Self-augmenting technical indicator with recurrent reinforcement learning: evidence from Forex markets

Abstract

Skilled human traders often outperform rule-based algorithmic systems by intuitively recognizing path-dependent patterns in technical indicators. 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 eight currency pairs, our approach demonstrates consistent improvement over the baseline in out-of-sample testing. The training period (2019–2021) encompasses diverse market conditions including the COVID-19 crisis, while the test period (2022–2025) covers the Ukraine war onset and monetary tightening, enabling assessment of cross-regime generalization. Statistical analysis reveals that RRL augmentation significantly improves mean out-of-sample performance relative to GA-optimized MACD for the majority of tested pairs. This research demonstrates that historical trajectories in technical indicators contain exploitable information that enhances trading performance beyond static rule-based approaches. Graphical abstract