This poster's question regarding the Robinson test leads to another. As
far as I can tell, the various routines for estimating the fractional
root in time series using Stata (there are several - lomodrs, gphudak,
modlpr, roblpr) do not permit one to predict residuals after estimating
the root. It would be useful to be able to do so, so that the residuals
can be modeled. Has anyone written code to permit this?
It is quite true that none of these routines currently support prediction.
Perhaps the most sensible thing would be to incorporate predicted values,
from which residuals could be generated. I have a RATS routine on SSC that
will generate out-of-sample (ex ante) forecasts for a (0,d,0) model that
could readily be translated into Stata; but the nature of the fractionally
differenced model is such that it is hard to construct in-sample forecasts
unless you have a goodly amount of presample data. If you estimate d from a
sample of size T, I don't know how to generate T predicted values. If you
want to forecast that series for an additional tau observations (and, if
they are historical, generate realised forecast errors), no problem. Let me
know if this would be useful.
