Attractor-Aware Hyperbolic Lipschitz-Constrained-Reinforcement Learning for FX Market: LLM-Structured Chain of Thought With Lyapunov–Entropy Dynamics
Open Access
- 23 October 2025
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 182977-183001
Abstract
Foreign exchange (FX) markets are non-stationary and event-driven, with heavy tails and regime shifts that complicate generalization and risk control.We introduce an explainable FX trading pipeline that integrates chaotic-systems diagnostics with geometry-aware learning. A Large Language Model (LLM) consolidates several dynamic heterogeneous evidence streams into a Structured Chain-of-Thought (s-CoT) that provides rules, supporting evidence, and uncertainties. The s-CoT is embedded in a hyperbolic space to preserve hierarchical dependencies and to render recurrent price levels geometrically salient. Chaotic and dynamical features, including Lyapunov exponents and permutation entropy, estimate local stability, forecastability horizons, and disorder. Recurrent magnet prices are identified as attractors on the price axis through recurrence density and cross-scale persistence, and they inform entry timing and risk budgeting. A Reinforcement Learning (RL) policy operates on the hyperbolic embedding under Lipschitz constraints that bound geodesic step sizes and stabilize policy updates under distribution shift. The result is an interpretable reasoning-to-action pipeline that produces daily decisions and explicit risk budgets. In pair-averaged non-compounding back-tests across nine major FX pairs from Jan. 2015 to July 2025, the system attains 2.22% monthly return with 10.39% maximum drawdown. An ablation study across components confirms the robustness of the full pipeline, indicating that geometry, Lipschitz control, and s-CoT jointly sustain performance while maintaining controlled drawdown. The empirical evaluation confirmed that encoding structured reasoning in hyperbolic space and steering actions toward magnet-price attractors yields regime-resilient FX trading with strong average performance and transparent risk control.Keywords
This publication has 53 references indexed in Scilit:
- Deep learning for time series forecasting: a surveyInternational Journal of Machine Learning and Cybernetics, 2025
- Hyperbolic Deep Learning in Computer Vision: A SurveyInternational Journal of Computer Vision, 2024
- Regularisation of neural networks by enforcing Lipschitz continuityMachine Learning, 2020
- Deep Reinforcement Learning for Foreign Exchange Trading2020
- On the Utility of Trading Criteria Based Retraining in Forex MarketsLecture Notes in Computer Science, 2013
- Iterated Maps on the Interval as Dynamical Systems2009
- Embeddings of Gromov hyperbolic spacesGeometric and Functional Analysis, 2000
- Metric Spaces of Non-Positive Curvature1999
- Lectures on Spaces of Nonpositive Curvature1995
- Detecting strange attractors in turbulence1981