Properties of high frequency DAX returns: Intraday patterns,

Jumps and their impact on subsequent volatility

Philippe Masset

March 2008
Abstract
This paper analyzes the behavior of the German DAX index intraday returns. We devote
particular attention to three related empirical issues. First we provide an up-to-date char-
acterization of the DAX intraday volatility patterns. They are mostly W-shaped with peaks
at the opening, at 2.30pm and before the closing. We nd some evidence suggesting that the
implied volatility also follows some deterministic patterns over the trading day. Second we
identify jumps in DAX returns. On jump days, they account on average for 15% to 25% of
the daily variance. Jumps also tend to cluster and are not evenly distributed throughout the
trading day. Third we estimate the impact of a price jump on volatility. We consider dierent
proxies for volatility: absolute returns, implied volatility and realized volatility. Our results
indicate that negative jumps trigger a strong upward correction in volatility. This correction
starts just after a jump occured and persists during up to 25 minutes. On the other hand,
positive jumps seem to have a much less signicant impact on volatility. These results hold
for all volatility proxies but they are more signicant when we consider the implied volatility.
JEL classication: G10; G12; G13
Keywords: Intraday patterns; price jumps; implied volatility; realized variance; asymetric
volatility.

Philippe Masset, Department of Finance and Accounting, University of Fribourg / Switzerland, Bd. de
Prolles 90, CH-1700 Fribourg, [email protected].
1
1 Introduction 2
1 Introduction
In this paper, we describe and investigate some phenomena and regularities that emerge when
observing high-frequency DAX returns. We do not focus on the sign of returns but rather on
their amplitude (or volatility). The sign of returns should indeed be unpredictable according to
the Ecient Market Hypothesis (Fama (1970)).
1
Furthermore, empirical evidence suggests that
their amplitude displays some regularities (Cont (2001) and Granger & Poon (2002)).
The amplitude of intraday returns mainly depends on the interaction between two driving vari-
ables: the trading activity and the volume of information that reaches the market at a particular
moment of the day. The rst mostly determines the speed at which prices are updated, while
the second has a direct eect on the amplitude of individual price changes. There is obviously
a link between these variables. In the absence of signicant news, less traders are active on the
market. Similarly, the way new information is reected in the prices depends on the number of
active traders. There is a vast literature that directly models either the time duration between
trades
2
or the process by which equilibrium prices are reached
3;4
. Unlike these, we adopt what
can be considered as a reduced form approach and directly study some particular features of
high-frequency price changes. Our aim is to characterize the so-called volatility intraday patterns
on the German market, to identify price discontinuities (jumps) and to analyze their statistical
properties and their impact on volatility.
The way trading activity spreads throughout the day is our rst object of study. Trading activity
and the amplitude of price changes are linked together: in general, the higher the activity at a
particular moment of the day, the higher is the intraday variance at that moment. The literature
on intraday volatility patterns traces back to Wood et al. (1985) and was mostly concerned
with US equity markets and FX markets as these are open 24 hours a day.
5
These studies
have shown that the patterns are typically U-shaped. The highest level of trading activity is
reached just after the opening and before the market closes. During lunchtime, there are only
few active traders and therefore the variance in this period remains rather low. The existence
of these regularities have both theoretical and practical consequences. Notably, it is crucial to
lter out intraday patterns from raw returns before trying to t a parametric volatility model;
otherwise the estimates might be severely biased (Andersen et al. (1999)). Moreover, intraday
traders have concerns about very short-term moves and are obviously aected by higher intraday
volatilities as they may lead to rapid losses.
1
This is not exactly the case for high-frequency returns as they are known to exhibit a signicant negative
rst-order autocorrelation. See, e.g., Andersen & Bollerslev (1997a) or Dacorogna et al. (1997).
2
Two recent contributions to this literature are Fernandes & Grammig (2005) and Meitz & Tersvirta (2006).
3
See, e.g., the survey article of Madhavan (2000).
4
A related literature, which traces back to Clark (1973), is concerned with the so-called mixture of distribution
hypothesis. See also Tauchen & Pitts (1983) and Bauwens et al. (2006).
5
See Wood et al. (1985) and Harris (1986) for some early discussions on intraday patterns for equity markets; a
recent contribution is Tian & Guo (2006), who provide an up-to-date literature review and also study intraday
patterns on the Shanghai Stock Exchange. To the best of our knowledge, the only articles studying intraday
patterns on the German market are Kirchner & Schlag (1998) and Ozenbas et al. (2002). For FX markets,
see, e.g., Muller et al. (1990), Baillie & Bollerslev (1991), Dacorogna et al. (1993) and Andersen & Bollerslev
(1997b).
1 Introduction 3
New information impacts the market dierently depending on its importance and whether it is
expected or not. For instance, there are recurrent news that reach the market each day with
a regular timing and deliver essentially the same kind of information. Notably, the overnight
evolution of US and Asian markets has an impact on the opening of European markets. Similarly,
the mid-afternoon period is characterized by the release of information from the USA and the
opening of the NYSE and the Nasdaq. Because of its recurrent nature, this kind of information
directly aects the intraday volatility patterns. On the other hand, there are many unexpected
pieces of information about rms, economic situation or investor sentiments, that have a unique
and important eect on prices. The release of such news often generates price jumps. Our second
objective is to identify and characterize those jumps on the German market. We make use of
recent advances in the study of so-called stochastic volatility semimartingale (SVSM) processes
to identify the jumps.
6
We then study their occurrences, size and timing. Jumps are typically
located in the tails of the return distribution and are thus very relevant for risk management
purposes. They also have a strong impact on the ability to properly hedge a derivative position
as it is not possible anymore to replicate it perfectly.
Our third objective is to analyze the relationship among returns and changes in volatility when
there are price discontinuities. Knowledge of this relationship is a key ingredient for eective
option pricing and it might also help establishing innovative trading strategies. In the continuous
case (i.e. in the absence of jumps), this relationship has already attracted much interest from
researchers. Volatility changes have typically been found to be negatively and asymmetrically
related to returns (Black (1976) and Christie (1982)). Two theories have been advanced for
explaining this eect: the leverage and the feed-back explanations. The rst assumes that past
returns have a predictive power on futures volatility changes, while the other implies the opposite
chain of reactions.
7
The causal relationship remains mostly unclear when considering daily data
(see Bouchaud et al. (2001)). When switching to a high-frequency framework, one observes
that the relation is mostly return-driven. Previous returns have an inverse impact on subsequent
volatility changes (Bollerslev et al. (2006) and Masset & Wallmeier (2007)). The relationship
between price jumps and subsequent volatility changes has not yet been investigated. Though
an often encountered (sometimes explicit but mostly implicit) assumption is that both price
and volatility jump simultaneously. We test if this assumption holds. In particular, we analyze
the impact of a price discontinuity on contemporaneous and subsequent volatility changes. If a
jump is expected to have a persistent impact on the magnitude of subsequent returns and if the
investors are fully rational, the implied volatility should jump (almost) simultaneously.
For our inferences, we make use of a high-quality database, which covers all transactions involving
options and futures on the German market from January 1995 to December 2005.
8
There is a
one-to-one relationship between future prices and actual index levels; hence we can estimate DAX
levels with a very high level of precision. This is obviously crucial for the analysis of intraday
6
See, e.g., Andersen et al. (2001b), Andersen et al. (2001a), Barndor-Nielsen & Shephard (2002), Andersen
et al. (2006b) and Bandi & Russell (2007).
7
For a detailled discussion of these two theories, see, e.g., Bekaert & Wu (2000), Bollerslev et al. (2006) and
Masset & Wallmeier (2007).
8
Over this period, DAX futures and options have represented the highest trading volume among all stock index
derivatives in Europe.
2 Data 4
