ñòð. 1 
EMPIRICAL
FINANCE
ELSEVIER Journal of Empirical Finance 4 (1997) 115158
Intraday periodicity and volatility persistence in
financial markets
Torben G. Andersen a, Tim Bollerslev b,c,.
a Department of Finance, J.L. Kellogg Graduate School of Management, Northwestern Universin,,
Evanston, IL 60208, USA
b Department of Economics, Rouss Hall, University of Virginia, Charlottesville, VA 22901, USA
c NBER, Cambridge, MA 02138, USA
Abstract
The pervasive intraday periodicity in the return volatility in foreign exchange and equity
markets is shown to have a strong impact on the dynamic properties of high frequency
returns. Only by taking account of this strong intraday periodicity is it possible to uncover
the complex intraday volatility dynamics that exists both within and across different
financial markets. The explicit periodic modeling procedure developed here provides such a
framework and thus sets the stage for a formal integration of standard volatility models with
market microstructure variables to allow for a more comprehensive empirical investigation
of the fundamental determinants behind the volatility clustering phenomenon. Â© 1997
Elsevier Science B.V.
JEL classification." C14; C22; GI4; G15
Keywords: Volatility; Intraday periodicity; Temporal aggregation; ARCH
1. I n t r o d u c t i o n
It is widely documented that return volatility varies systematically over the
trading day and that this pattern is highly correlated with the intraday variation of
trading volume and b i d  a s k spreads. Indeed, these strikingly regular patterns of
market activity measures have provided the impetus for much theoretical work. On
* Corresponding author.
09275398/97/$17.00 Copyright Â© 1997 Elsevier Science B.V. All rights reserved.
PII S09275 398(97)000042
116 T. G. Andersen, T. Boilers'let: / Journal of Empirical Finance 4 (1997) 115158
the other hand, the dynamics of the intraday return volatility process is mostly
ignored in the empirical market microstructure literature. This is quite surprising
given the notion that news arrivals and the resolution of their informational impact
are intimately related to the dynamics of the return volatility process 1. W e
conjecture that the intraday return dynamics is neglected primarily because
standard time series models of volatility have proven inadequate when applied to
high frequency returns data. In fact, previous results reported in the literature are
often contradictory and generally defy theoretical predictions. Consequently, there
is no well established paradigm for intraday volatility modeling, and at present its
inclusion in market microstructure research is tenuous.
In this paper we demonstrate that the difficulties encountered by standard
volatility models arise largely from the aforementioned systematic patterns in
average volatility across the trading day. W e further show how practical estimation
and extraction of the intraday periodic component of return volatility is both
feasible and indispensable for meaningful intraday dynamic analysis. Particular
attention is paid to the differing impact of the periodic pattern on the dynamic
return features at the various intraday frequencies. To illustrate the range of
applicability of the developed procedures, the analysis is conducted in parallel for
two different asset classes traded under widely different market structures, namely
the overthecounter foreign exchange interbank market and an organized ex
change for futures equity index contracts. Moreover, to bring out the distinct
character of the intraday returns process, the findings are contrasted to the
corresponding features of interdaily returns series for the identical assets.
The empirical evidence on the properties of average intraday stock returns dates
back to, at least, W o o d et al. (1985) and Harris (1986a) who document the
existence of a distinct Ushaped pattern in return volatility over the trading day i.e.
volatility is high at the open and close of trading and low in the middle of the day.
The existence of equally pronounced intraday patterns in foreign exchange mar
kets has been demonstrated by Miiller et al. (1990) and Baillie and Bollerslev
(1991) 2
Meanwhile, a separate time series oriented literature has modeled the dynamics
of the intraday return volatility directly, building on the A R C H methodology of
1For example, theoretical work stress issues such as the process of price discovery, the optimal
timing of trades designed to limit price impact, the differing price response to public versus private
information, the clustering of discretionary liquidity trading and the associated increase in market depth
when private information is shortlived and the particular market dynamics associated with periodic
market openings and closures.
2 Empirical work continues to refine and classify the regularities of high frequency returns in this
dimension. Recent studies include Barclay et al. (1990) and Harvey and Huang (1991) on return
variances over trading versus nontrading periods, Lockwood and Linn (1990) on overnight and
intraday return volatility and Ederington and Lee (1993) on the impact of macroeconomic announce
ments on inter and intraday return volatility.
117
T.G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158
Engle (1982). Most of these studies fall into one of three categories. Firstly, some
authors investigate the interrelation between returns in geographically separated
financial markets that trade sequentially, with a focus on the transmission of
information as measured by the degree of spillover in the mean returns a n d / o r
volatility from one market to the next 3. A second strand of this literature is
concerned with the l e a d  l a g relations between two or more markets that trade
simultaneously 4. Finally, a third group of papers explores the role of information
flow and other microstructure variables as determinants of intraday return volatil
ity 5
Direct comparison of these intraday volatility studies is complicated by the
different sampling frequencies employed. Nonetheless, as noted by Ghose and
Kroner (1994) and Guillaume (1994), the results regarding the implied degree of
volatility persistence appear puzzling and in stark conflict with the aggregation
results for A R C H models developed by Nelson (1990, 1992), Drost and Nijman
(1993) and Drost and Werker (1996). One potential explanation is that these
theoretical predictions about the relationship between parameter estimates at
different sampling frequencies do not generally apply in the face of strong intraday
periodicity, a fact that has gone largely unnoticed. The most comprehensive prior
attempt at direct modeling of this intraday heteroskedastic pattern in returns is
provided by a series of papers by the research group at Olsen and Associates on
the foreign exchange market e.g. Miiller et al. (1990, 1993) and Dacorogna et al.
(1993). They apply time invariant polynomial approximations to the activity in the
distinct geographical regions of the market over the 24hour trading cycle 6
Although this might be a reasonable assumption for the foreign exchange market,
we propose an alternative and more general methodology that allows the shape of
the periodic pattern to also depend on the current overall level of return volatility.
This feature makes the procedure readily applicable to the analysis of high
frequency financial data in general and turns out to be essential for our investiga
tion of the stock market. While our approach accounts for the pronounced intraday
patterns, we explicitly do not make any attempts to correct for the lower frequency
interdaily patterns that also exist e.g. dayoftheweek and holiday effects which
3 Examples of early contributions are Engle et al, (1990) for foreign exchange markets and Hamao et
al. (1990) for various national equity index returns.
4 See for example Baillie and Bollerslev (1991) and Chan et al. (1991).
5 This literature is exemplified by Bollerslev and Domowitz (1993), Locke and Sayers (1993), Laux
and Ng (1993), Foster and Viswanathan (1995) and Goodhart et al. (1993). This research is partially
motivated by an attempt to identify the economic origins of the volatility clustering phenomenon as
motivated by the mixture of distributions hypothesis; see for example Clark (1973), Tauchen and Pitts
(1983), Harris (1986b, 1987), Gallant et al. (1991), Ross (1989) and Andersen (1994, 1996).
6 One may note that the devolatilization procedure proposed by Zhou (1992) implicitly adjusts for
the intraday periodicity in the adaptive calculation of the volatility increments from tickbytick
observations. Along similar lines, the notion of time deformation in modeling time varying volatility in
financial markets has recently been advocated by Ghysels and Jasiak (1994).
T.G. Andersen, T. Bollerslev /Journal of Empirical Finance 4 (1997) 115158
118
are most certainly present in both of the data sets analyzed here. These interdaily
features are clearly less significant and not critical for the high frequency analysis
pursued here. Yet, in analyses of longer run phenomena, accounting for these
effects may be equally important and could in principle be incorporated along the
same lines.
The remainder of the paper is organized as follows. Section 2 describes our
data and summarizes the intraday average return patterns. Section 3 contains an
analysis of the correlation structure of both raw and absolute 5minute returns, as
well as a comparison to the corresponding properties of the two daily time series.
The impact of periodic heteroskedasticity on the 5minute correlations is strong,
while the evidence of standard conditional heteroskedasticity, although evident at
the daily level, appears weak at many intraday frequencies. This motivates our
simple model of intraday returns that renders formal assessments of the relation
between the intra and interdaily correlation patterns feasible. Section 4 investi
gates the properties of temporally aggregated intraday returns. Estimates of the
degree of volatility persistence at the various sampling frequencies are contrasted
to the theoretical aggregation results. Our estimation strategy for characterizing the
intraday periodicity is presented in Section 5. A relatively simple model that
allows for a direct interaction between the level of the daily volatility and the
shape of the intradaily pattern provides a close fit to the average intradaily
volatility patterns for both return series, with the interaction effect being less
significant for the foreign exchange market. The corresponding time series proper
ties of the filtered returns obtained by extracting the estimated volatility patterns
from the raw series is also explored. Estimation results for these returns are much
more in line with the theoretical predictions. Moreover, this analysis strongly
suggests that several distinct component processes affect the volatility dynamics.
