You are on page 1of 21

MARKET RISK MODELS FOR INTRADAY DATA

Pierre Giot
1
August 2002
Abstract
In this paper, we quantify market risk at an intraday time horizon using normal GARCH, Student
GARCH, RiskMetrics and high-frequency duration (Log-ACD) models set in the framework of the
conditional VaR methodology. Because of the small time horizon of the intraday returns (15 and 30
minute returns in this paper), an evaluation of intraday market risk can be useful to market par-
ticipants (traders, market makers) involved in frequent trading. As expected, the volatility features
an important intraday seasonality, which must be removed prior to using the market risk models.
The four models are applied to intraday returns data for three stocks traded on the New York Stock
Exchange and it is shown that the Student GARCH model performs best. We also comment on the
use of price durations as a measure of risk on time.
Keywords: Intraday market risk, Value-at-Risk, Duration models, NYSE
JEL classication: C22, C41, C53, G10
1
Department of Business Administration & CEREFIM, University of Namur, Rempart de la Vierge 8, B-5000 Namur,
Belgium, Tel.: +3281724887 and Center for Operations Research and Econometrics, Universite catholique de Louvain,
Belgium; email: pierre.giot@fundp.ac.be
While remaining responsible for any errors in this paper, the author would like to thank Luc Bauwens, Franz Palm, Olivier
Scaillet, Neil Shepard for useful remarks and suggestions.
1 Introduction
On April 4th 2000 around midday (EST), the Nasdaq and Dow Jones stock indexes plunged by 14% and
4.8% respectively from their opening levels. At the closing bell (4pm EST), the two indexes had recouped
most of their losses, closing down only 1.8% and 0.5%. An outside observer focusing on successive closing
prices would conclude that this had been a normal trading day, but it certainly was not for most active
traders and market makers. Over the last ve years, such intraday price variations have occurred more
and more frequently, highlighting the importance of intraday price movements in nancial assets.
1
Because of the increasing availability of so-called high-frequency databases for prices of stocks and
other nancial assets
2
, there has been a growing number of research papers focusing on intraday volatility:
Andersen and Bollerslev (1997, 1998 and 1999), Giot (2000) or the work by the Olsen & Associates
research group now summarized in Dacorogna, Gencay, Muller, Olsen, and Pictet (2001). However, to
our knowledge and with the exception of Muller, Dacorogna, and Pictet (1996), no study has been made
on the large intraday price movements and no work has been done on the intraday performance of market
risk models (of the VaR type for example).
In this paper we wish to address this issue by studying large intraday price movements for three
actively traded stocks on the NYSE. To quantify market risk at very short time horizons, we use the
conditional VaR framework and study the performance of the one-step ahead VaR predicted by normal
GARCH, Student GARCH, RiskMetrics and Log-ACD models. As we focus on market risk for intradaily
returns, our time horizon is extremely short as we deal with 15 and 30 minute returns. This is quite
dierent from the usual VaR models which often work with daily returns and then compute the 10 day
VaR (which is the VaR requested for regulatory reasons according to the Basel Capital Accord, see Jorion,
2000). Nevertheless, a study of market risk based on the VaR methodology for high-frequency returns is
useful for market participants (such as intraday traders and market makers) involved in frequent intraday
trading. As indicated in Christoersen and Diebold (2000), volatility forecastability (such as featured
by conditional volatility models) decays quickly with the time horizon of the forecasts. An immediate
consequence is that volatility forecastability is relevant for short time horizons. Thus in this paper, we
are consistent with these characteristics of volatility forecastability as we focus on intraday returns and
the quantication of market risk at the intraday level.
As in Giot (2000), we also use a high-frequency duration model (i.e. the Log-ACD model) on so-
called price durations
3
to characterize volatility, which is an alternative to ARCH type models. All models
are estimated on a sub-sample (estimation sample) of the original datasets and are then evaluated on
1
Most of these large intraday price variations can usually be traced back to external events, such as the Asian nancial
crisis in 1997, the Russian default on some of its bonds in 1998 and the subsequent near-failure of LTCM, a hedge fund.
2
Most exchanges, such as the New York Stock Exchange (NYSE) or the Paris Bourse, distribute price databases on
CD-ROMs which contain most intraday market characteristics, such as the time of the trade, the price, the volume,
. . . Regarding FOREX trading, the Swiss consultancy Olsen & Associates has recorded intraday quoted prices on the
Reuters screens for a large time period.
3
By denition, a price duration is the minimum amount of time needed for the price of an asset to increase or decrease
by a given amount. See Section 3.
1
the remaining portion of the dataset (forecast sample) by comparing the empirical failure rate to the
theoretical value. Anticipating on the results, it is shown that the Student GARCH model performs
much better than the other models. Thus, previous results which recognize the importance of the fat
tails of the returns distribution on daily data seem to extend to intraday data, provided that intraday
seasonality for the volatility is taken into account prior to estimating the models. Quite surprisingly, the
Log-ACD model for the price durations performs quite poorly when market risk is to be forecasted in a
xed interval framework (i.e. the framework of the other models). Thus we also comment on the use of
price durations and duration models to quantify market risk in an irregularly time-spaced framework.
The rest of the paper is organized in the following way. In Section 2, we briey review what is VaR.
In Section 3, we characterize the intraday VaR and intraday volatility as measures of market risk at very
short time horizons. In Section 4, we present the VaR models that are applied to our intraday data.
These models are then applied to three stocks traded on the NYSE in Section 5. Section 6 concludes.
2 Market risk models and Value-at-Risk
Broadly speaking Value-at-Risk is a quantitative tool whose goal is to assess the possible loss that can
be incurred by a trader or bank over a given time period and for a given portfolio of assets. Over the last
ten years, this technique has been increasingly used by banks and regulators all over the world as a way
to estimate possible losses related to the trading of nancial assets, i.e. as a tool designed to quantify
and forecast market risk.
4
Furthermore the VaR level (at a 10 day horizon) is directly related to the
amount of capital that the bank needs to put aside to cushion possible market losses (Basel Capital
Accord). Further general information about VaR techniques can be found in Jorion (2000) or Danielsson
and de Vries (2000).
From a statistical and quantitative point of view and working in the framework of market risk, the
VaR can be easily dened if a sample of past returns on the portfolio of assets is available. Indeed, once
the time series y
i
of the returns is known and the VaR level is specied, the VaR at level for the given
sample is the likely loss at the percent probability level, which can simply dened as the empirical
quantile at %. Hereafter, this empirical quantile is called z