patterns and the identication of price discontinuities. We compute the implied volatility from
option prices. Implied volatilities present several advantage over other volatility proxies (like
squared returns or realized variance). They do not depend on actual index returns and they
allow us to accurately determine the point in time when changes in volatility occur. It is thus
possible to gauge precisely the impact of a price jump on the volatility process.
This paper contributes to the existing literature in several ways. First, we provide a precise and
up-to-date description of the intraday patterns for squared returns and for implied volatility
on the German market. The patterns for squared returns are mostly \shaped: returns are
especially large at the opening, at 2:30 p.m. and at the closing. We also nd clear-cut patterns
for the implied volatility; they have a Jshape: implied volatility is typically lower during the
lunch than during the rest of the day and it usually reaches a high when the market closes.
Second, we characterize the occurrences of jumps and their statistical properties. On jump
days, price discontinuities account on average for 15% to 25% of the daily return variance and,
on some days, a jump might even explain as much as 80% of the daily variance. As already
noticed by Andersen et al. (2007), jumps tend to cluster. Furthermore, the probability of a jump
seems to be higher in the early morning, at 2:30 p.m. and in the late afternoon. Third, we show
that the impact of price discontinuities on both implied and realized volatility is considerable.
Negative jumps trigger a strong increase in volatility. In particular, about 20 to 25 minutes are
needed by the implied volatility process to fully incorporate the eects of such a jump. On the
other hand, positive jumps have only a limited impact on volatility.
Our paper is organized as follows. In the next section, we describe our dataset and discuss some
issues related to high-frequency frictions. The methodology used to estimate intraday patterns
and to identify price discontinuities is presented in section 3. Intraday patterns for squared
returns and for implied volatility are discussed in section 4. In section 5, we study the number
of jumps, their importance and their timing. Section 6 is devoted to the analysis of the impact
that a jump has on volatility. Section 7 concludes.
2 Data
Our data come from the joint German and Swiss options and futures exchange, Eurex.
9
The
Eurex is the worlds largest futures and options exchange and is jointly operated by Deutsche
Brse AG and SWX Swiss Exchange. The database contains all reported transactions of options
and futures on the German stock index DAX from January 1995 to December 2005. The average
daily trading volume of DAX options (ODAX) and futures (FDAX) in December 2005 was of
166, 886 and 117, 388 contracts, respectively. The options are European style. At any point
in time during the sample period, at least eight option maturities were available. However,
trading is heavily concentrated on the nearby maturities. The contract values amount to 5 euros
(ODAX). Trading hours changed several times during our sample period, but both products were
traded at least from 9:30 a.m. to 4:00 p.m.
9
We are very grateful to the Eurex for providing the data.
2.1 Computation of DAX and implied volatility levels 5
7-Mar-2000 12-Mar-2003
2000
3000
4000
5000
6000
7000
8000
9000
D
A
X
Figure 1: Period denition.
We split the observation period in eight subsamples according to two criteria: (i) market con-
ditions (bull or bear) and (ii) trading hours. This has no impact on the identication of jumps
because the test statistics used to detect jumps consider each day separately. It also makes no
sense to use data that have dierent opening or closing hours when estimating intraday patterns.
Furthermore, this gives us the opportunity to test if intraday patterns and jumps share the same
features in dierent market conditions. Between 1995 and 2005, we identify three broad periods
of either bullish or bearish market, periods 1 to 3 in gure 1. The rst period corresponds to
the long bull market between 1995 and 2000, which ended with the burst of the tech bubble on
March 10
th
, 2000. The second period, which covers years 2000 to 2003, saw the market suer
severe losses, caused by the post-bubble correction and was further reinforced by the September
11
th
terrorist attacks. The beginning of the Iraqi war on March 20
th
, 2003 marked the beginning
of a new bull market, which is still running (as of December 2005). We nally split once again
each period such that all observation days in a subperiod have the same trading hours.
10
2.1 Computation of DAX and implied volatility levels
We use the relationship between futures and spot prices to estimate the underlying DAX index
levels.
11
Similarly, we calculate the implied volatility from the Black & Scholes (1973) formula.
In order to obtain correct implied volatilities, it is crucial to accurately match the option price
and the corresponding underlying price. As we use time-stamped tick-by-tick data, matching of
option and DAX index levels is straightforward. We apply the method of Hafner & Wallmeier
(2001) to account for dividend eects and to ensure put-call-parity consistent estimates of implied
volatilities. We remove all options that violate the arbitrage bounds or have implied volatilities
10
The start and end dates of each subperiod are the followings: 01/01/1995 - 05/19/1995 (Period 1A), 05/22/1995
- 03/27/1997 (1B), 04/01/1997 - 09/17/1999 (1C), 09/20/1999 - 03/07/2000 (1D), 03/08/2000 - 06/01/2000
(2A), 06/02/2000 - 03/12/2003 (2B), 03/13/2003 - 11/18/2005 (3A), 11/21/2005 - 12/30/2005 (3B).
11
We use futures to infer DAX levels as some evidence suggests that futures should react more rapidly to news
than the underlying (Hasbrouck (2003)). Furthermore, there is a direct link between options and futures as
the latter are usually used to hedge the rsts (Liu et al. (2007)).
2.1 Computation of DAX and implied volatility levels 6
higher than 150% (Hafner & Wallmeier (2001)). We consider only options with a time-to-
maturity longer than 10 calendar days to avoid expiration-day eects and because estimates for
implied volatility are known to be unstable for shorter maturities.
12
Options with a very distant
maturity are much less liquid and it is dicult to account properly for term structure eects.
Therefore we keep only options that have a time-to-maturity no longer than 3 months. There
are several diculties to accurately estimate the implied volatility of deep in-the-money (ITM)
and deep out-of-the-money (OTM) options. Such options are prone to be aected by errors in
measurement (Hentschel (2003)). Therefore we do not use options with a moneyness lower than
0.85 or larger than 1.15. This moneyness interval is relatively wide as we want to have the
largest number of valid observations in our sample.
Implied volatilities are not constant across moneyness levels and also vary with the time-to-
maturity. OTM put options typically have higher implied volatilities than at-the-money (ATM)
put options. This phenomenon is known as the volatility smile or volatility smirk in the nance
literature (see, e.g., Rebonato (2004)). The existence of an implied volatility term structure is
a related issue: the implied volatility tends to decrease when the option approaches maturity.
Consequently, one cannot directly use the implied volatilities estimated from options with dier-
ent features (maturity and strikes) to construct a homogeneous time-series of implied volatility
for the underlying. This implies that a change in implied volatility between two subsequent
trades can either be due to a change of its fundamental level or to a dierent moneyness or
time-to-maturity. Hereafter, we detail how we proceed to rule out smile and term structure
eects.
Smile correction
We rst estimate the smile structure each day following the cubic regression approach described
in Hafner & Wallmeier (2001) and Hafner & Wallmeier (2007) and, then, we use the tted
smile function to remove the impact of moneyness on implied volatilities. More specically, let
1 denote the strike price of an option with time to maturity T t. Each trade is assigned a
moneyness according to:
'(t, T, 1) =
ln
_
K
F
t
(T)
_
_
T t
,
where 1
t
(T) is the forward price at time t for maturity T. Thus, ATM options are characterized
by a moneyness of 0. Suppressing the arguments of moneyness, we chose the cubic regression
function:
o = ,
0
+,
1
' +,
2
'
2
+,
3
1 '
3
+-, (1)
where o is the implied volatility, ,
i;
i = 0, 1, 2, 3 are regression coecients, - is a random error,
and 1 is a dummy variable dened as:
1 =
_
1 , ' 0
0 , ' _ 0
.
12
These points are discussed in De Jong & Donders (1997). In other studies, the last 7 to 10 days are usually
removed. For instance, Dennis et al. (2006) removed options with a maturity of one week or less; Masset &
Wallmeier (2007) took only options with at least 10 days to maturity.
2.2 High-frequency frictions 7
The dummy variable accounts for an asymmetry of the pattern of implied volatilities around
the ATM strike (' = 0).
Let o
imp
(', t) denote the implied volatility of an option with moneyness ' traded at time t.
Then, the corresponding ATM implied volatility o
ATM
imp
(t) is calculated as
o
ATM
imp
(t) = o
imp
(', t)
_

