Online Program Home
My Program

Abstract Details

Activity Number: 478 - Scalable Bayesian Models for Time Series and Dynamic Networks: Making an Impact in Business and Socio-Economic Applications
Type: Invited
Date/Time: Wednesday, July 31, 2019 : 10:30 AM to 12:20 PM
Sponsor: Section on Bayesian Statistical Science
Abstract #300040 Presentation 1 Presentation 2 Presentation 3
Title: Bayesian Forecasting of High-Dimensional Count-Valued Time Series: Massive Data in Consumer Sales Forecasting
Author(s): Lindsay Berry* and Mike West and Paul Helman
Companies: Duke University and Duke University and 84.51°
Keywords: decouple/recouple; forecast assessment; multi-step forecast; supermarket sales forecasting; probabilistic forecast; state-space model

Interest in forecasting many time series of counts arises in numerous areas, including consumer sales contexts. With a focus on multi-step forecasting of daily sales of supermarket items, we have developed new classes of models based on the concept of decouple/recouple applied to multiple series that are each represented via novel univariate state-space models. The latter involve dynamic generalized linear models for binary and conditionally Poisson time series, random effects for over-dispersion, and covariates in both binary and non-zero count components. New multivariate models enable information sharing across series when aggregated data provide more incisive inferences on shared patterns such as trends and seasonality. These cross-series linkages insulate the parallel estimation of univariate models, hence scaling efficiently in the number of series. A case study on sales of related supermarket items showcases forecasting of multiple series, with discussion of forecast accuracy metrics and probabilistic forecast assessment. Examples demonstrate improved forecast accuracy in a range of metrics, and illustrate the benefits of full probabilistic models for forecasting.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2019 program