On Sat, 16 Oct 2010, Allin Cottrell wrote:
> On Fri, 15 Oct 2010, Henrique Andrade wrote:
>
> > Let me explain better: I would like to make the forecasts for the
> > 1972:1-1991:4 period. For 1972:1-1991:1 I would like to use the static
> > option (because I have the data), and for 1991:2-1991:4 I would like
> > to use the dynamic option...
>
> As you say, it looks as if just doing
>
> fcast 1972:1 1991:4
>
> should produce the same results as first doing an explicitly
> static forecast over the estimation range, then calling fcast
> again, with the --out-of-sample flag. I'll look into this.
Henrique saw a difference between the out of sample VAR forecast
values (a) when using the --out-of-sample option with the "fcast"
command and (b) when giving fcast a range of dates which included
both in-sample and out-of-sample periods.
This was due to an ambiguity over exactly when, in the case of a
dynamic forecast, we should start using prior forecast values for
the regressors as opposed to actual data values -- this point
ended up being different in cases (a) and (b) above.
This ambiguity is now resolved. If a dynamic forecast (or the
dynamic portion of a longer forecast) starts in, say, 1991:2 what
we do when computing the forecast for t is this: for each lag, k,
of the right-hand side endogenous variables, check whether t - k
is before 1991:2. If so, use the actual data; if not, use a
previously computed forecast value instead.
This is equivalent to setting all values of the right-hand side
endogenous variables to NA from the starting period of the dynamic
forecast onward (if they're not already NA), and, in forming
forecasts, using actual data whenever possible and prior forecasts
otherwise.