Out-of-sample gravity predictions and trade policy counterfactuals

Gravity equations are often used to evaluate counterfactual trade policy scenarios, such as the effect of regional trade agreements on trade flows. In this paper, we argue that the suitability of gravity equations for this purpose crucially depends on their out-of-sample predictive power. We propose a methodology that compares different versions of the gravity equation, both among themselves and with machine learning-based forecast methods such as random forests and neural networks. We find that the 3-way gravity model is difficult to beat in terms of out-of-sample average predictive performance, especially if a flexible specification is used. This result further justifies its place as the predominant tool for applied trade policy analysis. However, when the goal is to predict individual bilateral trade flows, the 3-way model can be outperformed by an ensemble machine learning method.