I have done a hearty search and so far and I have not been able to find
the answer to my question.
For a model with auto-correlated errors, when you run GLS or ARIMA and
fit() or predict(), you get the unconditional fitted values, as opposed to
the conditional or dynamic values. Or in other words, the correlation
structure of the error is omitted. Most stat packages provide this as far
as I know (eViews, or SAS's AUTOREG procedure). And so far with R, I have
been doing the correction manually, but I would love to know if this is
built in somewhere.
C.F. asked a similar question about 5 years ago under the title "problems
with the fitted values of the function gls" and it went unanswered. I hope
some new light can be shed here.
