Re(Visiting) Large Language Models in Finance

Abstract

This study evaluates the effectiveness of specialised large language models (LLMs) developed for accounting and finance. Empirical analysis demonstrates that these domain-specific models, despite being nearly 50 times smaller, consistently outperform state-of-the-art general-purpose LLMs in return prediction. By pre-training the models on year-specific financial datasets from 2007 to 2023, the study also mitigates look-ahead bias, a common limitation of general-purpose LLMs. The findings highlight the critical importance of addressing look-ahead bias to ensure reliable results. Extensive robustness checks further validate the superior performance of these models.

Keywords: Natural Language Processing, Large Language Models, Asset Pricing, Return Prediction, Machine Learning

JEL Classification: G10, G11, G14, C22, C23, C45, C55, C58