Re: [Gretl-devel] ARMA developments
On Sun, 29 Apr 2018, Sven Schreiber wrote:
Am 28.04.2018 um 11:29 schrieb Riccardo (Jack) Lucchetti:
>> Am 27.04.2018 um 23:43 schrieb Allin Cottrell:
>>> Thanks go to Oleh for pushing us in this direction!
> The best way I can think for making the speed improvement useful would be
> adding some kind of "auto" option to the arima command to perform automatic
> selection of the arma orders.
Or Oleh, what was again your original motivation for wanting a faster ARMA
code?
I'm pretty sure it was just what Jack describes: automatic ar(i)ma
selection. As Oleh framed this task, it involved estimating a
potentially large number of ar(i)ma models having a sum of AR and MA
orders less than or equal to such-and-such. In that context many of
the specifications are likely to be over-parameterized (perhaps
grossly so); and with the maximizer requiring a large number of
iterations, "slow" computation of the likelihood will be mercilessly
exposed.
I take your basic point, in that our old, "slow" computation was
unlikely to be a real problem with sane ARMA specs -- and so
speeding it up was not a top priority -- but it's nice that we've
been able to come up with what I think is now the fastest ARMA
around (faster than x12a/x13a, and much faster than R's arima()).
Allin