Exploring financial sentiment analysis via fine-tuning large language model and attributed graph neural network.

Area of Science:

• Computational Finance
• Natural Language Processing
• Machine Learning

Background:

• Financial sentiment analysis (FSA) faces challenges with pre-trained LLMs due to domain specificity and data schema adaptation.
• Existing LLMs often neglect cross-asset impacts, focusing solely on individual stock information for price prediction.

Purpose of the Study:

• To develop a novel framework synergizing LLMs and GNNs for improved stock price dynamics modeling using financial news sentiment.
• To enhance LLM sensitivity to financial sentiment and effectively model cross-asset dependencies.

Main Methods:

• Utilized the Llama-3-8B model, fine-tuned with SFT and DPO for financial sentiment sensitivity.
• Developed a GNN to process LLM sentiment outputs, creating text-attributed graphs to model cross-asset dependencies and time-varying correlations.

Main Results:

• Demonstrated that financial sentiment significantly influences stock price variations in the Chinese A-share market.
• The proposed framework outperformed existing baselines, achieving an average Sharpe ratio improvement of 50%.

Conclusions:

• The integrated LLM-GNN framework effectively captures financial sentiment and cross-asset dynamics for stock price prediction.
• This approach offers a significant advancement in applying NLP and graph-based methods to financial market analysis.