Global vs. Regional Forecasting in Asset Pricing: A Machine Learning Approach
Artificial Intelligence & Machine Learning
This study tackles a longstanding question in international asset pricing: are global or regional factor models more effective at predicting stock returns? Departing from prior research that relies heavily on ex-post comparisons and linear frameworks, the authors adopt an ex-ante forecasting methodology combined with a diverse suite of machine learning (ML) models. These include both linear techniques (OLS, Elastic Net) and non-linear algorithms (Random Forest, Gradient-Boosted Regression Trees, Neural Networks). Leveraging data from 24 developed markets from 1990 to 2021, the authors explore how model complexity influences the relative benefits of global versus local training strategies.
The key finding is that regional models yield slightly higher returns when using simple linear methods, although they fail to generate statistically significant alpha. In contrast, more complex ML algorithms consistently perform better when trained on global data, both economically and statistically. This result challenges earlier conclusions by Griffin (2002), Fama and French (2012, 2017), and Hollstein (2022), offering a nuanced perspective on financial market integration and the predictability of stock returns.
Strengths
A core strength of this paper lies in its systematic comparison of predictive performance across a wide range of ML models under both local and global training paradigms. The authors use an expanding window approach to maintain out-of-sample validity, carefully segmenting training, validation, and testing periods. The variety of models—from basic OLS to advanced ensemble neural networks—allows for robust analysis of how complexity impacts the effectiveness of global data pooling.
The empirical design further enhances the study’s credibility. By using a balanced, quality-filtered dataset from MSCI-classified developed markets—and implementing rigorous data preprocessing steps such as excluding microcaps and capping extreme market caps—the authors minimize common biases in financial ML research. Their consistent use of value-weighted, quintile-sorted long-short portfolios adheres to established norms in empirical asset pricing.
Notably, the spanning regressions provide additional insight by testing whether forecasts from one model encompass those of another. For instance, globally trained Gradient-Boosted Regression Trees and Neural Networks yield significant positive alpha even after controlling for local forecasts (e.g., Neural Networks: 0.20% per month, t-stat 3.45). Robustness checks across four global subregions further demonstrate that globally trained Neural Networks outperform regional models, including in markets typically seen as anomalies, such as Japan (Sharpe ratio: 0.68 global vs. 0.42 local).
Areas for Improvement
Despite the paper’s rigor and contribution, several areas could benefit from further elaboration:
- Model Selection and Hyperparameters
While the paper references prior literature (e.g., Gu et al., 2020) for hyperparameter settings, it would benefit from a clearer justification of the chosen ranges for model architectures, such as number of layers, neurons, and learning rates in Neural Networks. Including a sensitivity analysis or robustness check using alternative hyperparameter grids would address overfitting concerns. - Economic Relevance of Results
Although the study reports statistical significance in returns and Sharpe ratios, a deeper discussion of economic implementability—such as transaction costs, turnover rates, and real-world constraints—would make the findings more actionable. High turnover associated with ML-driven quintile sorting may reduce net returns, especially for Neural Networks. - Lack of Emerging Market Analysis
The study focuses exclusively on developed markets. Given prior divergent findings in emerging markets (e.g., Cakici, Fabozzi, and Tan, 2013), applying the same methodology to those markets would test the generalizability of the conclusions. While the authors acknowledge this gap, it is not empirically addressed. - Interpretability and Model Drivers
The analysis prioritizes predictive accuracy but overlooks interpretability. Identifying which firm characteristics drive return forecasts, particularly in Neural Networks, would enhance the study’s relevance. Tools like SHAP values or partial dependence plots could illuminate how input variables interact differently in global vs. regional contexts. - Performance Across Market Regimes
The dataset spans several major market events (e.g., 2008 financial crisis, COVID-19). Evaluating model performance across different regimes—crisis vs. recovery—could provide practical guidance for institutional investors considering the use of global ML models.
Conclusion
This paper presents a robust, forward-looking study that bridges machine learning and empirical asset pricing. It shows convincingly that globally trained models outperform regional ones when leveraging advanced algorithms, offering new insight into the interplay between geographic scope and model complexity.
Future research could expand the study’s impact by addressing economic frictions, extending analysis to emerging markets, and improving model interpretability. Nonetheless, the current work makes a significant contribution to both finance and applied machine learning by offering a more sophisticated approach to international return forecasting.
Written by Tamson Pham
Read the full paper:
Join us at:
