6:23 a.m.
Hi,Below is an example for one-step ahead forecasting that I use the first 5 obs. for
initial sample and then forecast one-step-ahead by rolling ahead.My question is that how
should I change my code to get 5-step-ahead rolling forecast.Thanks,
open denmark.gdtset verbose offseries frcst = NAloop i=1..20 -q #out-of-sample (20
observations)# smpl 1+i 5+i #initial sample (5 observations)# ols LRM const
LRY fcast 6+i 6+i #one-step-ahead-rolling-windows-forecasting# frcst[6 + i]
= $fcastendloop
7:30 a.m.
Am 07.01.2021 um 20:23 schrieb Burak Korkusuz:
Hi,
Below is an example for one-step ahead forecasting that I use the first
5 obs. for initial sample and then forecast one-step-ahead by rolling ahead.
My question is that how should I change my code to get 5-step-ahead
rolling forecast.
Thanks,
open denmark.gdt
set verbose off
series frcst = NA
loop i=1..20 -q #out-of-sample (20 observations)#
smpl 1+i 5+i #initial sample (5 observations)#
ols LRM const LRY
fcast 6+i 6+i #one-step-ahead-rolling-windows-forecasting#
frcst[6 + i] = $fcast
endloop
Hi, two questions first: Do you really want to have a rolling sample
with a constant sample length of 5 obs? Or rather move only the end obs?
Secondly, do you really want to have no lagged variables in your
equation? While technically feasible, a "forecast" from an equation
where all regressors except the constant are unknown for the next period
doesn't give anything really meaningful.
cheers
sven
7:50 a.m.
I have a different but similar problem where 2 series appear to be
recursively linked Yt=f(Xt ...) ; Xt=g(Yt-1....) . How to generate
dynamic roll-forward forecasts from the two separate equations when both Tt
and Xt are of interest.
On Thu, 7 Jan 2021 at 20:30, Sven Schreiber <svetosch(a)gmx.net> wrote:
Am 07.01.2021 um 20:23 schrieb Burak Korkusuz:
> Hi,
> Below is an example for one-step ahead forecasting that I use the first
> 5 obs. for initial sample and then forecast one-step-ahead by rolling
ahead.
> My question is that how should I change my code to get 5-step-ahead
> rolling forecast.
> Thanks,
>
> open denmark.gdt
> set verbose off
> series frcst = NA
> loop i=1..20 -q #out-of-sample (20 observations)#
> smpl 1+i 5+i #initial sample (5 observations)#
> ols LRM const LRY
> fcast 6+i 6+i #one-step-ahead-rolling-windows-forecasting#
> frcst[6 + i] = $fcast
> endloop
>
Hi, two questions first: Do you really want to have a rolling sample
with a constant sample length of 5 obs? Or rather move only the end obs?
Secondly, do you really want to have no lagged variables in your
equation? While technically feasible, a "forecast" from an equation
where all regressors except the constant are unknown for the next period
doesn't give anything really meaningful.
cheers
sven
_______________________________________________
Gretl-users mailing list -- gretl-users(a)gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave(a)gretlml.univpm.it
Website:
https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
8:35 a.m.
Am 07.01.2021 um 21:50 schrieb Brian Revell:
I have a different but similar problem where 2 series appear to be
recursively linked Yt=f(Xt ...) ; Xt=g(Yt-1....) . How to generate
dynamic roll-forward forecasts from the two separate equations when both
Tt and Xt are of interest.
Need to specify a (bivariate) system.
-s
8:16 a.m.
I just randomly wrote my code as an example for that dataset to be clear. Yes, I want to
have a rolling initial sample with a constant. Actually in my real dataset I use the
lagged variables of my dependent variable (it is a time series regression). I can consider
only the first lag of dependent variable and constant in the equation.
On Thursday, January 7, 2021, 08:30:36 PM GMT, Sven Schreiber <svetosch(a)gmx.net>
wrote:
Am 07.01.2021 um 20:23 schrieb Burak Korkusuz:
Hi,
Below is an example for one-step ahead forecasting that I use the first
5 obs. for initial sample and then forecast one-step-ahead by rolling ahead.
My question is that how should I change my code to get 5-step-ahead
rolling forecast.