This finding may help shed new light on the longmemory feature in low
frequency return volatility documented by a number of recent studies. Section 6
contains concluding remarks. Details regarding the construction of the 5minute
foreign exchange and equity returns employed throughout and the flexible non
parametric procedure used in the estimation of the intraday periodicity are
contained in the appendices.
2. Intraday return periodicity
Our primary data set consists of 5minute returns for the DeutschemarkU.S.$
( D M  $ ) exchange rate from October 1, 1992 through September 30, 1993,
comprising 74,880 observations, and the Standard and Poor's 500 ( S & P 500)
composite stock index futures contract from January 2, 1986, through December
31, 1989, consisting of a total of 79,280 observations. A more detailed description
of the data sources and the calculation of the 5minute returns is provided in
Appendix A. In addition, we use two daily time series of 3,649 spot D M  $
T. G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158 119
exchange rates from March 14, 1979 through September 29, 1993 and 9,558
observations on the S &P 500 cash index from January 2, 1953 through December
31, 1990 7. All the empirical work is done in parallel, with tables and figures for
the foreign exchange and equity data labelled 'a' and 'b', respectively.
2.1. The Deutschemark U.S. dollar foreign exchange data
The sample mean of the 5minute Deutschemark appreciation of 0.000175% is
indistinguishable from zero at standard significance levels given the sample
standard deviation of 0.047% 8. However, the returns are clearly not normally
distributed. For example, the sample skewness of 0.367 and the sample kurtosis of
21.5 are both highly statistically significant 9. At the same time, the maximum and
minimum 5minute returns of 1.24% and  0 . 6 3 7 % do not suggest the presence of
sharp discontinuities in the series. A small negative first order autocorrelation
coefficient of  0 . 0 4 provides some support for the hypothesis that foreign
exchange dealers position their quotes asymmetrically relative to the perceived
'true market price' as a way to manage their inventory positions, thus causing the
midpoint of the quoted prices to move around in a fashion similar to the bidask
bounce often observed on organized exchanges 10. A more detailed set of sum
mary statistics are available in Andersen and Bollerslev (1994).
In order to evaluate the intraday periodicity of the returns, Fig. l a plots the
average sample mean for each 5minute interval. The average returns are centered
around zero but numerous violations of the constant 5% confidence band for the
null of an i.i.d, series occur between 09.00 GMT and 18.00 GMT (interval range
108 to 216). Allowing for different return sample variances across the day
produces a more realistic timevarying confidence band that is violated at seem
ingly random points in time and at a frequency consistent with the 5% band (13
violations over 288 intervals). Thus, there appears to be little evidence for any
systematic D M  $ appreciation or depreciation through the regular trading day
cycle 11
7 The initial observation on March 14, 1979 for the exchange rates corresponds to the beginning of
the EMS. The Standard 90 was replaced by the broader S and P 500 composite index on January 2,
1953. Also, for reasons discussed below, the estimates for the S and P 500 data exclude the October
1987 crash period.
8 Assuming the returns to be uncorrelated, the standard deviation for the mean equals
0.047/(74,880) ˜/2 = 0.000172%.
9 The standard errors of these statistics in their corresponding asymptotic normal distributions are
(6//T) 1/2 and ( 2 4 / T ) I/2, or 0.009 and 0.018, respectively (see e.g. Jarque and Bera, 1987).
J0 The coefficient is small in an economic sense but given the large sample size it is highly
statistically significant. Bollerslev and Domowitz (1993) also report a negative first order autocorrela
tion in 5minute D M  $ returns over 3 months in 1989, but find the correlation for artificially
constructed 5minute pseudo transactions price returns to be positive.
Jl This is counter to Ito and Roley (1987) who found evidence for systematic dollar appreciation
during the U.S. segment of the market but dollar depreciation during the European trading hours.
T. G. Andersen, T. Bollersleu / Journal of Empirical Finance 4 (1997) 115158
120
0.03 ˜ ˜
0.02
0.01
0.00
˜0.01
<
0.02
(a)
0.03
40 80 120 160 200 240 280 320
Five Minute I n t e r v a l
0.03
0.02
0.01
"˜ o.oo
v v.˜v ˜/,vv ˜W,X/v,
" IA 
˜0,01
0.02
(b)
0,03 L i i i t i i h i i i i i
lO 20 30 40
50 60 70 80 90
Five Minute Intervat
Fig. I. Intraday average returns, (a) D M  $ , (b) S&P 500.
In contrast, the exchange rate volatility fluctuates dramatically over the daily
cycle. The average absolute returns over the 5minute intervals are depicted in Fig.
2a. It reveals a pronounced difference in the volatility over the day, ranging from a
low of around 0.01% at 04.00 GMT (interval 48) to a high of around 0.05% at
15.00 GMT (interval 180) 12. This pattern is closely linked to the cycle of market
activity in the various financial centers around the globe. The volatility starts out
at a relatively high level followed by a slow decay up to around 03.00 GMT
(interval 36). The strong drop between intervals 40 and 60 corresponds to the
12 Since the average standard error for the absolute returns is 0.0022%, these differences are highly
statistically significant.
T. G. Andersen, T. Bollersleu / Journal of Empirical Finance 4 (1997) 115158 121
0.08
(a)
0.07
0.06
0.05
˜ 0.04
2
0.0`3
. 0.02
0.01
0.00 8hO J , J
, , ˜ , i , ,
520
40 120 160 200 240 280
Five Minute Interval
0.12
(b)
0.11
0.10
0.09
0.08
o 0.07
â€¢ 0.06
0.05
0.04 i i i , i , i L i i ˜ i i i i
90
0 10 20 30 40 50 60 70 80
Five Minute Interval
(a) D M  $ , (b) S&P 500.
Fig. 2. Intraday average absolute returns,
lunch hour in the Tokyo and Hong Kong markets. Activity then picks up during
the afternoon session in the Far Eastern markets and is further fueled by the
opening of the European markets around 07.00 GMT (interval 84). The market
volatility then declines slowly until the European lunch hour at 11.30 GMT
(interval 138), before it increases sharply during the overlap of afternoon trading
in Europe and the opening of the U.S. markets around 13.00 GMT, or 7.00 a.m.
New York (interval 156). After the European markets close volatility declines
monotonically until trading associated with the Far Eastern markets starts to pick
up again around 21.00 GMT (interval 252). The robustness of this intraday
volatility pattern is confirmed by the subsample analysis and the sorting of days
according to volatility levels reported in the more detailed analysis in Andersen
and Bollerslev (1994), which is also consistent with earlier findings in Wasser
T.G. Andersen, T. BollersleL,/ Journal of Empirical Finance 4 (1997) 115158
122
fallen (1989), Miiller et al. (1990), Baillie and Bollerslev (1991) and Dacorogna et
al. (1993) ˜3. Standard summary statistics further verify the overwhelming signifi
cance of this intraday volatility pattern. In particular, the first order autocorrelation
coefficient for the absolute 5minute returns of pA = 0.309 exceeds the 1 / x / T
asymptotic standard error by almost a factor of one hundred, while the L j u n g  B o x
statistic for up to tenth order serial correlation in ]Rt,,,] equals QA(10) = 36,680 J4
2.2. The standard and poor's 500 stock index futures data
The basic features of the 5minute S & P 500 are qualitatively similar to those of
the 5minute D M  $ returns. Perhaps, the most notable difference is that the
standard deviation for the stock index futures return of 0.104% is more than
double the value for the foreign exchange market. However, since the overnight
returns for the S & P 500 are excluded, the average 5minute standard deviation
corresponds to active trading on the CME only, whereas the foreign exchange
returns cover the entire 24hour trading cycle and therefore include periods of
relatively slow activity. Even so, when judged by auxiliary statistics such as the
sample m i n i m u m and m a x i m u m of 2.22% and  2 . 7 6 % , the equity market
exhibits the more volatile returns. Another distinguishing feature is the virtual
absence of autocorrelation in the futures returns. Although the first ten autocorrela
tion coefficients are highly significant, the coefficients are economically small and
have unpredictable signs ˜5. This lack of correlation contrasts sharply with results
reported by most studies on the intraday S & P 500 cash market, where nonsyn
chronous trading effects imply that stale prices may enter the calculation of the
index (see e.g. Chart et al., 1991) 16
The intraday periodic patterns over the eighty 5minute intraday intervals are
depicted in Fig. l b and Fig. 2b. Apart from the positive returns over the initial
5minute interval from 8.35 to 8.40 a.m. and towards the end of the trading day,
˜3We follow Dacorogna et al. (1993) in using GMT time scale throughout our analysis. Daylight
savings time is observed in Europe and North America, but not in East Asia. From the subsample
analysis in Andersen and Bollerslev (1994) this gives rise to a one hour difference in the peaks
associated with the regular release of U.S. macroeconomicannouncementsat 08.30 a.m. corresponding
to interval 162 for winter time and interval 150 for summer time. Ederington and Lee (1993) and
Harvey and Huang (1991) also suggest that macroeconomic announcementeffects have a distinct
impact on the average volatilityin early Friday morningtrading in the U.S. segmentof the market. We
do not pursue this or any other dayoftheweekeffects any further here, however.
