If you’re training a model to do finance work, or buying data for a team that is, the first thing to sort out is which kind of financial data you mean. Market data teaches a model what prices did. The work of analysts, bankers and accountants teaches it how to do the job, and that second kind is what labs are short of. Read on to find out the five kinds of financial data labs train on, why finance leads their lists, what to check before you license any of it, and where to get it.
The five kinds of financial data
People use “financial data” for five different things, and each comes with its own rights question.
| Kind of data | What it teaches a model | The rights question |
|---|---|---|
| Market and price data | What prices, rates and volumes did | Vendors and exchanges own it, and their licenses often rule out redistribution or derivative use |
| Financial news and researchFinancial news and research | How events get reported and interpreted | The publisher owns it, and a subscription isn’t a training license |
| Public filings and disclosures | What companies report | Open to everyone, so every lab already has it |
| Firm work product (models, memos, diligence files, workpapers) | How finance work gets done, argued and checked | The firm owns its own work, and client, vendor and non-public material has to come out first |
| Personal financial records | How individuals save, borrow and spend | Regulated, and mostly not licensable |
The fourth row is the scarce one. Prices and filings are easy to find. The model with its assumptions, the memo that argued for it and the result a year later sit inside firms and were never published.
Why finance leads the list
Buyers tell us they pay a premium for full company datasets from industries built on high-value knowledge work, and finance is the first industry they name. The work is expensive, it takes years to learn, and almost none of it is on the public internet. A textbook explains what a discounted cash flow model is. A firm holds real ones.
OpenAI’s GDPval benchmark, published in September 2025, shows which jobs labs measure their models against. It tests models on real tasks from 44 occupations across the nine sectors that each contribute more than 5 percent of US GDP, and finance and insurance is one of the nine. Five of the 44 occupations sit inside it, with accountants and auditors listed under professional services. OpenAI built each task from the work of a professional with 14 years of experience on average, and a task takes an expert about seven hours. The deliverables are spreadsheets, documents and slide decks.
Frontier labs like OpenAI and xAI and enterprise AI companies like Cohere and WRITER are building models for this kind of work, and labs have hired former bankers and analysts to write and grade finance tasks. Agents raise the stakes again, because an agent has to finish the task inside the tools. Our guide to training data for AI agents covers the workflow records agents learn from.
Finance work can be checked, which is why it trains well
Buyers tell us that the more objectively a piece of work can be checked, the more readily they license it, and they point to financial models as the example. A model has inputs, formulas and an answer. A forecast can be compared with what happened. A reconciliation ties out or it doesn’t.
That gives a training team something most knowledge work can’t. The record carries its own outcome, so a model can learn from whether the work was right as well as from how it was written. The value is highest where the records connect, like a financial model linked to the memo that used it and the thread where a partner questioned it. Our guide to legal training data for AI covers the same idea for legal work.
What to check before you license financial data
Financial data has more ways to go wrong than most, and the problems tend to surface late. Six checks catch them early.
- Is third-party data hiding inside it? A firm’s spreadsheets are full of market data pasted in from a vendor, and that data belongs to the vendor. The data a firm made itself may be licensable. The data it bought usually isn’t.
- Does it hold non-public information about public companies? That can’t be licensed to anyone.
- Have names and personal details been replaced with pseudonyms to a documented standard? Client identities, account details and personal financial information are regulated.
- Was any of it received under a confidentiality agreement? Deal materials from a counterparty stay out.
- Does the license cover commercial model training in writing? A subscription or a research license doesn’t, and your legal team will ask.
- Can you follow a piece of work from start to outcome? Ask for the written explanation next to the model, and for records that link to each other.
This isn’t legal advice, and your counsel should review the scope of any financial dataset. Our guide to how AI data licensing deals work covers the terms that get negotiated, and our guide to rights-cleared training data covers what clearing a dataset takes.
Where to get financial training data
You have four routes, and they get you different rows of the table above.
- License market data from the vendor or exchange. Check the license for model training before you assume it’s covered.
- License news and research from the publisher. Publishers have signed public training deals, and the license has to name training as a permitted use.
- Pay experts to produce it. Former bankers, analysts and accountants write tasks and solutions to your brief. You control the format, and every hour is billed. Our guide to AI data collection companies covers the vendors that work to order.
- License real work from the firms that did it. Funds, advisory firms and accounting practices hold years of models, memos and workpapers with real outcomes, and they can license what they own once the client, vendor and non-public material is out. Our guide for financial firms licensing their data covers what a firm can license and what has to stay out.
One public price exists for a company’s operating history that included finance records. Spirit Airlines’ records drew bids of 10 million and 12.5 million dollars in bankruptcy court, and the package included the airline’s finance and revenue-management records. Our breakdown of the Spirit Airlines data deal covers what was in it, and our guide to who buys company data for AI covers the five kinds of buyer.
Where Troveo fits
Troveo is a data licensing marketplace, and we’ve paid rights holders more than 20 million dollars. Financial firms can license their own work through us (financial models and valuations, investment memos, due diligence files, client reports and accounting workpapers). Before anything is exported, the firm takes out vendor data, non-public information and anything under a confidentiality agreement, and we replace names and personal details with pseudonyms and document what’s included in a manifest. You sign one agreement with us and never negotiate with each firm. Tell us what finance work your model needs to learn and we’ll tell you what we hold.