Autonomous Risk Agents for Derivatives Risk Management

Abstract

The growing complexity of derivatives markets and the limitations of classical hedging strategies have motivated a search for more adaptive and robust risk management approaches. This study proposes an autonomous multi-agent framework for derivatives risk management tailored to the Iranian market. The framework decomposes the risk management problem into four specialized agents — Pricing, Hedging, Execution, and Risk Monitoring — that interact within a simulated market environment calibrated to Tehran Stock Exchange (TSE) and Iran Mercantile Exchange (IME) data. The framework is augmented with a unified explainability layer that integrates SHAP values, attention visualization, LIME, counterfactual explanations, and large language model (LLM)-generated rationales to produce human-readable justifications for agent decisions, enabling regulatory compliance with the transparency requirements of the Securities and Exchange Organization (SEO) and the Central Bank of Iran (CBI). The proposed framework is evaluated against classical Delta-Gamma hedging, single-agent deep reinforcement learning, and risk-sensitive reinforcement learning baselines. Experimental results demonstrate that the proposed framework achieves a Sharpe ratio of 1.24, representing a 47.6% improvement over the best single-agent baseline and a 195% improvement over Delta-Gamma hedging. The framework also reduces maximum drawdown to 9.8% (25.8% reduction), CVaR (95%) to 0.041 (36.9% reduction), hedging error to 0.029 (29.3% reduction), and transaction costs to 0.24% (22.6% reduction), while improving Greeks stability by 38.2%. Expert evaluation of the explainability layer reveals that LLM-generated rationales receive the highest rating (4.42/5), followed by SHAP values (4.17/5). The findings have important implications for regulators, brokerage firms, investment funds, and proprietary trading desks in Iran, and provide a foundation for future research on autonomous risk agents in emerging markets.