News-Aware Direct Reinforcement Trading for Financial Markets

TL;DR

The paper tackles the challenge of incorporating news into quantitative trading by directly integrating LLM-derived sentiment signals with raw price/volume into reinforcement learning agents. It evaluates DDQN and GRPO using sequence-based encoders (LSTM/Transformer) on 1-minute BTC/USDT data, demonstrating that news-aware, end-to-end RL can outperform BTC benchmarks without handcrafted features. Key findings show that incorporating news signals enhances performance, with LSTM-based architectures consistently benefiting from sentiment information, while Transformer-based models are less sensitive to it in this setup. This work highlights the potential of time-series-aware, news-informed RL for practical crypto trading and motivates further research on more sophisticated architectures and decision-making frameworks.

Abstract

The financial market is known to be highly sensitive to news. Therefore, effectively incorporating news data into quantitative trading remains an important challenge. Existing approaches typically rely on manually designed rules and/or handcrafted features. In this work, we directly use the news sentiment scores derived from large language models, together with raw price and volume data, as observable inputs for reinforcement learning. These inputs are processed by sequence models such as recurrent neural networks or Transformers to make end-to-end trading decisions. We conduct experiments using the cryptocurrency market as an example and evaluate two representative reinforcement learning algorithms, namely Double Deep Q-Network (DDQN) and Group Relative Policy Optimization (GRPO). The results demonstrate that our news-aware approach, which does not depend on handcrafted features or manually designed rules, can achieve performance superior to market benchmarks. We further highlight the critical role of time-series information in this process.

Figures (4)

• Figure 1: Schematic of the proposed news-aware reinforcement learning framework for financial trading. The system comprises an LLM that analyzes financial news; its output signals are combined with market data and then input to an RL agent. The RL agent utilizes DDQN/GRPO algorithms with various network architectures to process the integrated time-series data, and the action head's output actions interact with the trading environment to generate rewards.
• Figure 2: The format of input prompt that guides LLMs to generate sentiment scores and risk scores. It consists the task specification, the scoring system and the output format example.
• Figure 3: BTC (1-minute) Prices timeline division in our framework. Blue part: training set. Orange part: validation set. Green part: test set.
• Figure 4: Generalization performance of validation-selected Top-K agents on the BTC test set: DDQN (top row) and GRPO (bottom row). Each panel marks cumulative return on the test set in USDT terms (left-side axis) and percentage terms (right-side axis). Columns correspond to $K\in\{1,10\}$; for $K=10$, curves are averaged across the top-$K$ agents ranked by validation performance. The gray curve denotes the BTC prices baseline, while colored curves indicate distinct network architectures (MLP, LSTM, Transformer) and the presence/absence of an LLM-derived news-sentiment signals.