A Comparative Analysis of Internal Migration Models: Traditional Gravity versus Constrained Machine Learning Approaches

Abstract

This paper analyzes internal migration patterns and explores how they can be modeled using explanatory features. The goal is to validly depict district-level migration flows using causal migration models based on district-level features, for the purpose of analyzing the impact of regional interventions. Thus, sole black-box approaches like time-series extrapolation are not regarded. The models are trained and evaluated for Austria using different time periods as training data and different time lags in features. The models are ranked by how accurately they match the data between 2016 and 2023. Finally, the three best, that is, the extended gravity model of migration, an Autogluon Ensemble model, and constrained gradient boosted trees, are applied to simulate hypothetical migration scenarios involving two selected rural districts in Austria. Compared to other work on the topic of internal migration modeling, we assess not only the model's quantitative accuracy by computing established error metrics on model predictions against the ground truth, but also qualitative behavior by examining the plausibility of simulated intervention scenarios. While the present study focuses on this regression-based comparative analysis, the resulting models are designed for integration into the agent-based population simulation framework GEPOC, where they will be used to sample individual migration decisions based on agent characteristics and regional features. This integration enables future work to investigate the long-term effects of environmental and policy changes on migration behavior in a fully dynamic simulation setting.