[Paper Review] Integrating Stock Features and Global Information via Large Language Models for Enhanced Stock Return Prediction

Yujie Ding, Shuai Jia+5 more|arXiv (Cornell University)|Oct 9, 2023

Stock Market Forecasting MethodsDecision Sciences3 citations

TL;DR

This paper proposes SCRL-LG, a novel framework that integrates stock features and global information via Large Language Models (LLMs) for enhanced stock return prediction. By employing a Local-Global model to decompose returns and Self-Correlated Reinforcement Learning to align LLM-generated news embeddings with quantitative stock features in a shared semantic space, the method achieves superior performance in Rank Information Coefficient and abnormal returns, especially over longer holding periods in the China A-share market.

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.

Motivation & Objective

• To address the underutilization of semantic knowledge in LLMs for stock return prediction.
• To solve the misalignment between LLM-generated news embeddings and quantitative stock features in different semantic spaces.
• To improve predictive performance by integrating local (stock-specific) and global (market, industry, policy) information through a unified framework.
• To demonstrate the effectiveness of LLMs in multi-modal stock return prediction beyond sentiment analysis.

Proposed method

• The Local-Global (LG) model decomposes stock returns into a local component (α) from volume and price features and a global component (β) from market, industry, and policy information.
• Three strategies for global modeling are introduced: (1) stock feature-based, (2) LLM-native, and (3) hybrid, enabling flexible integration of global context.
• Self-Correlated Reinforcement Learning (SCRL) aligns LLM-generated news embeddings with stock features in a shared semantic space through a contrastive learning objective.
• The framework uses a joint loss function combining ranking loss and contrastive loss to optimize alignment and predictive accuracy.
• Backtesting is conducted on China A-share data from 2022, evaluating performance across varying holding periods.
• The model leverages LLMs not only for sentiment but for rich semantic representation extraction from news headlines.

Experimental results

Research questions

• RQ1 Can LLMs be effectively leveraged beyond sentiment analysis to extract comprehensive semantic knowledge for stock return prediction?
• RQ2 How can embeddings from LLMs and quantitative stock features be aligned in a shared semantic space to improve predictive performance?
• RQ3 Does integrating global information (market, industry, policy) via LLMs mitigate performance degradation in long-horizon stock return predictions?
• RQ4 How does the SCRL-LG framework compare to models relying solely on stock features in terms of Rank Information Coefficient and abnormal returns?

Key findings

• The SCRL-LG model significantly outperforms models using only stock features in both Rank Information Coefficient and abnormal returns on China A-share data.
• The model maintains or improves performance with increasing prediction horizons, while the Local model alone shows a marked decline in annual returns and Sharpe ratio.
• The LG-LLM and LG-STOCK variants demonstrate robustness to longer holding periods, with minimal performance degradation.
• The SCRL-LG model exhibits a decreasing turnover rate alongside performance improvements, indicating efficient and stable trading signals.
• The hybrid global modeling strategy (combining stock features and LLMs) proves most effective in capturing multi-level (macro, meso, micro) influences on stock returns.
• The framework confirms that aligning LLM-generated embeddings with stock features in a unified semantic space enhances predictive accuracy.