Long memory in volatility and trading volume

Jeff Fleming
a
, Chris Kirby
b,
a
Jesse H. Jones Graduate School of Business, Rice University, Houston, TX 77005, USA
b
John E. Walker Department of Economics, Clemson University, Clemson, SC 29634, USA
a r t i c l e i n f o
Article history:
Received 19 May 2010
Accepted 14 November 2010
Available online 20 November 2010
JEL classication:
C32
C58
G12
Keywords:
Realized volatility
Fractional integration
Strongly autocorrelated
Bivariate mixture model
Long-range dependence
a b s t r a c t
We use fractionally-integrated time-series models to investigate the joint dynamics of equity trading vol-
ume and volatility. Bollerslev and Jubinski (1999) show that volume and volatility have a similar degree
of fractional integration, and they argue that this evidence supports a long-run view of the mixture-of-
distributions hypothesis. We examine this issue using more precise volatility estimates obtained using
high-frequency returns (i.e., realized volatilities). Our results indicate that volume and volatility both dis-
play long memory, but we can reject the hypothesis that the two series share a common order of frac-
tional integration for a fth of the rms in our sample. Moreover, we nd a strong correlation
between the innovations to volume and volatility, which suggests that trading volume can be used to
obtain more precise estimates of daily volatility for cases in which high-frequency returns are
unavailable.
2010 Elsevier B.V. All rights reserved.
1. Introduction
There is a consensus among nancial econometricians that
volatility is characterized by long memory. The consensus began
to take shape with reports of hyperbolic decay in the autocorrela-
tions of stock index and currency absolute returns (Taylor, 1986;
Ding et al., 1993). It gained momentum as fractionally-integrated
GARCH models made inroads into the volatility modeling literature
(Baillie et al., 1996; Bollerslev and Mikkelsen, 1996). More recently,
studies of realized volatility (i.e., cumulative squared intraday
returns), such as Andersen, Bollerslev, Diebold, and Ebens (hence-
forth ABDE) (2001), Andersen, Bollerslev, Diebold, and Labys
(henceforth ABDL) (2001), and Andersen et al. (2003) have pro-
duced compelling evidence of long memory for both equities and
currencies. Realized volatility appears to display all the hallmarks
of a fractionally-integrated process with a degree of fractional inte-
gration in the 0.30.5 range.
Bollerslev and Jubinski (1999) build on the evidence of long
memory in volatility to investigate the extent to which volume
and volatility share common long-run dependencies. A theoretical
result called the mixture-of-distributions hypotheses (MDH)
predicts that returns and trading volume are jointly subordinate
to an unobserved directing variable that measures the rate at
which information arrives to the market (Tauchen and Pitts,
1983). Although early tests of the MDH were generally supportive,
subsequent work using more sophisticated methods revealed that
a single latent directing variable cannot account for the observed
short-run dynamics of volume and volatility.
1
In light of this nd-
ing, and the evidence of long memory in volume as well as volatility,
Bollerslev and Jubinski (1999) argue that the focus on short-run
dynamics may be misplaced. They propose that the MDH might be
better viewed as long-run proposition.
To assess the empirical plausibility of a long-run MDH,
Bollerslev and Jubinski (1999) use a semiparametric frequency-do-
main approach to estimate the fractional order of integration of the
absolute return and trading volume series for each rm in the S&P
100 index. Their results suggest that volume and volatility have
remarkably similar memory characteristics. The estimates of the
integration order for the absolute returns have a mean value of
0.41, while the estimates for detrended trading volume have a
mean value of 0.40. Moreover, only eight of the 100 rms in the
S&P index produce evidence against the hypothesis that volume
0378-4266/$ - see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/[Link].2010.11.007

Corresponding author. Tel.: +1 864 656 0553; fax: +1 864 656 4192.
E-mail addresses: jeming@[Link] (J. Fleming), cmkirby@[Link] (C.
Kirby).
1
Empirical studies of the MDH typically focus on the relation between volume and
volatility in equity markets. Although there is an emerging literature on the relation
between volume in derivative markets and the returns and volatility for the
underlying asset (see, e.g., Chuang et al., 2009, 2010), these studies are not motivated
by the MDH.
Journal of Banking & Finance 35 (2011) 17141726
Contents lists available at ScienceDirect
Journal of Banking & Finance
j our nal homepage: www. el sevi er . com/ l ocat e/ j bf
and volatility share a common order of fractional integration. This
nding differs sharply with the ndings of studies that focus on the
short-run dynamics of volume and volatility, leading Bollerslev and
Jubinski (1999) to conclude that allowing for differing short-lived
news impacts, while imposing a common long-memory compo-
nent, may provide a better characterization of the joint volume
volatility relationship in US equity markets.
We investigate the empirical performance of a model that is
structured along the lines suggested by Bollerslev and Jubinski
(1999). Specically, we consider a trend-stationary fractionally-
integrated model for volume and volatility that allows for VAR(p)
dynamics after fractional differencing. The idea is to strike a middle
ground between the semiparametric estimation approach of
Bollerslev and Jubinski (1999) and the type of fully parametric ap-
proach that would be needed to incorporate all the structural and
distributional assumptions of the MDH. Although we assume en-
ough parametric structure to allow the long- and short-memory
components to compete on an equal footing, we do not impose a
common order of fractional integration, nor do we specify the joint
distribution of the errors. This makes it straightforward to assess
the relative importance of each component in explaining the joint
dynamics of volume and volatility, while reducing the potential for
model misspecication.
Even if volume and volatility are not fractionally integrated, our
model still has the potential to provide a useful characterization of
their joint dynamics. It is well established that mechanisms such as
occasional structural breaks can give rise to long-range depen-
dence (Diebold and Inoue, 2001; Granger and Hyung, 2004). Hence,
long-range dependence could be induced under the MDH by occa-
sional changes in the mean and/or volatility of the intra-day vol-
umes and intra-day returns that are generated by the reaction of
traders to information events. Although the resulting data generat-
ing process would fall outside of our framework, our modeling
strategy should be relatively robust to this form of misspecica-
tion. Diebold and Inoue (2001) show, for example, that in the pres-
ence of structural breaks, fractionally-integrated models provide a
useful description of volatility dynamics because they effectively
allow the unconditional variance to slowly change over time. Sim-
ilarly, Hyung et al. (2006) point out the difculty of forecasting vol-
atility breaks and nd that, in the absence of a way to identify the
breaks before they occur, fractionally-integrated models provide
the best volatility forecasts.
The dataset for our empirical analysis consists of daily observa-
tions of realized volatility and trading volume for the 20 rms in
the Major Market Index (MMI). Using realized volatilities rather than
absolute returns should lead to more precise inferences. We con-
struct the realized volatilities for the full day, as suggested by Hansen
and Lunde (2005), by choosing the linear combination of the trading-
day realized variance and the overnight squared return that is the
most efcient estimator of the integrated variance. This accounts
for the possibility that the overnight return contains relevant infor-
mation and the trading-day realized volatility is a biased estimator
of the full-day integrated volatility. We construct the trading-day
realized variance using Newey and West (1987) weights to account
for serial correlation in the intraday returns induced by microstruc-
ture effects, and we determine the optimal combination of the sam-
pling frequency for the intraday returns and the lag truncation
parameter for the weights by evaluating the effects of these choices
on the bias and efciency of the resulting estimator.
When we t our model to the log volatility and log volume ser-
ies for each rm, the results clearly support the view that volatility
and volume display long memory: the average estimated order of
fractional integration is about 0.35 in both cases. Nonetheless,
the evidence on whether the two series share a common fractional
order of integration is mixed. We can reject this hypothesis for 20%
of the MMI rms. A rejection rate of 20%, which is considerably
higher than the 8% rate reported by Bollerslev and Jubinski
(1999), raises serious doubts about the ability of a long-run version
of the MDH to adequately characterize the data. It suggests that the
dynamics of volume and volatility are simply too complex to be
produced by a single latent information arrival process or to be
captured empirically by a model with a common long-memory
component. On the other hand, fractional differencing is sufcient
to remove almost all of the predictability in the data. The estimated
rst-order autocorrelation of the fractionally-differenced series
ranges from 0.05 to 0.12 for the volatility series and 0.01 to
0.22 for the volume series. Thus short-memory components play
only a minor role in explaining the dynamics of realized volatility
and trading volume.
We also nd that the innovations in the two series are highly
correlated. This suggests that volume shocks are informative about
contemporaneous volatility shocks, raising the possibility that vol-
ume can be used in conjunction with daily returns to obtain more
precise volatility estimates for cases in which the high-frequency
return data needed to construct realized volatilities are unavail-
able. Our analysis indicates that the increase in precision could
be substantial. For example, if we regress the log realized variance
on estimates of the variance that condition on lagged information
(i.e., past realized volatility and volume observations) plus
contemporaneous volume, the mean-squared-error is 30% lower
on average than the MSE obtained using variance estimates that
condition on lagged information alone.
The remainder of the paper is organized as follows: Section 2
introduces our long-memory specication and discusses time-do-
main methods of estimation and inference. Section 3 describes the
dataset and details of our construction of the realized volatilities.
Section 4 presents the model tting results. Section 5 investigates
the informationcontent of volume innovations for volatility estima-
tion and forecasting. Section 6 offers some concluding remarks.
2. Model and econometric methodology
Let d denote the memory parameter of a fractionally-integrated
process. It is often convenient to estimate d using frequency-do-
main methods. Perhaps the most widely used approach is the log
periodogram regression rst proposed by Geweke and Porter-Hu-
dak (1983). Indeed, Bollerslev and Jubinski (1999) use a bivariate
log periodogram regression in their study of long memory in vol-
ume and volatility. They motivate this approach by noting that
the log-periodogram estimator is much simpler to implement than
the exact maximum likelihood estimator and that it imposes min-
imal structure beyond that required to ensure that the spectral
density is well behaved. In our case, however, the drawbacks of
the estimator outweigh its advantages.
First, the log-periodogram estimator of d converges at less than
a

T
_
rate. This raises concerns about efciency because the high-
frequency data needed to construct the realized volatilities are
unavailable before 1993. The efciency of the log-periodogram
estimator varies somewhat depending on the data generating pro-
cess and the choice of truncation parameter that determines the
Fourier frequencies used to construct the estimator, but we know
that it is considerably less efcient than parametric alternatives.
For example, Bollerslev and Jubinski (1999) use truncation param-
eters that are proportional to

T
_
. As a result, their estimator
achieves only a T
1/4
rate of convergence under the assumption of
Gaussian observations.
Second, the asymptotic theory for the log-periodogram estima-
tor is not well established for cases in which the observations used
to compute the periodogram are non-Gaussian. This is important
because we want to compare the results of estimating d using
the realized volatilities to the results obtained using absolute
J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726 1715
returns. Deo and Hurvich (2001) study the asymptotics of the log-
periodogram estimator based on the rst m Fourier frequencies for
the case in which returns are generated by a simple univariate sto-
chastic volatility model. They nd that for consistency and asymp-
totic normality to obtain, the value of m must increase at a rate less
than T
4d/(1+4d)
. Thus, unlike in the Gaussian case, the choice of an
appropriate m depends on the (unknown) value of d. This suggests
that robustness is a concern if the data exhibit clear departures
from normality.
Finally, the log-periodogram estimator focuses exclusively on
long-run frequencies. Bollerslev and Jubinski (1999) argue that this
is an advantage because it allows them to estimate d without mak-
ing any assumptions about the short-run dynamics of the process.
However, our objective is to provide evidence on both the short-
and long-run relations between volume and volatility. In order to
accomplish this, we need to use a more exible approach that al-
lows the short- and long-memory components of the process to
compete on an equal footing. The methodology we adopt represents
a middle ground between the semiparametric framework of the
log-periodogram estimator and the type of fully parametric ap-
proach that would be needed to incorporate all the structural and
distributional assumptions of the MDH. We turn nowto the details.
2.1. Multivariate linear regression with fractionally-integrated errors
Suppose for generality that we observe an N 1 vector time
series y
t

T
t=1
that is described by the multivariate linear regression
model
y
t
= bx
t
e
t
; (1)
where x
t
is a K 1 vector of deterministic components, b is an un-
known N K matrix, and e
t
is an unobserved N 1 error vector. Let
g
t
denote an N 1 vector of white noise innovations with
E[g
t
g
/
t
[ = X. We focus on specications in which the regression
errors are generated by a process of the form
U(L)D(L)e
t
= g
t
1(t > 0); (2)
where U(L) is a N N matrix polynomial of order p in the lag oper-
ator L with all roots outside the unit circle, D(L) is an N N diagonal
matrix with nth diagonal element
(1 L)
dn
=

j=0
d
n
j
_ _
(L)
j
;
(3)
and 1() is the indicator function. In other words, after fractional dif-
ferencing, the errors are described by a VAR(p) process with white
noise innovations. We assume that 0 6 d
n
< 0.5 for all n. Under this
restriction, the error process is asymptotically stationary and dis-
plays long memory unless all elements of d = (d
1
, . . . , d
N
)
/
are zero.
The regression model in Eq. (1) allows for deterministic trends
that are independent of the long-memory characteristics of the er-
rors. This conforms with both predictions of market microstructure
theories and empirical evidence. Trading volume often displays
strong time trends that are qualitatively consistent with the theo-
retical consequences of growth in the number of traders in the
market (see, e.g., Tauchen and Pitts, 1983), and a number of recent
studies suggest that the volatility of individual stock returns has
increased over time (e.g., Campbell et al., 2001). Since most of
the models studied in the volume and volatility literature assume
an absence of time trends, it is common to detrend the data before
conducting the econometric analysis (see, e.g., Gallant et al., 1992;
Andersen, 1996; Bollerslev and Jubinski, 1999). We explicitly
consider the potential impact of this detrending on the procedures
used for estimation and inference.
A potential alternative to detrending would be to model trading
volume as a unit root process. However, the empirical evidence does
not favor this approach. Bessembinder and Seguin (1993), for exam-
ple, investigate the dynamic properties of trading volume for a wide
range of currency, nancial, and commodity futures contracts. They
nd that the hypothesis of a unit root in trading volume is rejected
for every contract examined. Similarly, Darbar and Deb (1995) test
the hypothesis of a unit root in trading volume for 22 stocks traded
onNewYork Stock Exchange. They report that the evidence strongly
suggests that daily trading volume is stationary around a linear
trend. These ndings are consistent with the emphasis on deter-
ministic volume trends in the empirical literature.
2.2. Model tting and hypothesis testing
First consider the case in which b = 0. Let h = (d
/
, /
/
)
/
where /
= vec(U
1
U
2
. . . U
N
) is an Np 1 vector that contains the VAR(p)
parameters. After concentrating out X, we can express the log like-
lihood function for a model with Gaussian g
t
as
L
T
(h) =
TN
2

TN
2
log(2p)
T
2
log [X(h)[; (4)
where
X(h) =
1
T

T
t=1
[U(L)D(L)e
t
[[U(L)D(L)e
t
[
/
: (5)
If g
t
is non-Gaussian, then L
T
(h) is a quasi-log likelihood function.
Intuitively, quasi-maximum likelihood should be a reasonable esti-
mation strategy provided that suitable regularity conditions are
satised.
Nielsen (2004b), for example, analyzes the properties of the
QML estimator under the assumption that g
t
is i.i.d. with posi-
tive-denite covariance matrix and nite fourth moments. He
shows that there exists a local maximizer
^
h = (
^
d
/
; ^ u
/
)
/
of L(h) that
is a

T
_
-consistent and asymptotically normal estimator of h. Spe-
cically, we have

T
_
(
^
h h)
d
N(0; R); (6)
with
R =
p
2
6
XX
1
J
/
(K
/
I)
(KI)J C X
1
_ _
1
; (7)
where C is the covariance matrix of (e
/
t
; . . . ; e
/
tp1
)
/
, J = (vece
1
e
/
1
; . . . ;
vece
N
e
/
N
) with e
i
denoting the ith unit N-vector, and
K = (K
/
1
; . . . ; K
/
p
)
/
with K
i
=

j=i
j
1
W
ji
and W
i
denoting the coef-
cient matrix on e
ti
in the Wold representation of {e
t
}.
In light of these results, testing whether d is equal to a known
vector is straightforward. Consider a null hypothesis of the form
H
0
:d = d
0
where d
0
is a prespecied N 1 vector and let
~
/ denote
the estimate of / obtained by imposing H
0
. The quasi-likelihood ra-
tio (LR), Wald (W), and Lagrange multiplier (LM) statistics are
LR = 2(L
T
(
^
d;
^
/) L
T
(d
0
;
~
/)); (8)
W = T(
^
d d
0
)
/
p
2
6
XX
1
(XK
/
C
1
KX) X
1
_ _
(
^
d d
0
) (9)
LM = S
T
(d
0
;
~
/)
/
p
2
6
XX
1
(XK
/
C
1
KX) X
1
_ _
S
T
(d
0
;
~
/) (10)
where S
T
(d
0
;
~
/) =

T
_
T1
j=1
j
1
J
/
(I X(d
0
;
~
/)
1
)(vec

C(j)) with

C(j) =
(1=T)

T
t=j1
~ g
t
~ g
/
tj
denotes the score vector evaluated under H
0
.
Under a local alternative of the form d = d
0
d=

T
_
with d a xed
N-vector, Nielsen (2004b) shows that the statistics in Eqs. (8)
(10) are asymptotically equivalent and distributed as v
2
N
(f) with
f = d
/
p
2
6
XX
1
(XK
/
C
1
KX) X
1
_ _
d. The LR statistic, which
does not require estimation of the asymptotic covariance matrix
1716 J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726
of
^
d, is particularly convenient for applications in which there is no
desire to compute standard errors.
We can also apply these results for the case in which b 0 pro-
vided that we estimate b in a way that does not inuence our
asymptotic inference on d. To illustrate, consider a univariate mod-
el with H
0
:d = d
0
. Nielsen (2004a) shows that if we estimate b by
^
b =

T
t=1
y
+
t
x
+/
t
_ _

T
t=1
x
+
t
x
+/
t
_ _
1
; (11)
where x
+
t
= (1 L)
d
0
x
t
and y
+
t
= (1 L)
d
0
y
t
, then we can treat the
residual y
t

^
bx
t
, which equals e
t
(b
^
b)x
t
under H
0
, as observed
for the purpose of conducting asymptotic inference on d. This pre-
sumably holds in the multivariate setting as well.
Of course, this strategy works only if d
0
is prespecied. In our
empirical application, we are mainly interested in testing hypoth-
eses of the form H
0
:d = d
0
i where d
0
is an unspecied scalar and i is
an N 1 vector of ones, i.e., our null is that the observed series
share a common unknown order of integration. One way to deal
with this situation is to estimate b by
^
b =

T
t=1
y
t
x
/
t
_ _

T
t=1
x
t
x
/
t
_ _
1
; (12)
and then estimate h by maximizing
L
T
(
^
b; h) =
TN
2

TN
2
log(2p)
T
2
log [X(
^
b; h)[; (13)
where X(
^
b; h) is computed as in Eq. (5) except that we replace e
t
with the vector of residuals
^
e
t
= y
t

^
bx
t
. This is equivalent to
extracting a trend via OLS regression prior to tting the model to
the data.
To justify this approach, we need to argue that replacing L
T
(b; h)
with L
T
(
^
b; h) has no effect on the limiting distribution of
^
h. The
analysis of Giraitis and Koul (1997) supports this argument. They
consider maximum likelihood estimation of a multiple linear
regression model in which the errors are a non-decreasing function
of a stationary long-memory Gaussian process. For a fairly broad
class of deterministic regressors (e.g., bx
t
is a polynomial in t),
the estimator of the parameter vector that determines the memory
characteristics of errors is

T
_
-consistent and asymptotically nor-
mal with an asymptotic covariance matrix that attains the Cra-
mer-Rao lower bound. Since multiple linear regression with
Gaussian long-memory errors is covered by these results, using
OLS to detrend the data is asymptotically sound for g
t
- NID (0,
X). To our knowledge, the question of whether this holds for a
Gaussian QML approach with i.i.d. innovations has not been ad-
dressed in the literature. But it seems likely that an extension is
possible because the non-Gaussian log likelihoods studied by the
Giraitis and Koul (1997) have a quadratic structure similar to that
of the QML objective function.
3. Data
The dataset consists of daily realized volatilities and trading
volumes for the 20 rms in the MMI.
2
We construct the dataset
from two sources: intradaily transaction records from the Trade
and Quote (TAQ) database of the New York Stock Exchange (NYSE),
and information on stock splits, stock dividends, and cash dividends
from the Center for Research in Security Prices (CRSP) daily stock
price le. The sample period is January 4, 1993 to December 31,
2003 (2771 observations).
3
We exclude records from the TAQ database that have an out-of-
sequence time stamp, a zero price, a correction code greater than
two (indicating errors and corrections), or a condition code (indi-
cating nonstandard settlement). We also apply two lters to iden-
tify obvious price reporting errors. First, we exclude records that
have a reported price more than 20% different than the previous
transaction price. Second, we ag records that imply a price change
greater than 2% in magnitude from the previous transaction price
and which are followed by a price reversal greater than 2%. We ex-
clude the record if the implied price change is more than two times
greater than the next largest price change on that day, or if the re-
ported price falls outside the days high-low range (ignoring the
agged price) by more than the next largest price change.
We construct the daily trading volume for each rm by aggre-
gating the transaction volumes for all of the remaining TAQ records
on a given day. We adjust the daily volumes as necessary to ac-
count for stock splits and stock dividends using the information re-
ported in the CRSP le. We construct the daily realized volatility for
each rm using the ltered TAQ transaction prices and the proce-
dure explained in Section 3.1. Implementing this procedure also re-
quires the daily overnight return. We compute this return using
the last transaction price on a given day, and the rst transaction
price on the following day, adjusted for cash dividends and stock
distributions reported in the CRSP le.
3.1. Constructing the realized volatilities
Merton (1980) introduced the concept of realized volatility. In
its simplest form, the realized variance on day t is the sum of the
squared intradaily returns, r
t
i;m
; i = 1, . . . , m, over m equally-spaced
intervals,
RV
t
=

