This is probably too general of a question, but are there base or SAS/STAT procedures that would allow me to do some basic time series analyses? We do not have ETS and I have a need to do work on some trending and seasonal data.
If there is a seriously cyclical nature to the data, then I would use/group like sample periods: every Saturday, every Sunday, every Monday, every Tuesday, etc. for example to see if there is a trend across all the Saturday's, Sunday's, Monday's, etc. This could be done across years by month or week or day as well.
You would need to play around with the data more, and do more visual inspection to personally recognize the patterns.
I don't often disagree with you Chuck, but I'm afraid this is one of those times.
Faced with a desire to predict call loads on an inbound Call Centre, I was analysing a number of years of data. The population could be divided into four groups, but not very evenly. An underlying group was gradually increasing, and would present a slow increase over time. The duration of the calls varied quite a lot.
A second group would present monthly in line with a processing cycle. Their call durations varied moderately. A third group would present with long duration calls on a weekly basis. A fourth group were largely unpredictable, but were of medium to long duration on the occasions when they presented.
For privacy reasons, no identifying data existed to show which calls belonged to which group, so they were inferred from the day in which the call presented.
Over the top of this data were a series of interventions. If the call centre had been an investment house, which it wasn't, then you might imagine that news of fund collapses or sub prime lending issues would produce a spike of calls from worried investors. These needed to be removed from the model since they were unpredictable and added unwanted variance to the models.
Couple these issues with corrections for heteroscedasticity, and you can see that the model becomes far too complex for a simple regression. We used ETS, and the excellent SAS Press book co-authored by Dr John Brocklebank "SAS for forecasting time series" to define our approach.
In the end, we still could not get a robust model for prediction, although knowing the underlying cycles in the business processes did allow us to identify expected work loads in the week ahead. This assisted with resourcing, but took a lot of administrative magic by the Call Centre team leaders to effectively implement.
A colleague well versed in ETS from west coast US suggested that a large sprinkling of magic incantations was often needed to make it all work!!
I don't disagree with you in the least. When real time series analysis is the right way to go, then it is the correct approach -- "the right tool for the job". I interpreted the question to be "can we get something somewhat meaningful against time dependent seasonal data without explicitly using ETS?" and I offered the best advise that I could give to try to help. I spent years doing forecasting without ETS, and now that I have ETS, having proc forecast and proc arima are godsends.
Message was edited by: Chuck