A Risk Management Agent with Uncertainty Quantification for Financial Trading

Financial markets exhibit non-stationarity and heavy-tailed return distributions with pronounced uncertainty, making adaptive risk management a critical challenge for algorithmic trading systems. While recent advances in machine learning and reinforcement learning have improved trading signal generation, risk control is still predominantly handled through static heuristics or tightly coupled objectives, limiting robustness under regime shifts and extreme market conditions. This paper proposes a modular, learning-based risk management agent designed to operate independently from trading decision logic. Market uncertainty is modeled through distribution-aware risk representations that combine volatility-based regime indicators with quantile-derived tail-risk estimates obtained via non-parametric quantile regression. A dedicated reinforcement learning policy is trained in a standalone risk environment to dynamically regulate exposure, position sizing, and protective parameters such as stop-loss and take-profit levels, without interfering with upstream trading decisions. A comprehensive ablation study is conducted across multiple asset classes, including equities, foreign exchange, and cryptocurrencies, to isolate the contribution of each risk component. Experimental results show that quantile-based tail-risk information is the dominant driver of performance and robustness, while regime indicators play a secondary stabilizing role. Furthermore, the integration of a fuzzy logic fusion layer yields benefits under extreme-risk conditions, particularly in cryptocurrency markets, where it significantly reduces drawdowns and trading turnover while improving risk-adjusted performance.

keywords: Risk-aware reinforcement learning, Uncertainty quantification, Trading risk management, Quantile methods, Multi-agent systems, Fuzzy rule-based inference system