2:49 a.m.
Is there an option to compute the rolling correlation between multiple
time series? I'm seeing some mentioning on this topic in 2016 between
Fernando and Sven but not sure what happened to it. Thanks
7:06 a.m.
On Mon, 16 Jan 2017, Tomas Nilsson wrote:
Is there an option to compute the rolling correlation between
multiple
time series? I'm seeing some mentioning on this topic in 2016 between
Fernando and Sven but not sure what happened to it. Thanks
you mean, given n series, return n*(n+1) series of covariances computed on
a fixed observation window?
If so, try this:
<hansl>
function list rolling_vcv(list X, scalar length)
list ret = null
n = nelem(X)
m = n*(n+1)/2
scalar T = $nobs
set warnings off
matrix wrk = mshape(NA, T, m)
matrix mX = {X}
ini = 1
fin = length
loop t = length .. T --quiet
matrix cX = mX[ini:fin,]
wrk[t,] = vech(mcov(cX))'
ini++
fin++
endloop
k = 1
loop i=1..n --quiet
sname = sprintf("var_%02d", i)
ret += genseries(sname, wrk[, k])
k++
loop j = i+1..n --quiet
sname = sprintf("cov_%02d_%02d", i, j)
ret += genseries(sname, wrk[, k])
k++
endloop
endloop
return ret
end function
open AWM --quiet
list RATES = STN LTN
list R = rolling_vcv(RATES, 16)
</hansl>
hope this helps
-------------------------------------------------------
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
-------------------------------------------------------
8:14 a.m.
Am 17.01.2017 um 08:06 schrieb Riccardo (Jack) Lucchetti:
On Mon, 16 Jan 2017, Tomas Nilsson wrote:
> Is there an option to compute the rolling correlation between multiple
> time series? I'm seeing some mentioning on this topic in 2016 between
> Fernando and Sven but not sure what happened to it. Thanks
I re-checked, and I think what happened is that Fernando found a
solution for himself ;-)
you mean, given n series, return n*(n+1) series of covariances
computed
on a fixed observation window?
If so, try this:
Jack, sorry, but your code must look just horrible for regular gretl
users. Why wouldn't a simple loop be enough?
(untested:)
<hansl>
list X = mylist # define variable list here
N = 10 # set how many subsamples here
loop N
corr(X) # prints out the correlation matrix
smpl +1 +1 # moves the subsample forward by one
endloop
</hansl>
If you want to capture the numbers for further use, you can do that with
an array of matrices:
(untested)
<hansl>
matrices MC = array(N)
loop i = 1..N
matrix MC[i] = mcorr( {X} )
smpl +1 +1
endloop
</hansl>
HTH,
sven
1:30 p.m.
On Tue, 17 Jan 2017, Sven Schreiber wrote:
Am 17.01.2017 um 08:06 schrieb Riccardo (Jack) Lucchetti:
> On Mon, 16 Jan 2017, Tomas Nilsson wrote:
>
>> Is there an option to compute the rolling correlation between multiple
>> time series? I'm seeing some mentioning on this topic in 2016 between
>> Fernando and Sven but not sure what happened to it. Thanks
I re-checked, and I think what happened is that Fernando found a solution for
himself ;-)
> you mean, given n series, return n*(n+1) series of covariances computed
> on a fixed observation window?
>
> If so, try this:
Jack, sorry, but your code must look just horrible for regular gretl users.
Why wouldn't a simple loop be enough?
(untested:)
<hansl>
list X = mylist # define variable list here
N = 10 # set how many subsamples here
loop N
corr(X) # prints out the correlation matrix
smpl +1 +1 # moves the subsample forward by one
endloop
</hansl>
If you want to capture the numbers for further use, you can do that with an
array of matrices:
(untested)
<hansl>
matrices MC = array(N)
loop i = 1..N
matrix MC[i] = mcorr( {X} )
smpl +1 +1
endloop
</hansl>
You mean it's too long and complex? Hm, maybe you're right.
Well, storing the results as series does have certain advantages. For
example, it's easy to plot them through time. Besides, I think that the
script I posted is reasonably general while not being horribly
inefficient. Finally, it's a nice showcase for lesser-known features like
the genseries() function.
-------------------------------------------------------
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
-------------------------------------------------------
3:44 p.m.
Thanks the advice; t'was very helpful. Now to another dumb question but
with this time with a twist. I want to compute the correlation (of change)
between two time series of different frequencies, monthly oil prices and
quarterly manufacturing shipments. I would interpret the recommendation
from my old econometrics textbook to turn the monthly data into quarterly;
and then compute the rolling correlation of change. According to the new
age, is there a better way to do this? Is there a package in GRETL that
allows users to examine time series patterns with data of different
frequencies?
On Mon, Jan 16, 2017 at 7:49 PM, Tomas Nilsson <tknilsso(a)gmail.com> wrote:
Is there an option to compute the rolling correlation between
multiple
time series? I'm seeing some mentioning on this topic in 2016 between
Fernando and Sven but not sure what happened to it. Thanks
4:42 p.m.
Am 17.02.2017 um 16:44 schrieb Tomas Nilsson:
Thanks the advice; t'was very helpful. Now to another dumb
question but
with this time with a twist. I want to compute the correlation (of
change) between two time series of different frequencies, monthly oil
prices and quarterly manufacturing shipments. I would interpret the
recommendation from my old econometrics textbook to turn the monthly
data into quarterly; and then compute the rolling correlation of change.
According to the new age, is there a better way to do this? Is there a
package in GRETL that allows users to examine time series patterns with
data of different frequencies?
Gretl can handle time series of different frequencies. This was just
implemented a while ago. Have a look at the MIDAS manual:
http://gretl.sourceforge.net/midas/midas_gretl.pdf
However, I am pretty sure there is no ready-made package for computing
some rolling statistics though, yet.
Best,
Artur
3178
days inactive
3209
days old
5 comments
4 participants
participants (4)
-
Artur Tarassow -
Riccardo (Jack) Lucchetti -
Sven Schreiber -
Tomas Nilsson