Determinants of their relative impact
on T-Bond futures
Dieter Hess
Center ofFinance and Econometrics(CoFE)
Universityof Konstanz,D-78457 Konstanz, Germany
email: dieter.hess@uni-konstanz.de
tel:+49 7531 883164
fax: +49 7531 883559
March 2001
ForvaluablecommentsI amgratefulto JanBeran,AngelikaEymann,YuanhuaFeng,Bernd
Fitzenberger,Gunter Franke,FrankGerhard,Nikolaus Hautsch,JoachimInkmann,StefanKlotz,
MichaelLechner,ErikLuders,JurgenMeckl,Winfried Pohlmeier,JackWahl, andPeterWinker.
This paper has also beneted from comments of participants of the Seventh Annual Meeting
of the German Finance Association (DGF) as well as workshop participantsat the Universities
of Konstanz, Mannheim, and St. Gallen. Data on analysts' forecasts were generously provided
by Standard & Poors Global Markets. Research support by the Centre for European Economic
Research (ZEW) and nancial support by the Deutsche Forschungsgemeinschaft (DFG, project
HE 3180/1)isgratefully acknowledged.
Determinants of their relative impact on T-Bond futures
Abstract
This paper investigates the intraday response of CBOT T-bond futures
prices to surprises in headline gures contained in scheduled U.S. macroe-
conomic news releases. While several previous studies try to nd out which
releaseshaveasignicantimpactonpricesandvolatilityinnancialmarkets,
considerably less eort has been devoted to the question what makes some
releases important in contrast to others that seem to attract no attention
at all. Inorder to identifythe factorsdeterminingthe relative importanceof
releases,thetimeseriespropertiesandtheinformationcontentofthemacroe-
conomicnewsowareinvestigated.Inparticular,severaltypesofinformation
regardinginationand economicstrengtharedistinguished.Theexplanatory
powerof the type of information istested against the alternative hypothesis
that the timeliness of a release determines its impact. The results indicate
that the value of the information contained in a release decreases with the
numberof previouslyreleased gureshighlightingsimilar aspects. Thus, the
priceimpactof arelease decreasesastheadditionalinformationcontainedin
a release becomes smaller.
Keywords:Macroeconomic news;scheduledannouncements;informationprocessing;price
formation; Treasurybonds; futuresmarkets
JEL classication:E44,G14
Every month, a variety of macroeconomic reports, such as monthly employment
gures, consumer prices, and building permits, are released providing brand new
information about the state of the economy. While several studies try to nd out
whichreleaseshaveasignicantimpactonpricesoronvolatilityinnancialmarkets,
considerably less eorthas been devotedtothe question why some releasesprovoke
signicant marketreactions while others seem to be ignored.
Based on an analysis of the time series properties of the macroeconomic news ow
andtheinformationcontentofreports,thispaperpresentsandtestssomealternative
hypothesesexplainingtherelativeimportanceofreleases.Unlikearegulartimeseries
of say temperature measurements obtained from a single meteorological station at
the end of each month, several measurements of economic conditions in a given
period are taken from dierent perspectives. Moreover, instead of announcing all
these gures simultaneously, macroeconomic reports are released one afteranother
{ some with apronounced time lag.Hence, one hypothesis argues that the value of
the information provided by a release, and thus its impact on prices, depends on
its timeliness.On the other hand, macroeconomic reports contain dierent types of
informationaboutthestateoftheeconomy,especiallydierentindicatorsofination
as well as economic strength. Therefore, an alternative hypothesis argues that the
type of information is crucial in determining the relative importance of releases.
More precisely, itstates that the informationvalue of arelease diminishes with the
numberofpreviouslyreleasedreportsprovidingasimilarinformationcontent.These
hypotheses are tested inthis paper.
From thepreviousliterature itiswellknown thatinformationarrivalhas animpact
an overview). Since information arrival is a rather broad concept, previous studies
have stressed various aspects of news arrival by employing dierent measurement
concepts.
1
Scheduledmacroeconomicannouncementsstandoutfromthesteadyow
of informationwhichhitsnancialmarkets. Severalstudies showthat thesereleases
have a very distinct impact on volatility. Their importance is underlined, for ex-
ample, by the ndings of Fleming and Remolona (1997) that out of the 25 largest
intradaypricechangesintheU.S.treasurymarketallbutoneoccurredaftersuchan
announcement.ThisisconrmedbyBollerslev,Cai,and Song(2000)forT-bondfu-
tures. Constructingdummy variables fromthe schedule of macroeconomicreleases,
Ederington and Lee (1993), Crain and Lee (1995), and others nd that quite a
number of releases have a signicant impact on volatility in bond and foreign ex-
change markets. Using robust tests, Franke and Hess (2000a) nd an even larger
spectrum of releases to be signicant in the T-bond futures market. The increase
in volatility seems tobe rather short-lived, although this periodmayvary substan-
tially across releases.
2
Providinga comprehensive study of high-and low-frequency
volatility components in T-Bond futures returns, Bollerslev, Cai, and Song (2000)
restrictthe volatilityresponsehorizonfor allmacroeconomicreleases exceptfor the
employment report to one hour. The estimated pattern suggests that ten minutes
after anannouncement the initialimpactis reduced by one half.
3
1
Mitchell and Mulherin (1994), for example, use the daily number of news announcements
reported byDowJones&Company (Broadtape) inorder to explainpatternsin tradingactivity.
Chang and Taylor(1996) investigate intraday volatility employing a keyword count in Reuters
headlines.
2
Volatility seems to persist longer in the more liquid futures markets (Christie-David and
Chaudhry1999).
3
Notonlyvolatilityisaectedbyannouncements.Balduzzi,Elton,andGreen(1997)andFlem-
ing and Remolona (1999a) present evidence for the U.S. Treasury market that trading volume
surgesandbid-askspreadswidenafterarelease.Moreover,FrankeandHess(2000a)ndthatthe
formation. In general, the impact of the mere existence of an announcement is
investigated using dummy variables. In contrast, another branch of the literature
investigatesthe impact ofsurprises in announcements onthe level ofprices, mainly
in bond and foreign exchange markets. Usually, these studies measure the magni-
tude of surprises employing survey data on analysts' forecasts for certain headline
gures containedin macroeconomic reports. Early studiesfocus on asmallnumber
of releases investigating their impact on daily returns. Berkman (1978), Urich and
Wachtel(1981, 1984),and others analyzemoney growth announcements,Cook and
Korn (1991) and Prag (1994), for example, focus on employment reports. Since in
the early1980s the FederalReserve deemphasized monetaryaggregates toguideits
policy actions, Dwyer and Hafer (1989) among others, examine whether the Fed's
focus oncurrent economic conditions leads tosignicant interest rate changes after
surprises invarious macroeconomic reports.Interestingly,they nd that theimpact
of money growth announcements diminishes in the mid-1980s. Consequently, other
studies, such as Hardouvelis (1988) and Edison (1996), nd a growing inuence of
employment gures, releases of consumer or producer prices, durable goods orders
and retail sales. While these studies examine daily interest rate changes, Becker,
Finnerty, and Kopecky (1996) and Fleming and Remolona (1997, 1999b) focus on
narrowintraday windows around the announcements. This should help to separate
the impact of scheduled announcements from other not explicitly observed news
which may arrive occasionally over the course of a trading day. As a consequence,
Fleming and Remolonand more releases that have a signicant impact on prices
correlationofintradaypricechangesofT-bondandBundfutures issignicantlyincreased,Chris-
tiansen(2000)ndssignicantlyhigherconditionalcorrelationsofdailybondreturnsfordierent
maturities.
