 1 () = 0, () = ( 41 2 )2 , 1 () = 2/( ) / , and