1:26 a.m.
Dear Allin and Sven,
Many thanks for your quick responses. I do agree with Sven that it
shouldn't matter qualitatively, but I think that it should be considered
normal to expect quantitatively consistent results when running the same
test, even though it's through different procedures.
By the way, I downloaded and installed the current snapshot (1.10.90.cvs
build date 2015-05-30) but the results are still different.
adf runs the test-down on the common data set (N-1-maxlag) for lag
selection, but reports the "optimal" lag ADF results on the full data set
(N-1-bestlag).
coint reports the ADF results directly from the test-down run for the
"optimal" lag (N-1-maxlag).
Koray
7:03 a.m.
New subject: test-down option reports different results in adf
Am 03.06.2015 um 03:26 schrieb Koray Simsek:
By the way, I downloaded and installed the current snapshot
(1.10.90.cvs
build date 2015-05-30) but the results are still different.
adf runs the test-down on the common data set (N-1-maxlag) for lag
selection, but reports the "optimal" lag ADF results on the full data
set (N-1-bestlag).
coint reports the ADF results directly from the test-down run for the
"optimal" lag (N-1-maxlag).
I don't have time right now, but if Allin says the code is supposed to
handle this consistently and it doesn't, then of course it's a problem.
-sven
12:51 p.m.
New subject: test-down option reports different results in adf
On Wed, 3 Jun 2015, Koray Simsek wrote:
Dear Allin and Sven,
Many thanks for your quick responses. I do agree with Sven that it
shouldn't matter qualitatively, but I think that it should be considered
normal to expect quantitatively consistent results when running the same
test, even though it's through different procedures.
By the way, I downloaded and installed the current snapshot (1.10.90.cvs
build date 2015-05-30) but the results are still different.
adf runs the test-down on the common data set (N-1-maxlag) for lag
selection, but reports the "optimal" lag ADF results on the full data set
(N-1-bestlag).
coint reports the ADF results directly from the test-down run for the
"optimal" lag (N-1-maxlag).
I'm not sure I understand this. For reference I'm appending a script
and its output. I'm seeing identical single-variable ADF tests on the
variable LRM in the context of "adf" and "coint". In both cases we
start with the specified max of 5 lags and find that AIC is minimized
at 5 lags, so we lose 6 observations.
When it comes to the cointegrating regression we run that on all
available data, and testing down on the residuals from this
regression we find that the optimal lag order is 0, so we lose only
one observation in the final (A)DF regression.
I can see there are certain points in the process where we might do
things differently -- e.g. restrict the cointegrating regression to
the same sample as the initial ADF regressions? But I'm not seeing
anything I would describe as inconsistent.
<hansl>
open denmark.gdt
adf 5 LRM --c --test-down --verbose
coint 5 LRM LRY --test-down --verbose
</hansl>
<output>
gretl version 1.10.90cvs
Copyright Ramu Ramanathan, Allin Cottrell and Riccardo "Jack" Lucchetti
This is free software with ABSOLUTELY NO WARRANTY
Current session: 2015-06-03 10:45
? run simsek.inp
/home/cottrell/stats/esl/gretl/build/cli/simsek.inp
? open denmark.gdt
