Gravity-informed machine learning improves prediction of global cereal virtual water trade

This dataset accompanies the manuscript "Machine learning versus gravity models for bilateral cereal virtual water trade under leakage-free validation" and contains the data and code needed to reproduce its results. It is built from the CWASI (Coping with Water Scarcity in a Globalized World) database and comprises a panel of all ordered pairs of 198 countries for 1993–2023 (1.21 million country pair-years, 16% non-zero). Variables cover bilateral cereal virtual water flows, trade costs, economic size, cereal production, water scarcity, and cropland-weighted drought and climate anomalies. The deposit includes the analysis-ready panel, derived variables, and the scripts implementing the PPML gravity models, Poisson XGBoost, gravity-informed hybrid and two-part models, evaluated under forward-chaining, country-blocked and random validation, together with the SHAP/ALE driver analyses and shock decompositions into tonnage and water per ton. The accompanying README documents file contents, variable definitions, data sources and reproduction steps.