Prediction and control for a time-series count data model: K. BrännäsInternational Journal of Forecasting11 (2), 263–270 (June 1995)

Abstract

Time series of count data are becoming more widely available. In a recently suggested class of models, the serial correlation between counts can conveniently be accounted for. In this paper, an easily calculated linear predictor is introduced. Control solutions for average count and for probabilities of specified events are given. An illustration based on a road accident frequency model for a Swedish county is included. Time series of count data are becoming more widely available. The number of bankrupt firms in a region, the number of passengers using a certain mode of transportation, etc. are exam- ples of such count data. In analysing count data, their specific nature should be recognized and taken into account in choosing between or in developing appropriate inferential techniques. In this paper we study prediction and control topics based on Zeger's (1988) extension of the Poisson regression model. This model extension allows for serial correlation between successive time- series observations, but distributionaUy it is known only up to the first- and second-order moments. The properties of estimators and test statistics for this model were recently studied by Brfinnfis and Johansson (1994). For policy studies, we may use prediction conditionally on purely exogeneous variables and on variables (instruments) that can be controlled by a policy-maker (or a group of policy-makers).