ARMA model rolling window dynamic forecast (STATA 12)

[Alexander Koll]

I use the BOX Jenkins methodology to forecast US quarterly GDP and I want dynamic forecasts (4 quarters ahead)
The regular predict command is fine but I would like to use a rolling and recursive window forecast combined with the dynamic forecast

Here are the commands that I have so far (first part is just the regular way the predict command is implemented)
install estatsto and freduse if not installed yet, commands provided


** group project "Topics in Econometrics Semester B 2014 HK"
**forecasting US real quarterly GDP
clear
set more off
**define working directory (use online Dropbox)

**install estatso and freduse program if not installed (delete the stars in the following line)
**ssc install estatsto
**ssc install freduse
**open relevant file
freduse GDPC96, clear
gen q=quarter(daten)
drop daten
gen time=tq(1947q1)+_n-1
format time %tq
gen t=_n
tsset t
gen lGDP=log(GDP)
**generate linear, quadratic and cubic trend variable (may not all be needed)
gen trend=_n
gen trend2=trend^2
**engage in forecast of the AR2(and for completeness for AR3 as well)
local estend 260
local fbegin 261
arima lGDP trend trend2 if tin(,`estend'), arima(2,0,0) vce(robust)
predict fcGDPAR2, dynamic(261)
predict fce, residual
**obtain FC error interval only for AR2 which is superior
predict fcGDPerr, mse dynamic(261)
gen fcGDPupper=fcGDPAR2+1.96*sqrt(fcGDPerr) if tin(`fbegin',)
gen fcGDPlower=fcGDPAR2-1.96*sqrt(fcGDPerr) if tin(`fbegin',)
list fcGDPupper lGDP fcGDPAR2 fcGDPlower if tin(`fbegin',)
tsline lGDP fcGDPAR2 fcGDPupper fcGDPlower in 230/268, legend(label(1 "lGDP") label(2 "dynamic FC") label(3 "upper bound") label(4 "lower bound")) saving(lGDPintFC1.png, asis replace)



************************************************** ************************************************** ************************************************** *****
** attempts for using rolling and recursive window (this command works, but is only 1 step ahead, I would like to get dynamic forecasts)
gen recursiveFC= .
gen rollingFC= .
forvalues i = 200(1)268 {
reg lGDP trend trend2 L(1/2).lGDP if t<`i', robust
predict p if t==`i'
replace recursiveFC = p if t==`i'
drop p
}

**for rolling window
forvalues i = 200(1)268 {
reg lGDP trend trend2 L(1/2).lGDP if t>=`i'-200 & t<`i', robust
predict p if t==`i'
replace rollingFC = p if t==`i'
drop p
}

tsline lGDP rollingFC recursiveFC in 230/268


I need help because I cannot figure out how to implement the dynamic forecast =predict fcGDPAR2, dynamic

The regular command needs a start date, but I have a time varying. Simply implementing `i' instead of a start date does not work.



Thanks for any advice. Help on how to include Interval forecast bands that grow over time (not just the 1 step ahead forecast error) would be nice as well.


Alex

[Hsin-Yun Kuan]

Dear Alexander,

I am currently exploring the possibility of conducting time series rolling forecasts, much like you. Have you managed to find a solution to this challenge?

If so, would it be possible for you to share some sample code? I would greatly appreciate your assistance.

Thank you for taking the time to read my message.

Warm regards,
Hsin-Yun