On Tue, 19 Feb 2019, Artur T. wrote:
Am 17.02.19 um 15:48 schrieb Allin Cottrell:
> On Sat, 16 Feb 2019, Artur T. wrote:
>
>> I am trying to do some forecasting exercise using a panel data set. I am a
>> bit puzzled as the fcast command does not behave as I would have thought.
>> However, I am not sure I am using fcast correctly in this example -- the
>> manual doesn't include anything on the fcast command in connection with
>> panel data currently. Please, see the sample script below.
>
> True, the "fcast" facilities we have for time-series data are not available
> for panel data. That's no doubt something we should work on. In the
> meantime one can do an out-of-sample forecast manually. I show an example
> below, for pooled OLS, fixed effects and random effects. (But note that the
> latter is not consistent in this case.)
Thanks for your sample script on this, Allin -- very helpful!
Thanks, I'm glad to hear it was helpful.
Meanwhile, I've been working on making out-of-sample forecasting
more transparent in the case of panel data and I should have
something to share before long.
A brief comment for now. In the panel context "out of sample" could
mean one or both of two things: (a) out of sample in the unit or
group or individual dimension, and (b) out of sample in the time
dimension.
Case (a) is obviously going to be problematic other than in simple
cases such as pooled OLS. If the estimator involves "individual
effects", either fixed or random, we won't have estimates of these
out of sample. Case (b) is more promising, but if the regression
specification includes time dummies we have what looks like an
insurmountable problem there too.
Nonetheless, if we confine ourselves to case (b) and exclude
specifications that include time dummies, then with some clever
book-keeping we should be able to offer automatic out-of-sample
forecasts for at least pooled OLS, fixed effects and random effects
(plus, with a bit more work, arbond/dpanel). So that's what is in
the works right now.
Allin