m
i=1
r
2
t
i;m
: (14)
Under certain conditions, the realized variance should be close
to the true variance. For example, if returns are generated by a con-
tinuous-time process with instantaneous volatility r
t
, then it is
natural to use the integrated variance IV
t
=
_
1
0
r
2
ts
ds as a measure
of the daily variance. Andersen et al. (2001) and Barndorff-Nielsen
and Shephard (2002) show that under weak regularity conditions,
RV
t
IV
t
?0 almost surely as m?. Thus, by increasing the sam-
pling frequency of returns, we can construct consistent nonpara-
metric estimates of the integrated volatility that in principle are
arbitrarily efcient.
In practice, realized volatilities can be biased by market micro-
structure effects, and these effects are typically more pronounced
as m increases. A number of strategies have been proposed to deal
with this issue. Andersen and Bollerslev (1997) argue that using 5-
min returns strikes a reasonable balance between the bias and the
efciency gains associated with more frequent sampling.
4
Bandi
and Russell (2008) derive the optimal sampling frequency under a
mean-squared-error criterion. Andersen et al. (2001) construct real-
ized volatility using the residuals from an MA(1) model tted to re-
turns. Andersen et al. (2001) use linear interpolation to estimate the
prices used in constructing the returns. Each of these strategies,
2
These rms are American Express (AXP), AT&T (T), ChevronTexaco (CVX), Coca-
Cola (KO), Disney (DIS), Dow Chemical (DOW), DuPont (DD), Eastman Kodak (EK),
Exxon-Mobil (XOM), General Electric (GE), General Motors (GM), International
Business Machines (IBM), International Paper (IP), Johnson & Johnson (JNJ), McDon-
alds (MCD), Merck (MRK), 3M (MMM), Philip Morris (MO), Procter and Gamble (PG),
and Sears (S).
3
Philip Morris has only 2770 observations. Its stock did not open on May 25, 1994
in advance of a board meeting regarding a proposal to split the rms food and
tobacco businesses.
4
This strategy is employed by some recent papers in this journal, e.g., Wright and
Zhou (2009) and Chuli et al. (2010).
J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726 1717
however, either places an upper bound on m or does not sufciently
eliminate the serial correlation in returns when m is large. We adopt
an alternative approach suggested by Hansen and Lunde (2004)
which directly accounts for serial correlation in returns when con-
structing the realized variances. Specically, we construct the trad-
ing-day realized variance for day t using a Newey and West (1987)
estimator of the form,
RV
t[o[
=

m
i=1
r
2
t
i;m
2

q
j=1
1
j
q 1
_ _

mj
i=1
r
t
i;m
r
t
j;m
; (15)
where q denotes the window length for the autocovariance terms.
This estimator guarantees nonnegativity and is consistent in the
presence of serial correlation, which allows us to sample returns
at very high frequencies.
Table 1
Model tting results using the standard estimator of realized volatility.
Firm ^
d
1
^
d
2
se(
^
d
1
) se(
^
d
2
)
^
/
1
^
/
2
^
/
3
^
/
4
t(
^
/
1
) t(
^
/
2
) t(
^
/
3
) t(
^
/
4
)
L
T
LR
Panel A: Long-memory and VAR parameters
AXP 0.45 0.39 0.02 0.02 0.06 0.10 0.03 0.01 1.78 2.48 1.64 0.27 587.5 5.19
*
CVX 0.42 0.29 0.02 0.02 0.10 0.05 0.00 0.07 3.33 1.46 0.02 1.94 182.0 23.15
*
DD 0.43 0.33 0.02 0.02 0.07 0.05 0.00 0.12 2.15 1.19 0.20 3.19 289.7 11.05
*
DIS 0.45 0.36 0.02 0.02 0.15 0.01 0.05 0.11 5.04 0.29 3.40 3.06 369.2 11.44
*
DOW 0.45 0.32 0.02 0.02 0.13 0.10 0.02 0.02 4.61 2.75 1.00 0.62 1059.4 24.81
*
EK 0.39 0.34 0.02 0.02 0.11 0.01 0.08 0.18 3.44 0.22 4.57 4.54 1070.0 6.19
*
GE 0.43 0.37 0.02 0.02 0.08 0.06 0.06 0.03 2.43 1.79 2.60 0.78 173.6 5.89
*
GM 0.41 0.36 0.02 0.02 0.10 0.02 0.03 0.11 3.16 0.50 2.02 2.94 631.2 4.95
*
IBM 0.46 0.39 0.02 0.02 0.10 0.04 0.07 0.11 2.76 0.88 3.04 2.77 249.1 11.61
*
IP 0.42 0.32 0.02 0.02 0.12 0.02 0.02 0.14 4.19 0.41 1.69 3.99 619.1 11.51
*
JNJ 0.43 0.33 0.02 0.02 0.12 0.01 0.05 0.11 3.74 0.19 2.79 3.12 218.9 15.18
*
KO 0.43 0.25 0.02 0.03 0.09 0.06 0.04 0.22 2.75 1.45 2.71 5.73 74.7 39.39
*
MCD 0.42 0.34 0.02 0.02 0.10 0.04 0.01 0.09 3.10 1.07 0.55 2.52 479.4 22.62
*
MMM 0.41 0.34 0.02 0.02 0.08 0.03 0.04 0.10 2.61 0.65 2.07 2.68 560.8 24.47
*
MO 0.41 0.33 0.02 0.02 0.13 0.04 0.10 0.19 3.92 1.08 4.95 5.01 744.1 11.67
*
MRK 0.42 0.30 0.02 0.02 0.07 0.01 0.03 0.19 2.23 0.24 1.68 4.96 168.0 18.73
*
PG 0.43 0.33 0.02 0.02 0.05 0.05 0.03 0.09 1.45 1.20 1.56 2.57 282.3 12.89
*
S 0.41 0.38 0.02 0.02 0.08 0.04 0.03 0.08 2.46 1.04 2.19 2.23 927.7 1.84
T 0.42 0.39 0.02 0.02 0.07 0.03 0.03 0.11 2.17 0.64 1.90 3.16 531.3 0.91
XOM 0.46 0.27 0.02 0.02 0.11 0.07 0.02 0.10 3.75 2.13 0.95 2.94 452.5 49.71
*
Innovations to logRV
t
Innovations to logV
t
Firm ^
b
11
^
b
12
^
b
13
^
b
21
^
b
22
^
b
23

X
11

X
12

X
22
CS CK Q
5
Q
10
CS CK Q
5
Q
10
Panel B: Trend parameters and innovation diagnostics
AXP 3.50 0.16 0.00 1.27 0.41 0.15 0.05 0.04 0.13 0.34 1.72 5.44 12.24 0.13 1.65 3.65 10.31
CVX 2.77 1.77 1.75 0.09 0.19 1.48 0.04 0.03 0.10 0.37 1.87 6.91 21.94
*
0.40 3.14 4.03 11.14
DD 2.95 1.43 1.20 0.40 1.38 0.73 0.05 0.03 0.11 0.28 1.18 9.10 16.68 0.42 2.37 1.64 20.05
*
DIS 3.32 0.14 0.46 1.15 0.71 0.20 0.05 0.03 0.12 0.62 2.29 7.75 13.38 0.79 2.98 7.83 13.96
DOW 2.96 0.25 0.31 0.45 0.76 0.13 0.06 0.04 0.14 0.38 2.17 13.68
*
21.43
*
0.47 3.11 4.14 6.64
EK 3.26 0.02 0.33 0.07 1.16 2.26 0.06 0.06 0.18 0.58 2.29 20.15
**
29.22
**
1.10 4.51 11.59
*
15.52
GE 2.76 1.70 1.21 2.41 0.14 0.57 0.05 0.03 0.08 0.35 1.48 3.80 19.40
*
0.34 1.84 0.67 14.56
GM 3.43 0.27 0.23 0.83 0.96 1.79 0.05 0.04 0.13 0.37 1.40 9.30 29.78
**
0.30 1.58 5.21 13.16
IBM 3.29 0.79 0.78 2.18 0.36 0.74 0.06 0.05 0.11 0.33 1.05 7.90 15.07 0.57 2.27 7.60 22.93
*
IP 2.70 2.91 2.52 0.41 1.80 0.31 0.05 0.03 0.13 0.20 1.07 3.50 17.65 0.23 0.64 0.46 12.64
JNJ 3.34 0.60 0.37 1.53 0.76 1.44 0.05 0.03 0.10 0.23 1.31 10.61 18.84
*
0.40 1.90 6.59 19.64
*
KO 3.16 0.39 0.51 0.89 0.59 0.11 0.04 0.03 0.10 0.30 1.53 6.39 11.15 0.38 1.68 3.57 7.58
MCD 3.12 0.50 0.38 0.90 0.72 0.11 0.05 0.03 0.13 0.51 2.50 5.04 9.11 0.55 1.78 4.59 12.69
MMM 2.55 2.58 2.29 0.08 1.25 0.37 0.05 0.04 0.13 0.21 1.07 7.46 15.18 0.71 3.22 1.53 6.74
MO 3.13 1.20 1.14 1.84 0.13 0.26 0.07 0.05 0.13 1.06 4.44 3.21 4.62 0.95 3.66 2.41 6.64
MRK 3.47 0.81 0.60 1.69 0.64 0.82 0.05 0.04 0.10 0.34 2.85 11.89
*
15.21 0.57 2.55 12.94
*
20.83
*
PG 3.04 1.48 1.70 0.26 1.30 0.31 0.05 0.03 0.11 0.27 1.23 4.87 16.06 0.76 6.44 3.94 21.57
*
S 3.23 1.03 0.67 0.14 0.25 1.13 0.05 0.04 0.15 0.25 1.41 16.31
**
27.32
**
0.62 2.54 5.00 12.50
T 3.01 1.36 0.74 0.94 3.07 0.64 0.05 0.03 0.12 0.49 2.24 3.20 9.98 0.85 3.83 2.51 16.94
XOM 2.68 1.64 1.39 1.22 1.63 0.40 0.04 0.02 0.08 0.32 1.48 4.66 16.09 0.15 2.06 2.00 10.67
The table reports the results of tting a bivariate linear regression model with deterministic regressors and long-memory errors to the daily realized variance and daily
trading volume for the MMI rms. The model is
y
t
= bx
t
e
t
;
U(L)D(L)e
t
= g
t
;
where y
t
= (logRV
t
, logV
t
)
/
with RV
t
based on a 5-min sampling frequency and a window length of zero, x
t
= (1, t/T, t
2
/T
2
)
/
, U(L) = I UL, D(L) = diag((1 L)
d1
; (1 L)
d2
), and
g
t
- NID (0,X) with g
t
= 0 "t 6 0. We estimate d and / = vec(U) by Gaussian quasi-maximum likelihood, which is equivalent to minimizing [

X(d; /)[ where

X(d; /) =

T
t=1
[U(L)D(L)(y
t

^
bx
t
)[[U(L)D(L)(y
t

^
bx
t
)[
/
with
^
b =

T
t=1
y
t
x
/
t
_ _

T
t=1
x
t
x
/
t
_ _
1
:
Panel A reports the elements of
^
d and their asymptotic standard errors, the elements of
^
/ and their asymptotic t-ratios, the quasi-log likelihood value (LT ), and the quasi-like-
lihood-ratio statistic (LR) for H
0
: d
1
= d
2
. Panel B reports the elements of
^
b and summary statistics for the tted innovation sequence g
t

T
t=1
implied by the parameter estimates.
The summary statistics are the unique elements of

