Finance AI benchmark directory
Every finance-AI benchmark we track, with links to the originals. Rows marked Moneda Bench are Moneda's own. Spot something wrong? Send a correction.
Updated 2 Nov 2026
52 benchmarks
| Top score | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| AA-AnalystAgent | Financial spreadsheets and accounting | Artificial Analysis | Gemini 3.7 Flash (High) | 60 pass^5 % | Not stated | none (only the example tasks shown on the page) | Not stated | Not stated | Not stated |
| AccountingBench | Financial spreadsheets and accounting | Penrose | 2025 | none (write-up, sample agent transcripts and metric charts only) | Not stated | none | Not stated | ||
| Advanced Financial Reasoning at Scale (CFA Level III evaluation) | Finance knowledge | GoodFin, NYU and Rochester Institute of Technology authors (Shetty et al.) | o4-mini | 79.1 overall % (mean of MCQ and essay) | 2025 | none (paywalled AnalystPrep mock material) | Not stated | none | No |
| Agent Market Arena (AMA) | Financial trading and portfolios | The Fin AI, with Stevens Institute of Technology, Columbia, Harvard, University of Manchester, Université de Montréal, University of Florida, Georgia Tech, PAAL AI, DeepKin | InvestorAgent / GPT-4.1 | 40.8 cumulative return % (TSLA) | 2025 | Live dashboard only; no code or dataset release linked | Not stated | maker-only | Not stated |
| AI-Trader | Financial trading and portfolios | The University of Hong Kong (HKUDS) | MiniMax-M2 | 9.6 cumulative return % | 2025 | Code and evaluation data on GitHub; live platform at ai4trade.ai | MIT (code, per README badge) | open | Yes |
| Alpha Arena | Financial trading and portfolios | Nof1 | Not stated | Not stated | Not stated | Not stated | Not stated | ||
| APEX-Accounting | Financial spreadsheets and accounting | Mercor, in partnership with Ramp | Gemini 4 Argon | 71.5 Mean Score (% criteria passed) | 2026 | One open world of 10 tasks (prompts, rubrics, files) on Hugging Face; the full task set stays private | Not stated | via-maker | Not stated |
| BizFinBench | Finance knowledge | HiThink Research | ChatGPT-o3 | 73.9 average score across 9 tasks | 2025 | All 6,781 questions on Hugging Face (test split per subset) | CC-BY-NC-4.0 (data, research use only); Apache-2.0 (code) | maker-only | No |
| BlueFin | Financial spreadsheets and accounting | Longitude Labs Inc., Carnegie Mellon University, Cornell University | GPT-5.5 | 49.6 criteria-weighted score % (overall) | 2026 | 11 public tasks; 120 held out for the main evaluation | CC-BY-NC-4.0 (public data, HF); paper CC BY-NC-SA 4.0 | none | Not stated |
| CAIA (Crypto AI Agent Benchmark) | Live financial market data | Surf AI / Cybertino Lab and Princeton University | GPT-5 | 67.4 accuracy % (majority vote, 5 runs) | 2025 | All 178 questions (benchmark.csv on GitHub; dataset on Hugging Face). The release uses anonymized mock tools. | CC-BY-4.0 (data); Apache-2.0 (code) | Not stated | Not stated |
| ClaimBench | Financial Verifiability | Vals AI | d1-ultra | 100 accuracy % | Not stated | 10 public examples | Not stated | Not stated | Not stated |
| CorpFin v2 | Financial filings Q&A | Vals AI | Claude Opus 5 | 73.2 accuracy % | Not stated | 20 questions from 1 document, available on request | proprietary | none | Not stated |
| Deep FinResearch Bench | Analyst Financial research | AI Research, JPMorganChase | OpenAI | 86 overall factual claim rate % | 2026 | none (the paper links no data or code; the professional reports came from a paid Yahoo Finance subscription) | Not stated | none | Yes |
| Excel Modeling Benchmark (EMB) | Financial spreadsheets and accounting | Vals AI | Claude Fable 5.1 | 76.7 accuracy % (v1.1) | Not stated | 1 task | proprietary | Not stated | Not stated |
