6:56 a.m.
Many thanks Artur,
However, I was puzzled mostly by the different behavior of different
versions, in a case that seems to me they should behave in the same
manner. In any case, I have another question: if one modifies the model
so that there is no dynamic behavior, the --static option still gives
one period forecast, though it is not strictly necessary. It is
documented behavior, but is what one would expect? Your modified example
is attached (test_fcast2.inp).
Concerning the original example, in test_fcast3.inp (attached), I have
increased the out of sample forecast period to 10 periods. In 2019a
version the script works, in 2019c not (now on Win 10 both, though I
think the OS is irrelevant). I am assuming that all forecasts are
constructed in the correct way, i.e. using the last period forecast as
the lag of the endogenous variable from the second forecast period
onward. What created the difference?
Many thanks again,
Andreas
7:20 a.m.
New subject: fcast behavior
Am 15.11.2019 um 07:56 schrieb ΑΝΔΡΕΑΣ ΖΕΡΒΑΣ:
Many thanks Artur,
However, I was puzzled mostly by the different behavior of different
versions, in a case that seems to me they should behave in the same
manner.
Hi, it would help if you could describe what you expected and why do you
think it should work. Note the doc for fcast: "If an out-of-sample
forecast is requested but no relevant observations are available, an
error is flagged." That's what's happening. Perhaps older gretl versions
had a bug there. (Certainly the changelog shows a somewhat related crash
bug fixed in 2019c.)
In any case, I have another question: if one modifies the model
so that there is no dynamic behavior, the --static option still
gives
one period forecast, though it is not strictly necessary. It is
documented behavior, but is what one would expect? Your modified example
is attached (test_fcast2.inp).
At least the comments in that script are partly wrong. "smpl 1 199" and
then "smpl ; -10" does not imply "Train on sample T=1..190". (Easy to
see also in the ols output.)
Concerning the original example, in test_fcast3.inp (attached), I
have
increased the out of sample forecast period to 10 periods. In 2019a
version the script works, in 2019c not (now on Win 10 both, though I
think the OS is irrelevant).
It is indeed interesting if the script works in earlier versions. But so
far the current behavior seems fine to me, so I would tend to view it as
some kind of bug in the older version. If you think the old behavior is
correct, please provide more details about the old results. (Since we
want to avoid installing old versions for testing if not absolutely
necessary.)
thanks
sven
9:07 a.m.
New subject: fcast behavior
In test_fcast3.inp, the exogenous variable is available for all out of estimation sample
periods, so I suppose that forecasts should be generated in all different ways, except
perhaps in the --static option case. If this is not the expected behavior, shouldn't
this be clearly stated in the documentation?
Thanks again,
Andreas
10:14 a.m.
New subject: fcast behavior
Am 15.11.2019 um 10:07 schrieb anzervas(a)yahoo.com:
In test_fcast3.inp, the exogenous variable is available for all out
of estimation sample periods, so I suppose that forecasts should be generated in all
different ways,
Ah, now I see what you mean. Indeed the estimated model uses obs until
190 (because the innovation u is only available up to there), and then
"fcast 191 200" and "fcast --out-of-sample" should be equivalent, I
agree.
Perhaps gretl is confused by the fact that here $t2 != $model.t2, i.e.
that the estimation sample does not correspond to the active workfile
sample. But it shouldn't be.
Can somebody test on 2019b?
thanks
sven
1:48 p.m.
New subject: fcast behavior
On Fri, 15 Nov 2019, Sven Schreiber wrote:
Am 15.11.2019 um 10:07 schrieb anzervas(a)yahoo.com:
> In test_fcast3.inp, the exogenous variable is available for all out of
> estimation sample periods, so I suppose that forecasts should be generated
> in all different ways,
Ah, now I see what you mean. Indeed the estimated model uses obs until
190 (because the innovation u is only available up to there), and then
"fcast 191 200" and "fcast --out-of-sample" should be equivalent, I
agree.
Perhaps gretl is confused by the fact that here $t2 != $model.t2, i.e.
that the estimation sample does not correspond to the active workfile
sample. But it shouldn't be.
Can somebody test on 2019b?
I can confirm that Andreas's test_fcast3.inp worked in gretl 1.10.1.
And I guess Sven has diagnosed the problem: a new ambiguity or
confusion over what "out of sample" means. I'll take a look at the
fcast code when I get a chance.
Allin
8:31 p.m.
New subject: fcast behavior
On Fri, 15 Nov 2019, anzervas(a)yahoo.com wrote:
In test_fcast3.inp, the exogenous variable is available for all
out of estimation sample periods, so I suppose that forecasts
should be generated in all different ways, except perhaps in the
--static option case.
That's now fixed in git and snapshots. The static out-of-sample
option produces just a single forecast, which is right since we
immediately run out of actual lagged y values, but the dynamic
forecast is available to the end of the dataset.
Allin
1:12 p.m.
New subject: fcast behavior
Am 15.11.2019 um 21:31 schrieb Allin Cottrell:
On Fri, 15 Nov 2019, anzervas(a)yahoo.com wrote:
> In test_fcast3.inp, the exogenous variable is available for all out of
> estimation sample periods, so I suppose that forecasts should be
> generated in all different ways, except perhaps in the --static option
> case.
That's now fixed in git and snapshots. The static out-of-sample option
produces just a single forecast, which is right since we immediately run
out of actual lagged y values, but the dynamic forecast is available to
the end of the dataset.
Thanks Allin.
This may be the first serious bug that was found in 2019c, maybe
warranting a new release?
cheers
sven
1:16 p.m.
New subject: fcast behavior
On Sat, 16 Nov 2019, Sven Schreiber wrote:
Am 15.11.2019 um 21:31 schrieb Allin Cottrell:
> On Fri, 15 Nov 2019, anzervas(a)yahoo.com wrote:
>
>> In test_fcast3.inp, the exogenous variable is available for all out of
>> estimation sample periods, so I suppose that forecasts should be
>> generated in all different ways, except perhaps in the --static option
>> case.
>
> That's now fixed in git and snapshots. The static out-of-sample option
> produces just a single forecast, which is right since we immediately run
> out of actual lagged y values, but the dynamic forecast is available to
> the end of the dataset.
Thanks Allin.
This may be the first serious bug that was found in 2019c, maybe
warranting a new release?
Yes, I think it's about time for a release. BTW, that partial
breakage of "fcast" came in with 2019b.
Allin
1894
days inactive
1895
days old
7 comments
4 participants
participants (4)
-
Allin Cottrell -
anzervas＠yahoo.com -
Sven Schreiber -
ΑΝΔΡΕΑΣ ΖΕΡΒΑΣ