X, the coefcients of skewness (CS) and excess kurtosis (CK) for the tted innovations, and the Leung-Box statistics com-
puted from the rst ve and 10 sample autocorrelations of the tted innovations (Q
5
and Q
10
). An asterisk indicates a likelihood ratio or Leung-Box statistic that is statistically
signicant at the 5% level. The realized variances are expressed as squared annualized percentage rates (assuming 252 trading days per year). Trading volume is measured in
millions of shares. The sample period is January 1, 1993 to December 31, 2003.
1718 J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726
The trading-day realized variance may be a biased estimator of
the daily integrated variance because it ignores returns during
nontrading periods overnight and on weekends. We construct the
full-day realized variance by combining RV
t[o]
with the squared
nontrading-period return, r
2
t[c[
using the weighting scheme pro-
posed by Hansen and Lunde (2005). They consider the class of con-
ditionally unbiased estimators that are linear in RV
t[o]
and r
2
t[c[
and
show that the following weights deliver the lowest mean squared
error:
RV
t
= u
l
l
o
RV
t[o[
(1 u)
l
l
c
r
2
t[c[
; (16)
where
u =
l
2
o
1
2
c
l
o
l
c
1
oc
l
2
c
1
2
o
l
2
o
1
2
c
2l
o
l
c
1
oc
; (17)
and l=E(r
2
t
); l
o
=E(RV
t[o[
); l
c
=E(r
2
t[c[
); 1
2
o
=var(RV
t[o[
); 1
2
c
=var r
2
t[c[
_ _
,
and 1
oc
=cov RV
t[o[
; r
2
t[c[
_ _
.
5
The ratios l/l
o
and l/l
c
scale the variance
estimates to match the unconditional mean of the squared close-to-
close returns and u determines the relative weights on the trading-
and nontrading-period variance estimates. In general, u should be close
to one because r
2
t[c[
is a relatively imprecise estimator of the nontrading-
period variance and this variance is typically lower than the variance
during the trading day. These effects can easily be seen in Eq. (17) for
the case in which 1
oc
= 0.
To implement Eq. (15), we need to specify the sampling fre-
quency (m) and the window length for the autocovariance terms
(q). Since the optimal values are not known a priori, we consider
a range of candidate values and evaluate their effect on the prop-
erties of the realized variances. The candidate values for m corre-
spond to returns sampled as nely as every 30 seconds, and the
values for q correspond to four different window lengths: 0, 15,
30, and 60 min. We construct the returns used to implement Eq.
(15) by applying the Andersen and Bollerslev (1997) linear interpo-
lation scheme to the ltered intraday transaction prices from the
TAQ database.
It is common practice to construct realized variances using 5-
min returns and with q = 0. We nd that realized variances con-
structed using this combination of parameter values are severely
biased, with the average realized variance 13% greater than the
average squared open-to-close return. The magnitude of bias in-
creases rapidly with the sampling frequency, but increasing the
window length counteracts the bias. Using a 30-min window pro-
duces realized variances that, for any choice of sampling frequency,
are within 2% of the average squared open-to-close return. Increas-
ing the window length further (e.g., 60 min) substantially increases
the standard deviation of the realized variances. Therefore, we con-
clude that the best combination of parameter values is a 30-sec-
onds sampling frequency and a 30-min window length.
We construct the full-day realized volatilities by substituting
the sample analogs of l; l
o
; l
c
; 1
2
o
; 1
2
c
; and 1
oc
into Eqs. (16)
and (17). Hansen and Lunde (2005) suggest removing outliers from
the estimation to avoid obtaining a negative weight on r
2
t[c[
. Accord-
ingly, we exclude days in which either RV
t[o]
or r
2
t[c[
is among the
largest 0.5% of the observations for a given stock. The average u
estimate for the 20 stocks is 0.92. By comparison, the ratio of the
average squared close-to-close return to the average squared
close-to-open return indicates that 20% of the daily variance occurs
during the nontrading period. The u estimate gives less weight
than this to the nontrading-period variance estimate because the
trading-period variance estimate is much more precise.
4. Model tting results
We t our long-memory model using three different proxies for
the volatility of daily returns. First, to provide a set of baseline re-
sults, we t the model with y
t
= (logRV
t
, logV
t
)
/
where RV
t
is con-
structed with m = 78 and q = 0, the parameter choices most
commonly employed in prior research (i.e., 5-min sampling fre-
quency, without the Newey-West correction). We refer to this
measure as the standard estimator of realized volatility. Next, we
t the model replacing RV
t
with the estimator obtained using a
30-s sampling frequency and a 30-min window length. We refer
to this measure as the higher-frequency, biased-corrected (HFBC)
estimator of realized volatility. Finally, we t the model using daily
squared returns in place of RV
t
to investigate how the precision of
the volatility estimates impacts our inference.
4.1. Results using the standard estimator
Table 1 reports the model tting results obtained using the
standard estimator of realized volatility. In parameterizing the
model, we assume that a VAR(1) process is sufcient to capture
the short-run dynamics of the fractionally-differenced series. Panel
A of the table reports the memory parameter estimates and their
asymptotic standard errors, the VAR coefcient estimates and their
asymptotic t-ratios, the value of the quasi-log likelihood, and the
likelihood ratio statistic for the hypothesis d
1
= d
2
. Panel B reports
the trend parameter estimates, the innovation covariance matrix
estimates, and selected model diagnostics.
As expected, the estimates of d
1
point to long memory in the
volatility process. The mean of the estimates is 0.43 and the range
is relatively narrow (0.39 for Eastman Kodak to 0.46 for IBM and
Exxon-Mobil). This nding is broadly consistent with the evidence
of Bollerslev and Jubinski (1999), who report that the absolute re-
turns for the S&P 100 rms have a mean estimated order of frac-
tional integration of 0.41. Of course, our memory parameter
estimates should be much more precise. All of the standard errors
in Table 1 are equal to 0.02, but the standard error of the log-peri-
odogram-based estimator is three times as large using the
maximum sample size considered by Bollerslev and Jubinski
(1999).
6
Thus, the QML estimator of d
1
appears to have a substantial
efciency advantage.
The estimates of d
2
also point to long memory in trading vol-
ume. However, the mean of the estimates (0.34) is lower than
the mean for the volatility process and the range is wider as well
(0.25 for Coca Cola to 0.39 for American Express, IBM, and AT&T).
This nding represents a signicant departure from Bollerslev
and Jubinski (1999). They nd that detrended trading volume has
a mean estimated order of fractional integration of 0.40, while 13
of our estimates of d
2
are more than two standard errors below
0.40. Moreover, they report only eight rejections of the hypothesis
d
1
= d
2
for the S&P 100 rms, while we reject this hypothesis for 18
of the 20 MMI rms. There are many factors, such as different esti-
mators, sample periods, volatility measures, and rms, that could
contribute to the disparity in ndings. We provide more evidence
in this regard shortly.
5
Strictly speaking, this estimator is not optimal under a MSE criterion if we
construct RV
t[o]
using Newey-West weights since this does not produce a condition-
ally unbiased estimator. As a practical matter, however, the magnitude of the bias
appears to be small. Hansen and Lunde (2005) favor a different weighting scheme
which leads to conditionally unbiased estimators, but it does so by admitting the
possibility of negative estimates. Using their weighting scheme produces an average
of over 40 negative variance estimates per stock over our sample period.
6
In particular, the asymptotic standard error is p=

24
_
_ _
T
1=4
and the maximum
value of T is 8440 daily observations.
J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726 1719
The estimates of Usuggest that long memory is the main source
of time variation in the conditional means of logRV
t
and logV
t
, i.e.,
fractional differencing eliminates most of the serial correlation in
the data. Our estimate of the rst-order autocorrelation in
(1 L)
d
1
logRV
t
ranges from 0.05 (Procter and Gamble) to
0.15 (Disney), while the range for (1 L)
d
2
logV
t
is 0.01 (Ameri-
can Express) to 0.22 (Coca-Cola). Although most of the autocorre-
lation estimates are statistically signicant at the 5% level, the
level of predictability implied by the VAR(1) specication is small.
Apparently the short-memory component of the model plays only
a minor role in explaining the dynamics of realized volatility and
trading volume.
Panel B of the table shows that most rms display some evi-
dence of trends of one form or another. This follows by comparing
the estimated means of logRV
t
and logV
t
at the start of the sample
(i.e.,
^
b
11
and
^
b
21
) to the estimated means at the end of the sample
(i.e.,