| FAITH | Financial Verifiability | Asian Institute of Digital Finance, National University of Singapore | Claude Sonnet 4 | 95.6 overall accuracy % | 2025 | Pilot and main splits plus eval code on GitHub | Apache-2.0 (GitHub repo, which says it holds code and data). The paper footnote says CC BY-NC-SA. | none | No |
| Finance Agent Benchmark | Analyst Financial research | Vals AI | Claude Opus 5.5 | 67.8 Partial Credit accuracy % (v2.1) | 2025 | 27 public questions plus the agent harness in the GitHub repo. The v1 paper released 50 public validation questions (CC BY 4.0). | MIT (code, finance-agent-v2 repo). The v1 paper lists CC BY 4.0 for the public data and MIT for the code. | Not stated | Not stated |
| FinanceBench | Financial filings Q&A | Patronus AI, with Contextual AI and Stanford University | Numa 1 | 67 % correct, no filing given (150-question sample) | 2023 | 150 annotated questions, with source PDFs and the paper's human-graded model outputs. The full set is available on request from Patronus AI. | CC-BY-NC-4.0 (data, per Hugging Face card) | via-maker | Not stated |
| FinanceQA | Financial filings Q&A | AfterQuery | o1 | 0.487 accuracy (total, 0-1) | 2025 | 148-row test split on Hugging Face | Apache-2.0 (data) | none | Not stated |
| FinBen | Finance knowledge | The Fin AI (Qianqian Xie, Jimin Huang et al., 34 authors) | 2024 | Datasets, code and results through the PIXIU repo and Hugging Face. Each dataset keeps its own licence. | MIT (code); data licences vary per dataset (MIT, CC BY 4.0, CC BY-NC 4.0, CC BY-SA 3.0, CC BY-NC-SA 4.0) | Not stated | No | ||
| Finch: Finance & Accounting Cowork Benchmark | Financial spreadsheets and accounting | Haoyu Dong et al. (UCAS/Microsoft, South China University of Technology, Zhongguancun Academy, Harvest Fund, Fudan University, Duke University, Hugging Face) | GPT 5.1 Pro | 38.4 workflow pass rate % (human eval) | 2025 | Full set of 172 workflows published on Hugging Face (single 'test' split) | CC-BY-3.0 (data, per HF tag; paper Appendix E says CC BY 3.0 US) | none | Yes |
| FinDeepResearch | Analyst Financial research | National University of Singapore, 6Estates, Asian Institute of Digital Finance, Hefei University of Technology, USTC (OpenFinArena) | FinAgentTest | 41.9 overall accuracy score (0-100) | 2025 | Full set public on Hugging Face, including analytical-structure templates and expert-annotated reference reports | Apache-2.0 (data, HF); paper CC BY 4.0 | open | Yes |
| FinLFQA | Financial Verifiability | New York University; Yale University; NYU Shanghai | GPT-4o | 56.5 evidence attribution F1 | 2025 | Dev and test splits on Hugging Face, plus code on GitHub. The paper says test answers are kept private, which was not checked against the HF files. | MIT (Hugging Face dataset card). The GitHub repo declares no licence. | Not stated | No |
| FinResearchBench | Analyst Financial research | Stepfun; FinStep | Gemini | 76.6 Final Score (0-100) | 2025 | none (no data or code linked) | Not stated | none | Yes |
| FinSafetyBench | Financial advice and safety | Shanghai University of Finance and Economics; Beihang University; Southern University of Science and Technology | LLaMA-3 | 32.3 attack success rate % (financial crimes, English) | 2026 | Instructions only (data/FinsafetyBenchmark.csv) plus judge.py on GitHub, for research use | Not stated | none | No |
| FinSearchComp | Live financial market data | ByteDance Seed (Columbia Business School listed as advisor) | Grok 4 | 68.9 Global subset average accuracy % | 2025 | The paper calls it fully open-source and the dataset is on Hugging Face. The paper does not state a held-out split. | CC-BY-4.0 (data, per Hugging Face card) | Not stated | Yes |
| FinTrust | Financial advice and safety | NYU Shanghai; National University of Singapore; Yale University; NYU Center for Data Science | o4-mini | 1 safety harmfulness score (1-10, GPT-4.1 mini judge) | 2025 | Full dataset on Hugging Face. The GitHub repo has per-dimension scripts and 100-case samples. | MIT (Hugging Face dataset card). The GitHub repo declares no licence. | none | No |
| FrontierFinance | Analyst Financial research | Samaya AI | Samaya Sterling | 58.9 rubrics qualification rate % | 2026 | All 220 queries and rubrics, on Hugging Face (gated with automatic approval; you agree not to reshare) | CC-BY-4.0 (data); Apache-2.0 (grader code) | Not stated | Yes |
| InvestorBench | Financial trading and portfolios | Stevens Institute of Technology; The Fin AI; Columbia University; Harvard University | Qwen2.5-72B-Instruct | 46.2 cumulative return % | 2024 | Paper says datasets and environments are open source; no repository URL found on any fetched primary page | MIT (data, per paper's ethics statement) | none | Not stated |
| IPO Finance Agent | Financial filings Q&A | Mostapha Benhenda | GLM-5.2 | 79.8 mean rubric score % | 2026 | 70 questions on the SpaceX (SPCX) S-1. The rest, on further S-1s across sectors, are private. | MIT (code, per GitHub repo) | Not stated | Not stated |
| K-FinHallu | Financial Verifiability | KAIST AI; Financial Tech Lab, KakaoBank Corp. | Gemini 2.5 Flash | 0.860 binary detection F1 | 2026 | The paper says the benchmark is released, but no repository or dataset URL was found | CC-BY-NC-4.0 (stated in the paper; no release link found) | none | No |
| KTD-Fin | Financial trading and portfolios | Stepfun; Tsinghua University (with FinStep, Shanghai Jiao Tong University, Adelaide University) | Qwen3.6-Plus | 85.3 total return % | 2026 | none found (paper says it is released as a template, but gives no URL) | Not stated | none | Not stated |
| LiveTradeBench | Financial trading and portfolios | University of Illinois Urbana-Champaign (U Lab) | GPT-4.1 | 6.3 cumulative return % | 2025 | Code and PyPI package; live site | PolyForm Noncommercial 1.0.0 (code); separate commercial licence | maker-only | Not stated |
| Market-Bench | Financial trading and portfolios | AfterQuery | Grok 4 | 443.2 mean MAE | 2025 | Not stated | Not stated | Not stated | Not stated |
| Moneda AdviceBench | Financial advice and safety | Moneda | 2026 | A public sample of about 1 item in 10; the rest is a private test set, the only one ranked | Items CC BY 4.0, harness MIT | Open harness at release | Yes | ||
| Moneda CiteBench | Financial Verifiability | Moneda | Numa 1 First results, one system scored | 63.3 Factual share of linked claims % | 2026 | Yes, companies rotate each release | Items CC BY 4.0, harness MIT | Open harness at release | Yes |
| Moneda FilingsBench | Financial filings Q&A | Moneda | Numa 1 First results, one system scored | 91.6 % correct (private split) | 2026 | Yes, about 1 question in 10; the rest are private and only those are ranked | Items CC BY 4.0, harness MIT | Open harness at release | Yes |
| Moneda GroundBench | Live financial market data | Moneda | 2026 | A public sample of about 1 item in 10; the rest is a private test set, the only one ranked | Items CC BY 4.0, harness MIT | Open harness at release | Yes | ||
| Moneda KnowledgeBench | Finance knowledge | Moneda | 2026 | A public sample of about 1 item in 10 with a randomised-answer twin to detect training on it; the rest is a private test set, the only one ranked | Items CC BY 4.0, harness MIT | Open harness at release | No | ||
