Integrating Large Language Models and Reinforcement Learning for Sentiment-Driven Quantitative Trading

Battle Count: 156

The paper is directly and highly relevant to quantitative trading. It proposes a complete trading system integrating LLM-generated sentiment signals with technical indicators via reinforcement learning. It addresses practical concerns including transaction costs, look-ahead bias, out-of-sample evaluation, and factor decomposition. The comparison between rule-based and RL-driven approaches provides actionable insights for signal integration in systematic trading strategies.

High complexity due to the multi-stage pipeline: (1) news summarization with LLaMA 3.1 8B using vLLM, (2) sentiment classification with FinGPT, (3) technical indicator computation, (4) TD3 reinforcement learning training with custom gymnasium environment, (5) portfolio construction and backtesting. Requires GPU resources for LLM inference and RL training, access to commercial financial data (Thomson Reuters, CRSP, Bloomberg), and expertise in both NLP and reinforcement learning. The 512-epoch training with 5.1×10^7 updates is computationally intensive.

The paper provides detailed hyperparameters for TD3 (learning rate, discount factor, replay buffer size, batch size, policy noise, delay, total updates), network architecture (2 FC layers of 256 units with ReLU), and data processing pipeline. However, no code repository is mentioned, FinGPT model version details are limited, and the specific fine-tuning procedure on Google Colab is not fully described. The use of commercial data sources (Thomson Reuters, CRSP, Bloomberg) limits full reproducibility.