Re: [Gretl-devel] arima issues
I'd like to start from example
function scalar get_maxgrad(void)
outfile mybuf --buffer --quiet
set bfgs_maxgrad
end outfile
mybufs = strsplit(strstrip(mybuf))
n = nelem(mybufs)
return atof(mybufs[n])
end function
open greene5_1.gdt -q
logs *
old_maxgrad = get_maxgrad()
reallyhuge = $huge
set bfgs_maxgrad reallyhuge
arima 1 1 1; l_realcons --nc
lnl1 = $lnl
# ? eval $coeff
# 1.0000
# -0.99937
set bfgs_maxgrad old_maxgrad
arima 1 1 1; l_realcons --nc --x-12-arima
lnl2 = $lnl
# ? eval $coeff
# 1.0000
# -0.99864
eval lnl1 - lnl2
#0.0014629666
In fact, 197 works better than x-12!
Estimates are very similar, and 197
gives greater likelihood
Here polynomial roots indicate
missing constant in model (linear trend in data)
and not missing second unit root (we have d=1)
arima 2 1 0; l_realcons
modtest --autocorr
Behaves quite well
Ljung-Box Q' = 3.29461,
with p-value = P(Chi-square(2) > 3.29461) = 0.1926
Insofar I have found only very similar
instances of non-convergence
It seems, before examples of non-convergence
with well-behaved roots are found everything
is needed is some additional text in the help entry/User's guide
Oleh
It seems the
16 жовтня 2018, 00:49:53, від "Allin Cottrell" <cottrell(a)wfu.edu>:
On Sat, 13 Oct 2018, Sven Schreiber wrote:
> Am 13.10.2018 um 13:35 schrieb oleg_komashko(a)ukr.net:
>> Dear all,
>> Below is a script illustrating
>> some problems with arima command
>
> Hi Oleh,
> excellent testing! This must come from the internal switch to the AS197
> algorithm. Good that you found those bugs.
Good find, yes, but I think there's little if anything here
associated with the switch from native Kalman to AS197: the problems
are mostly either (a) of long standing and somehow never before
revealed or (b) associated with our "new" method of fixing up
out-of-bounds initializations of the ARMA parameters.
I might add that the new(ish) initialization fix-up was shown by
Jack and myself to give much better results in most of the ARIMA
test cases posted by Oleh a while back. However, it seems we missed
some counter-examples (notably including the crash consequent upon
giving an all-zero initialization, which is now fixed).
Allin
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel