The Model-Free Implied Volatility and Its Information Content

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/5217049
The Model-Free Implied Volatility and Its Information

Content
Article in Review of Financial Studies · May 2005

DOI: 10.1093/rfs/hhi027 · Source: RePEc
CITATIONS READS
538 3,362
2 authors, including:
Yisong S. Tian
York University
51 PUBLICATIONS 2,286 CITATIONS
SEE PROFILE
All content following this page was uploaded by Yisong S. Tian on 16 May 2014.
The user has requested enhancement of the downloaded file.

The Model-Free Implied Volatility and Its
Information Content
George J. Jiang
Eller College of Management, University of Arizona
Yisong S. Tian
Schulich School of Business, York University
Britten-Jones and Neuberger (2000) derived a model-free implied volatility under the
diffusion assumption. In this article, we extend their model-free implied volatility to
asset price processes with jumps and develop a simple method for implementing it
using observed option prices. In addition, we perform a direct test of the informa-
tional efficiency of the option market using the model-free implied volatility. Our
results from the Standard & Poor’s 500 index (SPX) options suggest that the model-
free implied volatility subsumes all information contained in the Black–Scholes (B–S)
implied volatility and past realized volatility and is a more efficient forecast for future
realized volatility.
There has been considerable research on the forecasting ability and

information content of the Black–Scholes (B–S; Black and Scholes,
1973) implied volatility. Since option prices reflect market participants’
expectations of future movements of the underlying asset, the volatility
implied from option prices is widely believed to be informationally super-
ior to the historical volatility of the underlying asset. If the option market
is informationally efficient and the B–S model is correct, implied volatility
is expected to subsume all information contained in historical volatility
and provides a more efficient forecast for future volatility.
Early studies find that implied volatility is a biased forecast of future
volatility and contains little incremental information beyond historical
volatility. For example, Canina and Figlewski (1993) found that the
implied volatility from the Standard & Poor’s (S & P) 100 index options
is a poor forecast for the subsequent realized volatility of the underlying
index. Based on an encompassing regression analysis, they found that
implied volatility has virtually no correlation with future realized volati-
lity and thus does not incorporate information contained in historical
We thank Yacine Aı̈t-Sahalia (the editor), an anonymous referee, Stewart Hodges, Chris Lamoureux,
Nathaniel O’Connor, Wulin Suo, Shu Yan, Hao Zhou, and seminar participants at the Federal Reserve
Board, the University of Toronto, and the University of Warwick for helpful comments and suggestions.
The financial support of the Social Sciences and Humanity Research Council of Canada is gratefully
acknowledged. Address correspondence to Yisong S. Tian, Finance Area, Schulich School of Business,
York University, 4700 Keele Street, Toronto, ON M3J 1P3, or e-mail: ytian@schulich.yorku.ca.
ª The Author 2005. Published by Oxford University Press. All rights reserved. For Permissions, please email:
journals.permissions@oupjournals.org
doi:10.1093/rfs/hhi027 Advance Access publication May 25, 2005
The Review of Financial Studies / v 18 n 4 2005
volatility. In contrast, Day and Lewis (1992), Lamoureux and Lastrapes

(1993), Jorion (1995) and Fleming (1998) report evidence supporting the
hypothesis that implied volatility has predictive power for future volati-
lity. They also found that implied volatility is a biased forecast for future
realized volatility.
More recent research attempts to correct various data and methodologi-
cal problems in earlier studies. These studies [e.g., Christensen and Prabhala
(1998), Christensen, Hansen, and Prabhala (2001), Blair, Poon, and Taylor
(2001), Ederinton and Guan (2002), and Pong et al. (2004)] consider longer
time series to take into account possible regime shift around the October
1987 crash, use instrumental variables (IVs) to correct for the errors-
in-variable (EIV) problem in implied volatility, adopt high-frequency
asset returns to provide a more accurate estimate for realized volatility, or
use nonoverlapping samples to avoid the ‘‘telescoping overlap’’ problem.
Collectively, these studies present evidence that implied volatility is a more
efficient forecast for future volatility than historical volatility.
However, nearly all previous research on the information content of
implied volatility focuses on the B–S implied volatility from at-the-money
options. Granted, at-the-money options are generally more actively
traded than other options and are certainly a good starting point. By
concentrating on at-the-money options alone, however, these studies fail
to incorporate the information contained in other options. More impor-
tantly, tests based on the B–S implied volatility are joint tests of market
efficiency and the B–S model. These studies are thus subject to model
misspecification errors.
In an important departure from previous research, we perform direct
tests of the informational efficiency of the option market using an alter-
native implied volatility measure that is independent of option pricing
models. This model-free implied volatility was derived by Britten-Jones
and Neuberger (2000), building on the pioneering work of Breeden and
Litzenberger (1978), Derman and Kani (1994, 1998), Rubinstein (1994,
1998), Derman, Kani, and Chriss (1996), and Ledoit and Santa-Clara
(1998) on implied distributions. Unlike the traditional concept of implied
volatility, their model-free implied volatility is not based on any specific
option pricing model. Instead, it is derived entirely from no-arbitrage
conditions. In particular, Britten-Jones and Neuberger (2000) showed
that the risk-neutral integrated return variance between the current date
and a future date is fully specified by the set of prices of options expiring
on the future date (their Proposition 2).
However, Britten-Jones and Neuberger (2000) derived the model-free
implied volatility under diffusion assumptions. It is unclear whether it is
still valid when the underlying asset price process includes jumps. This
can be a serious limitation since random jumps are an important aspect
of the price dynamics of many financial assets. In this article, we extend
1306
The Model-Free Implied Volatility
Britten-Jones and Neuberger (2000) and demonstrate that their model-

free implied volatility is still valid even if the underlying asset price
process has jumps. We first provide a simpler derivation under diffusion
assumptions and then generalize it to processes with jumps. Our result
ensures the generality of the model-free implied volatility.
In addition, we develop a simple method for implementing the model-
free implied volatility using observed option prices. Defined as an integral
of option prices, we show that it can be accurately calculated using
trapezoidal integration. However, a greater challenge in practice is that
options are traded only over a finite range of strike prices while an infinite
range is required. We provide theoretical upper bounds for truncation
errors when a finite range is used. We also demonstrate how truncation
errors vary with the range of strike prices and identify the strike price
range required to control such errors. More importantly, we show that
the model-free implied volatility can be calculated accurately from
observed option prices using a curve-fitting method and extrapolation
from endpoint implied volatilities.
Finally, we investigate the forecasting ability and information content
of the model-free implied volatility. The model-free implied volatility
facilitates a direct test of the informational efficiency of the option
market, rather than a joint test of market efficiency and the assumed
option pricing model. It also aggregates information across options with
different strike prices and should be informationally more efficient. We
conduct our empirical tests using the S & P’s 500 index (SPX) options
traded on the Chicago Board Options Exchange (CBOE). Following
prior research, we minimize measurement errors by using tick-by-tick
data, commonly used data filters, nonoverlapping samples, and realized
volatility estimated from high-frequency index returns.
Consistent with previous research, we find that the B–S implied vola-
tility contains more information than the historical volatility of the under-
lying asset but is an inefficient forecast of future realized volatility. In
contrast, we find that the model-free implied volatility subsumes all
information contained in both the B–S implied volatility and the past
realized volatility and is a more efficient forecast for future realized
volatility. Our results provide support for the informational efficiency
of the option market. Information in historical volatility is correctly
incorporated into option prices, although the use of a single option is
not sufficient to extract all relevant information. These findings are
robust to alternative estimation methods, volatility series over different
horizons, the choice of actual or implied index values, and alternative
methods for calculating realized volatility.
The rest of the article proceeds as follows. The next section describes
and generalizes the model-free implied volatility. A simple method
for implementing the model-free implied volatility is also developed.
1307
Numerical examples are provided to demonstrate the validity and accu-

racy of this method. In Section 2, we discuss the data used in this study
and the calculation of volatility measures. The forecasting ability and
information content of the model-free implied volatility are investigated
in Section 3 using monthly nonoverlapping samples. Section 4 conducts
robustness tests to ensure the generality of our results. Conclusions are in
the final section.
1. Model-Free Implied Volatility

Britten-Jones and Neuberger (2000) derived the model-free implied vola-
tility under diffusion assumptions. In this section, we first provide a
simpler derivation under diffusion assumptions and then generalize it to
processes with jumps. We also examine various implementation issues.
1.1 Diffusions and jumps

Suppose call options with a continuum of strike prices (K) for a given
maturity (T) are traded on an underlying asset. Following Dumas,
Fleming, and Whaley (1998) and Britten-Jones and Neuberger (2000),
we consider the forward asset price and forward option price, denoted
by Ft and CF (T,K), respectively, under the forward probability measure.
The forward price Ft is a martingale under the forward measure even
when both interest rates and dividends are stochastic. Britten-Jones and
Neuberger’s (2000) model-free implied volatility is defined in the follow-
ing proposition:
Proposition 1. The integrated return variance between the current date 0

and a future date T is fully specified by the set of prices of call options
expiring on date T:
"ð # ð1 F
F
T
dFt 2 C ðT,KÞ maxð0,F0 KÞ
E0 ¼2 dK ð1Þ
0 Ft 0 K2
where the superscript F denotes the forward probability measure.

Although Britten-Jones and Neuberger (2000) relied on the diffusion
assumption to derive Proposition 1, we show in the Appendix that it also
holds when asset prices contain jumps.1 Since a martingale can be decom-
posed canonically into the orthogonal sum of a purely continuous mar-
tingale and a purely discontinuous martingale [Jacod and Shiryaev (1987)
and Protter (1990)], we generalize Britten-Jones and Neuberger’s result to
all martingale processes. Equation (1) is thus valid for a very general class
of asset price processes and provides a model-free relationship between
1
We are indebted to an anonymous referee for providing the idea and outline for the proof to Proposition 1.
1308
asset return variance and option prices. Following Britten-Jones and

Neuberger (2000), the right-hand side (RHS) of Equation (1) will be
referred to as the model-free implied variance and its square root the
model-free implied volatility.
1.2 Implementation issues

As shown on the RHS of Equation (1), the model-free implied volatility is
defined as an integral of option prices over an infinite range of strike
prices. If option prices are available for all strike prices, the required
integral is straightforward to calculate using numerical integration meth-
ods. Of course, only a finite number of strike prices are actually traded in
the marketplace which may lead to inaccuracies in the calculation of the
model-free implied volatility. Limited availability of strike prices and
other implementation issues are discussed below.
1.2.1 Truncation errors. Suppose the interval [Kmin, Kmax] defines

the range of available strike prices. To avoid trivial cases, we further
require that 0 < Kmin < F0 < Kmax < + 1. In order to focus on trunca-
tion errors, we also assume that all strike prices in the interval are
available. Truncation errors are present when we ignore the tails of the
distribution and approximate the RHS of Equation (1) by the following
integral:
ð Kmax F
C ðT, KÞ maxð0, F0 KÞ
2 dK: ð2Þ
Kmin K2
In the Appendix, we show that the truncation errors are related to the
local variations in the tails of the asset return distribution. The results are
summarized in the following proposition:
Proposition 2. The right and left truncation errors beyond the strike price
range [Kmin, Kmax] have the following upper bounds, respectively:
ð þ1 F " #
C ðT, KÞ maxð0, F0 KÞ F FT Kmax 2
2 dK E0 FT > Kmax ,
Kmax K2 Kmax
ð3Þ
and
ð Kmin " #
C F ðT, KÞ maxð0, F0 KÞ FT Kmin 2
2 dK E0F FT < Kmin :
0 K2 FT
ð4Þ
1309
Both upper bounds are quite intuitive as they reflect the local variations
in the tails of the return distribution. Although these upper bounds are
shown subsequently to be quite tight, they are not model free. We also
derive model-free upper bounds for truncation errors in the Appendix.
However, the model-free upper bounds are not as tight as the ones in
equations (3) and (4).
To illustrate typical truncation errors, we consider the following sto-
chastic volatility and random jump (SVJ) model:
1=2
dFt =Ft ¼ Vt dWt þ Jt dNt J dt,
1=2 ð5Þ
dVt ¼ ðv v Vt Þdt þ v Vt dWtv ,
dWt dWtv ¼ dt,