Thanks,
open denmark.gdt
set verbose off
series frcst = NA
loop i=1..20 -q #out-of-sample (20 observations)#
smpl 1+i 5+i #initial sample (5 observations)#
ols LRM const LRY
fcast 6+i 6+i #one-step-ahead-rolling-windows-forecasting#
frcst[6 + i] = $fcast
endloop
Hi, two questions first: Do you really want to have a rolling sample
with a constant sample length of 5 obs? Or rather move only the end obs?
Secondly, do you really want to have no lagged variables in your
equation? While technically feasible, a "forecast" from an equation
where all regressors except the constant are unknown for the next period
doesn't give anything really meaningful.
cheers
sven
_______________________________________________
Gretl-users mailing list -- gretl-users(a)gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave(a)gretlml.univpm.it
Website: https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
8:35 a.m.
Am 07.01.2021 um 22:16 schrieb Burak Korkusuz:
I just randomly wrote my code as an example for that dataset to be
clear. Yes, I want to have a rolling initial sample with a constant.
Hm, still don't really understand what a rolling initial sample is; I'm
asking because gretl's builtin --recursive option might be relevant. But
anyway, just define your horizon, e.g.:
horizon = 5
and replace your fcast lines with the adapted ones:
fcast 5+i+horizon 5+i+horizon
frcst[5+i+horizon] = $fcast
Notice, however, that in a model of your type (ols LRM const LRY), the
RHS variable LRY is treated as exogenous. So your "forecast" in this
case does nothing else than take the observed value of LRY five steps
ahead and plugs it into the estimated equation. But that comes from your
model specification, not from the fcast command.
hth,
sven
9:07 a.m.
Thanks for your reply,maybe, that one is clearer. This is for one-step ahead forecasting.
I am trying to generate 5 (and 22) -step-ahead out-of-sample forecasts using the equation
(ols djclose const djclose(-1)) starting of the sample period 1+i – 100+i. However, the
equation is estimated recursively every time and forecasted every time using the new
equation fitted values.
open djclose.gdtset verbose offseries frcst = NAloop i=1..200 -q #out-of-sample
(10 observations)# smpl 1+i 100+i #initial sample (5 observations)# ols
djclose const djclose(-1) fcast 101+i 101+i
#one-step-ahead-rolling-windows-forecasting# frcst[101 + i] = $fcastendloop
On Thursday, January 7, 2021, 09:35:14 PM GMT, Sven Schreiber <svetosch(a)gmx.net>
wrote:
Am 07.01.2021 um 22:16 schrieb Burak Korkusuz:
I just randomly wrote my code as an example for that dataset to be
clear. Yes, I want to have a rolling initial sample with a constant.
Hm, still don't really understand what a rolling initial sample is; I'm
asking because gretl's builtin --recursive option might be relevant. But
anyway, just define your horizon, e.g.:
horizon = 5
and replace your fcast lines with the adapted ones:
fcast 5+i+horizon 5+i+horizon
frcst[5+i+horizon] = $fcast
Notice, however, that in a model of your type (ols LRM const LRY), the
RHS variable LRY is treated as exogenous. So your "forecast" in this
case does nothing else than take the observed value of LRY five steps
ahead and plugs it into the estimated equation. But that comes from your
model specification, not from the fcast command.
hth,
sven
_______________________________________________
Gretl-users mailing list -- gretl-users(a)gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave(a)gretlml.univpm.it
Website: https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
9:19 a.m.
Am 07.01.2021 um 23:07 schrieb Burak Korkusuz:
Thanks for your reply,
maybe, that one is clearer. This is for one-step ahead forecasting. I am
trying to generate 5 (and 22) -step-ahead out-of-sample forecasts using
the equation (ols djclose const djclose(-1)) starting of the sample
period 1+i – 100+i. However, the equation is estimated recursively
every time and forecasted every time using the new equation fitted values.
open djclose.gdt
set verbose off
series frcst = NA
loop i=1..200 -q #out-of-sample (10 observations)#
smpl 1+i 100+i #initial sample (5 observations)#
ols djclose const djclose(-1)
fcast 101+i 101+i #one-step-ahead-rolling-windows-forecasting#
frcst[101 + i] = $fcast
endloop
Well inserting the FC horizon specification should still work here.
Since you're forecasting out-of-sample in every step the forecast should
be dynamic by default (check the fcast reference), but if you want to be
absolutely sure you may append the --dynamic option to the fcast line.
cheers
sven
9:32 a.m.
