11:53 a.m.
Hi,
I've just revisited the available tools for diagnostic checking of VARs
(as you may have guessed from my earlier email this morning). To be
honest, I've found them a little lacking.
Attached is an implementation of the Rao-style F-statistic for VAR
residual autocorrelation that is described in Helmut Luetkepohl's "New"
book (also more than 10 years old already), section 4.4.4 (Lagrange
Multiplier Tests for Checking the Whiteness of the Residuals).
A simple test case is also included (at the bottom of the script). Note
the different p-values for order 4 compared to the built-in tests.
Comments and corrections welcome. Eventually I would propose to adapt
this as a built-in "modtest" variant (if no problems arise with this).
thanks,
sven
6:43 p.m.
On Tue, 14 Mar 2017, Sven Schreiber wrote:
I've just revisited the available tools for diagnostic checking
of VARs (as
you may have guessed from my earlier email this morning). To be honest, I've
found them a little lacking.
Attached is an implementation of the Rao-style F-statistic for VAR residual
autocorrelation that is described in Helmut Luetkepohl's "New" book (also
more than 10 years old already), section 4.4.4 (Lagrange Multiplier Tests for
Checking the Whiteness of the Residuals).
A simple test case is also included (at the bottom of the script). Note the
different p-values for order 4 compared to the built-in tests.
Comments and corrections welcome. Eventually I would propose to adapt this as
a built-in "modtest" variant (if no problems arise with this).
Thanks for the nice hansl prototype. I've now replaced the built-in
"modtest --autocorr" for VARs with a C implementation of this system
test. If anyone's in a position to check against known-good results
that would be helpful.
Our ARCH test for VARs is still just per-equation. Maybe at some point
we can upgrade that too.
Allin
8:34 p.m.
Am 26.03.2017 um 20:43 schrieb Allin Cottrell:
Thanks for the nice hansl prototype. I've now replaced the
built-in
"modtest --autocorr" for VARs with a C implementation of this system
test. If anyone's in a position to check against known-good results that
would be helpful.
I haven't checked out Eviews's tests there in ages (and it's not clear
at all they would be "known good") but maybe I could do that.
Our ARCH test for VARs is still just per-equation. Maybe at some
point
we can upgrade that too.
Yes, that would be good, too. Any particular test incarnation you have
in mind (haven't looked)?
Thanks for the translation hansl-to-gretl-C!
cheers,
sven
8:45 p.m.
On Sun, 26 Mar 2017, Sven Schreiber wrote:
Am 26.03.2017 um 20:43 schrieb Allin Cottrell:
> Thanks for the nice hansl prototype. I've now replaced the built-in
> "modtest --autocorr" for VARs with a C implementation of this system
> test. If anyone's in a position to check against known-good results that
> would be helpful.
I haven't checked out Eviews's tests there in ages (and it's not clear at all
they would be "known good") but maybe I could do that.
> Our ARCH test for VARs is still just per-equation. Maybe at some point
> we can upgrade that too.
Yes, that would be good, too. Any particular test incarnation you have in
mind (haven't looked)?
No, I don't have anything specific in mind. Does Lutkepohl discuss a
multivariate ARCH test? (Sorry, I don't have his book.)
Thanks for the translation hansl-to-gretl-C!
No problem. This is how Jack and I often work: if a prototype
function can be written in hansl, it generally won't take too long
to translate to gretl-C.
Allin
8:57 p.m.
Am 26.03.2017 um 22:45 schrieb Allin Cottrell:
>> Our ARCH test for VARs is still just per-equation. Maybe at
some point
>> we can upgrade that too.
>
> Yes, that would be good, too. Any particular test incarnation you have
> in mind (haven't looked)?
No, I don't have anything specific in mind. Does Lutkepohl discuss a
multivariate ARCH test? (Sorry, I don't have his book.)
Yeah, he is basically saying he copied that from the PcFiml/PcGive
manual by Doornik&Hendry. (multivariate ARCH-LM type test, I mean)
Should also be fairly straightforward, but not right now...
thanks,
sven
7:12 a.m.
Am 26.03.2017 um 22:57 schrieb Sven Schreiber:
Am 26.03.2017 um 22:45 schrieb Allin Cottrell:
> No, I don't have anything specific in mind. Does Lutkepohl
discuss a
> multivariate ARCH test? (Sorry, I don't have his book.)
Yeah, he is basically saying he copied that from the PcFiml/PcGive
manual by Doornik&Hendry. (multivariate ARCH-LM type test, I mean)
Should also be fairly straightforward, but not right now...
