We apply machine learning techniques to predict international stock returns using firm characteristics. Market-specific training is important, as neural network models (NNs) achieve stronger results when they are trained in each market separately than in a global model trained with U.S. data. NNs outperform linear models in predicting stock return rankings and forming profitable portfolios. In contrast, regression trees underperform linear models when the number of observations is low. We also show that adding variables constructed from U.S. firm characteristics, which may contain information beyond the characteristics of international stocks, further enhances the return predictability of market-specific NNs

An extensive literature studies interactions of stock market anomalies using double-sorted portfolios. But given hundreds of known candidate anomalies, examining selected interactions is subject to a data mining critique. In this paper, we conduct a comprehensive analysis of all possible double-sorted portfolios constructed from 102 underlying anomalies. We find hundreds of statistically significant anomaly interactions, even after accounting for multiple hypothesis testing. An out-of-sample trading strategy that invests in the top backward-looking double-sort strategy generates equal-weighted (value-weighted) monthly average returns of 4% (2.7%) at an annualized Sharpe ratio of 2 (1.38), on par with state-of-the-art anomaly-based machine learning strategies