Enhancing stock timing predictions based on multimodal architecture: Leveraging large language models (LLMs) for text...

Area of Science:

• Financial Technology
• Natural Language Processing
• Computational Finance

Background:

• Online investor sentiment analysis faces challenges with data quality, redundancy, and authenticity.
• Traditional quantitative methods often overlook qualitative insights from social media.
• Accurate stock timing predictions are crucial for investment decision-making.

Purpose of the Study:

• To enhance stock timing predictions using large language models (LLMs) for analyzing online investor comments.
• To develop and evaluate a multimodal architecture integrating LLM-processed sentiment with financial data.
• To assess the efficacy of GPT-4 in filtering and analyzing unstructured financial commentary.

Main Methods:

• Utilized GPT-4 to filter and analyze investor comments from Chinese bank data.
• Developed a multimodal architecture combining filtered comment data with stock prices and technical indicators.
• Compared GPT-4 filtering against four baseline models and evaluated performance using financial metrics.

Main Results:

• GPT-4 significantly improved key financial metrics, including profit-loss ratio, win rate, and excess return rate.
• The proposed multimodal architecture outperformed baseline models in stock timing prediction.
• Effective preprocessing of comment data by LLMs enhanced the integration with quantitative financial information.

Conclusions:

• Large language models, particularly GPT-4, offer a powerful tool for improving financial forecasting accuracy.
• The multimodal architecture provides a robust framework for integrating qualitative sentiment data with quantitative financial analysis.
• The methodology demonstrates potential for broader application in diverse financial markets, aiding investor decision support.