3

I have weekly forecasts. I would like to also obtain monthly forecasts. Should I forecast using monthly data or just average my weekly forecasts?

Alex
  • 2,011
  • I suggest using monthly data to forecast ... Even better if you have daily data that can be used to predict not only days , but weeks . – IrishStat Nov 14 '17 at 15:05
  • Are you asking if it's better to take the same weekly data you used for your weekly forecasts and use that to directly make a monthly forecast, or would you have additional data purely for months? That is, are you essentially asking is it better to aggregate your data and forecast or to forecast then aggregate your forecasts? (Or as IrishStat suggests, do you actually have daily data, etc?) – Wayne Nov 14 '17 at 15:08

1 Answers1

2

Neither one.

Instead, do both. The forecasts will not be sum consistent. Now consolidate them, using optimal reconciliation. This is the Multi Aggregation Prediction Algorithm (Kourentzes et al., 2014, IJF). See the MAPA package for R, and Kourentzes' blog.

This presupposes that you have grouped your weeks into months in a nested way, e.g., (4, 4, 5) weeks to a "month". If your weeks are not nested within months, you may need to adapt a few things, but it's doable.

Stephan Kolassa
  • 123,354
  • Do you have a source that shows that forecasts are not sum-consistent? – Alex Nov 17 '17 at 15:24
  • 1
    No. It's just the normal case in hierarchical forecasting: if you forecast low granularity data, then aggregate the forecasts, you will usually get different aggregate forecasts than if you first aggregated historical data, then forecasted aggregates directly. Try it! – Stephan Kolassa Nov 18 '17 at 17:14