. Because z

is such that F(z

) = , where
F(x) =

x

f(u) du is the cumulative density function of x, it is also the case that the probability area
right of z

is equal to 1 . This allows for an intuitive explanation of the VaR z

: with probability
1 the returns will be larger than the VaR.
From an empirical point of view, the computation of the VaR of a portfolio of assets thus requires the
computation of the empirical quantile at level of the distribution of the future returns of the portfolio.
If the returns are assumed to be identically distributed, the predicted VaR (i.e. VaR for future returns)
is usually based on the time-series of past returns. Broadly speaking, the VaR can be computed using
two kinds of models, parametric and non-parametric models, see van den Goorbergh and Vlaar (1999),
or Jorion (2000). A parametric model species a certain type of distribution for the returns (for example
4
Recent developments in quantitative nance aim at modelling other types of risks, such as credit risk, liquidity risk or
operational risk. See for example Saunders (2000) or Saunders (1999).
2
the normal distribution) and the empirical quantile is computed from the theoretical formula. If the
non-parametric method is chosen, then the empirical quantile is directly computed from the available
data without any model tted on the returns. In this paper we use parametric VaR models and among
those models we focus on conditional VaR models. Prior to detailing these models, we rst need to
address the issue of quantifying market risk for intraday returns using intraday VaR (or VaR applied to
intraday returns) which is the focus of this paper.
3 Intraday market risk, volatility and VaR
In this paper we characterize market risk on an intraday basis using the concept of intraday VaR, which
is an extension of VaR to intraday returns. As mentioned in the introduction, the main goal of VaR is
to assess the possible loss that can be incurred by a trader or nancial institution over a given period
of time. Because of regulatory reasons, the time horizon is usually a 10 day period and the models are
evaluated on daily returns.
5
However, for active market participants such as high frequency traders,
oor traders or market makers, the time horizon of their returns is much shorter and the corresponding
trading risk must therefore be assessed on such short time intervals. At the New York Stock Exchange for
example, all trading in a given stock is monitored by a single market maker (called the specialist). The
specialist must maintain an orderly market (which requires buys and sells of the stock by the specialist
for his own account) and thus continuously interacts with the oor traders and other traders. At the
NASDAQ which features a multiple market maker trading mechanism, there is a constant interaction
between the market makers and the traders.
6
Furthermore, with the growing use of the Internet as a
vehicle of trading, there is an increasing number of individuals engaged in so-called intraday trading.
Intraday returns are based on intraday data (prices at which the trades were made or posted quotes by
the market makers) of which main feature is that the recorded market characteristics of the trades (prices,
traded volume) and quotes (bid and ask prices, depths) are non regularly time-spaced. Indeed, because
trades and quotes are recorded continuously, the available observations are no longer equidistantly time-
spaced and the times between the trades or quotes may convey important information. As indicated by
the recent literature on high-frequency duration models, modelling the times between the market events
has become an important topic with applications in econometrics and market microstructure (Engle and
Russell, 1998; Bauwens and Giot, 2000).
While the data is presented in Section 5, an important question immediately arises as to the modelling
of the returns and their associated volatility. When intraday volatility is to be modelled, two methods
can be used (see Giot, 2000). Firstly, the non regularly time-spaced data can be re-sampled along a
pre-specied time grid, yielding equidistantly time-spaced observations. Calling P
i
= (B
i
+ A
i
)/2 the
sampled prices, raw returns on the bid-ask quotes can be computed as Y
i
= ln(P
i
) ln(P
i1
), for
i = 1 . . . N, with B
i
and A
i
the closest bid and ask prices to time t
i
. By denition, the prices P
i
are
5
For an estimation of the 10 day VaR not based on daily returns but on intraday returns, see Beltratti and Morana
(1999).
6
See Bauwens and Giot (2001) for a review of trading mechanisms in nancial markets.
3
sampled every s seconds, so that t
i
t
i1
= s. Then, standard volatility models (such as the GARCH
model for example) can be used provided that intraday seasonality is taken into account (Andersen and
Bollerslev, 1997; Giot, 2000).
7
In this paper, we assume a deterministic seasonality in the intraday
volatility.
8
Deseasonalized returns y
i
are computed from the raw returns Y
i
as
y
i
=
Y
i

