Alpha Go Everywhere: Machine Learning and International Stock Returns

We apply machine learning techniques and use stock characteristics to predict the cross-section of stock returns in 33 international markets. We conduct a stringent out-of-sample test to examine concerns about overfitting: the models are trained with past U.S. data and used to predict international stock returns. Neural network models outperform linear models in terms of both predicting returns and generating profits. We achieve even stronger results when the models are trained separately for each market, allowing for market-specific return-characteristic relationships. However, regression tree models are prone to overfitting and underperform linear models when the number of observations is low.