Financial information extraction with Large Language Models

Abstract

Stock market predictions are highly sensitive to the information contained in financial articles released daily. While large language models (LLMs) excel on financial tasks such as named entity recognition, their effectiveness for specialized information extraction of financial types (monetary financial aspect of a company such as revenues, dividends, and cash flows) remains understudied. This work evaluates methods to improve financial information extraction (IE) with LLMs. We examine how model size and inclusion of domain specific information in prompting impact performance. We compare a smaller (GPT-4o-mini) and larger (GPT-4o) model using two approaches: (1) constrained generation, where models are constrained to select financial types from a list, and (2) inclusion of definitions in prompts to clarify ambiguous terms. Our findings reveal that constraining the model’s output does not outperform post-processing an unconstrained model, definitions in prompts are only beneficial for larger models, and smaller models can sometimes outperform larger models.