h i
where Nt~i.i.d. Poisson (l) and lnð1 þ Jt Þi:i:d: N lnð1þmJ Þ 12 s2 J ; sJ .
If l = 0, the SVJ model reduces to the Heston (1993) model. We use the
following two sets of parameters to illustrate the results: (I) kv = 1,
sv = 0.25, r = 0, l = 4, mJ = 0.0375, yv = 0.18542Kv, V0 = yv/kv and (II)
kv = 1, sv = 0.25, r = 0, l = 0.5, mJ = 0.075, yv = 0.18542kv, V0 = yv/kv.
These parameters are consistent with empirical estimates from SPX
options by Bakshi, Cao, and Chen (1997) and are chosen to have suffi-
cient variations in implied volatility patterns (Figure 1).2 Both sets of
parameters imply an annualized volatility of 20%. We fix the initial
forward price (F0) at $100. The implied volatility function and truncation
errors are plotted in Figure 1 using these parameters and various option
maturities.
For each set of parameters, we first calculate call prices using the SVJ
model for various maturities and strike prices. The B–S implied volatility
is then calculated and plotted against option moneyness (panels A and C).
Option moneyness is defined as the ratio of strike price over asset price.
Both smile and smirk patterns are illustrated in these plots, reflecting the
behavior of the asset return distribution (such as fat tails and skewness).
Next, we calculate the true truncation errors [the left-hand side (LHS) of
equations (3) and (4)] for a given truncation interval. The left and right
truncation points, Kmin and Kmax, are both expressed as multiples of
standard deviations (SDs) from the initial forward price (F0), and the
truncation errors are plotted against the corresponding multiples. As
shown in panels B and D of Figure 1, the truncation error declines
monotonically and diminishes as the truncation point moves away from
the initial forward price. In particular, the left and right truncation errors
2
The skewness and kurtosis of the daily return distribution implied by the first set of parameters are -0.03
and 7.43, respectively. The corresponding numbers for the second set are -2.22 and 41.87.
1310
Figure 1
Truncation errors in calculating the model-free implied volatility
The left and right truncation errors of the model-free implied volatility are plotted against the respective
truncation point (Kmin or Kmax). The stochastic volatility with random jump (SVJ) model is assumed for the
1=2 1=2 v
underlying asset price: dFt =Ft ¼ Vt dWt þ Jt dNt mJ ldt, dVt ¼ ðyv kv Vt Þdt þ sv V
t dWt , and
dWt dWtv ¼ rdt where Nt ~ i.i.d. Poisson (l) and lnð1 þ Jt Þ iid N lnð1 þ mJ Þ 12 s2J ; sJ . Two sets of
2
parameters are used: (I) kv = 1, sv = 0.25, r = 0, l = 4, mJ = 0, sJ = 0.0375, yv = 0.1854 kv, V0 = y/kv
and (II) kv = 1, sv = 0.25, r = 0, l = 0.5, mJ = 0.075, sJ = 0.075, yv = 0.18542kv, V0 = y/kv. The two panels
on the left (panels A and C) illustrate the shape of the Black–Scholes (B–S) implied volatility function for
one- and six-month maturities while the two panels on the right (panels B and D) show the truncation errors
at various truncation levels. The truncation points are stated in multiples of standard deviations (SDs).
are of similar magnitude (panel B) and exhibit nearly identical convergence

properties when the return distribution is nearly symmetric (panel A). In
comparison, the left truncation error is larger than the corresponding
right truncation error (panel D) when the return distribution is skewed to
the left (panel C). In this case, the left truncation errors converges at a
1311
Figure 1
Continued
slower rate than the right truncation error does. With a fatter left tail, a
larger range of strike prices is needed on the left of F0 if we wish to have
identical truncation errors from both sides. Nevertheless, the differences
between left and right truncation errors are relatively small and conver-
gence is rapid in all cases. In general, the truncation error is negligible if
the truncation points are more than two SDs from F0.
Figure 1 also shows that the model-free implied volatilities calculated
from one-month and six-month maturities have similar truncation
errors. With a longer maturity, however, the same multiple of SDs
translates into a larger range of strike prices. This is because integrated
1312
variance is an increasing function of maturity. As a result, a larger range

of strike prices are needed to control the truncation errors for longer
maturities.
We also calculate the upper bounds for truncations errors [the RHS of
Equations (3) and (4)] for both the one-month and six-month maturities.
In general, the difference between the upper bound and the true trunca-
tion error is small. In particular, the upper bound is almost indistinguish-
able from the true truncation error when the truncation point is more
than two SDs from F0. For this reason, the upper bounds are not plotted
in Figure 1.
1.2.2 Discretization errors. In addition to truncation errors, the imple-

mentation of the model-free implied volatility also involves discretization
errors due to numerical integration. Consider the numerical integration of
Equation (2) using the trapezoidal rule:
ð Kmax F X m
2 2
dK ½gðT, Ki Þ þ gðT, Ki1 ÞDK,
Kmin K i¼1
ð6Þ
where DK = (KmaxKmin)/m, Ki = Kmin + iDK for 0 i m, and
gðT; Ki Þ ¼ ½C F ðT; Ki Þ maxð0; F0 Ki Þ=Ki2 : For a finite m (or DK >
0), the numerical integration scheme in Equation (6) leads to discretiza-
tion errors.
Figure 2 illustrates typical discretization errors at various levels of DK
We set the truncation points at 3.5 SDs from F0. As shown in Figure 1,
truncation errors are virtually zero beyond 3.5 SDs. We use the second
set of parameter values to plot Figure 2 since it exhibits a more severe
volatility smile or smirk pattern. Discretization errors are smaller if we
use the first set of parameter values. In order to provide a useful bench-
mark for discretization errors, we measure DK in SD units. As shown in
Figure 2, discretization errors diminish quickly as DK decreases. For
both one-month and six-month maturities, discretization errors are
negligible when DK 0.35 SDs (or m 20). With the initial asset
price at $100 and the annualized volatility of 0.2, 0.35 SD translates
into a strike price increment of roughly $2 and $5 for one-month and
six-month options, respectively. This is generally consistent with actual
strike price increments in the marketplace. Since we use a much smaller
DK (or larger m) in our empirical implementation, discretization errors
are unlikely to have any impact on the calculation of the model-free
implied volatility.
1.2.3 Spot prices versus forward prices. Equation (1) is stated in forward
prices. When applications require the use of spot prices, appropriate
1313
Figure 2
Discretization errors in calculating the model-free implied volatility
The discretization errors of the model-free implied volatility are plotted against strike price increment
(DK, measured in standard deviation [SD] units) for options with one- and six-month maturities. The
1=2
SVJ model is assumed for the underlying asset price: dFt =Ft ¼ Vt dWt þ Jt dNt mJ ldt,
1=2 v v
dVt ¼ ðyv kv Vt Þdt
þ s v V t dW t , and
dW t dW t ¼ rdt where N t i.i.d. Poisson (l) and
lnð1 þ Jt Þ iid N lnð1 þ mJ Þ 12 s2J ;sJ . The second set of parameters with a more pronounced volatility
skew is used: kv = 1, sv = 0.25, r = 0, l = 0.5, mJ = 0.075, mJ = 0.075, yv = 0.18542kv, V0 = y/kv. The
model-free implied volatility is calculated using the trapezoidal integration method. To minimize trunca-
tion errors, we choose truncation points sufficiently far from the forward price (3.5 SDs on both sides of
the forward price F0). The discretization error is the difference between the calculated volatility and the
true volatility.
adjustments are needed. Under the assumption of deterministic interest

rates and dividends, we write the forward asset price and forward option
price at time t as Ft = St/B(t, T) and CF(T, K) = C(T, K)/B(t, T), respec-
tively, where St is the spot asset price minus the present value of all
expected future dividends to be paid prior to the option maturity, B(t, T)
is the time t price of a zero-coupon bond that pays $1 at time T, and
C(T,K) is the spot option price. By substitution, Equation (1) changes to:
"ð # ð1
F
T
dSt 2 CðT,KÞ=Bð0,TÞ max½0,S0 =Bð0,TÞ K
E0 ¼2 dK:
0 St 0 K2
Through a change of variables, the above equation is restated as:

"ð # ð1
F
T
dSt 2 C½T,K=Bð0,TÞ maxð0,S0 KÞ
E0 ¼2 dK: ð7Þ
0 St 0 K2
1314
Equation (2) provides an alternative definition of the model-free

implied variance in spot prices.
1.2.4 Limited availability of strike prices. Up to this point, we have

assumed that prices are available for options with any strike price in a
given range [Kmin, Kmax]. This is, however, not realistic because only strike
prices at a fixed increment are listed for trading in the marketplace. We
must deal with not only a limited range but also a sparse set of discrete
strike prices. Following prior research, we apply a curve-fitting method to
interpolate between available strike prices. To illustrate this method, we
consider the strike price structure of SPX options on September 23, 1988
and use it as a prototype.3 On this date, the closing index value is
approximately 270 (S0) and the listed strike prices (K) are 200, 220 to
310 with a 5-point increment, 325 and 350.
We continue to use the SVJ model with the second set of parameter
values as described previously. Without loss of generality, interest rates
and dividends are assumed to be zero. We calculate the model prices of
call options with the listed strike prices and a specified maturity (T). We
use model prices (instead of market prices) as inputs in the implementa-
tion of the model-free implied volatility so that the true volatility is
known and the approximation error due to our numerical implementa-
tion can be accurately measured.
To calculate the model-free implied volatility, we need to evaluate the
RHS of Equation (6). This requires prices of call options with strike prices
Ki for 0 i m. Because these options are not listed on our sample date,
their prices are not directly available (even though we could calculate
their prices using the SVJ model). Instead, their prices must be inferred
from the prices of listed options. Among the approaches used in previous
research, the curve-fitting method is the most practical and effective.
Although some studies apply the curve-fitting method directly to option
prices [e.g., Bates (1991)], the severely nonlinear relationship between
option price and strike price often leads to numerical difficulties. Follow-
ing Shimko (1993) and Aı̈t-Sahalia and Lo (1998), we apply the curve-
fitting method to implied volatilities instead of option prices. Prices of
listed calls are first translated into implied volatilities using the B–S
model. A smooth function is then fitted to the implied volatilities. We
then extract implied volatilities at strike prices Ki from the fitted function.
The B–S model is used once more to translate the extracted implied volati-
lities into call prices. Note that this curve-fitting procedure does not make the
assumption that the B–S model is the true model underlying option prices. It
3
The choice of the sampling date is not important because we do not use observed option prices in our
implementation here. We only need the set of strike prices listed on that date to provide a sense of
availability of strike prices in the marketplace.
1315
is merely used as a tool to provide a one-to-one mapping between option

prices and implied volatilities. With the extracted call prices, the model-free
implied volatility is calculated using the RHS of Equation (6). Following
Bates (1991) and Campa, Chang, and Reider (1998), we use cubic splines in
the curve-fitting of implied volatilities. Using cubic splines has the advantage
that the obtained volatility function is smooth everywhere and provides an
exact fit to the known implied volatilities.4
It is also important to note that the curve-fitting method is only
effective for extracting option prices between the maximum and minimum
available strike prices. For options with strike prices beyond the available
range, we use the endpoint implied volatility to extrapolate their option
values. In other words, the volatility function is assumed to be constant
beyond the maximum and minimum strike prices. Extrapolation may be
necessary in empirical applications as the range of available strike prices
may not be sufficiently large on all trading days. Of course, this extra-
polation procedure introduces an approximation error that is different
from the truncation error due to Equation (2). Whereas the truncation
method ignores options with strike prices beyond the available range, the
extrapolation method incorporates these options by approximating their
prices using the implied volatility at the endpoint strike price (either Kmin
or Kmax).
Table 1 presents the approximation errors from both the truncation
method (panel A) and the extrapolation method (panel B) for various
maturities (T) and truncation intervals ([Kmin, Kmax]). The available strike
price range of [200, 350] on September 23, 1988 is also included in the
table, which is equivalent to a truncation interval of [0.74S0, 1.30S0]. The
discretization parameter (m) is fixed at 100. The approximation error is
calculated as the estimated annualized volatility minus the true volatility.
As expected, the magnitude of the approximation error from both meth-
ods is positively related to maturity and negatively related to truncation
interval. In addition, the extrapolation method is in general more accu-
rate than the truncation method. Consider for example the longest matur-
ity of 180 days and the smallest truncation interval of [0.9S0,1.1S0]. The
approximation error from the truncation method is –0.0325 or 16.3% of
the true volatility of 0.2. The large error is not surprising because the
truncation points are roughly 0.7 SD from the initial asset price. From the
results in Figure 1, we know that the truncation error may not diminish
until the truncation interval is at least plus/minus 2 SDs. In comparison,
the approximation error from the extrapolation method is much smaller.
As shown in panel B, the corresponding error is only –0.0033 or 1.7% of
the true volatility. Additional calculations show that similar results are
4
For a more detailed review of curve-fitting and other methods, see the survey article by Jackwerth (1999).
1316
Table 1
Approximation errors in the calculation of the mode-free implied volatility
Maturity (days)
Kmin Kmax 30 45 60 75 90 120 180
Panel A: the truncation

