A new arXiv paper argues that financial NLP benchmarks have a timing problem: corporate disclosures do not happen at a single instant. Announcements, prepared remarks, and Q&A discussion each arrive at different points, and returns move in multiple high-frequency intervals around them. This makes it unclear where the 'post-call' period truly begins, which can skew any benchmark that labels text or prices relative to the call.
The authors propose an 'availability contract' — a formal rule for defining when post-call information becomes available. The abstract notes that benchmarks 'can suffer from a within-event' issue, though the full details are cut off. The core idea is that without a consistent, explicit contract, different researchers may label the same event differently, making results hard to compare.
The paper is positioned as a methodological intervention rather than a new model or dataset. If adopted, such a contract could make earnings-call NLP benchmarks more reliable and reproducible, especially for tasks that link language to market reactions. The abstract does not provide experimental results, so the practical impact remains to be seen.