Scholarlink

<p>The evaluation of quantitative strategies relies on fragmented metrics creating dangerous “evaluation traps,” concealing catastrophic tail risk (the “Sortino Trap”) or masking pathological volatility (the “Calmar Trap”). This deficiency is especially consequential in Deep Reinforcement Learning (DRL), where agents relentlessly optimize given objective functions. This paper introduces a unified, five-principle framework, deriving two robust metrics: the Generalized Robust Trading Strategy Score (GRTSS), integrating path-dependency and survival via a hyperbolic penalty, and the Log-Transformed Market Adaptability Index (MAIL), addressing the “luck vs. skill” dilemma by decoupling decision consistency from outcome magnitude. Validation via Monte Carlo simulations, real-world backtests, sensitivity analysis, and a rigorous multi-asset comparative analysis (2020–2024) covering Stocks, Gold, Forex, and Crypto demonstrates the framework’s superiority over industry standards, including STARR, CVaR, Omega, Kappa, Ulcer Index, DSR, Gain-Loss, Rachev, and CDaR. While competitors favored Bitcoin despite a 76% drawdown, GRTSS correctly prioritized capital preservation, identifying Gold as the superior risk-adjusted asset. Furthermore, while traditional metrics computationally break down in pathological scenarios (e.g., returning − 4.82e + 14), our framework remains mathematically stable. Finally, we empirically implement a differentiable Soft-GRTSS surrogate objective. By training a Proximal Policy Optimization (PPO) agent in a non-stationary synthetic market with embedded Black Swan events, we demonstrate that the agent successfully learns survival-aware policies, avoiding catastrophic ruin while achieving positive equity growth. This confirms Soft-GRTSS as a tractable, non-exploding gradient signal for the gradient-based training of risk-aware AI trading systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating survival into reward functions for reinforcement learning in financial trading

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Reza Roshanpour,
  • Mohammadreza Parsanejad,
  • Mir Saman Pishvaee

摘要

The evaluation of quantitative strategies relies on fragmented metrics creating dangerous “evaluation traps,” concealing catastrophic tail risk (the “Sortino Trap”) or masking pathological volatility (the “Calmar Trap”). This deficiency is especially consequential in Deep Reinforcement Learning (DRL), where agents relentlessly optimize given objective functions. This paper introduces a unified, five-principle framework, deriving two robust metrics: the Generalized Robust Trading Strategy Score (GRTSS), integrating path-dependency and survival via a hyperbolic penalty, and the Log-Transformed Market Adaptability Index (MAIL), addressing the “luck vs. skill” dilemma by decoupling decision consistency from outcome magnitude. Validation via Monte Carlo simulations, real-world backtests, sensitivity analysis, and a rigorous multi-asset comparative analysis (2020–2024) covering Stocks, Gold, Forex, and Crypto demonstrates the framework’s superiority over industry standards, including STARR, CVaR, Omega, Kappa, Ulcer Index, DSR, Gain-Loss, Rachev, and CDaR. While competitors favored Bitcoin despite a 76% drawdown, GRTSS correctly prioritized capital preservation, identifying Gold as the superior risk-adjusted asset. Furthermore, while traditional metrics computationally break down in pathological scenarios (e.g., returning − 4.82e + 14), our framework remains mathematically stable. Finally, we empirically implement a differentiable Soft-GRTSS surrogate objective. By training a Proximal Policy Optimization (PPO) agent in a non-stationary synthetic market with embedded Black Swan events, we demonstrate that the agent successfully learns survival-aware policies, avoiding catastrophic ruin while achieving positive equity growth. This confirms Soft-GRTSS as a tractable, non-exploding gradient signal for the gradient-based training of risk-aware AI trading systems.