method
0.9S0 1.1S0 0.0048 0.0079 0.0113 0.0146 0.0177 0.0234 0.0325
0.8S0 0.2S0 0.0001 0.0001 0.0005 0.0011 0.0018 0.0034 0.0072
200 350 0.0000 0.0002 0.0005 0.0010 0.0014 0.0027 0.0056
0.7S0 1.3S0 0.0001 0.0001 0.0000 0.0001 0.0002 0.0006 0.0018
0.6S0 1.4S0 0.0001 0.0000 0.0000 0.0001 0.0002 0.0004 0.0010
Panel B: the extrapolation
method
0.9S0 1.1S0 0.0013 0.0015 0.0017 0.0019 0.0021 0.0026 0.0033
0.8S0 1.2S0 0.0005 0.0003 0.0001 0.0000 0.0001 0.0004 0.0011
200 350 0.0005 0.0003 0.0002 0.0001 0.0000 0.0003 0.0008
0.7S0 1.3S0 0.0005 0.0003 0.0002 0.0001 0.0000 0.0003 0.0008
0.6S0 1.4S0 0.0005 0.0003 0.0002 0.0001 0.0000 0.0003 0.0008
The approximation error is defined as the estimated annualized volatility minus the true volatility (0.2).
The SVJ model is implemented with the following parameters: kv = 1, sv = 0.25, = 0, l = 0.5, J =
0.075, sJ = 0.075, v = 0.18542 kv, V0 = yv/kv. The current asset price (S0) is 270 and the listed strike
prices range from 200 to 350, based on options listed on the Standard & Poor’s (S & P ’s) 500 index on
September 23, 1988. The discretization parameter (m) used is 100.
obtained over a wide range of parameters for the SVJ model. The accu-
racy and robustness of our extrapolation method is thus assured.
As shown in Table 1, the extrapolation method (panel B) leads to
smaller approximation errors than the truncation method (panel A)
does. The truncation method ignores strike prices beyond the truncation
interval, which leads to the underestimation of the true volatility. The
extrapolation method corrects the underestimation error by assuming a
flat implied volatility function beyond the truncation point. In other
words, implied volatilities outside the truncation interval are extrapolated
from the implied volatility at the respective truncation point. We further
argue that this correction mechanism in the extrapolation method is likely
to lead to smaller approximation errors in most empirical settings. Con-
sider, for example, the three most commonly observed shapes of the
implied volatility function: smile, smirk, and skew. With a volatility
smile, the implied volatility rises as the strike price increases or decreases
from the asset price. By extrapolating on the implied volatility at the
truncation points, the extrapolation method underestimates the true
volatility beyond the truncation interval. The underestimation is, how-
ever, less severe than that in the truncation method, leading to a more
accurate approximation. With a volatility smirk, the implied volatility
rises faster at low strike prices than at high strike prices. The extrapola-
tion method is again more accurate than the truncation method. In this
case, however, extrapolation creates a more severe kink at the left truncation
1317
point, which may lead to larger approximation errors if Kmin is too close
to S0. Numerical examples in Table 1 show that the extrapolation method
is, nevertheless, reasonably accurate even in the presence of a severe
volatility smirk. Finally, we consider a volatility skew. A typical volatility
skew exhibits high implied volatility for low strike prices and low implied
volatility for high strike prices. This means that the extrapolation method
underestimates the volatility beyond the minimum strike price but over-
estimates it beyond the maximum strike price. As the two effects offset or
partially offset each other, the extrapolation method should be more
accurate than the truncation method. Consequently, we adopt the extra-
polation method in our subsequent empirical tests.
2. Data and Volatility Calculation

Data used in this study are from several sources. Intraday data on SPX
options are obtained from the CBOE. Daily cash dividends are obtained
from the S & P’s DRI database. High-frequency data at 5-minute and
30-minute intervals for the SPX are extracted from the contemporaneous
index levels recorded with the quotes of SPX options. In addition, daily
Treasury bill yields (our risk-free rates) are obtained from the Federal
Reserve Bulletin. Our sample period is from June 1988 to December 1994.
Following Harvey and Whaley (1991, 1992a, 1992b), we use daily cash
dividends instead of constant dividend yield. We use mid bid-ask quotes
instead of actual transaction prices in order to avoid the bid-ask bounce
problem [Bakshi, Cao, and Chen (1997, 2000)].
To select our final sample of SPX options, several data filters are
applied. First, option quotes less than 3/8 are excluded from the sample.
These prices may not reflect true option value due to proximity to tick
size. Second, options with less than a week remaining to maturity are
excluded from the sample. These options may have liquidity and market
microstructure concerns. Third, option quotes that are time-stamped
later than 3:00 P.M. (Central Standard Time) are excluded. As the stock
market closes at 3:00 P.M., these option prices cannot be matched with
simultaneously observed index values. Next, following Aı̈t-Sahalia and
Lo (1998), we exclude in-the-money options from the sample. In-the-
money options are more expensive and often less liquid than at-the-
money or out-of-the-money options. We define in-the-money options as
call options with strike prices less than 97% of the asset price and put
options with strike prices more than 103% of the asset price. In addition,
options violating the boundary conditions are eliminated from the sam-
ple. These options are significantly undervalued and the B–S implied
volatility is in fact negative for these options. Finally, only option
quotes from 2:00 to 3:00 P.M. are included in the sample. To reduce
computational burden, we construct one implied volatility surface for
1318
each day in our sample. As option quotes over a range of strike prices
and maturities are needed, we use all option quotes in the final trading
hour before the stock market closes. The last trading hour is used
because trading volume is typically higher than at other times of the day.
Our empirical tests require the construction of implied volatility surface
from available option prices. In theory, this is straightforward to do if the
set of traded options covers a sufficient range of strike prices and matu-
rities. For each available option, we use the B–S model to back out the
implied volatility from option quote. If option quotes are available for all
strike prices and maturities, the calculated implied volatilities naturally
form a surface. However, only a limited number of strike prices and
maturities are listed for trading. A curve-fitting method is needed to
extract the implied volatility surface from available option prices. To fit
a smooth surface to the available B–S implied volatilities, the curve-fitting
method is implemented in two steps. Cubic splines are first applied in the
moneyness dimension to fit a smooth curve to the B–S implied volatilities
as in Section 1.2.4. That is, implied volatility as a smooth function of
moneyness (x), s ^ðxjTÞ, is obtained for each option maturity (T). In the
second step, we fix moneyness and apply cubic splines in the maturity
dimension. The result is a smooth implied volatility surface, ðx;TÞ.
To avoid the telescoping overlap problem described by Christensen,
Hansen, and Prabhala (2001), we extract implied volatilities from the
implied volatility surface at predetermined fixed maturity intervals. The
model-free implied volatility is then calculated using the RHS of Equa-
tion (6) for fixed maturities of 30, 60, 120, and 180 (calendar) days. The
B–S implied volatility is extracted directly from the implied volatility
surface for the same fixed maturities and moneyness levels from 0.94 to
1.06.
In addition to implied volatilities, our empirical analysis also requires
monthly series of realized volatility and historical volatility. Realized
volatility is needed to assess the forecasting ability and information con-
tent of implied volatilities. We calculate the realized volatility over peri-
ods of 30 to 180 days, matching the maturities of the corresponding
implied volatilities. Historical volatility serves as a competing forecast,
to implied volatilities, for future realized volatility. We use the realized
volatility on the latest trading day as a proxy for historical volatility since
the latest day’s volatility is likely to contain the most relevant information
for future volatility. In other words, the lagged daily realized volatility is
our proxy for historical volatility.
Previous research adopts different sampling frequencies of asset returns
for the calculation of realized volatility. Some earlier studies such as
Canina and Figlewski (1993) and Christensen and Prabhala (1998) calcu-
lated these volatilities using daily returns. More recent studies argue that
realized volatility should be calculated using intraday returns instead of
1319
daily returns. For example, Andersen and Bollerslev (1998), Andersen,

