Large-Scale Forecasting of Hierarchically Structured Data

• Activity Number: 21
• Type: Topic Contributed
• Date/Time: Sunday, August 1, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Business and Economic Statistics Section
• Author(s): Jonathan R.M. Hosking*+
• Companies: IBM T.J. Watson Research Center

Abstract

Enterprise-level business data are commonly stored in a "cube" that has several dimensions (e.g. product, geographic area, customer class), each of which can have a hierarchical structure (e.g. a product hierarchy could be line-of-business/brand/product-type/product). Each combination of dimensions (a "cell" of the cube) contains a time series or observations (e.g. monthly sales). Cubes can contain hundreds of thousands of cells and millions of observations. A business requirement is to make forecasts of future observations. I describe the software implementation of such a forecasting procedure, focusing on statistical issues including the choice of forecasting methods in different cells and the use of statistical models to achieve, as nearly as possible, conflicting aims of consistency and proportionality of point forecasts across different levels of combinations of hierarchies.