Answers are grounded in this paper and cite their sources.
Continuous portfolio optimization balances capital growth with risk exposure. Deep Reinforcement Learning (DRL) and Graph Neural Networks (GNNs) have been applied to this domain as they can represent the interconnected structure of financial markets. However, existing DRL frameworks struggle during non-stationary market crashes because they rely on static, backward-looking correlation graphs and reduce financial news to a single scalar sentiment score. They also lack downside risk controls to handle high volatility. In this paper, we propose a Graph Neural Network-Soft Actor–Critic (GNN-SAC) framework for continuous portfolio allocation that integrates structured events with a risk-aware GNN. A three-stage Natural Language Processing (NLP) pipeline extracts structured event tuples from raw headlines and instantiates them as transient nodes within a daily updated heterogeneous market graph. To manage downside exposure, the Soft Actor–Critic (SAC) architecture is extended with a dual-head Implicit Quantile Network (IQN) critic, Conditional Value-at-Risk (CVaR) curriculum learning, and momentum masking. We evaluate the framework through ablation studies, benchmark backtesting, and an out-of-sample stress test. The results show that the proposed framework achieves competitive risk-adjusted returns with Sharpe Ratio of 2.11 and a maximum drawdown of -5.60%. During the 2020 market crash, the agent contained tail-risk and reduced crash losses of -26.23%. Explainable Artificial Intelligence (XAI) analysis shows this resilience stems from the agent routing capital to stable defensive clusters. Integrating structured, entity-linked event embeddings into a risk-aware GNN produces an interpretable framework for balancing capital growth with explicit downside protection.
No takes yet. Share an insight, caveat, or question.
Pavithra et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: