R devel - Nov 2006 - bug in arima? (PR#9404)

I don't think arima works exactly the way one would expect when there is
differencing.  What I think should happen is that by
default the mean of the differenced series is estimated and if include.mean=F,
then it is not.  This is not what happens.  Instead
when there is differencing the include.mean argument is ignored.

Now I guess, someone could argue that the mean of the original series
doesn't exist when using a model with differencing.  In this
case, then the arima function doesn't conveniently estimate models with
deterministic drift like the ARIMA(0,1,1) with drift.  Such
models are occassionally important.

As a workaround, one can do the differencing outside of arima.  But I don't
think things should work like that!

Most other time series software doesn't do this.

Ian McLeod

On 11/30/2006 2:20 PM, aim at stats.uwo.ca wrote:
> I don't think arima works exactly the way one would expect when there
is differencing.  What I think should happen is that by
> default the mean of the differenced series is estimated and if
include.mean=F, then it is not.  This is not what happens.  Instead
> when there is differencing the include.mean argument is ignored.
But it's working as documented, so this is not a bug.  I'd be very 
reluctant to change the meaning of that parameter.

Duncan
> Now I guess, someone could argue that the mean of the original series
doesn't exist when using a model with differencing.  In this
> case, then the arima function doesn't conveniently estimate models with
deterministic drift like the ARIMA(0,1,1) with drift.  Such
> models are occassionally important.
> 
> As a workaround, one can do the differencing outside of arima.  But I
don't think things should work like that!
> 
> Most other time series software doesn't do this.