Reliable Financial Named Entity Recognition under Domain Shift

Confidence estimation for financial NER degrades significantly under domain shifts between SEC filings and social media. While whole-output probability detects errors well in-domain, entity-span probability and self-consistency maintain superior robustness across out-of-distribution tasks when evaluating BERT and LoRA-tuned Qwen2.5 (0.5B/1.5B) models.