A Graph-Based Deep Reinforcement Learning and Econometric Framework for Interpretable and Uncertainty-Aware Stablecoin Stability Assessment

Abstract

The instability of algorithmic and hybrid stablecoins has become a systemic concern in decentralized finance. This paper proposes a unified, interpretable, and uncertainty-aware framework that integrates graph-based deep reinforcement learning, GARCH econometric modeling, and Bayesian inference. Multi-stage reinforcement learning agents simulate interactions between arbitrageurs and protocol mechanisms. GARCH models capture volatility dynamics, while Bayesian methods provide confidence intervals for peg deviation forecasts, enabling adaptive prediction and transparent risk interpretation. The framework is validated using over eight million on-chain and off-chain records across 120 scenarios involving USDT, USDC, and TerraUSD. It achieves 89 per cent crisis prediction accuracy and 83 per cent reflexivity modeling performance, significantly outperforming six benchmark models. Notably, the system issued early warnings up to 72 hours before the TerraUSD collapse. Ablation studies confirm the unique contribution of each module. In addition to technical improvements, the framework outputs a stability index and dynamic reserve recommendations to support policy response and supervisory planning. Compared to existing approaches, this is the first framework to combine dynamic simulation, interpretability, and probabilistic forecasting in a single architecture. It offers practical value for stablecoin monitoring and establishes a methodological foundation for future research in digital asset risk assessment.