3
i=1
^
b
1i
and

3
i=1
^
b
2i
). With the exception of IBM, the estimated
volume trend is positive, which is not particularly surprising given
the growth in equity market participation over time. However,
there are also indications of positive volatility trends for 14 of
the 20 rms. This could be a symptom of increasing idiosyncratic
volatility in recent years (see, e.g., Campbell et al., 2001).
The estimates of X imply that the log volatility and log volume
innovations typically have a correlation between 40% and 60%. It
follows, therefore, that volume shocks are quite useful for predict-
ing contemporaneous volatility shocks, which is consistent with
the evidence reported by Fleming et al. (2006). They use linear
state-space methods to t a factor model to daily trading volume
and daily squared returns for the MMI rms, and nd that the non-
persistent component of volume is highly correlated with the con-
temporaneous nonpersistent component of volatility. Specically,
this component of volume typically captures 6080% of the varia-
tion in nonpersistent volatility.
The diagnostics for the model suggest that it is reasonably well
specied. The innovations to logRV
t
tend to have small estimated
coefcients of skewness and excess kurtosis, with most of the
skewness values less than 0.5, and most of the excess kurtosis val-
ues less than 2. There is some evidence of statistically signicant
serial correlation in the innovations. The Leung-Box statistic based
on the rst ve autocorrelations produces four rejections at the 5%
level, and the statistic based on the rst 10 autocorrelations pro-
duces seven rejections at the 5% level. Nonetheless, treating the
log volatility innovations as Gaussian white noise is probably a rea-
sonable approximation.
Normality is somewhat less plausible for the log volume inno-
vations. The coefcient of skewness ranges from 0.13 (American
Express) to 1.10 (Eastman Kodak), while the coefcient of excess
kurtosis ranges from 0.64 (International Paper) to 6.44 (Procter
and Gamble). Nonetheless, the Leung-Box statistics produce about
the same number of rejections as for realized volatility, which sug-
gests we could reasonably treat the volume innovations as non-
Gaussian white noise. Using the rst ve autocorrelations produces
two rejections at the 5% level, and using the rst 10 autocorrela-
tions produces ve rejections. We could reduce the number of
rejections by tting a higher-order VAR specication, but standard
criteria for selecting the dimension of a model, such as the Bayes-
ian information criterion, favor the more parsimonious VAR(1)
parameterization.
4.2. Results using the HFBC estimator
Table 2 reports the model tting results obtained using the
HFBC estimator of realized volatility. These results provide addi-
tional insights into the results reported in Table 1. In general, our
previous results bolster the case for using long memory specica-
tions to model the dynamics of volatility and volatility. However,
the one nding at odds with prior research is the lack of empirical
support for the hypothesis that the two series share a common
long-memory parameter. We now assess whether these results,
and this last nding in particular, are robust to our choice of real-
ized volatility measure.
The estimates of d
1
reported in Table 2, like our earlier results,
point to long memory in the volatility process. However, a key dif-
ference is that the estimates are now much lower; the mean of the
estimates is 0.36, and the range is 0.330.40. Since these estimates
are much closer to the estimates of d
2
and the standard errors are
essentially unchanged, the number of rejections of the hypothesis
d
1
= d
2
falls from 18 in Table 1 to only four in Table 2. It seems that
using a more precise realized volatility estimator has a substantial
impact on our inferences regarding the empirical plausibility of a
common long-memory parameter. In all other respects the results
in Panel A of Table 2 are similar to those reported in Panel A of Ta-
ble 1.
The diagnostics reported in Panel B of Table 2 suggest a mod-
est improvement in the specication of the model. In particular,
the volatility innovations tend to have smaller estimated coef-
cients of skewness and excess kurtosis than in Table 1. Most of
skewness values are less than 0.4, and most of the excess kurto-
sis values are less than 1. Although there is still evidence of sta-
tistically signicant serial correlation in the innovations, the
number of rejections based on the rst ve autocorrelations falls
from four to three, and the number of rejections based on the
rst 10 autocorrelations falls from seven to four. These changes,
while relatively minor, suggest that there are gains to using an
improved realized volatility measure to t the long memory
model.
Although there is more support for the proposition that vol-
ume and volatility share a common order of fractional integration
using our preferred realized volatility estimator, we still nd less
support than Bollerslev and Jubinski (1999). We reject the
hypothesis for 20% of the MMI rms, while they reject for only
8% of the S&P 100 rms. It is unlikely that we would we see a
rejection for one in ve of the MMI rms under circumstances
in which a long-run version of the MDH holds. A rejection rate
of this magnitude points to volume and volatility dynamics that
are too complex to be captured by a model with a common
long-memory component. This nding lends further credibility
to the rejections of the MDH obtained using short-memory spec-
ications, and indicates that the evidence against the MDH is
quite robust.
4.3. Results using daily squared returns
Bollerslev and Jubinski (1999) conduct their analysis using trad-
ing volume and absolute returns. To see how our ndings change
when we use their proxy for volatility, we t the model a third
time using daily squared returns in place of realized volatilities.
7
Table 3 reports the results. As in both of the previous tables, the esti-
mates of d
1
and d
2
support the hypothesis that volume and volatility
display long memory. However, the d
1
estimates now range from
0.07 (3M) to 0.17 (American Express), which is substantially lower
than the estimates reported in Tables 1 and 2, while the magnitude
of the d
2
estimates are relatively unchanged. Consequently, the re-
sults in Table 3 provide strong evidence against the hypothesis of
a common order of fractional integration. We reject the hypothesis
for all of the rms at the 5% level.
The diagnostics in Panel B of Table 3 suggest a possible explana-
tion for the disparity between the different sets of d estimates. The
7
Note that this produces the same estimates of d
1
and d
2
as using absolute returns
because log r
2
t
= 2log [rt [.
1720 J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726
estimated variance of the innovation to logr
2
t
is typically an order
of magnitude larger than the estimated variance of the innovations
to logRV
t
and logV
t
. Since this implies a much more difcult signal-
extraction problem, we should expect to obtain less precise
information about the dynamics of the conditional variance from
the squared returns than from the realized variances. Moreover,
the innovation to logr
2
t
displays more serious departures from nor-
mality than the innovations to logRV
t
and logV
t
. It seems probable
that the performance of the QML estimator would deteriorate un-
der such circumstances.
Another concern is that the standard errors in Table 3 may be
unreliable if the limiting distribution of the QML estimator is a
Table 2
Model tting results using the higher-frequency, bias-corrected (HFBC) estimator of realized volatility.
Firm ^
d
1
^
d
2
se(
^
d
1
) se(
^
d
2
)
^
/
1
^
/
2
^
/
3
^
/
4
t(
^
/
1
) t(
^
/
2
) t(
^
/
3
) t(
^
/
4
)
L
T
LR
Panel A: Long-memory and VAR parameters
AXP 0.39 0.39 0.02 0.02 0.07 0.10 0.03 0.01 2.01 3.09 1.36 0.40 1206.0 0.01
CVX 0.33 0.29 0.02 0.02 0.05 0.02 0.00 0.08 1.55 0.75 0.10 2.14 963.4 2.86
DD 0.37 0.34 0.02 0.02 0.08 0.06 0.02 0.09 2.37 2.01 0.91 2.51 1099.6 1.18
DIS 0.38 0.37 0.02 0.02 0.09 0.00 0.03 0.10 2.65 0.03 1.23 2.66 1172.2 0.38
DOW 0.38 0.32 0.02 0.02 0.09 0.05 0.00 0.02 2.94 1.87 0.01 0.59 1917.9 5.10
*
EK 0.33 0.36 0.02 0.02 0.09 0.02 0.09 0.15 2.46 0.58 3.35 3.55 1733.6 1.19
GE 0.38 0.36 0.02 0.02 0.11 0.05 0.06 0.04 3.36 1.75 2.08 0.98 371.7 1.01
GM 0.33 0.35 0.02 0.02 0.07 0.02 0.03 0.12 2.14 0.66 1.11 3.01 1364.7 0.49
IBM 0.40 0.38 0.02 0.02 0.09 0.03 0.07 0.12 2.50 0.85 2.13 2.86 876.0 1.53
IP 0.37 0.33 0.02 0.02 0.09 0.01 0.01 0.13 2.88 0.31 0.45 3.52 1291.7 2.03
JNJ 0.36 0.32 0.02 0.02 0.10 0.01 0.05 0.12 2.93 0.23 1.87 3.16 997.4 2.56
KO 0.37 0.25 0.02 0.02 0.09 0.05 0.05 0.21 2.77 1.62 1.88 5.37 628.1 22.94
*
MCD 0.33 0.32 0.02 0.02 0.07 0.01 0.00 0.11 2.00 0.19 0.10 2.90 1254.6 0.09
MMM 0.33 0.33 0.02 0.02 0.08 0.01 0.02 0.11 2.24 0.32 0.92 2.90 1385.2 0.00
MO 0.35 0.31 0.02 0.02 0.08 0.01 0.11 0.22 2.15 0.41 3.24 5.23 1422.9 2.93
MRK 0.38 0.30 0.02 0.02 0.08 0.03 0.01 0.18 2.43 0.82 0.26 4.41 768.6 10.39
*
PG 0.36 0.33 0.02 0.02 0.09 0.02 0.02 0.10 2.75 0.59 0.92 2.60 1014.2 1.85
S 0.36 0.39 0.02 0.02 0.07 0.05 0.01 0.06 2.08 1.65 0.63 1.55 1814.3 1.67
T 0.37 0.40 0.02 0.02 0.05 0.05 0.01 0.09 1.53 1.46 0.52 2.34 1175.7 1.61
XOM 0.40 0.27 0.02 0.02 0.12 0.07 0.01 0.10 3.84 2.82 0.38 2.71 309.3 28.82
*
Innovations to logRV
t
Innovations to logV
t
Firm ^
b
11
^
b
12
^
b
13
^
b
21
^
b
22
^
b
23

X
11

X
12

X
22
CS CK Q
5
Q
10
CS CK Q
5
Q
10
Panel B: Trend parameters and innovation diagnostics
AXP 3.05 1.11 0.76 1.27 0.41 0.15 0.09 0.06 0.13 0.29 1.22 4.55 7.81 0.13 1.64 3.59 10.15
CVX 2.55 2.32 2.08 0.09 0.19 1.48 0.09 0.04 0.10 0.17 0.80 3.46 9.41 0.40 3.18 3.45 10.68
DD 2.63 2.33 1.84 0.40 1.38 0.73 0.09 0.05 0.11 0.33 1.11 8.20 15.53 0.41 2.35 1.84 19.61
*
DIS 3.00 0.50 0.10 1.15 0.71 0.20 0.09 0.06 0.12 0.46 1.15 9.27 16.21 0.79 2.98 7.91 14.16
DOW 2.71 0.91 0.15 0.45 0.76 0.13 0.12 0.06 0.14 0.28 0.98 9.13 13.99 0.47 3.08 4.09 6.87
EK 2.77 1.10 0.45 0.07 1.16 2.26 0.11 0.09 0.18 0.61 1.77 19.16
**
33.08
**
1.10 4.50 13.05
*
17.10
GE 2.51 2.17 1.40 2.41 0.14 0.57 0.08 0.05 0.08 0.23 0.55 0.77 12.12 0.34 1.83 1.53 15.27
GM 3.13 0.46 0.15 0.83 0.96 1.79 0.10 0.06 0.13 0.20 0.45 10.14 20.18
*
0.30 1.57 4.99 12.84
IBM 3.01 1.43 1.23 2.18 0.36 0.74 0.10 0.07 0.11 0.30 0.50 8.43 10.47 0.56 2.27 6.31 21.14
*
IP 2.61 2.82 2.30 0.41 1.80 0.31 0.08 0.05 0.13 0.18 0.72 1.92 14.79 0.23 0.64 0.33 12.84
JNJ 3.05 0.22 0.20 1.53 0.76 1.44 0.10 0.05 0.10 0.28 0.81 8.32 22.23
*
0.41 1.91 5.98 19.60
*
KO 2.73 1.60 1.29 0.89 0.59 0.11 0.08 0.05 0.10 0.26 0.76 2.91 7.36 0.38 1.68 4.24 8.18
MCD 2.80 1.07 0.54 0.90 0.72 0.11 0.10 0.06 0.13 0.40 1.04 3.20 6.31 0.55 1.77 4.12 11.81
MMM 2.35 2.99 2.55 0.08 1.25 0.37 0.10 0.06 0.13 0.24 1.02 4.46 9.88 0.70 3.21 1.01 6.56
MO 2.81 1.79 1.49 1.84 0.13 0.26 0.12 0.08 0.13 0.79 2.58 2.86 7.23 0.96 3.69 2.63 6.94
MRK 3.05 0.46 0.25 1.69 0.64 0.82 0.09 0.06 0.10 0.42 1.51 14.47
*
16.03 0.58 2.53 12.09
*
20.21
*
PG 2.70 2.18 2.19 0.26 1.30 0.31 0.09 0.05 0.11 0.64 3.62 7.59 15.38 0.76 6.42 3.96 21.20
*
S 2.94 1.40 0.81 0.14 0.25 1.13 0.11 0.07 0.15 0.27 0.84 17.65
**
22.76
*
0.62 2.51 5.38 12.43
T 2.63 1.97 0.96 0.94 3.07 0.64 0.09 0.06 0.12 0.60 1.98 4.69 13.70 0.86 3.86 2.34 17.04
XOM 2.41 2.30 1.79 1.22 1.63 0.40 0.07 0.04 0.08 0.27 0.55 4.90 13.13 0.15 2.08 1.95 10.34
The table reports the results of tting a bivariate linear regression model with deterministic regressors and long-memory errors to the daily realized variance and daily
trading volume for the MMI rms. The model is
y
t
= bx
t
e
t
;
U(L)D(L)e
t
= g
t
;
where y
t
= (logRV
t
,logV
t
)
/
with RV
t
based on a 30-s sampling frequency and 30-min window length, x
t
= (1, t/T, t
2
/T
2
)
/
, U(L) = I UL, D(L) = diag((1 L)
d1
; (1 L)
d2
), and g
t
-
NID (0,X) with g
t
= 0 "t 6 0. We estimate d and / = vec(U) by Gaussian quasi-maximum likelihood, which is equivalent to minimizing [

