Integrating Stock Features and Global Information via Large Language Models for Enhanced Stock Return Prediction

In one sentence

Summary

The paper addresses the challenge of integrating unstructured financial news with structured quantitative features for stock return prediction. It identifies two main issues: insufficient utilization of LLM semantic information and the misalignment between LLM embeddings and pre-existing stock features. The authors propose a novel framework called Self-Correlated Reinforcement Learning with Local-Global model (SCRL-LG). The Local-Global (LG) model decomposes returns into idiosyncratic alpha (local stock features) and beta (global market/news information). To bridge the semantic gap, the SCRL component uses Proximal Policy Optimization (PPO) to align LLM-generated news embeddings with stock feature spaces, treating the LLM as a feature selector and the quantitative model as a critic to ensure stable policy updates. Experiments are conducted on the China A-share market using 3,506 stocks and 342 daily price-volume features from 2019 to 2022. The LLM used is Llama 7B, fine-tuned on Chinese news corpora. The study compares SCRL-LG against baselines including a local-only model, a global model using only stock features, and a global model using only LLM embeddings. Evaluation metrics include Rank Information Coefficient (Rank IC), annual returns, Sharpe ratio, and maximum drawdown. The backtesting strategy involves daily sorting into deciles, trading the top 10% with a 0.3% transaction cost. Results show that SCRL-LG outperforms all baselines, achieving a Rank IC of 0.152 and a Sharpe ratio of 1.24, compared to 0.132 and 0.56 for the local-only model. The integration of global information significantly enhances predictive accuracy, particularly for longer-term horizons (10-20 days), where noise is averaged out. The method demonstrates robustness and improved risk-adjusted returns. However, limitations include the potential for overfitting due to high-dimensional alignment and the reliance on specific Chinese market data, which may limit generalizability to other markets or asset classes without further adaptation.

Research problem

The paper tackles the difficulty of effectively integrating unstructured financial news, processed by Large Language Models (LLMs), with structured quantitative stock features for accurate stock return prediction. It specifically addresses the challenges of underutilizing LLM semantic information and the misalignment between LLM embeddings and traditional quantitative feature spaces.

Core contributions

• Proposes the Local-Global (LG) model to decompose stock returns into idiosyncratic alpha (local features) and beta (global news/market) components.
• Introduces Self-Correlated Reinforcement Learning (SCRL) using PPO to align LLM-generated news embeddings with stock feature embeddings in a shared semantic space.
• Demonstrates that LLMs can act as effective feature selectors when integrated with quantitative models via reinforcement learning.
• Provides empirical evidence on the China A-share market showing superior predictive performance and risk-adjusted returns compared to baselines.

Method

• Local-Global decomposition: Separates alpha (stock-specific) and beta (market/news) components.
• LLM Embedding: Uses Llama 7B to generate daily news vectors from headlines.
• SCRL Alignment: Uses PPO to align LLM embeddings with stock features via a sparse vector selector.
• Reward Function: Combines prediction reward with KL divergence to stabilize training.
• Feature Selection: LLM acts as a critic to select relevant global information for the actor (quantitative model).
• Periodic Retraining: LLM is retrained every 60 trading days on selected news to adapt to market changes.

Results & findings

• SCRL-LG achieves a Rank IC of 0.152, outperforming the local-only model (0.132) and LG-STOCK (0.142).
• Annual returns for SCRL-LG are 28.8%, significantly higher than the local model's 11.2%.
• Sharpe ratio improves from 0.56 (local) to 1.24 (SCRL-LG), indicating better risk-adjusted performance.
• Performance gaps widen for longer prediction horizons (10-20 days), suggesting global information helps filter noise.
• The method reduces turnover rates while maintaining or improving returns compared to shorter-horizon models.

Data & experiments

• Dataset: 3,506 China A-share stocks with 342 daily price-volume features from 2019-2022.
• News Data: 162,845 daily news headlines used to generate LLM embeddings.
• LLM: Llama 7B fine-tuned on Chinese news corpora (CLUECorpusSmall, CSL).
• Baselines: Local-only, LG-STOCK (features only), LG-LLM (LLM only), and SCRL-LG (hybrid).
• Metrics: Rank IC, annual returns, Sharpe ratio, max drawdown, turnover.
• Backtesting: Top 10% quantile strategy with 0.3% transaction costs.

Limitations & caveats

• Potential overfitting due to high-dimensional alignment between LLM and stock features.
• Reliance on Chinese market data may limit generalizability to other regions.
• LLM fine-tuning requires significant computational resources.
• News data quality and coverage may impact performance.
• The method assumes news is released before market open, which may not always hold.

Curator notes

Discovery status: round3_promoted. Source links: Automate Strategy Finding with LLM in Quant investment; Large Language Models in equity markets: applications, techniques, and insights