Empirical comparison of parametric and nonparametric trade gravity models

A systematic comparison is made of parametric (i.e., ordinary leastsquares regressions and related generalizations) and nonparametric (i.e., kernel regressions and regression trees) log-linear gravity models for reproducing international trade. Experiments were conducted to estimate a log-linear gravity model reproducing import and export trade fows in quantity between Italy and 13 world economic zones, based on a panel estimation data set. The best parametric regression model was estimated to defne a baseline reference model. Some spec-ifcations of nonparametric models, belonging to the categories of kernel regressions and regression trees, were also estimated. The performance of parametric and nonparametric models is contrasted through a comparison of goodness-of-ft measures (R2, mean absolute percentage error) both in estimation and in hold-out sample validation. To assess the differences in model elasticity and forecasts, both parametric and nonparametric models are applied to future scenarios and the corresponding results compared.