,
1
' +

,
2
'
2
+

,
3
1 '
3
_
, (2)
where

,
i
are the estimated regression coecients.
Term structure correction
We estimate the term structure of implied volatility using the linear regression model:
13
o
ATM
imp
(t) = c
0
+c
1
(T t) +c
2
1
2
(T t) +n, (3)
where c
i;
i = 0, 1, 2 are regression coecients, n is a random error, and 1
2
is a dummy variable
dened as:
1
2
=
_
1 , T t 60
0 , otherwise
.
The dummy variable gives more exibility to the specication as it allows the trend of the term-
structure to be dierent for options with the longest time-to-maturity. The implied volatility
corresponding to a 30-day constant time-to-maturity option can then be calculated as:
o
ATM;CTtM
imp
(t) = o
ATM
imp
(t) [ c
1
(T t 30) + c
2
1
2
(T t 30)] , (4)
where c
i
are the estimated regression coecients.
2.2 High-frequency frictions
Previous operations leave us with two time-series, one of implied volatility and another of DAX
levels. Both series are recorded at irregular time intervals. In order to get homogeneous time-
series, we resample the data by aggregating them over a particular sampling frequency. There
is always a trade-o between the amount of information that can be inferred from the data
and the noise that they might contain (see Hansen & Lunde (2006)). In order to keep as much
information as possible from the original data, the highest possible sampling frequency should
be considered. However, this is at a cost as it exacerbates the impact of high-frequency frictions
and thus augments the noise-to-signal ratio.
14
DAX futures are very liquid and have a small
13
We also consider non linear models but the nal results remain mostly unaltered.
14
For a discussion of optimal sampling procedures, see, e.g., Ait-Sahalia et al. (2005), Zhang et al. (2005) and
Bandi & Russell (2008).
2.2 High-frequency frictions 8
bid-ask spread. Thus, a very high sampling frequency could be selected. Unfortunately, DAX
options are less liquid and have a larger bid-ask spread. We therefore decide to sample the
data each 5 minutes. This consensual choice is indeed quite common in the literature (see, e.g.,
Bollerslev et al. (2006)).
The estimated DAX level for each 5-minute interval is set equal to its average over the last
minute of the interval. In very few cases, there is no observation in the last minute of an interval
and thus the DAX level cannot be estimated, implying that the corresponding return cannot
be computed. This is an issue as the procedure we employ to identify price discontinuities (see
section 3.2) requires a complete series of returns. In order to solve this, we complete the original
series and estimate the prevalent DAX level by interpolating between the average DAX levels
over the previous and the next minute with valid observations.
15
As stated above, options are
less liquid than futures: in many intervals, there is no observation at all and consequently it
is not possible to calculate an implied volatility. We do not replace these missing values for
two reasons. First, we do not necessarily need a complete series of implied volatilities as they
are not required for identifying price jumps. Second, when analyzing the relationship among
DAX jumps and volatility changes, we want to avoid spurious results caused by the presence of
ctitious (interpolated) values. Consequently, to conduct our inferences in the next sections, we
employ only the intervals in which we have a valid observation for the implied volatility.
Log-returns of implied volatility and of the underlying stock index are calculated as:
1
v;t
j
= ln
_
o
ATM
imp
(t
j
)

ln
_
o
ATM
imp
(t
j1
)

and 1
S;t
j
= ln [o(t
j
)] ln[o(t
j1
)] ,
where o
ATM
imp
(t
j
) is the implied volatility and o(t
j
) denotes the index level in the ,-th 5-minute
interval on day t. 1
v;t
j
is only calculated if we have an observation at times t
j
and t
j1
. We
remove overnight returns (Gwilym & Buckle (2001)) and the rst return of each day (Bollerslev
et al. (2004)).
We eventually devote some consideration to the impact of the bid-ask spread on returns. As is
well known, the bid-ask bounce leads to a negative rst order autocorrelation of returns (Roll
(1984)). This spurious autocorrelation comes from successive trades, where one is executed
at the bid, the other at the ask price. As already stated, futures have a narrower bid-ask
spread than options. It is thus not surprising to nd out that this eect is very pronounced
for volatility returns but absent from DAX returns.
16
We use the standard method to remove
spurious autocorrelation from volatility returns only. This method consists of ltering returns
with a MA(1) process (see, e.g., Stephan & Whaley (1990), Easley et al. (1998) and Gwilym &
Buckle (2001)):
1
v;t
j
= j
v
+c
v;t
j
0
v
c
v;t
j1
,
where 1
v;t
j
is the observed implied volatility return, j
v
the unconditional mean of the observed
returns series, 0
v
the moving average coecient, and c
v;t
j
the innovation of the process. Since
the innovations from the MA(1) process are uncorrelated, we can use them as bid-ask bounce
corrected returns. Through this correction, raw implied volatility returns 1
v;t
j
are transformed
into adjusted log-returns r
v;t
j
, i.e. r
v;t
j
= c
v;t
j
.
15
A robustness check shows that this procedure has no impact on the estimation of the intraday patterns.
16
First order autocorrelation coecients are 0:4159 and 0:0170 for volatility returns and DAX returns.
3 Estimation of the intraday patterns and jump identication 9
3 Estimation of the intraday patterns and jump identication
3.1 Methodology for estimating intraday patterns
Intraday patterns are just a graphical and quantied representation of how the activity on the
market uctuates during the trading day. The level of activity depends on the number of active
traders and the quantity and type of information reaching the market at a particular time. When
there are more active traders, prices change more quickly and the returns variance computed
over a xed-length interval increases.
17
This implies that intraday patterns can be estimated on
the basis of the amplitude of price changes computed at dierent moments of the day. These
patterns can also be considered as a time deformation in the sense of Clark (1973) as they
deliver some information about the speed at which business time ows. When dealing with a
return series sampled in the usual clock time, it is crucial to lter intraday patterns out from
raw returns in order to avoid biases in the modelling of conditional volatility (Andersen et al.
(1999)).
To study and estimate intraday patterns, their inuence has to be isolated from the original
series of returns. A usual assumption is that observed high-frequency returns can be split into
two parts: an unpredictable (innovation) component and a volatility component. The volatility
component itself is the product of both the prevalent daily volatility and an intraday volatility
factor. The return in interval , on day t can thus be expressed as:
1
t
j
= o
t
`
j
.
t
j
, (5)
where o
t
is a measure of the latent volatility on day t, `
j
is a factor that accounts for the
proportion of a tradings day return variance that is attributed to interval ,, and .
t
j
is an
innovation term with mean zero and variance one. Intraday patterns should reect some typical,
inherent, characteristic of the activity ow during trading hours; therefore they are not expected
to change from one day to another. We dene standardized returns r
t
j
as r
t
j
=
R
t
j

t
, where o
t
can be estimated either from a parametric model (e.g. a Stochastic Volatility or a GARCH
model) or a non-parametric one (e.g., the realized volatility, see section 3.2). Assuming no
covariation among o
t
, `
j
and .
t
j
, the variance of 1
t
j
can be calculated as \ ar(1
t
j
) = o
2
t
`
2
j
(as \ ar(.
t
j
) = 1) and the variance of r
t
j
is:
\ ar(r
t
j
) = `
2
j
. (6)
Further, .
t
j
has mean zero and thus the variance can be equally computed from:
\ ar(r
t
j
) = 1(r
2
t
j
) =
1
N

t=1
r
2
t
j
= `
2
j
. (7)
17
Unless returns from successive trades cancel out, the overall price variation will indeed be higher. Note also
that this does not mean that the volatility of a single trade should increase.
3.1 Methodology for estimating intraday patterns 10
This expression suggests a simple way for estimating the intraday volatility factors. As an
alternative, one may instead focus on the expression

r
t
j

, which is less inuenced by extreme

values. If .
t
j
comes from a normal distribution, we have that 1(

r
t
j

) =
_
2

`
j
and thus we get

`
j
=
_

2
1(

r
t
j

) as an estimate for `
j
. Nevertheless, the assumption that the innovations
.
t
j
are normally distributed is very disputable. Therefore, we opt for an estimation of the
intraday patterns based on (7).
The summation in (7) considers a set of successive days; the larger , the more precise the
estimate for

`
j
will be. Though, if the patterns are not stable through time
18
, the complete
sample has to be split into homogeneous subsamples. Eventually, the factors estimates remain
noisy and sensitive to outliers. Some erratic changes in the patterns when moving from an
interval , to the next interval , + 1 do not have much economic content. They can be caused
by some large returns in a few days, either in interval , or , + 1. Therefore, it is preferable to
smooth the original estimates. Andersen et al. (2001c) recommend Fourier exible functions
(FFFs):
19
`
j
=
K

k=0
j
k
,
k
+
D

i=1
c
i
1
j=d
i
(8)
+
P

p=1
[c
p
cos(
2,j

) +c
p
sin(
2,j

)].
The FFF model (8) consists of three parts. The rst is based on a polynomial structure,
which aims at modelling the general trend of the patterns over the trading day. If assuming
a unique homogeneous trend for the whole trading day seems too unrealistic, the polynomial
can be broken in several subpolynomials in order to account for dierent trading regimes. The
second part shows a summation of dummies; their purpose is to model particular phases of the
trading day which exhibit unusual characteristics and are thus dicult to capture through the
polynomial part (e.g. the opening or the closing of the market). The last part is based on a
series of sinusoidal functions, which work best when the series to be modelled exhibits cyclical
behaviors. To obtain a smooth estimate of the intraday patterns, we further add an error term
on the right hand-side of model (8) and then estimate the complete model by Ordinary Least
Squares (OLS). Robust standard errors for the coecients can be obtained using the Newey &
West (1987) method.
20
18
This would be the case either if the trading hours in the home country or abroad (e.g. in the USA) or if the
investors trading habits have changed over time.
19
See also Andersen & Bollerslev (1997b).
20
Dacorogna et al. (1993) propose another way to deal with intraday patterns. They suggest resampling the
data in so-called business time. In their setting, time-intervals are dened as some proportion of the daily
returns quadratic variation. This procedure mechanically removes the eects of intraday patterns.
3.2 Realized variance and identication of jumps 11
3.2 Realized variance and identication of jumps
Volatility is said to be latent as it is not directly observable. Nevertheless, recent advances in the
theory of quadratic variation have demonstrated that it can be estimated in a robust manner.
21
The realized variance on day t is dened as:
1\
t
=
M

j=1
1
2
t
j
, (9)
where 1
t
j
is the ,-th intraday return on day t. In a given day, the number of intervals sums to
'. If this number tends to innity (' ), i.e. if the sampling frequency is increased such
that the interval between two return observations become innitesimal, the realized variance
becomes asymptotically equivalent to the quadratic variation of the process:
1\
t

M!1
Q\
t
, (10)
where Q\
t
is the quadratic variation of the process. Furthermore, if the underlying does not
exhibit any discontinuity, Q\
t
will be an unbiased estimator of the integrated variance, 1\
t
:
1\
t
=
t
_
t1
o
2
(:)d:. (11)
However, with jumps in the process, Q\
t
will equal the integrated variance plus the sum of
squared jumps, i.e.:
Q\
t
=
t
_
t1
o
2
(:)d: +
N
t

n=1
i
2
t
n
, (12)
where
t
is the number of jumps on a given day (
t
_ ') and i
t
n
is the size of the :-th jump.
In order to isolate the jump component from the quadratic variation, we need an estimator of
the integrated variance, which remains consistent even in the presence of jumps in the process.
Barndor-Nielsen & Shephard (2004) propose another measure, namely the bipower variation
(as opposed to the simple power or quadratic variation):
1\
t
= j
2
1
M

j=2

1
t
j1

1
t
j

, (13)
where j
1
=
_
2

= 1([7[) denotes the mean of the absolute value of standard normally distrib-
uted random variable 7.
21
See, e.g., Andersen et al. (2001b), Andersen et al. (2001a), Barndor-Nielsen & Shephard (2002), Andersen
et al. (2006b) and Bandi & Russell (2007).
3.2 Realized variance and identication of jumps 12
More generally,
j
a
=
2
a=2
(1,2(a + 1))
(1,2)
= 1([7[
a
), (14)
where is the Gamma function. Barndor-Nielsen & Shephard (2004) further prove that:
1\
t

M!1
1\
t
. (15)
The intuition behind this result is the following (Huang (2004)). A jump might occur either in
t
j1
or in t
j
but not in t
j1
and in t
j
. This is because the number of jumps in a given day
is nite. Hence the probability to observe two successive jumps is zero when the number of
partitions goes to innity (i.e. when ' ). Jumps may still enter into the calculation of
the bipower variation through the cross-product

1
t
j1

1
t
j

; nevertheless, their contribution

to the 1\
t
process also becomes negligible when ' goes to innity. This is the case for two
reasons: (i) there is only a nite number of jumps a day and (ii) the magnitude of the returns
from the continuous part of the process tends to zero when ' goes to innity. Consequently, as
the number of partitions increases (' ), the contribution of the jump component to 1\
t
vanishes. Moreover, when the sampling frequency is very high (i.e. when ' ), returns
from successive intervals become increasingly close to each other and, therefore, 1\
t
consistently
estimates the integrated variance.
The result (15) implies that one can estimate either the part of the variance due to the continuous
volatility component or the part due to the jump component by considering 1\
t
or by taking the
dierence between 1\
t
and 1\
t
.
22
Obviously, this dierence, computed from empirical data,
can be negative. This is counter-intuitive as the variance has to be positive. Thus we set the
daily variance due to jumps J\
t
equal to max(0, 1\
t
1\
t
).
23
Using the procedure described above, one invariably nds that there are frequent but mostly
small or even negligible jumps in the return process. To test if a signicant jump occurred
during a given day we use the test statistics derived by Barndor-Nielsen & Shephard (2004),
Barndor-Nielsen & Shephard (2006) and Andersen et al. (2007). We adopt the logarithmic
test specication of Huang & Tauchen (2005) and Andersen et al. (2006a):
\
t
=
ln1\
t
ln1\
t
_
1
M
(j
4
1
+ 2j
2
1
5) max(1,
TP
t
BV
2
t
)
, (16)
and T1
t
is the realized quarticity, which can be estimated consistently using either a tri- (An-
dersen et al. (2007)) or a quadripower quarticity (Barndor-Nielsen & Shephard (2004)). We
focus on the tripower quarticity, which is estimated through the following formula:
22
It is possible to estimate the variance of the Brownian increments using multipower variations:
PV
t
=
2
p
M
P
j=2=p
2=p
Q
k=0

R
t
jk

p
.
As long as the returns are raised to a power inferior to two (p < 2), PV
t
will be a consistent estimator of the
variance of the continuous part of the process (Barndor-Nielsen & Shephard (2004)).
23
Barndor-Nielsen & Shephard (2004) and Andersen et al. (2007) adopt the same adjustment.
4 Intraday volatility patterns 13
T1
t
= 'j
3
4=3
M

j=3

1
t
j2

4=3

1
t
j1

4=3

1
t
j

4=3
. (17)
Under the null of absence of jumps, the test statistics \
t
from (16) is asymptotically standard
normally distributed.
4 Intraday volatility patterns
4.1 Intraday patterns for squared returns
The intraday volatility factors

