8:44 p.m.
Dear Gretl Team,
I would like to share with you some thoughts about the automatic creation
of series in Gretl. The first point I would like to address is apparently
inofensive.
When we perform ADF and KPSS tests we see a different Gretl's behavior. To
ilustrate this, please use the "australia.gdt" data.
# First case: ADF test
adf 4 E --difference --test-down
No variable is created.
# Second case: KPSS test
kpss 4 E --difference
The variable d_E is created.
The second point is a little bit more "dangerous". Yet using the
"australia.gdt" data, please perform this script:
<hansl>
open australia.gdt
"Teste 1" <- ols E const PAU PAU(-1) PAU(-2) E(-1) E(-2)
print PAU* --byobs
series PAU = 0
print PAU* --byobs
</hansl>
One can observe that "PAU(-1)", and "PAU(-2)" are not equal to zero,
what,
in my humble opinion, is not a desirable behavior.
In my humble opinion (again), the variables "PAU(-1)" and "PAU(-2)"
are not
new ones, they are just elements of the vector "PAU". I know this is not a
consensus, but it is not the most important part of my explanation [:-)] So
let's continue...
Now try this new regression:
<hansl>
"Teste 2" <- ols E const PAU PAU(-1) PAU(-2) E(-1) E(-2)
print PAU* --byobs
</hansl>
Now Gretl tells that "PAU", "PAU(-1)", and "PAU(-2)" are all
equal to zero.
I noticed this "problem" when colleagues, working with a Gretl session file
(.gretl), here from my work tried to make all the values of a specific
variable equal zero to check the sensitivity of their model and Gretl did
not behaved well. What they tried was something like that:
# Regress X against its lagged values and Y:
X(t) = b0 + b1*X(t-1) + b2*X(t-2) + b3*Y(t)
# Set all X values to zero:
X(t) = 0
# Regress X again:
X(t) = b0 + b1*X(t-1) + b2*X(t-2) + b3*Y(t)
I believe the expected behavior is to make X(t)=X(t-1)=X(t-2)=0, but Gretl
didn't make that. So they get exactly the same results on the two exercises.
I'm really afraid that my English skills combined with the difficult of
describe this situation (even in Portuguese [:-)]), so I attached a script
file with the complete example.
Best regards,
Henrique Andrade
4:26 p.m.
On Wed, 21 Mar 2012, Henrique Andrade wrote:
I would like to share with you some thoughts about the automatic
creation
of series in Gretl. The first point I would like to address is apparently
inofensive.
When we perform ADF and KPSS tests we see a different Gretl's behavior. To
ilustrate this, please use the "australia.gdt" data.
# First case: ADF test
adf 4 E --difference --test-down
No variable is created.
# Second case: KPSS test
kpss 4 E --difference
The variable d_E is created.
Agreed, that's inconsistent: if we're going to delete any
automatically created lags with the adf command we should do
the same for kpss. That is now fixed in CVS (or will be real
soon).
The second point is a little bit more "dangerous". Yet
using the
"australia.gdt" data, please perform this script:
<hansl>
open australia.gdt
"Teste 1" <- ols E const PAU PAU(-1) PAU(-2) E(-1) E(-2)
print PAU* --byobs
series PAU = 0
print PAU* --byobs
</hansl>
One can observe that "PAU(-1)", and "PAU(-2)" are not equal to zero,
what,
in my humble opinion, is not a desirable behavior.
OK, this is a bit subtle, and should be clearly documented
somewhere, but here goes...
1) When you issue a command such as
ols E const PAU PAU(-1) PAU(-2) E(-1) E(-2)
[note: using the notation "varname(-lag)"] this results in the
automatic creation of series whose names are "PAU_1", "PAU_2"
and so on.
2) These named series are "hard-wired" in the sense that if
you change the definition of PAU, PAU_1 and so on are not
automatically redefined.
3) So if you say
print PAU*
you're going to see your redefined PAU along with the original
values of the series PAU_*.
4) However, if in any subsequent command you again use the
notation "varname(-lag)" (not "varname_lag") you will get lags
updated to match the redefined parent series. For example,
print PAU PAU(-1) PAU(-2)
So: the notation "x(-p)" says to gretl "give me lag p of
series x, as x is currently defined", whereas the notation
"x_p" simply says "give me the series named x_p".
5) When you use the "x(-p)" notation, a side effect is that
the series named x_p is updated (if need be).
Allin Cottrell
5:20 p.m.
Dear Allin,
Ok, we will "kill" the automatic series creation in the KPSS test, but what
about the OLS estimations? In my opinion, Gretl should never create series
automatically :-)
Thanks for your (habitual!) attention.
Best,
Henrique
6:44 p.m.
