Insights About Causality Detection in Financial Text - Towards an Informed Approach

We perform a linguistic investigation of causality in financial reports, and find that there are a number of intricacies specific to this domain, making it hard for a machine learning model to accurately detect causal statements. Specifically, cause and effect clauses are oftentimes very subtle or implicit. Additionally, some degree of world knowledge and reasoning is necessary to successfully identify many of those statements. We apply our findings by prompting GPT-4o with the acquired knowledge to improve its predictive capabilities. The results suggest that an informed approach can help enhance the performance of a causality detection system, possibly allowing for more intelligent and light-weight solutions in the future.