*
* Example from section 6.2, through example 6.2.1
*
open data wine.dat
calendar(m) 1980
data(format=free,org=columns) 1980:1 1991:10 wine
*
graph
# wine
*
set lwine = log(wine)
*
* Contrary to the labeling in the book, the graphs use levels, not logs
*
* Do the classical decomposition, removing a trend and a period 12 seasonal.
*
@ClassicalDecomp(trend=linear,irreg=detwine) wine
graph(footer="Figure 6.11 Irregular Component of red wine data")
# detwine
*
* Apply seasonal differencing instead
*
diff(sdiffs=1) wine / sdwine
graph(footer="Figure 6.12 Seasonally Differenced data")
# sdwine
@bjident(method=yule,number=40) sdwine
*
diff(sdiffs=1) lwine / sdwine
diff(center) sdwine
@yulelags(table) sdwine
*
* The estimates of the AR(12) look somewhat different than the estimates in the
* book because RATS uses a tighter convergence criterion.
*
boxjenk(ar=12,maxl) sdwine
boxjenk(ar=||1,5,8,11,12||,maxl) sdwine
@bjautofit(pmax=3,qmax=15) sdwine
*
* The fully parameterized ARMA(1,12) model proves to be a bit hard to fit without
* improving the initial guess values. That's done with the options PMETHOD
* (Preliminary METHOD) and PITER (Preliminary ITERations). The simplex method is
* a derivative-free algorithm which tries to "ooze" up the likelihood surface.
*
boxjenk(ar=1,ma=12,maxl,pmethod=simplex,piters=10,method=bfgs) sdwine
*
* The subset model is much better behaved
*
boxjenk(ar=1,ma=||2,5,8,10,12||,maxl) sdwine