On Thu, 22 Mar 2012, Henrique Andrade wrote:
Ok, we will "kill" the automatic series creation in the
KPSS test, but what
about the OLS estimations? In my opinion, Gretl should never create series
automatically :-)
Ah, you reckon we should destroy any newly created lags after,
for example "ols y 0 x(-1)"?
What do others think of that?
Allin
7:23 p.m.
Off the top of my head, I think this makes sense. I can't think of a situation where
I'd need to keep x_1 in a data set as a separate series if it can be generated on the
fly.
PS
-----Original Message-----
From: gretl-devel-bounces(a)lists.wfu.edu [mailto:gretl-devel-bounces@lists.wfu.edu] On
Behalf Of Allin Cottrell
Sent: Thursday, March 22, 2012 2:45 PM
To: Gretl development
Subject: Re: [Gretl-devel] Series Creation: An important issue
On Thu, 22 Mar 2012, Henrique Andrade wrote:
Ok, we will "kill" the automatic series creation in the
KPSS test, but
what about the OLS estimations? In my opinion, Gretl should never
create series automatically :-)
Ah, you reckon we should destroy any newly created lags after, for example "ols y 0
x(-1)"?
What do others think of that?
Allin
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
7:32 p.m.
On Thu, 22 Mar 2012, Summers, Peter wrote:
Off the top of my head, I think this makes sense. I can't
think of a situation where I'd need to keep x_1 in a data
set as a separate series if it can be generated on the fly.
It does make good sense in the context of some sort of monte
carlo study or other repetitive exercise. But in that case you
could define the lags ahead of time and refer to them by their
series names rather than using the "x(-p)" notation.
I'm not going to change this right away, but it's worth
considering.
Allin
-----Original Message-----
From: gretl-devel-bounces(a)lists.wfu.edu [mailto:gretl-devel-bounces@lists.wfu.edu] On
Behalf Of Allin Cottrell
Sent: Thursday, March 22, 2012 2:45 PM
To: Gretl development
Subject: Re: [Gretl-devel] Series Creation: An important issue
On Thu, 22 Mar 2012, Henrique Andrade wrote:
> Ok, we will "kill" the automatic series creation in the KPSS test, but
> what about the OLS estimations? In my opinion, Gretl should never
> create series automatically :-)
Ah, you reckon we should destroy any newly created lags after, for example "ols y 0
x(-1)"?
What do others think of that?
Allin
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
--
Allin Cottrell
Department of Economics
Wake Forest University, NC
8:49 a.m.
On 03/22/2012 08:32 PM, Allin Cottrell wrote:
On Thu, 22 Mar 2012, Summers, Peter wrote:
> Off the top of my head, I think this makes sense. I can't
> think of a situation where I'd need to keep x_1 in a data
> set as a separate series if it can be generated on the fly.
I tend to agree with Peter (even though it's backward-incompatible ;-).
I seem to remember a discussion about these automatically generated
series quite some time ago, and the result was (I think) the
introduction of the little arrows/triangles which can be toggled on/off
to show or hide the created lags. I don't remember why it wasn't decided
to discard those on-the-fly series, maybe it could be worth to search
the list archive.
It does make good sense in the context of some sort of monte
carlo study or other repetitive exercise. But in that case you
could define the lags ahead of time and refer to them by their
series names rather than using the "x(-p)" notation.
Absolutely. For the same reason I have always felt a bit uneasy about
the 'lags' command as well (and 'diff' also, I guess). The names of the
created series were a little difficult to predict (e.g. because of name
truncation), so I think it's better to recommend that script authors
explicitly declare the needed series, including their names.
cheers,
sven
3:29 a.m.
Em 22 de março Allin escreveu:
On Thu, 22 Mar 2012, Henrique Andrade wrote:
> Ok, we will "kill" the automatic series creation in the KPSS test, but
what
> about the OLS estimations? In my opinion, Gretl should never create
series
> automatically :-)
Ah, you reckon we should destroy any newly created lags after,
for example "ols y 0 x(-1)"?
What do others think of that?
Allin, looking further at this situation I found additional "misbehaviors"
related to the automatic series creation. Please estimate the following
model via the GUI:
ols PAU const PUS(-1) PAU(-1) --robust
You will notice that no series were created. Now, if you try the same
estimation via the command line (CLI), Gretl creates the new series "PUS_1"
and "PAU_1". Additionally, it is not possible to delete (via the GUI)
"PUS_1" and "PAU_1", even if no models using these variables.
We can see the same behavior in the ARIMA modelling:
arima 1 0 0 ; E PAU(-1) PUS(-1) 0
When we perform VAR and VEC estimations (using GUI or CLI), no variables
are generated:
var 4 E PAU PUS --robust
vecm 4 1 E PAU PUS
In conclusion, this plurality of behavior is the main reason I believe the
automatic creation of series should be avoided.
