Bayesian nonparametric forecasting for INAR models

A nonparametric Bayesian method for producing coherent predictions of count time series with the nonnegative integer-valued autoregressive process is introduced. Predictions are based on estimates of h-step-ahead predictive mass functions, assuming a nonparametric distribution for the innovation process. That is, the distribution of errors are modeled by means of a Dirichlet process mixture of rounded Gaussians. This class of prior has large support on the space and probability mass functions and can generate almost any kind of count distribution, including over/under-dispersion and multimodality. An efficient Gibbs sampler is developed for posterior computation, and the method is used to analyze a dataset of visits to a web site. (C) 2015 Elsevier B.V. All rights reserved.