Read datafile /opt/esl/share/gretl/data/misc/denmark.gdt
periodicity: 4, maxobs: 55
observations range: 1974:1 to 1987:3
Listing 5 variables:
0) const 1) LRM 2) LRY 3) IBO 4) IDE
? adf 5 LRM --c --test-down --verbose
k = 5: AIC = -198.839
k = 4: AIC = -197.883
k = 3: AIC = -195.102
k = 2: AIC = -196.848
k = 1: AIC = -187.891
k = 0: AIC = -189.709
Augmented Dickey-Fuller test for LRM
including 5 lags of (1-L)LRM
(max was 5, criterion AIC)
sample size 49
unit-root null hypothesis: a = 1
test with constant
model: (1-L)y = b0 + (a-1)*y(-1) + ... + e
1st-order autocorrelation coeff. for e: 0.035
lagged differences: F(5, 42) = 4.012 [0.0046]
estimated value of (a - 1): -0.0300808
test statistic: tau_c(1) = -0.775124
asymptotic p-value 0.8255
Augmented Dickey-Fuller regression
OLS, using observations 1975:3-1987:3 (T = 49)
Dependent variable: d_LRM
coefficient std. error t-ratio p-value
-------------------------------------------------------
const 0.357614 0.454896 0.7861 0.4362
LRM_1 -0.0300808 0.0388078 -0.7751 0.8255
d_LRM_1 0.183648 0.145021 1.266 0.2124
d_LRM_2 0.304352 0.143462 2.121 0.0398 **
d_LRM_3 -0.0249137 0.153967 -0.1618 0.8722
d_LRM_4 0.317505 0.153274 2.071 0.0445 **
d_LRM_5 -0.261515 0.161820 -1.616 0.1136
AIC: -198.839 BIC: -185.597 HQC: -193.815
? coint 5 LRM LRY --test-down --verbose
Step 1: testing for a unit root in LRM
k = 5: AIC = -198.839
k = 4: AIC = -197.883
k = 3: AIC = -195.102
k = 2: AIC = -196.848
k = 1: AIC = -187.891
k = 0: AIC = -189.709
Augmented Dickey-Fuller test for LRM
including 5 lags of (1-L)LRM
(max was 5, criterion AIC)
sample size 49
unit-root null hypothesis: a = 1
test with constant
model: (1-L)y = b0 + (a-1)*y(-1) + ... + e
1st-order autocorrelation coeff. for e: 0.035
lagged differences: F(5, 42) = 4.012 [0.0046]
estimated value of (a - 1): -0.0300808
test statistic: tau_c(1) = -0.775124
asymptotic p-value 0.8255
Augmented Dickey-Fuller regression
OLS, using observations 1975:3-1987:3 (T = 49)
Dependent variable: d_LRM
coefficient std. error t-ratio p-value
-------------------------------------------------------
const 0.357614 0.454896 0.7861 0.4362
LRM_1 -0.0300808 0.0388078 -0.7751 0.8255
d_LRM_1 0.183648 0.145021 1.266 0.2124
d_LRM_2 0.304352 0.143462 2.121 0.0398 **
d_LRM_3 -0.0249137 0.153967 -0.1618 0.8722
d_LRM_4 0.317505 0.153274 2.071 0.0445 **
d_LRM_5 -0.261515 0.161820 -1.616 0.1136
AIC: -198.839 BIC: -185.597 HQC: -193.815
Step 2: testing for a unit root in LRY
k = 5: AIC = -217.928
k = 4: AIC = -219.841
k = 3: AIC = -220.730
k = 2: AIC = -222.395
k = 1: AIC = -224.395
k = 0: AIC = -225.512
Augmented Dickey-Fuller test for LRY
including 0 lags of (1-L)LRY
(max was 5, criterion AIC)
sample size 49
unit-root null hypothesis: a = 1
test with constant
model: (1-L)y = b0 + (a-1)*y(-1) + e
1st-order autocorrelation coeff. for e: 0.154
estimated value of (a - 1): -0.128707
test statistic: tau_c(1) = -2.47723
p-value 0.1271
Dickey-Fuller regression
OLS, using observations 1975:3-1987:3 (T = 49)
Dependent variable: d_LRY
coefficient std. error t-ratio p-value
-------------------------------------------------------
const 0.772640 0.309740 2.494 0.0162 **
LRY_1 -0.128707 0.0519562 -2.477 0.1271
AIC: -225.512 BIC: -221.729 HQC: -224.077
Step 3: cointegrating regression
Cointegrating regression -
OLS, using observations 1974:1-1987:3 (T = 55)
Dependent variable: LRM
coefficient std. error t-ratio p-value
--------------------------------------------------------
const 0.929031 0.848846 1.094 0.2787
LRY 1.81866 0.142595 12.75 8.80e-18 ***
Mean dependent var 11.75441 S.D. dependent var 0.152357
Sum squared resid 0.308048 S.E. of regression 0.076238
R-squared 0.754248 Adjusted R-squared 0.749611
Log-likelihood 64.54124 Akaike criterion -125.0825
Schwarz criterion -121.0678 Hannan-Quinn -123.5300
rho 0.856564 Durbin-Watson 0.298798
Step 4: testing for a unit root in uhat
k = 5: AIC = -170.487
k = 4: AIC = -172.373
k = 3: AIC = -174.126
k = 2: AIC = -176.069
k = 1: AIC = -176.656
k = 0: AIC = -177.323
Augmented Dickey-Fuller test for uhat
including 0 lags of (1-L)uhat
(max was 5, criterion AIC)
sample size 54
unit-root null hypothesis: a = 1
model: (1-L)y = (a-1)*y(-1) + e
1st-order autocorrelation coeff. for e: -0.087
estimated value of (a - 1): -0.143436
test statistic: tau_c(2) = -1.95648
p-value 0.5548
Dickey-Fuller regression
OLS, using observations 1974:2-1987:3 (T = 54)
Dependent variable: d_uhat
coefficient std. error t-ratio p-value
-------------------------------------------------------
uhat_1 -0.143436 0.0733134 -1.956 0.5548
AIC: -192.742 BIC: -190.753 HQC: -191.975
There is evidence for a cointegrating relationship if:
(a) The unit-root hypothesis is not rejected for the individual variables, and
(b) the unit-root hypothesis is rejected for the residuals (uhat) from the
cointegrating regression.
Done
</output>
Allin Cottrell
3:43 p.m.
New subject: test-down option reports different results in adf
Am 03.06.2015 um 14:51 schrieb Allin Cottrell:
On Wed, 3 Jun 2015, Koray Simsek wrote:
>
> adf runs the test-down on the common data set (N-1-maxlag) for lag
> selection, but reports the "optimal" lag ADF results on the full data set
> (N-1-bestlag).
>
> coint reports the ADF results directly from the test-down run for the
> "optimal" lag (N-1-maxlag).
I'm not sure I understand this. For reference I'm appending a script and
its output. I'm seeing identical single-variable ADF tests on the
variable LRM in the context of "adf" and "coint". In both cases we
start
with the specified max of 5 lags and find that AIC is minimized at 5
lags, so we lose 6 observations.
I think Koray's point applies to the case where the maxlag and the best
lag differ, whereas in your example they're equal. When they're
different there are two different possible samples, one that starts at
t0+maxlag and one that starts at t0+bestlag. His point was that adf and
coint choose different samples then. (If I understood him correctly.)
cheers,
sven
11:47 p.m.
New subject: test-down option reports different results in adf
Dear Allin,
I am sorry that the two lines I sent in my original post do not replicate
the issue in the current snapshot as they were run in an older gretl where
MAIC was used as the test-down criteria and selected a smaller lag value
for test-down in LRM.
Sven explained my point perfectly. If you run ADF for LRY (not LRM) as
follows, you'll notice that AIC below the regression output is different
from the AIC for k=0 in the test-down procedure. This is because the
regression is run on the full-sample whereas the test-down is done for the
lag 5 compatible sample.
adf 5 LRY --c --test-down --verbose
k = 5: AIC = -217.928
k = 4: AIC = -219.841
k = 3: AIC = -220.730
k = 2: AIC = -222.395
k = 1: AIC = -224.395
k = 0: AIC = -225.512
Augmented Dickey-Fuller test for LRY
including 0 lags of (1-L)LRY
(max was 5, criterion AIC)
sample size 54
unit-root null hypothesis: a = 1
test with constant
model: (1-L)y = b0 + (a-1)*y(-1) + e
1st-order autocorrelation coeff. for e: 0.195
estimated value of (a - 1): -0.0484706
test statistic: tau_c(1) = -1.00236
p-value 0.7463
Dickey-Fuller regression
OLS, using observations 1974:2-1987:3 (T = 54)
Dependent variable: d_LRY
coefficient std. error t-ratio p-value
-------------------------------------------------------
const 0.291153 0.287768 1.012 0.3163
LRY_1 −0.0484706 0.0483563 −1.002 0.7463
AIC: -241.464 BIC: -237.486 HQC: -239.93
Now if you look at the step 2 of the coint you ran above, you will notice
that the regression output is different and it belongs to the one for
test-down sample.
I might be wrong, but I think adf was changed at some point to behave like
this. Perhaps coint was forgotten along the way :)
Best,
Koray
On Wed, Jun 3, 2015 at 6:43 PM, Sven Schreiber <svetosch(a)gmx.net> wrote:
Am 03.06.2015 um 14:51 schrieb Allin Cottrell:
> On Wed, 3 Jun 2015, Koray Simsek wrote:
>
>> adf runs the test-down on the common data set (N-1-maxlag) for lag
>> selection, but reports the "optimal" lag ADF results on the full data
set
>> (N-1-bestlag).
>>
>> coint reports the ADF results directly from the test-down run for the
>> "optimal" lag (N-1-maxlag).
>>
>
> I'm not sure I understand this. For reference I'm appending a script and
> its output. I'm seeing identical single-variable ADF tests on the
> variable LRM in the context of "adf" and "coint". In both cases
we start
> with the specified max of 5 lags and find that AIC is minimized at 5
> lags, so we lose 6 observations.
>
I think Koray's point applies to the case where the maxlag and the best
lag differ, whereas in your example they're equal. When they're different
there are two different possible samples, one that starts at t0+maxlag and
one that starts at t0+bestlag. His point was that adf and coint choose
different samples then. (If I understood him correctly.)
cheers,
sven
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
4:34 p.m.
New subject: test-down option reports different results in adf
On Thu, 4 Jun 2015, Koray Simsek wrote:
Dear Allin,
I am sorry that the two lines I sent in my original post do not replicate
the issue in the current snapshot as they were run in an older gretl where
MAIC was used as the test-down criteria and selected a smaller lag value
for test-down in LRM.
Sven explained my point perfectly. If you run ADF for LRY (not LRM) as
follows, you'll notice that AIC below the regression output is different
from the AIC for k=0 in the test-down procedure. This is because the
regression is run on the full-sample whereas the test-down is done for the
lag 5 compatible sample.
OK, thanks, now I see. This should be fixed in current CVS and
snapshots.
Allin Cottrell
4:45 p.m.
New subject: test-down option reports different results in adf
Am 06.06.2015 um 18:34 schrieb Allin Cottrell:
On Thu, 4 Jun 2015, Koray Simsek wrote:
> Dear Allin,
>
> I am sorry that the two lines I sent in my original post do not replicate
> the issue in the current snapshot as they were run in an older gretl
> where
> MAIC was used as the test-down criteria and selected a smaller lag value
> for test-down in LRM.
>
> Sven explained my point perfectly. If you run ADF for LRY (not LRM) as
> follows, you'll notice that AIC below the regression output is different
> from the AIC for k=0 in the test-down procedure. This is because the
> regression is run on the full-sample whereas the test-down is done for
> the
> lag 5 compatible sample.
OK, thanks, now I see. This should be fixed in current CVS and snapshots.
And which of the two samples is used now, starting at t0 + maxlag or at
t0 + bestlag?
thanks,
sven
4:50 p.m.
