[1Yin-A] Poster 1

[1Yin-A-02] Expected Return Prediction of Country Equity Indices Using XGBoost and Its Application to Portfolio OptimizationAn Empirical Study of Machine-Learning-Based Return Forecasts with Mean–Variance and CVaR Optimization

〇Satoshi Kimura¹ (1. Mitsubishi UFJ Asset Management Co.,Ltd)

Keywords:

XGBoost,Portfolio Optimization

This study proposes a portfolio construction framework that integrates machine-learning-based expected return forecasts into classical optimization methods. Using monthly equity index data from major developed countries, one-month-ahead expected returns are estimated with XGBoost. For each country, an independent model is trained in a rolling-window setting using Fama–French style factors—market, value, profitability, investment, and dividend yield—and their lagged values. The predicted returns are incorporated into mean–variance and Conditional Value-at-Risk (CVaR) optimization frameworks. Their out-of-sample performance is compared with benchmark strategies that do not rely on return forecasts, including minimum variance, equal risk contribution, and equally weighted portfolios. Empirical results show that the portfolio combining XGBoost-based expected returns with mean–variance optimization achieves the highest annualized return and risk-adjusted performance. CVaR minimization effectively reduces drawdowns but leads to more conservative returns. Overall, the results indicate that non-linear expected return structures captured by machine learning models provide economically meaningful information for international portfolio allocation when directly embedded into return-sensitive optimization frameworks.