This study examines the performance of international asset allocation strategies based on predictions of stock market returns and foreign exchange rates. Specifically, we predict returns for the Korea Composite Stock Price Index (KOSPI) and Standard & Poor's (S&P)500 index, and for the KRW/USD exchange rate over the period from January 2004 to July 2019 using a machine learning method, and build portfolios that consist of assets delivering those returns. For the predictions, we collect various economic variables and employ a machine learning algorithm known as the light gradient boosting method (LGBM). The out-of-sample results reveal significantly better performance for portfolios based on LGBM, which implies that LGBM has superior predictive power regarding stock market indices and exchange rates. Based on the LGBM's forecasts, we provide supportive evidence for the portfolio diversification effect from international asset allocations for Korean investors. Moreover, incorporating exchange rate predictions leads to better performance in constructing international portfolios.​

JEL Classification: C55, F31, F37, G11, G15
Keywords: international asset allocation, return prediction, exchange rate, portfolio diversification, machine learning​