Forecasting global international student mobility with artificial intelligence

A confirmatory multi-horizon study

Abstract

Forecasting international student mobility is increasingly important for institutional planning, national education policy and equitable internationalisation, yet evidence remains limited on whether artificial intelligence improves upon simple temporal benchmarks. This study develops GlobalStudent-AI, a horizon-dependent framework for forecasting inbound international student shares using global mobility records and socioeconomic, demographic, technological and higher education indicators. The analysis comprised 9,987 exact calendar-aligned observations from 159 countries across one- to six-year forecast horizons. Models were selected using development data from 2021–2023 and evaluated on an untouched 2024 confirmation year. Persistence, drift, historical-mean and pooled linear baselines were compared with direct and persistence-residual machine-learning models using multiple error measures, country-clustered bootstrap confidence intervals, multiplicity-adjusted tests, feature ablation, robustness analysis and conformal uncertainty estimation. Machine learning did not reliably outperform persistence at one year, but gains increased with forecast length. At six years, the persistence-residual Extra Trees model reduced mean absolute error from 3.77 to 2.34, a 38.02% improvement that remained statistically significant after multiplicity correction and robust to exclusions of microstates, extreme changes and countries with limited histories. Several countries showed positive expected changes by 2027, although all conformal intervals crossed zero. Artificial intelligence was therefore most useful for medium- and long-term forecasting.