|
Abstract:
|
Over the past two decades count time series models have received significant attention by statisticians and researchers working with integer-valued data. Although these models have been adopted in many application areas, some practitioners remain either unfamiliar or unconvinced of their value. As a result, they may ignore the count nature in their data by employing classic ARIMA models, or they may discount the temporal dependence by using standard GLM methods. Both strategies are known to yield suboptimal inferences.
In this talk we aim to deal with this issue. Following a succinct literature review, we will switch our attention to a novel model with many desirable traits that existing count time series models fail to produce in tandem. Specifically, the proposed model allows for any prescribed marginal distribution, can produce both positive and negative correlations, permits effortless inclusion of covariates and finally can be fit through a computationally feasible and well performing inference procedure. We conclude with a review of available software implementations and with performance comparisons between the proposed and competitor models.
|
