Bayesian predictions of low count time series

The application of traditional forecasting methods to discrete count data yields forecasts that are non-coherent. That is, such methods produce non-integer point and interval predictions, which violate the restrictions on the sample space of the integer variable. This paper presents a Bayesian methodology for producing coherent forecasts of low count time series. The forecasts are based on estimates of the p-step ahead predictive mass functions for a family of distributions nested in the integer-valued first-order autoregressive (INAR(1)) class. The predictive mass functions are constructed from convolutions of the unobserved components of the model, with uncertainty associated with both parameter values and model specification fully incorporated. The methodology is used to analyse Canadian wage loss claims data.