GRAPH-AUGMENTED HYBRID PORTFOLIO RISK MANAGEMENT USING GRAPH NEURAL NETWORKS, HIERARCHICAL RISK PARITY, AND REINFORCEMENT LEARNING WITH XGBOOST-BASED CRASH ANTICIPATION

Abstract

The extreme volatility and susceptibility to abrupt crashes inherent in cryptocurrency markets present significant challenges to conventional risk management and portfolio optimization techniques. This paper proposes a novel hybrid machine learning framework designed to enhance resilience and optimize risk-adjusted returns for cryptocurrency portfolios. Recognizing the limitations of traditional models—which often fail to capture the complex, dynamic interdependencies and unique market microstructure characteristics (e.g., pronounced sentiment influence, regulatory uncertainty, security vulnerabilities, and manipulation risks) of digital assets. Our approach integrates multiple advanced methodologies. Specifically, Graph Neural Networks (GNNs) model complex inter-cryptocurrency relationships to uncover latent market structure. Hierarchical Risk Parity (HRP) utilizes this structural insight for robust, correlation-aware diversification. Reinforcement Learning (RL) dynamically optimizes asset allocation in response to real-time market shifts. Furthermore, XGBoost-generated crash signals provide an early-warning mechanism for proactive risk mitigation. Extensive evaluation demonstrates that the proposed GNN-RL hybrid framework significantly outperforms conventional HRP-based strategies, achieving a 25.3% reduction in annual volatility and minimized maximum drawdowns while maintaining competitive returns. Key improvements include superior adaptability across diverse market regimes, with the framework's advantages stemming from the GNN's relational analysis beyond simple correlation metrics and the RL agent's capacity for adaptive, performance-driven allocation. This work constitutes a significant advancement in cryptocurrency risk management, offering investors a powerful, AI-driven tool for navigating market uncertainty. It contributes to financial computing literature by demonstrating the efficacy of integrating structural analysis, optimized diversification, dynamic control, and crash prediction within a unified system for digital asset portfolios.