LSTM-based Portfolio Optimization Strategy for SP500

Abstract

Portfolio optimization is a perennial topic in the field of finance and recent breakthrough in deep learning techniques offers a new perspective to tackle it. This study selects 30 stocks of SP500 in different sectors through various constraints and deploys Long Short-Term Memory and Ledoit-Wolf Shrinkage to estimate returns and covariance respectively. The target portfolio is then obtained by inputting the predicted results into the mean-variance model, which is dynamically updated on a daily basis given evolving market information. The results show that the target model this study proposed surpasses the market benchmark (SP500), 1/N portfolio, and other mean-variance variants in terms of numerous financial metrics. Moreover, the target model exhibits volatility invariance and the capability to mitigate risk while extracting returns. This study showcases the revolutionary and promising applications of deep learning in the financial industry, shedding light on novel portfolio allocation strategies for risk-averse investors seeking stable positive returns in turbulent markets.