时间序列分析及VAR模型Word下载.docx
《时间序列分析及VAR模型Word下载.docx》由会员分享,可在线阅读,更多相关《时间序列分析及VAR模型Word下载.docx(17页珍藏版)》请在冰点文库上搜索。
一,+_i+f2t
Tliebeautyoftliismodelisthatwedon'
tneedtopredefinewhetherxoryareendogenous(thedependentvariable).Infact,wecantestwhetherx(y)isendogenousorexogenoususingGrangercausalitytests•TheideaofGrangercausalityisthatpastobservations(laggeddependentvariables)caninfluencecuiTentobservations一butnotviceversa・Sotheideaisrathersimple:
thepastaffectsthepresent,andthepresentdoesnotaffectthepast.STATAprovidesGrangercausalitytestsafterconductingaVARanalysis,whichisbasedontestingthejointhypothesisthatpastrealisationsdonotGrangercausethepresentrealisationofthedependentvariable.
Inmanyapplications,VARmodelsmakealotofsense,asacleardirectionofcausalitycannotbepredefined・Forinstance,thereisasubstantialliteratureonthebenefitsofinternationalisation(e.g.enteringforeignmarketthroughcross-borderM&
A).Thereisevidencethatmultinationalsoutperformlocalpeersduetothebenefitsofoperatinginmanycountries・Atthesametime,weknowthathigh-performingcompaniesaremorelikelytoenterforeignmarketsduetotheirownershipspecificadvantages・ThisargumentisbasedontheResource-basedViewandtheOLSframeworkdevelopedbyDunningandRugman(ReadingSchoolofInternationalBusiness).TheVARmodelallowsyoutoincorporatebotheffects:
infactyoucantestwhetherperformancedrivesinternationalisationorinternationalisationdrivesperformance.
BeforeyoustartusingaVARmodeLyouhavetomakesurethatthetimeseriesarestationary.SothefirststepistocheckwhetherthetimeseriesissStionjryusingDickey-FullertestsandKPSStests.Thesecondstepistospecifytheoptimallaglength(p)ofthemodel.Thisisdonebycomparingdifferentmodelspecificationsusinginformationcriteria.ApartfromusingAkaike(AIC)andBayesianSchwarz(BIC),theHannan-Quinn(HQIC)iscommonlyused.MostappliedeconometriciansfavourtheHannan-Quinn(HQIC)criterion・STATAwillhelpyoutomakeagoodchoice*Afterspecifyingyourmodel,youneedtocheckstabilityconditions.ThecoefficientmatrixofthereducedformVARhastoensurethattheiterationsequenceconvergestoalong-termvalue・STATAwillhelpyouincheckingstability.
Tobeprecise,youneedtoshowthattheeigenvaluesofthecoefficientmatrixliewithintheunitcircle・Thereasonbehinditcanbeonlyunderstoodwhenyouunderstandthemethodofdiagonalizingamatrix.
VARmodelsofferano什nicefeature:
impulseresponsefunctions・VARmodelscapturethedynamicsoftwo(ormore)stationarytimeseries;
hence,wecanassessthedynamicimpactofamarginalchangeofonevariableonanother・ThestandardOLSregressionprovidescoefficients,andcoefficientsrefertothepartialimpactofanexplanatoryvariableonthedependentvariable・InthecaseofVARmodels,therelationshipbecomesdynamic,asachangeofonevariable
(sayx)intcanaffectxandyint+1•Theimpactonxandyint+linturnaffectsxandyint+2andsoonuntiltheimpactdiesout.Impulseresponsefunctionsareveryusefulinillustratingtheshort-termdynamicsinamodel.
LefslookatanexampletoseehowVARmodellingworks・InLecture5,wetriedveryhardtounderstandgoldprices・Weextendourunivariatemodelbyexploringtherelationshipsbetweengoldandsilverprices・Linkingtwo(similar)assetsorsecuritiesisaveiycommontradingstrategy,whichiscalledpairs-tradin^•
Beforewedoanysophisticatedmodelling,itisalwaysbeneficialtolookatsomelinecharts.Figure1showstheindexedtimeseriesofnominalgoldandsilverpricesfrom1900to2010.
Figure1:
Nominalgoldandsilverprices,indexedA1900-2010
Wecanseethatthereisacertaindegreeofco-movement,whichwemightbeabletoexploitforourtradingstrategy.BeforewecanuseVAR.weneedtoensurethatbothtimeseriesarestationary.ItisobviousfromFigure1thatgoldandsilverpricesarenotstationary.However,aftertakingafirst-differencewecanshowthatpricechangesarestationary・Sobothtimeseriesare1
(1).
Thenextstepistodeterminetheoptimallaglengthusinginformationcriteria・Table1showsdifferentspecificationsusingthevarsoccommand・
Table1:
Determiningtheoptimallaglengthusinginformationcriteria
lag
LL
LR
df
P
FPE
AIC
HQIC
SBIC
104・118
・00049
■1.94511
•丄•92463
■JL-89456
1
113.551
JIR-866
4
0・00l
・000442
■2-0486
-1.98714
■JI・89694*
2
120.015
12.927
0.01?
•00042?
ifK
-2.09552*
■1.9931*
-1.R4276
3
120-ft19
1-6086
0-807
・00044R
■2.03465
■:
I・89126
-1.6R079
124-R26
8.0136
0・09l
・00044ft
-2.03478
-1.85042
-1.5798?
5
130.126
10.599*
0.031
・000438
-2.05954
-1.83421
-1.50347
Numberofobs
Endogenous:
return.greturn^sExogenous:
_cons
Selectioreordercriteria
Sample:
1906-2010
105
BasedontheAICandHQIC,twolagsareoptimal;
however,the(S)BICprefersonlyonelag.IwouldpreferHQICandtrytwolagsfirst.Ifthesecondlagdoesnotexhibitsignificantcoefficient,wecouldtrytoreducethelagleng什】inlinewith(S)BIC-
WerunaVARwithtwolagstoexplaincurrentpricechangesingoldandsilver-Table2providestheOLSestimates.
Table2:
VARmodelwithtwolags
1903-2010
IISI#I:
1^X%^1
dnoAdOO
・0004
Det(Sigma_ml)
二・OOOM323
NO-ofobsAIC
=108=・2・148455二・2・04776=•丄•90Oil
Parms
Fcmmlien
return.g5
RMSE
・1269270.2425
chi2P>
chi2
34.57860.0000
Vectorautoregression
return^s
5・1965690.130616.227630-0027
coef.
Std・Err.
P>
|Z|
[95%Conf.
Interval]
return^g
return.a
11-
・4ftR4in7
•l29QREA
2・QA
nnon
・7274SR1
L2-
-0139809
・122817
-0.11
0-909
-.2546979
・2267361
return.s
I:
I
-006R126
•6R05903
-0.0R
0.933
-164766R
・1511415
L2・
■・207786
・0807151
-2.57
0.010
-.3659847
■・0495874
_cons
・0277213
・0124857
2.22
O・026
・0032497
・0521929
return^sreturnAg
・如427弘
・1Q04R4ft
1AS
nOQQ
-0SRQ9S7
・AR7AR9A
・1085011
・1902038
0.57
0.568
-2642915
・4812937
・1094293
・124B0阳
0・RR
0.3R1
-.1351905
・354049
-3201805
・1250015
-2.56
-.5651789
■・0751821
・024511
.0193363
1.27
0-205
-.0133875
・0624095
Weseethatsilverprices(lag2)affectcurrentgoldprices,andwecanestablishautocorrelationinbothtimeseries*TotestwhethergoldGrangercausessilverorviceversa,werunGrangercausalitytestsreportedinTable3.
Table3:
Grangercausalitytests
GrangercausalityWaldtests
Equation
Excluded
chi2
dfProb>
chi2
return_g
returns
6.765
0.034
ALL
3-9615
0.138
Hence,weconfirmthatpastchangesinsilverpricescanpredictfuturegoldpricechanges.Thisisveryinteresting,asitcanbeusedtodevelopatradingstrategy.Finally,weneedtoshowthattheVARisstable(seeTable4).
Table4:
StabilityconditionoftheVAR
Eigenvaluestabilitycondition
Eigenvalue
Modulus
・2367286+
・362415?
.43288
•2367286-
・3624157
・43288
・06119136+
・3747777?
•37974
・06119136-
•37477771
・37974
Alltheeigenvalueslieinsidetheunitcircle・VARsatisfiesstabilitycondition・
Finally,wecanillustratetheimpactofsilverpricechangesonfuturegoldpricechangesusinganimpulseresponsefunction*Figure2showstheimpulseresponsefunctionandconfidenceintervalsderivedfrombootstrapping.Ifsilverpricesincreasetodayby1%,weshouldexpectasignificantdeclineingoldpricesintwoyearsby0.2%•
Figure2:
Impulseresponsefunction
r,return_s,return_g2•
024
step
95%Climpulseresponsef