10:16 p.m.
Thank you very much!
1. Adding 0 worked. I got the bootstrap output.
2. There is a red flag for the short and long run restrictions, though. I put the two
restrictions simultaneously but the results are widely different from what my coauthor
obtained from RATS and GAUSS.
The result from gretl shows a C matrix that looks like one that just superimposed short
and long run matrices in a way that any 0 values in either of the two matrices replaced
the relevant element of the C.
In addition, the bootstrap did not run adequately. Any advice would be greatly
appreciated! Jay
Sent via the Samsung GALAXY S® 5, an AT&T 4G LTE smartphone
-------- Original message --------
From: "Riccardo (Jack) Lucchetti" <r.lucchetti(a)univpm.it>
Date: 7/30/2016 5:17 AM (GMT-05:00)
To: Jay Ryu <jayeryu(a)yahoo.com>, Gretl list <gretl-users(a)lists.wfu.edu>
Subject: Re: [Gretl-users] SVAR bootstrap output table etc.
On Fri, 29 Jul 2016, Jay Ryu wrote:
Dear gretl friends:
I have been running some SVAR models based on:
Jack Lucchetti's "svar_gretl.pdf" as of February 2015
I ran the following bootstrap from the manual below:
[...]
I found the IRF outputs but could not get the output table (like page
10 in Lucchetti).
Question 1: How can I get the bootstrap output table in page 10?
Uh, you're right. There's an undocumented fourth parameter to SVAR_boot:
it's an optional boolean "quiet" flag, which defaults to 1.
Supplying a 0 as fourth parameter to SVAR_boot, as in
<hansl>
SVAR_boot(&Model, 1000, 0.95, 0)
</hansl>
should do the trick.
I'll update the documentation asap.
Question 2: Lucchetti's manual shows how to impose
"both" short- and
long-run restrictions simultaneously. Does this simultaneous restriction
function work?
AFAIK, it should. Does the example script "BlQuah_w_sr.inp" (available in
the "Resources" directory) work for you?
-------------------------------------------------------
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
-------------------------------------------------------
11 p.m.
New subject: SVAR bootstrap output table etc.
Am 03.08.2016 um 00:16 schrieb jayeryu:
2. There is a red flag for the short and long run restrictions,
though.
I put the two restrictions simultaneously but the results are widely
different from what my coauthor obtained from RATS and GAUSS.
Thanks for reporting your case.
The result from gretl shows a C matrix that looks like one that just
superimposed short and long run matrices in a way that any 0 values in
either of the two matrices replaced the relevant element of the C.
It would be useful if you could copy&paste your restrictions and the C
matrix that gretl shows.
In case you're not using the latest gretl version please also add that
information.
Thanks,
sven
1:56 a.m.
New subject: SVAR bootstrap output table etc.
Thank you for your prompt response!Below, I copied the gretl syntax (restrictions and
bootstrap). I also copied the gretl output. And yes, I installed the latest version from
the gretl site.Sincerely,Jay
BQModel = SVAR_setup("C", X, exog, maxlag)BQModel["horizon"]=20
SVAR_restrict(&BQModel, "lrC", 1,2,0)SVAR_restrict(&BQModel,
"lrC", 1,3,0)SVAR_restrict(&BQModel, "lrC",
1,4,0)SVAR_restrict(&BQModel, "lrC", 1,5,0)
SVAR_restrict(&BQModel, "C", 1,3,0)SVAR_restrict(&BQModel,
"C", 1,4,0)SVAR_restrict(&BQModel, "C",
1,5,0)SVAR_restrict(&BQModel, "C", 3,5,0)SVAR_restrict(&BQModel,
"C", 4,5,0)SVAR_restrict(&BQModel, "C", 5,3,0)
SVAR_cumulate(&BQModel, 1)SVAR_cumulate(&BQModel, 2)SVAR_cumulate(&BQModel,
3)SVAR_cumulate(&BQModel, 4)SVAR_cumulate(&BQModel, 5)
SVAR_estimate(&BQModel)