(t
i
)
(1)
where (t
i
) is the deterministic intraday seasonal component. The latter is dened as the expected
volatility conditioned on time-of-day, where the expectation is computed by averaging the squared raw
returns over s second long intervals for each day of the week. Because intraday seasonality has been
taken into account, the volatility models are applied to the deseasonalized returns y
i
, but the VaR is
later computed for the original returns Y
i
by re-including the seasonal component (see below). We thus
avoid the pitfall of directly specifying the volatility and VaR models on the Y
i
returns, which would be
incorrect as the intraday seasonality is a main feature of the data and would distort estimation results.
Secondly, intraday volatility can be directly modelled by using high-frequency duration models applied
to the so-called price durations. Price durations X
p,i
= t

i
t

i1
are dened as the time needed to witness
a given cumulative price change c
p
in the price of the asset (see Bauwens and Giot, 2001 or Engle and
Russell, 1997).
9
Once a high-frequency duration model of the ACD type is tted on these price durations,
it can be shown that intraday volatility is a function of the conditional hazard of the ACD model (see
Section 4.4). The price durations feature a strong time-of-day eect akin to the intraday seasonality for
the volatility dened on the regularly sampled returns (Engle and Russell, 1998). To take into account
this deterministic intraday seasonality, we compute time-of-day standardized price durations, which are
dened as
x
p,i
=
X
p,i

p
(t

i
)
(2)
where X
p,i
is the raw ltered price duration with respect to the minimum price change c
p
,
p
(t

i
) is the
time-of-day eect and x
p,i
denotes the time-of-day standardized price duration. The deterministic time-
of-day eect is dened as the expected price duration conditioned on time-of-day, where the expectation
is computed by averaging the durations over thirty minutes intervals for each day of the week. Cubic
splines are then used on the thirty minutes intervals to smooth the time-of-day function. The Log-ACD
model is then applied to the newly dened time-of-day standardized price durations and the VaR is
computed for the original returns Y
i
(see below).
7
One of the main contributions of the papers by Andersen and Bollerslev (1997, 1998, 1999) is the identication of the
three sources of intraday volatility: long term (i.e. daily, but which aect intraday returns because of the aggregation
properties of the ARCH model) ARCH eects, intraday seasonality and intraday news announcements.
8
Similar deterministic techniques are also used in Andersen and Bollerslev (1997, 1998, 1999). Beltratti and Morana
(1999) dene a stochastic seasonality for the intraday volatility (which is much more complicated to estimate) and the
results are hardly better than those obtained with the more simple deterministic seasonality.
9
We use the notation t

i
to mark the beginning and end of price durations to distinguish them from t
i
which give the
times at which the regularly time-spaced returns are recorded. Thus, the t
i
are regularly time spaced, while the t

i
are not.
4
4 Market risk models for intraday data
For all models except the Log-ACD, we use the following notation. The original sample is S, which is
the collection of sampled (and thus equidistantly time-spaced) raw and deseasonalized returns Y
i
and y
i
(recorded at times t
i
) with i = 1 . . . N and N is the total number of observed returns. S is split in two
sub-samples, an estimation sample S
E
(for i = 1 . . . N
1
) and a forecast sample S
F
(for i = N
1
+1 . . . N).
The models are estimated on S
E
and their VaR performance is assessed on S
F
.
For the Log-ACD model, the original sample is S

, which is the collection of deseasonalized price


durations x
p,i
(recorded at times t

i
) at the pre-specied c
p
threshold, with i = 1 . . . N

and N

is the
total number of price durations. S

is split in two sub-samples, an estimation sample S

E
and a forecast
sample S

F
. The estimation sample contains x
p,i
, i = 1 . . . N

1
, with N

1
such that the time index of x
p,N

1
coincides with the time index of the last return of S
E
. The sub-sample S

E
is then the equivalent of
S
E
, but for price durations; they refer to the same time period. The high-frequency duration model
is estimated on the price durations contained in the estimation sample S

E
and its VaR performance is
assessed on the equidistantly time-spaced returns Y
i
in the S
F
dataset.
4.1 Normal GARCH
The ARCH model (Engle, 1982) models asset returns by allowing temporal dependence between the
squares of the returns. The GARCH(1,1) model introduced by Bollerslev (1986) can be written as
y
i
= +e
i
(3)
e
i
=
i

h
i
(4)
where is the expected return,
i
is drawn from an IID N(0, 1) distribution and h
i
is dened as
h
i
= +
1
e
2
i1
+
1
h
i1
(5)
where > 0,
1
0 and
1
0. To take into account a possible serial correlation in the returns, one
substitutes y
i
= +y
i1
+e
i
to Equation (3). If the parameters of the model have been estimated on
the S
E
sample, than, at index j corresponding to time t
j
in the forecast sample, the one-step ahead VaR
for the original returns is equal to
V aR
G
(j) = +