New subject: test-down option reports different results in adf
On Sat, 6 Jun 2015, Sven Schreiber wrote:
Am 06.06.2015 um 18:34 schrieb Allin Cottrell:
> On Thu, 4 Jun 2015, Koray Simsek wrote:
>
>> Dear Allin,
>>
>> I am sorry that the two lines I sent in my original post do not replicate
>> the issue in the current snapshot as they were run in an older gretl
>> where
>> MAIC was used as the test-down criteria and selected a smaller lag value
>> for test-down in LRM.
>>
>> Sven explained my point perfectly. If you run ADF for LRY (not LRM) as
>> follows, you'll notice that AIC below the regression output is different
>> from the AIC for k=0 in the test-down procedure. This is because the
>> regression is run on the full-sample whereas the test-down is done for
>> the
>> lag 5 compatible sample.
>
> OK, thanks, now I see. This should be fixed in current CVS and snapshots.
And which of the two samples is used now, starting at t0 + maxlag or at
t0 + bestlag?
The same as the stand-alone ADF test: t0 + bestlag.
Allin
1:40 p.m.
New subject: test-down option reports different results in adf
Dear Allin,
Thanks for quickly acting on this, but when I installed the current
snapshot, I noticed that the coint Steps 1,2, and 4 completely disappeared.
Here's what my output looks like when I run coint with --verbose:
Step 1: testing for a unit root in X1
Step 2: testing for a unit root in X2
Step 3: cointegrating regression
Cointegrating regression -
OLS, using observations 2012-12-14:2013-12-13 (T = 250)
Dependent variable: X1
coefficient std. error t-ratio p-value
---------------------------------------------------------
const 0.331381 0.119003 2.785 0.0058 ***
X2 0.932445 0.0138352 67.40 1.83e-161 ***
Mean dependent var 8.297023 S.D. dependent var 0.962697
Sum squared resid 11.94726 S.E. of regression 0.219487
R-squared 0.948229 Adjusted R-squared 0.948020
Log-likelihood 25.38522 Akaike criterion −46.77044
Schwarz criterion −39.72752 Hannan-Quinn −43.93587
rho 0.912250 Durbin-Watson 0.174442
Step 4: testing for a unit root in uhat
There is evidence for a cointegrating relationship if:
(a) The unit-root hypothesis is not rejected for the individual variables,
and
(b) the unit-root hypothesis is rejected for the residuals (uhat) from the
cointegrating regression.
I hate to say this, but my comment might have led to the introduction of a
bug.
Can I also have a feature request while this is being fixed? Using
--verbose produces the AIC and BIC values for all lags if selected as
test-down criteria, but when tstat is selected for test-down, nothing is
produced. I know tstat test-down stops when >1.645 is observed so that
t-stats for lower lags are not available, but it could be nice to see other
t-stats above that lag again for the sake of consistency in the --verbose
option.
Best,
Koray
On Sat, Jun 6, 2015 at 7:50 PM, Allin Cottrell <cottrell(a)wfu.edu> wrote:
On Sat, 6 Jun 2015, Sven Schreiber wrote:
Am 06.06.2015 um 18:34 schrieb Allin Cottrell:
>
>> On Thu, 4 Jun 2015, Koray Simsek wrote:
>>
>> Dear Allin,
>>>
>>> I am sorry that the two lines I sent in my original post do not
>>> replicate
>>> the issue in the current snapshot as they were run in an older gretl
>>> where
>>> MAIC was used as the test-down criteria and selected a smaller lag value
>>> for test-down in LRM.
>>>
>>> Sven explained my point perfectly. If you run ADF for LRY (not LRM) as
>>> follows, you'll notice that AIC below the regression output is different
>>> from the AIC for k=0 in the test-down procedure. This is because the
>>> regression is run on the full-sample whereas the test-down is done for
>>> the
>>> lag 5 compatible sample.
>>>
>>
>> OK, thanks, now I see. This should be fixed in current CVS and snapshots.
>>
>
> And which of the two samples is used now, starting at t0 + maxlag or at
> t0 + bestlag?
