Modelos bayesianos zero-modificados para séries temporais de contagem

Abstract

This work presents two Bayesian zero-modified (ZM) models for count time series: zero-modified Poisson ARMA and zero-modified COM-Poisson ARMA. The latter allows a greater flexibility since it has an aditional parameter which accomodates greater subdispersion or overdispersion in comparison with the ZM Poisson ARMA model. The models are applied to simulated data and two real data sets. Both ZM Poisson ARMA and ZM COM-Poisson ARMA performed very well in zero-modified data. The goodness of fit was studied using posterior predictive checks. Model comparison was done using the deviance information criterion (DIC). Finally, a forecast study of six-steps-ahead was performed. In general, the ZM COM-Poisson model, although having an aditional parameter in comparison with the ZM Poisson ARMA model, showed DIC values similar to the DIC values of the ZM Poisson ARMA model. Since the ZM COM-Poisson ARMA model has the ZM Poisson ARMA model as a particular case, having the advantage of being more flexible, the ZM COM-Poisson ARMA model is proposed as an alternative to zero-modified count data.