The More, the Better? Predicting Stock Returns with Local and Global Data

Abstract

We investigate the utility of local and global data in cross-sectional asset pricing. Using machine learning and three decades of stock data from 45 markets, we evaluate the performance of locally and globally trained models across country, regional, sector, and industry dimensions. We find limited added value in global over local data, as both yield comparable predictive performance. The gains from global training are modest, often insignificant, and mainly benefit smaller markets with high idiosyncratic risk. Stock return drivers are broadly consistent across markets, with global alignment in asset pricing increasing over time. Thus, local data is typically sufficient for accurate return forecasting, with incremental benefits of global datasets likely diminishing over time.

Keywords: asset pricing, return predictability, market integration, machine learning, cross-section of stock returns JEL Codes: F37, G12, G14, G15