X(d; /)[ where

X(d; /) =

T
t=1
[U(L)D(L)(y
t

^
bx
t
)[[U(L)D(L)(y
t

^
bx
t
)[
/
with
^
b =

T
t=1
y
t
x
/
t
_ _

T
t=1
x
t
x
/
t
_ _
1
:
Panel A reports the elements of
^
d and their asymptotic standard errors, the elements of
^
/ and their asymptotic t-ratios, the quasi-log likelihood value (LT ), and the quasi-like-
lihood-ratio statistic (LR) for H
0
: d
1
= d
2
. Panel B reports the elements of
^
b and summary statistics for the tted innovation sequence g
t

T
t=1
implied by the parameter estimates.
The summary statistics are the unique elements of

X, the coefcients of skewness (CS) and excess kurtosis (CK) for the tted innovations, and the Leung-Box statistics com-
puted from the rst ve and 10 sample autocorrelations of the tted innovations (Q
5
and Q
10
). An asterisk indicates a likelihood ratio or Leung-Box statistic that is statistically
signicant at the 5% level. The realized variances are expressed as squared annualized percentage rates (assuming 252 trading days per year). Trading volume is measured in
millions of shares. The sample period is January 1, 1993 to December 31, 2003.
J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726 1721
poor approximation to its nite-sample distribution. Nielsen
(2004b) provides some preliminary evidence on the nite sample
performance of the likelihood ratio test for models with Gaussian
innovations. He nds that the test has good size in samples as
small as 100 observations, and that the rejection frequencies under
the alternative are close to the asymptotic local power implied by
the asymptotic distribution theory. Since the innovations for the
model in Table 2 display only mild departures from normality,
the evidence suggests that the standard errors reported in Table
2 are reliable. This may not be true, however, for the standard
Table 3
Model tting results using daily squared returns.
Firm ^
d
1
^
d
2
se(
^
d
1
) se(
^
d
2
)
^
/
1
^
/
2
^
/
3
^
/
4
t(
^
/
1
) t(
^
/
2
) t(
^
/
3
) t(
^
/
4
)
L
T
LR
Panel A: Long-memory and VAR parameters
AXP 0.17 0.37 0.02 0.02 0.13 0.03 0.11 0.04 4.51 4.27 1.71 1.11 5083.4 50.16
*
CVX 0.10 0.28 0.02 0.02 0.08 0.02 0.10 0.07 2.84 3.81 1.41 2.29 4864.9 38.94
*
DD 0.12 0.29 0.02 0.03 0.04 0.03 0.10 0.15 1.32 4.51 1.42 4.41 4946.4 25.98
*
DIS 0.11 0.31 0.02 0.02 0.07 0.02 0.22 0.14 2.31 2.43 2.92 4.00 5120.5 50.36
*
DOW 0.15 0.31 0.02 0.02 0.11 0.04 0.11 0.03 3.72 5.24 1.81 0.93 5339.5 35.50
*
EK 0.10 0.30 0.02 0.03 0.08 0.02 0.20 0.20 2.47 1.93 3.02 5.44 5640.8 47.74
*
GE 0.12 0.32 0.02 0.02 0.10 0.02 0.14 0.08 3.33 4.21 1.75 2.48 4399.3 52.63
*
GM 0.14 0.31 0.02 0.02 0.07 0.03 0.03 0.12 2.46 4.03 0.42 3.65 5117.7 32.50
*
IBM 0.15 0.34 0.02 0.02 0.15 0.03 0.25 0.14 5.22 4.74 3.35 4.16 4951.5 49.15
*
IP 0.12 0.31 0.02 0.03 0.06 0.01 0.13 0.14 2.16 2.00 2.01 4.12 5104.2 42.75
*
JNJ 0.13 0.31 0.02 0.02 0.07 0.02 0.01 0.11 2.31 3.95 0.13 3.34 4807.2 39.08
*
KO 0.12 0.25 0.02 0.03 0.10 0.03 0.27 0.21 3.46 3.97 3.56 5.79 4684.1 16.98
*
MCD 0.11 0.30 0.02 0.02 0.07 0.02 0.07 0.13 2.27 2.37 0.96 3.67 5162.0 33.52
*
MMM 0.07 0.31 0.02 0.02 0.05 0.02 0.13 0.12 1.62 3.17 1.81 3.55 5302.1 66.02
*
MO 0.15 0.27 0.02 0.03 0.11 0.01 0.36 0.26 3.83 1.69 4.96 6.88 4944.4 22.41
*
MRK 0.15 0.27 0.02 0.03 0.09 0.02 0.16 0.20 2.92 3.62 2.12 5.56 4790.8 17.05
*
PG 0.11 0.32 0.02 0.02 0.10 0.02 0.19 0.10 3.59 3.03 2.53 2.99 5001.3 52.67
*
S 0.12 0.36 0.02 0.02 0.07 0.01 0.04 0.10 2.26 1.09 0.53 2.91 5456.8 68.34
*
T 0.16 0.35 0.02 0.03 0.07 0.01 0.07 0.15 2.44 2.37 0.80 4.40 5517.9 33.18
*
XOM 0.14 0.24 0.02 0.02 0.11 0.02 0.12 0.13 3.95 4.37 1.44 4.01 4331.2 13.53
*
Innovations to log r
2
t
Innovations to logV
t
Firm ^
b
11
^
b
12
^
b
13
^
b
21
^
b
22
^
b
23

X
11

X
12

X
22
CS CK Q
5
Q
10
CS CK Q
5
Q
10
Panel B: Trend parameters and innovation diagnostics
AXP 2.35 1.82 1.34 1.27 0.41 0.15 1.14 0.11 0.13 1.15 2.24 5.33 6.38 0.13 1.69 3.51 10.45
CVX 2.04 2.05 1.85 0.09 0.19 1.48 1.23 0.09 0.10 1.33 3.88 3.71 10.56 0.39 3.13 2.88 10.40
DD 2.04 2.65 2.19 0.40 1.38 0.73 1.16 0.10 0.11 1.00 1.13 5.75 8.00 0.40 2.34 1.71 20.72
*
DIS 2.25 1.33 0.58 1.15 0.71 0.20 1.29 0.13 0.12 1.42 4.71 15.21
**
20.24
*
0.80 3.02 8.64 14.83
DOW 2.13 0.97 0.15 0.45 0.76 0.13 1.23 0.12 0.14 1.22 2.81 8.48 16.21 0.47 3.16 4.12 6.83
EK 2.29 0.91 0.51 0.07 1.16 2.26 1.35 0.19 0.18 1.10 1.83 0.64 6.41 1.08 4.43 9.91 14.33
GE 1.88 2.66 1.85 2.41 0.14 0.57 1.12 0.09 0.08 1.12 1.80 15.44
**
26.14
**
0.34 1.91 5.75 19.68
*
GM 2.54 0.56 0.22 0.83 0.96 1.79 1.21 0.13 0.13 1.07 1.44 7.24 14.49 0.29 1.66 3.39 10.32
IBM 2.40 1.50 1.22 2.18 0.36 0.74 1.24 0.13 0.11 1.19 3.10 13.42
*
14.94 0.59 2.35 3.71 17.98
IP 2.06 2.83 2.30 0.41 1.80 0.31 1.15 0.11 0.13 1.14 1.77 3.28 5.72 0.23 0.63 0.87 12.03
JNJ 2.42 0.64 0.68 1.53 0.76 1.44 1.25 0.11 0.10 1.28 2.94 10.26 25.37
**
0.39 1.89 4.61 20.40
*
KO 2.06 2.51 2.33 0.89 0.59 0.11 1.14 0.11 0.10 1.04 1.22 12.90
*
22.52
*
0.37 1.76 3.85 7.52
MCD 2.17 1.40 0.92 0.90 0.72 0.11 1.25 0.12 0.13 1.09 1.41 5.83 10.37 0.54 1.75 3.86 11.64
MMM 1.81 2.64 2.18 0.08 1.25 0.37 1.37 0.12 0.12 1.27 2.98 4.19 9.64 0.69 3.16 0.50 5.92
MO 2.17 2.38 2.13 1.84 0.13 0.26 1.13 0.16 0.13 0.84 1.35 9.98 19.09
*
0.95 3.69 4.97 10.23
MRK 2.42 1.02 0.86 1.69 0.64 0.82 1.22 0.11 0.10 1.18 2.17 4.33 11.39 0.58 2.57 12.58
*
21.36
*
PG 2.07 2.61 2.68 0.26 1.30 0.31 1.27 0.12 0.11 1.33 3.43 2.67 11.28 0.76 6.43 4.44 21.26
*
S 2.52 0.99 0.65 0.14 0.25 1.13 1.30 0.15 0.15 1.12 2.15 7.24 12.13 0.62 2.56 4.15 12.67
T 1.75 2.73 1.46 0.94 3.07 0.64 1.66 0.14 0.12 1.44 2.39 2.71 4.52 0.85 3.85 4.05 16.67
XOM 1.83 2.40 1.84 1.22 1.63 0.40 1.12 0.08 0.08 1.08 1.61 6.17 12.22 0.15 2.04 1.07 10.74
The table reports the results of tting a bivariate linear regression model with deterministic regressors and long-memory errors to the daily squared demeaned returns and
daily trading volume for the MMI rms. The model is
y
t
= bxt et ;
U(L)D(L)e
t
= g
t
;
where y
t
= log r
2
t
; log Vt
_ _
/
; xt = (1; t=T; t
2
=T
2
)
/
; U(L) = I UL; D(L) = diag((1 L)
d1
; (1 L)
d2
), and g
t
- NID (0,X) with g
t
= 0 "t 6 0. We estimate d and / = vec(U) by Gaussian
quasi-maximum likelihood, which is equivalent to minimizing [

X(d; /)[ where

X(d; /) =

T
t=1
[U(L)D(L)(y
t

^
bxt )[[U(L)D(L)(y
t

^
bxt )[
/
with
^
b =

T
t=1
y
t
x
/
t
_ _

T
t=1
xt x
/
t
_ _
1
:
Panel A reports the elements of
^
d and their asymptotic standard errors, the elements of
^
/ and their asymptotic t-ratios, the quasi-log likelihood value (LT ), and the quasi-like-
lihood-ratio statistic (LR) for H
0
: d
1
= d
2
. Panel B reports the elements of
^
b and summary statistics for the tted innovation sequence g
t

T
t=1
implied by the parameter estimates.
The summary statistics are the unique elements of

X, the coefcients of skewness (CS) and excess kurtosis (CK) for the tted innovations, and the Leung-Box statistics com-
puted from the rst ve and 10 sample autocorrelations of the tted innovations (Q
5
and Q
10
). An asterisk indicates a likelihood ratio or Leung-Box statistic that is statistically
signicant at the 5% level. The returns are expressed as annualized percentage rates and multiplied by