`
j
have been estimated on the basis of equations (5) and (6). They
are reported in gure 2. Full lines show the t from the FFF regression model (8), which has
been specied as follows. First, the original polynom has been broken into three subpolynoms
to account for dierent trading regimes, which correspond to (1) the morning; (2) the lunch
time and the rst part of the afternoon (i.e. before US-opening); and (3) the rest of the day.
Second, two dummies have also been added to the regression model in order to account for the
rst and the last interval of the trading day. The order of the polynomials has been set to ve.
This ensures a satisfactory trade-o between the smoothness of the tted intraday patterns and
the risk of overtting.
24
As in Andersen & Bollerslev (1997b),we leave out the sinusoidal part
in (8).
Many news are released before the market opens. As a consequence, the rst minutes are highly
nervous. The activity remains intense during the next few intervals and then begins to slightly
decrease. It typically reaches a low between 12:00 a.m. and 1:00 p.m. In Europe, the mid-
afternoon is usually rather nervous because of information releases from the USA. In particular,
news about the coming US-opening leads the intraday volatility to skyrocket between 2:30 p.m.
and approximately 2:45 p.m. After that, the market calms down for a short while. When the
closing time is approaching, investors trade again very intensively. This eventually brings the
intraday volatility to new highs.
The patterns we get for the dierent periods look very similar. The precision of the tted patterns
obviously depends on the number of observation days in each period. It seems that intraday
volatility patterns are neither aected by investor sentiments nor by market conditions. For
instance, period 2-B was strongly bearish (the index lost about 70% of its value), while period
3-A was extremely bullish (the index more than doubled during this period). Though the
patterns for these two periods remain very similar.
24
For instance, Andersen et al. (2001c) use a third order polynomial. Nevertheless, they use FX data, which
display much simpler and smoother U-shaped patterns. As a matter of comparison, we also run the regression
models with third order polynomials. Overall very similar shapes emerge but the ability of the model to t
peaks in activity is severely weakened.
4.2 Intraday patterns for implied volatility 14
10 12 14 16
0.8
1
1.2
1.4
1.6
Period 1-A
10 12 14 16
0.8
1
1.2
1.4
1.6
Period 1-B
10 12 14 16
1
1.5
Period 1-C
10 12 14 16
0.6
0.8
1
1.2
1.4
1.6
Period 1-D
10 12 14 16
0.6
0.8
1
1.2
1.4
1.6
Period 2-A
10 12 14 16 18 20
0.6
0.8
1
1.2
1.4
1.6
1.8
Period 2-B
10 12 14 16 18 20
0.6
0.8
1
1.2
1.4
1.6
1.8
Period 3-A
10 12 14 16 18 20 22
0.4
0.6
0.8
1
1.2
Period 3-B
Figure 2: Intraday patterns for squared returns. The dierent panels show the patterns
for each period of the sample.
4.2 Intraday patterns for implied volatility
As a matter of comparison, we check if the implied volatility might also display some kind of
intraday patterns. We consider each period separately and proceed as follows. For each day,
we estimate a midday implied volatility level, which is dened as the average implied volatility
observed between 12:00 a.m. and 1:00 p.m. This midday estimate is then used to standardize
the implied volatility levels in the other 5-minute intervals of the day. This standardization
allows to directly compare the implied volatility measured in the same interval in dierent days.
It thus permits to calculate for each period an average standardized implied volatility for each
5-minute interval.
Figure 3 displays the intraday patterns for implied volatility. The full line shows the t from a
fth order polynomial regression (without any dummy or break). The implied volatility tends
to decrease just after the opening. It then stabilizes during the second part of the morning
and nally raises again during the whole afternoon. These patterns are clear-cut but remain
less pronounced than the ones we get for the squared returns: the highest point of the intraday
patterns is about 1% to 1.5% larger than the lowest one. Nevertheless this observation might
4.2 Intraday patterns for implied volatility 15
have important implication for option pricing as it means that, everything else held constant,
the same option could well be cheaper when bought between 10:00 a.m. and 11:00 a.m. than in
the late afternoon. In particular if the option is OTM, the price dierence might be substantial.
Consider for instance a call option written on an index that currently quotes at 8, 500 points.
The interest rate amounts to 3.5% yearly. The option strikes at 9, 000 points and has a time-
to-maturity of 30 calendar days. The implied volatility level is of about 20% at 10:00 a.m.
According to the Black and Scholes formula [1973], the option price may increase by as much
as 14% after a 1% increase in implied volatility (i.e., from 20% to 20.2%).
In order to quantify the importance of this eect from an economic viewpoint, we build a
trading strategy based on a strangle. This is an indirect way to invest in volatility. We purchase
simultaneously both an OTM call and an OTM put at 10:00 a.m. The position is closed at 4:00
p.m. This strategy has been implemented over the last ve years of the sample. By the end of
2005, we achieve a raw return of about 0.62% per day. This number is much larger than the
average daily return of the DAX over the same period but it remains lower than the eective
transaction costs on the EUREX; in particular, the bid-ask spread is typically larger than 1%
for OTM options.
10 12 14 16
0.998
1
1.002
1.004
1.006
Period 1-A
10 12 14 16
0.995
1
1.005
Period 1-B
10 12 14 16
1
1.005
1.01
Period 1-C
10 12 14 16
0.996
0.998
1
1.002
1.004
1.006
Period 1-D
10 12 14 16
0.995
1
1.005
Period 2-A
10 12 14 16 18 20
1
1.01
Period 2-B
10 12 14 16
0.998
1
1.002
1.004
1.006
1.008
Period 3-A
10 12 14 16
0.995
1
1.005
Period 3-B
Figure 3: Intraday patterns for implied volatility. The dierent panels show the patterns
for each period of the sample.
5 Price discontinuities 16
5 Price discontinuities
5.1 Number and importance of jumps
Table 2 reports descriptive statistics of jump occurrences. In order not to confuse a genuine jump
with a burst of activity, intraday patterns have been preliminary ltered out from raw returns.
We follow Huang & Tauchen (2005) and consider primarily the 99% and 99.9% condence levels
to identify price discontinuities. As a matter of comparison, we also use the 99.99% level (Lahaye
et al. (2007)). Obviously the number of jumps decreases when we consider a higher condence
level. For instance, at the 99% and 99.9% condence levels, 346 and 81 jumps are identied. At
the highest condence level, we can still identify 22 individual jumps and 20 jump days. These
results are in line with those of Lahaye et al. (2007).
25
When the lowest condence level is
considered, there is quite a larger number of days in which more than one jump took place.
Period Signif. # jumps # days with jumps Max. Min. Mean Std.
99% 119 80 (5.96%) 0.87% -0.90% -0.0346% 0.3491%
1 99.9% 29 26 (1.94%) 0.87% -0.90% -0.0164% 0.4700%
99.99% 7 5 (0.37%) 0.43% -0.78% -0.1552% 0.4854%
99% 84 51 (6.70%) 1.16% -2.17% -0.1079% 0.5522%
2 99.9% 16 12 (1.58%) 0.45% -0.97% -0.3512% 0.4328%
99.99% 3 3 (0.39%) 0.36% -0.97% -0.3412% 0.6655%
99% 143 80 (11.20%) 0.86% -1.74% 0.0309% 0.3151%
3 99.9% 36 31 (4.34%) 0.86% -1.74% -0.0284% 0.4513%
99.99% 12 12 (1.68%) 0.63% -1.74% -0.2112% 0.6329%
Table 2: Descriptive statistics about jump occurrences. For each period and condence
level (Signif.. in the table), the number of jumps and the number of days with jumps are
reported. The two columns called Max. SJ and Min. SJ show the most positive and most
negative returns due to a single jump. The last two columns, Mean and Std. report the
mean and the standard deviation of all returns due to jumps.
On average, returns due to jumps are slightly negative and their standard deviations are large
(between 50% and 120% in annualized terms, depending on the period under consideration and
the signicance level). There are some striking dierences between the three periods. Notably,
there are more days with jumps in the last period than in the other two. The second period
was strongly bearish and highly volatile, while the last one was bullish and rather calm.
26
This
dierence seems therefore to be counterintuitive. Nevertheless, in a high volatility environment,
relatively larger price changes are needed for the test statistics in (16) to become signicant.
25
Lahaye et al. (2007) study the number of jump occurences for the Dow Jones, the Nasdaq and the S&P
500. They use data sampled each 15 minutes and consider a 99:99% condence level. They nd out that the
proportion of jump days for these indices was respectively 1:43%, 0:7% and 1:70%.
26
The average intradaily (i.e. without considering overnight returns) realized volatility was 15.22% in the rst
period, 25.04% in the second, and 14.94% in the last.
5.2 Timing of jumps 17
These results might indicate that the prices declined rather steadily during period two and that
only a few returns were driven by sudden price jumps. Moreover, rare jumps were on average
bigger during period two than during the other periods. When looking at the Max. and Min.
columns of Table 2, one may also notice that the amplitude of the largest jumps tend to decrease
when a higher condence level is considered. This means that the largest price changes are not
necessarily driven by jumps.
Table 3 shows the contribution of jumps to the overall daily variance. When considering all
days (i.e. also non-jump days), their contribution remains quite restrained. But if we focus only
on those days in which at least one jump takes place, the picture changes dramatically. In this
case, about 15% to 40% of the daily variance is due to jumps. On some extreme days, more
than 80% of the daily variance can be explained through jumps. It also seems that the impact
of jumps on daily variance was bigger in the rst period than in the next two periods.
Period Signif. All days Jump days Min. Max.
99% 1.51% 25.38% 6.78% 79.47%
1 99.9% 0.61% 31.62% 12.66% 80.82%
99.99% 0.16% 42.24% 18.49% 71.34%
99% 1.13% 16.85% 4.89% 49.18%
2 99.9% 0.22% 13.75% 4.66% 25.85%
99.99% 0.06% 15.00% 10.82% 22.52%
99% 1.98% 17.70% 2.12% 50.00%
3 99.9% 0.69% 15.93% 4.61% 47.78%
99.99% 0.34% 20.31% 7.32% 49.11%
Table 3: Part of the daily realized variance due to jumps. For each period and condence
level (Signif. in the table), the average part of the daily variance that is due to jumps is
reported (column All days). The last three columns shows the average (Jump days), minimal
(Min.) and maximal (Max.) contribution of jumps to the daily variance for days in which
at least a jump takes place.
5.2 Timing of jumps
In this subsection, we rst analyze whether jumps occur randomly throughout the trading day.
We then study the duration between two successive jumps.
Distribution of jumps throughout the trading day
We consider a typical trading day and study the probability for a jump to occur in each 5-minute
interval, from the opening to the closing of the market. In most of these intervals, only a handful
of jumps (or even no jump at all) have been identied. We therefore aggregate them into broader
20-minute intervals in order to get a more precise picture of their distribution throughout the
day. We consider each 20-minute interval separately and calculate the proportion of days in
5.2 Timing of jumps 18
which a jump took place in that interval. The results are reported in gure 4. We focus on the
99% (left panel of the gure) and 99.9% (right panel) condence levels as not enough jumps
have been identied at the highest condence level (99.99%) for conducting precise inferences.
10 12 14 16
0
1
2
3
4
5
6
x 10
-3
99% confidence level
P
e
r
c
e
n
t
a
g
e

