de Jesus, Luiz Carlos , Fernandez-Navarro, Francisco , CARBONERO RUZ, MARIANO
No
Comput. Econ.
Article
Científica
2
2
03/10/2026
001892398500001
2-s2.0-105052311949
We introduce a novel strategy that incorporates Topological Data Analysis (TDA) into a Deep Reinforcement Learning (DRL) system tailored for financial trading tasks. Standard econometric techniques and deep learning models frequently struggle to capture the nonlinear dependencies and regime shifts characteristic of asset price behaviour. Although DRL has gained popularity in financial applications, it typically depends on statistical or technical signals, which may limit interpretability and resilience. TDA, by contrast, provides a mathematical framework for capturing the evolving geometric patterns in time series through persistent homology. Our method employs time-delay embedding with sliding windows to convert one-dimensional price sequences into point clouds. These are transformed into persistence diagrams, from which quantitative descriptors-such as entropy, amplitude, and feature counts-are extracted. These topological features are incorporated into the state representation of a Deep Q-Network (DQN) agent, enabling it to leverage the structural properties of financial time series. We test the proposed model across multiple datasets spanning both cryptocurrencies and traditional financial markets, and under varying regime conditions. Comparative experiments show that augmenting the agent's input space with TDA-derived metrics improves performance in terms of Sharpe ratio, Sortino ratio, total economic return, and final net worth, outperforming baselines that use purely technical or statistical inputs. The code is available at .
Topological data analysis; Financial time series; Reinforcement learning; Deep Q-network