This papercontributesto the previousliterature by analyzingthe informationcon-
tent of releases and the time series properties of the announcement cycle. On the
basis of this analysis, competing hypothesis concerning the relative importance of
releasesarederived:Theimpactisdeterminedby thetimeliness(orthesequence) of
announcements, and/orby the type of information ina report. Performing a series
of tests forthese hypotheses shows that timeliness alone isnot suÆcientto explain
dierences inthe impactof releases. Better resultsare obtainedtaking intoaccount
thatmacroeconomicreportsprovidedierenttypesofinformation,i.e.dierentindi-
cators ofinationand economic strength.Whilethe previousliterature isprimarily
concerned with the question which releases have a signicant impact on prices (or
volatility) ina regression framework, this study tackles the issue of the underlying
factors driving the relative impact of macroeconomic releases. This should help to
resolve some contradictory results of previous studies. For example, Fleming and
Remolona(1999b)nd asignicantimpactoftheIndexof LeadingIndicators(LI),
while Flemingand Remolona (1997) do not. Since the main components of LI are
available well in advance of the announcement of the index, and in addition, LI
comes rather late inthe release cycle, this paperwould decide the case for Fleming
and Remolona(1997)despitethe factthat theirresultsare based onashortersam-
ple period. Althoughhere the focus is onrst moments (i.e. signed price changes),
the results do have implications for the analysis of volatility as well. For example,
the nding of a signicantly higher volatility around the announcement of a given
report has tobeinterpretedmore carefullyif this reportdoesnot leadtoconsistent
and signicant changes in the price level due to the fact that the report contains
rather outdated information.
structure of the macroeconomic release cycle and the content of major releases.
Moreover, some hypotheses concerning the relative importance of reports are pre-
sented.Insection3,thesehypothesesaretestedonthebasisofasystemofequations
describing the impactofsurprises inheadlinegures onprices. Section4concludes.
2 The impact of information arrival on prices
2.1 T-Bond futures price changes and surprises
Using narrow ve-minute windows around announcement times, T-Bond futures
price reactions to non-anticipated information in U.S. macroeconomic announce-
mentsareinvestigated.ThisfuturescontractislistedattheChicagoBoardofTrade
(CBOT) and calls for deliveryof a T-bond withat least 15years tomaturity.Note
thatT-bond futurespricesareby farmoresensitivetochangesinlong-terminterest
rates than to short rate movements.
4
Let P
i
denote the change of the futures price in anarrowtime intervalaround t
i .
More precisely, this isthe dierencebetween the lasttrading price observed before
t
i
and the last price observed within the interval (t
i
;t
i
+t], where t equals ve
minutes. Thisprice changeismodeledasalinearfunctionof distinctpiecesof news
arriving during this period,especiallyheadlinegures in scheduled macroeconomic
reports.Theseheadlineguressummarizetheinformationcontainedinsuchareport
(e.g. the overall unemployment rate inthe employment report; see table 5for more
details). Since they are closely watched by market participants,analysts' forecasts
4
Due to the high duration of the underlying bonds and the short contract maturity of the
investigated front month contract, bond price movements outweigh the cost-of-carry impact on
futures pricesbyfar.Fordetailssee,forexample,Jarrow(1996).
('consensus forecast') of analysts surveyed by Standard & Poors Global Markets
(also known as MMS) is employed here. Let F
j
denote this forecast for headline j
and A
j
its announced value. D
j;i
is a dummy variable equal to one if A
j
becomes
available duringthe time interval(t
i
;t
i
+t], and zero otherwise. The future price
changemaythenbewrittenasafunctionofsurprisesinheadlinegures,i.e.A
j F
j ,
P
i
= X
j
j (A
j F
j )D
j;i +"
i
: (1)
Any otherinformationarrivingbetween t
i and t
i
+twhichmightsurprisemarket
participantsandothereectsonpricesare reectedbytheerrorterm"
i
in(1).Since
pricereactionsinverynarrowtimewindowsaroundannouncementsareinvestigated,
the probability that other information besides the observed releases arrives and
aects prices signicantly should be fairly small. Eq. (1) is analyzed for the three
majorreleasetimesof scheduledmacroeconomicannouncements,i.e.8:30,9:15,and
10:00 a.m. ET (EasternTime).
According to the well known eÆcient market hypothesis one would expect that
the impact of a surprise is incorporated rapidly in prices, especiallysince the price
responseofoneofmostactivelytradedfuturescontractstowidelyanticipatedhead-
line gures is analyzed here.
5
Hence, it would be rather astonishing to nd that a
surprisein an8:30headlinestillhas animpactonve-minute price changesaround
9:15. Since the remainder of this paperanalyzes the value ofinformationcontained
inreleases, itisratheressentialthatwelookata marketthat processesinformation
eÆciently. Therefore, the following hypothesis is testedas a prerequisite:
5
InformationprocessingintheopenoutcrysystemoftheCBOTshould beveryeÆcient.Fora
discussionofinformationdiusioninelectronicandoortradingsystemssee,forexample,Franke
andHess(2000b).
Ifmarkets processinformationeÆciently, pricechangesshouldnot beaected
systematicallybypreviousannouncements(i.e. reports beingreleased45or90
minutes earlier duringthe day).
Since non-anticipated information is measured by the deviation of announcements
from analysts' forecasts, another prerequisite for the analysis of the impact of non-
anticipated informationis that market participantsare actually surprised by these
deviations, i.e.that they are not predictable. Several previousstudies provide tests
on the performance of analysts' forecasts (see, for example,Pearce and Roley 1985
or Beckeret al1996). They suggestthat these forecasts are not always eÆcient,es-
peciallyifshorttestperiodsareused.Butatleast,mostofthetimetheyoutperform
commonly used time series models (Hardouvelis 1988, Moersch 2001). Test results
provided in table 7 conrm these results largely. On the 1% level, the eÆciency of
analysts' forecastscan be rejectedforonly1 out ofthe 24headlineguresanalyzed
here, i.e. for GDP
1
. Splitting the sample period into halves, no consistent pattern
of predictability remains.
2.2 The content of macroeconomic reports
ThisstudyinvestigatesU.S.macroeconomicreportswhicharereleasedonamonthly
orquarterly schedulewithinoortradinghoursof CBOTT-bond futures(see table
5).
6
To gain a better understanding of the relative importance of these reports, in
the remainder of this section headline gures are classied according to the type
of information they provide. Unexpected macroeconomic news may lead agents to
6
Foradescriptionofindividualreleasesandheadlineguressee,forexample,Rogers(1994).
tion rates or higher real rates. Previous studies have tried to identify whether an
announcement provides information that might alter market participants expecta-
tions of real rates or ination rates.
7
Following Edison (1996), headline gures are
categorized into two broad content groups: gures that provide ination measures
(C1), and others that indicate higher or lower levels of real activity (C2). Note
that incontrast tostudies like Hardouvelis (1988) or Dornau and Schroder (2000),
the purpose of this classicationis neitherto nd out whichmacroeconomic model
market participants might have in mind nor to assess the empirical relevance of
dierent models. The sole purpose is to identify reports with a similar information
content.Identifyingrelativelyhomogeneoussetsofinformationconstitutesthe basis
for testing whether the information content helps to explain the relative impact of
releases.
Higher levels of real economic activity may be associated with higher real interest
rates. If increasing economic activity is coupled with increasing investments, and
thus with a higher demand for capital, interest rates should rise given a nite elas-
ticity ofcapitalsupply.Informationabout highereconomicactivity mightalsoalter
agents' expectations of future ination rates, since ination could be spurred by
an overheating economy. Thus, an unexpected increase in real activity could drive
interest rates up through higher real rates and/or higher ination expectations.
Headline gures about economic activity are classiedas C1. Since they are rather
heterogeneous, three subcategories, i.e.(a) to(c), are distinguished.
7
Otherstudies, forexampleDwyerandHafer(1989),investigatealsomonetaryphenomenons.
Sincemoneysupplyguresarereleasedafteroortradinghours,i.e.at4:30p.m.ET,theyarenot
includedhere.
(a) Overall productionlevel:NAPM
1 , IP
1
, DGO
1
,GDP
1 , LI
1 , FI
2 .
(b) Demandfor consumption goods: CC
1 , R S
1 , PI
2 .
(c) Demand inhousingsector: HS
1
, NHS
1 , CS
1 .
The rst subcategory in C1 includes headline gures that provide evidence about
the overall production level: The industrial production gure (IP
1
, see table 5 in
the appendix), the level of the gross domestic product (GDP
1
), the index of the
National Association of Purchasing Managers (NAPM) 8
, the index of leading in-
dicators (LI),durablegoodsorders (DGO),and factory orders(FI
2 ).
9
The second
subcategoryofguresprovidesspecicinformationaboutconsumerdemand,e.g.the
retail sales gure (R S) and personal consumptionexpenditures (PI
2
). In addition,
consumer condence(CC)may permit someconclusions about thefuture spending
behavior of consumers. The thirdgroup of relatedgures covers the demand in the
housingsector,i.e.the numberofhousingstarts(HS),newhomesales (NHS),and
construction spending (CS).
C2: Ination expectations (and subcategories)
(a)Measures of past price changes: PPI
2 , CPI
2
,GDP
2 .
(b) Earlyinationindicators: E
2 , IP
2 ,ECI
1 , FI
1 , PC
1 ,BI
1 .
Classication(C2)includesmeasuresofination.Twosubcategories aredistinguish
8
This composite index is based ona questionnairecoveringseveralareas of business activity,
among themthe currentlevelof production, newordersfrom customers,and employmentin the
manufacturingsector.
9
Durablegoodsordersmeasure orders,shipments,andunlled ordersplacedwithU.S.manu-
facturersforgoodswithalifeexpectancyofatleastthreeyears,factoryordersincludenon-durable
goodsaswell.
tains gures measuring past price changes at the very end of the production pro-
cess, i.e.inationinnishedoralmostnishedgoods. Amongthemarethe monthly
consumer and producer price indices (PPI
2 , CPI
2 )
10
as well as the price deator
contained in the quarterly GDP report (GDP
2
). The second subcategory contains
indications of price pressures atearlier stages of the productionprocess and short-
ages of productionfactors. Whileseveral reports include such informationon price
pressures, e.g. raw material prices included in the producer price report, only for
two headline gures analysts' forecasts are available. These are labor costs (ECI)
and productivity (PC). Both, higher than expected wages and lower productivity,
mightsuggestthat inationpressures are buildingup,especiallyif wages rise faster
than productivity.Shortages of production factors which mighttranslateintoprice
pressures of input factors are indicated, for example, by a stretched capacity uti-
lization(IP
2
)orby lowinventories(BI
1 ,FI
1
).Furthermore,ifatightlabormarket
givesemployees more bargaining power, a lowerthan expected unemployment rate
(E
2
) may foreshadow higherwages and, thus, ination pressures.
2.3 Determinants of the relative impact of releases
This sectionderivessome hypothesesinordertoinvestigatethedeterminantsof the
relative impact of releases. The rst hypothesis (H2) follows immediately from the
very special timeseries properties of surprises inmacroeconomicreleases. Compare
these reports, for example, to a regular time series of temperature measurements
obtained at the end of each month. Instead of drawing one observation for each
10
Previousstudies usethe overall consumerand producerpricesindices. Instead, heretheless
volatilecoreinationnumbersareemployedwhichexcludefoodandenergy.
same time,but for sometechnical reason,these guresare not released atthe same
time. A similar structure is found for macroeconomic reports. There are several
macroeconomic reports referring to the same periodand measuring similar aspects
ofeconomic strengthand ination.Again, theseguresarenot released atthe same
time but witha moreor less extended timelagtothe reporting period(see gure1
in the appendix). It seems reasonable to assume that the price impact of the non-
anticipated informationinareportdependsonthe timelagtothe reportingperiod,
and thus, itsimpact onprices (hypothesis H2).
H2: Timeliness
The price impact of non-anticipatedinformation in a release depends on the
time between the announcement of the report and the end of the reference
period.
H2restatesanobservationpreviouslymadeby FlemingandRemolona(1997).They
ndthatthefourmostrecentavailablegovernmentreports(E,CPI,PPI,andR S)
have the highest impact on ve year T-note prices. From this they conclude that
the time between the end of the periodcovered by a report and the announcement
helps to explain the impactof a release.
It may be questioned whether the timeliness of a report is best measured by the
numberof days determiningthe time lag.Instead, areport may become 'outdated'
due to the factthat other reports whichprovidesimilar informationare released in
advance. This is statedby hypothesis H3:
The price impact of non-anticipated information in a release decreases with
the numberof previously released reports for agiven reference period.
Note that H2 and H3 are rather similar since both imply a monotonically declin-
ing impact of subsequently released reports. However, if one does not nd such a
strict relation, H2 will provide another testable implication: Reports with a time
lag of more than a month should have almost no impact since reports for the fol-
lowingcalendar monthare alreadyavailable.This shouldbe the case forLI andall
subsequently released reports (see gure 1).
Both, H2 and H3 ignore any dierences in the type of information. However, it
seems more reasonable to assume that market participants dierentiate between
variousaspectsof economicstrength andinationinordertoassessthe equilibrium
long-terminterest rate. Inthis case, a formulationof H3 that buildsonthe content
analysis oftheprevioussection(seeC1 andC2) seemsmoreappropriate.Therefore,
H4 explicitlydierentiates between the typeof information,hypothesizing that the
additionalinformation provided by a report for a given monthdiminishes with the
numberof already released reports with a similar content. For example,this would
imply that a gure like housing starts HS could have quite a signicant impact
since it is the rst gure providing evidence on the demand in the housing sector.
In contrast, accordingtoH3 (andH2)the impactof HS shouldberathermoderate
sinceseveralothergures-althoughhighlightingdierentaspects-wouldbealready
available.
The price impact of non-anticipatedinformation in a release depends on the
numberof previously released reports with asimilar content.
A strong argument in favor of this hypothesis comes from the fact that certain
gures repeat to some extent information contained in previously released reports.
A rather outdated gure in this sense is the factory orders number (FI
2
) since an
earlier estimatecan be derived fromboth, the reportondurablegoodorders which
account for over 50% of total factory orders, and the new orders component in the
NAPM report. Even worse, the NAPM reportfor the subsequent month is already
availablewhen FI comes out.
3 Empirical results
3.1 Data description
This study analyzes surprises in 24 headline gures contained in 19 dierent U.S.
macroeconomic reports over a 6 year period, i.e. January 1994 to December 1999.
Thisincludesmonthlyaswellasquarterlyreportsscheduledduringtheoortrading
hours of T-Bond futures at the Chicago Board of Trade. These are reports which
are released ateither8:30a.m.,9:15 a.m.,or10:00a.m.ET (EasternTime).Dueto
strict lock-up conditions, described for example in Fleming and Remolona (1997),
reports are released precisely according to the schedule.
