Can Machine Learning Beat Gravity in Flow Prediction?

Understanding geospatial flows, such as the movement of goods or people between locations, is critical for a wide range of policy questions. Various formulations of the gravity equation have been commonly used to model these flows. But can this equation predict future geospatial flows with high accuracy, and how do more complex machine learning models stack up against it? This chapter evaluates the out-of-sample predictive accuracy of four classes of models—standard gravity equations, random forests, neural networks, and graph neural networks—across three distinct data sets: international trade, inter-state mobility in the U.S., and intra-state human mobility. By most metrics, machine learning models only marginally outperform the gravity equation. The high explanatory power achieved by all models is primarily due to their ability to explain cross-sectional variation rather than time-series changes. Our findings provide nuanced insights into the strengths and weaknesses of different modelling approaches for geospatial flows, informing future research and policy considerations.