A risk-constrained distributional reinforcement learning portfolio strategy integrating chain-of-thought logical priors and time-varying topological structures

Time series of financial asset returns typically exhibit pronounced non-ergodicity and heavy-tailed, spiky distributions. Traditional meanvariance models and expectation-based Deep Reinforcement Learning (DRL) approaches struggle to effectively capture distributional shifts induced by exogenous logical shocks. To address this challenge, this paper proposes a risk-constrained distributional reinforcement learning framework that integrates large language modelbased logical reasoning with dynamic graph dependencies (LLM-G-DRL). At the perception level, rather than relying on conventional sentiment polarity classification, this study introduces large language models that support Chain-of-Thought (CoT) reasoning (e.g., DeepSeek-R1) to construct a Bayesian logical belief updating mechanism. This mechanism maps unstructured financial texts into high-dimensional latent logical states embedding causal transmission paths, thereby correcting predictive biases arising from exclusive dependence on historical price and volume data. At the structural level, to characterize the nonlinear contagion of systemic risk, the framework abandons the assumption of static adjacency matrices and employs dynamic graph attention networks (Dynamic GAT) to reconstruct time-varying topological dependencies among assets, enabling explicit modeling of risk propagation channels. At the decision-making level, the portfolio optimization problem is formulated as a Constrained Markov Decision Process (CMDP). Implicit Quantile Networks (IQN) are adopted to approximate the full probability distribution function while preserving higher-order moment information. Furthermore, based on Lagrangian duality theory, a hard constraint on Conditional Value at Risk (CVaR) is introduced, transforming tail-risk control into a dynamic penalty mechanism governed by dual variables. Theoretical analysis demonstrates that, through a closed-loop design combining logical priors, structural contagion, and distributional decision-making, the proposed framework exhibits stronger mathematical robustness than traditional point-estimation models. It effectively identifies and avoids tail losses under extreme market conditions, offering a statistically interpretable new paradigm for intelligent asset allocation in non-stationary markets.