Y
j1
+z

h
j
(t
j
). (6)
In the above expression, we put the

(t
j
) factor to re-introduce the seasonality component of the
volatility (which was removed prior to the GARCH estimation on the y
i
returns).
5
4.2 RiskMetrics
The RiskMetrics analysis of VaR was introduced by JP Morgan in 1994 as a simple practical risk model
which requires almost no empirical computations. In its most simple form, it can be shown that the
basic RiskMetrics model is equivalent to an IGARCH model where the autoregressive parameter
1
is
set at a prespecied value 1 and the coecient of e
2
i1
is equal to . Thus it is (3) and (4) with
h
i
= e
2
i1
+ (1 )h
i1
. (7)
Once and

are known, the one-step ahead VaR at index j is equal to
V aR
RM
(j) = +

Y
j1
+z

h
j
(t
j
) (8)
as in the GARCH(1,1) case.
4.3 Student GARCH
While the simple GARCH(1,1) model often matches the empirical properties of the data quite well, it
usually cannot fully take into account the fat tails of the returns distribution. Indeed, it can be shown
that GARCH models allow for a kurtosis coecient larger than 3, but empirical daily or intradaily data
exhibit a still much larger kurtosis coecient. To alleviate this problem, the Student GARCH (or t
GARCH) is introduced, which species the underlying e
i
as being drawn from a t(0, 1, ) distribution.
At index j in the forecast sample, the one-step ahead VaR is then equal to
V aR
t
(j) = +

y
j1
+t
,

h
j
(t
j
). (9)
4.4 Log-ACD
As indicated above, the GARCH and RiskMetrics models have been specically designed for the analysis
of time-series which are made up of equidistantly time-spaced returns (for example daily or weekly
returns). In this sub-section, we highlight an alternative way of modelling intraday volatility which is
directly based on the price durations.
The Log-ACD model (Bauwens and Giot, 2000) for the standardized price durations (i.e. the price
durations of which the intraday seasonality has been removed, see Section 3) is dened by
x
p,i
=
e
i
(1 + 1/)

i
(10)

i
= +
i1
+
i1
(11)
where the
i
are IID and follow a Weibull(1,) distribution and
i
is the logarithm of the conditional
expectation of x
i
, so that
i
= lnE(x
i
|I
i1
). For covariance stationarity of
i
, || must be smaller than
one. As rst indicated in Engle and Russell (1998), there is a direct link between the instantaneous
6
intraday volatility and the conditional hazard of the price durations. Giot (2000) provides the link for
the Log-ACD models as

2
(t|I
i1
) =

c
p
P(t

i1
)

2
1
e
i

p
(t

i1
)
(12)
where
2
(t|I
i1
) is the conditional instantaneous intraday volatility for the raw returns, P(t

i1
) is the
bid-ask quote midpoint at time t

i1
, e
i
is the conditional expectation of price duration x
p,i
and
p
(t

i1
)
is the time-of-day eect. This equation provides a direct estimation of the intraday volatility once the
Log-ACD model has been estimated. Strictly speaking, this is the instantaneous volatility as forecasted
at time t

i1
. In an empirical application, this translates into the discretized volatility forecasted at all
times starting and ending a price duration. Thus in our VaR framework,
2
(t|I
i1
) is taken to be the
forecasted volatility at time t

i1
and valid up to t

i
, which is the end of price duration x
p,i
. Using the
normal distribution which underlies the diusion equation, the demeaned VaR forecasted at time t

i1
for the raw returns is
V aR
LogACD
(t

i1
) = z


2
(t|I
i1
). (13)
As such, it is not possible to compare this VaR to the other measures of VaR given by the models based
on regularly time-spaced data (indeed, the t

i
are not regularly time-spaced). To circumvent this diculty
and because we wish to compare this special VaR to the VaRs predicted by the other models using
regularly time-spaced returns, we compute the average of the V aR
LogACD
(t

i1
) for all times t

i1
on
the time interval [t
i1
, t
i
]. One of the motivation behind this sampling rule is that there are a lot more
t

i
than t
i
: in other words, the volatility based on the Log-ACD model is updated much more often than
in the regularly time-spaced framework.
10
5 Empirical application
5.1 Data and intraday seasonality
Our dataset is based on the TAQ CD-ROMs of the NYSE and contains intraday trade and quote data for
some selected stocks. In addition to the original irregularly time-spaced trades and quotes, our dataset
also contains regularly sampled quotes. In this paper we focus on three actively traded stocks on the
NYSE, BOEING, EXXON and IBM, for the January-May 1997 period. We make use of both types of
10
In a previous version of this paper, we resampled V aR
LogACD
(t

i1
) at each times t
i
dening the regularly time-
spaced returns. Thus, we used
V aR
LogACD
(t
i1
) = V aR
LogACD
(t

k
) (14)
where t

k
is the closest (from below) time to t
i1
. However this seems unfair for the Log-ACD model as its forecast of
volatility is going to be updated many times on the interval [t
i1
, t
i
], hence our dierent sampling rule in this version of
the paper.
7
datasets (i.e. regularly and irregularly time-spaced data) as the GARCH and RiskMetrics models need
the regularly time-spaced data, while the Log-ACD model is estimated on the irregularly spaced data.
As detailed in Section 4, the original sample is split into an estimation sample S
E
(and S