Attached is a prototype. Now this produces some results in the test
case, but it is really untested, please don't make this the standard
VAR-ARCH test output just yet.
Based on section 16.5.1 (essentially equation 16.5.1) from Helmut
Lütkepohl.
cheers,
sven
1:59 p.m.
On Mon, 27 Mar 2017, Sven Schreiber wrote:
Am 26.03.2017 um 22:57 schrieb Sven Schreiber:
> Am 26.03.2017 um 22:45 schrieb Allin Cottrell:
>> No, I don't have anything specific in mind. Does Lutkepohl discuss a
>> multivariate ARCH test? (Sorry, I don't have his book.)
>
> Yeah, he is basically saying he copied that from the PcFiml/PcGive
> manual by Doornik&Hendry. (multivariate ARCH-LM type test, I mean)
>
> Should also be fairly straightforward, but not right now...
Attached is a prototype. Now this produces some results in the test case, but
it is really untested, please don't make this the standard VAR-ARCH test
output just yet.
Based on section 16.5.1 (essentially equation 16.5.1) from Helmut Lütkepohl.
Very nice script, thanks.
Now, this is slightly OT, because I realise Sven's script is just meant as
a prototype and clarity is more important than efficiency, but just for
future reference to script writers who may be reading this thread: we have
a very nice (and little known) function called "hdprod" that, in cases
like this, makes the loop contruct unnecessary. The lines
<hansl>
matrix vechU = zeros(0, K * (K+1) / 2)
loop tau = 1..T -q
vechU |= transp( vech(mU[tau, ]'mU[tau, ]) )
endloop
</hansl>
could be replaced by
<hansl>
idx = vec(lower(ones(K,K)))
vU = selifc(hdprod(mU, mU), idx')
</hansl>
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
10:42 p.m.
On Mon, 27 Mar 2017, Sven Schreiber wrote:
Am 26.03.2017 um 22:57 schrieb Sven Schreiber:
> Am 26.03.2017 um 22:45 schrieb Allin Cottrell:
>> No, I don't have anything specific in mind. Does Lutkepohl discuss a
>> multivariate ARCH test? (Sorry, I don't have his book.)
>
> Yeah, he is basically saying he copied that from the PcFiml/PcGive
> manual by Doornik&Hendry. (multivariate ARCH-LM type test, I mean)
>
> Should also be fairly straightforward, but not right now...
Attached is a prototype. Now this produces some results in the test case, but
it is really untested, please don't make this the standard VAR-ARCH test
output just yet.
Based on section 16.5.1 (essentially equation 16.5.1) from Helmut Lütkepohl.
Thanks, Sven. This looks less daunting that one might have feared!
I'll have a go at turning it into gretl-C but I won't replace the
current per-equation ARCH test for VARs until we're satisfied that
the multivariate version is right.
Allin
10:28 p.m.
On Sun, 26 Mar 2017, Sven Schreiber wrote:
Am 26.03.2017 um 20:43 schrieb Allin Cottrell:
> Thanks for the nice hansl prototype. I've now replaced the built-in
> "modtest --autocorr" for VARs with a C implementation of this system
> test. If anyone's in a position to check against known-good results that
> would be helpful.
I haven't checked out Eviews's tests there in ages (and it's not
clear at all they would be "known good") but maybe I could do
that.
How about Helmut's example on page 174 of his "New introduction"? (I
still don't have the book itself, but google books gives me quite a
lot of it.) We can now replicate this -- and just as well that I
tried because the trial exposed a nasty bug (my bad), now fixed in
git.
It would be quite nice to include replication of this example in the
VAR chapter of our User's Guide. Do you think we're OK just grabbing
the dataset (oldish West German data in wgmacro.dta, at
https://ideas.repec.org/p/boc/bocins/lutkepohl.html) and making it
available for gretl? Or if you think that's dodgy, would you mind
asking Helmut for permission?
Thanks,
Allin
5:43 a.m.
Am 28.03.2017 um 00:28 schrieb Allin Cottrell:
How about Helmut's example on page 174 of his "New
introduction"? (I
still don't have the book itself, but google books gives me quite a lot
of it.) We can now replicate this -- and just as well that I tried
because the trial exposed a nasty bug (my bad), now fixed in git.
Well if you already replicated this, that's great! (I'm talking about
table 4.8 on p.174)
It would be quite nice to include replication of this example in the
VAR
chapter of our User's Guide. Do you think we're OK just grabbing the
dataset (oldish West German data in wgmacro.dta, at
https://ideas.repec.org/p/boc/bocins/lutkepohl.html) and making it
available for gretl? Or if you think that's dodgy, would you mind asking
Helmut for permission?
It very much looks like it's there to be shared. I mean under Biblio
Info it says "paper provided by Boston College Department of Economics
in its series Instructional Stata datasets for econometrics" and it has
an open link. I would just use it.
cheers,
sven
5:59 a.m.
Am 28.03.2017 um 07:43 schrieb Sven Schreiber:
Am 28.03.2017 um 00:28 schrieb Allin Cottrell:
> It would be quite nice to include replication of this example in
the VAR
> chapter of our User's Guide. Do you think we're OK just grabbing the
> dataset (oldish West German data in wgmacro.dta, at
> https://ideas.repec.org/p/boc/bocins/lutkepohl.html) and making it
> available for gretl? Or if you think that's dodgy, would you mind asking
> Helmut for permission?
It very much looks like it's there to be shared. I mean under Biblio
Info it says "paper provided by Boston College Department of Economics
in its series Instructional Stata datasets for econometrics" and it has
an open link. I would just use it.
BTW, there seems to be some overlap with the datasets on:
http://www.jmulti.de/data_atse.html
Maybe those are useful for other examples in the book.
-s
6:25 a.m.
On Tue, 28 Mar 2017, Sven Schreiber wrote:
Am 28.03.2017 um 07:43 schrieb Sven Schreiber:
> Am 28.03.2017 um 00:28 schrieb Allin Cottrell:
>> It would be quite nice to include replication of this example in the VAR
>> chapter of our User's Guide. Do you think we're OK just grabbing the
>> dataset (oldish West German data in wgmacro.dta, at
>> https://ideas.repec.org/p/boc/bocins/lutkepohl.html) and making it
>> available for gretl? Or if you think that's dodgy, would you mind asking
>> Helmut for permission?
>
> It very much looks like it's there to be shared. I mean under Biblio
> Info it says "paper provided by Boston College Department of Economics
> in its series Instructional Stata datasets for econometrics" and it has
> an open link. I would just use it.
BTW, there seems to be some overlap with the datasets on:
http://www.jmulti.de/data_atse.html
Maybe those are useful for other examples in the book.
Helmut has been very sympathetic to our project for quite a long time now.
I suppose we could just put together a collection of gdt files like we
have for other textbooks, that cover both his books. Then, we ask him for
permission to make them available. My subjective probability evaluation of
him giving us the green light is 99.99%.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
6:15 a.m.
Am 26.03.2017 um 20:43 schrieb Allin Cottrell:
Thanks for the nice hansl prototype. I've now replaced the
built-in
"modtest --autocorr" for VARs with a C implementation of this system
test. If anyone's in a position to check against known-good results that
would be helpful.
One more remark: This test can fail in small samples if the order up to
which AC is tested becomes too large, simply because of missing
degrees-of-freedom. So perhaps some fallback would also be good, maybe
just the portmanteau statistic.
cheers,
sven
6:24 a.m.
On Mon, 27 Mar 2017, Sven Schreiber wrote:
Am 26.03.2017 um 20:43 schrieb Allin Cottrell:
> Thanks for the nice hansl prototype. I've now replaced the built-in
> "modtest --autocorr" for VARs with a C implementation of this system
> test. If anyone's in a position to check against known-good results that
> would be helpful.
One more remark: This test can fail in small samples if the order up to which
AC is tested becomes too large, simply because of missing degrees-of-freedom.
So perhaps some fallback would also be good, maybe just the portmanteau
statistic.
True. Although one may question the whole point of using inferential
procedures at all, when degrees of freedom get exhausted. In those cases,
I would find it hard, statistically speaking, to consider those "tests"
more than descriptive statistics. Perhaps in those cases we may just spew
out an error message saying that the statistic is not available because
the sample size is too small.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
3:57 p.m.
On Mon, 27 Mar 2017, Sven Schreiber wrote:
Am 26.03.2017 um 20:43 schrieb Allin Cottrell:
> Thanks for the nice hansl prototype. I've now replaced the built-in
> "modtest --autocorr" for VARs with a C implementation of this system
> test. If anyone's in a position to check against known-good results that
> would be helpful.
One more remark: This test can fail in small samples if the order up
to which AC is tested becomes too large, simply because of missing
degrees-of-freedom.
OK, we'll now (in git) catch that possibility in advance and flag an
error.
Allin
2826
days inactive
2840
days old
14 comments
3 participants
participants (3)
-
Allin Cottrell -
Riccardo (Jack) Lucchetti -
Sven Schreiber