Bollerslev, Diebold, and Labys (2001, 2003) and Barndorff-Nielsen and
Shephard (2003) find that there is considerable advantage in using high-
frequency data over daily data in estimating realized volatility. In parti-
cular, Andersen and Bollerslev (1998) show that the typical squared
returns method for calculating realized volatility produces inaccurate
forecasts if daily returns are used. The inaccuracy is a result of noise in
daily returns. They further show that the impact of the noise component
is diminished if high-frequency returns (e.g., 5-minute returns) are used.
These studies, however, draw their conclusions using evidence from for-
eign exchange rates. Intraday data on foreign exchange rates are relatively
well behaved and exhibit little or no autocorrelation. This is not true for
stock or stock index returns. Stoll and Whaley (1990) show that 5-minute
returns on the SPX during the sample period from July 23, 1984 to
December 31, 1986 have a first-order autocorrelation of 0.45 and a
second-order autocorrelation of 0.14. The positive correlation is mainly
due to the infrequent trading of the stocks in the index. As market
liquidity tends to improve over time, we should expect this positive
correlation to become smaller in more recent period. Indeed, in our
sample period from June 1, 1988 to December 31, 1994, the first-order
autocorrelation drops to 0.31 while the second-order autocorrelation
drops to 0.04. Nevertheless, the correlation is still significant and a
naive estimate of realized volatility based on 5-minute returns is likely
to be downward biased. In the extreme case when returns are sampled
continuously, Zhang, Mykland, and, Aı̈t-Sahalia (2003) show that the
standard realized volatility estimator captures the volatility of the market
microstructure noise rather than the true integrated volatility if the noise
is not accounted for.
The optimal sampling frequency of intraday returns and the appropri-
ate procedure for dealing with the related market microstructure problem
are the subject of several recent studies. Aı̈t-Sahalia, Mykland, and Zhang
(2005) demonstrate that more data does not necessarily lead to a better
estimate of realized volatility in the presence of market microstructure
noise. They show that the optimal sampling frequency is jointly deter-
mined by the magnitude of the market microstructure noise and the
horizon of the realized volatility. For a given level of noise, the realized
volatility for a longer horizon (e.g., one month) should be estimated with
less frequent sampling than the realized volatility for a shorter horizon
(e.g., one day). In addition, they also show that correcting for market
microstructure noise may restore the first-order statistical effect that it is
optimal to sample as often as possible.
Consistent with the findings from Aı̈t-Sahalia, Mykland, and Zhang
(2003), we use 30-minute index returns to calculate the realize volatility
over one month or longer horizons and 5-minute index returns to calculate
1320
the lagged daily realized volatility (our proxy for historical volatility). To
correct for the bias in estimated realized volatility due to autocorrelation
in intraday returns, we adopt a correction method suggested, in various
forms, by French, Schwert, and Stambaugh (1987), Zhou (1996), and
Hansen and Lunde (2004). In this correction method, the annualized
realized variance over the period [t, t + t] is calculated as:
l
1X n
2X n Xnh
Vt,t ¼ R2i þ Ri Riþh , ð8Þ
t i¼1 t h¼1 n h i¼1
where Ri is the index return during the i-th interval, n is the total number
of intervals in the period, and l is the number of correction terms
included.5 When the volatility (e.g, the lagged daily realized volatility) is
calculated with 5-minute returns, we use Equation (8) with one correction
term (i.e., l = 1). This is because 5-minute returns in our sample have a
first-order autocorrelation of 0.31 while higher-order autocorrelations are
much smaller. In contrast, we do not correct for autocorrelation if vola-
tility (e.g., the one-month realized volatility) is calculated with 30-minute
returns (i.e., l = 0). These less frequent returns have a first-order auto-
correlation of only 0.06 in our sample while higher-order autocorrelations
are all negligible. Although all reported results in this study are based on
realized volatility calculated this way, we do consider alternative correc-
tion terms subsequently in robustness tests and the results are not materi-
ally affected.
Table 2 provides summary statistics for the model-free implied volati-
lity, the B–S implied volatility and the realized volatility. All volatility
measures are estimated monthly on the Wednesday immediately follow-
ing the expiry date of the month. For the B–S model, implied volatility is
reported for at-the-money options only. For other moneyness levels, the
results are similar but exhibit the well-known smile pattern. The realized
volatility is calculated using 30-minute index returns for the period
matching the maturity of the corresponding option(s) in the implied
volatility calculation. As shown in Table 2, both the B–S and model-
free implied volatilities are on average higher than the realized volatility
over all horizons. These implied volatilities are thus likely biased forecast
for realized volatility, with a slightly larger bias for the latter. The larger
bias from the model-free implied volatility is expected as the B–S implied
volatility is known to be a downward biased measure of risk-neutral
expected variance under stochastic volatility by Jensen’s inequality. By
correcting this downward bias, the model-free implied volatility tends to
5
Dividends are ignored in the above calculation since daily dividends on the SPX are typically small and
much less volatile than the index itself. Unreported results confirm that adjustment for dividends does not
lead to material change to realized volatility estimates.
1321
Table 2
Summary Statistics of monthly volatility series
Standard Excess
t N Mean deviation Skewness kurtosis Minimum Maximum
Panel A: sBS
30 78 14.78 4.396 1.1258 2.0636 6.912 31.21
60 79 15.71 4.399 1.0210 0.9198 8.325 29.90
120 78 16.82 4.254 0.8038 0.6587 8.761 30.81
180 78 17.31 3.844 0.7213 0.3137 10.62 29.21
Panel B: sRE
30 79 13.96 4.078 1.4165 2.9300 6.953 28.78
60 79 14.08 3.751 1.2724 2.1930 7.350 26.92
120 79 14.10 3.682 1.3448 2.2043 8.248 26.54
180 79 14.13 3.241 1.1327 0.9593 9.208 25.32
Panel C: sMF
30 78 15.68 4.450 1.5966 3.8305 8.589 34.61
60 79 16.11 3.964 1.1685 1.3999 10.16 28.72
120 78 17.51 3.976 0.7028 0.1716 10.06 28.24
180 78 17.33 3.566 0.4522 0.1985 9.666 25.80
Panel D: ln(sBS)
30 78 2.653 0.284 0.1563 0.2628 1.933 3.441
60 79 2.719 0.264 0.3500 0.1444 2.119 3.397
120 78 2.792 0.246 0.1050 0.0401 2.170 3.427
180 78 2.828 0.216 0.1901 0.3050 2.363 3.374
Panel E: ln(sRE)
30 79 2.596 0.272 0.3959 0.6496 1.938 3.360
60 79 2.612 0.255 0.2439 0.0728 1.995 3.292
120 79 2.621 0.244 0.5847 0.3935 2.110 3.278
180 79 2.608 0.217 0.6220 0.0051 2.220 3.154
Panel F: ln(sMF)
30 78 2.718 0.255 0.6607 0.6865 2.150 3.544
60 79 2.752 0.229 0.5321 0.1377 2.318 3.357
120 78 2.838 0.221 0.1528 0.2150 2.308 3.341
180 78 2.832 0.205 0.0519 0.2103 2.368 3.250
sBS, sRE, and sMF are the at-the-money Black–Scholes (B–S) implied volatility, the realized volatility, and
the model-free implied volatility, respectively. t is the time horizon and N is the sample size. All
volatilities are expressed in annualized percentage terms.
be higher than the B–S implied volatility. In addition, the reported skew-
ness and excess kurtosis in Table 2 reveal that the log volatility is the most
conformable with the normal distribution. Regressions based on the log
volatility are thus statistically better specified than those based on vola-
tility or variance.
Table 3 summarizes the correlation matrix of the monthly 30-day
volatility series for the B–S implied volatility (from options with different
moneyness levels), the model-free implied volatility, and the realized
volatility. Correlations for monthly volatility series with longer maturities
have similar properties and are not reported. Overall, all B–S implied
volatility series are highly correlated with the model-free implied volati-
lity. As expected, the B–S implied volatility from at-the-money options
has the highest correlation with the model-free implied volatility (94.1%).
While all implied volatility series are highly correlated with the corre-
sponding realized volatility, the model-free implied volatility has the
1322
Table 3
Correlation matrix of monthly 30-day volatility series
sBS(0.94) sBS(0.97) sBS(1.00) sBS(1.03) sMF sRE
Panel A: correlation matrix

of volatility (N = 60)
sBS(0.94) 1.000 — — — — —
sBS(0.97) 0.879 1.000 — — — —
sBS(1.00) 0.861 0.887 1.000 — — —
sBS(1.03) 0.864 0.872 0.898 1.000 — —
sMF 0.922 0.927 0.941 0.936 1.000 —
sRE 0.767 0.778 0.814 0.803 0.857 1.000
Panel B: correlation matrix ln(sBS) (0.94) ln(sBS) (0.97) ln(sBS)(1.00) ln(sBS)(1.03) ln(sMF) ln(sRE)
of log volatility (N = 60)
ln(sBS) (0.94) 1.000 — — — — —
ln(sBS) (0.97) 0.847 1.000 — — — —
BS
ln(s )(1.00) 0.813 0.852 1.000 — — —
ln(sBS)(1.03) 0.828 0.829 0.872 1.000 — —
MF
ln(s ) 0.889 0.903 0.918 0.916 1.000 —
ln(sRE) 0.721 0.755 0.764 0.771 0.843 1.000
sBS, sRE, and sMF are the Black–Scholes (B–S) implied volatility, the realized volatility and the model-
free implied volatility, respectively. Numbers in parentheses after sBS are option moneyness. The sBS
(1.06) is excluded in the calculation of correlation matrix as it has fewer observations.
highest correlation (85.7%), followed by the at-the-money B–S implied

volatility (81.4%). We also test and reject the null hypothesis that the B–S
implied volatility is an unbiased estimator of the model-free implied
volatility at the 1% significance level. Figure 3 plots the two implied
volatility time series over our sample period. For the B–S implied volati-
lity, we only illustrate the at-the-money series. It is clear that the two
volatility series track one another fairly closely. The B–S implied volatility
tends to be lower than the model-free implied volatility and exhibits
greater local fluctuations than the latter does. Overall, these results
imply that while both the B–S and model-free implied volatilities have
similar features, the two implied volatility series do contain sufficiently
different information content.
3. The Information Content of Implied Volatility

Prior research has extensively examined the information content of the B–S
implied volatility. While early studies produce mixed results, recent
empirical studies seem to agree that the B–S implied volatility is a more
efficient forecast for future realized volatility than historical volatility.
They also find that the B–S implied volatility does not subsume all
information contained in historical volatility and is an (upward) biased
forecast for future realized volatility. However, these studies typically
adopt a point estimate of the B–S implied volatility obtained from the
price of a single option or prices of several similar options. Since informa-
tion contained in other options is discarded, tests based on such volatility
1323
Figure 3
Time-series plot of the model-free and at-the-money Black–Scholes (B–S) implied volatilities
The figure plots the time series of the 30-day model-free and at-the-money B–S implied volatilities over the
sample period from June 1988 to December 1994. The two implied volatilities are calculated using the
implied volatility surface. The implied volatility surface is constructed by interpolating implied volatilities of
traded options using cubic splines, first in the moneyness dimension and then in the maturity dimension.
measures are likely biased towards rejecting informational efficiency. In

this study, we examine the information content of the model-free
implied volatility. Because it aggregates information from options
across all strike prices, the model-free implied volatility should be infor-
mationally more efficient. To nest previous research within our frame-
work, we compare and contrast three competing volatility forecasts—
the model-free implied volatility, the B–S implied volatility, and the
historical volatility.
Following Christensen and Prabhala (1998) and Christensen, Hansen,
and Prabhala (2001), we test our null hypotheses using monthly nono-
verlapping samples. This avoids the so-called telescoping overlap problem
which may render the t-statistics and other diagnostic statistics in the
regression analysis invalid. To extract our monthly nonoverlapping sam-
ple, we choose the Wednesday immediately following the expiration date
of the month. If it happens to be a nontrading day, we go to the following
Thursday then the proceeding Tuesday. We select the monthly sample
this way because option trading seems to be more active during the week
following the expiration date and Wednesday has the fewest holidays
among all weekdays. When calculating the monthly volatility measures,
1324
we fix the option maturity at 30 calendar days. Time series obtained this
way are neither overlapping nor telescoping.
Consistent with prior research [e.g., Canina and Figlewski (1993) and
Christensen and Prabhala (1998)], we employ both univariate and encom-
passing regressions to analyze the information content of volatility fore-
casts. In a univariate regression, the realized volatility is regressed against
a single volatility forecast. In comparison, two or more volatility forecasts
are used as explanatory variables in encompassing regressions. While the
univariate regression focuses on the forecasting ability and information
content of one volatility forecast, the encompassing regression addresses
the relative importance of competing volatility forecasts and whether one
volatility forecast subsumes all information contained in other volatility
forecast(s). Since a univariate regression is a restricted version of the
corresponding encompassing regression, we only state the encompassing
regressions. These regressions are specified as follows, in three alternative
specifications:
MF MF BS BS LRE LRE
sRE
t,t ¼ ax,t þ bx,t st,t þ bx,t st,x,t þ bx,t st,t þ et,x,t , ð9Þ
RE
Vt,t ¼ ax,t þ bMF MF BS BS LRE LRE
x,t Vt,t þ bx,t Vt,x,t þ bx,t Vt,t þ et,x,t , ð10Þ
MF BS LRE
ln sRE MF BS LRE
t,t ¼ ax,t þ bx,t ln st,t þ bx,t ln st,x,t þ bx,t ln st,t þ et,x,t , ð11Þ
where s and V are asset return volatility and variance, respectively. The
superscripts RE, MF, BS, and LRE stand for REalized, Model-Free,
Black–Scholes, and Lagged REalized, respectively. The subscripts t, x,
and t are observation date, moneyness, and maturity, respectively. Uni-
variate regressions are obtained if two of the three regressors are dropped.
Lagged realized volatility sLRE is our proxy for historical volatility.
Previous studies [e.g., Canina and Figlewski (1993) and Christensen and
Prabhala (1998)] adopted the realized volatility over a matching period
(30 calendar days here) immediately proceeding the current observation
date as the lagged realized volatility. As the volatility process is likely a
Markov process, the latest day’s realized volatility is the most relevant for
forecasting future volatility. Following this logic, we adopt the realized
volatility on the latest trading day as the lagged realized volatility. This
lagged realized volatility is calculated from 5-minute index returns using
Equation (8) with one correction term. The robustness of our results with
respect to alternative correction terms and other proxies for historical
volatility is examined subsequently.
Table 4 summarizes the OLS regression results from both univariate
and encompassing regressions using the monthly nonoverlapping sample
described previously. Results from the three specifications are presented
in separate panels in the table. While the estimate for the other two
volatility measures is only a function of maturity, the estimate for the
1325
1326

Table 4
Univariate and encompassing regressions of 30-day volatility (OLS)
N a MF BS LRE Adjusted R2 Durbin-Watson 2 test(a) 2 test(b)
Panel A: st
78 2.49 (1.15) — 0.77 (0.080) — 0.64 1.98 14.85 (0.000) —
78 2.66 (1.16) — 0.70 (0.093) 0.06 (0.027) 0.65 2.04 — 11.91 (0.003)