E
for the
price durations) and a forecast sample S
F
(and S

F
for the price durations). In our case, this means
that the ve months dataset for the three stocks is split in a three-month long estimation sample and a
two-month long forecast sample. We deal with two sampling frequencies which yield 15 and 30 minute
returns respectively and use the deseasonalization procedures given in Section 3 to transform the raw
returns Y
i
into intraday seasonally adjusted (isa) returns y
i
. Price durations are computed at the $1/8
level by ltering the original bid-ask quotes in the TAQ database and extracting those mid-quotes giving
a price change at least equal to $1/8. We then compute the time-of-day standardized price durations as
explained in Section 3.
Information on the isa returns is given in Table 1. As indicated in this table, the mean intraday return
is extremely small for all stocks, and much lower than its standard deviation.
11
Moreover, intraday
returns exhibit fat tails as their kurtosis is higher than 3 (this is conrmed by QQ-plots versus the
normal distribution). Focusing again on the 15 minute returns for the IBM stock, we plot in Figure 1
the autocorrelation functions (ACF) of the raw returns (Y
i
), squared raw returns (Y
2
i
), isa returns (y
i
)
and squared isa returns (y
2
i
). These gures are similar to those given in Bauwens and Giot (2001) and
Andersen and Bollerslev (1997, 1998, 1999) and clearly indicate the need to deseasonalize the data prior
to estimating any stochastic model. As expected, the squared raw returns exhibit cycles in their ACF
due the strong seasonality in the volatility (because of the recurrent increasing volatility at the open
and close of trading). The deseasonalized returns y
i
do not feature such cycles and exhibit a slowly
decreasing ACF.
Figure 2 plots the (annualized) deterministic intraday seasonal component for the volatility of the
15 minute returns (i.e. annualized

(t
i
)). The U-shaped pattern for the intraday seasonal component
is striking, as volatility is much higher at the open and close of the exchange than around lunch time.
The deterministic intraday pattern for the price durations at $1/8 is shown in the bottom of Figure 2
and is quite similar to what is given in Engle and Russell (1998) or Bauwens and Giot (2001): price
durations are much shorter around the open and close of the exchange than around lunch time, which
implies (according to equation (12)) that volatility is much higher at these times. This is in agreement
with the top of Figure 2 and the comment given above.
5.2 Estimated intraday VaR
In this section, we compute dierent measures of the intraday VaR for the returns sampled at several
frequencies. For each model presented in Section 4 and for two sampling frequencies (15 and 30 minute
returns), we compute the one-step ahead VaR on the returns Y
i
in S
F
. All models (except the Log-ACD
model) are estimated using the isa returns y
i
in S
E
while the Log-ACD model makes use of the price
11
Because the mean intraday return is so small, all reference to the mean of the returns () could be removed in the
preceding VaR formulae and the results would be almost identical.
8
Figure 1: Autocorrelation functions for the 15 minute returns (top left), isa returns (top right), squared
returns (bottom left) and squared isa returns (bottom right), January-May 1997, IBM stock.
durations given in S

E
; in both cases, the intraday seasonality in the volatility is re-introduced in the
VaR computation once the models have been estimated. More precisely:
(1) The GARCH (normal and Student) models are estimated on the returns y
i
in S
E
with N
1
xed at
1,586 (15 minute returns) and 793 (30 minute returns), yielding the estimated parameters ,

, ,

1
,

1
and . Equation 6 (for the normal GARCH) and equation 9 (for the Student GARCH) are
then used on the returns Y
i
in S
F
to compute the one-step ahead VaR.
(2) For the RiskMetrics method, equation 8 is used on the returns Y
i
in S
F
to compute the one-step
ahead VaR. The starting value for the iterative computation of h
i
is set at the unconditional
variance of the returns in S
E
.
(4) To compute the VaR of the Log-ACD model, we use the methodology given in Section 4.4. First,
we estimate the Log-ACD model on the price durations in S

