Self-augmenting Technical Indicator with Recurrent Reinforcement Learning

Skilled human traders often outperform rule-based algorithmic systems by intuitively recognizing path-dependent patterns in technical indicators - a cognitive advantage that current algorithms fail to capture. We introduce a reinforcement learning framework that extracts sequential information solely from technical indicator histories, specifically targeting the additional information that static rules might ignore. Applied to MACD signals across four major currency pairs using rigorous out-of-sample testing with 20 random seeds per asset, our approach demonstrates non-inferior performance with the best models consistently outperforming traditional rule-based strategies across all tested assets. Statistical analysis using two-sided Wilcoxon tests demonstrates significant Sharpe ratio improvements for GBP/USD (mean improvement from -0.0007 to 0.0019, p ≈ 0) and USD/JPY (from -0.0041 to 0.0056, p ≈ 0), while EUR/USD and USD/CHF show non-significant differences before addressing overfitting concerns. This research bridges human intuitive decision-making and algorithmic trading, demonstrating that historical trajectories in technical indicators contain exploitable information that enhances trading performance beyond static rule-based approaches. JEL Classification: C45 , C63 , G11 , G15 , G17 MSC Classification: 68T05 , 68T07 , 91G70 , 91G80