Corn OLS Corn Mean
Figure 6.8 CAARs of WTO Complaint
Table6.8 Testing the WTO Complaint
ConstantMean OLS Theil
1-day 3-day 1-day 3-day 1-day 3-day
SoybeanFutures
T-Test
Prob(CAAR>0) .198 .157 .181 .156 .198 .190
Prob(CAAR=0) .395 .315 .361 .311 .395 .380
RankTest .492 .481 .481 .484 .462 .489
Prob(CAAR=median)
SignTest .207 .207 .207 .051 .207 .207
Prob(CAAR=0)
EU CornFutures
T-Test
T-Test
Prob(CAAR>0) .110 .135 .177 .132 .105 .177
Prob(CAAR=0) .220 .270 .356 .264 .210 .353
RankTest .464 .488 .479 .493 .471 .486
Prob(CAAR=median)
SignTest .5 .5 .207 .5 .5 .207
Prob(CAAR=0)
Note: Values in ells denote probability under null hypothesis, whih is
statedinitalis. Probabilitiesbelow0.5arein bold.
6.3 Time Series Regression
Thissetion presentsthe resultsof atimeseries regressionof North
Ameri-ansoybean futuresreturnsandEuropeanornfuturesreturnsrespetively.
Although the data struture suggests an event study approah the events
are very lose together in time, they `luster'. Therefore, it is diult to
nd a suitable estimation window and hypothesis tests with a time series
regression, inaddition,seem advisable.
6.3.1 Soybean Futures
I hoose a T-GARCH model over a GARCH or an EGARCH in view of
the Akaike Information Creterion (AIC). I speify for the soybean futures
returns the following model: As mentioned earlier there are ARCH eets
in the series so a GARCHmodelis mandated. I inlude two ARCH terms
and one GARCH term and the GARCH(2,1) model. A lagged-dependent
variable would be insigniant. I speify Student-T distributed errors
be-ause the model t is better in terms of kurtosis of the residuals. I donot
inlude a onstant beause ommodity futures `returns' do not represent
a real return on a investment in terms of real prodution but rather the
hedgingand speulating behaviourof investors. I inludetwo market
vari-ables namelythe DowJones Indexand the CRB CommodityIndex while I
determinethe lagstruture viewofthe lowerAIC, whihindiatesabetter
goodness-of-t 2
. The variane equation of the modelontains several
on-trolvariablesasdisussedinSetion5.5.3namelyadummyfortheplanting
season and adummy for the rollover.
The results for soybean futures returns are in line with my previous
ndings(see Table6.9). News onregulatorymeasures bythe EU thathave
tigthenedregulation(CONTRA-GMOEU)havedereasedpriesbyaround
half a perent on the event day eteris paribus (see Column II). Pivotal
deisions (VETOPOINT) do not bear an inuene (see Column III), but
the adaption dereases soybean futures pries by around one and a half
perent ontheeventday plus thetwofollowingdays(see ColumnIV). The
2
Note that if both dependent and independent variables are expressed in logs, the
interpretation is anelastiity, meaning that `a oneperent hange in the independent
variables ause all other variables held onstant aX perent hange in the dependent
variable'. Furthermore,thenullhypothesishangesto
H 0 : β = − 1
. Thedefaultoutputof statistial software is usually atest for
H 0 : β = 0
and the unaware researh ouldanomalyof positivereturns aroundPro-GMOannounements (PRO-GMO
GOVERNMENT)isrepeatedinthe regression framework(see ColumnV).
The WTO omplaint does not have a signiant impat on returns (see
ColumnVI)
Therolloverdereasesthe varianeandduringtheplantingseasonthe
variane is signiantly higher but the oeients are tiny in omparison
with the movingaverage of the variane. Most of the variane isexplained
by the moving average. Shoks to the variane quikly die down. There
is an aysemtry in shoks. The negative sign indiatesthat positive shoks
havea larger impaton the variane thannegative shoks.
The testdiagnostisindiatethatthe residualsthat arefreeofARCH
eets told by Q-Statstis and an ARCH-LM test of up to 10 terms. The
residualsare free of serial autoorrelationtold by Q-Stastiis of severllag
lenghts. A Jarque-Bera test of the residuals is, nevertheless, strongly
re-jeted, whih is ommon for nanial time series. Non-normal residuals
violate an assumption of the GARCH model and the results ought to be
treatedwith due aution.
6.3.2 Corn Futures
The GARCH model for the orn futures series enompasses two ARCH
terms and one GARCH term and distinguishes between positive and
neg-ative shoks, a so alled TARCH model. Only then I annot detet any
more ARCH eets as told by ARCH-LM tests of various lag lengths. An
autoregressive element of orderone AR(1)is inluded toombat serial
au-toorrelation in the errors. Thus I speify an AR(1)-GARCH(2,1) model.