78 10.8 (0.80) — — 0.23 (0.036) 0.15 1.65 204.6 (0.000) —
78 1.30 (0.79) 0.80 (0.053) — — 0.74 1.98 70.37 (0.000) —
78 1.36 (0.78) 0.83 (0.116) 0.04 (0.123) — 0.74 2.01 — 15.17 (0.001)
78 1.32 (0.80) 0.80 (0.060) — 0.01 (0.033) 0.74 2.00 — 15.12 (0.001)
78 1.39 (0.84) 0.84 (0.122) 0.05 (0.126) 0.01 (0.032) 0.74 2.02 — 16.21 (0.001)
Panel B: Vt
78 32.1 (18.9) — 0.75 (0.063) — 0.62 2.07 15.10 (0.000) —
78 32.8 (19.1) — 0.73 (0.104) 0.02 (0.029) 0.62 2.08 — 10.68 (0.005)

78 176.2 (16.6) — — 0.15 (0.037) 0.11 1.75 652.2 (0.000) —
78 19.9 (13.4) 0.71 (0.058) — — 0.73 1.95 67.86 (0.000) —
78 18.8 (15.9) 0.69 (0.119) 0.03 (0.152) — 0.73 1.90 — 48.11 (0.000)
78 20.4 (14.4) 0.73 (0.072) — 0.02 (0.020) 0.73 1.91 — 49.32 (0.000)
78 19.3 (14.7) 0.71 (0.121) 0.03 (0.152) 0.02 (0.025) 0.73 1.91 — 49.14 (0.000)
Panel C: In Vt
78 0.55 (0.22) — 0.77 (0.080) — 0.60 1.90 23.20 (0.000) —
78 0.40 (0.21) — 0.70 (0.082) 0.14 (0.051) 0.62 1.91 — 13.37 (0.001)

78 1.88 (0.18) — — 0.29 (0.071) 0.20 1.51 111.4 (0.000) —
78 0.10 (0.16) 0.92 (0.057) — — 0.75 1.94 71.21 (0.000) —
78 0.11 (0.16) 0.99 (0.135) 0.09 (0.120) — 0.74 1.96 — 2.262 (0.323)
78 0.09 (0.16) 0.89 (0.065) — 0.04 (0.042) 0.75 1.97 — 2.541 (0.281)
78 0.09 (0.16) 0.95 (0.147) 0.07 (0.123) 0.03 (0.043) 0.74 1.98 — 3.625 (0.305)
The at-the-money Black–Scholes (B–S) implied volatility and one-day lagged historical volatility are used in the regressions. The numbers in parentheses beside the parameter
estimates are the standard errors, which are computed following a robust procedure taking into account of the heteroscedast1icity [see White (1980)]. The w2 test(a) is for the joint
hypothesis H0: a = 0 and j = 1 (j = MF, BS, LRE) in univariate regressions, and the w2 test(b) is for the joint hypothesis H0: BS = 1 and LRE = 0 or H0: MF = 1, and
BS = LRE = 0 in encompassing regressions. The test statistics are reported with the p-values in the brackets beside the statistic.
, , and indicate that the leading term coefficient of the regressionis significantly different from one at the 10%, 5%, and 1% level, and , , and indicate
that the coefficient of the remaining terms is significantly different from zero at the 10%, 5%, and 1% level, respectively.
B–S implied volatility is a function of both maturity and moneyness. To

keep it manageable, we only present the results for at-the-money options
(x = 1) since the B–S implied volatility from these options has the highest
correlation with the realized volatility (Table 3). Results for other money-
ness levels are not materially different. Numbers in brackets beside
the parameter estimates are the standard errors, which are estimated
following a robust procedure taking into account of heteroscedasticity
[White (1980)]. The Durbin-Watson (DW) statistic is not significantly
different from two in most regressions, indicating that the regression
residuals are not autocorrelated.
Similar to Christensen and Prabhala (1998), we formulate and test
several testable hypotheses associated with the information content of
volatility measures. We begin our discussion with the results from uni-
variate regressions. First, if a volatility forecast contains no information
about future volatility, the slope coefficient () for the volatility forecast
should be zero. This leads to our first testable hypothesis H0: ¼ 0.
Results in Table 4 strongly reject this hypothesis. In all univariate regres-
sions, the slope coefficient () is positive and significantly different from
zero at all conventional significance levels. This implies that all three
volatility measures contain important information for future volatility.
Secondly, if a given volatility forecast is an unbiased estimator of future
realized volatility, the slope coefficient () should be one and the intercept
(a) should be zero. This testable hypothesis is formulated as a joint
hypothesis H0: a ¼ 0 and ¼ 1. A w2 test is used to examine this hypoth-
esis, with test statistics reported in the column with the heading ‘‘w2
test(a).’’ Numbers in brackets beside the test statistics are p-values. The
null hypothesis is strongly rejected by the w2 test for every volatility
forecast in every regression specification. This result is not surprising
because summary statistics in Table 2 indicate that both the model-free
and B–S implied volatilities are on average greater than the realized
volatility. The evidence is also consistent with the existing option pricing
literature which documents that stochastic volatility is priced with a
negative market price of risk (or equivalently a positive risk premium).
The volatility implied from option prices is thus higher than their counter-
part under the objective measure due to investor risk aversion.
Some notable differences also exist in univariate regressions across
model specifications and volatility measures. For instance, the R2 is the
highest for the model-free implied volatility regressions ranging from
73 to 74% while it is the lowest for the lagged realized volatility regres-
sions ranging from 11 to 20%. This evidence suggests that, among the
three volatility measures, the model-free implied volatility explains the
most of the variations in future realized volatility. In other words,
the model-free implied volatility contains the most information among
the three volatility measures while the lagged realized volatility the least.
1327
The higher R2 in the model-free implied volatility regression also implies

that previous studies based on the B–S implied volatility have under-
estimated the information content of implied volatility. By aggregating
information across options with different strike prices, the model-free
implied volatility retains more information for future realized volatility
than the B–S implied volatility does.
We next consider the results from encompassing regressions involving
only the B–S implied volatility and the lagged realized volatility. In these
regressions, we investigate the informational efficiency of the B–S
implied volatility relative to the lagged realized volatility. If the B–S
implied volatility is informationally efficient, we should expect the
lagged realized volatility to be statistically insignificant. This leads to
the following null hypothesis H0:LRE = 0. If this hypothesis is sup-
ported, the historical volatility of the underlying asset is redundant
and its information content has been impounded in the B–S implied
volatility. As shown in Table 4 (the second regression in each panel), the
null hypothesis is rejected at the 5% level in two out of three specifica-
tions. Only the variance regression (panel B) fails to reject the null
hypothesis. In addition, the informational efficiency of the B–S implied
volatility can be formulated as a joint hypothesis H0: BS = 1 and
LRE = 0. It states that the B–S implied volatility is not only informa-
tionally efficient but also fully subsumes the information contained in
the lagged realized volatility. Test statistics from a w2 test [column ‘‘w2
test(b)’’] strongly reject this null hypothesis in all three specifications.
These results, together with those from the univariate regressions, pro-
vide evidence that the B–S implied volatility is a biased forecast for
future realized volatility and does not subsume all information con-
tained in past realized volatility. This finding is consistent with previous
research [e.g., Lamoureux and Lastrapes (1993), Jorion (1995), and
Christensen and Prabhala (1998)].
Finally, we examine the results from encompassing regressions invol-
ving the model-free implied volatility. A total of nine encompassing
regressions involving the model-free implied volatility are run and ana-
lyzed for alternative regression specifications and choices of volatility
measures. Combining the results from all nine regressions (the last three
regressions in each panel), we find strong evidence in support of the
hypothesis that the model-free implied volatility subsumes all information
contained in both the B–S implied volatility and the lagged realized
volatility and is a more efficient forecast for future realized volatility.
To begin with, once the model-free implied volatility is included in the
regression, the addition of either the B–S implied volatility or the lagged
realized volatility or both does not improve the regression goodness-of-fit
(adjusted R2) at all. This is clear from a direct comparison between
univariate regressions and the related encompassing regressions involving
1328
the model-free implied volatility. Secondly, in all nine encompassing

regressions involving the model-free implied volatility, the t-statistics
cannot reject the hypothesis that the slope coefficients for the B–S
implied volatility and the lagged realized volatility are zero at any con-
ventional significance levels. In fact, the estimated coefficients of the B–S
implied volatility and the lagged realized volatility are all close to zero in
the nine encompassing regressions. In addition, the t-test statistics in the
log encompassing regressions cannot reject the hypothesis that the slope
coefficient for the model-free implied volatility is one at the 5% signifi-
cance level. Furthermore, the w2 test [column ‘‘w2 test(b)’’] in the log
encompassing regressions does not reject the null hypothesis that the
slope coefficient is one for the model-free implied volatility and zero for
all other slope coefficients. These results provide evidence that the
model-free implied volatility fully subsumes information contained in
both the B–S implied volatility and the lagged realized volatility and is a
more efficient forecast for future realized volatility.
4. Robustness Tests
Results from the previous section provide strong support for the informa-
tional efficiency of the model-free implied volatility. Unlike the B–S
implied volatility, the model-free implied volatility extracts information
from options across all available strike prices. By aggregating information
contained in individual options, the model-free implied volatility exhibits
superior forecasting ability and is informationally more efficient. These
findings are striking and provide new insight on volatility forecasting and
market efficiency. We now conduct robustness tests to ensure the general-
ity of our findings.
4.1 IV regressions
Christensen and Prabhala (1998) employ IV regressions to correct for
potential EIV problems. They recognize that the B–S implied volatility
(their volatility forecast) may contain significant measurement errors
due to either the early exercise premium in the American style S & P 100
index options, the possible nonsynchronous observations of option
quotes and index levels in their data set, or the misspecification error
of the B–S model. It is well known that the EIV problem tends to drive
the slope coefficient downward (biased toward zero). This may explain
why the coefficient of the B–S implied volatility in their univariate and
encompassing regressions is below one. Christensen and Prabhala (1998)
found substantial differences between the estimation results from the
OLS and IV regressions, supporting the existence of measurement errors
in the B–S implied volatility.
1329
Compared to Christensen and Prabhala (1998) and other previous stu-

dies, our data sample is arguably less prone to measurement errors for
reasons discussed previously. In fact, our regressions involving the B–S
implied volatility, both univariate and encompassing, have much higher
R2 than the corresponding figures in previous studies. Nevertheless, it is
still important to investigate whether there is any EIV problem in the B–S
and model-free implied volatilities in our sample and whether the IV
regressions support the findings from the OLS regressions.
Following Christensen and Prabhala (1998), we apply a two-stage least
squares regression to implement the IV estimation procedure. For the B–S
implied volatility, we use the lagged realized volatility and lagged B–S
implied volatility as IVs. Similarly, we use the lagged realized volatility,
the lagged model-free implied volatility and the lagged B–S implied
volatility as IVs for the model-free implied volatility. Table 5 reports
the results from the IV regressions under the log specification as it is
econometrically better specified than the other two specifications. Results
from the OLS regressions in the first stage are reported in panel A while
those from the second stage are reported in panel B.
Consider first the forecasting ability and information content of the B–S
implied volatility. The IV regression results are found in the top half of
panel B in Table 5. From the univariate regression, the joint hypothesis
H0: a = 0 and BS = 1 is strongly rejected by the w2 test [column ‘‘w2
test(a)’’], consistent with the result from the OLS regression reported in
Table 4. This result confirms that the B–S implied volatility is a biased
forecast for future realized volatility. In comparison, the results from the
encompassing regression contradict the corresponding results from the
OLS regression. The w2-statistic [column ‘‘w2 test(b)’’] cannot reject the
joint hypothesis H0: BS = 1 and LRE = 0 at any conventional significance
levels. This is evidence that the B–S implied volatility is informationally
efficient and subsumes all information in lagged realized volatility. The
t-statistics also support this conclusion. However, estimates of both coeffi-
cients have noticeably larger standard errors. The much increased stan-
dard errors make it more difficult to reject the null hypothesis. The
Hausman (1978) w2-statistics (one degree of freedom) for testing the
EIV problem are 6.58 and 6.65 with p-values of 0.0103 and 0.0099 for
the two regressions, respectively. These test statistics strongly suggest the
presence of measurement errors and the IV regressions are less efficient
than the OLS regressions. Taken together, these results suggest that the
B–S implied volatility is more efficient than historical volatility in fore-
casting future volatility but does not subsume all information contained
in historical volatility. This is consistent with the findings in Christensen
and Prabhala (1998) from the S & P 100 index options.
Next, we examine the forecasting ability and information content of the
model-free implied volatility. The IV regression results are found in the
1330
Table 5
Univariate and encompassing regressions of 30-day log volatility instrumental variable (IV)
N LBS or LMF LRE LBS Adjusted R2 DW
Panel A: first stage regression

