AI Trading Agent Achieves 124.5% Return, Outperforming Bitcoin by 68.5%

A new research paper, "News-Aware Direct Reinforcement Trading for Financial Markets," co-authored by Rohan Paul, details the development of an AI-driven trading agent that significantly surpasses market baselines by integrating news sentiment with price data. The study, published on arXiv.org, demonstrates a cumulative return of 124.5% on a held-out test set, compared to Bitcoin's 56% return over the same period. This represents a 68.5 percentage point outperformance. The innovative approach utilizes large language models (LLMs) to extract sentiment scores from financial news, which are then combined with raw price and volume data. This integrated data feeds into sequence-based reinforcement learning agents, such as Long Short-Term Memory (LSTM) networks or Transformers, to make trading decisions without relying on handcrafted features or manually designed rules. The agents employ Double Deep Q Network (DDQN) and Group Relative Policy Optimization (GRPO) to determine long, short, or hold positions. According to the research team, "News signals, when fed directly with prices, make simple sequence based agents meaningfully stronger." The paper emphasizes that sequence models consistently outperformed simpler multilayer perceptrons, underscoring the importance of temporal patterns over static snapshots in financial markets. Specifically, adding news data stably boosted LSTM agents, while plain Transformers showed less pronounced gains in this experimental setup. The study serves as a proof-of-concept, highlighting the feasibility and effectiveness of directly incorporating LLM-derived news sentiment into reinforcement learning frameworks for financial trading. This development marks a significant step towards more sophisticated and autonomous trading systems, particularly in volatile markets like cryptocurrency, where the research was applied using BTC/USDT data. The findings suggest a promising direction for future research in AI-driven financial strategies, aiming for minimal manual intervention.