Forecasting Hierarchical Time Series
Professor Rob Hyndman, Department of Econometrics and Business Statistics; Monash University
10th Oct 2013 11:30 am - Room 498 Merewether Bldg H04
Hierarchical time series occur when there are multiple time series that are hierarchically organized and can be aggregated at several different levels in groups based on dimensions such as product, geography, or some other features. A common application occurs in manufacturing where forecasts of sales need to be made for a range of different products in different locations. The forecasts need to add up appropriately across the levels of the hierarchy.
Historically, forecasting of hierarchical time series has been done using either the "bottom-up" method, various "top-down" methods, or some combination of the two known as "middle-out" approaches.
I will describe a framework for studying such methods which leads naturally to an optimal combination approach based on a large ill-conditioned regression model.
While the model leads to optimal forecasts, the ill-conditioning and size of the model make computation difficult or impossible. I will describe a solution to this problem that make the forecasts fast to compute even for problems involving hundreds of thousands of time series.