>
The same as the stand-alone ADF test: t0 + bestlag.
Allin
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
3:58 p.m.
New subject: test-down option reports different results in adf
I can't replicate this problem. Could you give me an example I could
try running? Thanks.
(I'll try to get to the accompanying feature request before long.)
Allin
On Thu, Jun 11, 2015 at 3:40 PM, Koray Simsek <korays(a)gmail.com> wrote:
Dear Allin,
Thanks for quickly acting on this, but when I installed the current
snapshot, I noticed that the coint Steps 1,2, and 4 completely disappeared.
Here's what my output looks like when I run coint with --verbose:
Step 1: testing for a unit root in X1
Step 2: testing for a unit root in X2
Step 3: cointegrating regression
Cointegrating regression -
OLS, using observations 2012-12-14:2013-12-13 (T = 250)
Dependent variable: X1
coefficient std. error t-ratio p-value
---------------------------------------------------------
const 0.331381 0.119003 2.785 0.0058 ***
X2 0.932445 0.0138352 67.40 1.83e-161 ***
Mean dependent var 8.297023 S.D. dependent var 0.962697
Sum squared resid 11.94726 S.E. of regression 0.219487
R-squared 0.948229 Adjusted R-squared 0.948020
Log-likelihood 25.38522 Akaike criterion −46.77044
Schwarz criterion −39.72752 Hannan-Quinn −43.93587
rho 0.912250 Durbin-Watson 0.174442
Step 4: testing for a unit root in uhat
There is evidence for a cointegrating relationship if:
(a) The unit-root hypothesis is not rejected for the individual variables,
and
(b) the unit-root hypothesis is rejected for the residuals (uhat) from the
cointegrating regression.
9:48 p.m.
New subject: test-down option reports different results in adf
Allin,
I restarted gretl and reloaded the data file I have been using and you are
right -- it works as expected. Sorry for bothering you with this. I should
have double-checked it. Many thanks for all your help again.
Koray
On Thu, Jun 11, 2015 at 6:58 PM, Cottrell, Allin <cottrell(a)wfu.edu> wrote:
I can't replicate this problem. Could you give me an example I
could
try running? Thanks.
(I'll try to get to the accompanying feature request before long.)
Allin
On Thu, Jun 11, 2015 at 3:40 PM, Koray Simsek <korays(a)gmail.com> wrote:
> Dear Allin,
>
> Thanks for quickly acting on this, but when I installed the current
> snapshot, I noticed that the coint Steps 1,2, and 4 completely
disappeared.
> Here's what my output looks like when I run coint with --verbose:
>
> Step 1: testing for a unit root in X1
> Step 2: testing for a unit root in X2
> Step 3: cointegrating regression
>
> Cointegrating regression -
> OLS, using observations 2012-12-14:2013-12-13 (T = 250)
> Dependent variable: X1
>
> coefficient std. error t-ratio p-value
> ---------------------------------------------------------
> const 0.331381 0.119003 2.785 0.0058 ***
> X2 0.932445 0.0138352 67.40 1.83e-161 ***
>
> Mean dependent var 8.297023 S.D. dependent var 0.962697
> Sum squared resid 11.94726 S.E. of regression 0.219487
> R-squared 0.948229 Adjusted R-squared 0.948020
> Log-likelihood 25.38522 Akaike criterion −46.77044
> Schwarz criterion −39.72752 Hannan-Quinn −43.93587
> rho 0.912250 Durbin-Watson 0.174442
>
> Step 4: testing for a unit root in uhat
>
> There is evidence for a cointegrating relationship if:
> (a) The unit-root hypothesis is not rejected for the individual
variables,
> and
> (b) the unit-root hypothesis is rejected for the residuals (uhat) from
the
> cointegrating regression.
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
3494
days inactive
3502
days old
10 comments
4 participants
participants (4)
-
Allin Cottrell -
Cottrell, Allin
-
Koray Simsek -
Sven Schreiber