bfail = SVAR_boot(&BQModel, 2000, 0.9, 0)
?SVAR_estimate(&BQModel)
UnconstrainedSigma:
0.00004 0.00000 0.00071 0.00059 0.00014
0.00000 0.00000 0.00010 0.00364 0.00000
0.00071 0.00010 0.11088 0.24597 0.00347
0.00059 0.00364 0.24597 159.92282 -0.00971
0.00014 0.00000 0.00347 -0.00971 0.00084
Long-run matrixC(1) =
0.7736 -1.5321 0.0135 0.0001 0.1158
0.1695 1.4970 0.0076 0.0001 0.0254
-72.3749 -90.8705 2.9837 0.0254 15.0368
305.4149 -1451.6624 -17.4910 0.7748 -14.4628
-0.4012 -13.7671 0.0605 0.0006 1.6212
Optimizationmethod = Scoring algorithm
coefficient std. error z p-value
--------------------------------------------------------------
C[ 1; 1] 0.00520873 0.000629939 8.269 1.36e-016 ***
C[ 2; 1] -0.000263129 0.000231818 -1.135 0.2563
C[ 3; 1] 0.223805 0.0378243 5.917 3.28e-09 ***
C[ 4; 1] 1.58688 1.62616 0.9758 0.3291
C[ 5; 1] 0.0237545 0.00304730 7.795 6.43e-015 ***
C[ 1; 2] -0.00320062 0.000292175 -10.95 6.33e-028 ***
C[ 2; 2] -0.000529455 0.000225448 -2.348 0.0189 **
C[ 3; 2] 0.141552 0.0290911 4.866 1.14e-06 ***
C[ 4; 2] 2.39856 1.60483 1.495 0.1350
C[ 5; 2] -0.00460859 0.00209923 -2.195 0.0281 **
C[ 1; 3] 0.00000 0.00000 NA NA
C[ 2; 3] 0.000120530 0.000476143 0.2531 0.8002
C[ 3; 3] 0.114847 0.0437330 2.626 0.0086 ***
C[ 4; 3] -11.2260 1.65104 -6.799 1.05e-011 ***
C[ 5; 3] 0.00000 0.00000 NA NA
C[ 1; 4] 0.00000 0.00000 NA NA
C[ 2; 4] 0.00131884 0.000186939 7.055 1.73e-012 ***
C[ 3; 4] 0.166040 0.0362217 4.584 4.56e-06 ***
C[ 4; 4] 5.06246 3.24872 1.558 0.1192
C[ 5; 4] -0.00717976 0.00194941 -3.683 0.0002 ***
C[ 1; 5] 0.00000 0.00000 NA NA
C[ 2; 5] 0.00107498 9.81318e-05 10.95 6.33e-028 ***
C[ 3; 5] 0.00000 0.00000 NA NA
C[ 4; 5] 0.00000 0.00000 NA NA
C[ 5; 5] 0.0142210 0.00129819 10.95 6.33e-028 ***
Log-likelihood = 413.11
? bfail = SVAR_boot(&BQModel, 2000, 0.9, 0)
Bootstrapping model (2000 iterations)
errcode = 32
errcode = 32
errcode = 32
errcode = 32
errcode = 32
errcode = 32
errcode = 32
errcode = 32
Warning: couldn't improve criterion (gradient = 2.25191)
errcode = 32
errcode = 32
errcode = 32
errcode = 32
errcode = 32
Bootstrap results (2000 replications, 0 failed)
coefficient std. error z p-value
----------------------------------------------------------
C[ 1; 1] 0.00153498 0.00411606 0.3729 0.7092
C[ 2; 1] 2.14368e-05 0.000533545 0.04018 0.9680
C[ 3; 1] 0.0748155 0.171772 0.4356 0.6632
C[ 4; 1] 2.02507 1.62786 1.244 0.2135
C[ 5; 1] 0.00678030 0.0184051 0.3684 0.7126
C[ 1; 2] -0.000957296 0.00313491 -0.3054 0.7601
C[ 2; 2] 4.25591e-05 0.000707832 0.06013 0.9521
C[ 3; 2] 0.0331823 0.101026 0.3285 0.7426
C[ 4; 2] 2.80113 2.28712 1.225 0.2207
C[ 5; 2] -0.00259911 0.00885520 -0.2935 0.7691
C[ 1; 3] 0.00000 0.00000 NA NA
C[ 2; 3] -0.000110869 0.000565895 -0.1959 0.8447
C[ 3; 3] 0.0287889 0.143152 0.2011 0.8406
C[ 4; 3] -7.72066 3.47706 -2.220 0.0264 **
C[ 5; 3] 0.00000 0.00000 NA NA
C[ 1; 4] 0.00000 0.00000 NA NA
C[ 2; 4] 0.000248349 0.000768865 0.3230 0.7467
C[ 3; 4] 0.0170406 0.147616 0.1154 0.9081
C[ 4; 4] 5.01946 3.53079 1.422 0.1551
C[ 5; 4] -0.00276044 0.00696356 -0.3964 0.6918
C[ 1; 5] 0.00000 0.00000 NA NA
C[ 2; 5] 0.000545628 0.000746872 0.7306 0.4651
C[ 3; 5] 0.00000 0.00000 NA NA
C[ 4; 5] 0.00000 0.00000 NA NA
C[ 5; 5] 0.0114515 0.00396175 2.891 0.0038 ***
On Tuesday, August 2, 2016 7:01 PM, Sven Schreiber <svetosch(a)gmx.net> wrote:
Am 03.08.2016 um 00:16 schrieb jayeryu:
2. There is a red flag for the short and long run restrictions,
though.
I put the two restrictions simultaneously but the results are widely
different from what my coauthor obtained from RATS and GAUSS.
Thanks for reporting your case.