Dependent variable: in VtBS
77 0.93 (0.22) 0.65 (0.082) — — 0.39 2.38 — —
77 0.93 (0.23) 0.67 (0.107) 0.03 (0.072) — 0.39 2.33 — —
Dependent variable: in VtMF
77 0.67 (0.15) 0.75 (0.057) — — 0.56 2.25 — —
77 0.68 (0.15) 0.79 (0.074) 0.05 (0.065) — 0.56 2.19 — —
77 0.66 (0.16) 0.64 (0.150) — 0.11 (0.156) 0.56 2.28 — —
77 0.67 (0.16) 0.68 (0.165) 0.05 (0.064) 0.12 (0.154) 0.56 2.22 — —
N MF BS LRE adjusted R2 DW 2 test(a) 2 test(b)
Panel B: second stage IV estimates

Instrument variable for the B–S implied volatility
77 0.08 (0.40) — 0.94 (0.146) — 0.35 1.90 9.855 (0.007) —
77 0.21 (0.41) — 0.79 (0.194) 0.10 (0.080) 0.36 2.05 — 2.474 (0.290)
Instrument variables for both model-free and B–S implied volatility
77 0.06 (0.32) 0.93 (0.116) — — 0.58 1.94 31.13 (0.000) —
77 0.19 (0.34) 1.10 (0.267) –0.22 (0.288) — 0.64 1.94 — 0.774 (0.679)
77 0.12 (0.35) 0.89 (0.173) — 0.02 (0.082) 0.63 1.95 — 0.505 (0.776)
77 0.23 (0.35) 0.99 (0.294) –0.16 (0.305) 0.04 (0.076) 0.64 1.94 — 1.180 (0.758)
The at-the-money Black–Scholes (B–S) implied volatility and one-day lagged historical volatility are used in the regressions. The numbers in parentheses beside the parameter
estimates are the standard errors, which are computed following a robust procedure taking into account of the heteroscedastic and autocorrelated error structure [see Newey and
West (1987)]. The w2 test(a) is for the joint hypothesis H0: = 0 & j = 1 (j = MF, BS, LRE) in univariate regressions, and the w2 test(b) is for the joint hypothesis H0: BS = 1 and
LRE = 0 or H0: MF = 1 and BS = 1 and LRE = 0 in encompassing regressions. The test statistics are reported with the p-values in the brackets beside the statistic. Durbin-
Watson (DW) denotes the Durbin-Watson statistic.
1331
bottom half of panel B in Table 5. In both univariate and encompassing

regressions, we find no material change in statistical inferences between
IV and OLS regressions. What is changed is the increased standard errors
and reduced regression R2. However, the magnitude of the change is
relatively small compared to that in corresponding IV regressions asso-
ciated with the B–S implied volatility. The Hausman (1978) w2 test also
confirms that there is no evidence of any EIV problem in the model-free
implied volatility. As a result, the IV regressions continue to support the
hypothesis that the model-free implied volatility subsumes all information
contained in both the B–S implied volatility and the lagged realized
volatility. Our findings for the model-free implied volatility are thus
robust to the estimation method used. In comparison, there is evidence
that the B–S implied volatility contains measurement errors and these
errors are likely due to model misspecification and the use of the B–S
implied volatility from a single option.
4.2 Samples with longer maturities

In addition to the monthly nonoverlapping samples of 30-day options, we
also extract monthly samples of volatility measures for options with a
fixed maturity of 60, 120 or 180 days. We use the same monthly series of
implied volatility surface constructed in Section 2 to extract the required
implied volatility series. Again, the realized volatility is calculated using
30-minute index returns over the matching time horizon and the lagged
realized volatility is calculated using 5-minute index returns on the latest
trading day. The latter is corrected for autocorrelation using Equation (8)
with one correction term.
The monthly samples obtained this way exhibit varying degrees of
overlap. The longer the option maturity is, the more overlapped the
sample becomes. We are interested in finding out whether or not regres-
sion results change if samples with longer horizons are used in the analysis
instead of the nonoverlapping samples used previously. Overlapping
samples may lead to severe serial correlation and render the OLS test
statistics invalid. As pointed out by Richardson and Smith (1991), how-
ever, correct test statistics can be computed using the generalized method
of moments (GMM). It is also important to note that our monthly over-
lapping samples do not have the telescoping overlap problem described
by Christensen, Hansen and Prabhala (2001). Although these monthly
samples are overlapping, they are not telescoping because the time to
maturity rather than the time of maturity is fixed.
Table 6 summarizes the OLS results from the univariate and encom-
passing regressions for the log volatility measures over the 60-day, 120-
day and 180-day maturity horizons. Because the log regression is econo-
metrically better specified, we again concentrate on the log regressions in
the analysis. As indicated by the DW statistics, there is an increased
1332
Table 6
Univariate and encompassing regressions of 60-, 120- and 180-day log volatility (OLS)
N MF BS LRE Adjusted R2 Durbin-Watson 2 test(a) 2 test(b)
Panel A: In Vt (t = 60)
79 0.51 (0.21) — 0.76 (0.076) — 0.57 1.66 55.56 (0.000) —
79 0.41 (0.22) — 0.72 (0.073) 0.09 (0.051) 0.58 1.64 — 10.77 (0.004)
79 2.14 (0.20) — — 0.18 (0.085) 0.10 1.09 129.7 (0.000) —
79 0.03 (0.19) 0.93 (0.068) — — 0.70 1.61 114.6 (0.000) —
79 0.03 (0.19) 0.87 (0.127) 0.07 (0.108) — 0.70 1.62 — 1.313 (0.518)
79 0.01 (0.20) 0.92 (0.065) — 0.03 (0.038) 0.70 1.69 — 1.433 (0.488)
79 –0.00 (0.20) 0.85 (0.128) 0.07 (0.110) 0.03 (0.039) 0.70 1.72 — 2.401 (0.493)
Panel B: In Vt (t = 120)
78 0.77 (0.21) — 0.67 (0.076) — 0.50 1.51 67.57 (0.000) —
78 0.67 (0.22) — 0.63 (0.074) 0.08 (0.049) 0.50 1.49 — 19.55 (0.000)
78 2.21 (0.19) — — 0.16 (0.77) 0.08 0.83 160.4 (0.000) —
78 0.20 (0.19) 0.87 (0.097) — — 0.65 1.25 118.7 (0.000) —

78 0.18 (0.19) 0.80 (0.121) 0.08 (0.086) — 0.65 1.27 — 3.954 (0.138)
78 0.19 (0.21) 0.86 (0.091) — 0.01 (0.034) 0.65 1.25 — 3.499 (0.173)
78 0.18 (0.20) 0.79 (0.127) 0.08 (0.086) 0.01 (0.033) 0.64 1.27 — 4.251 (0.235)
Panel B: In Vt (t = 180)
78 0.62 (0.20) — 0.70 (0.073) — 0.45 0.91 142.8 (0.000) —
78 0.55 (0.21) — 0.68 (0.076) 0.05 (0.037) 0.44 0.89 — 12.82 (0.002)
78 2.29 (0.13) — — 0.13 (0.055) 0.06 0.53 354.0 (0.000) —
78 0.19 (0.27) 0.85 (0.092) — — 0.61 0.94 214.3 (0.000) —
78 0.20 (0.25) 0.79 (0.194) 0.06 (0.122) — 0.61 0.94 — 4.778 (0.091)
78 0.25 (0.27) 0.86 (0.096) — 0.04 (0.035) 0.60 0.98 — 4.341 (0.114)
78 0.21 (0.25) 0.82 (0.195) 0.05 (0.125) 0.03 (0.034) 0.60 0.96 — 4.423 (0.219)
The numbers in parentheses beside the parameter estimates are the standard errors, which are computed following a robust procedure taking into account of the heteroscedastic
and autocorrelated error structure [see Newey and West (1987)]. The w2 test(a) is for the joint hypothesis H0: = 0 and j = 1 (j = MF, BS, LRE) in univariate regressions, and the
w2 test(b) is for the joint hypothesis H0: BS = 1 and LRE = 0 or H0: MF = 1 and BS = LRE = 0 in encompassing regressions. The test statistics are reported with the p-values in
the brackets beside the statistic.
, , and indicate that the leading term coefficient of the regression is significantly different from one at the 10%, 5%, and 1% level and the ,,and indicate
that the coefficient of the remaining terms is significantly different from zero at the 10%, 5%, and 1% level, respectively.
1333
autocorrelation in regression errors from the overlapping samples. Not

surprisingly, the DW statistic declines as the forecast horizon increases
from 60 to 180 days, indicating that the degree of autocorrelation
increases as the sample becomes more overlapped. The standard errors
in all regressions are thus estimated following a robust procedure taking
into account of both heteroscedasticity and autocorrelation [Newey and
West (1987)]. The number of lags used in the estimation is set equal to the
number of overlapping periods in the regression.
The most striking finding from Table 6 is that parameter estimates and
statistical inferences from the overlapping samples are quite similar to
those from the nonoverlapping samples reported in Table 4. For example,
the slope coefficient from univariate regressions is consistently positive
and statistically significant, supporting the hypothesis that all three vola-
tility measures contain information for future realized volatility. In addi-
tion, the encompassing regressions using overlapping samples continue to
show that the B–S implied volatility does not subsume all information
contained in the lagged realized volatility. The lagged realized volatility is,
however, only marginally significant, which is not unexpected since the
daily realized volatility is much more volatile than integrated volatility
over 60 to 180 days. On the other hand, the model-free implied volatility
is found to subsume all information contained in both the B–S implied
volatility and lagged realized volatility. As long as the model-free
implied volatility is included in the regression, the B–S implied volatility
and the lagged realized volatility are no longer statistically significant.
These findings are consistent with the results from the nonoverlapping
samples.
4.3 Implied index values

Previous research [e.g., Longstaff (1995)] found that the price of the
underlying asset implied by option prices may be substantially different
from the corresponding actual market price. These differences may be
caused by market frictions (e.g., commission fees, bid-ask spread, illiquid-
ity, and taxes), measurement errors, or market inefficiency. In our empiri-
cal tests, the B–S and model-free implied volatilities are both calculated
using actual index values rather than implied index values. It is prudent to
investigate whether the regression results change materially if implied
index values are used to calculate implied volatility.
Longstaff (1995) calculated the implied value of the S & P 100 index
from option prices using either the B–S model or more generalized models
with a four-parameter distribution function. Implied index values
obtained this way are nevertheless subject to model misspecification
errors. Following Aı̈t-Sahalia and Lo (1998), we use the put-call parity
to calculate the implied index value from option prices. Model misspeci-
fication errors are avoided as the put-call parity is model free. In addition,
1334
Aı̈t-Sahalia and Lo (1998) documented that the put-call parity is rarely

violated for the SPX options.
To use the put-call parity, we need to find matched pairs of call and put
options. For each observed call price, we find the closest temporally
matched price from a corresponding put option with identical maturity
and strike price. Matched pairs with option prices observed more than
5 minutes apart are eliminated. Nonsynchronous trading is unlikely to be
a problem with this matching criterion. Finally, we minimize the effect of
illiquidity on option prices by only including options that are near the
money. Specifically, we only use options whose strike price is within 3% of
the index value. These options are traded more actively and their prices
are likely to be more efficient and reliable. For strike prices further away
from the index value, either the call or the put is deep in the money and
may not have sufficient liquidity.
Comparing the actual and implied index values over the sample period,
we find that the implied index value is within 0.64% of the actual index
value in 98% of the sample. The mean and median of the absolute
percentage difference are 0.12 and 0.08%, respectively.6 In addition, the
distribution of the percentage difference is also nearly symmetric with an
approximately 50-50 split between positive and negative differences.
These results indicate that the implied index value is generally consistent
with the actual index value. The choice of actual or implied index values
should have little impact on the calculation of the model-free implied
volatility. Unreported results indeed confirm this assertion. The choice of
actual or implied index value has little impact on the calculated model-
free implied volatility or the regression results.
4.4 Alternative measures for realized volatility