o
f

j
u
m
p

i
n
t
e
r
v
a
l
s
10 12 14 16
0
0.5
1
1.5
2
2.5
3
3.5
x 10
-3
99.9% confidence level
IP-filtered returns
Raw returns
IP-filtered returns
Raw returns
Figure 4: Distribution of jumps throughout the trading day. The numbers on the
ordinate indicate the probability for a jump to occur in a 20-minute interval. The left and right
panels are for the 99% and 99.9% condence level.
Jumps are found in almost every interval but they do not spread evenly around the clock.
Typically more jumps occur just after market opening, at about 2:30 p.m. and in the late
afternoon. The raw and ltered series exhibit similar features. However, there are less jumps
in the ltered series. This is because many large changes arise at particular moments of the
day, when the market gets nervous. Thus many of them can be explained through the intraday
patterns. This illustrates how crucial it is to lter out intraday patterns from raw returns before
identifying jumps. All in one, these patterns are remarkably similar to the ones we get when
studying intraday market activity (see section 4.1).
Duration between two successive jumps
Table 4 shows the typical duration between two successive jumps for each period and each
condence level. In the rst period and for the lowest condence level, there was, on average, a
jump each 15.79 (calendar) days. During the next two periods, jumps were more frequent and
the average duration decreased to 7.06 days in the third period.
5.2 Timing of jumps 19
Period Signif. Mean Std. Min. Max.
99% 15.79 22.62 0.0035 104.97
1 99.9% 65.08 74.49 0.0208 349.90
99.99% 251.01 340.83 0.0208 916.95
99% 13.07 18.50 0.0035 90.88
2 99.9% 66.87 69.13 0.0243 210.87
99.99% 322.07 438.28 12.1597 631.98
99% 7.06 11.26 0.0035 63.01
3 99.9% 27.98 39.46 0.0069 196.68
99.99% 78.73 69.07 4.0174 253.67
Table 4: Duration between two successive jumps. This table shows the duration between
two jumps. For each condence level and period, the average duration as well as the standard
deviation, the minimum (Min.) and the maximum (Max.) of the duration are reported.
Figure 5 is complementary to this analysis as it shows the complete time-series of jumps for the
99% (upper panel in the gure) and 99.9% (lower panel) condence levels. Sometimes there is
more than one jump on the same day. For the lower condence level, there are up to 12 jumps
a day. This result looks a bit excessive. Indeed if we consider the 99.9% condence level, the
maximum amount of jumps a day decreases to 4. Furthermore, jumps tend to cluster. This is
particularly striking when we focus on the 99.9% condence level.
27
1996 1998 2000 2002 2004
0
5
10
Time-series of jumps
1996 1998 2000 2002 2004
0
2
4
Figure 5: Time-series of jumps. The y-axis reports the number of jump occurences on each
day. The upper and lower panels are for the 99% and 99.9% condence levels.
27
Andersen et al. (2007) analyze jump occurences for the S&P500 and nd evidence of jumps clustering.
6 Impact of a jump on volatility 20
6 Impact of a jump on volatility
6.1 Impact on absolute returns and implied volatility
We follow the approach of Bollerslev et al. (2006) to examine the nature of the relationship
between index returns and volatility. However, unlike Bollerslev et al. (2006), who use absolute
returns as a proxy for volatility, we focus on implied volatility.
We calculate the correlation coecient of DAX returns in a 5-minute interval ,

with implied
volatility changes in 5-minute interval ,

+ /, where / 12, . . . , 12. We denote by ,

the interval in which a jump has occurred. As a matter of comparison, we also examine the
correlation between r
S;t
j

and the absolute return

r
S;t
j

+k

.
28
The number of 12 leads and lags
corresponds to 1 hour around the time at which the jump occurred. In order to quantify the
additional impact of a jump on volatility, we also compute the correlation coecients for the
complete sample of returns (i.e. without making any distinction between jump and non-jump
intervals). There is only a limited number of jumps in our sample. This might limit our ability
to make precise inferences. Therefore, we make use of the complete 11 year sample and consider
only jumps that have been identied at the 99% and 99.9% condence levels.
Figure 6 reports the correlation coecients between the dierent series. We rst discuss the lead-
lag eects among returns and absolute returns (left panel of the gure). Correlation coecients
for / 1 and / < 0 are rather erratic and mostly insignicant. Though, at least for / = 0
and / = 1 they are consistently negative (about 0.2 to 0.4 for the 99% and 99.9% condence
levels) and signicantly dierent from zero. By comparing these results with the ones we get
for the complete sample (upper panel), we see that a jump has a real impact on volatility.
Furthermore, this impact remains signicant in the next 5-minute interval.
The right panels of the gure show the lead-eects for the implied volatility returns. These
patterns are much clearer than those for the absolute returns. The correlation coecients are
signicantly dierent from zero even for / < 0. This means that lagged volatility returns
have a predictive power on future jump occurrences. We also nd a signicant and highly
negative correlation of DAX returns with contemporaneous volatility returns (/ = 0). At the
99% condence level, the correlation coecient is even more negative at / = 1 than at / = 0
and remains signicant up to / = 5. This shows that the impact of a jump on volatility is
important and also indicates that the market needs some time to fully incorporate the eects
of the jump in the volatility process. At the 99.9% condence level, the retarded reaction is
even more pronounced as the correlation remains lower than 0.5 up to / = 2. If we compare
these gures with the ones from the right upper panel, we see that the results mostly indicate
the same kind of causality, running from index returns to volatility. Though the patterns are
stronger in the case of jumps. This was expected because a large return caused by a jump should
have more economic content than an innovation, which is only slightly dierent from zero. It is
dicult to interpret the few signicant correlations for / < 0. They seem to indicate that the
market has some ability to predict a jump and its sign. If it is the case, we would expect to nd
28
This measure is closely related to the skew correlations of Engle & Lee (1993).
6.1 Impact on absolute returns and implied volatility 21
a positive autocorrelation between the returns recorded before the jump occurred and the jump
return. We check this, but do not nd any signicant relationship.
-10 -5 0 5 10
-0.04
-0.02
0
Complete sample
-10 -5 0 5 10
-0.2
0
0.2
Jump intervals (99%)
-10 -5 0 5 10
-0.6
-0.4
-0.2
0
0.2
Jump intervals (99.9%)
-10 -5 0 5 10
-0.15
-0.1
-0.05
0
-10 -5 0 5 10
-0.6
-0.4
-0.2
0
0.2
-10 -5 0 5 10
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
Figure 6: Lead-lag eects among DAX returns and absolute DAX returns (left
panel) / returns on implied volatility (right panel). The upper panels report the results
for the complete sample (i.e. without any distinction between jump and non-jump intervals),
while the lower panels report the results for the jump intervals, where jumps are identied at the
99% and 99.9% condence levels. The dotted lines correspond to the lower and upper bounds
for a 95% condence interval for each coecient.
In order to gain further insight into this issue, we dierentiate lead-lag eects for positive
and negative jumps. This distinction obviously reduces the number of observations we have.
Therefore we consider only jumps identied at the 99% condence level and focus on a shorter
time horizon of 30 minutes (i.e. / 6, . . . , 6). The correlation coecients are reported
in gure 7. The gure provides some evidence in favor of an asymmetric relationship. On
the one hand, positive jumps do not seem to have much impact on volatility. On the other
hand, negative jumps have an important eect on volatility. We also nd some sort of volatility
feed-back, as the correlation coecients for / = 2 and / = 1 are signicant. However, the
causality remains mostly return-driven.
29
29
We run another test in oder to see if a jump might have an impact not only on the implied volatility level
but also on the whole implied volatility smile structure. The results indicate that a jump does not have a
signicant eect neither on the slope nor on the curvature of the smile function.
6.2 Impact on realized volatility 22
-6 -4 -2 0 2 4 6
-0.4
-0.2
0
0.2
0.4
Positive jumps
-6 -4 -2 0 2 4 6
-0.5
0
0.5
Negative jumps
Figure 7: Lead-lag eects among positive (and negative, right panel) jump returns
and returns on implied volatility. The dotted lines correspond to the lower and upper
bounds for a 95% condence interval for each coecient.
6.2 Impact on realized volatility
In this section, we focus on the impact of a jump on the realized volatility. We rst estimate
the realized variance before and after a jump occurred:
1\
K;t
j
<t

