Automatic Model Selection for Forecasting Large Sets of Count Time Series

• Activity Number: 502
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Business and Economic Statistics Section
• Author(s): Ta-Hsin Li*+
• Companies: IBM Research
• Keywords: Count data ; model selection ; time series ; forecasting ; autoregression ; big data

Abstract

Many business and industrial applications require automatic procedures to forecast large sets of count time series. These time series often have diverse characteristics that are difficult to classify and cannot be handled effectively by a single family of statistical models. To cover the diversity, a large pool of candidate models have to be considered. However, the shear size of the data makes it difficult to apply the standard Box-Jenkins procedure and build manually a unique model for each time series. A practical solution demands a fully automated procedure that selects the "best" model from the pool with the aim of producing the most accurate forecast for each time series. In this paper, we present a case study which compares the performance of different criteria for model selection from a pool of predictive models that include the standard linear autoregressive model, the ordinary Poisson and negative binomial autoregressive models, and the zero-inflated Poisson and negative binomial models. The study is based on simulation as well as a real data set of demand for software development skills.