Machine Learning for Portfolio Optimization: Yet Another Approach

This work explores the integration of modern machine learning techniques with classical financial models for portfolio optimization. Building on the well- established Black-Litterman model, we incorporate predictions from hybrid LSTM-GRU neural networks to forecast stock prices and use these forecasts within the Black-Litterman framework to construct optimal portfolios. While using classical portfolio optimization techniques we also employ hierarchical risk parity (HRP) to enhance portfolio efficiency and diversification. Our approach aims to balance the robust theoretical foundations of traditional models with the adaptability and predictive power of machine learning. By comparing the performance of portfolios optimized through different methods we assess the effectiveness of these strategies under varying market conditions. We conducted extensive empirical analysis, using data from Yahoo Finance spanning from 2010 to 2020, encompassing daily values for S&P 500 stocks. The results indicate that portfolios constructed using the hybrid model outperform classical optimization methods and demonstrate resilience during periods of market volatility. Notably, the R2 (R-squared) measure proved more effective than MAPE (mean absolute percentage error) in this context, suggesting that it better suits the combined machine learning Black-Litterman framework. Our findings contribute to the ongoing dialogue on portfolio optimization, emphasizing the importance of integrating modern computational techniques with traditional financial theories to develop more reliable and effective investment strategies, offering a promising pathway for future studies in the field of finance