A profitable currency portfolio strategy: Learning from connectedness

This study proposes a profitable currency portfolio strategy integrating dynamic connectedness into machine learning (ML) predictions. The portfolio is constructed using consensus predictions of return levels from LSTM and MLP and return directions from SVM and RF. Our findings reveal that connectedness slightly enhances returns but significantly reduces return volatility, implying its role in risk management. Compared to eight classical currency trading strategies, ML-based portfolios outperform in returns and mitigating extreme losses. Notably, portfolios incorporating RF predictions achieve the highest average return and Sharpe ratio among all strategies. Additionally, ML-based portfolios exhibit significant differences from classical strategies in determining currency positions.