14The first ten autocorrelationsfor 4Rt.,,Iare 0.309, 0.256, 0.238, 0.214, 0.212, 0.199, 0.204, 0.182,
0.185 and 0.182.
15The first ten autocorrelationsof the returns are 0.009, 0.003, 0.009, 0.010, 0.004, 0.018,
0.009˜ 0.015, 0.013 and 0.008, respectively. The corresponding LjungBox statistic equals Q(10)=
87.6.
16The impact of nonsynchronoustrading has been explored extensively in the literature (see e.g.
Scholes and Williams, 1977; Lo and MacKinlay, 1990).
123
T. G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158
the violations of the 5% confidence bands for the average returns are dispersed
unpredictably over the trading day 17. Nonetheless, as was the case for the foreign
exchange market, the systematic return effects are dwarfed by the systematic
movements in the return volatility, here documented in Fig. 2b. The average
absolute returns attain the commonly observed intraday Ushape, starting out at
0.095% in the morning, followed by a smooth decline to a level of 0.055% around
noon and a steady rise to 0.105% towards the end of trading in the cash market.
The subsequent drop and rise over the last fifteen minutes corresponds to the post
cash market trading on the CME 18. The robustness of this intraday periodicity in
the S & P 500 returns is again underscored by the more detailed analysis in
Andersen and Bollerslev (1994) in which the full four year sample is divided into
calendar years as well as four daily volatility categories. The only discernible
difference across these subsample patterns is a tendency for the fight part of the
' U ' to occasionally rise above the left part, creating more of a 'J' shape.
Interestingly, this tendency appears to be concentrated on high volatility days. The
model proposed in Section 5 below explicitly accounts for this phenomenon.
Several recent studies have attempted to rationalize the pronounced Ushape
pattern in intraday stock market volatility by strategic interaction of traders around
market openings and closures (see e.g. Admati and Pfleiderer, 1988, 1989; Foster
and Viswanathan, 1990; Son, 1991; Brock and Kleidon, 1992). Even though the
foreign exchange market operates on a continuous basis, the volatility pattern for
the D M  $ depicted in Fig. 2a may be viewed, tentatively, as a sum of two
overlapping Ushapes corresponding to the Far East and European trading hours,
along with an inverted Ushape for the U.S. segment of the market. Hence, in spite
of obvious differences in market microstructures, the foreign exchange returns are
calculated from quotes in a 24hour overthecounter market while the equity
returns are obtained from transaction prices on an organized futures market with
well defined daily closings, the pattern of intraday periodicity in the two markets
share important common characteristics.
3. Characterization and modeling of the correlation structure in intraday
returns
3.1, I n t r a d a y r e t u r n c o r r e l a t i o n s
While the intraday volatility patterns documented in the preceding section may
be irrelevant for standard studies of the return dynamics based on price observa
17This is related to the findings in Harris (1986a) who reports that the average positive returns in the
equity markets tend to occur over the first 45 min of the trading day and the very last trade of the day.
Notice also, that there is no indication of any abnormal positive returns after the cash market is closed.
18This Ushaped pattern in the volatility of S&P 500 futures prices following the closure of the cash
market has also recently been documented by Chang et al. (1995).
T. G. Andersen, 72 Bollerslev / Journal of Empirical Finance 4 (1997) 115158
124
tions at daily frequencies, conclusions drawn from the recent surge of empirical
papers on return volatility and market microstructure variables at the intraday
frequencies are likely subject to severe distortions due to the strong periodicity in
returns. We therefore supplement the prior investigation of the unconditional
volatility patterns with an explicit look at the dynamic features of our two return
series.
Fig. 3a and b display the sample autocorrelations of the 5minute returns for up
to five days i.e. 1440 observations for the foreign exchange and 400 for the equity
returns. All values are small and beyond the first few lags the series resemble
realizations of white noise. Thus, we again detect little of interest in the mean
Five Days Correlogram
0.02
0.01
0.00
0.01
o 0,02
0.03 1
0.04 (a)
0 250 500 750 ˜000 1250 1500
Five Minute Lag
Five Days Correlogram
tJlJ..... ,lliLiil,iti,iil, iii
0.00 I ,
0.01
(b)
0.02 , , ' , ' , ' ' ' '
0 5'0 100 150 2;0 2;0 3;0 3;0 400
Five Minute Lag
Fig. 3. Five days correlogram of intraday returns, (a) D M  $ , (b) S&P 500.
T. G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158 125
returns. In contrast, the autocorrelation patterns for the absolute returns are
strikingly regular. Consider the series for the D M  $ exchange rate in Fig. 4a. The
strong intraday pattern induces a distorted Ushape in the sample correlogram ˜9
Notice also how the size of the autocorrelations at the daily frequencies decay
slowly over the first four days, only to increase slightly at the fifth, or weekly,
frequency. This signals the presence of a minor dayoftheweek effect, which we
ignore in the remainder. Fig. 4b for the S & P 500 futures returns is equally telling.
The slowly declining Ushape occupies exactly 80 intervals, corresponding to the
daily frequency.
3.2. Interpretation in terms o f a suggestive intraday return model
The pronounced systematic fluctuations in the return correlogram provide an
initial indication that direct A R C H modeling of the intraday return volatility would
be hazardous. Standard A R C H models imply a geometric decay in the return
autocorrelation structure and simply cannot accommodate strong regular cyclical
patterns of the sort displayed in Fig. 4. Instead, it seems intuitively clear that the
combination of recurring cycles at the daily frequency and a slow decay in the
average autocorrelations may be explained by the joint presence of the pronounced
intraday periodicity documented above coupled with the strong daily conditional
heteroskedasticity 20. The following stylized model provides a simple specification
of the interaction between these two components,
N 1 N
R, E R,,n = E soZ,,n. (1)
n=l n=l
Here, R t denotes the daily continuously compounded return calculated from the N
uncorrelated intraday return components, Rt, .. The conditional volatility factor for
day t is denoted by ˜rt, while s n refers to a deterministic intraday periodic
component and Z,, n is an i.i.d, mean zero, unit variance error term assumed to be
independent of the daily volatility process, {˜rt}. Both volatility components must
be nonnegative i.e. % > 0 a.s. for all t and s, > 0 for all n. The following
terminology for the normalized, deterministic sample means and covariances for
the periodic structure will prove convenient:
1 N 1 N 1 N
˜ E SnSn_i = M( ssi), ˜ E S2 = M( s2),
n = = M ( s) = 1,
˜
=1 n=l n=l
where s,,+j N = s n for any integer j and 0 < n < N.
19A corresponding figure is presented by Dacorogna et al. (1993). However, in their analysis the
correlations at the daily frequencies are sharply diminished due to a strong weekend effect. By
excluding weekend returns we have effectively eliminated this distortion.
20 The temporal variation in daily financial market volatility have been successfully modeled by
ARCH type processes (see Bollerslev et al. (1992) for a survey of this extensive literature).
126 T.G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158
0.35
0.30
0.25
o
t˜ 0.15
r˜ 0.10
0.05 i ˜ , ,
320
0 40 80 120 160 200 240 280
Five Minute Lag
Five Days C o r r e l o g r a m
0.35
0.30
0.25
0.20
:3
 0.15
0.10
0.05
0.00
0.05
0 250 500 750 1000 1250 1500
Five Minute Lag
(a)
Fig. 4. Intraday absolute returns, (a) DM$, (b) S&P 500.
In the absence of intraday periodicity ( s , = 1 for all n) the daily returns may be
represented in the form R t = o',N 1/2Y',,=1,NZt, ., where the return component
N1/2F,,= LNZ,,n is i.i.d, with mean zero and unit variance. Thus, Eq. (1) extends
the standard volatility model for daily returns to an intraday setting with indepen
dent return innovations and deterministic volatility patterns. Of course, this type of
periodicity is annihilated when the returns are measured at the daily frequency. In
particular, letting Z, denote an i.i.d, random variable with E(Z t) = 0 and Var(Z t)
= 1, we have
R t = M I / 2 ( S2 ) o ' , Z , , (2)
T.G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158 127
0.35
0.30
0.25
0.20
:,.7,
,$ 0.15
o 0.10
0.05
0.00
0.05 L i i
40
i'o 20 3o 5'o 8'o 7'o 8o
Five Minute La˜
Five Days Correlogram
0.35.
