*
* Example 8/2/6 from pp 415-418
*
open data housing.dat
calendar(m) 1983
data(format=free,org=columns) 1983:1 1989:10 housing construct rates
*
* AR1 is a more efficient instruction for doing maximum likelihood estimation of
* a regression with AR1 errors.
*
ar1(method=maxl,define=firsteq) housing * 1988:10
# constant construct{0 to 6} rates{0 to 9}
*
* Check residuals
*
@regcorrs
*
boxjenk(ar=1,constant,inputs=2,define=transfer,maxl) housing * 1988:10
# construct 5 0 0
# rates 3 0 5
*
uforecast(equation=firsteq,static) firstfore 1988:11 1989:10
uforecast(equation=transfer,static) transfore 1988:11 1989:10
@uforeerrors(title="Step 1 Model") housing firstfore
@uforeerrors(title="Transfer Function Model") housing transfore
*
* Simpler models cited in text
*
boxjenk(ar=2,sar=1,constant,maxl,define=arimaeq) housing * 1988:10
uforecast(equation=arimaeq,static) arimafore 1988:11 1989:10
@uforeerrors(title="SARIMA model") housing arimafore
*
* In order to get comparable results for exponential smoothing, we have to use
* two ESMOOTH instructions; one to estimate the parameters over the estimation
* range used for the other models, and another to get the one-step forecasts.
* Note that the forecasts that ESMOOTH produces directly are multiple step
* forecasts, not the series of one-step forecasts needed for this. Instead, we
* need the in-sample fitted values.
*
esmooth(seasonal=additive,estimate,constrain) housing * 1988:10
esmooth(seasonal=additive,alpha=%esalpha,gamma=%esgamma,fitted=esfore) housing * 1989:10
@uforeerrors(title="ES model") housing esfore 1988:11 1989:10