Finance benchmarks cover question answering, calculations, filing analysis, data extraction, and professional workflows. Source quality and data dates matter as much as the final answer.
Prefer tasks with traceable source documents, dated data, checkable calculations, and clear rules for when to abstain. Test again with your own filings, accounts, policies, and approval limits.
9
Benchmarks in this domain
Sources first · alphabetical collection order
Runnability describes access to a usable repository, dataset, or harness—not whether setup is easy.
EvasionBench
specializedrunnable
2026-11-03
Tests
Tests whether a model can tell if a manager answers an analyst question directly, partly, or not at all. Uses earnings-call Q&A (question and answer) pairs.
Best for
Evaluating financial-analysis systems that must distinguish disclosure from deflection.
Tests AI agents on the work of entry-level financial analysts. Agents use tools such as SEC (Securities and Exchange Commission) EDGAR search and web search to answer questions on public companies.
Tests LLMs (large language models) on 24 financial tasks from 36 datasets. Areas include extraction, text analysis, QA (question answering), generation, risk, forecasting, and stock trading.
Tests AI models on real work tasks from 44 occupations in the nine sectors that add most to US GDP (gross domestic product). Experts compare model output with expert work.
A leaderboard that tests LLMs (large language models) on financial tasks in seven groups, such as extraction, sentiment, risk, and forecasting. Uses zero-shot prompts.
Tests AI agents on quantitative finance work in a Docker sandbox. Agents must clean data, write and debug code, and produce numbers that tests can check.