Um abraço,
Henrique
2:20 p.m.
On Fri, 23 Mar 2012, Henrique Andrade wrote:
Allin, looking further at this situation I found additional
"misbehaviors"
related to the automatic series creation. Please estimate the following
model via the GUI:
ols PAU const PUS(-1) PAU(-1) --robust
You will notice that no series were created.
The two lag series are created. The issue here is just that
the main window is not being automatically updated to show the
lags. The workaround is to press "r" (refresh) in the main
window. Then you see them (under the "twisty") and can delete
them if you wish.
Allin
12:51 p.m.
Em 23 de março de 2012 Allin Cottrell escreveu:
On Fri, 23 Mar 2012, Henrique Andrade wrote:
> Allin, looking further at this situation I found additional
"misbehaviors"
> related to the automatic series creation. Please estimate the following
> model via the GUI:
>
> ols PAU const PUS(-1) PAU(-1) --robust
>
> You will notice that no series were created.
The two lag series are created. The issue here is just that
the main window is not being automatically updated to show the
lags. The workaround is to press "r" (refresh) in the main
window. Then you see them (under the "twisty") and can delete
them if you wish.
Ok, I tested the new snapshot and now we don`t need to press "r" (refresh)
anymore. But the others inconsistences/problems I pointed are still there:
(1) Even if you don't save any model you are not able to delete the new
variables ("PUS_1" and "PAU_1").
(2) If you estimate OLS, Arima or Garch models Gretl creates "new"
variables, but if you try VAR and VEC estimations, Gretl does not create
anything.
These are the commands I used to teste:
open australia.gdt
ols E const E(-1) PUS(-1) PAU(-1)
open australia.gdt
arima 1 0 0 ; E PAU(-1) PUS(-1)
open australia.gdt
garch 1 1 ; E PAU(-1) PUS(-1)
open australia.gdt
var 1 E PUS PAU
open australia.gdt
vecm 2 1 E PUS PAU
Best,
Henrique Andrade
11:50 p.m.
On Sun, 25 Mar 2012, Henrique Andrade wrote:
Em 23 de março de 2012 Allin Cottrell escreveu:
On Fri, 23 Mar 2012, Henrique Andrade wrote:
>
>> Allin, looking further at this situation I found additional
> "misbehaviors"
>> related to the automatic series creation. Please estimate the following
>> model via the GUI:
>>
>> ols PAU const PUS(-1) PAU(-1) --robust
>>
>> You will notice that no series were created.
>
> The two lag series are created. The issue here is just that
> the main window is not being automatically updated to show the
> lags. The workaround is to press "r" (refresh) in the main
> window. Then you see them (under the "twisty") and can delete
> them if you wish.
>
Ok, I tested the new snapshot and now we don`t need to press "r" (refresh)
anymore. But the others inconsistences/problems I pointed are still there:
Let me explain how gretl works at present. It may not be totally
intuitive but it is consistent.
1) If, in a context where a gretl list is wanted (for example, in
specifying the regressors of an OLS model) you use the syntax
"x(-p)" (for "x" a series in the dataset and "p" a valid lag
order)
this is equivalent to saying to gretl: please add a series holding
lag p of x to the dataset; call it "x_p" and put it into the list at
this point.
2) This explains why, if you do (for example)
ols E const E(-1) PUS(-1) PAU(-1)
you then find series named "E_1", "PUS_1" and "PAU_1" in the
dataset, so long as the command is successfully completed.
3) You do not find such series added to the dataset if you do (say)
var 1 E PUS PAU
Why not? Because you did not use the "x(-p)" syntax; rather you used
a command which knows how to handle lags automatically (the "var"
command with lag order 1).
4) After estimating the ad hoc model specified under point 2), the
command "delete E_1" fails, with the message: "No variables were
deleted (some data were in use)" -- why is that?
It's because you are trying to delete a variable that occupies the
position of regressor in the "last model estimated" (even if this
model was not saved by name). The "last model" is the default target
for many commands (diagnostics, "add", "omit", forecasting).
What do you suppose would happen if you (a) estimate a model; (b)
delete from the dataset one or more of the regressors used in this
model; then (c) issue the "add" command, asking gretl to re-estimate
the model with added regressors? It's not going to be pretty!
Bottom line: (a) lagged series should be added to the dataset
(without an explicit command on the user's part) only if you "ask
for them" by using the syntax "x(-p)" in a gretl list context; and
(b) gretl should be quite willing to delete such series unless they
are "in use", either by a model that was explicitly saved or by the
"last model".
Allin Cottrell
4414
days inactive
4418
days old
10 comments
4 participants
participants (4)
-
Allin Cottrell -
Henrique Andrade -
Summers, Peter -
Sven Schreiber