Investigating the Predictive Capabilities of Large Language Models in Day Trading by Leveraging Multimodal Data

This paper evaluates the predictive capabilities of six LLMs-GPT-4, GPT-4o, Llama 3, Claude 3.5, Mistral 0.3, and Gemma 2-in day trading using multimodal data. The LLMs process diverse inputs, including text-based price histories, news titles, and images. The lowest Mean Absolute Percentage Errors (MAPEs) (1.4%) were achieved by Claude 3.5 and Gemma 2 using only price history text and by Claude 3.5 and Mistral 0.3 with combined price and news history inputs, demonstrating LLMs' potential for accurate financial predictions through prompting without advanced technical expertise. Remarkably, GPT-4 and Claude 3.5 achieve MAPEs of just 1.7% and 1.5%, respectively, by processing only price history images. Furthermore, Gemma 2 achieves a MAPE of 1.5% using only news history inputs, without any information from the price history.