E
. Then we use the estimated param-
eters to determine V aR
LogACD
(t

i1
) at all price durations in S

F
. In a third step, we average
V aR
LogACD
(t

i1
) for all times t

i1
in [t
i1
, t
i
]. This gives us the regularly time-spaced VaR.
All parametric models are estimated using maximum likelihood (CML estimation in GAUSS) and
diagnostic tests are (successfully) performed on the residuals of the normal GARCH, Student GARCH
9
Table 1: Deseasonalized intraday returns
15 minute returns 30 minute returns
BOEING EXXON IBM BOEING EXXON IBM
N 2,703 2,703 2,703 1,351 1,351 1,351
N
1
1,586 1,586 1,586 793 793 793
Mean -0.030 0 0.004 -0.044 -0.004 0.005
S.E. 1 1 1 0.999 1 1
Kurtosis 6.590 4.795 4.791 7.131 4.588 4.299
Information on the isa 15 and 30 minute returns. The returns are dened on equidis-
tantly sampled intraday data extracted from the January-May 1997 TAQ CD-ROMs.
and Log-ACD models. The performance of each model is then assessed by computing the failure rate
12
for the returns Y
i
in S
F
. It should be stressed that all models (including the Log-ACD model) are
evaluated with this criteria. Thus, the Log-ACD based method (which deals directly with the irregularly
spaced data) is assessed using regularly spaced returns. An alternative would be to dene and assess a
VaR on irregularly spaced prices (see comments in Section 5.3).
The failure rates are given in Table 2 (BOEING), Table 3 (EXXON) and Table 4 (IBM). Generally
speaking, the results are quite good for all three stocks and for both types of returns as the empirical
failure rates are close to their theoretical counterparts. Nevertheless, a statistical test is needed in order
to ascertain the quality of the estimation methods. Because the computation of the empirical failure rate
denes a sequence of yes/no observations, it is possible to test H0 : f = against H0 : f = , where f is
the failure rate (estimated by

f, the empirical failure rate).
13
At the 5% level and if T yes/no observations
are available, a condence interval for