11
A major disruption of
11
FlemingandRemolona(1999b)cite twoexceptionsofthisrule.Thesearetwooccurrencesof
inadvertentlyearlyreleased reports,i.e. theNovember1998employmentreport andtheJanuary
1999PPIreport.Nevertheless,thestrictlockupconditionsnormallypreventaleakageofinforma-
tion before the oÆcialreleasetime (see e.g.,Flemingand Remolona 1999).This is conrmedby
EderingtonandLee(1995)whondnosignicantlypositivecorrelationbetweenreturnsinseveral
intervalspreceding anannouncementand returnsimmediatelyafter anannouncement.
shutdownofseveralfederalagenciesduetoafederalbudgetdispute.Sincethewhole
forecasting process might have been aected, all observations from December 1995
through February 1996 are excluded. The sample contains 69observations for each
of the 16 monthly reports and 23 observations for the 3 quarterly reports in our
sample. Out of the 1497 trading days,on 769 days at least one reportis released.
Consensus forecasts, i.e. median analysts' forecasts, of headlinegures were gener-
ously provided by Standard & Poors Global Markets (MMS) for the period 1995{
1999. Earlier forecasts as well as actual outcomes were obtained fromseveral print
sources, especially the Wall Street Journal, Barron's, Financial Times, and Busi-
ness Week. Surprises, i.e. non-anticipated information in announced headline g-
ures, are calculatedsubtracting consensusforecasts fromthe actualoutcomes.Note
that non-revised gures are used as they were available to market participants at
the time of announcement. These gures may dier substantially from those avail-
ablefromstatisticalagenciestoday. 12
Foreachheadlinegurestandardizedsurprises
are computed dividing surprises by the sample standard deviationof outcomes (i.e.
S
i
=Std(A
i
)). Descriptive statistics for both, outcomes and non-standardized sur-
prises are given in table 6.
Intraday data on CBOT T-Bond futures trading provided by the Futures Industry
Institute are used.
13
Focusing on the front month contract, i.e. the most actively
traded contract among the nearby and second nearby contract, price changes over
12
Many macroeconomic gures are subject to several revisions afterwards. For example, the
initially reportedunemploymentrateisrevisedeveryJanuaryforthepreviousveyears.
13
This is a so-called tick-by-tick data set containing a time-stamped record whenever aprice
changeisobserved.Transactionvolumesarenotrecorded.
ve minute intervals around 8:30, 9:15, and 10:00 releases are analyzed. For ex-
ample, ve-minute price changes around 8:30 are calculated using the price of the
last transactionin the front monthcontract recorded before8:30 and the lastprice
before8:35.
3.2 Estimation of the impact of surprises on price changes
InordertotesthypothesesH2throughH4,asystemofthreeequationsisestimated,
one for each releasetime.
P
8:30
=
1 +
15
X
i=1
i S
(8:30)
i
+"
1
P
9:15
=
2 +
15
X
i=1
i S
(8:30)
i +
17
X
i=16
i S
(9:15)
i
+"
2
P
10:00
=
3 +
15
X
i=1
i S
(8:30)
i +
17
X
i=16
i S
(9:15)
i +
24
X
i=18
i S
(10:00)
i
+"
3
Here,P
8:30 ,P
9:15
,andP
10:00
denoteve-minutepricechangesaround8:30,9:15,
and 10:00 releases, respectively. S (time)
()
denotes standardized surprises occurring at
a given release time. For example, S (8:30)
1
denotes the surprise in headline gure
E
1
, i.e. nonfarm payrolls contained in the employment report which is released at
8:30 a.m. ET. These variables are zero if no such report is announced during a
givenve-minuteinterval.Notethatpriceresponsestosignedsurprisesareanalyzed
according to the hypothesized T-bond future reactions which are detailed in table
5.Forexample, sinceitis hypothesized that T-bondfutures pricesshould fallif the
announced non-farmpayrollgure(A
E1
)ishigherthan itsforecast(F
E1
)asurprise
in E
1
is calculated as (A
E
1 F
E
1
). Hence, a positive S
()
should be "good news"
for futures prices.
14
Untilaround3weeksbeforeexpiration,thenearbycontractisthemostactivelytradedone.
Signed Release 8:30 9:15 10:00
variable time equation equation equation
CC
1
10:00
18
11.287
NAPM
1
10:00
19
13.763
E
1
8:30
1
13.201
1
-.610
1
.201
+E
2
8:30
2
22.732
2
1.108
2
.527
PPI
2
8:30
3
6.725
3
.349
3
-.327
R S
1
8:30
4
5.801
4
-.065
4
.317
CPI
2
8:30
5
5.883
5
.315
5
-.212
IP
1
9:15
16
.197
16
-1.428
IP
2
9:15
17
17.274
17
2.557
HS
1
8:30
6
5.494
6
-.208
6
-.687
DGO
1
8:30
7
4.353
7
.018
7
-.355
ECI
1
8:30
8
3.086
8
.750
8
-.039
GDP
1
8:30
9
6.575
9
1.462
9
-.135
GDP
2
8:30
10
6.320
10
.005
10
1.927
PI
1
8:30
11
1.937
11
.487
11
-.905
PI
2
8:30
12
-1.776
12
-.105
12
-.903
NHS
1
10:00
20
7.345
LI
1
8:30,10:00
13
.154
13
-.576
13
4.151
CS
1
10:00
21
-.873
+FI
1
10:00
22
-.402
FI
2
10:00
23
4.559
+PC
1
10:00
24
-.139
+BI
1
8:30,10:00
14
.189
14
-.558
14
-1.053
+TR D
1
8:30
15
.349
15
.043
15
.050
AdjustedR 2
.309 .120 .226
BG(30) 17.484 26.877 28.270
Fiveminutepricechanges areregressed onstandardized signedsurprises (i.e. S
i;t
=Std(A
i
)),based
on asystem of three equations,one foreach releasetime, i.e. 8:30, 9:15,and 10:00.Parameters of
the systemare estimatedaccountingforheteroskedasticityand contemporaneouscorrelationin the
errors acrossequations. Estimated constants are omittedsince they are insignicant(see table 2).
Parameter tests are based on the heteroscedasticity consistent White variance-covariance matrix.
Signicance at the 1%, 5%, and 10%level is indicatedby
,
, and
, respectively. The sample
period is 1/94{12/99, including 1497 trading days. There are 69 observations for each monthly
report, and 23 for each quarterly report, resulting in 769 days on which at least one report is
released. Breusch-Godfrey (BG) tests have been performed for several lag lengths. Since none of
these is ableto detect autocorrelation in residuals,test statisticsare displayedonly forthe largest
lag length,i.e.30days.AdjustedR 2
sare providedforeachequation. Thesystem-wideR 2
measure
accordingtoMcElroy(1977)isR 2
=:226.
interval, and in addition, on surprises occurring earlier at a given day.
15
The
i
coeÆcients capture the immediate price impact of a release, i.e. the price change
occurring in the ve minute interval around the announcement. The
i s (
i
s) cap-
ture the impact of headlines being released 45 (90) minutes earlier at a given day.
If markets process information eÆciently one would expect that the impact of a
surprise is incorporated rapidly into prices. Hence, if release i has an impact, one
should ndasignicant
i
while
i and
i
should notbe signicantly dierent from
zero.
The system of the three equations isestimated by aseemingly unrelated regression
(SUR). Generalized least squares estimates are used in order to account for het-
eroskedasticity across trading days and contemporaneous correlation in the errors
across equations.The employed estimationtechnique yieldsresultsthat are asymp-
totically eÆcient.
16
Parameter estimates are provided in table 1 in the appendix.
