Model complexity and the performance of global versus regional models

We assess the predictive performance of global versus regional models in cross-sectional asset pricing using linear and machine learning algorithms. Based on data from 24 developed market countries, we find that the relative performance of globally versus locally trained models depends on algorithmic complexity. Regional models yield marginally higher long-short portfolio returns for linear methods, but spanning tests show no statistically significant alpha. In contrast, global models significantly outperform their local counterparts for more complex algorithms. These findings challenge earlier studies that rely on ex-post comparisons and linear methods, highlighting that global models outperform regional approaches when model complexity increases.