Financial Document Causality Detection Task (FinCausal 2026)

Financial analysis depends on factual data and explanations of data variability. While data present facts, they do not reveal why these facts occurred. It’s important to examine the narrative behind the data presentation. Recognising causality is essential for understanding decision-making processes.
The Financial Document Causality Detection Task aims to develop the ability to explain, from external sources, why a transformation occurs in the financial landscape as a preamble to generating accurate and meaningful financial narrative summaries. Its goal is to evaluate which events or chains of events can cause a financial object to be modified or an event to occur, regarding a given context.
Participants will need to identify the cause or effect within the given segments. There are two subtasks: one in English and one in Spanish. The English dataset has been sourced from various 2017 UK financial annual reports from the corpus provided by UCREL at Lancaster University and from the English version of the 2018 FinT-esp corpus. The Spanish dataset has been extracted from a corpus of Spanish financial annual reports from 2014 to 2018. These datasets are comparable across both languages to facilitate testing of multilingual models. For the 2026 edition, the previous dataset has been expanded with 500 examples in each language, including complex cause-and-effect structures.
Traditionally, the extraction of cause-effect relationships has been extractive, as seen in the first editions of FinCausal. In 2025, it was presented as a generative AI task, in which abstractive questions about causes or effects are posed, and the system’s responses are evaluated using exact-matching and similarity metrics.
In FinCausal 2026, we have introduced more features and innovations:

• Complete review of the previous datasets, removing ambiguous or very simple cases. More than 500 new fragments with complex relationships have been added (such as a causal chain of three or more elements).
• Rephrasing of the abstractive questions in 10% of the cases, to create a dataset that requires more advanced reasoning and is less dependent on similarity to the original text.
• Random partitioning of the training and test sets, based on the new 2026 dataset. In this way, the innovations are distributed evenly.
• Introduction of an LLM-as-a-judge–based evaluation metric, which scores system responses on a 1–5 scale according to their adequacy. This metric replaces the previous SAS + Exact Match evaluation scheme and aligns FinCausal with current practices adopted in recent shared tasks and competitions.