Paper: arXiv 2310.05627

Authors: Yujie Ding, Shuai Jia, Tianyi Ma, Bingcheng Mao, Xiuze Zhou, Liuliu Li, Dongming Han

Abstract

The remarkable achievements and rapid advancements of Large Language Models (LLMs) such as ChatGPT and GPT-4 have showcased their immense potential in quantitative investment. Traders can effectively leverage these LLMs to analyze financial news and predict stock returns accurately. However, integrating LLMs into existing quantitative models presents two primary challenges: the insufficient utilization of semantic information embedded within LLMs and the difficulties in aligning the latent information within LLMs with pre-existing quantitative stock features. We propose a novel framework consisting of two components to surmount these challenges. The first component, the Local-Global (LG) model, introduces three distinct strategies for modeling global information. These approaches are grounded respectively on stock features, the capabilities of LLMs, and a hybrid method combining the two paradigms. The second component, Self-Correlated Reinforcement Learning (SCRL), focuses on aligning the embeddings of financial news generated by LLMs with stock features within the same semantic space. By implementing our framework, we have demonstrated superior performance in Rank Information Coefficient and returns, particularly compared to models relying only on stock features in the China A-share market.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 5.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces advanced mathematical concepts like attention mechanisms, reinforcement learning, and multi-component decomposition within a deep learning framework, scoring high in math complexity. It presents empirical results on the China A-share market with performance metrics like Rank Information Coefficient, indicating backtest-ready methodology, but relies on LLM embeddings without explicit code or dataset sharing, placing it at moderate empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Integrate LLMs with Stock Features<br>for Enhanced Return Prediction"] --> B["Data & Inputs<br>China A-share Market<br>Stock Features & Financial News"]
  B --> C["Methodology: Local-Global Model<br>1. Stock Feature Strategy<br>2. LLM-based Strategy<br>3. Hybrid Strategy"]
  C --> D["Methodology: Self-Correlated Reinforcement Learning<br>Aligns LLM Embeddings with Stock Features<br>in Semantic Space"]
  D --> E["Computational Process<br>Framework Training & Validation"]
  E --> F["Key Findings/Outcomes<br>Superior Rank IC & Returns<br>vs. Stock Feature-only Models"]

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 Self-Correlated Reinforcement Learning with Local-Global model (SCRL-LG) that aligns LLM news embeddings with stock features to improve stock return prediction.

Source sentence
To demonstrate the effectiveness of LLMs in predicting stock returns, we have introduced a novel approach known as Self-Correlated Reinforcement Learning with Local-Global model (SCRL-LG).

Assumes 6

  • All trades are executed at the closing price with a flat transaction cost of 0.3% and no slippage. execution
    Source sentence
    All transactions were executed at the closing prices, with a transaction cost of 0.3%.
  • News used for alignment are selected from the previous 60 trading days. data
    Source sentence
    Finally, we select the news related to the selected stock features from the previous 60 trading days in the news corpus.
  • The LLM embeddings can be aligned with stock feature space via KL divergence and PPO reinforcement learning. model
    Source sentence
    We propose a Self-Correlated Reinforcement Learning (SCRL) framework, incorporating the Proximal Policy Optimization (PPO) paradigm
  • The market universe is limited to China A‑share stocks that were not suspended and listed before 2020. data
    Source sentence
    We exclude suspended stocks or the stocks listed after 2020.
Show 2 more
  • Predicted returns are sorted daily into ten quantiles and the top 10% is bought while others are liquidated. execution
    Source sentence
    Our strategy involved purchasing an equal allocation of stocks from the top 10% quantile while liquidating portfolios not belonging to this quantile.
  • The LLM (Llama 7B) fine‑tuned on Chinese corpora provides meaningful representations for news headlines. model
    Source sentence
    ... continued training Llama for an additional 2 epochs on 15 GB of Chinese corpus

Data

  • China A‑share dataset market-prices unknown
    01/01/2019 to 12/31/2022 · 3506 stocks
    Source sentence
    The experiments are conducted on the China A‑shares market using a dataset comprising 3506 stocks and 342 daily features constructed from stock price‑volume data ranging from 01/01/2019 to 12/31/2022.
  • News headlines corpus text-news unknown
    01/01/2019 to 12/31/2022
    Source sentence
    Additionally, we collect news headlines during the same period, resulting in a dataset size of 162,845 headlines and 2,526,174 stock data points for training (from 01/01/2019 to 12/31/2021) and 849,173 data points for testing.

Costs

Modeled Costs are part of the reported headline results.

A flat transaction cost of 0.3% per trade is applied to all transactions.

Source sentence
All transactions were executed at the closing prices, with a transaction cost of 0.3%.

Finds 4

  • SCRL‑LG outperforms all baselines, achieving higher Rank IC, annual return, Sharpe ratio, and lower max draw‑down compared to LG‑STOCK and LG‑LLM. performance
    Source sentence
    The results show that SCRL‑LG yields noticeable improvements in all metrics compared to the methods above for generating global information.
  • SCRL‑LG shows increasing performance with longer prediction horizons, contrary to the Local model which degrades. performance
    Source sentence
    Surprisingly, the SCRL‑LG model even exhibits performance improvements with increasing prediction horizons, accompanied by a decrease in turnover rate.
  • Incorporating global information (LG‑STOCK, LG‑LLM) mitigates the decline in predictive performance observed in the Local model. predictive-accuracy
    Source sentence
    These results collectively indicate that the three proposed approaches for incorporating global information can effectively mitigate the decline in predictive performance observed in the Local model.
  • The proposed framework demonstrates superior Rank Information Coefficient and returns compared to models relying only on stock features in the China A‑share market. performance
    Source sentence
    By implementing our framework, we have demonstrated superior performance in Rank Information Coefficient and returns, particularly compared to models relying only on stock features in the China A‑share market.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 51 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 9 papers in the archive (8 by independent authors): TradingAgents: Multi-Agents LLM Financial Trading Framework (uses its method or data)

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.