Since signed surprises are used, only positive
i
s are in line with the hypothesized
immediate price response. Interestingly, with only one exception the signs of the
signicant
i
sare indeed positive.This exception isPC
1
. However, the impact ofa
15
The announcement timefor tworeports (i.e. LI and BI)changesfrom 8:30 to 10:00 within
our sample. Thus, the immediate impact of these releases has be captured in two equations. If
theseguresarereleasedat8:30,an
i
coeÆcientisincludedinthe8:30equation,otherwiseinthe
10:00 equation.However,thecorrespondingcoeÆcientsare restrictedtotakeonthe samevalues
in bothequations.
16
EÆciencygainsareprimarilytobeexpectedfromthefactthatrestrictionsareimposedacross
equations.Otherwise,iftheerrortermsin theabovegivenequationsareuncorrelated,aseparate
leastsquaresestimationoftheequationsyieldseÆcientparameterestimates,assumingwellbehaved
data(seee.g.,DwivediandSrivastava1978).Thecorrelationofsingleequationsresidualsestimated
with ordinaryleastsquares isindeed small but signicantlydierentfrom zeroat the10%level.
Hence,itcannotbetakenforgrantedthat theequationsareactuallyunrelated.Thisisconrmed
on the10%level byatest onthediagonalityof thevariance-covariancematrixof therst-stage
residuals(seee.g.,Baltagi 1999,Ch.10).
surprise in PC
1
is quite small. Note that the right hand variables in table 1 are
sorted according to the median report time lag (see gure 1). Thus, the ordering
reects the release sequence of the reports.
3.3 Tests of hypotheses
A prerequisite for ananalysis of the value of information is totest whether the T-
bond futures market processes information eÆciently. As a minimum requirement
previous releases should have no systematic impact on current prices. Thus, the
i
's and
i
's shouldbezero (hypothesis H1).Indeed, most of
i
s and
i
s are rather
small, but 15out of 32parameters are signicantlydierentfrom zero.In contrast,
17 out of the 24
i
s are signicant, 12 of them at the 1% level. The arising rst
impression that futures prices respond immediatelyto surprises in macroeconomic
announcements isconrmed by a Wald test (see table 2,row 3).
18
This test cannot
reject thehypothesisthatthe
i
sand
i
sarezero asagroup.Thus, after45(aswell
as90) minutes nosignicantsystematicimpactcan be found.In contrast,a teston
the immediate impact of surprises reveals that the hypothesis ^
i
=0, 8i =1;::24,
is strongly rejected (table 2, row 2).This is not surprising, since one would expect
that non-anticipated informationleads toan immediateprice response.
A much more criticalquestion for our analysis is whether this impact diers across
releases. Whetherthe immediate impactofsurprises ismore orless the sameacross
releasescanbetestedonthebasisoftherestriction^
i
=
,i.e.whethertheindivid-
ual rst-stage parameters
i
can be replaced by a commonsecond-stage parameter
17
Asurpriseofthemagnitudeofonesamplestandarddeviationoftheobservedoutcomesleads
to apricereactionofaneightofonetick.
18
Thesameresultisobtainedtestingseparatelywhetherthe
i
'sarezeroandwhetherthe
i 's
arezero(Hess2000).
. Takinginto account the variance-covariance matrix of parameters estimated in
the rst stage, an asymptoticallyconsistent and eÆcient estimate of
can be ob-
tainedbymeansofasymptoticleastsquares(seee.g.,GourierouxandMonfort1995,
Ch. 9).Results of this estimation are given in the right hand panel of table 2 (row
2). Beingnot able to reject this set of restrictions would implythat allthe releases
have virtually the same impact on prices, and thus would provide strong evidence
againstanyof thethreehypotheses(H2toH4).This isnot thecase. Therestriction
^
i
=
is strongly rejected by the highly signicant 2
statistics. In contrast, the
results of the Wald test for the lagged impact are conrmed by this test (row 3).
The restriction
^
i
;^
i
=
holds while the estimated common parameter
^
is not
signicantly dierentfrom zero.
Table 2: Waldtests and Asymptotic Least Squares.
Wald tests Asymptotic leastsquares estimates
2
test on 2
nd
stage
2
teston
Est. Imposed imposed Imposed parameter imposed
no. restrictions restrictions restrictions estimates restrictions
(1) ^
i
=0
2
(3)
= 2.48 ^
i
=
^
= 0:014
2
(2)
= 2.39
(2) ^
i
=0
2
(24)
=1261.10
^
i
=
^
= 3:266
2
(23)
=795.25
(3)
^
i
;^
i
=0
2
(32)
= 38.13
^
i
;^
i
=
^
= 0:004
2
(31)
= 38.13
Each line displays test results for a given set of parameters. The left hand side panel shows
Waldtestsrestrictingthegivenparametersetto zero,i.e.^
i
=0,fori=1;2;3(row1), ^
i
=0,
i = 1;:::;24 (row 2), and
^
i
= 0, i = 1;:::;17 as well as ^
i
= 0, i = 1;:::;15 (row 3). The
2
()
test statistic with degrees of freedom is given as well. The right hand panel displays
an alternative test: Instead of restricting a set of parameters to zero, these parameters are
restricted to a common value
which is estimated on the basis of asymptotic least squares.
Signicance tests of
^
are constructed from the asymptotic variance-covariance matrix of
restrictedparameters.A testof thenullhypothesis thatthe setof restrictionsholds isobtained
on thebasisofthe asymptotically 2
()
distributed statistic(fordetails see e.g.,Gourierouxand
Montfort1995, Ch. 9, 18).
,
, and
indicates signicance at the 1%, 5%, and 10% level,
respectively.
implies that the impact of a surprise declines with the number of previously re-
leased reports, H2 relates the impact to the time lag of a release. In contrast, H4
conditions on the sequence within content categories. In order to obtain a formal
test of these hypotheses, again a asymptotic least squares estimation is performed
imposing certain constraints on the parameters estimated in the rst stage. For
example, a somewhat strict form of H2 (timeliness) postulates that the impact of
releases declines linearly with the time lag of a release
i
. This results in the re-
strictions
i
=
0 +
1
i
, 8i=1;:::;24.Given that these restrictions hold, one can
test whether
^
1
is signicantly negative, as it is suggested by H2. Results of this
estimation are given intable 3, line(1). Line (2)contains results for H3 (sequence)
statingthattheimpactofsurprises decreaseswiththenumberofpreviouslyreleased
reports forthe same reportingperiod,n
i
,i.e.
i
=
0 +
1 n
i
.Line (3)to (7)provide
results for H4 (sequence within content categories), i.e.
i
=
0 +
1 c
i;j
, where c
i;j
represents the number of previously released gures fallinginto the same category
j as headlinei.
Judging from the estimated slope coeÆcients
^
1
in table 3, the impact of these
releases seems to be decreasing. These coeÆcients are all negative and all but one
highlysignicant.Thiswouldsupportallthreehypothesesifnotseveralofthesetsof
restrictions were rejected by the corresponding 2
statistics. Especially,hypotheses
H2andH3arestronglyrejected.Thissuggestthatalthoughtheimpactofsuccessive
releasesmay bedecreasingwith n
i
as wellast
i
,assumingthatthe impactdecreases
linearly istoorestrictive.In contrast, the resultsfor H4are slightlymore favorable.
Atleasttherestrictionofalinearlydecayingimpactwithincontentcategoriescannot
berejectedfortwo ofthe vecategories (C1b'Demandfor consumptiongoods' and
this restriction istoo strong.
