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

Abstract

This study aims to enhance stock timing predictions by leveraging large language models (LLMs), specifically GPT-4, to filter and analyze online investor comment data. Recognizing challenges such as variable comment quality, redundancy, and authenticity issues, we propose a multimodal architecture that integrates filtered comment data with stock price dynamics and technical indicators. Using data from nine Chinese banks, we compare four filtering models and demonstrate that employing GPT-4 significantly improves financial metrics like profit-loss ratio, win rate, and excess return rate. The multimodal architecture outperforms baseline models by effectively preprocessing comment data and combining it with quantitative financial data. While focused on Chinese banks, the approach can be adapted to broader markets by modifying the prompts of large language models. Our findings highlight the potential of LLMs in financial forecasting and provide more reliable decision support for investors.