A0889
Title: A Bayesian approach to model bivariate count data and its application in forecasting soccer matches
Authors: Anirban Nath - Columbia University (United States) [presenting]
Soudeep Deb - Indian Institute of Management Bangalore (India)
Abstract: Discrete-valued time-series data can appear in a lot of social and political scenarios that include accidents, the size of the infected population, etc. A time-varying auto-regressive model is developed with covariate dependence for bivariate integer-valued time series data. The model is implemented in a Bayesian framework that leverages the concepts of Hamiltonian Monte Carlo techniques for estimation procedure. The method is applied to a soccer dataset, and the match outcomes are forecasted by modeling the number of goals of the two teams. The importance of forecasting predictive distribution instead of using point forecasts and the utility of proper scoring rules is stressed. The method is compared with two other existing methods for similar setups in terms of forecasting accuracy. Properties of bivariate Poisson and Skellam distribution are used extensively.