Table 3: Test of hypotheses H2 toH4 by means of asymptoticleast squares
2 nd
stage
2
teston
Est. Imposed parameterestimates imposed
no. restrictions
^
0
^
1
restrictions
(1) ^
i
=
0 +
1
i
9:646
0:244
2
(22)
=329:47
(2) ^
i
=
0 +
1 n
i
9:468
0:478
2
(22)
=298:39
(3) ^
i
=
0 +
1 c
i;1
7:539
1:398
2
(4)
=132:23
(4) ^
i
=
0 +
1 c
i;2
12:217
6:882
2
(1)
= 0:67
(5) ^
i
=
0 +
1 c
i;3
9:398
4:951
2
(1)
= 21:55
(6) ^
i
=
0 +
1 c
i;4
6:594
0:497
2
(1)
= 0:28
(7) ^
i
=
0 +
1 c
i;5
6:717
1:707
2
(4)
= 83:21
Each line displaysresultsofan asymptoticleastsquares estimationfor agivenset of linear
restrictions. Line (1) and (2) provide tests for hypothesis H2 and H3, respectively, by
restricting estimated rst stage parameters ^
i
as a linear function of
i
, i.e. the time lag
of arelease, and n
i
, i.e. the numberof previousreleases. Results for H4 are given in lines
(3) to (7) testing whether the ^
i
's may be expressed as a linear function of c
i;j
, i.e. the
number of previous releases within a given content category j = 1;:::;5 (corresponding
to classicationC1a, C1b, C1c, C2a, and C2b,respectively). A test of the nullhypothesis
whether theconstraintshold isobtainedonthebasisof theasymptotically 2
()
distributed
statistic with degreesof freedom, t-statistics for
^
i
are constructed from the asymptotic
variance-covariancematrixofrestrictedparameters(seee.g.GourierouxandMontfort1995,
Ch.9,18).
,
,and
indicates signicanceatthe1%,5%,and10%level,respectively.
In order toobtain moreevidence, some less rigorousimplications ofthe hypotheses
are testedin theremainder ofthis section.Forexample,hypothesisH2 impliesthat
releaseswithatimelaglargerthanamonthshouldhavenoimpactsinceinformation
for the subsequent report period are already available. Interestingly, allbut one of
the reports being released within one month after the end of the reporting period
(i.e. CC to NHS, see gure 1) are signicant, most of them at the 1% level (see
table 1).In contrast,out ofthe remainingseven headlinegureswhichare released
preliminary evidence in favor of a somewhat loose interpretation of hypothesis H2.
Nevertheless,onthebasisofaWaldtestthehypothesisthattheirimpactasagroup
is zero has to be rejected. So, according to this test H2 has to be rejected again.
Note that the implications forhypothesis H3 are quitesimilar.
Turning tohypothesisH4, insteadof imposinglinear restrictions onthe coeÆcients
one mayperformaseries ofpairwiset-testsonthe dierenceofcoeÆcientswithina
content category.If the impactof reports decreases strictly within acontentgroup,
for each pair of successive reports the dierence between consecutive coeÆcients
should be signicantly positive. It is not surprising that this very strong result is
obtainableforonlyoneof theve contentcategories,since therestrictionofalinear
decay in the coeÆcients was already rejected for three of them. Nevertheless, very
strong evidenceofadecreasingprice impactcanbefoundinvestigatingasomewhat
less strict formulation of H4: The rst or second release within a content category
should have ahigherpriceimpact thanallthe subsequent releases. Table4displays
the results of pairwise comparisons of the rst and the second headline gure in
a given content category with the subsequently released headlines falling into the
same category. Tests for hypotheses H2 are H3 omittedsince the rst two releases,
i.e.NAPM andCC haveasignicantlylowerimpactthanthefollowingones.Thus,
H2 are H3 are easily rejected againby this test.
C1a: Overall production C2a: Past pricechanges
NAPM
1
IP
1
PPI
2
CPI
2
IP
1
13.565
CPI
2
0.841
DGO
1
9.410
-4.155 GDP
2
0.405 -0.436
GDP
1
7.188
-6.377
LI
1
13.608
0.043 C2b: Earlyindicators
FI
2
9.204
-4.361 E
2
IP
2
IP
2
5.458
C1b: Consumptiongoods ECI
1
19.646
14.188
CC
1
R S
1
FI
1
23.134
17.676
R S
1
5.486
PC
1
22.871
17.413
PI
2
13.063
7.577
BI
1
22.543
17.085
C1c: Housingsector
HS
1
NHS
1
NHS
1
-1.851
CS
1
6.366
8.217
Tests onadecreasingimpactof subsequentreleasesin individual contentcategories.The
rstand secondheadline gureiscompared with allsubsequentlyreleasedgureswithin
agivencontentcategory.DierencesbetweenestimatedcoeÆcientsaredisplayedforeach
ofthevecontentcategories.A positiveentryindicatesthat theimpactofthepreviously
released headline (i.e. theheadline guregiven ontop) is larger than the impact of the
report releasedsubsequently (i.e. theheadline gureto theleft). A signicantlypositive
(negative) dierence according to a one-sided t-test at the 1%, 5%, and 10% level is
indicatedby
,
,and
,respectively.Standarderrorsofthedierencesareconstructed
fromtheestimatedWhitevariance-covariancematrixofparameters.
As can be seen from table 4, in none of the categories the rst release's impact
is outweighed by subsequent releases. At worst, its impact is insignicantly lower
than thatofthesecondrelease(categoryC1c'Housingsector'). Forthreeofthe ve
categories, the rst release has a signicantly higher impact than all others. In a
fourth category (again,C1c),there is nosignicantdierencebetween the rst and
second release, but both dominate the third. There remains only one category in
whichsubsequent releases donot havea signicantlylowerimpact(C2a 'past price
H4 since the impact of subsequent releases is decreasing, too, although not signi-
cantly. Hence, the sequences of pairwise t-tests given intable 4 supporthypothesis
H4.
Overall,hypothesisH4cannotberejectedifonedoesnotdemand thattheimpactof
subsequent releasesdeclinesstrictly linearly.Asimilarresult cannotbeobtainedfor
Hypothesis H2 are H3 since both, the rst and second release have a considerably
lowerimpact than several subsequent releases. Thus, the type of informationplays
a substantial role in explaining the relative impact of non-anticipated information
in macroeconomic releases onT-bond futures price changes.
4 Summary and conclusions
T-bond futures prices like bond prices are driven mainly by market participants'
expectations of real interest rates and future ination rates. Therefore, the set of
headlineguresinscheduled macroeconomic releasesisdivided intotwobroad con-
tentcategories, newsrelatedtoinationexpectations andnews relatedtoeconomic
activity. Among them ve subcategories are distinguished. For example, ination
related news are dierentiated according to their time horizon, i.e. measurements
of past price changesin nished goodsversus indications of price pressures further
down in the production channel. Interestingly, all but one of the signicant coeÆ-
cients capturing the immediate futures price response tosurprises show the correct
sign, i.e. the introduced informationclassication may wellexplain the directionof
futures price changes.
Investigating the sequence of releases without dierentiating for content, at best
reports matters(hypotheses H3 and H2).On the one hand, the rather strict impli-
cation of a monotonically linearly decaying impact has to be rejected. But on the
other hand, the response to releases coming out in the rst monthafter the end of
the reportingperiodissomewhatstronger thanthe impactofreleasesannouncedin
thesecondmonth.Nevertheless,thereleasesinthesecondmontharestillsignicant
asagroupalthoughthis informationshouldberatheroutdatedsinceseveral reports
for the subsequent monthare already available.
Test results are much more favorable if headline gures are dierentiated by the
type of information. The rather strong hypothesis of a strictly linearly decreasing
impact within content categories is only rejected for three of the ve categories.