| Moneda MarketBench | Live financial market data | Moneda | 2026 | A public sample of about 1 item in 10; the rest is a private test set, the only one ranked | Items CC BY 4.0, harness MIT | Open harness at release | Yes | ||
| Moneda PortfolioBench | Financial trading and portfolios | Moneda | Numa 1 First results, one system scored | 96 Portfolio tasks, % correct | 2026 | Yes; forecasts resolve after they are asked, so there is nothing to memorise | Items CC BY 4.0, harness MIT | Open harness at release | No |
| MortgageTax | Financial filings Q&A | Vals AI | Claude Opus 5 | 72.1 accuracy % | Not stated | No public download; validation sets are available on request or for purchase, the test set is never shared | proprietary | Not stated | Not stated |
| Open FinLLM Leaderboard | Finance knowledge | FinAI / SecureFinAI Lab (Columbia) with the Linux Foundation, FINOS and Hugging Face | GPT-4o | 46.0 balanced score (mean of text/vision/audio averages) | 2025 | Scores (scores.json) and the Space source are public. Datasets are on Hugging Face under their own cards. | Apache-2.0 (Space code); the paper says the project follows the OpenMDW License | via-maker | No |
| PolyBench | Financial trading and portfolios | Sichuan University | MiMo-V2-Flash | 17.6 Confidence-Weighted Return (CWR) % | 2026 | Pipeline code on GitHub; dataset via a OneDrive link in the README | Not stated | none | Not stated |
| PortBench | Financial trading and portfolios | AgenticFinLab; The Hong Kong University of Science and Technology (Guangzhou); Yantai Research Institute of Harbin Engineering University | GLM-5.1 | 0.384 CEPS (0-1) | 2026 | Code on GitHub; dataset as a Hugging Face collection; raw experimental results as a Hugging Face dataset | Apache-2.0 (code); MIT (PortBench-RawData results dataset on HF) | none | Not stated |
| PRBench Finance | Financial advice and safety | Scale AI | Muse Spark 1.3 | 59.5 weighted rubric score (0-100) | 2025 | All tasks and rubrics on Hugging Face (not gated) | CC-BY-4.0 (data) | Not stated | Not stated |
| Ramp Accounting Bench | Financial spreadsheets and accounting | Ramp Labs | Gemini 4 Argon | 61.1 mean reward % (criteria, mean of 3 attempts) | Not stated | One sample task is shown on the page; the task set is not published | Not stated | Not stated | Not stated |
| Reasoning Models Ace the CFA Exams (CFA mock-exam evaluation) | Finance knowledge | SecureFinAI Lab, Columbia University (with RPI and UNC Chapel Hill) | Gemini 3.0 Pro | 97.6 Level I accuracy % | 2025 | none (official CFA Institute practice packs and paywalled AnalystPrep mocks) | Not stated | none | No |
| StockBench | Financial trading and portfolios | Tsinghua University; Beijing University of Posts and Telecommunications | Qwen3-235B-Think | 2.5 final return % | 2025 | Code on GitHub; results on the project page | Apache-2.0 (code); paper says academic, non-commercial research use only | maker-only | Not stated |
| Tax Agent Bench | Finance knowledge | Vals AI | Qwen 3.8 Max 0902 | 49.7 accuracy % (all-pass x citation factor) | Not stated | A public split exists (it includes example P-004), but its size is not stated | proprietary | Not stated | Not stated |
| TaxEval v2 | Finance knowledge | Vals AI | Muse Spark 1.2 | 80.4 accuracy % (v2) | Not stated | No public download; validation sets are available on request or for purchase | proprietary | Not stated | Not stated |
| Vals Finance Index | Analyst Financial research | Vals AI | Gemini 4 Carbon | 66.3 index accuracy % | 2026 | none (inherits the public samples of its components) | Not stated | Not stated | Not stated |
| τ³-Banking | Finance knowledge | Sierra | Qwen 3.8 Max | 55.2 pass^1 % | 2026 | Code and tasks in the open tau2-bench repository | MIT (code) | Submissions to the τ-bench leaderboard | Not stated |
