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

Sammanfattning

This thesis investigates whether U.S. financial variables contain informationabout one-month-ahead excess stock returns in eight developed foreign equity markets, and whether nonlinear machine learning methods improve forecast accuracyrelative 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 generatedrecursively using an expanding window and evaluated against each country’s ownhistorical 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 extendedsample, suggesting that the predictability documented by Rapach et al. (2013) doesnot extend to more recent decades. Univariate regressions on the thirteen GoyalWelch predictors confirm that no individual U.S. variable reliably beats the historical average across the eight markets. Multivariate OLS performs clearly worsethan the historical average, and performance deteriorates further as more predictors are added, consistent with estimation noise dominating any genuine predictivecontent. Ridge regression improves substantially over unrestricted OLS, but thecross-validation procedure consistently selects the largest available penalty, indicating that the data favor forecasts close to the unconditional mean. The nonlinearmodels do not reverse this conclusion. Random forest produces moderately negativeresults, gradient boosting performs substantially worse, and the neural network produces the least negative results of the three nonlinear models yet still fails to beatthe historical average. A pooled extension that stacks observations across countriesyields modest improvements in some linear specifications but does not improve thenonlinear models.The overall evidence points toward a weak and unstable predictive signal ratherthan an overly restrictive functional form. The results suggest that machine learningmethods do not automatically improve forecast performance when the underlyingsignal is weak, the sample is limited relative to model complexity, and the forecastingrelationship requires information to transmit from U.S. markets to foreign equitymarkets with a lag