Boththe varianeand the meanequation inludeaontrolvariableforthe
rollover; the variable inthe meanfuntion is a dummy and in the variane
funtionaindexountingtomaturity. Notethelargeandpositiveoeient
of the lag dependent variable in the mean funtion indiates that returns
bear inertia. The variane funtion inludes as well a dummy variable for
the planting season.
Table6.11presentstheresultsofGARCHregressionofEuropeanorn
futures returns. The regression indiates that around half a perentage
point of the positive returns is attributable to news that the EU is losing
its market to GMOs (see Column II CONTRA EU). None of the other
politial events does bear aninuene on average onreturns (see Columns
Table6.9 GARCH of Soybean Futures Returns
DOWJONES
t − 2
0.046*** 0.047*** 0.047***(0.016) (0.016) (0.016)
COMMODITY INDEX 0.033 0.032 0.033
(0.033) (0.032) (0.033)
Log likelihood 12887 12893 12891
Akaikeinfo riterion -5.957 -5.958 -5.957
Shwarzriterion -5.943 -5.940 -5.941
LB(2) .003 .001 .002
Note: *,**,***denotesignianeat the10-,5-and1-perentlevel.
Parenthesisbehindvariablesindiatetheeventwindowsize
(b, a)
.Table 6.9 GARCH of Soybean Futures Returns (ontinued)
DOWJONES
t − 2
0.046** 0.047*** 0.047***(0.016) (0.016) (0.016)
COMMODITY INDEX 0.033 0.033 0.032
(0.032) (0.033) (0.033)
Log likelihood 12895 12893 12891
Akaikeinfo riterion -5.959 -5.958 -5.958
Shwarzriterion -5.943 -5.940 -5.941
LB(2) .003 .002 .002
Note: *,**,***denotesignianeat the10-,5-and1-perentlevel.
Parenthesisbehindvariablesindiatetheeventwindowsize
(b, a)
.Table 6.11 GARCH of Corn Futures Returns
COMMODITY INDEX 0.047** 0.05** 0.046**
(0.022) (0.022) (0.023)
EUROSTOXX
t − 1
0.009 0.011* 0.010(0.006) (0.006) (0.007)
β t − 1
0.000005*** 0.00005*** 0.000004***(0.0000006) (0.0000006) (0.0000006)
Log likelihood 6707357 6709558 6707191
Akaikeinfo riterion -7.371 -7.371 -7.370
Shwarzriterion -7.338 -7.332 -7.333
LB(2) .022 -.001 .020
Note: *,**,***denotesignianeat the10-,5-and1-perentlevel.
Parenthesisbehindvariablesindiatetheeventwindowsize
(b, a)
.6.11 GARCH of Corn Futures Returns (ontinued)
COMMODITY INDEX 0.047** 0.048** 0.046**
(0.023) (0.022) (0.023)
EUROSTOXX
t − 1
0.010 0.009 0.011(0.006) (0.006) (0.006)
β t − 1
0.00005*** 0.000005*** 0.000005***(0.0000006) (0.0000006) (0.0000006)
Log likelihood 6707020 6707670 6706291
Akaikeinfo riterion -7.369 -7.369 -7.369
Shwarzriterion -7.333 -7.330 -7.332
LB(2) .002 .009 .004
Note: *,**,***denotesignianeat the10-,5-and1-perentlevel.
Parenthesisbehindvariablesindiatetheeventwindowsize
(b, a)
.Theontrolvariablesinthevarianefuntionaresigniantbutsmall
in magnitude. Note that the predition error of two days ago is the best
preditorof the today'sreturndeviation. This meansthat volatilityisvery
spiky and doeshardly ontain any long-termomponent. Corn futures
re-turnsreatdierentlytopositiveandnegativeshoksasdosoybeanfutures,
but orn futures reat more strongly tonegativeshoks.
The diagnostis of the model are mixed. ARCH and serial
autoor-relation are taken are of by the ontrol variables, still the residuals are
not normally distributed. Mean and standard deviation are as expeted,
but the kurtosis of the residuals is high an exess kurtosis of around 7
indiating that the variane funtion annot predit spikes in returns.
Therefore, the Jarque-Bera test of normality is rejeted, asting doubt on
the goodness-of-tof the model.
Volatility Analysis As stated in Hypothesis 3 and 4 that I expet the
events toinuene the variane as well. However, I refrain from analysing
thevarianebeauseIdonotonsiderthevarianeequationsuientlywell
speiedespeiallyfor orn futures returns. The orn futures'volatility
re-ats mostly to short term inuenes while the volatilityof soybean futures
is mostly determined by the moving average. Any other regressor in the
variane equation is dwarfed in omparison with the ARCH and GARCH
terms. Bothtimeseriessuer fromhighkurtosisofthe residuals,whih
vio-lates the assumptionof normallydistributed errorsfor the GARCH model.
Surges in the variane are hardly preditable by the regressors. Speifying
Student-Tdistributederrorshasnotremediedtheproblem. Theoeients