f is given by

f 1.96

f(1

f)/T,

f + 1.96

f(1

f)/T

. In
Tables 2, 3 and 4, empirical failure rates are set in bold when the theoretical quantile does not belong
to the corresponding condence interval: this denotes a failure of the VaR model.
14
A summary of the
number of successes for each model as a function of the level of is given in Table 5. These results
indicate that:
(1) When is at the 5%, 2.5% or 1% level, the performance of models based on the normal distribution
is quite satisfactory (although the RiskMetrics model completely breaks down at the 1% level).
The Student GARCH is rejected twice at the 5% level because it is too conservative in one case
and the Log-ACD is rejected in most cases (see below for comments regarding this model).
(2) When < 1%, the normal GARCH and RiskMetrics models clearly underperform. When = 0.5%
12
By denition, the failure rate is the number of times returns exceed the forecasted VaR. If the VaR model is correctly
specied, the failure rate should be equal to the prespecied VaR level.
13
In the literature on VaR models, this test is also called the Kupiec LR test, if the hypothesis is tested using a likelihood
ratio test. See Kupiec (1995).
14
Because this is a bilateral test, a model can fail if it under or overestimates the true VaR. Thus, too conservative
models are also rejected.
10
for example, these models are rejected 5 times out of 6. In contrast, the Student GARCH model
fails only once. Of course this was to be expected as the Student GARCH model takes into account
the fat tails feature of the distribution of returns. The Log-ACD is rejected in most cases.
(3) The Log-ACD model performs rather poorly when the one-step ahead VaR is to be forecasted (in a
regularly time-spaced framework). As indicated in Table 5, the price durations based model fails
most of the time and for all stocks. Empirical results given in Tables 2, 3 and 4 however show
that the Log-ACD VaR tracks quite well the GARCH VaR while being always somewhat lower,
hence the larger failure rate. Thus, while Equation 12 allows the computation of the instanta-
neous volatility based on the price durations, there does seem to be a slight discrepancy with the
volatility computed from regularly time-spaced returns. However, one should notice that: (a) the
instantaneous volatility has to be discretized to be compared with the observed returns and can
only be computed when a price duration ends, (b) the instantaneous volatility is based on the
normal approximation and hence uses the critical values of this distribution. We further comment
on price durations and market risk in the next section.
As an illustration of the performance of the Student GARCH model, we give in Figure 3 the level of
market risk (i.e. forecasted intraday VaR) for an investment of $100, 000 relative to 15 minute returns.
We plot this VaR for the April 7th - April 11th week, which is the second week of the forecast sample,
and for the three stocks in our dataset. Thus, the right scale of the graphs gives the likely loss (at the
1% level) at a time horizon of 15 minutes. The strong seasonality in the market risk is immediately
apparent and seems to be a dominant factor in the risk analysis; indeed, the market risk level is 2 to
3 times higher around the open and closing hours of the exchange than around noon. This factor is of
course common to the three stocks as they are inuenced by the same global features of the exchange.
5.3 Price durations as a measure of market risk
The empirical results given in the previous section clearly indicate that the performance of the Log-ACD
model is quite disappointing in forecasting the intraday VaR in a regularly time-spaced framework. This
is quite puzzling as:
(1) Formula (12) is conceptually sound as it is based on the usual diusion model for an asset price;
(2) The deseasonalization techniques used for regularly time-spaced returns and price durations are
the same, hence the multiplicative factors re-introducing the seasonality should be neutral to the
analysis;
(3) Diagnostic tests performed on the residuals of the GARCH and Log-ACD models indicate that both
models successfully model their respective datasets;
(4) Prigent, Renault, and Scaillet (1999) successfully price stock options using a high-frequency duration
model of the Log-ACD type applied to price durations.
11
Table 2: VaR results for BOEING
15 minute returns
5% 2.5% 1% 0.5% 0.25%
normal GARCH 4.21 2.69 1.34 1.08 0.90
Student GARCH 4.57 2.51 1.07 0.54 0.18
RiskMetrics 4.30 2.96 2.24 1.52 1.25
Log-ACD 7.44 4.75 2.51 1.79 1.25
30 minute returns
5% 2.5% 1% 0.5% 0.25%
normal GARCH 3.95 2.87 1.44 1.44 1.08
Student GARCH 4.67 2.33 1.26 0.54 0.18
RiskMetrics 5.21 3.41 1.97 1.44 1.08
Log-ACD 5.92 3.59 2.33 1.61 1.26
Failure rates for the normal GARCH, t GARCH, RiskMetrics and
Log-ACD measures of intraday VaR. The estimation sample is the
January-March 1997 period, and the forecast sample is the April-May
1997 period (BOEING stock). A bold gure indicates that the corre-
sponding VaR is signicantly dierent (LR test) from the theoretical
value.
Table 3: VaR results for EXXON
15 minute returns
5% 2.5% 1% 0.5% 0.25%
normal GARCH 6.18 3.85 2.24 1.43 1.25
Student GARCH 6.63 3.31 1.43 0.90 0.27
RiskMetrics 5.56 3.94 1.79 1.43 1.07
Log-ACD 7.26 5.29 3.22 2.24 1.52
30 minute returns
5% 2.5% 1% 0.5% 0.25%
normal GARCH 6.10 3.59 2.87 1.61 0.72
Student GARCH 6.28 3.41 1.97 0.36 0.18
RiskMetrics 4.67 3.59 2.33 1.26 0.72
Log-ACD 7.90 5.92 3.77 2.33 1.79
Failure rates for the normal GARCH, t GARCH, RiskMetrics and
Log-ACD measures of intraday VaR. The estimation sample is the
January-March 1997 period, and the forecast sample is the April-May
1997 period (EXXON stock). A bold gure indicates that the corre-
sponding VaR is signicantly dierent (LR test) from the theoretical
value.
12
Table 4: VaR results for IBM
15 minute returns
5% 2.5% 1% 0.5% 0.25%
normal GARCH 3.58 2.41 1.61 1.07 0.90
Student GARCH 3.58 2.24 0.98 0.18 0.09
RiskMetrics 4.03 3.14 1.97 1.52 0.90
Log-ACD 7.97 5.64 3.85 2.78 1.88
30 minute returns
5% 2.5% 1% 0.5% 0.25%
normal GARCH 4.31 2.87 1.61 0.54 0.54
Student GARCH 4.31 2.51 0.54 0 0
RiskMetrics 4.31 3.95 2.33 0.90 0.72
Log-ACD 8.62 4.85 3.59 2.63 2.15
Failure rates for the normal GARCH, t GARCH, RiskMetrics and
Log-ACD measures of intraday VaR. The estimation sample is the
January-March 1997 period, and the forecast sample is the April-May
1997 period (IBM stock). A bold gure indicates that the correspond-
ing VaR is signicantly dierent (LR test) from the theoretical value.
However, this method relies on the normal distribution while the best performing methods use a
fat-tail density distribution. Furthermore, a lot of time transformations are needed to switch from
the irregularly time-spaced framework to the regularly time-spaced world where the VaR is assessed.
Indeed, the volatility forecast based on the Log-ACD model is always below the volatility forecast of
the GARCH model. A good example of this feature is shown in Figure 4, which displays the normal
GARCH and Log-ACD forecast of intraday VaR for 15 minute returns (April 7th - April 11th week).
Both processes are driven by the same seasonality and the Log-ACD VaR is close to the GARCH VaR,
but it is always below it. Furthermore, the sampling rule used to convert the irregularly time-spaced
