Researchers have developed a method to improve financial named entity recognition (NER) under domain shift. They tested their approach on three types of text: SEC filings, financial news, and social media posts. The results show that confidence estimation and selective prediction can help reduce errors when the input distribution changes. A staged deployment strategy is proposed to detect severe domain shift before applying prediction-level confidence gating. --- Why it matters: This matters because financial AI systems often struggle with domain shift, leading to inaccurate predictions. By developing methods to handle this issue, researchers can improve the reliability of financial NER and reduce errors in automated decision-making processes. Source: https://arxiv.org/abs/2608.19558