This study explores a key question in international asset pricing: whether global or regional factor models are more effective in predicting stock returns. Moving beyond traditional methods that rely on retrospective comparisons and linear frameworks, the authors employ a forward-looking forecasting approach using a variety of machine learning models. These range from linear techniques like OLS and Elastic Net to non-linear algorithms such as Random Forest, Gradient-Boosted Regression Trees, and Neural Networks. Analyzing data from 24 developed markets between 1990 and 2021, the research examines how model complexity affects the advantages of global versus local training strategies.
The study finds that regional models produce slightly higher returns with simple linear methods, though they lack statistically significant alpha. In contrast, complex machine learning algorithms consistently perform better when trained on global data, both economically and statistically. This challenges previous findings by Griffin, Fama and French, and Hollstein, providing a nuanced view on financial market integration and stock return predictability.
A major strength of the paper is its systematic comparison of predictive performance across a wide range of machine learning models under both local and global training paradigms. The authors use an expanding window approach to ensure out-of-sample validity, carefully segmenting training, validation, and testing periods. The diverse model selection—from basic OLS to advanced ensemble neural networks—allows for a robust analysis of how complexity impacts the effectiveness of global data pooling.
The study’s empirical design further bolsters its credibility. By utilizing a balanced, quality-filtered dataset from MSCI-classified developed markets and implementing rigorous data preprocessing steps, the authors minimize common biases in financial machine learning research. Their consistent use of value-weighted, quintile-sorted long-short portfolios adheres to established norms in empirical asset pricing.
Despite the paper’s rigor, there are areas for improvement. The study could benefit from clearer justification of chosen model architectures and hyperparameters, a deeper discussion on economic implementability, and an analysis of emerging markets. Additionally, a focus on interpretability and model drivers, as well as performance across different market regimes, could enhance the study’s relevance.
In conclusion, this paper offers a robust and forward-looking study that bridges machine learning and empirical asset pricing. It demonstrates that globally trained models outperform regional ones when using advanced algorithms, providing new insights into the interplay between geographic scope and model complexity. Future research could extend the study’s impact by addressing economic frictions, exploring emerging markets, and improving model interpretability. Nonetheless, this work makes a significant contribution to both finance and applied machine learning by presenting a more sophisticated approach to international return forecasting.
