Forecasting sales with micro-panels: Empirical Bayes approach. Evidence from consumer goods sector

Most of the existing non-Bayesian studies on dynamic panel models mainly focus on consistent estimation of the common auto-regressive parameters, while treating the heterogeneous intercept as a nuisance and removing it from the model, e.g. by first-differentiating. Bayesian literature proposes methods to obtain posterior distributions of the unit-specific intercept based on Markov Chain Monte Carlo sampling, which can be slow for large data sets. This paper investigates an empirical-Bayes procedure for obtaining the individual-specific intercept parameters without resorting to MCMC sampling. A set of simulation studies and empirical applications show that the proposed model outperforms benchmark models in some settings, especially for long-run forecasting of highly persistent data. Additionally, this paper proposes a bootstrap-based method for computing prediction intervals in the analysed modelling framework.