Graph Neural Networks for Multi-Asset Portfolio Optimization: Dynamic Correlations and Cost-Aware Regularization

Roshane Shahbaz

Keywords: Portfolio optimization, graph neural networks, transaction costs, multi-asset strategies, deep reinforcement learning.

Abstract: Portfolio optimization confronts two persistent challenges: modeling dynamic asset correlations during market shocks and mitigating transaction costs that erode 10-60% of profits. This systematic review synthesizes 42 studies (2018-2025) evaluating deep learning solutions, including Graph Neural Networks (GNNs) that capture non-linear dependencies and reinforcement learning frameworks incorporating cost penalties. Empirical evidence demonstrates GNNs achieve 15-30% higher Sharpe ratios than traditional methods by dynamically weighting asset relationships, while regularization techniques reduce turnover by 20-40% without compromising returns. Hybrid architectures (e.g., GNN-LSTM combinations) further enhance adaptability across multi-asset portfolios including ETFs, futures, and cryptocurrencies. Despite these advances, critical gaps persist in real-time deployment where sub-50ms execution remains challenging, and interpretability for regulatory compliance. We propose lightweight graph architectures via neural pruning and explainable AI through attention heatmaps as essential solutions. These innovations bridge academic research and practical implementation, enabling robust portfolio management in volatile markets through improved correlation sensitivity and cost efficiency.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 4148