Graph neural networks for deep portfolio optimization

Abstract

There is extensive literature dating back to the Markowitz model on portfolio optimization. Recently, with the introduction of deep models in finance, there has been a shift in the trend of portfolio optimization toward data-driven models, departing from the traditional model-based approaches. However, deep portfolio models often encounter issues due to the non-stationary nature of data, giving unstable results. To address this issue, we advocate the utilization of graph neural networks to incorporate graphical knowledge and enhance model stability, thereby improving results in comparison with state-of-the-art recurrent architectures. Moreover, we conduct an analysis of the algorithmic risk-return trade-off for deep portfolio optimization models, offering insights into risk for fully data-driven models. © 2023, The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature.