*
* Forecasts from ARMA models for Canadian employment data
* pp 200-204
*
cal(q) 1961
open data caemp.dat
data(format=prn,org=columns) 1961:1 1994:4
*
* Models with MA's will often produce somewhat different estimates from different
* software because there isn't a single "correct" way to deal with the
* unobservable pre-sample lagged residuals. Because the MA(4) isn't really a good
* model for this data set (in fact, this exercise is designed to demonstrate
* that), the differences are more pronounced than you typically see.
*
boxjenk(ma=4,constant,maxl,define=ma4eq) caemp * 1993:4
ufore(equation=ma4eq,stderr=ma4std) ma4fore 1994:1 1994:4
set upper = ma4fore+ma4std*1.96
set lower = ma4fore-ma4std*1.96
set forezone * 1996:4 = t>=1994:1
graph(footer="Figure 8.1 Employment History and Forecast: MA(4) Model",shading=forezone) 4
# caemp 1990:1 1993:4
# ma4fore
# upper / 3
# lower / 3
*
ufore(equation=ma4eq,stderr=ma4std) ma4fore 1994:1 1996:4
set upper 1994:1 1996:12 = ma4fore+ma4std*1.96
set lower 1994:1 1996:12 = ma4fore-ma4std*1.96
graph(footer="Figure 8.2 Employment History and Long-Horizon Forecast: MA(4) Model",shading=forezone) 4
# caemp 1990:1 1993:4
# ma4fore
# upper / 3
# lower / 3