252
_
. Trading volume is measured in millions of shares. The sample
period is January 1, 1993 to December 31, 2003.
1722 J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726
errors in Table 3. Studies that t short-memory stochastic volatility
models to log squared returns via Gaussian QML report that this
produces highly inefcient inferences (see, e.g., Alizadeh et al.,
2002). This probably holds for long-memory models as well.
As an additional robustness test, we investigate whether tting
the model to squared returns rather than log squared returns alters
the basic message of the analysis. This test is motivated by a well-
known drawback of the logarithmic transformation: it translates
any squared return sufciently close to zero into a negative outlier.
In general, the parameter estimates (not reported) are similar to
those obtained using log squared returns, and again the estimates
of d
1
are substantially smaller than those in Table 2. As an aside, we
should point out that the assumption of i.i.d. innovations is almost
certain to be violated for the squared-return specication.
8
Although this would be problematic from the standpoint of conduct-
ing formal hypothesis tests, it is not a signicant concern for the lim-
ited purposes of the application considered here.
5. Model forecasting performance
We now investigate the volatility forecasting performance of
our long memory models. We are specically interested in two is-
sues. First, we want to measure the incremental forecasting perfor-
mance associated with using realized variances instead of daily
squared returns to t the model. Second, we want to measure
the incremental performance of volatility estimates obtained by
conditioning on contemporaneous trading volume. These esti-
mates are not forecasts of volatility in the usual sense. Like the
log realized variances, they are estimates of the unobserved log
variance based on both past and contemporaneous information.
We evaluate these estimates to gain some preliminary insights into
the potential usefulness of volume data in the econometric analy-
sis of realized volatility.
Suppose, for example, that realized volatility is modeled as the
true unobserved volatility plus measurement error. In this case, we
can obtain more precise volatility estimates by employing a lter-
ing technique. Barndorff-Nielsen and Shephard (2002) propose a
simple approach based on the Kalman lter. A similar strategy
could be employed in a long-memory setting. For instance, we
might model the log realized variances as a fractionally-integrated
process subject to additive measurement error, and then estimate
a joint system that includes log trading volume via linear state-
space methods.
9
This would allow us to exploit the contemporane-
ous correlation between volatility and volume shocks to extract
more efcient smoothed estimates of the latent volatility process.
5.1. Methodology
Our methodology is based on the Andersen et al. (2003) regres-
sion approach for comparing the performance of volatility fore-
casts produced by different models. Let h
t[t1
denote the forecast
of the log variance for day t given the (model specic) information
observed through day t 1. We use an OLS regression of the form
log RV
t
= a bh
t[t1
u
t
; (18)
to assess how much of the variation in the log volatility is captured
by a given model. Although the regression R-squared is biased to-
wards zero because the variance of logRV
t
is greater than the vari-
ance of the true log volatility (see Andersen et al. (2005) for details),
this is irrelevant for evaluating relative forecasting performance be-
cause the ratio of the R-squared values produced by different mod-
els is bias free.
We use this approach to compare the volatility forecasts from
our long memory model t to the HFBC estimates of realized vola-
tility to those from the model t to daily squared returns. For the
realized volatility specication, the value of h
t[t1
is simply the t-
ted conditional mean of logRV
t
produced by the model. For the
squared return specication, we adjust the tted values from the
model to account for the difference between the unconditional
mean of log r
2
t
and the unconditional mean of the log variance.
The standard way to do this is to assume that we can express the
squared demeaned return as r
2
t
= h
t
z
2
t
, where h
t
is the unobserved
variance and z
t
is Gaussian white noise. Under this approach, it fol-
lows that E[log h
t
[ = E[log r
2
t
[ 1:27. Thus, we add 1.27 to each of
the tted log variances to obtain the log variance forecasts.
We use a similar strategy to investigate the relation between
the unexpected components of trading volume and volatility. Sup-
pose that
log RV
t
= h
t[t1
g
1t
; (19)
log V
t
= m
t[t1
g
2t
; (20)
where h
t[t1
and m
t[t1
are the conditional means implied by our
long memory model and g
t
= (g
1t
, g
2t
)
/
is distributed i.i.d. N(0, X).
Now let 1
t1
= V
1
; RV
1
; . . . ; V
t1
; RV
t1
denote the day t 1 infor-
mation set. With Gaussian innovations, we have
log RV
t
[1
t1
; V
t
- N(h
t[t
; X
11[2
); (21)
where h
t[t
= h
t[t1
X
12
X
1
22
(log V
t
m
t[t1
) and X
11[2
= X
11

