Christoph Drobner

Christoph Drobner compares the predictive performance of four model types—gravity equations, random forests, neural networks, and graph neural networks—on geospatial flows across three datasets: international trade, U.S. interstate mobility, and intra-state human mobility. Machine learning models only slightly outperform traditional gravity models, with most of the predictive power coming from capturing cross-sectional patterns rather than changes over time. His findings highlight that while complex models offer marginal gains, the gravity model remains a strong baseline, offering valuable insights for policy and future research.