All reported results so far are based on realized volatility calculated from
30-minute index returns and lagged daily realized volatility (our proxy for
historical volatility) calculated from 5-minute index returns. The latter is
corrected for autocorrelation with one correction term while the former is
not. Although recent research on high frequency data and our sample
estimates for autocorrelations of 5-minute and 30-minute returns suggest
that this is a reasonable approach, it is prudent to consider alternative
measures for realized volatility and find out whether our regression
results are robust to them. As high frequency index returns typically
have positive autocorrelation due to the infrequent trading problem, the
estimate for realized volatility would be downward biased if no correction
is made. We now investigate the robustness of our regression results over
alternative estimates for realized volatility.
6
Compared to the corresponding numbers of 40.0 and 31.1% reported in Longstaff (1995), these differ-
ences are negligible.
1335
We first examine the impact of different number of correction terms on

our regression analysis. Since autocorrelation more or less disappears
after the first lag in our intraday return series, we examine the effect of
including zero to three correction terms in Equation (8) for both 5-minute
and 30-minute index returns. Realized volatility and lagged realized
volatility series are recalculated and the results in Table 4 are reproduced
using the new volatility estimates. Unreported results show that test
statistics continue to support our key finding that the model-free implied
volatility subsumes all information contained in the B–S implied volatility
and lagged realized volatility and is a more efficient forecast for future
realized volatility. There are also some noticeable differences as more
correction terms are added. The regression R2 is reduced and the lagged
realized volatility becomes less significant. This is an indication of an
increasing level of noise in the volatility series as more correction terms
are included. We interpret these results as evidence supporting our pre-
vious findings.
In addition, previous studies such as Andersen, Bollerslev, Diebold,
and Ebens (2001) suggest that intraday returns should be cleaned up
using an MA(1) filter before they are used to calculate realized volati-
lity. The filtered returns have much reduced autocorrelation and are
thus better suited for realized volatility calculation. Applying the
MA(1) filter to both 5-minute and 30-minute returns, we recalculate
realized volatility and lagged realized volatility series using the filtered
series and rerun our univariate and encompassing regressions. Unre-
ported results show that all parameter estimates and test statistics are
similar to those reported in Table 4 and there is no material change in
statistical inferences.
Finally, we also consider another proxy for historical volatility com-
monly used in previous research [e.g., Christensen and Prabhala (1998)]—
lagged realized volatility over a matching horizon immediately proceed-
ing the observation date. For the 30-day realized volatility calculated on
date t, the matching lagged realized volatility is calculated over the 30-day
period proceeding date t. In comparison, the lagged realized volatility on
the latest trading day is used as our proxy for historical volatility. To see
whether the regression results are influenced by our choice of the proxy,
we reproduce results in Table 4 using the lagged 30-day realized volatility
calculated using 30-minute index returns. Unreported results show that
parameter estimates and test statistics are similar when the lagged
monthly volatility is used as the proxy for historical volatility. In parti-
cular, the lagged monthly volatility is still significant in encompassing
regressions involving only the B–S implied volatility but remains insignif-
icant in encompassing regressions involving the model-free implied vola-
tility. Our main findings are thus robust to the choice of lagged realized
volatility as well.
1336
5. Conclusions
In this article, we empirically test the forecasting ability and information
content of implied volatility. Instead of relying on the B–S implied
volatility as in previous research, we implement the model-free implied
volatility derived by Britten-Jones and Neuberger (2000). This new
implied volatility has several advantages over its predecessor. First, it is
independent of any option pricing model, whereas the commonly used B–
S implied volatility is model-specific. Second, the model-free implied
volatility extracts information from options across all strike prices instead
of a single option as in the case of the B–S implied volatility. By aggregat-
ing information across options, the model-free implied volatility should
be informationally more efficient than the B–S implied volatility. Third,
tests based on the model-free implied volatility are direct tests of market
efficiency instead of joint tests of market efficiency and the assumed
option pricing model. Evidence on market efficiency from the model-
free implied volatility is thus not subject to model misspecification errors.
Our research makes several contributions to the related literature. We
provide a simpler derivation of the model-free implied volatility under
Britten-Jones and Neuberger’s (2000) diffusion assumption and then
generalize it to processes with jumps. We thus establish the validity of
the model-free implied volatility and ensure that it applies to general asset
price processes. In addition, we develop a simple method for implement-
ing the model-free implied volatility using observed option prices. We
provide theoretical bounds on truncation errors due to the finite range of
available strike prices and illustrate the minimum range of strike prices
required to control such errors. We show that the model-free implied
volatility can be calculated accurately using our implementation method.
Finally, we use univariate and encompassing regressions to investigate the
forecasting ability and information content of the model-free implied
volatility. Our findings from the SPX options support the hypothesis that
the model-free implied volatility subsumes all information contained in the
B–S implied volatility and past realized volatility and is a more efficient
forecast for future realized volatility. These results are robust to alternative
estimation methods, volatility series over different horizons, the choice of
actual or implied index values, and alternative methods for calculating
realized volatility. Our findings also provide theoretical and empirical
support for the recent decision by the CBOE to modify its widely watched
VIX index. The new VIX index is based on the model-free implied volatility
instead of the B–S implied volatility of at-the-money options.
Appendix
Proof of Proposition 1. We first prove that Equation (1) holds under the diffusion assump-
tion. The no-arbitrage argument implies that there exists a forward measure F such that:
1337
pffiffiffiffiffi
dFt =Ft ¼ Vt dWt , ðA1Þ
where Vt is the instantaneous variance and Wt is a standard F-Brownian motion. Applying

Ito’s lemma, we have:
1 pffiffiffiffiffi
d ln Ft ¼ Vt dt þ Vt dWt :
2
Integrating over time and taking expectations, we further have

ð T

E0F Vt dt ¼ 2 ln F0 E0F ðln FT Þ ,
0
2
ÐT
where the LHS is equivalent to E0F 0 dF
Ft
t . Thus we only need to prove that:
ð1 F
dK ¼ ln F0 E0F ðln FT Þ: ðA2Þ
0 K2
Applying integration by parts to the LHS of (A2), we have

ð1 1
C F ðT, KÞ maxð0, F0 KÞ C F ðT, KÞ maxðF0 K, 0Þ
dK ¼
0 K2 K
ð1 F 0
CK ðT, KÞ þ 1F0 > K
þ dK, ðA3Þ
0 K
where CKF(T,K) is the partial derivative of option price with respect to strike price.
The first term on the RHS of Equation (A3) vanishes under the mild condition that
the density of the forward price distribution is bounded. Since CKF ðT; KÞ ¼

qmaxð0; FT KÞ
E0F ¼ E0F ð1FT > K Þ, the second term on the RHS of Equation (A3)
qK
can be simplified to:
ð1
CKF ðT, KÞ þ 1F0 > K
dK ¼ ln F0 E0F ðln FT Þ:
0 K
Equation (A2) thus holds. This completes the proof for diffusion processes.
We now consider processes with jumps. Under certain regularity conditions, the forward
price as a martingale can be decomposed canonically into two orthogonal components: a
purely continuous martingale and a purely discontinuous martingale [see Jacod and Shiryaev
(1987) and Protter (1990)]. Incorporating the discontinuous component (i.e., jumps), we
restate the forward price process as:
pffiffiffiffiffi
dFt =Ft ¼ Vt dWt þ Jt dNt t t dt, ðA4Þ
where Nt is a pure jump process with time-varying intensity lt, Jt, is the jump size
with instantaneous mean mt, and the jump component is uncorrelated with the diffusion
component.
Applying Ito’s lemma, we have:
1 pffiffiffiffiffi
d ln Ft ¼ Vt dt þ Vt dWt þ ln ð1 þ Jt ÞdNt t t dt:
2
Since ln ð1 þ Jt Þ Jt 1 Jt2 , we further have:

2
1338
ð T
ð T
1
E0F d ln Ft E0F ðVt þ lt Jt2 Þdt ,
0 2 0
or equivalently:
ð T

E0F ðVt þ lt Jt2 Þdt 2 ln F0 E0F ðln FT Þ :
0
Since the LHS of the above equation is the integrated return variance, we now have:
"ð #
T
dFt 2
E0F 2 ln F0 E0F ðln FT Þ : ðA5Þ
0 F t
Note that Equation (A2) holds even when asset returns include jumps because its
derivation does not require any knowledge of the asset return process. Combining Equations
(A2) and (A5), we complete the proof for processes with jumps.7
Proof of Proposition 2. We first consider the right truncation error,

Ð þ1 C F ðT; KÞ maxð0; F0 KÞ
2 Kmax dK. From Equation (A3), this error can be rewritten as:
K2
þ1 ð þ1
C F ðT,KÞ maxð0,F0 KÞ F
2 þ2 CK ðT,KÞ þ 1F0 > K d ln K, ðA6Þ
K Kmax
Kmax
which further simplifies to:

C F ðT, Kmax Þ maxð0, F0 Kmax Þ FT

2 2E0F max 0, ln
Kmax
Kmax
F0
þ max 0, ln :
Kmax
With Kmax > F0, we have:

ð þ1 ð þ1
C F ðT, KÞ maxð0, F0 KÞ FT Kmax FT Kmax

2 dK ¼ 2 ln 1 þ fðFT ÞdFT ,
Kmax K2 Kmax Kmax Kmax
where F(FT) is the density of the forward price. Using the properties of the Taylor series
expansion for the log function, we have the following upper bound for the right truncation
error:
ð þ1 ð þ1
C F ðT, KÞ maxð0, F0 KÞ FT Kmax 2
2 dK fðFT ÞdFT : ðA7Þ
Kmax K2 Kmax Kmax

where the RHS can be rewritten as E0F FT Kmax 2 F > K , which proves Equation
Kmax T max
(3). The upper bound for the left truncation error in Equation (4) is derived in a similar
fashion.
7
Note the Equation (A5) does not hold exactly. The approximation error is due to ignoring higher order
(3) terms in the ln (1+Jt) expansion, which are mainly determined by the third moment of the jump size
Jt. When the jump size is negatively skewed, the model-free implied volatility [the RHS of Equation (A5)]
tends to overestimate the asset price variance. However, our numerical evaluation suggests that the
approximation error is negligible in commonly encountered cases (see the results in Figure 1 and Table 1).
1339
It is also straightforward to derive model-free upper bounds based on available option

prices. Invoking the monotonicity and convexity of option prices, we have
ð þ1
C F ðT, KÞ maxð0, F0 KÞ 2C F ðT, Kmax Þ
2 2
dK ,
Kmax K Kmax
ð þ1
C F ðT,KÞ maxð0, F0 KÞ 2PF ðT, Kmin Þ
2 2
dK ln ðKmin =K Þ þ eðK Þ,
Kmin K Kmin