0.30 [
0.25 I
o.15
r˜ 0.10
0.05
0.00
0.05 i I i i
80 160 240 320 400
Five Minute Lag
(b)
Fig. 4 (continued).
so that the expected absolute return equals M1/2(s2)˜tEIZ, I. Since M1/2($2) > 1,
the expected daily absolute return is an increasing function of the fluctuations in
the intraday periodic pattern. However, this effect is limited to a scale factor.
Letting c = (EIZ, I)2  1 > 0, it follows that for t 4= ˜',
Cov( o,, o;_)
(3)
Corr(lRtl, ]RTI) = Var(o.t) + cE(˜rt2) "
Hence, the presence of periodic components reduces the overall level of the
T.G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158
128
interdaily return autocorrelations, without affecting the autocorrelation pattern 21
In contrast, the periodicity may have a strong impact on the autocorrelation pattern
for the absolute intraday returns. Straightforward calculations reveal
Corr(IR,,,,I, IR .... I)
m(ssn_m)COv(o't, ˜) q CoV(Sn, Sm)E2(o't)
(4)
M ( s 2 ) V a r ( % ) + CNE( Crt2)M( s 2) + E2( crr)Var( s) '
where Var(s) ˜ m ( s 2)  M2(s), Cov(sn, s,,) = M(ss,,_,n) M 2 ( s ) and C N

E  2 I Z t , , {  1. Eq. (4) illustrates the interaction between the periodicity in absolute
returns at the intradaily level and the conditional heteroskedasticity at the daily
level. For adjacent trading days the impact of the positive correlation in the daily
return volatility, captured by Cov(˜,, oT), is strong and induces positive depen
dence in the absolute returns, but as the distance between t and ˜ grows this effect
becomes less important which is consistent with the slow decay in the correlo
grams in the bottom panels of Fig. 4. At the same time, the correlograms are
affected by the strong intraday periodicity. For example, consider the display for
the absolute S & P 500 returns in Fig. 4b. The correlations attain their lowest
values around lag forty, or half a trading day. This corresponds to the bottom of
the Ushape for the average absolute returns depicted in Fig. 2b. Clearly, the
population covariance, Cov(s n, Sm), is minimized and significantly negative, at
this frequency. Eq. (4) verifies that the negative correlation between the 5minute
absolute returns, realized about halfaday apart, translates into a negative contri
bution to the corresponding correlogram at the 3  4 hour frequencies. Likewise,
Fig. 2a indicates that there is strong negative correlation between the absolute
foreign exchange returns in the intervals 80225 (covering about halfaday) and
all the remaining 5minute returns. Not surprisingly, the lower panel of Fig. 4a
verifies that this again results in highly significant troughs in the correlogram
around the 12 hour frequency (and its harmonics). Indeed, the impact is now
sufficiently strong that the absolute return autocorrelations turn negative. This is
truly remarkable given the very large positive autocorrelations found at the daily
frequency and it is testimony to the profound impact of the periodic structure on
the intraday return dynamics. In terms of the specification in Eq. (4), the size of
the second, negative, term of the numerator exceeds the first, positive, term around
the 12 hour frequency.
21 Consistent with the findings of Granger and Ding (1996), informal investigations reveal that the
dynamic dependencies are significantly more pronounced for the absolute as opposed to the squared
sample returns. Consequently, our intraday modeling focuses on the patterns in ]Rt] and ]Rt,,,], rather
than R2 and R,2,,,.
129
T.G. Andersen. T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158
3.3. Longrun implications and comparison to daily returns
To further assess the descriptive accuracy of the formulation in Eq. (1), we now
investigate the longrun implications for the correlogram. It is convenient to focus
on the daily frequency i.e. n = m,
C o v ( ˜ , a'˜) + ( V a r ( s ) / M ( s 2 ) ) E 2 ( o  , )
IR.,.I) = Var(˜rt) + CNE( O't2) Jr (Var( s) /M( s2) )E2( o',) "
Corr(IRt,.I,
(5)
Fig. 5a and b display the first forty autocorrelations for the absolute returns of the
two daily time series on the D M  $ spot exchange rate and the S & P 500 cash
(a) o.zo ...................
 DaLly, R˜ ]
/ "˜ t'˜
]
 R˜. n
O. 1 5 t˜ PlveMinute.
˜t x
IvvvvlIIlllIIIIIIIIIIIIVIIVllttl
[tiJfst[[tIttlffltlfttl
0.10 ˜ ,˜ ,˜ d,
˜ ˜ , ˜ / ,
,o
iiii vvv
vv
0.10 ' ' ' ' ' ' ' ' 2'0 '
0 4 8 12 16 24 28 32 36 40
D a i l y Lag
(b) 0.20
0.1 .. "I
[ ˜' , ,
i 0.05
0,00
0.05
0.10
4 8 12 16 20 24 28 32 36 40
Daily Lag
Fig. 5. Forty days correlogram of absolute returns,(a) DM$, (b) S&P 500.
T.G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158
130
index, along with the corresponding 5minute intraday autocorrelations out to lag
11,520 and 3,200, respectively 22. Direct comparison between the empirically
estimated Corr(IRtl, IR˜I) in Eq. (3) and the expression for Corr(lR,,nl, IR,,,[)
above are complicated by the different sample periods required for reliable
inference. Nonetheless, it is clear that the decay rate in the local maxima of the
intradaily absolute correlogram and the daily return autocorrelations should be
qualitatively comparable, as the Cov(%, o term governs both. Comparing the
7)
peaks in the intraday correlograms to the daily autocorrelations in Fig. 5a and b
confirms this implication of our stylized model; the dominant rate of decay is
strikingly similar for both markets ,_3
The findings in this section demonstrate a strong correspondence between the
qualitative implications of the model outlined in Eq. (1) and the stylized empirical
facts. It suggests that our model, stressing the conditional heteroskedasticity at the
daily level along with the strong deterministic periodicity at the intradaily level,
may serve as a good starting point for high frequency volatility modeling and as
such it constitutes the basis for the subsequent analysis in Section 5.
4. Implications for volatility modeling and high frequency return aggregation
Section 3 demonstrates that the distinct intraday periodicity has a strong impact
on the autocorrelation patterns of the 5minute returns. The question therefore
arises whether more formal time series modeling of return volatility is similarly
affected by the presence of periodic features, and if so, whether some observation
intervals are preferable relative to others for the purpose of drawing inference
concerning the dynamic features of interest. In order to address these issues this
section presents an extensive analysis of the properties of the return series
obtained at a range of different intradaily and interdaily frequencies.
4.1. Characterization of the intraday returns at the various frequencies
Summary statistics for the foreign exchange market are provided in Table la
for all seventeen possible intraday returns with a 24hour periodicity. The returns
are continuously compounded i.e. the nth return on day t for the series at
( k . 5)minute intervals is defined by Rtk,  52i=˜n_ l)k+ l.nkRt,i, t = 1, 2 . . . . . 260,
22 The sample autocorrelations are generally negatively biased and become less precise as the lag
length grows; see Percival (1993) who point out that the sum of all the sample autocorrelations by
construction equals zero. We therefore limit our analysis to laglengths which may appear large, but
nonetheless constitute a modest fraction of the total intraday sample.
23 The slow rate of decay in the autocorrelation functions is also in accordance with the apparent
longmemory feature of asset return volatility documented by a number of recent studies (see e.g.
Baillie et al., 1996; Ding et al., 1993).
131
T. G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158
n = 1, 2 . . . . . K where K  2 8 8 / k refers to the number of returns per day. Note
that while the 5minute return series consists of 74,880 observations, the hourly
series contains only 6,240 observations and the 1 / 2  d a y return series has a mere
520 observations. These differences should be kept in mind when interpreting the
evidence.
The standard deviations in Table l a grow at a rate almost proportional to the
square root of the sampling frequency. This is consistent with the 5minute returns
being approximately uncorrelated, although there is a small, but highly significant,
negative first order autocorrelation coefficients at the higher frequencies. As
mentioned, the weak negative correlation may be the result of spread positioning
by dealers causing mean reversion in the quote midpoints; an effect similar to a
b i d  a s k bounce in transactions data 24. In line with this explanation, the pj
coefficients generally turn insignificant at the 40minute and lower frequencies.
Further corroborating evidence along these lines is provided by the variance ratio
statistics,
vR = x . VarT(R ,n)
k '
iRt,n)
Varr (Eft= (6)
where V a r r ( R ˜ , ) and Varr(˜2,=l, k R,, n) denote the sample variances for the
k
intraday and daily returns, respectively. Expanding the daily variance estimate in
the denominator demonstrates that a value of the VRstatistic below unity will
result from positive autocorrelation between adjacent return components, while a
statistic above one is indicative of predominantly negatively correlated intraday
returns zs. Finally, it is worth noting from Table la, that the kurtosis of the D M  $
returns increases almost monotonically with the sampling frequency.
The first order autocorrelations of the absolute returns, pA are, not surprisingly,
all highly significant for the shorter intervals. However, beyond the 2hour
sampling frequency the autocorrelations drop off very sharply and in fact turn
negative at the 8 and 12 hourly frequencies ( k = 96, 144). This is, of course,
consistent with the negative region of the 5minute absolute return correlogram in
Fig. 4a. The VRAstatistic reported in the final column o f Table l a is calculated by
replacing Rtk,. with ]R˜,.] in the definition of VR in Eq. (6) 26 The statistic starts
out at 0.05 for the 5minute returns and rises almost monotonically to 0.69 for the
24 Note that the standard deviation of the 5minute returns is less than the average quoted bidask
spread. According to Bollerslev and Melvin (1994), more than half of the DM$ quotes are posted
with a spread of 0.10%, while the second most common and lowest regularly posted spread of 0.05%
accounts for about a quarter of the quotations.
25 Formal tests for serial correlation based on the VRstatistic may be calculated as outlined in Lo and
MacKinlay (1989).
26 Note that the denominator in this VRAstatistic involves the variance of the sum of the absolute
returns rather than the absolute value of the sum of the returns. The expected value of the VRstatistic
would not equal unity under the latter definition.
132 T.G. Andersen, T. BollersleL,/ Journal of Empirical Finance 4 (1997) 115158
Table 1
k T/k Mean St. Dev. Skew. Kurtosis Pl Q(10) VR pA QA(10) VRa
(a) Summary statistics for intraday DM$ exchange rate
l 74,880 0.018 0.047 0.368 21.5 0.040 281 1 . 1 9 4 0.309 36,680 0.054
2 37,440 0.035 0.066 0.363 1 6 . 6 0.070 263 1 . 1 6 2 0.313 17,563 0.071
3 24,960 0.053 0.079 0.200 1 3 . 6 0.089 220 1 . 1 1 8 0.307 10,710 0.086
4 18,720 0.070 0.089 0.276 1 4 . 0 0.082 154 1.084 0.287 6,296 0.099
6 12,480 0.105 0.107 0.534 1 2 . 6 0.043 36.5 1 . 0 2 3 0.268 2,757 0.127
8 9,360 0.140 0.121 0.135 9 . 1 1 0.023 22.0 0.994 0.272 1,736 0.149
9 8,320 0.158 0.126 0.345 10.1 0.002 1 8 . 9 0.948 0.251 1,212 0.161
12 6,240 0.210 0.148 0.326 1 1 . 0 0.001 12.7 0.978 0.229 609 0.193
16 4,680 0.280 0.170 0.318 8.77 0.032 13.1 0.968 0.246 425 0.235
18 4,160 0.315 0.178 0.489 10.6 0.058 21.9 0.947 0.193 219 0.260
24 3,120 0.420 0.212 0.166 8.94 0.011 16.7 1 . 0 1 2 0.164 159 0.311
32 2,340 0.560 0.246 0.326 9.15 0.018 43.3 1 . 0 1 8 0.171 238 0.373
36 2,080 0.630 0.253 0.329 7.89 0.047 37.7 0.954 0.097 109 0.416
48 1,560 0.840 0.300 0.400 5.92 0.002 30.5 1 . 0 0 7 0.075 66.6 0.487
72 1,040 1.261 0.373 0.319 6.59 0.019 20.6 1 . 0 4 2 0.007 67.60.679
96 780 1.681 0.423 0.389 5.19 0.022 1 8 . 2 1.004 0.025 53.20.653
144 520 2.521 0.520 0.192 4 . 3 1 (l.021 12.2 1.012 0.033 28.2 0.692
(b) Summary statistics for intraday S&P 500 returns
1 79,280 0.064 0.104 0.597 29.3 0.009 87.6 0.774 0.292 32,425 0.099
2 39,640 0.128 0.150 1.212 30.9 0.009 81.5 0.801 0.285 12,641 0.128
4 19,820 0.255 0.212 l.609 33.2 0.014 53.4 0.795 0.232 3,374 0.179
5 15,856 0.319 0.234 1.755 33.3 0.032 60.6 0.780 0.243 2,323 0.205
8 9,910 0.511 0.299 1.478 22.2 0.047 44.6 0.793 0.207 1 , 2 9 5 0.261
10 7,928 0.638 0.339 1.417 21.2 0.039 41.7 0.819 0.211 973 0.300
16 4,955 1.021 0.437 1.803 26.6 0.040 35.2 0.849 0.135 437 0.405
20 3,964 1.277 0.499 1.869 27.4 0.016 24.4 0.884 0.114 268 0.463
40 1,982 2.553 0.728 1.541 16.2 0.026 22.6 0.942 0.148 175 0.673
12 h o u r l y r e t u r n s . T h e r e s u l t s f o r the m u l t i p l e d a y r e t u r n s r e p o r t e d in A n d e r s e n
a n d B o l l e r s l e v ( 1 9 9 4 ) c o n t i n u e t h i s n e a r m o n o t o n e a s c e n t , r e a c h i n g 1.94 f o r t h e
b i w e e k l y s a m p l i n g interval. T h e s m o o t h increase s u g g e s t s that a c o m m o n c o m p o 
n e n t a c c o u n t s for a s u b s t a n t i a l p a r t o f t h e p o s i t i v e h i g h e r o r d e r d e p e n d e n c e in all
o f t h e r e t u r n series. T h e c o r r e s p o n d i n g p statistics of 0.123 and 0.118 for the
w e e k l y a n d b i w e e k l y s a m p l i n g f r e q u e n c i e s a l s o t e s t i f y to the i m p o r t a n c e o f t h e
i n t e r d a y h e t e r o s k e d a s t i c i t y 27
Hence, the VR A statistics convey a coherent message about the degree of conditional heteroskedas
27
ticity in the series. As a set of simple diagnostics, these statistics may therefore be more informative
about the nature of the volatility process than the standard LjungBox statistics for tenth order serial
correlation in the absolute returns, Q˜'(IO), which appear both erratic and highly dependent on the
sample size.
T.G. Andersen, T. Bollerslez:/ Journal of Empirical Finance 4 (1997) 115158 133
T h e s u m m a r y statistics for the S & P 5 0 0 i n d e x futures r e t u r n s in T a b l e l b
l a r g e l y p a r a l l e l t h o s e for the D M  $ returns. H o w e v e r , in c o n t r a s t to the results for
the e x c h a n g e rates, the first o r d e r a u t o c o r r e l a t i o n s a n d the V R  s t a t i s t i c s in T a b l e
l b all i n d i c a t e a slight p o s i t i v e i n t r a d a y d e p e n d e n c e . M o r e o v e r , the e q u i t y r e t u r n s
are n e g a t i v e l y s k e w e d a n d d i s p l a y v e r y s i g n i f i c a n t e x c e s s k u r t o s i s 28. F i n a l l y , the
i n t r a d a y r e t u r n p e r i o d i c i t y , h e r e d e p i c t e d in Fig. 2b, a g a i n h a v e a s t r o n g effect on
the c o r r e l a t i o n s for the a b s o l u t e i n t r a d a y r e t u r n s , a l t h o u g h the d e c a y in the pA
c o e f f i c i e n t s for the l o w e r f r e q u e n c i e s is less p r o n o u n c e d t h a n for the e x c h a n g e
rates.
4.2. Specification o f the uolatili˜ model and the associated persistence measures
N u m e r o u s r e c e n t studies h a v e r e l i e d o n m o r e f o r m a l t i m e series t e c h n i q u e s in
the a n a l y s i s o f h i g h f r e q u e n c y r e t u r n d y n a m i c s b o t h w i t h i n a n d a c r o s s d i f f e r e n t
m a r k e t s . T h e m o s t c o m m o n l y e m p l o y e d f o r m u l a t i o n is the G A R C H ( 1 , 1) m o d e l
p r o p o s e d i n d e p e n d e n t l y b y B o l l e r s l e v ( 1 9 8 6 ) a n d T a y l o r (1986). T h u s , in o r d e r to
e v a l u a t e the p o t e n t i a l i m p a c t o f the s t r o n g i n t r a d a y p e r i o d i c i t y in this c o n t e x t w e
Notes to Table l:
(a) The percentage returns are based on interpolated 5minute logarithmic average bidask quotes for
the DeutschemarkU.S. dollar spot exchange rate from October 1, 1992 through September 29, 1993.
Quotes from Friday 21.00 Greenwich mean time (GMT) through Sunday 21.00 GMT have been
excluded, resulting in a total of 74,880 return observations. The length of the different intraday return
sampling intervals equals 5k minutes. Each time series has a total of T / k nonoverlapping return
observations. The sample means have been multiplied by one hundred. The columns indicated by Pl
and pA give the first order autocorrelations for the returns and the absolute returns. The Ljung and Box
(1978) portmanteau test for up to tenth order serial correlation in the returns and the absolute returns
are denoted by Q(10) and QA(I0), respectively. The variance ratio's for the different sampling
frequencies versus the daily return variance are denoted by VR. The corresponding variance ratio
statistics for the absolute returns are given in the VR A column.
(b) The returns are based on 79,280 interpolated 5minute futures transactions prices for the Standard
and Poor's 500 composite index. The sample period ranges from January 2, 1986 through December
31, 1989, excluding the period from October 15, 1987 through November 13, 1987. Overnight five
minute returns have also been deleted, resulting in a total of 80 intraday return observations from 08.35
through 15.15 for each of the 991 days in the sample. The length of the different intraday return
sampling intervals equals 5.k minutes. Each time series has a total of T / k nonoverlapping return
observations. The sample means have been multiplied by one hundred. The columns indicated by p˜
and pA give the first order autocorrelations for the returns and the absolute returns. The Ljung and Box
(1978) portmanteau test for up to tenth order serial correlation in the returns and the absolute returns
are denoted by Q(10) and QA(10), respectively. The variance ratio's for the different sampling
frequencies versus the daily return variance are denoted by VR. The corresponding variance ratio
statistics for the absolute returns are given in the VR A column.
2s The negative skewness may be interpreted as evidence of the socalled 'leverage' and/or
'volatility feedback' effects discussed by Black (1976), Christie (1982) and Nelson (1991), and
Campbell and Hentschel (1992), respectively.
T.G. Andersen, T. Bollerslec / Journal of Empirical Finance 4 (1997) 115158
134
present MA(1)GARCH(1, 1) estimation results for each of the intradaily sam
pling frequencies in Table 2a and b. Formally, the model is defined by
R˜,˜ = tx(h ) + O(k)etk,,_ 1 + ek
t,n˜
and
where E,,˜_ l(e,˜..)= 0 and Et, ˜_ l[(e˜.) 2 ] = (o˜k.) 2 denotes the conditional return
variance over the subsequent intraday period, with the subscript (t, 0) defined to
equal (t  1, K). The reported parameter estimates for a(k ) and fl(k) are obtained
Table 2
k T/k o˜(k˜ tick) a˜k) + tick) Half life Mean lag Median lag
(a) Persistence of MA(1)GARCH(I, 1) models for intraday D M  $ exchange rate
k _
Rt,n = 100"˜i=(n 1)k+l,nkRt,i = I't'(k) + OIk)Etk, n I
+ ,k (o.,˜o)2 = ,o˜k˜ +a(k˜(et.,,_
k j) 2 +fl˜k)(o't,n_l ) 2 t = l , 2 ,
k ...,260, n=1,2, ...,288/k
1 74,880 0.193 (0.0l 1) 0.822 (0.009) 1.015 ˜ ˜ w
2 37,440 0.229 (0.012) 0.774 (0.008) 1.003 ˜ co
3 24,960 0.273 (0.018) 0.708 (0.014) 0.981 533 725 488
4 18,720 0.287 (0.019) 0.677 (0.016) 0.964 375 488 320
6 12,480 0.322 (0.035) 0.579 (0.033) 0.901 200 233 138
8 9,360 0.286 (0.028) 0.581 (0.037) 0.868 195 207 108
9 8,320 0.306 (0.035) 0.521 (0.042) 0.828 165 167 81
12 6,240 0.311 (0.047) 0.395 (0.069) 0.706 119 105 35
16 4,680 0.261 (0.039) 0.456 (0.074) 0.718 167 136 < 40
18 4,160 0.270 (0.061) 0.246 (0.124) 0.516 94 67 < 45
24 3,120 0.018 (0.015) 0.969 (0.026) 0.988 6,771 5,919 1,878
32 2,340 0.016 (0.008) 0.975 (0.013) 0.991 12,159 11,219 4,318
36 2,080 0.011 (0.004) 0.978 (0.005) 0.989 11,311 8,293 266
48 1,560 0.011 (0.004) 0.979 (0.005) 0.990 17,084 13,229 1,748
72 1,040 0.007 (0.005) 0.987 (0.004) 0.987 19,585 10,153 < 180
96 780 0.014 (0.008) 0.969 (0.007) 0.983 19,637 13,202 < 240
144 520 0.010 (0.010) 0.960 (0.007) 0.970 16,329 5,988 < 360
(b) Persistence of MA(1)GARCH(I, 1) models for intraday S&P 500 returns
k __ k
Rt,n = 100"˜i=(n I ) k + 1 , n k R t , i = t˜(k) 4 O(k)˜'t,n 1
4e,., (Â°'t,n)w(k)+Â°l(k)(et.n 1)2 4 /3˜k˜(o't,,_l)2, t = l , 2
k k2 k k . . . . . 991, n = 1 , 2 . . . . . 8 0 / k
1 79,280 0.137 (0.004) 0.838 (0.005) 0.975 137 168 105
2 39,640 0.180 (0.010) 0.765 (0.011) 0.945 121 138 79
4 19,820 0.223 (0.024) 0.664 (0.036) 0.887 116 118 57
5 15,856 0.230(0.067) 0.630(0.123) 0.861 116 112 49
8 9,910 0.053 (0.027) 0.935 (0.036) 0.988 2,213 2,602 1,559
10 7,928 0.048 (0.018) 0.940 (0.023) 0.988 2,947 3,437 2,043
16 4,955 0.148 (0.333) 0.764 (0.694) 0.912 606 575 240
20 3,964 0.060 (0.049) 0.890 (0.092) 0.951 1,376 1,124 246
40 1,982 0.108 (0.158) 0.798 (0.315) 0.906 1,397 1,128 228
T.G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158 135
by quasimaximum likelihood methods assuming the innovations to be condition
ally normally distributed. The corresponding robust standard errors for the esti
mates are provided in parentheses (see Bollerslev and Wooldridge, 1992). We note
that, although it usually represents a reasonable approximation, the GARCH(1, 1)
model is not necessarily the preferred specification for the return generating
process in all, or even most, instances. However, estimating the same model across
both asset classes and all return frequencies facilitates meaningful comparisons of
the findings. Moreover, it corresponds to the class of models for which theoretical
aggregation results are available. The MA(1) term is included to account for the
economically minor, but occasionally highly statistically significant, first order
autocorrelation in the returns.
Unfortunately, an unambiguous characterization of the estimated volatility
dynamics and the associated persistence properties is not possible in this nonlin
ear setting (see Bollerslev and Engle (1993), Bollerslev et al. (1994) and Gallant et
al. (1993) for further discussion of these issues). Hence, we supplement the
parameter estimates for a(k ) and /3˜k) in Table 2a and b with three additional
summary measures for the implied degree of volatility persistence. In particular, if
a(˜) +/3(k ) < 1, the jstep ahead prediction for the conditional variance may be
written as
k 2 j ˜2
where 02  o˜(k)(1  a(k >  / 3 ( ˜ ) ˜ equals the unconditional variance of the
Notes to Table 2:
(a) The returns are based on 288 interpolated five minute logarithmic average bidask quotes for the
DeutschemarkU.S. dollar spot exchange rate from October 1, 1992 through September 29, 1993.
Quotes from Friday 21.00 Greenwich mean time (GMT) through Sunday 21.00 GMT have been
excluded, resulting in a total of 74,880 return observations. The length of the different intraday retum
sampling intervals equal 5. k minutes. The model estimates are based on T˜ k nonoverlapping return
observations. The ot(k) and /3˜k) columns give the Gaussian quasimaximum likelihood estimates for
the GARCH(1, 1) parameters. Robust standard errors are reported in parentheses. The half life of a
shock to the conditional variance at frequency k is calculated as log(2)/log(a˜k)+/3tk )) and
converted into minutes. The mean lag of a shock to the conditional variance is given by a(k ) +/3ck ) > 1.
The median lag of a shock to the conditional variance is calculated by Â½ +[log(l  / 3 ( k ˜ )  1og)a(k))
log(2)]/1og(a(k ) + /3(k)) and reported in number of minutes. For 2a˜k ) < 1 /3(k ) the median lag is
less than Â½. The median lag is also not defined for a(t) +/3(k ˜ > 1.
(b) The returns are based on 79,280 interpolated five minute futures transactions prices for the Standard
and Poor's 500 composite index. The sample period ranges from January 2, 1986 through December
31, 1989, excluding the period from October 15, 1987 through November 13, 1987. Overnight five
minute returns have also been deleted, resulting in a total of 80 intraday return observations from 08.35
through 15.15 for each of the 991 days in the sample. The length of the different intraday return
sampling intervals equal 5. k minutes. The model estimates are based on T / k nonoverlapping return
observations. The ˜(k) and /3(k) columns give the Gaussian quasimaximum likelihood estimates for
the GARCH(I, 1) parameters. See (a) for the definition of the half life, mean lag and median lag
statistics.
136 T.G. Andersen, T. Bollerslet,/ Journal of Empirical Finance 4 (1997) 115158
return innovations. The 'halflife' of the volatility process is then defined as the
number of time periods it takes for half of the expected reversion back towards or 2
to occur i.e.  l o g ( 2 ) . l o g ( a ( k ) +/3(k )) i. Alternatively, by defining the condi
tional heteroskedastic squared return innovations, v ˜ , ,  ( ˜ n ) 2  ( ˜ , ) 2, the
GARCH(1 , 1) model may be expressed as an infinite M A model for (E tkn )2 with ,
positive coefficients, 0˜ k,
E k = 0"2"˜ o/.(k)E(og(k)q[˜(k)) i lp k q_ 1) k ˜ 02} E o k p k
t,n t,n i t,n t t,n t"
i=1 i=0
This specification suggests the corresponding 'mean lag', a(˜)(1  o˜(˜ 2/3˜k) +
oe(k)/3(k ) +/3(k)) i and 'median lag', Â½ + [log(1 /3(k ˜)  log(oe(k ))  log(2)] â€¢
log(a(k ) +/3(k˜) 1 , as additional measures for characterizing the degree of volatil
ity persistence and the duration of the dynamic adjustment process in squared
returns across the different sampling frequencies 29 Neither the mean nor the
median lag is defined for ce(k) +/3˜k ˜> 1. Also, the median lag is less than 1 / 2 for
2a(˜) +/3(k ) < 1.
4.3. Interpretation o f the G A R C H results .for different return frequencies
This section summarizes the evidence from fitting standard G A R C H models to
the return series at different frequencies. Particular emphasis is placed on the type
of distortions that may be induced by the strong periodic intraday patterns which
are ignored in these models. There are a couple of indirect ways to gauge the
effect. First, there are theoretical predictions about the relation between the
parameters at various frequencies. If these are most obviously violated at the
particular frequencies where the intraday periodicity are expected to assert the
maximal impact, this is therefore consistent with the periodic pattern being a
dominant source of misspecification for these models. Second, to the extent that
the periodic pattern is a strictly deterministic intraday phenomenon as suggested in
Section 3, the distortions should be absent from models estimated at daily or
multipleday frequencies. Consequently, if the theoretical aggregation results work
satisfactorily at the multipleday frequencies but break down intradaily then this is
further evidence of a significant impact of the periodic pattern on the dynamic
properties of the intraday volatility process. We also relate our findings to the prior
estimates reported from intraday volatility modeling. The comparison shows that
our results are fully consistent with the diverse set of estimates reported in the
literature once we control for the different return frequencies employed in the
studies. Finally, the explicit incorporation of the cyclical pattern in Section 5
verifies that most of the distortions attributable to the intraday volatility cycle may
29 The mean lag is given by Ei o.˜iOi whereas the median lag, m, is implicitly defined by
˜,
Y˜i=0,,,0,k = l/2.E˜_a˜O ˜ (see Harvey, 1981).
T.G. Andersen, T. Bollerslet / Journal of Empirical Finance 4 (1997) 115 158 137
be eliminated. Hence, our findings apply readily to the majority of the prior high
frequency studies in the literature, and, in particular, provide an indication of the
magnitude of their potential biases due to the neglect of the intraday periodicity in
the volatility process.
The M A ( 1 )  G A R C H ( 1 , 1) results for the intraday foreign exchange rates are
given in Table 2a. The implied persistence measures reveal an alarming degree of
irregularity across the different sampling frequencies. For the longer intraday
intervals the estimates, converted into minutes, point to half lives around 18,000,
or about 12Â½ trading days and mean lags of around 8  9 days. However, the
1 I
corresponding measures collapse at the intermediate 5  1 g hour frequencies
(k = 6  1 8 ) , becoming less than 4 hours, only to resurrect again at the lowest,
5  1 0 minute, intervals ( k = 1, 2) where violations of the a(k / +/3ck ) < 1 inequality
cause the estimated processes to be covariance nonstationary.
These intraday results contrast sharply with the findings for the interdaily
D M  $ returns reported in Table 3a i.e. R =  ˜ r = ( t _ l ) k + l , t k R . r , t = 1, 2 . . . . .
[3,649/k], k = 1, 2 . . . . . 10 where [.] denotes the integer value. Here, the
persistence measures appear quite consistent over the different return intervals,
with the half lives and mean lags fluctuating around 20 and 15 days, respectively 3o
As for the intraday returns, the median lag is always substantially lower than the
mean lag and measured with some imprecision resulting in numerous violations of
the inequality governing the lower bound of the statistic, particularly for the
smaller sample sizes.
A formal framework for assessment of the parameter estimates obtained at the
various sampling frequencies is available from the results on temporal aggregation
in A R C H models provided by Nelson (1990, 1992), Drost and Nijman (1993) and
Drost and Werker (1996). Specifically, assuming that the GARCH(1, 1) model
serves as a reasonable approximation to the returns process at the daily frequency,
it follows from Drost and Nijman (1993) that the estimates for the corresponding
weak GARCH(1, 1) models at the lower interdaily frequencies should be related to
the daily parameters via the simple formula a ˜ + / 3 ˜ k ) = ( C e ˜ l ) + fl˜l/) k. This
implies that the estimated half lives, when converted to a common unit of
measurement as in our tables, should be stable across the frequencies 31. Our
evidence in Table 3a is in line with this prediction and it is also consistent with
30 The intraday measures in Table 2a are converted to minutes whereas the interdaily results in Table
3a are given in days. Furthermore, recall that the weekend returns have been excluded from the
intraday series. This may induce a distortion in the return dynamics but, again, our informal analysis
found this effect to be inconsequential.
3t Note that any serial dependence in the mean will generally increase the order of the implied low
frequency weak GARCH model beyond that of the high frequency GARCH(1, 1) model (see Drost and
Nijman (1993) for further details). However, the estimate for the MA(l) term for the daily DM$
GARCH(I, 1) model is only 0.034 with an asymptotic standard error of 0.018.
138 T.G. Andersen, T. Bollerslec / Journal of Empirical Finance 4 (1997) 115158
Table 3
Half life Mean lag Median lag
k [T/k] ot(k ) tick) Â°˜C˜)+ tick)
(a) Persistence of MA(I)GARCH(I, l) models for daily DM$ exchange rates
R t  100 ")ZT=(˜ l)k+ I,tk Rr = tx(k˜ + O(˜et I + etk (crt) 2 = w(k)
k= k k
+ ot(k)(etk_ 1)2 +/3(k)(o't ˜_1)2, t = 1, 2 . . . . . [ T / k ]
l 3,649 0.105 (0.015) 0.873 (0.015) 0.978 31.2 37.7 23.2
2 1,824 0.150 (0.024) 0.784 (0.026) 0.934 20.6 21.4 10.9
3 1,216 0.106 (0.021) 0.813 (0.037) 0.919 24.8 21.2 6.1
4 912 0.167 (0.036) 0.713 (0.042) 0.879 21.6 19.2 6.5
5 729 0.182 (0.049) 0.611 (0.081) 0.794 15.0 11.4 < 2.5
6 608 0.19l (0.049) 0.646 (0.060) 0.838 23.5 20.9 5.6
7 521 0.129 (0.049) 0.674 (0.071) 0.803 22.2 14.2 < 3.5
8 456 0.170 (0.051) 0.563 (0.1761 0.733 17.8 11.7 < 4.0
9 405 0.133 (0.067) 0.641 (0.230) 0.774 24.3 14.7 < 4.5
10 364 0.174 (0.079) 0.434 (0.186) 0.607 13.9 7.8 < 5.0
(b) Persistence of MA(I)GARCH(1, 1) models for daily S&P 500 returns
1 9,558 0.089 (0.019) 0.906 (0.0t8) 0.995 147 202 137
2 4,779 0.087 (0.015) 0.902 (0.015) 0.990 135 175 114
3 3,186 0.108 (0.016) 0.870 (0.018) 0.979 98 119 74
4 2,389 0.093 (0.016) 0.889 (0.(/191 0.983 158 194 121
5 1 , 9 1 1 0.101 (0.015) 0.900 (0.025) 1.000 :c ˜
6 1,593 0.127 (0.037) 0.838 (0.044) 0.965 117 135 79
7 1,365 0.177 (0.085) 0.776 (0.076) 0.953 100 117 69
8 1,194 0.137 (0.050) 0.821 (0.044) 0.958 129 145 83
9 1,062 0.123 (0.035) 0.805 (0.052) 0.924 84 79 33
10 955 0.173 (0.066) 0.768 (0.030) 0.941 114 127 71
(a) The returns are based on 3,649 daily quotes tor the DeutschemarkU.S. dollar spot exchange rate
from March 14, 1979 through September 29, 1993. Weekend and holiday quotes have been excluded.
