Re: [Gretl-users] R: R: Rolling parameter estimate
On Fri, 27 May 2011, Oscar Soppelsa wrote:
I want to express my congratulations to Professor A. Cottrell and
Professor R. Lucchetti for the awesome gretl software. I'm studying the
manual and it's impressive.
Thank you!
I have written this code in order to estimate an ARMA(1,1) model on
I(0)
time series using a moving window which uses only the last 20 values
from the time series, "forgetting" gradually all the past information
older than last 20 values.
Apart from the technical details: please bear in mind that the standard
estimation method for ARMA is Maximum Likelihood, whose _only_
justification is asymptotic. Nobody knows what the properties of ML are on
a sample with 20 observations. So: I would consider it a very brave idea
to use a sample of anything less than 50-60 observations to estimate an
ARMA(1,1) model, and even 50-60 would make me shudder. If, for some
reason, you absolutely must use such a short window size, then you may
want to use CML, which at least should be quicker and numerically more
robust.
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti