A new arXiv preprint proposes a benchmark that evaluates large language models on their ability to interpret the Management Discussion and Analysis (MD&A) sections of Form 10-Q filings. These sections blend long-form management narrative with accounting data and forward-looking statements, making them a rich but challenging source of financial information.
The authors argue that existing benchmarks for financial filings typically focus on question answering or numerical reasoning. That leaves unclear whether LLMs can actually capture the signals that drive stock market movements and analyst forecast revisions. The new benchmark directly tests this by scoring models on their predictions of market reactions and analyst forecast changes.
Because the abstract is truncated, the full methodology and results are not available from the source alone. Still, the framing suggests a meaningful step toward evaluating LLMs on real-world financial inference rather than static comprehension tasks.