R devel - Apr 2007 - predict.ar() produces wrong SE's (PR#9614)

Full_Name: Kirk Hampel
Version: 2.4.1
OS: Windows
Submission from: (NULL) (144.53.251.2)


Given an AR(p) model, the last p SE's are wrong.

The source of the bug is that the C code (ver 2.4.0) assumes *npsi is the length
of the psi vector (which is n+p), whilst the predict.ar function in R passes out
as.integer(npsi), where npsi <- n-1.

Some R code following reproduces the error. Let p=4, n=6, then
> ar<-c(-0.5,0.25,-0.125,0.0625)
> Mod(polyroot(c(1,-ar))) # ar model is stationary
[1] 1.037580 2.444163 2.444163 2.581298
> #Pass in values as predict.ar does
>
cumsum(c(1,.C("artoma",as.integer(4),as.double(ar),double(9),as.integer(5))[[3]][1:5]^2))
[1] 1.000000 1.250000 1.312500 1.328125 1.332031
1.332031
> #Pass in value that C code expects
>
cumsum(c(1,.C("artoma",as.integer(4),as.double(ar),double(9),as.integer(9))[[3]][1:5]^2))
[1] 1.000000 1.250000 1.500000 1.750000 2.000000 2.219727

Thank you.  Here is a reproducible example (since yours is not of the 
actual code):

example(ar)
predict(sunspot.ar, n.ahead=25)
predict(sunspot.ar, n.ahead=20)

and compare, say, prediction 20.

Fixed for 2.5.0.

On Tue, 17 Apr 2007, kirk.hampel at abs.gov.au wrote:
> Full_Name: Kirk Hampel
> Version: 2.4.1
> OS: Windows
> Submission from: (NULL) (144.53.251.2)
>
>
> Given an AR(p) model, the last p SE's are wrong.
>
> The source of the bug is that the C code (ver 2.4.0) assumes *npsi is the
length
> of the psi vector (which is n+p), whilst the predict.ar function in R
passes out
> as.integer(npsi), where npsi <- n-1.
>
> Some R code following reproduces the error. Let p=4, n=6, then
>
>> ar<-c(-0.5,0.25,-0.125,0.0625)
>> Mod(polyroot(c(1,-ar))) # ar model is stationary
> [1] 1.037580 2.444163 2.444163 2.581298
>> #Pass in values as predict.ar does
>>
cumsum(c(1,.C("artoma",as.integer(4),as.double(ar),double(9),as.integer(5))[[3]][1:5]^2))
> [1] 1.000000 1.250000 1.312500 1.328125 1.332031 1.332031
>> #Pass in value that C code expects
>>
cumsum(c(1,.C("artoma",as.integer(4),as.double(ar),double(9),as.integer(9))[[3]][1:5]^2))
> [1] 1.000000 1.250000 1.500000 1.750000 2.000000 2.219727
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>
-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595