Wednesday, October 23, 2019 | 2:00pm-3:30pm | Room 337, HSBC Business School Building

Abstract

We apply machine learning techniques and use stock characteristics to predict the cross-section of stock returns in 33 international markets, where data availability is lower than the U.S. Using only 12 variables (based on past returns, size, volume, and accounting information) as inputs, in most markets we are able to generate out-of-sample R2 and Sharpe ratios that are comparable to those in previous studies predicting U.S. stock returns with hundreds of stock characteristics and macroeconomic variables. These strategies are protable even in countries where size, book-to-market ratio, and momentum are shown to be weak linear predictors. Neural network models, which allow for complex interactions among the predictors, produce the best results among various machine learning methods. Finally, we find that different markets generally share a similar return structure: the strategies remain protable if we train our models with past U.S. data and run them on international stocks.