Optimal forecasting with heterogeneous panels: a Monte Carlo study

This paper reports the results of a series of Monte Carlo exercises to contrast the forecasting performance of several panel data estimators, divided into three main groups (homogeneous, heterogeneous and shrinkage/Bayesian). The comparison is done using different levels of heterogeneity, alternative panel structures in terms of T and N and using various error dynamics specifications. We also consider the presence of various degrees of cross sectional dependence among units. To assess the predictive performance, we use traditional measures of forecast accuracy (Theilís U statistics, RMSE and MAE), the Diebold and Marianoís (1995) test, and the Pesaran and Timmermanís (1992) statistics on the capability of forecasting turning points. The main finding of our analysis is that in presence of heterogeneous panels the Bayesian procedures have systematically the best predictive power independently of the model’s features.