Yes, now it is ok, thank you very much
it gives error without appending --dynamic option
Cheers,
On Thursday, January 7, 2021, 10:19:48 PM GMT, Sven Schreiber <svetosch(a)gmx.net>
wrote:
Am 07.01.2021 um 23:07 schrieb Burak Korkusuz:
Thanks for your reply,
maybe, that one is clearer. This is for one-step ahead forecasting. I am
trying to generate 5 (and 22) -step-ahead out-of-sample forecasts using
the equation (ols djclose const djclose(-1)) starting of the sample
period 1+i – 100+i. However, the equation is estimated recursively
every time and forecasted every time using the new equation fitted values.
open djclose.gdt
set verbose off
series frcst = NA
loop i=1..200 -q #out-of-sample (10 observations)#
smpl 1+i 100+i #initial sample (5 observations)#
ols djclose const djclose(-1)
fcast 101+i 101+i #one-step-ahead-rolling-windows-forecasting#
frcst[101 + i] = $fcast
endloop
Well inserting the FC horizon specification should still work here.
Since you're forecasting out-of-sample in every step the forecast should
be dynamic by default (check the fcast reference), but if you want to be
absolutely sure you may append the --dynamic option to the fcast line.
cheers
sven
_______________________________________________
Gretl-users mailing list -- gretl-users(a)gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave(a)gretlml.univpm.it
Website: https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
9:37 a.m.
Am 07.01.2021 um 23:32 schrieb Burak Korkusuz:
Yes, now it is ok, thank you very much
it gives error without appending --dynamic option
Hm, I see what you mean. The way I read the fcast documentation, this
option should be automatically implied in this context.
@Allin, what are we missing?
thanks
sven
3:54 p.m.
On Thu, 7 Jan 2021, Sven Schreiber wrote:
Am 07.01.2021 um 23:32 schrieb Burak Korkusuz:
> Yes, now it is ok, thank you very much
>
> it gives error without appending --dynamic option
Hm, I see what you mean. The way I read the fcast documentation, this
option should be automatically implied in this context.
@Allin, what are we missing?
The problem here is that "fcast t1 t2" doesn't start the forecast
machinery until t1. So when attempting a dynamic forecast according
to
fcast 100+i+h 100+i+h # with h > 1
gretl first looks for a prior forecast of y at obs 100+i+h-1, but
doesn't find one. Here's a version of the script which works
properly:
<hansl>
open djclose.gdt
set verbose off
series frcst = NA
h = 5
loop i=1..50 -q
smpl 1+i 100+i
ols djclose const djclose(-1)
fcast 100+i+1 100+i+h
frcst[100+i+h] = $fcast[h]
endloop
</hansl>
Now you might say that gretl should be clever enough to realize that
it needs to forecast the intermediate values (at least if the
--dynamic option is given), even if they will be discarded in the
output, but right now that doesn't happen.
Why did the original version "work" with the --dynamic option and
only complain about missing values if that option was not given? A
bug: with --dynamic, gretl fell back on using actual y in place of
the missing forecast lag (so in fact giving a static forecast) but
in the "auto" case that didn't happen -- which is probably the
reverse of what should occur.
Stuff needs fixing here, but note that the problem is specific to
the case where there's a temporal gap between the estimation sample
and the start of the dynamic "fcast" range.
Allin
10:58 p.m.
Am 08.01.2021 um 05:54 schrieb Allin Cottrell:
Why did the original version "work" with the --dynamic
option and only
complain about missing values if that option was not given? A bug: with
--dynamic, gretl fell back on using actual y in place of the missing
forecast lag (so in fact giving a static forecast) but in the "auto"
case that didn't happen -- which is probably the reverse of what should
occur.
Aha, good to have discovered that.
Stuff needs fixing here, but note that the problem is specific to
the
case where there's a temporal gap between the estimation sample and the
start of the dynamic "fcast" range.
Understood - so the purpose of giving the starting period of the
forecast is to make it possible to start before the end of the
estimation sample, if I understand correctly?