j
=
K

k=1
r
2
t

jk
and 1\
K;t
j
>t

j
=
K

k=1
r
2
t

j+k
,
where 1\
K;t
j
<t

j
and 1\
K;t
j
>t

j
are respectively the realized variance before and after a jump
took place. Both measures are estimated over a horizon of 1 5-minute intervals. t

j
is the time
at which a jump occurred. We consider various horizons, ranging from 5 minutes (1 = 1) to 1
hour (1 = 12). In order to render these measures more tractable, we reexpress them in terms
of annualized volatility, i.e. 1 =
_
252 (100,1) 1\ (as we have on average 100 5-minute
intervals per day).
We then study the impact of a jump on the realized volatility over an horizon of 1 periods by
taking the dierence between 1
K;t
j
>t

j
and 1
K;t
j
<t

j
. We dierentiate between positive and
negative jumps and get the following expressions:
1
+
K
= 1
K;t
j
>t

j
+
1
K;t
j
<t

j
+
and 1

K
= 1
K;t
j
>t

1
K;t
j
<t

,
where 1
+
K
and 1

K
respectively quantify the impact of a positive and a negative jump on
realized volatility, t

j
+
(t

) is the time at which a positive (negative) jump has occurred.

Figure 8 shows that the realized volatility tends to increase after a negative jump has occurred.
On an annualized basis, the increase amounts to about 49% for jumps identied at the 99%
condence level. The impact of the jumps identied at the 99.9% condence level is even stronger
(up to 62% increase in the realized volatility). However, this result is no longer signicant at the
95% level. This could be due to the scarcity of signicant jump observations. Positive jumps do
not have a similar impact on the realized volatility. The most signicant positive jumps seem to
trigger a decrease in the realized volatility, while less signicant jumps seem to have an opposite
impact on the realized volatility. Yet, in both cases, the results are far from being signicant at
the standard levels.
7 Conclusion 23
2 4 6 8 10 12
-0.6
-0.4
-0.2
0
0.2
0.4
Positive jumps
2 4 6 8 10 12
-0.5
0
0.5
Negative jumps
2 4 6 8 10 12
-1
-0.5
0
0.5
1
2 4 6 8 10 12
-1
-0.5
0
0.5
1
(99%)
(99.9%)
Figure 8: Impact of a jump on realized volatility over a horizon of K periods. The
left (right) panel reports the results for positive (negative) jumps. The dotted lines show the
upper and lower 95% condence bounds.
7 Conclusion
In this paper, we examine three related empirical questions. First, we characterize the intraday
patterns for both squared returns and implied volatilities. Second, we identify price discontinu-
ities and study their statistical properties. Third, we analyze the impact of a price jump on the
subsequent volatility level.
Our results indicate that intraday patterns for squared returns are typically \shaped and have
remained very stable over the sample period (1995-2005). The intraday volatility is particularly
high at the markets opening, at 2:30 p.m. and at the closing. We also observe very clear-cut
Jshaped patterns for implied volatility. These are more moderate than those for squared
returns but, as indicated by a small empirical experiment (see section 4.2), their economic
implications could be potentially important.
We identify some rare jumps. They occur on 0.37% and 11.20% of all days, depending on the
period under consideration and the condence level used to test their signicance. The impact
of jumps on the daily volatility is far from being negligible. On jump days, price discontinuities
account on average for 15% to 25% of the daily return variance and, on some particular days,
they explain as much as up to 80% of the daily return variance. We also notice that jumps
tend to cluster. The length of our sample allows studying precisely the occurrences of jumps
throughout the trading day. We nd that they are distributed according to patterns that are
very similar to the ones for squared returns. That is, there is a higher probability for a jump to
occur in the early morning, at 2:30 p.m. and in the late afternoon.
7 Conclusion 24
Finally we nd that price discontinuities have a considerable impact on volatility. A jump leads
to a direct adjustment of the implied volatility level. This adjustment is upward or downward
depending on whether the jump is negative or positive. Negative jumps have a much stronger
eect on volatility than positive ones. The information conveyed by a jump needs some time to
be fully integrated in the volatility process. That is, the impact of a jump on volatility remains
signicant during up to 25 minutes. Moreover, the subsequent impact of a jump on volatility
tends to be even more important than its direct impact. We nd almost similar results for
absolute returns and realized volatility but these results are less signicant. To sum up, we
document (i) a direct impact of a jump on volatility; (ii) a retarded (or causal) eect of a jump
on subsequent volatilities; (iii) a notable asymmetry in the relationship among jump returns and
volatility changes.
References 25
References
Ait-Sahalia, Y., Mykland, P.A., & Zhang, L. 2005. How Often to Sample a Continuous-Time
Process in the Presence of Market Microstructure Noise. Review of Financial Studies, 18(2),
351416.
Andersen, T., Bollerslev, T., Diebold, F.X., & Labys, P. 2001a. The Distribution of Realized
Exchange Rate Volatility. Journal of the American Statistical Association, 96(1), 4255.
Andersen, T., Bollerslev, T., Diebold, F.X., & Ebens, H. 2001b. The Distribution of Realized
Stock Return Volatility. Journal of Financial Economics, 61(1), 4376.
Andersen, T.A., Bollerslev, T., & Diebold, F.X. 2007. Roughing It Up: Including Jump Com-
ponents in the Measurement, Modeling and Forecasting of Return Volatility. The Review of
Economics and Statistics, 89(4), 701720.
Andersen, T.G., & Bollerslev, T. 1997a. Heterogeneous Information Arrivals and Return Volatil-
ity Dynamics: Uncovering the Long-Run in High Frequency Returns. Journal of Finance,
52(3), 9751005.
Andersen, T.G., & Bollerslev, T. 1997b. Intraday Periodicities and Volatility Persistence in
Financial Markets. Journal of Empirical Finance, 4(2-3), 115158.
Andersen, T.G., Bollerslev, T., & Lange, S. 1999. Forecasting Financial Market Volatility:
Sample Frequency Vis--Vis Forecast Horizon. Journal of Empirical Finance, 6(5), 457477.
Andersen, T.G., Bollerslev, T., & Das, A. 2001c. Variance-Ratio Statistics and High-Frequency
Data: Testing for Changes in Intraday Volatility Patterns. Journal of Finance, 56(1), 305327.
Andersen, T.G., Bollerslev, T., Frederiksen, P.H., & Nielsen, M.O. 2006a. Continuous-Time
Models, Realized Volatilities, and Testable Distributional Implications for Daily Stock Returns.
Working Paper, Cornell University.
Andersen, T.G., Bollerslev, T., & Diebold, F.X. 2006b. Handbook of Financial Econometrics.
Vol. Forthcoming. Amsterdam: North-Holland. Chap. Parametric and Nonparametric Volatil-
ity Measurement, in L.P. Hansen and Y. At-Sahalia (Eds.).
Baillie, R., & Bollerslev, T. 1991. Intra-Day and Inter-Market Volatility in Foreign Exchange
Rates. Review of Economic Studies, 58(3), 565585.
Bandi, F. M., & Russell, J. 2007. Handbooks of Financial Engineering. eds. John R. Birge and
Vadim Linetsky, Elsevier Science. Chap. Volatility.
Bandi, F.M., & Russell, J.R. 2008. Microstructure Noise, Realized Volatility, and Optimal
Sampling. Review of Economic Studies, 75(2), 339369.
