Aprendizado por Reforço com LSTM e DDPG otimizados através de comitê PSO para análise preditiva de séries temporais financeiras

Extracting predictive signals in financial time series is challenging due to inherent noise. This paper proposes a Deep Reinforcement Learning architecture based on DDPG and LSTM (Long Short-Term Memory) networks, structured into a committee of specialist agents across multiple scales (daily, 1h, 30m, and 5m). The actions recommended by individual agents are weighted using Particle Swarm Optimization. Backtesting results demonstrate that the proposed framework achieved a Sharpe Ratio of 0,81 and 22,29% return, outperforming standard market benchmarks with significantly reduced maximal drawdown.