So Burak, please take note that your seemingly working latest script
version needs an update.
cheers
sven
11:49 p.m.
oh yes, my version is a little old, i need to update. Thanks,
On Friday, January 8, 2021, 11:58:32 AM GMT, Sven Schreiber <svetosch(a)gmx.net>
wrote:
Am 08.01.2021 um 05:54 schrieb Allin Cottrell:
Why did the original version "work" with the --dynamic
option and only
complain about missing values if that option was not given? A bug: with
--dynamic, gretl fell back on using actual y in place of the missing
forecast lag (so in fact giving a static forecast) but in the "auto"
case that didn't happen -- which is probably the reverse of what should
occur.
Aha, good to have discovered that.
Stuff needs fixing here, but note that the problem is specific to
the
case where there's a temporal gap between the estimation sample and the
start of the dynamic "fcast" range.
Understood - so the purpose of giving the starting period of the
forecast is to make it possible to start before the end of the
estimation sample, if I understand correctly?
So Burak, please take note that your seemingly working latest script
version needs an update.
cheers
sven
_______________________________________________
Gretl-users mailing list -- gretl-users(a)gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave(a)gretlml.univpm.it
Website: https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
11:50 p.m.
Am 08.01.2021 um 13:49 schrieb Burak Korkusuz:
oh yes, my version is a little old, i need to update.
I don't mean to update the gretl version (yet), but your script -- from
what I understand it will give wrong results (due to the gretl bug that
was discovered in this context).
-s
12:59 a.m.
Is there a solution for that?
Sent from Yahoo Mail on Android
On Fri, 8 Jan 2021 at 12:51, Sven Schreiber<svetosch(a)gmx.net> wrote: Am
08.01.2021 um 13:49 schrieb Burak Korkusuz:
oh yes, my version is a little old, i need to update.
I don't mean to update the gretl version (yet), but your script -- from
what I understand it will give wrong results (due to the gretl bug that
was discovered in this context).
-s
_______________________________________________
Gretl-users mailing list -- gretl-users(a)gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave(a)gretlml.univpm.it
Website: https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
3:19 a.m.
On Fri, 8 Jan 2021, Sven Schreiber wrote:
Am 08.01.2021 um 14:59 schrieb Burak Korkusuz:
> Is there a solution for that?
>
Well, see Allin's script a few postings back. Essentially using +1 as
the start and not +horizon.
In git, using +h (h > 1) for the start of a dynamic forecast will
now work. (We generate but do not report the required intervening
forecast values.) That's not yet in snapshots but will be soon.
Allin
12:06 a.m.
Will the revised gretl version due to be released now include incorporation
of exogenous variables in the Structural TS forecast option?
On Thu, 7 Jan 2021 at 20:30, Sven Schreiber <svetosch(a)gmx.net> wrote:
Am 07.01.2021 um 20:23 schrieb Burak Korkusuz:
> Hi,
> Below is an example for one-step ahead forecasting that I use the first
> 5 obs. for initial sample and then forecast one-step-ahead by rolling
ahead.
> My question is that how should I change my code to get 5-step-ahead
> rolling forecast.
> Thanks,
>
> open denmark.gdt
> set verbose off
> series frcst = NA
> loop i=1..20 -q #out-of-sample (20 observations)#
> smpl 1+i 5+i #initial sample (5 observations)#
> ols LRM const LRY
> fcast 6+i 6+i #one-step-ahead-rolling-windows-forecasting#
> frcst[6 + i] = $fcast
> endloop
>
Hi, two questions first: Do you really want to have a rolling sample
with a constant sample length of 5 obs? Or rather move only the end obs?
Secondly, do you really want to have no lagged variables in your
equation? While technically feasible, a "forecast" from an equation
where all regressors except the constant are unknown for the next period
doesn't give anything really meaningful.
cheers
sven
_______________________________________________
Gretl-users mailing list -- gretl-users(a)gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave(a)gretlml.univpm.it
Website:
https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/
3:33 a.m.
Am 16.01.2021 um 14:06 schrieb Brian Revell:
Will the revised gretl version due to be released now include
incorporation of exogenous variables in the Structural TS forecast option?
Brian, pleeease start new threads for your new questions, don't just
reply to different postings!
The answer is no because the StrucTiSM package is a contributed package
and thus separate from the main gretl codebase. An update of the
StrucTiSM package can (and will) happen independently of the gretl releases.
cheers
sven
2006
days inactive
2015
days old
18 comments
4 participants
participants (4)
-
Allin Cottrell -
Brian Revell -
Burak Korkusuz -
Sven Schreiber