Trading Strategies in the Peruvian Foreign Exchange Market with Dynamic Optimization: A Hybrid HMM–Deep Reinforcement Learning Approach

Resumen

Furthermore, this study develops and evaluates a hybrid architecture for algorithmic trading in the USD/PEN foreign exchange market, where a Gaussian Hidden Markov Model (HMM) infers latent market regimes and a deep reinforcement learning (DRL) agent optimizes Buy/Sell/Hold actions within an MDP formulation. Moreover, a reproducible data pipeline is implemented to integrate heterogeneous financial and macroeconomic sources, producing a consolidated dataset with 4,126 observations and 1,113 variables spanning 2010-02-02 to 2025-12-05, with continuity checks to support stable learning. The inferred regime signal (e.g., bearish/volatile, sideways, bullish) is incorporated as exogenous context to mitigate market non-stationarity and improve policy learning. Finally, out-of-sample benchmarking shows that the RL-based strategy outperforms supervised baselines and Buy & Hold, achieving the highest cumulative return (28.64%) and improved risk-adjusted performance when regime context is included (Sharpe 0.116 for RL+HMM vs 0.081 for RL-only), while remaining robust in adverse periods where the passive benchmark incurs losses (-14.56%). In addition, training dynamics suggest PPO yields more stable convergence than alternative DRL methods in a noisy financial environment. Keywords: USD/PEN; Hidden Markov Models; market regimes; deep reinforcement learning; PPO; algorithmic trading; risk management; CRISP-DM..