Optimizing Predictions in Heterogeneous Panels with Pooling Averaging

This paper is an extension of the research done in Wang et al. (2017). For static linear panel models with heterogeneous coefficients across individuals, the predictive performance of pooling averaging methods is investigated. Latent group structure identification using Classifier-Lasso with Partial Profile Likelihood, as proposed by Su et al. (2016), is used to obtain pooling specifications to average over. In a Monte Carlo experiment, it is shown that Mallows pooling averaging combined with latent group structure identification using Classifier-Lasso has the best performance in terms of MSPE when a moderate or large degree heterogeneity is present across individuals. Especially when the amount of individuals in the panel is large, the performance of the newly proposed way to obtain pooling specification using Classifier-Lasso has clear advantages over the method proposed by Wang et al. (2017). When a low degree of heterogeneity is present, the Classifier-Lasso estimator as proposed by Su et al. (2016) has the best performance. For this estimator, its finite sample performance is also investigated. The newly proposed method is also used to predict changes in sovereign credit default swap spread for a cross-country panel as an illustration.