A risk-constrained SARSA–FIS hybrid decision architecture with adaptive exploration control

Abstract

Algorithmic trading systems operate in highly dynamic and uncertain environments where learning-based decision agents must balance adaptability with strict risk control. Reinforcement learning (RL) methods provide adaptive policy optimization but often suffer from unstable exploration and limited interpretability in financial markets. This study proposes a risk-constrained SARSA–FIS hybrid decision architecture with adaptive exploration control for algorithmic trading. The framework integrates a compact SARSA-based reinforcement learning environment with a Sugeno-type fuzzy inference system (FIS) that converts reinforcement signals into interpretable trading decisions. Exploration follows a decaying ε-greedy policy with a drawdown-triggered reset mechanism to maintain bounded risk exposure during learning. The system was implemented as a MetaTrader 5 Expert Advisor and evaluated on the GBPUSD currency pair using historical market data. Experimental results show that the hybrid framework improves trading performance compared with a rule-based baseline. During a six-month out-of-sample evaluation, the system achieved a net profit of 90 USD and a profit factor of 1.35, compared with 10 USD and 1.02 for the baseline. Extended one-year testing confirmed stable profitability and controlled drawdown behavior. The results demonstrate that integrating reinforcement learning, fuzzy decision mapping, and explicit risk constraints provides a practical approach for developing adaptive trading agents.