Forecasting refugee migration with high-dimensional covariate space

Abstract Forecasting refugee migration is challenging, exacerbated by the high dimensional and dynamic nature of its drivers, such as climatic, economic, and political stressors. This article introduces a novel forecasting framework based on the Dynamic Elastic Net (DynENet) algorithm, which incorporates a time-varying regularization and a new model selection criterion: the penalized deviance ratio (PDR). Unlike conventional metrics such as the deviance ratio (DR), which emphasize in-sample fit, PDR explicitly penalizes model complexity, enhancing generalization in high-dimensional covariate setting. We apply this framework to forecast asylum-seeker rates (ASR) from Somalia to EU member states, leveraging a comprehensive set of district-level predictors. Extensive validation demonstrates that PDR-tuned models consistently outperform DR-based benchmarks in out-of-sample accuracy, reducing average point prediction errors by 40% and improving interval forecasts by 79%. Furthermore, we demonstrate how the DynENet framework supports explanatory insights at multiple levels—origin district, destination, and temporal—revealing both persistent and transient nature of migration drivers. The proposed methodology not only advances forecasting accuracy under high-dimensional covariate conditions but also enhances the interpretability of complex and evolving migration systems.