Nonlinear International Stock Return Predictability from U.S. Variables: A Machine Learning Approach

Abstract

This thesis investigates whether U.S. financial variables contain information about one-month-ahead excess stock returns in eight developed foreign equity mar kets, and whether nonlinear machine learning methods improve forecast accuracy relative to standard linear models. The sample covers Australia, Canada, France, Germany, Italy, the Netherlands, Sweden, and the United Kingdom over 1982–2024, with an out-of-sample evaluation period of 1992–2024. All forecasts are generated recursively using an expanding window and evaluated against each country’s own historical average benchmark. The results are generally weak. A univariate regression based on lagged U.S. returns produces small and mostly negative out-of-sample values in the extended sample, suggesting that the predictability documented by Rapach et al. (2013) does not extend to more recent decades. Univariate regressions on the thirteen Goyal Welch predictors confirm that no individual U.S. variable reliably beats the his torical average across the eight markets. Multivariate OLS performs clearly worse than the historical average, and performance deteriorates further as more predic tors are added, consistent with estimation noise dominating any genuine predictive content. Ridge regression improves substantially over unrestricted OLS, but the cross-validation procedure consistently selects the largest available penalty, indicat ing that the data favor forecasts close to the unconditional mean. The nonlinear models do not reverse this conclusion. Random forest produces moderately negative results, gradient boosting performs substantially worse, and the neural network pro duces the least negative results of the three nonlinear models yet still fails to beat the historical average. A pooled extension that stacks observations across countries yields modest improvements in some linear specifications but does not improve the nonlinear models. The overall evidence points toward a weak and unstable predictive signal rather than an overly restrictive functional form. The results suggest that machine learning methods do not automatically improve forecast performance when the underlying signal is weak, the sample is limited relative to model complexity, and the forecasting relationship requires information to transmit from U.S. markets to foreign equity markets with a lag