st: Panel data, statsby, dynamic forecasting

[Wesley Burnett <burnettwesley@gmail.com>]

I'm attempting to run a series of arima regressions within each region
for a panel data set in long form. Following the arima regressions I'm
attempting to do dynamic predictions so that I can iteratively
forecast out of sample. Basically, the algorithm does a bunch of
individual time series regressions and predictions. I have written a
"foreach" loop with the "statsby" command to run the arima regressions
across each region. The looping algorithm works; i.e., it runs the
separate regressions and then performs the predictions, but the
predictions are not calculated out of sample--it only performs the
within sample predictions.

If I estimate the arima regression and prediction within a single
region (i.e., omitting the rest of the panel) then the code works fine
and the entire out-of-sample predictions are calculated. I think the
problem may be related to how the time series command is automated in
Stata. That is, if one uses a single region then the "tsset year"
command is invoked, where year denotes my time observations which are
annual. If one uses the "foreach" and "statsby" algorithm on a panel
data set then the "tsset regions year" command must be invoked;
otherwise, if you use the "tsset year" command on the panel then Stata
provides the error code: "repeated time values in sample."

I would appreciate any feedback on how to fix this problem. The code
is here, where the dynamic prediction is given a value of 2012 to tell
Stata when to start the out-of-sample predictions. The term "lprod"
designates a variable within my data set that is the log of production
examined across time.


*Invoke the time series command for region (fips) and time observation (year)
tsset fips year

*Add 24 years of annual time observations to the end of each region
for the out-of-sample predictions
tsappend, add(24)

*foreach algorithm
levelsof fips, local(fipsid)
foreach c of local fipsid {
local rr: label region `c'
quietly arima lprod if fips==`c', arima(1,1,0)
quietly predict p_prod_`c' if fips==`c', dyn(2012) y
quietly predict fev_`c' if fips==`c', mse
g upper_`c' = p_prod_`c' + 1.96*sqrt(fev_`c')
g lower_`c' = p_prod_`c' - 1.96*sqrt(fev_`c')
}
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/