X
2
12
X
1
22
.
10
Thus, we can use a regression of the form
log RV
t
= a bh
t[t
u
t
; (22)
to investigate how conditioning on contemporaneous trading vol-
ume improves our volatility estimates. In particular, we can use
the ratio of the R-squared values for the regressions in Eqs. (18)
and (22) as a measure of the information content of unexpected vol-
ume. If we nd that this ratio is less than one, then we can conclude
that a strategy of using volume data to improve volatility estimates
warrants further investigation.
5.2. Empirical results
Table 4 summarizes the results of the volatility regressions. Pa-
nel A focuses on the realized volatility specication of the long
memory model. The rst ve columns report the intercept and
slope estimates, as well as their standard errors, and the R-squared
estimates for the regression in Eq. (18), the next ve columns re-
port the same quantities for the regression in Eq. (22), and the last
column reports the ratio of the R-squared values for the two
regressions. Panel B reports the same set of results for the squared
returns specication of the long memory model.
The results in Panel A show that the one-step-ahead log vari-
ance forecasts produced by the realized volatility specication ex-
hibit little evidence of bias. The intercept estimates are close to
zero and the slope estimates are close to one. This is expected gi-
ven the nature of the rst-order conditions imposed during the
QML estimation procedure. Moreover, the R-squared values for
the forecasting regression range from 31% for General Motors to
58% for AT&T, which indicates that the long memory model pro-
duces forecasts that capture a sizeable fraction of the time-series
8
To see this, write r
2
t
= ht mt , where ht = E[r
2
t
[1
t1
[; mt = r
2
t
=ht , and 1
t1
denotes the
date t 1 information set. Since log r
2
t
= log ht log mt , it is reasonable to assume that
a log linear model has i.i.d. innovations. Note, however, that r
2
t
= ht zt , where
z
t
= h
t
(m
t
1). Even if m
t
is i.i.d., z
t
is not unless h
t
is constant.
9
Chan and Palma (1998) show that these methods perform well for univariate
long-memory specications.
10
More generally, for models with non-Gaussian innovations, h
t[t
represents the
minimum-mean-square linear estimate of logRV
t
given V
t
after we rst condition on
1
t1
.
J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726 1723
variation in the log realized variances. This is consistent with the
evidence presented in Andersen et al. (2003).
The results in Panel B indicate that the one-step-ahead log var-
iance forecasts produced by the squared returns specication per-
form more poorly. The intercept and slope estimates suggest that
the forecasts are biased, and the R-squared values now range from
19% for General Motors and Johnson & Johnson to 48% for AT&T.
Comparing the R-squared values in Panels A and B, there is a reduc-
tion in explanatory power for every rm, and the reduction in R-
squared is typically on the order of 0.1 or 0.2. These comparisons
provide an indication of the value of using high-frequency return
data to construct volatility forecasts.
Now consider the impact of updating the volatility forecasts
from the realized volatility specication (Panel A) using the con-
temporaneous realization of trading volume. The forecasts still
show little evidence of bias, but there is a marked increase in the
explanatory power for most rms. The regression R-squared values
range from 50% for Johnson & Johnson to 71% for General Electric
and AT&T. The ratio of the R-squared for the regression in Eq.
(18) to the R-squared for the regression in Eq. (22) is, on average,
0.73. Since this ratio is unaffected by measurement error, it should
provide an accurate reection of the extent to which observing
contemporaneous volume resolves the uncertainty about volatility.
Another way to look at this is to compute the reduction in the MSE
obtained by updating the log variance forecasts to account for the
observed trading volume. On average, the reduction is 30%. This
nding indicates that volume contains substantial information
which can be used to increase the precision of contemporaneous
volatility estimates.
We observe a similar increase in explanatory power by updat-
ing the forecasts from the squared returns specication (Panel B)
with contemporaneous volume. In this case, the average increase
in the R-squared value is 0.12, and the average ratio of R-squared
values for the regressions in Eqs. (18) and (22) is 0.72. Interest-
ingly, for most rms, the R-squared for the updated forecasts is
comparable to the R-squared for the one-step-ahead forecasts from
the realized volatility specication. Thus, the realization of con-
temporaneous trading volume is quite informative about unpre-
Table 4
Realized volatility regressions.
Firm One-step ahead After observing V
t
^ a ^
b
se(^ a)
se(
^
b)
R
2
^ a ^
b
se(^ a)
se(
^
b)
R
2
Ratio
Panel A: Log variance estimates based on realized volatilities
AXP 0.20 1.03 0.14 0.02 0.48 0.11 1.02 0.11 0.02 0.64 0.75
CVX 0.03 1.00 0.15 0.02 0.39 0.02 1.00 0.11 0.02 0.52 0.75
DD 0.10 1.02 0.13 0.02 0.49 0.11 1.02 0.10 0.02 0.63 0.78
DIS 0.00 1.00 0.13 0.02 0.49 0.05 1.01 0.10 0.01 0.64 0.76
DOW 0.07 1.01 0.11 0.02 0.53 0.05 1.01 0.10 0.02 0.63 0.84
EK 0.16 0.97 0.15 0.02 0.39 0.02 1.00 0.09 0.01 0.64 0.61
GE 0.07 1.01 0.11 0.02 0.56 0.04 1.01 0.08 0.01 0.71 0.79
GM 0.06 1.01 0.19 0.03 0.31 0.07 1.01 0.12 0.02 0.52 0.59
IBM 0.01 1.00 0.15 0.02 0.43 0.09 1.01 0.10 0.01 0.68 0.64
IP 0.10 1.02 0.12 0.02 0.53 0.06 1.01 0.10 0.02 0.63 0.85
JNJ 0.14 1.02 0.19 0.03 0.32 0.06 1.01 0.13 0.02 0.50 0.64
KO 0.19 1.03 0.13 0.02 0.48 0.08 1.01 0.10 0.02 0.63 0.76
MCD 0.05 0.99 0.16 0.03 0.36 0.03 1.00 0.11 0.02 0.55 0.66
MMM 0.10 1.02 0.13 0.02 0.46 0.03 1.01 0.09 0.02 0.62 0.75
MO 0.13 1.02 0.15 0.02 0.42 0.07 1.01 0.10 0.02 0.65 0.65
MRK 0.03 0.99 0.17 0.03 0.34 0.09 1.01 0.11 0.02 0.58 0.59
PG 0.24 1.04 0.13 0.02 0.46 0.10 1.02 0.11 0.02 0.61 0.75
S 0.02 1.00 0.16 0.02 0.41 0.04 1.01 0.12 0.02 0.57 0.72
T 0.04 0.99 0.10 0.02 0.58 0.00 1.00 0.08 0.01 0.71 0.82
XOM 0.10 1.02 0.11 0.02 0.54 0.06 1.01 0.08 0.01 0.65 0.83
Panel B: Log variance estimates based on squared demeaned daily returns
AXP 0.71 0.87 0.18 0.03 0.30 1.00 0.70 0.13 0.02 0.46 0.65
CVX 0.58 1.08 0.21 0.03 0.27 1.29 0.64 0.14 0.02 0.33 0.81
DD 0.06 0.96 0.16 0.02 0.35 0.57 0.74 0.12 0.02 0.46 0.78
DIS 0.70 0.87 0.17 0.03 0.32 1.83 0.59 0.11 0.01 0.41 0.76
DOW 0.23 1.01 0.15 0.02 0.40 0.23 0.78 0.14 0.02 0.46 0.87
EK 1.45 1.21 0.24 0.04 0.28 1.97 0.57 0.10 0.01 0.42 0.65
GE 0.00 0.97 0.13 0.02 0.45 0.44 0.75 0.11 0.01 0.57 0.79
GM 1.07 0.82 0.22 0.03 0.19 1.95 0.58 0.12 0.01 0.37 0.50
IBM 0.11 0.96 0.20 0.03 0.29 1.59 0.63 0.11 0.01 0.47 0.60
IP 0.28 1.02 0.15 0.02 0.43 1.09 0.68 0.11 0.01 0.45 0.95
JNJ 0.37 0.92 0.24 0.04 0.19 2.13 0.54 0.12 0.02 0.32 0.57
KO 0.88 0.84 0.16 0.03 0.29 1.89 0.57 0.10 0.01 0.41 0.71
MCD 0.27 0.94 0.19 0.03 0.25 1.54 0.62 0.12 0.02 0.39 0.64
MMM 0.93 1.14 0.19 0.03 0.35 0.58 0.74 0.12 0.02 0.44 0.78
MO 0.61 0.88 0.17 0.03 0.31 1.40 0.64 0.10 0.01 0.52 0.60
MRK 0.74 0.86 0.21 0.03 0.22 1.68 0.60 0.11 0.01 0.42 0.53
PG 0.45 0.91 0.17 0.03 0.29 1.54 0.61 0.12 0.02 0.41 0.71
S 2.51 1.35 0.27 0.04 0.31 1.64 0.63 0.14 0.02 0.36 0.87
T 1.82 0.73 0.09 0.01 0.48 1.83 0.61 0.07 0.01 0.58 0.83
XOM 0.08 0.96 0.14 0.02 0.38 0.77 0.70 0.11 0.02 0.45 0.85
The table summarizes the results of tting a linear regression of logRV
t
, where RV
t
is constructed using our HFBC estimator of realized volatility, on the log variance estimates
obtained from our tted long memory models. Panel A reports the results for the long memory model t based on realized volatilities and Panel B reports the results for the
long memory model t based on daily squared returns. For each model we consider two realized volatility regressions. The explanatory variable in the rst regression is a one-
step-ahead forecast of the log variance. The explanatory variable in the second regression is an estimate of the log variance obtained by updating the one-step-ahead forecast
to reect the information revealed by the realization of contemporaneous trading volume. For each regression we report the estimated intercept (^ a) and estimated slope (
^
b)
along with heteroscedasticity-robust standard errors and the sample R-squared statistic (R
2
). The nal column reports the ratio of the two R-squared values.
1724 J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726
dictable shocks to return volatility, even at the daily sampling fre-
quency. This suggests that daily trading volume can be used in con-
junction with daily returns to obtain more precise estimates of
daily return volatility for situations in which the intradaily returns
data needed to construct realized volatilities are unavailable.
6. Concluding remarks
Recent empirical work suggests that return volatility and trad-
ing volume exhibit similar long-run dynamics. Building on this re-
search, we examine the performance of a long-memory model for
return volatility and trading volume that permits both common
long-run dependencies and exible short-run interactions. In par-
ticular, we assume that the logarithms of volatility and volume
are described by a trend-stationary fractionally-integrated process
that displays VAR(p) dynamics after fractional differencing. Since
our approach allows the trend, short-memory, and long-memory
components to compete on an equal footing, it should be well sui-
ted to analyzing the relative importance of each component in
explaining the observed characteristics of the data.
When we t the model to daily realized volatilities and daily
trading volumes for the 20 MMI rms, the results paint an interest-
ing picture of the role of long memory in volatility and volume
dynamics. Both the log realized volatilities and log trading volumes
appear to be fractionally integrated with an order of integration of
about 0.35. However, the evidence on whether the two series share
a common order of fractional integration is mixed, with one fth of
the rms producing rejections of this hypothesis at the 5% signi-
cance level. This is a considerably higher rejection rate than that
reported previously for the S&P 100 rms. Although drawing den-
itive conclusions about this issue would require a better under-
standing of the nite sample performance of the tests used in
this paper and in previous work, our ndings suggest that it is pre-
mature to treat common long-run dependencies as a stylized fact
of the volumevolatility literature.
With respect to short-run dynamics, we nd that they play only
a minor role in explaining the serial dependence in the data. The
main short-run phenomenon is a strong contemporaneous correla-
tion between the log volatility and log volume innovations. In light
of this correlation, it may be possible to use trading volume data to
improve estimates of daily return volatility for cases in which the
high-frequency return data needed to construct realized volatilities
are unavailable. On the other hand, it seems that there is little to be
gained from modeling short-run dynamics if the objective is to im-
prove short-range volatility forecasts. Fractional differencing by it-
self is sufcient to remove most of the serial dependence in the
realized volatilities.
There are a couple of issues that hold potential for future
research in this area. First, it would be useful to document the
nite-sample size and power of asymptotic tests in the long-
memory setting. This would facilitate comparison between studies
that use different estimation techniques, and would help shed light
on why the QML estimates of the memory parameter for volatility
are sensitive to whether we use realized volatilities or squared
returns to t the models. Second, it would be interesting to study
the implications of long memory for long-range volatility forecasts.
If these forecasts can be improved by adopting long-memory spec-
ications, then we can potentially improve the pricing and hedging
of long-term nancial contracts.
Acknowledgements
We thank Scott Baggett for helping us construct the price and
volume datasets from TAQ data and Joel Hasbrouck and James
Weston for providing many useful insights regarding the use of
TAQ data. We are also grateful to seminar participants at the
University of Georgia and the editor (Ike Mathur) and an anony-
mous referee of this journal for providing valuable comments on
earlier drafts.
References
Alizadeh, S., Brandt, M., Diebold, F., 2002. Range-based estimation of stochastic
volatility models. Journal of Finance 57, 10411558.
Andersen, T., 1996. Return volatility and trading volume: an information ow
interpretation of stochastic volatility. Journal of Finance 51, 169204.
Andersen, T., Bollerslev, T., 1997. Intraday periodicity and volatility persistence in
nancial markets. Journal of Empirical Finance 4, 115158.
Andersen, T., Bollerslev, T., Diebold, F., Ebens, H., 2001. The distribution of stock
return volatility. Journal of Financial Economics 61, 4376.
Andersen, T., Bollerslev, T., Diebold, F., Labys, P., 2001. The distribution of realized
exchange rate volatility. Journal of the American Statistical Association 96, 4255.
Andersen, T., Bollerslev, T., Diebold, F., Labys, P., 2003. Modeling and forecasting
realized volatility. Econometrica 71, 579625.
Andersen, T., Bollerslev, T., Meddahi, N., 2005. Correcting the errors: volatility
forecast evaluation using high-frequency data and realized volatilities.
Econometrica 73, 279296.
Baillie, R., Bollerslev, T., Mikkelsen, H., 1996. Fractionally integrated generalized
autoregressive conditional heteroskedasticity. Journal of Econometrics 74, 330.
Bandi, F., Russell, J., 2008. Microstructure noise, realized volatility, and optimal
sampling. Review of Economic Studies 75, 339369.
Barndorff-Nielsen, O., Shephard, N., 2002. Econometric analysis of realized volatility
and its use in estimating stochastic volatility models. Journal of the Royal
Statistical Society Series B 64, 253280.
Bessembinder, H., Seguin, P., 1993. Price volatility, trading volume, and market
depth: evidence from futures markets. Journal of Financial and Quantitative
Analysis 28, 2139.
Bollerslev, T., Jubinski, D., 1999. Equity trading volume and volatility: latent
information arrivals and common long-run dependencies. Journal of Business
and Economic Statistics 17, 921.
Bollerslev, T., Mikkelsen, H., 1996. Modeling and pricing long memory in stock
market volatility. Journal of Econometrics 73, 151184.
Campbell, J., Lettau, M., Malkiel, B., Xu, Y., 2001. Have individual stocks become
more volatile? An empirical exploration of idiosyncratic risk. Journal of Finance
56, 143.
Chan, N., Palma, W., 1998. State space modeling of long-memory processes. Annals
of Statistics 26, 719740.
Chuang, C., Hsieh, P., Lai, H., 2009. Do informed option investors predict stock
returns? Evidence from the Taiwan stock exchange. Journal of Banking and
Finance 33, 757764.
Chuang, C., Hsieh, P., Wang, Y., 2010. Information content of options trading volume
for future volatility: evidence from the Taiwan options market. Journal of
Banking and Finance 33, 174183.
Chuli, H., Martens, M., van Dijk, D., 2010. Asymmetric effects of federal funds target
rate changes on S&P 100 stock returns, volatilities and correlations. Journal of
Banking and Finance 34, 834839.
Darbar, S., Deb, P., 1995. Does trading volume have a unit root? Applied Economics
Letters 2, 144147.
Deo, R., Hurvich, C., 2001. On the log periodogram regression estimator of the
memory parameter in long memory stochastic volatility models. Econometric
Theory 17, 686710.
Ding, Z., Granger, C., Engle, R., 1993. A long memory property of stock market
returns and a new model. Journal of Empirical Finance 1, 83106.
Diebold, F., Inoue, A., 2001. Long memory and regime switching. Journal of
Econometrics 105, 131159.
Fleming, J., Kirby, C., Ostdiek, B., 2006. Stochastic volatility, trading volume, and the
daily ow of information. Journal of Business 79, 15511590.
Gallant, R., Rossi, P., Tauchen, G., 1992. Stock prices and volume. Review of Financial
Studies 5, 199242.
Geweke, J., Porter-Hudak, S., 1983. The estimation and application of long memory
time series models. Journal of Time Series Analysis 4, 221238.
Giraitis, L., Koul, H., 1997. Estimation of the dependence parameter in linear
regression with long-range dependent errors. Stochastic Processes and their
Applications 71, 207224.
Granger, C., Hyung, N., 2004. Occasional structural breaks and long memory with an
application to the S&P 500 absolute stock returns. Journal of Empirical Finance
11, 399421.
Hansen, P., Lunde, A., 2004. An Unbiased Measure of Realized Variance. Working
Paper, Stanford University.
Hansen, P., Lunde, A., 2005. A realized variance for the whole day based on
intermittent high-frequency data. Journal of Financial Econometrics 3, 525554.
Hyung, N., Poon, S., Granger, C., 2006. A Source of Long Memory in Volatility.
Working Paper, University of California, San Diego.
Merton, R., 1980. On estimating the expected return on the market: an exploratory
investigation. Journal of Financial Economics 8, 323361.
Newey, W., West, K., 1987. A simple, positive semi-denite, heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica 55, 703708.
Nielsen, M., 2004a. Efcient likelihood inference in nonstationary univariate
models. Econometric Theory 20, 116146.
J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726 1725
Nielsen, M., 2004b. Efcient inference in multivariate fractionally integrated time
series models. Econometrics Journal 7, 6397.
Tauchen, G., Pitts, M., 1983. The price variabilityvolume relationship on
speculative markets. Econometrica 51, 485505.
Taylor, S., 1986. Modeling Financial Time Series. Wiley, Chichester.
Wright, J., Zhou, H., 2009. Bond risk premia and realized jump risk. Journal of
Banking and Finance 33, 23332345.
1726 J. Fleming, C. Kirby / Journal of Banking & Finance 35 (2011) 17141726