EViews.com

Posted: **Thu Jul 21, 2016 5:36 pm**

EV 9 32bit build 2016-06-29

for an explosive process, EV9 seems not able to find the true coefficient.


!T = 200   '400 will more explosive
wfcreate u !T
rndseed(type=mt) 12347
genr e = 0.6*nrnd
genr u = 0
genr u2 = 0
genr u3 = 0
smpl @first+1 @last
u = 0.4*u(-1)+e			'U(1)=0, starting from unconditional mean
u2 = 1.04*u2(-1) +e   'explosive
u3 = 1.04*u3(-1) -e   'explosive, u3 = -u2 !
equation eq01.ls u ar(1)
if @vernum>=9 then
	equation eq02.ls(arma=cls) u2 ar(1)	'produce EV8 results in EV9; discard arma=cls in EV8
else
	equation eq02.ls u2 ar(1)
endif
equation eq029.ls u2 ar(1)		'ml, ar1<1
equation eq029o.ls(optmethod=opg) u2 ar(1)

show eq02
show eq029
if @vernum>=9 then
	show eq02.resids
	show eq029.resids
endif

you will see that EV9 use ML to estimate ARMA model, from eq029.resids, the residuals are not normal.
My question: does EV9 restrict the AR(1) coef to <=1 in estimation?

Posted: **Thu Jul 21, 2016 5:37 pm**

Posted: **Thu Jul 21, 2016 7:01 pm**

for this data set, NLS seems better than ML

Code: Select all


c(1) = 1.04
c(2) = 0.6
equation eq02u.ls(armastart=user,showopts) u2 ar(1)

even given the true value as starting value, eq02u estimates the same values as eq029

Posted: **Thu Jul 21, 2016 8:32 pm**

You should use nls on it then

EViews.com

ARMA and explosive process

ARMA and explosive process

Re: ARMA and explosive process

Re: ARMA and explosive process

Re: ARMA and explosive process