Paper: arXiv 2510.19173

Authors: Qing-Yu Lan, Zhan-He Wang, Jun-Qian Jiang, Yu-Tong Wang, Yun-Song Piao

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.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced sequence models (RNNs, Transformers) and specific RL algorithms (DDQN, GRPO) for time-series processing, constituting moderate to high mathematical complexity. However, the methodology is described conceptually without code, specific dataset details, or statistical performance metrics, suggesting it is more of a theoretical framework than a backtest-ready implementation.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Incorporate news into<br>trading via RL"] --> B{"Data Collection"}
  B --> C["Raw Market Data<br>(Price & Volume)"]
  B --> D["News Data"]
  D --> E["LLM Sentiment Scoring"]
  E --> F["News-Aware Inputs"]
  C --> F
  F --> G{"Reinforcement Learning<br>Process"}
  G --> H["Algorithm: DDQN & GRPO<br>using Transformers/RNNs"]
  H --> I["End-to-End Trading<br>Decisions"]
  I --> J["Key Outcomes"]
  J --> K["Superior Performance<br>vs Benchmarks"]
  J --> L["No Handcrafted<br>Features Required"]
  J --> M["Time-Series<br>Critical"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a news-aware reinforcement learning framework that directly uses LLM-derived news sentiment scores together with raw price and volume data as inputs processed by sequence models for end‑to‑end trading decisions.

Source sentence
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.

Assumes 4

  • Leaving out news makes the market only partially observable and increases non‑stationarity. market
    Source sentence
    Leaving out news makes the whole market condition only partially observable and increases non-stationarity.
  • The dataset split chronologically (70% train, 15% validation, 15% test) prevents look‑ahead bias and yields realistic performance evaluation. data
    Source sentence
    We divided the whole dataset into training, validation and test area using a chronological split to prevent look-ahead bias and ensure realistic performance evaluation
  • 1‑minute OHLCV data from Binance for BTC/USDT and news scraped from Yahoo Finance are sufficient to represent the cryptocurrency market and its news flow. data
    Source sentence
    The whole dataset, including market 1-minute OHLCV (i.e. open, high, low, close, volume) time-series sourced from Binance Exchange
  • Sequence‑based agents (LSTM, Transformer) can capture temporal dependencies better than MLP agents. model
    Source sentence
    confirming the value of modeling prices and sentiment signals as continuous time-series rather than isolated inputs

Data

Backtest sample: BTC/USDT · crypto · intraday

  • Binance BTC/USDT 1‑minute OHLCV · Binance Exchange market-prices public
    2019-12-31 00:00:00 to 2024-01-24 21:48: · BTC/USDT
    Source sentence
    The whole dataset, including market 1-minute OHLCV (i.e. open, high, low, close, volume) time-series sourced from Binance Exchange
  • Yahoo Finance Bitcoin news · Yahoo Finance text-news public
    2019-12-31 00:00:00 to 2024-01-24 21:48: · Bitcoin‑related news
    Source sentence
    news text scrapped from Yahoo Finance related to Bitcoin

Recognised sources: Binance Yahoo Finance

Costs

Not modeled A trading or execution result is reported without costs.

Finds 2

Reported performance (news-aware RL framework with LSTM): total return 124.5%

  • Incorporating LLM‑derived news‑sentiment inputs yields higher returns than the corresponding model without LLM signals, confirming the value of news analysis. performance
    Source sentence
    Incorporating LLM-derived news-sentiment inputs yields higher returns than the corresponding model without LLM signals, confirming the value of news analysis.
  • The proposed framework exceeds BTC market benchmarks, with LSTM‑based agents achieving up to 124.5% top‑1 return for DDQN, outperforming both MLP agents and models without news signals. performance
    Source sentence
    results illustrate that our proposed framework can exceed the BTC market benchmarks.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 13 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.