More importantly, a pairwise comparison of the impact of surprises reveals that
the rst and/or second release within a given content category has the highest
impact on prices. This leads to the conclusion that the type of information is an
importantdeterminantoftherelativeimpactofreleases(hypothesisH4).Thisresult
suggests thatmarketparticipantsconsidervariousaspectsofinationand economic
growth to be relevant in order to assess the equilibrium long-term interest rate.
Moreover, it impliesthat the information value of an additionalrelease for a given
reporting period decreases with the number of already available gures providing
a similar content. For example, marketparticipants seem tolearn enough from the
rst two housinggures about the strength of demand in that sector, and thus the
subsequently released gurehas almost noprice impact.
Table 5: Headline gures in macroeconomic reports and hypothesized reactions to
non-anticipated information.
Higheroutcomessignal Hypothesized
higherinterestratesdue to price
already higher supply response
Reporting higher consumer bottle- ofT-Bond
Abbr. Headlinegure agency a
prices demand necks futures
CC
1
Consumercondence index CB +
NAPM
1
OverallNAPMindex NAPM + + +
E
1
Non-farmpayrolls BLS +
E
2
Unemploymentrate +
PPI
2
PPIex.food andenergy BLS +
R S
1
Retailsales CENS +
CPI
2
CPIex.foodandenergy BLS +
IP
1
Industrialproduction FED +
IP
2
Capacityutilization +
HS
1
Housingstarts CENS + +
DGO
1
Durablegoodsorders CENS +
ECI
1
Employmentcostindex BEA +
GDP
1
RealGDP +
GDP
2
GDPdeator +
PI
1
Personalincome BEA +
PI
2
Consumptionexpenditures +
NHS
1
Newhomesales CENS +
LI
1
Indexofleadingindicators BEA
CS
1
Constructionspendings CENS +
FI
1
Factoryinventories CENS +
FI
2
Factoryorders +
PC
1
Productivity BEA +
BI
1
Businessinventories CENS +
TR D
1
Tradedecit CENS +
For each report, headline gures, reporting agency, and hypotheses concerning the reaction of
T-Bond futures prices to surprises in these gures are given. "+" (" ") indicates a positive
(negative) reactionto ahigherthan expected announcementof individual gures.For example,a
higherthan expectedconsumercondence indexsuggestsstrongerconsumerdemand which might
translateintopricepressures.Thus,apositiveimpactoninterestrates("+")andanegativeimpact
onT-bond futuresprices(" ")is tobeexpected.
a
BEA: Bureauof EconomicAnalysis,BLS: Bureau ofLaborStatistics,CB:Conference Board,
CENS: Bureau of the Census, FED: Federal ReserveBoard, TRES: Department of the Treasury,
NAPM: NationalAssociationofPurchasingManagers
Figure 1:Timeliness of reports
Foreach reportthenumberofcalendardaysbetweentheannouncementandtheendof
thereferencemonthisdisplayed(seetable5forabbreviations).Formonthly(quarterly)
releasesthisistheendofthecalendarmonth(quarter).Themediantimelagisindicated
byasquare.Asolidlinerevealstherangebetweentheminimumandmaximumnumber
of days.Announcementtimes arealsoprovided.Whilemostof thereportsare released
alwaysatthesametime,either8:30,9:15,or10:00a.m.ET,thetimescheduleofLIand
BI changes within the sample period, i.e. January1994 to December1999. Note that
Releasesduringthegovernmentshutdownperiodinearly1996areexcluded.Therelease
cycle is opened by thereport on consumer condence which is released usually during
thelast week ofthereference month.It isfollowedbytheNAPM report which usually
appearsattherstbusiness dayofthesucceedingmonth.
Table 6:Descriptive statisticsof released headlinegures and surprises
Outcomes Surprise
Headline Min. Mean Max. Std.dev. Min. Mean Max. Std.dev.
CC
1
80:8 113:9 139:0 17:51 7:5 1:0 13:3 4:06
NAPM
1
45:1 53:1 61:2 4:00 4:8 0:0 3:8 1:95
E
1
101:0 212:1 705:0 134:15 274:0 10:9 408:0 119:93
E
2
4:1 5:1 6:7 0:69 0:4 0:1 0:3 0:14
R S
1
0:8 0:3 1:5 0:48 1:1 0:1 1:1 0:39
PPI
2
0:5 0:1 0:8 0:20 0:6 0:0 0:4 0:18
CPI
2
0:1 0:2 0:4 0:08 0:2 0:0 0:2 0:08
IP
1
0:6 0:3 1:7 0:45 0:5 0:1 0:9 0:25
IP
2
80:1 82:9 85:7 1:47 0:6 0:1 0:7 0:30
HS
1
1:2 1:5 1:8 0:14 0:2 0:0 0:1 0:07
DGO
1
5:2 0:6 6:1 2:51 4:6 0:3 4:3 2:09
ECI
1
0:4 0:9 3:5 0:64 0:4 0:1 2:7 0:65
GDP
1
0:5 3:6 5:9 1:38 1:1 0:5 1:7 0:69
GDP
2
0:8 1:7 3:1 0:70 1:3 0:3 0:7 0:46
PI
1
0:3 0:5 1:4 0:30 0:6 0:0 0:5 0:16
PI
2
0:2 0:5 1:3 0:29 0:5 0:1 0:7 0:25
NHS
1
551:0 793:3 983:0 104:15 102:0 14:7 126:0 51:57
LI
1
0:6 0:1 1:3 0:30 0:2 0:0 0:2 0:10
CS
1
2:4 0:3 3:3 1:19 2:6 0:0 3:3 1:23
FI
1
0:9 0:2 1:0 0:33 0:7 0:1 0:8 0:35
FI
2
2:5 0:5 4:4 1:46 0:9 0:1 1:3 0:47
PC
1
0:2 1:9 4:5 1:61 1:4 0:3 1:5 0:91
BI
1
0:2 0:3 1:1 0:28 0:4 0:0 0:6 0:20
TR D
1
24:9 11:9 7:3 4:16 48:5 1:0 3:1 6:11
Displayedaretheminimum,mean,maximum,andstandarddeviationofinitiallyreleased
non-revised headline gures (left hand panel) and of non-standardized surprises (right
handpanel)forthetotalsampleperiod,i.e.1/1994to12/1999.Surprisesaremeasuredas
deviationsofannouncedguresfrom consensusforecasts.Consensusforecastsaredened
asthemedianofanalysts'forecastspolledbyStandard&PoorsGlobalMarkets.
