FinRank: An Evidence-Grounded Benchmark for Financial Question Answering and Retrieval over SEC Filings
FinRank tests whether financial QA systems can find the right SEC filing evidence, not just produce the right-looking answer.
The benchmark has 1,185 manually written question-answer records from 10-K and 10-Q filings for 22 companies. Each record includes gold supporting passages and hard negatives selected from easily confused disclosures across filings, periods, and peer firms. Baselines struggle: the strongest reported 7B instruction-tuned embedder reaches 44.8% Recall@10 on the pooled evidence corpus. The paper frames this as a benchmark for answer grounding in the correct entity, reporting period, and disclosure context. ArXiv · AI/CL/LG's note
The benchmark has 1,185 manually written question-answer records from 10-K and 10-Q filings for 22 companies. Each record includes gold supporting passages and hard negatives selected from easily confused disclosures across filings, periods, and peer firms. Baselines struggle: the strongest reported 7B instruction-tuned embedder reaches 44.8% Recall@10 on the pooled evidence corpus. The paper frames this as a benchmark for answer grounding in the correct entity, reporting period, and disclosure context. ArXiv · AI/CL/LG's note
score 4