This message has been cross posted to the following eGroups: Young Professionals Group and Statistical Consulting Section .
-------------------------------------------
Hi all,
I'm looking to use/develop a sound method of time-series forecasting in which not only the frequency of events, but also the magnitude of the events are accounted for. For example, workload performed by an agent which comes from a queue I don't believe should be merely based on frequency (volume of events). Rather, the duration of time it takes to complete an event in the queue will vary across all events. So in sum, I am defining workload as a composite of frequency AND magnitude.
The original scope of my strategy was to build an aggregate GLM model which incorporates both frequency and magnitude to determine what factor(s) predict 'workload'. I thought a Tweedie model would suffice for this, allowing us to look at frequency and magnitude simultaneously. My thought was then to take the combined model outputs and place into a time-series forecast. I cannot determine whether the output of the model would be interpretable. Would we take the time-series of each variable which individually contributed to the best model to input into a multivariate forecast or can we by chance take the time-series datapoints from the aggregate model, as sort of a univariate forecast (i.e. a derived 'composite' variable of 'workload'), to which we might add some macroeconomic variables to adjust for external effects.
I also thought about something a bit simpler, but I'm stumped on how I would implement this. I was thinking that I could incorporate magnitude in some other way. Measures of magnitude do not, in any case that I know of, have a time domain. So I thought about the basic pdf for a moment and thought, "why use a probability number, which really isn't that intuitive anyway?". Instead, why not look at the probability of the magnitude of a single event as a 1 in N day/week/month/yr event? In this case I 'think' (read my vagueness here) I could look at frequency and magnitude simultaneously as a single data point - but I can't figure out how this translates into a time-series for forecasting. Ach vell.... any thoughts? Mine are quite jumbled at this point.
-------------------------------------------
Phillip Middleton
Analyst / Student
Rackspace / University of Texas At San Antonio
-------------------------------------------