[Gretl-devel] Re: seasonal adjustment: gretl + X13 compared with JDemetra(+)
Am 04.02.2022 um 19:37 schrieb Allin Cottrell:
Thanks, I've now installed RJDemetra (RJD for short) and can confirm
the results you posted. In the cases where there's noticeable
disagreement (RSA3, RSA5c) the programs are either arriving at a
different ARIMA specification, choosing different calendar effects, or
finding different (but overlapping) outliers. Here's a summary,
gleaned from the verbose printouts from RJD and X13A.
Thanks for the analysis,
Allin.
RSA3
----
RJD:
arima (3 1 1)(0 1 1) in levels + 5 outliers
(outliers: TC (4-2020), AO (3-2020), LS (11-2008),
AO (5-2011), TC (2-2009)
Log likelihood = -898.2, aicc = 1819
X13A, b58:
arima (2 1 0)(0 1 1) in levels + 3 outliers
(outliers: LS (11-2008), TC (3-2020), TC (4-2020)
Log likelihood = -916.1, aicc = 1846.5
X13A, b39:
arima (3 1 1)(0 1 1) in levels + 3 outliers
Log likelihood = -914.5, aicc = 1847.5
So this is a case where the two X13A builds I've tried disagree. Build
58 struggles with RSA3 on these data -- takes over 20 times as long as
any of the other specs -- while build 39 completes in reasonable time.
What's
the background on those specific builds, I mean why did you
choose exactly 39 and 58? I suppose one is the latest, but the other?
If we believe RJD's loglikelihood and AICC, it's doing a
little better.
I don't think we should compare the likelihood over different
specifications, especially different number of outliers. OK, the AICC
may be a different story, but for outlier-style regressors it's still
not so clear.
RSA5c
-----
RJD:
arima (2 1 1)(0 1 1) in levels + 4 outliers + 8 calendar effects
(calendar: 6 daily coeffs plus Leap year + Easter[1])
Log likelihood = -799.1, aicc = 1634
X13A, b58 and b39 (identical):
arima (2 1 1)(0 1 1) in levels + 5 outliers + 2 calendar effects
(calendar: Weekday (1-coeff) and Easter[1])
Log likelihood = -795.1, aicc = 1617
In this case X13A does better wrt loglik and AICC.
Again, more outliers get a
better AICC. I wonder if that's already a
pattern.
I'm not sure if the respective programs are starting with the
same
"test-down-from" spec for calendar effects; it doesn't look like it.
Do you know what "Weekday (1-coeff)" means? The RSA5c spec starts with
different regressors for each day of the week AFAIK, and our
"trading_days" is transformed into X13's "td" if I remember
correctly.
Does it mean of those regressors only one survived? Or is there a
problem somewhere and it effectively switched to "td1coef" somehow
(which our "working_days" is supposed to trigger)?
In general there are a billion ways to implement this search strategy,
of course. In the X13 doc I read somewhere that they hold the Arima
model fixed for that search; this is very reasonable, but obviously a
simplification due to the computational burden. Jdemetra being more
recent, maybe they skip some of the simplifications.
(Plus, there's the notorious leap year handling...)
cheers
sven