The length of the return intervals equals k days, for a total of [ T / k ] observations, where [. ] denotes
the integer value. See Table 2a for the definition of the half life, mean lag, and median lag. These
measures are converted to trading days.
(b) The returns are based on 9,558 daily observations for the Standard and poor's 500 composite index
from January 2, 1953 through December 31, 1990. The length of the return intervals equals k days, for
a total of [ T / k ] observations, where [. ] denotes the integer value. See Table 2a for the definition of the
half life, mean lag, and median lag. These measures are converted to trading days.
e a r l i e r e v i d e n c e f o r o t h e r i n t e r d a i l y e x c h a n g e r a t e s r e p o r t e d in B a i l l i e a n d B o i l e r 
slev (1989).
T h e o b s e r v a t i o n s a b o v e s u g g e s t that t h e r e s u l t s f o r t h e i n t r a d a y e x c h a n g e r a t e s
in T a b l e 2 a a r e i n d i c a t i v e o f s e r i o u s m o d e l m i s s p e c i f i c a t i o n . F o r f u r t h e r a n a l y s i s ,
w e a g a i n u s e t h e e s t i m a t e s f o r t h e d a i l y G A R C H ( 1 , 1) m o d e l (&(z88˜ = 0 . 1 0 5 a n d
/3(2881 = 0 . 8 7 3 ) as a n a t u r a l b e n c h m a r k s i n c e t h e s e are u n a f f e c t e d b y t h e i n t r a d a y
periodicity. The results of Drost and Nijman (1993) and Drost and Werker (1996)
n o w i m p l y that t h e i n t r a d a y r e t u r n s s h o u l d f o l l o w w e a k G A R C H ( 1 , 1) p r o c e s s e s
w i t h a ( ˜ ) + / 3 ( k ˜ c o n v e r g i n g to u n i t y a n d a(k ) c o n v e r g i n g t o w a r d s z e r o as t h e
l e n g t h o f t h e s a m p l i n g i n t e r v a l , k, d e c r e a s e s . I n fact, N e l s o n ( 1 9 9 0 , 1 9 9 2 )
T.G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158 139
establishes general conditions under which GARCH(1, 1) models, even if misspec
ified at all frequencies, will satisfy the above convergence results and produce
consistent estimates for the true volatility process at the highest sampling frequen
cies. Unfortunately, these predictions do not allow for deterministic effects in the
volatility process. Yet, given the estimated standard errors, the 12 hourly through
2 hourly returns (k = 24144) are roughly in line with the qualitative predictions.
Beyond this point the theoretical results are strongly contradicted, however. The
most blatant violations are provided by the much lower volatility persistence, as
measured by a(˜) +/3(k ), for the models based on ˜  1 71 hourly returns (k = 618).
^ ^ 1
For the 515 minute returns (k < 3) the sum of the estimates for ˜(k) and /3(k) is
again near unity, but the relative size of the coefficients does not conform to the
theoretical predictions, as &(k) is tOO large.
Our intraday results in Table 2a are not unusual. They mirror the range of
estimates previously obtained in the literature over corresponding return frequen
cies. In particular, Engle et al. (1990) and Hamao et al. (1990) who primarily rely
on returns over six hours or longer find evidence of volatility persistence that is
consistent with estimates from daily data. In contrast, Baillie and Bollerslev (1991)
and Foster and Viswanathan (1995), on using hourly and halfhourly returns, find
much lower volatility persistence 32. However, the volatility persistence measures
appear to rebound at the higher frequencies e.g. Bollerslev and Domowitz (1993)
report 5minute GARCH(1, 1) estimates for a(k ) + /3(k) close to one but, as in
Table 2a, &(k) seems too large. For the very highest frequencies, Locke and
Sayers (1993) find that 1minute returns generally display little volatility persis
tence. Conversely, Goodhart et al. (1993) detect very strong persistence in
quotebyquote data, but also find a marked decline in the persistence once
information events are taken explicitly into account, illustrating how specific news
arrivals may overwhelm the underlying conditional heteroskedasticity at the
extremely high frequencies.
Our findings provide strong, albeit indirect, evidence in support of the conjec
ture that a contributing factor to the systematic variation in volatility estimates
across return frequencies is the interaction between the previously well docu
mented interdaily conditional heteroskedasticity and the intraday periodicity. For
the highest frequencies the change in the intraday pattern will generally appear
smooth between adjacent returns, and thus have little impact on the overall
estimated degree of volatility persistence. However, as argued more formally
below, the existence of shortlived intraday volatility components (in addition to
the intraday periodicity) will tend to increase the dependence of (O't,k,)2 on the
32 Interestingly, Laux and Ng (1993) deviate from these studies by finding high persistence in
halfhourly data for the CME currency futures. However, the futures market only operates during the
most active trading in the U.S. segment of the foreign exchange interbank market and this represents a
period of relative stability for the intraday volatilitypattern.
T. G. Andersen, T. BollersleL,/ Journal of Empirical Finance 4 (1997) 115158
140
lagged squared innovation, (etch )2, relative to the overall volatility level,
.
(˜r,!n_ 1)2, hence explaining the relatively large estimates for c˜k) at the shortest
1 1
return intervals. For the intermediate 7  1 ˜ hour return models the change in the
average volatility between sampling intervals will typically appear much more
abrupt, resulting in significantly lower persistence measures. Beyond the 2hour
intervals the periodic pattern is averaged over a substantial part of the 24hour
trading day, and the intraday exchange rate estimates are generally closer to the
implications obtained from daily models.
The results for the S & P 500 equity returns tell a similar story. The interdaily
estimates in Table 3b are again broadly consistent with the a priori predictions
based on the daily G A R C H ( I , 1) model 33. Although the volatility persistence is
higher than for the foreign exchange returns, &˜k)+/3˜k) again displays a general
smooth decline and the explicit persistence measures are fairly stable across the
different return horizons. The discrepancy between the half lives, mean lags and
median lags implied by the intradaily and interdaily returns are even stronger than
for the foreign exchange rate data, however 34. Moreover, the pattern in the
intraday estimates for a<˜) +/3˜k ) reported in Table 2b is again erratic, reaching
lows at the Â½day (k = 40) and 2 0  2 5 minute (k = 4, 5) return horizons, and highs
at the 4 0  5 0 minute (k = 8, 10) and 5minute ( k   1) horizons. We conclude that
the daily G A R C H models conform closely to the theoretical predictions, but the
strong intraday periodic patterns in volatility render the intradaily estimates highly
variable and generally hard to interpret.
5. The dynamics of filtered and standardized intraday returns
This section proposes a general framework for modeling of high frequency
return volatility that explicitly incorporates the effect of the intraday periodicity.
The preceding section suggests that this is a prerequisite for meaningful time
series analysis. Our approach is motivated by the stylized model in Section 3.
While the model almost certainly is overly simplistic, the previous analysis
suggests that the representation does capture the dominant features of our foreign
exchange and equity return series and thus may serve as a reasonable first
approximation.
33 In this case the estimate for the daily MA(1) term equals 0.186, which is highly significant when
judged by the corresponding asymptotic standard error of 0.012. Consequently the GARCH(1, 1)
models for the other frequencies are, at best, approximate representations of the data generating process
(see Drost and Nijman (1993) for a formal analysis).
34 The previous footnote about the deletion of weekend exchange rate returns are even more pertinent
here as both weekend and overnight equity returns are excluded. However, our informal analysis again
found this to be inconsequential (see Andersen and Bollerslev, 1994).
141
T. G. Andersen, T. Bollerslev / Journal of Empirical Finance 4 (1997) 115158
Specifically, consider the following decomposition for the intraday returns,
o"l st,rtZt, n
R,. n = E ( R t , n ) + (7)
U,/2 ,
where E(Rt, n) denotes the unconditional mean, and N refers to the number of
return intervals per day. Notice that this represents a generalization of the model in
Section 3, in that the periodic component for the nth intraday interval, st,,,, is
allowed to depend on the characteristics of trading day, t 35. Given the absence of
any economic theory for stipulating a particular parametric form for the intraday
periodic structure, a flexible nonparametric procedure seems natural. Although no
one procedure is clearly superior, the smooth cyclical patterns documented in Fig.
ñòð. 1 