Sam,

The rolling calculations are powerful when the underpinning patterns shift over time. In these cases the good old time series models go stale very fast. One strategy is to refit them every so often (once a week or once a day) which is kind of a rolling solution. I find that calculating the explanatory variables, their correlations, and autocorrelations on a rolling basis results in more power and interoperability. The last client I worked for had The Enterprise Miner, ETS and more. We through the works at the data and got very poor performance on the test data. By the time we applied the models the rules of the games had changed. It was a classic case of over fitting the training with lousy predictive value. The break through was when we just used Base, Stat & Graph to create a system that was based on rolling calculations fed into good old-fashioned statistical modelling tools (well I did have a good time using proc NLMIXED as it was not as straight forward as I pretend). Since the Licence is much cheaper if you concentrate on sas Analytics Pro, I do not have proc Expand. The data step and the procedures in sas/stat do very nicely for me thank you very much. But the bottom line is that one must fit the tool for the job.

With regard,

Eli