LSTM Predictions Versus Historical Data in Portfolio Construction

This article explores the integration of machine learning-based predictions, specifically LSTM networks, into the field of portfolio construction. We address the practical challenges inherent in existing approaches and present a methodology customized for practical applicability in real-world scenarios. The primary objective of this research is to apply a portfolio construction strategy based on previous studies [9, 10], incorporating forecasted asset returns into the portfolio construction process, and adapt the methodology to make it applicable in practice. To achieve this goal, we selected a list of assets that includes the S&P 500, 10-year US Treasuries, gold, oil, wheat, and the US dollar. Our findings reveal the performance traits of various strategies for constructing portfolios, highlighting the relationships between returns, volatility, and downside risk. Leveraging on a LSTM network to predict future returns for each asset, we integrate these predictions into portfolio calculations with various strategies. Comparing predicted data-based and historical data-based portfolios reveals consistent tendencies in both types of approaches. In contrast with previous findings, our analysis leads to the conclusion that in most cases the portfolios based on historical returns outperform their counterparts based on predicted returns.