The result from gretl shows a C matrix that looks like one that just
superimposed short and long run matrices in a way that any 0 values in
either of the two matrices replaced the relevant element of the C.
It would be useful if you could copy&paste your restrictions and the C
matrix that gretl shows.
In case you're not using the latest gretl version please also add that
information.
Thanks,
sven
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
7:58 a.m.
New subject: SVAR bootstrap output table etc.
On Wed, 3 Aug 2016, Jay Ryu wrote:
Thank you for your prompt response!Below, I copied the gretl syntax
(restrictions and bootstrap). I also copied the gretl output. And yes, I
installed the latest version from the gretl site.Sincerely,Jay
From your output, it seems that the order of magnitude of variable 4
is
very different from the other four. This could provoke, in principle,
ugly numerical problems.
Could you try what happens if you multiply variables 1, 2, 3 and 5 by 100?
-------------------------------------------------------
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
-------------------------------------------------------
2:29 a.m.
New subject: SVAR bootstrap output table etc.
Thank you for your expert advice!It worked but there is still one question...
The results differ from what my coauthor obtained from RATS/GAUSS. What we did was solving
non-linear equations rendered by imposing short and long-run restrictions simultaneously.
We computed the reduced form VAR and its covariance matrix. Then, we computed and
combined the short-run restriction (contemporaneous) matrix and long-run restriction
matrices in the Structural Moving Average presentation. Then, the two combined matrices
were compared with the covariance matrix to estimate needed coefficients. This part of
solving multiple non-linear equations was done by GAUSS.
gretl solutions look great but the C matrix results imply as if short-run restrictions are
just superimposed onto the long-run restrictions, leaving the C matrix in your syntax. I
was wondering if the gretl syntax does the above process of solving non-linear
equations.Thank you again,Jay
On Wednesday, August 3, 2016 3:58 AM, Riccardo (Jack) Lucchetti
<r.lucchetti(a)univpm.it> wrote:
On Wed, 3 Aug 2016, Jay Ryu wrote:
Thank you for your prompt response!Below, I copied the gretl syntax
(restrictions and bootstrap). I also copied the gretl output. And yes, I
installed the latest version from the gretl site.Sincerely,Jay
From your output, it seems that the order of magnitude of variable 4
is
very different from the other four. This could provoke, in principle,
ugly numerical problems.
Could you try what happens if you multiply variables 1, 2, 3 and 5 by 100?
-------------------------------------------------------
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
-------------------------------------------------------
9:43 a.m.
New subject: SVAR bootstrap output table etc.
On Thu, 4 Aug 2016, Jay Ryu wrote:
Thank you for your expert advice!It worked but there is still one
question...
The results differ from what my coauthor obtained from RATS/GAUSS. What
we did was solving non-linear equations rendered by imposing short and
long-run restrictions simultaneously. We computed the reduced form VAR
and its covariance matrix. Then, we computed and combined the short-run
restriction (contemporaneous) matrix and long-run restriction matrices
in the Structural Moving Average presentation. Then, the two combined
matrices were compared with the covariance matrix to estimate needed
coefficients. This part of solving multiple non-linear equations was
done by GAUSS. gretl solutions look great but the C matrix results imply
as if short-run restrictions are just superimposed onto the long-run
restrictions, leaving the C matrix in your syntax. I was wondering if
the gretl syntax does the above process of solving non-linear
equations.Thank you again,Jay
The approach in the SVAR package is essentially the same as the one
thouroghly explained in Amisano-Giannini (1997). I tried to give a quick
outline in section 4.1 of the pdf help to SVAR. Basically, all constraints
(short- and long-run) are expressed together as a system of equations
R vec(C) = d
where each row of R is a restriction. Short-run restrictions are just
restrictions on the elements of C, so typically the corresponding row of R
is a selection vector (all zeros and one 1). Long-run restrictions are
expressed by putting selected elements of the long-run matrix $\Theta(1)$
into the appropriate places of the corresponding row of the matrix R. This
is asymptotically justifiable as long as the information matrix is
block-diagonal between the mean and the variance parameters.
-------------------------------------------------------
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
-------------------------------------------------------
10:24 a.m.
New subject: SVAR bootstrap output table etc.
Am 04.08.2016 um 04:29 schrieb Jay Ryu:
Thank you for your expert advice!
It worked but there is still one question...
The results differ from what my coauthor obtained from RATS/GAUSS. What
we did was solving non-linear equations rendered by imposing short and
long-run restrictions simultaneously.
So if I understand correctly, when you say "from RATS/GAUSS" that means
with an algorithm that you implemented yourself? Initially I thought you
meant some prebuilt routine, but apparently that's not the case?
Thanks,
Sven
3375
days inactive
3377
days old
6 comments
4 participants
participants (4)
-
Jay Ryu -
jayeryu
-
Riccardo (Jack) Lucchetti -
Sven Schreiber