Barndor-Nielsen, O.E., & Shephard, N. 2002. Econometric Analysis of Realised Volatility and
its Use in Estimating Stochastic Volatility Models. Journal of the Royal Statistical Society
Series B, 64, 253280.
Barndor-Nielsen, O.E., & Shephard, N. 2004. Power and Bipower Variation with Stochastic
Volatility and Jumps. Journal of Financial Econometrics, 2(1), 137.
References 26
Barndor-Nielsen, O.E., & Shephard, N. 2006. Econometrics of Testing for Jumps in Financial
Economics Using Bipower Variation. Journal of Financial Econometrics, 4(1), 130.
Bauwens, L., Rime, D., & Sucarrat, G. 2006. Exchange Rate Volatility and the Mixture of
Distribution Hypothesis. Empirical Economics, 30(4), 889911.
Bekaert, G., & Wu, G. 2000. Asymmetric Volatility and Risk in Equity Markets. Review of
Financial Studies, 13(1), 142.
Black, F. 1976. Studies of Stock Price Volatility Changes. Proceedings of the meetings of the
American Statistical Association, Business and Economics Section, Chicago, 177-181.
Black, F., & Scholes, M. 1973. The Pricing of Options and Corporate Liablilities. Journal of
Political Economy, 81(3), 637654.
Bollerslev, T., Gibson, M., & Zhou, H. 2004. Dynamic Estimation of Volatility Risk Premia and
Investor Risk Aversion from Option-Implied and Realized Volatilities. Finance and Economics
Discussion Series 2004-56, Board of Governors of the Federal Reserve System (U.S.).
Bollerslev, T., Litvinova, J., & Tauchen, G. 2006. Leverage and Volatility Feedback Eects in
High-Frequency Data. Journal of Financial Econometrics, 4(3), 353384.
Bouchaud, J.P., Matacz, A., & Potters, M. 2001. The Leverage Eect in Financial Markets:
Retarded Volatility and Market Panic. Physical Review Letters, 87(22), (228701) 14.
Christie, A.A. 1982. The Stochastic Behavior of Common Stock Variances: Value, Leverage and
Interest Rate Eects. Journal of Financial Economics, 10(4), 407432.
Clark, C.P. 1973. A Subordinated Stochastic Process Model with Finite Variance for Speculative
Prices. Econometrica, 41(1), 13555.
Cont, R. 2001. Empirical Properties of Asset Returns: Stylized Facts and Statistical Issues.
Quantitative Finance, 1, 223236.
Dacorogna, M. M., Mller, U. A., Nagler, R. J., Olsen, R. B., & Pictet, O. V. 1993. A Geo-
graphical Model for the Daily and Weekly Seasonal Volatility in the FX Market. Journal of
International Money and Finance, 12(4), 413438.
Dacorogna, M.M., Dav, R.R., Guillaume, D.M., Mller, U.A., Olsen, R.B., & Pictet, O.V. 1997.
From the Birds Eye to the Microscope: A Survey of New Stylized Facts of the Intra-Daily
Foreign Exchange Markets. Finance and Stochastics, 1(2), 95129.
De Jong, F., & Donders, M.W.M. 1997. Intraday Lead-Lag Relationships Between the Futures-,
Options and Stock Market. European Finance Review, 1(3), 337359.
Dennis, P., Mayhew, S., & Stivers, C. 2006. Stock Returns, Implied Volatility Innovations,
and the Asymmetric Volatility Phenomenon. Journal of Financial and Quantitative Analysis,
41(2), 381406.
Easley, D., OHara, M., & Srinivas, P.S. 1998. Option Volume and Stock Prices: Evidence on
Where Informed Traders Trade. Journal of Finance, 53(2), 431465.
Engle, R., & Lee, G. 1993. A Permanent and Transitory Component Model of Stock Return
Volatility. Discussion Paper, University of California at San Diego.
References 27
Fama, E. 1970. Ecient Capital Markets: A Review of Theory and Empirical Work. Journal
of Finance, 25(2), 383417.
Fernandes, M., & Grammig, J. 2005. Nonparametric Specication Tests for Conditional Duration
Models. Journal of Econometrics, 127(1), 3568.
Granger, C., & Poon, S.-H. 2002. Forecasting Volatility in Financial Markets: A Review. Journal
of Economic Literature, 41(2), 478539.
Gwilym, O.A., & Buckle, M. 2001. The Lead-Lag Relationship Between the FTSE100 Stock
Index and its Derivative Contracts. Applied Financial Economics, 11(4), 385393.
Hafner, R., & Wallmeier, M. 2001. The Dynamics of DAX Implied Volatilities. International
Quarterly Journal of Finance, 1(1), 127.
Hafner, R., & Wallmeier, M. 2007. Volatility as an Asset Class: European Evidence. European
Journal of Finance, 13(7-8), 621644.
Hansen, P., & Lunde, A. 2006. Realized Variance and Market Microstructure Noise. Journal of
Business and Economic Statistics, 24(2), 127161.
Harris, L. 1986. A Transaction Data Study of Weekly and Intradaily Patterns in Stock Returns.
Journal of Financial Economics, 16(1), 99117.
Hasbrouck, J. 2003. Intraday Price Formation in U.S. Equity Stock Markets. Journal of Finance,
58(6), 23752400.
Hentschel, L. 2003. Errors in Implied Volatility Estimation. Journal of Financial and Quanti-
tative Analysis, 38(4), 779810.
Huang, X. 2004. The Behavior of Jump Statistics. Working Paper, Duke University.
Huang, X., & Tauchen, G. 2005. The Relative Contribution of Jumps to Total Price Variance.
Journal of Financial Econometrics, 3(4), 456499.
Kirchner, T., & Schlag, C. 1998. An Explorative Investigation of Intraday Trading on the
German Stock Market. Finanzmarkt und Portfoliomanagement, 12, 1331.
Lahaye, J., Laurent, S., & Neely, C.J. 2007. Jumps, Cojumps and Macro Announcements.
Working Paper, Federal Reserve Bank of St. Louis.
Liu, X., Shackleton, M.B., Taylor, S.J., & Xu, X. 2007. Closed-Form Transformations from Risk-
Neutral to Real-World Distributions. Journal of Banking and Finance, 31(5), 15011520.
Madhavan, A. 2000. Market Microstructure: A Survey. Journal of Financial Markets, 3(3),
205258.
Masset, P., & Wallmeier, M. 2007. A High Frequency Investigation of the Interaction Between
Volatility and DAX Returns. Working Paper, University of Fribourg.
Meitz, M., & Tersvirta, T. 2006. Evaluating Models of Autoregressive Conditional Duration.
Journal of Business and Economic Statistics, 24, 104124.
Muller, U.A., Dacorogna, M.M., Olsen, R.B., Pictet, O.V., & Schwarz, M. 1990. Statistical
Study of Foreign Exchange Rates, Empirical Evidence of a Price Change Scaling Law, and
Intraday Analysis. Journal of Banking and Finance, 14(6), 11891208.
References 28
Newey, W. K., & West, K. D. 1987. A Simple, Positive Semi-Denite, Heteroskedasticity and
Autocorrelation Consistent Covariance Matrix. Econometrica, 55(3), 703708.
Ozenbas, D., Schwartz, R.A., & Wood, R.A. 2002. Volatility in US and European Equity
Markets: An Assessment of Market Quality. International Finance, 5(3), 437461.
Rebonato, R. 2004. Volatility and Correlation: The Perfect Hedger and the Fox. Wiley, New-
York.
Roll, R. 1984. A Simple Impicit Measure of the Eective Bid-Ask Spread in an Ecient Market.
Journal of Finance, 39(4), 11271139.
Stephan, J.A., & Whaley, R.E. 1990. Intraday Price Change and Trading Volume Relations in
the Stock and Stock Option Markets. Journal of Finance, 45(1), 191220.
Tauchen, G., & Pitts, M. 1983. The Price Variability-Volume Relationship on Speculative
Markets. Econometrica, 51(2), 485505.
Tian, G.G., & Guo, M. 2006. Interday and Intraday Volatility: Additional Evidence from the
Shanghai Stock Exchange. Review of Quantitative Finance and Accounting, 28(3), 287306.
Wood, R., McInish, T., & Ord, J. 1985. An Investigation of Transaction Data for NYSE Stocks.
Journal of Finance, 40(3), 723741.
Zhang, L., Mykland, P. A., & At-Sahalia, Y. 2005. A Tale of Two Time Scales: Determining
Integrated Volatility with Noisy High Frequency Data. Journal of the American Statistical
Association, 100, 13941411.