VaR into the regularly time-spaced VaR (discussed in Section 4.4) does not seem to matter much as the
results given in this version of the paper are not very dierent from those given in a previous version
where another sampling rule was used.
While the Log-ACD model and price durations do not seem to help much when a market risk forecast
on a regularly time-space grid is required, forecasts of price durations do provide an important information
regarding the expected length of time before seeing a given price change (equal to $c
p
). Hence modelling
price durations is tantamount to modelling a risk on time and stability of the traders position. In this
sense, a very short price duration forecast indicates that the trader or market maker needs to perform
an active market monitoring as the trading environment is challenging. As an illustration, we give in
Figure 5 the forecast of the expected time before a price change of at least $1/8 for the IBM stock (April
7th - April 11th week, in transaction time). These forecasts are of course equal to the one-step ahead
13
Table 5: Number of successes at given percentage level
5% 2.5% 1% 0.5% 0.25%
normal GARCH 5/6 5/6 4/6 1/6 2/6
Student GARCH 4/6 6/6 5/6 5/6 5/6
RiskMetrics 6/6 4/6 0/6 1/6 2/6
Log-ACD 1/6 1/6 0/6 0/6 0/6
Number of successes for each VaR model, for all three stocks
and for 15 and 30 minute returns combined, i.e. number of
times the empirical failure rate is not signicantly dierent
from the corresponding percentage level.
forecasts of the price durations by the Log-ACD model. Because intraday data is by nature irregularly
time-spaced, a risk on time model is useful to a trader who faces the ow of irregularly time-spaced
trades and quotes.
6 Conclusion
In this paper, we quantify market risk at an intraday time horizon using normal GARCH, Student
GARCH, RiskMetrics and high-frequency duration (Log-ACD) models set in the framework of the con-
ditional VaR methodology. While VaR models are usually applied to daily data to help manage the
nancial risks of banks and nancial institutions, we use these VaR models to assess the market risk on
15 and 30 minute intraday returns. Our time horizon is thus much shorter than what is usually consid-
ered in the VaR literature but our framework can be helpful to market participants engaged in frequent
trading (such as market makers or day traders). As in previous papers by Giot (2000) or Andersen and
Bollerslev (1997, 1998, 1999), the empirical returns feature a strong intraday seasonality in the volatility,
which must be taken into account when estimating the models.
The four models are applied to intraday returns data for three stocks (BOEING, EXXON and IBM)
traded on the New York Stock Exchange and it is shown that the Student GARCH model performs best
when the intraday VaR is assessed in a regularly time-spaced framework. Furthermore, once intraday
seasonality in the volatility is taken into account, there do not seem to be any key dierences between
daily and intradaily VaR models as the results given in this paper show that usual VaR results on
daily data extend to intradaily returns. When assessed in a regularly time-spaced framework, the high-
frequency duration model (Log-ACD model) performs rather poorly. Although the forecasted intraday
volatility and VaR are quite close to their regularly time-spaced counterparts, the empirical failure rate
of the model is too high and the model is often rejected. This result is quite puzzling as it is not clear
why it is so, although the use of the normal density distribution and the need of complicated time
transformations to switch to the regularly time-spaced world are indeed drawbacks of this method.
Nevertheless we motivate the use of this type of model applied to price durations in order to model a
14
risk on time: short price duration forecasts indicate that the trader or market maker needs to perform
an active market monitoring as the trading environment is challenging. These periods are of course
characterized by a high regime of volatility.
References
Andersen, T., and T. Bollerslev (1997): Intraday periodicity and volatility persistence in nancial
markets, Journal of Empirical Finance, 4, 115158.
(1998): DM-dollar volatility: intraday activity patterns, macroeconomic announcements, and
longer-run dependencies, Journal of Finance, 53, 219265.
(1999): Forecasting nancial market volatility: sample frequency vis-`a-vis forecast horizon,
Journal of Empirical Finance, 6, 457477.
Bauwens, L., and P. Giot (2000): The Logarithmic ACD model: an application to the bid-ask quote
process of three NYSE stocks, Annales dEconomie et Statistique, 60, 117149.
(2001): Econometric modelling of stock market intraday activity. Kluwer Academic Publishers.
Beltratti, A., and C. Morana (1999): Computing value-at-risk with high frequency data, Journal
of Empirical Finance.
Bollerslev, T. (1986): Generalized autoregressive condtional heteroskedasticity, Journal of Econo-
metrics, 31, 307327.
Christoffersen, P., and F. Diebold (2000): How relevant is volatility forecasting for nancial risk
management?, Review of Economics and Statistics, 82, 111.
Dacorogna, M., R. Gencay, U. Muller, R. Olsen, and O. Pictet (2001): An introduction to
high-frequency nance. Academic Press.
Danielsson, J., and C. de Vries (2000): Value-at-Risk and extreme returns, Annales dEconomie
et Statistique, 3, 7385.
Engle, R. (1982): Autoregressive conditional heteroscedasticity with estimates of the variance of
United Kingdom ination, Econometrica, 50, 9871007.
Engle, R., and J. Russell (1997): Forecasting the frequency of changes in quoted foreign exchange
prices with the autoregressive conditional duration model, Journal of Empirical Finance, 4, 187212.
(1998): Autoregressive conditional duration; a new model for irregularly spaced transaction
data, Econometrica, 66, 11271162.
Giot, P. (2000): Time transformations, intraday data and volatility models, Journal of Computational
Finance, 4, 3162.
15
Jorion, P. (2000): Value-at-Risk. McGraw-Hill.
Kupiec, P. (1995): Techniques for verifying the accuracy of risk measurement models, Journal of
Derivatives, 2, 17384.
Muller, U., M. Dacorogna, and O. Pictet (1996): Heavy tails in high-frequency nancial data,
Olsen preprint.
Prigent, J., O. Renault, and O. Scaillet (1999): An Autoregressive Conditional Binomial Option
Pricing Model, THEMA Discussion Paper 9940.
Saunders, A. (1999): Credit risk measurement. Wiley.
(2000): Financial institutions management. McGraw-Hill.
van den Goorbergh, R., and P. Vlaar (1999): Value-at-Risk analysis of stock returns. Historical
simulation, tail index estimation?, De Nederlandse Bank-Sta Report, 40.
16
Figure 2: Time-of-day pattern for the annualized volatility (computed on squared 15 minute returns)
and price durations (at $1/8), January-May 1997, IBM stock.
17
Figure 3: Forecasted market risk level (for an investment of $100, 000 and 15 minute returns) according
to the Student GARCH(1,1) model for the BOEING, EXXON and IBM stocks. This gure refers to the
April 7th - April 11th week of the forecast sample.
18
Figure 4: Annualized VaR (as forecasted by the Log-ACD and normal GARCH models) for the 15 minute
returns of the IBM stock (April 7th - April 11th week).
19
Figure 5: Expected time before a price change of at least $1/8 for the IBM stock (April 7th - April 11th
week).
20

You might also like