Claude Code stock analysis: how we run it on free data

Updated on 2026-10-27

Educational only, not investment advice.

This is how we do Claude Code stock analysis on bests.one: the workflow, the free data, the checking pass, and where it went wrong. It is a method, not a stock pick. Everything below comes from the run records in our own research notes, and none of it says whether any stock is worth buying.

What we actually ran

Between 2026-10-20 and 2026-10-21, Claude Code produced five research notes for bests.one. Each note opens with a "How this report was made" box, and the table below copies only what those boxes and the notes' source lists say. Where a note did not record something, the cell says so.

The cost column is what each run reported for its own compute, not a price list. For current Claude prices, see Anthropic's pricing page.

Three things stand out from the table. First, the runs were short: the three that recorded time took about 2 to 4 minutes. Second, most of the effort sits in checking, not writing. The DCF note checked 17 figures from filings and 6 web figures separately, and the earnings review checked 24. Third, the two notes that relied most on outside material, the SpaceX review and the backtest, did not record time or cost, so we have no comparison for them.

Every published note was also reviewed by a person at bests.one before it went live. The checking process is described on How we test.

The setup

The stack has three parts.

Claude Code. Anthropic's docs describe it as an agentic coding tool that reads your codebase, edits files and runs commands, available in the terminal, IDE, desktop app and browser. Install it from the official Claude Code docs; we don't repeat the steps here because they change.

Anthropic's open-source finance skills. The first three notes say they used "Anthropic's open-source finance skills, adapted by bests.one." These come from the anthropics/financial-services repository, released under the Apache 2.0 license. Its core financial-analysis plugin covers comps, DCF, LBO and three-statement models, and the equity-research plugin adds earnings notes and model updates. Anthropic's install tutorial says the plugins work in both Claude and Claude Code. The repository itself says nothing in it is investment advice and that outputs are drafts for review by a qualified professional. Our adapted versions are not published on this page.

A free-data tool server. The notes call it us-free-finance, an MCP server (a tool interface the agent can call) that reads SEC EDGAR XBRL facts and filings, FRED series and unofficial Yahoo Finance prices. The comps note shows it in use, for example a tool that pulls peer candidates from a company's proxy statement and 10-K. If you want to build the same kind of tool, the endpoints it relies on are public and documented in our SEC EDGAR API guide.

The backtest note doesn't mention the finance skills or the tool server. It says Claude Code wrote and ran its own Python on a Mac mini against two free datasets.

Step by step: from ticker to checked note

This is the order the notes follow. Use TICKER as a placeholder until the whole loop works.

1. Pick a report type and a ticker

Decide on the output before you start. A comps table needs a peer set, a DCF needs a projection period and terminal assumptions, an earnings review needs the release and the prior guidance, and a backtest needs a membership history. Ask for one report type per run.

2. Pull the data from free primary sources

Point the agent at filings first. Our notes use SEC EDGAR filings and XBRL for every historical figure, FRED for rates and inflation, and Yahoo Finance for prices, labeled as unofficial. The SEC side is covered in the EDGAR guide linked above, so it isn't repeated here.

3. Let the agent write and run the code

Have Claude Code compute the tables in code it runs, not in prose. The DCF note, for example, keeps its full scenario rows in a CSV file, and the backtest publishes its raw output as a CSV, so the numbers can be audited.

4. Tag every figure by source

Our notes separate figures into SEC-verified (from filings or XBRL) and [WEB] (analyst consensus, call commentary, news), with the provider, date and link for each web figure. The comps note used no web figures at all. When a web figure came only from a search snippet, the NVIDIA note labeled it "not verified, not relied on."

5. Run a second checking pass

A separate AI pass compares a list of figures in the draft against the original filings and FRED pages. The notes record the count, such as 17 of 17 or 24 of 24, and say that everything matched "except where noted." Web figures are checked against the pages they cite.

6. Review it yourself before publishing

A person reads the draft, the data gaps and the known limitations before publishing. Our notes keep their limitations sections in the published version instead of trimming them.

Where it went wrong

The notes record their own mistakes and gaps. A few examples:

• A debt item the tool missed. In the comps note, ACIW's $14.1M of other current notes payable was not included in its debt, because the free-data tag was not mapped yet. The note judged the effect on the multiples negligible, but it is the kind of gap you only find by reading the balance sheet.
• Zero debt that was never checked. The same note treated debt as zero for three companies where no debt tag was found, and says only one of them was corroborated by total liabilities.
• Two methods, very different answers. In the DCF note, the perpetuity-growth and exit-multiple terminal-value methods gave very different results. The note also flags excluded finance-lease additions in capex as a material limitation.
• Unverified consensus. The NVIDIA note could read only one consensus source, so no two-provider cross-check was possible, and some figures came from search snippets it did not rely on. It also notes that NVIDIA changed its non-GAAP definition, so the comparison basis may differ.
• Missing history. The backtest says about 125 former S&P 500 members had no Yahoo history, so it relied only on each year's top stock.

What it can't do

• No consensus or forward data from free sources. The comps note says free SEC and Yahoo data have no forward multiples and no consensus estimates, so none were provided. Anything forward-looking has to come from the web, tagged and checked, or be left out.
• It gets things wrong. The gaps above are from runs that looked finished. That is why the checking pass and the human review are part of the method, not optional extras.
• It is not investment advice. These are educational samples of a method. Nothing here or in the linked notes recommends buying, selling or holding any security.

FAQ

Can Claude Code analyze stocks?

It can draft stock research such as comparable companies tables, DCF models and earnings reviews if you give it data tools. In our runs it wrote and ran its own code on free SEC, FRED and Yahoo Finance data. A separate checking pass and a human review were still needed before anything was published.

What data does it need?

At minimum, company filings and financial facts (we use SEC EDGAR filings and XBRL), macro series such as FRED, and a price source. Free sources do not include consensus estimates or forward data, so those either come from the web, tagged and checked, or are left out.

Is this investment advice?

No. This page and the research notes it links to are for education only. They describe a method and do not recommend buying, selling or holding any security.

Related on bests.one

• SEC EDGAR API in Python: free financial data for AI agents
• Read a 10-K with AI using the SEC EDGAR API
• Can AI pick stocks? How we grade an AI analyst
• How we test

Sources: Claude Code overview (Anthropic docs), anthropics/financial-services on GitHub, Install financial services plugins for Claude (Claude Academy), EDGAR Application Programming Interfaces (APIs), fja05680/sp500 on GitHub. Checked on 2026-10-27.