Re: [Gretl-devel] arma testing: creating a test suite
Dear all,
I think series from M4 competition
make a decent test suit for arima
Since the data set is somewhat large
I provide R code using my Rgretl package
Below is R code to
create gretl data bases with
data from Makridakis competition, generation 4 (M4-competition)
Main function is Rgretl::save_bin
<R>
# download and install from local disk
# https://cran.r-project.org/src/contrib/Archive/M4comp/M4comp_0.0.1.tar.gz
library(M4comp)
yrl = sapply(M4,FUN = function(x) (x$type ==
"ECONOMICS")&&(x$period=="YEARLY"))
YEARLY = M4[yrl]
yearly = sapply(YEARLY,FUN=function(x) x$past)
names(yearly) <- paste0("series",1:length(yearly))
library(Rgretl)
# creating gretl data base 'm4Eyearly.bin'
# default path is R current working directory
save_bin("m4Eyearly",yearly)
qtl = sapply(M4,FUN = function(x) (x$type ==
"ECONOMICS")&&(x$period=="QUARTERLY"))
QUARTERLY = M4[qtl]
quarterly = sapply(QUARTERLY,FUN=function(x) x$past)
names(quarterly) <- paste0("seriesEQ",1:length(quarterly))
save_bin("m4Equarterly",quarterly)
qtlbi = sapply(M4,FUN = function(x) (x$type ==
"BUSINESS-INDUSTRY")&&(x$period=="QUARTERLY"))
QUARTERLYbi = M4[qtlbi]
quarterlybi = sapply(QUARTERLYbi,FUN=function(x) x$past)
names(quarterlybi) <- paste0("seriesBIQ",1:length(quarterlybi))
save_bin("m4BIquarterly",quarterlybi)
mnth = sapply(M4,FUN = function(x) (x$type ==
"ECONOMICS")&&(x$period=="MONTHLY"))
MONTHLY = M4[mnth]
monthly = sapply(MONTHLY,FUN=function(x) x$past)
names(monthly) <- paste0("seriesEM",1:length(monthly))
save_bin("m4Emonthly",monthly)
mnthbi = sapply(M4,FUN = function(x) (x$type ==
"BUSINESS-INDUSTRY")&&(x$period=="MONTHLY"))
MONTHLYbi = M4[mnthbi]
monthlybi = sapply(MONTHLYbi,FUN=function(x) x$past)
names(monthlybi) <- paste0("seriesBIM",1:length(monthlybi))
save_bin("m4BImonthly",monthlybi)
<R>
Oleh
20 березня 2018, 02:05:26, від "Allin Cottrell" <cottrell(a)wfu.edu>:
In current git and snapshots there's a "secret" option
for the arma
command, --as197 or just -a, which invokes new code using Melard's
algorithm AS 197 in place of the default Kalman-based code when
estimation is via native exact ML.
The hope -- borne out by our testing to date -- is that AS 197 will
be a good deal faster than the Kalman code without sacrifice of
accuracy. But it would be good to have that confirmed or denied on a
wider range of cases. (However, note that AS 197 will not be used if
there are missing values in the sample range.)
We've also been thinking about revising our arma initialization
code, but at present both Kalman and AS 197 use the "same old"
initialization methods.
I might point out that you can expect slightly different
coefficients and/or standard errors from Kalman and AS 197: these
often change a little even when one and the same algorithm is
employed, with updates to the C compiler or blas/lapack. Our concern
is ensuring that we don't get a noticeably inferior maximized
loglikelihood with AS 197, or failure to converge when the Kalman
variant converges.
Allin
_______________________________________________
Gretl-devel mailing list
Gretl-devel@lists.wfu.eduhttp://lists.wfu.edu/mailman/listinfo/gretl-devel