This paper introduces FinRL-DeepSeek, a novel hybrid reinforcement learning (RL) and large language model (LLM) trading agent that integrates financial news insights for both action selection and risk management. The authors extend the Conditional Value-at-Risk Proximal Policy Optimization (CVaR-PPO) algorithm by incorporating LLM-generated stock recommendation scores and risk assessment scores from the FNSPID financial news dataset. The LLM scores are used to perturb RL actions (via recommendation scores) and adjust trajectory returns (via risk scores) in a risk-sensitive manner. The approach is backtested on the Nasdaq-100 index using three LLMs: DeepSeek V3, Qwen 2.5, and Llama 3.3.

Key results show that LLM infusion can improve performance in specific market conditions. In bull markets, standard PPO tends to outperform, while in bear markets, the LLM-infused CVaR-PPO (CPPO-DeepSeek) achieves better cumulative returns and risk metrics. The impact of LLM infusion strength is non-monotonic: for PPO, stronger infusion degrades performance, whereas for CPPO, stronger infusion improves performance. The paper also highlights the importance of training duration and memory optimization for scalability. Overall, the work demonstrates the potential of combining LLM-based feature extraction from news with risk-sensitive RL for algorithmic trading, though results are mixed and sensitive to market regimes and infusion parameters.