Table 7: Test onthe eÆciency of MMS consensus forecasts
TotalSample Subsample Subsample
1/1994{12/1999 1/1994{12/1996 1/1997{12/1999
Headline F R
2
BG F R
2
BG F R
2
BG
CC
1
0:70 0:02 5:89 0:63 0:05 3:46 1:86 0:09 7:32
NAPM
1
0:58 0:02 2:01 1:94 0:12 1:45 0:57 0:03 1:27
E
1
2:38 0:05 0:50 2:93
0:09 1:21 0:37 0:02 1:65
E
2
2:45
0:06 0:62 2:73
0:16 0:64 2:37 0:11 1:29
R S
1
0:14 0:01 0:87 0:43 0:04 1:06 0:11 0:01 0:51
PPI
2
1:09 0:03 1:26 0:55 0:04 0:95 0:83 0:04 1:92
CPI
2
2:86
0:08 0:21 3:61
0:15 0:26 1:28 0:09 0:48
IP
1
4:17
0:16 0:47 4:49
0:13 3:51 2:12 0:22 0:51
IP
2
0:61 0:01 4:75 5:03
0:24 4:29 0:62 0:05 7:99
HS
1
0:01 0:00 1:90 0:20 0:01 0:78 0:23 0:01 1:06
DGO
1
2:73
0:06 5:80 2:16 0:08 1:71 1:50 0:08 3:39
ECI
1
0:22 0:03 9:93
5:49
0:46 5:43 28:61
0:53 1:08
GDP
1
8:09
0:45 1:13 1818:7
0:98 0:45 4:30
0:49 0:49
GDP
2
2:46 0:16 0:93 3:29 0:32 1:35 4:40
0:38 4:36
PI
1
0:31 0:01 4:20 0:07 0:00 2:25 0:74 0:02 8:36
PI
2
0:70 0:04 1:10 0:62 0:05 1:04 2:10 0:09 1:19
NHS
1
1:65 0:03 5:68 5:53
0:07 3:06 1:21 0:07 1:88
LI
1
0:90 0:02 3:67 1:04 0:03 0:88 0:13 0:01 10:29
CS
1
2:10 0:06 0:79 0:09 0:01 0:71 3:68
0:18 0:43
FI
1
1:76 0:06 0:70 3:04
0:15 0:90 0:27 0:02 4:49
FI
2
0:96 0:03 3:60 0:89 0:06 0:69 0:25 0:02 4:02
PC
1
0:45 0:05 4:06 0:40 0:00 4:00 0:74 0:15 4:22
BI
1
4:63
0:13 4:59 2:01 0:11 0:95 6:96
0:28 7:12
TR D
1
0:62 0:12 4:22 4:92
0:27 0:63 0:73 0:15 3:17
Surprises in each headline gure, i.e. deviations of announced from forecasted values
(S
i;t
= A
i;t F
i;t
), are regressed on the corresponding previous two outcomes by means of
ordinary least squares: S
i;t
=
0 +
1 A
i;t 1 +
2 A
i;t 2
. These estimations are performed on
the total sample period (1/1994{12/1999) as well as on two three-year subsamples. Ordinary
coeÆcients of determination are displayed (R 2
) along with the resultsof an F-test (F) of the
nullhypothesis H
0 :
1
=
2
=0 basedon theheteroscedasticityconsistentWhite covariance
matrixof parameters.Inaddition, resultsof aBreusch-Godfrey(BG)test onautocorrelationin
residualsupto velagsare reported.Signicance at the1%,5%, and10%level isindicatedby
,
,and
,respectively.
Balduzzi, P., E. J. Elton, and C. Green (1997): \Economic News and the
Yield Curve: Evidence for the U.S. Treasury Market," Discussion paper, New
York University, Oktober1997.
Baltagi, B. H. (1999):Econometrics.Springer, 2nd edn.
Becker, K. G., J. E. Finnerty, and K. J. Kopecky (1996):\Macroeconomic
News and the EÆciency of International Bond FuturesMarkets," Journal of Fu-
tures Markets, 16,131{145.
Berkman, N. G. (1978): \On the Signicance of Weekly Changes in M1," New
England Economic Review, May-June, 5{22.
Bollerslev, T., J. Cai, and F. M. Song (2000): \Intraday periodicity, long
memory volatility,and macroeconomic announcement eects in the US Treasury
bond market," Journal of Empirical Finance,7, 37{55.
Chang, Y., and S. J. Taylor (1996): \Information Arrivals and Intraday Ex-
change Rate Volatility,"Discussion paper, Lancaster Management School.
Christiansen,C.(2000):\Macroeconomicannouncementeectsonthecovariance
structureofgovernmentbondreturns,"Journal ofEmpiricalFinance,7,479{507.
Christie-David, R., and M. Chaudhry (1999): \Liquidity and maturity eects
aroundnews releases," Journal of FinancialResearch, 22,47{67.
ployment Report: The Role of Policy Anticipations," Economic Review, Federal
Reserve Bank of Richmond, 77,3{12.
Crain, S. J., and J. H. Lee (1995): \Intraday Volatility in Interest Rate and
Foreign Exchange Spot and Futures Markets," Journal of Futures Markets, 15,
395{421.
Dornau, R., and M. Schr
oder (2000): \Do Forecasters UseMonetary Models?
AnEmpiricalAnalysisofExchangeRateExpectations,"DiscussionPaper00/12,
Center of Finance and Econometrics, University of Konstanz, forthcoming in:
Applied Financial Economics.
Dwivedi, T., and K. Srivastava (1978): \Optimality of least squares in the
seeminglyunrelated regressions model,"Journal of Econometrics,7, 391{395.
Dwyer, G.-P., and R. W. Hafer (1989): \Interest Rates and Economic An-
nouncements," Review, Federal Reserve Bank of St. Louis, 71,34{46.
Ederington, L. H., and J. H. Lee (1993): \How Markets Process Information:
News Releases and Volatility,"Journal of Finance, 48,1161{1191.
(1995): \Theshort-run dynamics of the price adjustment tonew informa-
tion,"Journal of Financialand Quantitative Analysis, 31, 117{134.
Edison, H. J. (1996): \The Reaction of Exchange Rates and Interest Rates to
NewsReleases," DiscussionPaper570,BoardofGovernorsoftheFederalReserve
System.
Economic Policy Review, Federal Reserve Bank of New York,December, 31{50.
(1999a):\PriceFormationand Liquidityinthe U.S.TreasuryMarket: The
response topublic information,"Journal of Finance,54, 1901{1915.
(1999b):\TheTermStructureofAnnouncementEects,"Discussionpaper,
Federal Reserve Bank of New York.
Franke, G., and D. Hess (2000a): \The Impact of Scheduled News Announce-
ments onT-Bond and BundFutures Trading,"in Institutional Arrangements for
Global Economic Integration, ed. by H.-J. Vosgerau, pp. 337{366, MacMillan,
London.
(2000b):\Informationdiusioninelectronicand oortrading," Journal of
Empirical Finance, 7,455{478.
Goodhart,C.A.E.,andM.O'Hara(1997):\HighFrequencyDatainFinancial
Markets: Issues and Applications," Journal of EmpiricalFinance, 4,73{114.
Gourieroux, C., and A. Montfort(1995):Statistics and Econometric Models.
Cambridge University Press, Cambridge.
Hardouvelis, G. A. (1988): \Economic News, Exchange Rates, and Interest
Rates," Journal of International Money and Finance,7, 23{35.
Hess,D.(2000):\Surprisesinscheduledreleases:Whydotheymovethebondmar-
ket?," Discussion Paper 00-61, Centre forEuropean EconomicResearch,ZEW.
Jarrow, R. A. (1996): Modelling Fixed Income Securities and Interest Rate Op-
tions. McGraw Hill.
Glahn'sR 2
xy
and Hooper's r 2
,"Journal of Econometrics, 6, 381{387.
Mitchell, M. L., and J. H. Mulherin (1994):\The Impactof Public Informa-
tion onthe Stock Market," Journal of Finance,49,923{49.
Moersch, M. (2001): \Predicting Market Movers: A Closer Look at Consensus
Forecasts," forthcoming in Business Economics.
Pearce, D. K., and V. V. Roley (1985): \Stock Prices and Economic News,"
Journal of Business,58, 49{67.
Prag, J. (1994): \The Response of Interest Rates to Unemployment Rate An-
nouncements: Is There aNatural Rate of Unemployment,"Journal of Macroeco-
nomics,16, 171{184.
Rogers, R. M.(1994): Handbook of Key EconomicIndicators.McGrawHill,New
York.
Urich, T., and P. Wachtel (1981): \Market Response to the Weekly Money
Supply Announcements inthe 1970s," Journal of Finance, 36,1063{1072.
(1984): \The Eects of Ination and Money Supply Announcemnets on
Interest Rates," Journal of Finance, 39,1177{1188.
