I've found a few interesting articles online on this topic, but none which appear to be too cut and dry.

My question is coming up with an accurate predictive forecast based on forecasting individual component parts, then adding then up (or whatever function they comprise) ... vs. forecasting at an aggregate level (which will miss some trends but may have less variance).

The example I'm working on: I'm trying to forecast phone and email contacts. The source of these contacts is customers ordering software products, and then have technical or payment issues, or what-have-you.

I could, at the most bottom-up level, have a forecast for the contacts for each disparate software group (Software A has much different contact patterns than Software B).

I could also split up forecasts between orders placed and contact rate (we may see contacts flat, but orders are increasing, and contact rate is decreasing).

These all really point to the same question – I have all the granular data – is forecasting more accurate at the aggregate level (what are my total contacts this month?) or would forecasting the disaggregated parts, then adding them be better?

  • 2
    $\begingroup$ A combination of both works better. See otexts.org/fpp/9/4 $\endgroup$ – Rob Hyndman May 21 '14 at 1:12
  • $\begingroup$ That makes sense -- still, there are so many different ways to 'split' the data --- by each software group, emails vs. phone calls, by country of contact origin (different trends), splitting contacts into orders and contact rate (both a sales boon, or a product flub/ complicaton, could lead to spikes). I will definitely begin combining multiple models --- feel like I have a lot to learn though. Thanks! $\endgroup$ – user45867 May 22 '14 at 16:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.