Researchers have found that temporal leakage in financial news NLP can significantly affect model performance. They audited a corpus of 49,799 articles across various feature-model combinations and found that random splits inflate MCC by 1.1 to 6.5 times, depending on model capacity and feature richness. The study also identified mergers and acquisitions (M&A) as the only category with a positive locked-test signal under near-temporal chronological evaluation. However, this signal does not transfer to other datasets, suggesting that it is specific to the 2024-2025 European-tilted M&A semantics used in the study. --- Why it matters: This research matters because it highlights the importance of temporal leakage audits in financial NLP benchmarks. Engineers and researchers need to be aware of this issue to ensure that their models are not relying on stale or predictable information, which can lead to biased results. Source: https://arxiv.org/abs/2608.17223