#338 deseas() segfault

[Artur Tarassow]

Hi,

While working on the replication of Hyndman's book, Claude found a bug.

deseas() with seats=1 segfaults for some sample windows of aus_production.gdt (attached) Gas (automatic ARIMA, logtrans=1)

Environment: gretl 2026d-git; x13as at /usr/bin/x13as (Debian package)

Steps to reproduce:

# Monthly data (T=202) -> works
open hamilton.gdt
bundle B = _(output="all", logtrans=1, seats=1)
series s = deseas(PC6IT, &B)

# Quarterly data (T=218) -> crash
open aus_production.gdt
# smpl obs >= 1980:1 --restrict --permanent # would run it fine!
bundle B = _(output="all", logtrans=1, seats=1)
series s = deseas(Gas, &B)

Expected: SEATS decomposition is returned, as it is with seats=0 on the same series.

Actual: gretl terminates with a segfault (exit code 139).

Sample windows tested (Gas, logtrans=1, seats=1, default ARIMA selection):

Sample	T	Result
1956:1–2010:2	218	crash
1956:1–1980:4	100	ok
1956:1–1981:4	104	ok
1956:1–1983:4	112	ok
1956:1–1984:4	116	crash
1956:1–1985:4	120	crash
1960:1–1990:4	124	ok
1980:1–2000:4	84	ok
1980:1–2010:2	122	ok
1956:1–2010:2, seats=0	218	ok

Further observations:

• From a 1956:1 start, every window from T=116 upward crashes. Sample length matters, but it is not the only factor: the 1960:1 and 1980:1 starts run at T=122 and T=124.
• The monthly hamilton.gdt series (T=202) runs without error.
• airline=1 on the full sample runs without error, as does logtrans=0.
• With output="sa" and verbose=1 on the full sample, the run fails with "Bei Fehler beendet" and a missing ~/.gretl/x12arima/Gas.s11 file. This may be a second symptom of the same problem.
• Note: Series Gas is of type integer (don't know this is really relevant)

Workarounds:

• airline=1
• logtrans=0
• Restrict the sample to start at 1980:1

Not yet investigated: whether the failing windows share a common cause, such as a particular observation or the model x13as selects.

Best,
Artur

[Riccardo "Jack" Lucchetti]

I just pushed to git what I believe is a fix: basically, in the case you reported x13 finished with an error, and we were trying to fill the returned series with values from a no-longer-existing matrix.

[Artur Tarassow]

Thank you, @jacklucchetti. I think also @allin has pushed something on this.

Overall, I now obtain an error message instead of a crash.

Thank you for the quick fix, Jack and Allin!

Shall we close the ticket, or anything left?

Artur

[Allin Cottrell]

Well, there's one thing left. The crash that Jack fixed was specific to the case where we get an error from x13as. Why did we get an error? It seems that x13as itself segfaulted when calling the SEATS code. Not exactly a bug in x13as, but due to out-of-date fortran code which gfortran used to work around but not (by default) any more. I'm in the process of replacing the gretl builds of x13as that exhibit this problem with new builds using the gfortran flag "-std=legacy", which reinstates the workaround. Plus, Claude pointed something out: in the specific crashing case there was also a "legitimate" error from SEATS: an inadmissible decomposition with negative spectrum. By default SEATS flags an error in such cases, but by setting noadmiss="yes" we could get it to reformulate the model to obtain an admissible decomp. Claude sugegsted we might want to add an option to that effect to our deseas() function.