I am required to forecast delinquency counts at the business day level. Each day, I need to produce a forecast of delinquency counts 1-29, 30-59, etc Days Past Due (DPD). The structure of the deliquency varies significantly for each bucket. 1-29 DPD is an exponential decline over the month. 30-59 DPD is only mildly exponential, 150DPD is fairly flat.
Each day the forecast of the daily counts needs to be produced for the rest of the month. I would like to include a lag of the dependent variable, possibly AR and MA terms, so the forecast must be dynamic, and eligible for automation.
In looking at the SAS literature, and previous SUGI papers, I first tried Proc Forecast, but it required daily continuous data, at least as I understood it. I have tried Proc Loess and Proc Model to forecast the series individually, and I am aware of Proc UCM to forecast the vector of counts by buckets. I also found the code to construct business day as a custom interval.
I only have 13 months of data, so I need something straightforward. Can any ETS forecasting procedure use the custom interval of business day? (eg 1 to 20,21) I would think that would be important in construction any lag, AR, or MA terms. Should I be using something more straightforward in base SAS?
Thank you for the time you take in considering my question.
My lecture notes from the SAS forecasting class include the following example:
"Hourly call center data - business hours are 7:00am to 8:00pm for a 13-hour day. The seasonal period of the data is 13." They go on to say, "The Call Center time series is not equally spaced in the conventional sense, but you can treat it as equally spaced by ignoring all time points for which the Call Center is closed."
I think this is similar to your case where non-business days are not of interest. You might have a seasonal component in day of the week (lag 7), but holidays (or other business closures) would throw a twist into that.
Proc ARIMA (I'm not sure about the others) doesn't even look at your date variable. It just assumes that your data is equally spaced. The forecast output would have the wrong dates in it, but you can change the dates in the output dataset and the resulting forecasts would still be appropriate.