where PF(T,K) is the forward put option price and K is a sufficiently small positive number
Ð K C F ðT; KÞ maxð0; F0 KÞ
such that eðK Þ ¼ 2 0 dK is negligible.
K2
References
Aı̈t-Sahalia, Y., and A. W. Lo, 1998, ‘‘Nonparametric Estimation of State-price Densities Implicit in
Financial Asset Prices,’’ Journal of Finance, 53, 499–547.
Aı̈t-Sahalia, Y., P. A. Mykland, and L. Zhang, 2003, ‘‘How Often to Sample a Continuous-Time Process
in the Presence of Market Microstructure Noise,’’ Review of Financial Studies, 18, 351–416.
Andersen, T. G., and T. Bollerslev, 1998, ‘‘Answering the Critics: Yes, ARCH Models Do Provide Good
Volatility Forecasts,’’ International Economic Review, 39, 885–905.
Andersen, T. G., T. Bollerslev, F. X. Diebold, and. H. Ebens, 2001, ‘‘The Distribution of Realized Stock
Return Volatility,’’ Journal of Financial Economics, 61, 43–76.
Andersen, T. G., T. Bollerslev, F. X. Diebold, and P. Labys, 2001, ‘‘The Distribution of Realized
Exchange Rate Volatility,’’ Journal of the American Statistical Association, 96, 42–55.
Andersen, T. G., T. Bollerslev, F. X. Diebold, and P. Labys, 2003, ‘‘Modeling and Forecasting Realized
Volatility,’’ Econometrica, 71, 579–625.
Barndorff-Nielsen, O. E., and N. Shephard, 2003, ‘‘Realized Power Variation and Stochastic Volatility
Models,’’ Bernouilli, 9, 243–265.
Bakshi, G., C. Cao, and Z. Chen, 1997, ‘‘Empirical Performance of Alternative Option Pricing Models,’’
Journal of Finance, 52, 2003–2049.
Bakshi, G., C. Cao, and Z. Chen, 2000, ‘‘Pricing and Hedging Long-Term Options,’’ Journal of Econo-
metrics, 94, 277–318.
Bates, D., 1991, ‘‘The Crash of ’87: Was it Expected? The Evidence from Options Markets,’’ Journal of
Finance, 46, 1009–1044.
Black, F., and M. Scholes, 1973, ‘‘The Pricing of Options and Corporate Liabilities,’’ Journal of Political
Economy, 81, 637–659.
Blair, B., S. H. Poon, and S. J. Taylor, 2001, ‘‘Forecasting S&P 100 Volatility: The Incremental Information
Content of Implied Volatility and High Frequency Index Returns,’’ Journal of Econometrics, 105, 5–26.
Breeden, D. T., and R. H. Litzenberger, 1978, ‘‘Prices of State-Contingent Claims Implicit in Option
Prices,’’ Journal of Business, 51, 621–651.
Britten-Jones, M., and A. Neuberger, 2000, ‘‘Option Prices, Implied Price Processes, and Stochastic
Volatility,’’ Journal of Finance, 55, 839–866.
Campa, J. M., K. P. Chang, and R. L. Reider, 1998, ‘‘Implied Exchange Rate Distributions: Evidence
from OTC Option Markets,’’ Journal of International Money and Finance, 17, 117–160.
Canina, L., and S. Figlewski, 1993, ‘‘The Informational Content of Implied Volatility,’’ Review of
Financial Studies, 6, 659–681.
1340
Christensen, B. J., C. S. Hansen, and N. R. Prabhala, 2001, ‘‘The Telescoping Overlap Problem in
Options Data,’’ Working paper, University of Aarhus and University of Maryland,
Christensen, B. J., and N. R. Prabhala, 1998, ‘‘The Relation between Implied and Realized Volatility,’’
Journal of Financial Economics, 50, 125–150.
Day, T. E., and C. M. Lewis, 1992, ‘‘Stock Market Volatility and the Information Content of Stock Index
Options,’’ Journal of Econometrics, 52, 267–287.
Derman, E., and I. Kani, 1994, ‘‘Riding on a Smile,’’ Risk, 7, 32–39.
Derman, E., and I. Kani, 1998, ‘‘Stochastic Implied Trees: Arbitrage Pricing with Stochastic Term and
Strike Structure of Volatility,’’ International Journal of Theoretical and Applied Finance, 1, 61–110.
Derman, E., I. Kani, and N. Chriss, 1996, ‘‘Implied Trinomial Trees of the Volatility Smile,’’ Journal of
Derivatives, 3, 7–22.
Dumas, B., J. Fleming, and R. E. Whaley, 1998, ‘‘Implied Volatility Functions: Empirical Tests,’’ Journal
of Finance, 53, 2059–2106.
Ederington, L. H., and W. Guan, 2002, ‘‘Is Implied Volatility an Informationally Efficient and Effective
Predictor of Future Volatility?’’ Journal of Risk, 4, 29–46.
Fleming, J., 1998, ‘‘The Quality of Market Volatility Forecast Implied by S&P 100 Index Option Prices,’’
Journal of Empirical Finance, 5, 317–345.
French, K. R., G. W. Schwert, and R. F. Stambaugh, 1987, ‘‘Expected Stock Returns and Volatility,’’
Journal of Financial Economics, 19, 3–30.
Hansen, P. R., and A. Lunde, 2004, ‘‘Realized Variance and Market Microstructure Noise,’’ forthcoming
Journal of Business and Economic Statistics.
Harvey, C. R., and R. E. Whaley, 1991, ‘‘S&P 100 Index Option Volatility,’’ Journal of Finance, 46, 1551–
1561.
Harvey, C. R., and R. E. Whaley, 1992a, ‘‘Market Volatility Prediction and the Efficiency of the S&P 100
Index Option Market,’’ Journal of Financial Economics, 31, 43–73.
Harvey, C. R., and R. E. Whaley, 1992b, ‘‘Dividends and S&P 100 Index Option Valuation,’’ Journal of
Futures Markets, 12, 123–137.
Hausman, J., 1978, ‘‘Specification Tests in Econometrics,’’ Econometrica, 46, 1251–1271.
Heston, S. L., 1993, ‘‘A Closed-Form Solution for Options with Stochastic Volatility with Applications to
Bond and Currency Options,’’ Review of Financial Studies, 6, 327–343.
Jackwerth, J. C., 1999, ‘‘Option-Implied Risk-Neutral Distributions and Implied Binomial Trees: A
Literature Review,’’ Journal of Derivatives, 6, 1–17.
Jacod, J., and A. N. Shiryaev, 1987, Limit Theorems for Stochastic Processes, Springer-Verlag, Berlin.
Jorion, P., 1995, ‘‘Predicting Volatility in the Foreign Exchange Market,’’ Journal of Finance, 50, 507–528.
Lamoureux, C. G., and W. D. Lastrapes, 1993, ‘‘Forecasting Stock-Return Variance: Toward an Under-
standing of Stochastic Implied Volatilities,’’ Review of Financial Studies, 6, 293–326.
Ledoit, O., and P. Santa-Clara, 1998, ‘‘Relative Pricing of Options with Stochastic Volatility,’’ Working
paper, University of California at Los Angeles.
Longstaff, F. A., 1995, ‘‘Option Pricing and the Martingale Restriction,’’ Review of Financial Studies, 8,
1091–1124.
Newey, W. K., and K. D. West, 1987, ‘‘A Simple Positive Definite Heteroskedasticity and Autocorrela-
tion Consistent Covariance Matrix,’’ Econometrica, 55, 703–708.
1341
Pong, S., M. B. Shackleton, S. J. Taylor, and X. Xu, 2004, ‘‘Forecasting Sterling/Dollar Volatility: A
Comparison of Implied Volatility and AR(FI)MA Models,’’ Journal of Banking and Finance, 28, 2541–2563.
Protter, P., 1990, Stochastic Integration and Differential Equations: A New Approach, Springer, Berlin.
Richardson, M., and T. Smith, 1991, ‘‘Tests of Financial Models in the Presence of Overlapping
Observations,’’ Review of Financial Studies, 4, 227–254.
Rubinstein, M., 1994, ‘‘Implied Binomial Trees,’’ Journal of Finance, 49, 771–818.
Rubinstein, M., 1998, ‘‘Edgeworth Binomial Trees,’’ Journal of Derivatives, 5, 20–27.
Shimko, D., 1993, ‘‘Bounds of Probability,’’ Risk, 6, 33–37.
Stoll, H. R., and R. E. Whaley, 1990, ‘‘The Dynamics of Stock Index and Stock Index Futures Returns,’’
Journal of Financial and Quantitative Analysis, 25, 441–468.
White, H., 1980, ‘‘A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for
Heteroskedasticity,’’ Econometrica, 48, 817–838.
Zhang, L., P. A. Mykland, and Y. Aı̈t-Sahalia, 2003, ‘‘A Tale of Two Time Scales: Determining
Integrated Volatility with Noisy High-Frequency Data,’’ forthcoming Journal of the American Statistical
Assocation.
Zhou, B., 1996, ‘‘High-Frequency Data and Volatility in Foreign-Exchange Rates,’’ Journal of Business
and Economic Statistics, 14, 45–52.
1342
View publication stats

The Model-Free Implied Volatility and Its Information Content

Uploaded by

Copyright:

Available Formats

The Model-Free Implied Volatility and Its Information Content

Uploaded by

Document Information

Original Title

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

The Model-Free Implied Volatility and Its Information Content

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

The Model-Free Implied Volatility and Its Information

Article in Review of Financial Studies · May 2005

The user has requested enhancement of the downloaded file.

There has been considerable research on the forecasting ability and

volatility. In contrast, Day and Lewis (1992), Lamoureux and Lastrapes

Britten-Jones and Neuberger (2000) and demonstrate that their model-

Numerical examples are provided to demonstrate the validity and accu-

1. Model-Free Implied Volatility

1.1 Diffusions and jumps

Proposition 1. The integrated return variance between the current date 0

where the superscript F denotes the forward probability measure.

asset return variance and option prices. Following Britten-Jones and

1.2 Implementation issues

1.2.1 Truncation errors. Suppose the interval [Kmin, Kmax] defines

dWt dWtv ¼ dt,

are of similar magnitude (panel B) and exhibit nearly identical convergence

variance is an increasing function of maturity. As a result, a larger range

1.2.2 Discretization errors. In addition to truncation errors, the imple-

adjustments are needed. Under the assumption of deterministic interest

Through a change of variables, the above equation is restated as:

Equation (2) provides an alternative definition of the model-free

1.2.4 Limited availability of strike prices. Up to this point, we have

is merely used as a tool to provide a one-to-one mapping between option

Kmin Kmax 30 45 60 75 90 120 180

Panel A: the truncation

2. Data and Volatility Calculation

daily returns. For example, Andersen and Bollerslev (1998), Andersen,

sBS(0.94) sBS(0.97) sBS(1.00) sBS(1.03) sMF sRE

Panel A: correlation matrix

highest correlation (85.7%), followed by the at-the-money B–S implied

3. The Information Content of Implied Volatility

measures are likely biased towards rejecting informational efficiency. In

The Review of Financial Studies / v 18 n 4 2005

N a  MF BS  LRE Adjusted R2 Durbin-Watson 2 test(a) 2 test(b)

B–S implied volatility is a function of both maturity and moneyness. To

The higher R2 in the model-free implied volatility regression also implies

the model-free implied volatility. Secondly, in all nine encompassing

Compared to Christensen and Prabhala (1998) and other previous stu-

N LBS or  LMF  LRE  LBS Adjusted R2 DW

Panel A: first stage regression

N  MF  BS  LRE adjusted R2 DW 2 test(a) 2 test(b)

Panel B: second stage IV estimates

bottom half of panel B in Table 5. In both univariate and encompassing

4.2 Samples with longer maturities

N  MF  BS  LRE Adjusted R2 Durbin-Watson 2 test(a) 2 test(b)

autocorrelation in regression errors from the overlapping samples. Not

4.3 Implied index values

Aı̈t-Sahalia and Lo (1998) documented that the put-call parity is rarely

4.4 Alternative measures for realized volatility

We first examine the impact of different number of correction terms on

where Vt is the instantaneous variance and Wt is a standard F-Brownian motion. Applying

Integrating over time and taking expectations, we further have

Applying integration by parts to the LHS of (A2), we have

Since ln ð1 þ Jt Þ Jt 1 Jt2 , we further have:

Proof of Proposition 2. We first consider the right truncation error,

which further simplifies to:

C F ðT, Kmax Þ maxð0, F0 Kmax Þ FT

With Kmax > F0, we have:

dWt dWtv ¼ dt,

N a MF BS LRE Adjusted R2 Durbin-Watson 2 test(a) 2 test(b)

N LBS or LMF LRE LBS Adjusted R2 DW

N MF BS LRE adjusted R2 DW 2 test(a) 2 test(b)

N MF BS LRE Adjusted R2 Durbin-Watson 2 test(a) 2 test(b)