Skip to content

How stocks-llm works

By stocks-llm · Last updated 2026-10-17

Research stocks in your own words.

What stocks-llm is, and is not

stocks-llm is an independent research tool for public companies. You ask in plain English and get real companies back, with the figures behind each one and the date of every figure.

  • It is not a broker: it does not hold money, place orders or route trades.
  • It is not an adviser: it does not know your circumstances and does not tell you what to buy or sell.
  • It is paid by no company it covers. There is no advertising on the site and no affiliate links: no listing, ranking or mention is bought.

Where the figures come from

Company filings and the financial-statement facts inside them come from SEC EDGAR (filings and XBRL). Insider transactions come from SEC Forms 3, 4 and 5. Trades by members of the U.S. House come from the Office of the Clerk of the U.S. House. Short interest comes from FINRA. Interest rates, inflation and currency rates come from FRED (Federal Reserve Bank of St. Louis). For UK companies, registry details come from Companies House and regulatory filings from the FCA National Storage Mechanism.

Prices, and some fundamentals and analyst figures, come from market data: delayed daily closes, never real-time. A figure from market data is labelled as such and is never presented as a public record.

Every figure carries its as-of date. Where we do not hold a figure, or hold one we cannot stand behind, the page says so (“n/a”, with the reason) rather than estimating it.

What the language model does, and does not do

A language model writes the answer: it reads your question, chooses which stored records and screens to look up, and writes the prose and the tables that describe what came back. It also drafts the short company descriptions, over the same stored records. The figures on the company cards beside an answer are read from the stored record, not written by the model. The figures the model writes into a table or a sentence are meant to be copied from the records it looked up, but apart from the checks below they are not machine-checked: a wrong P/E or RSI in a table the model wrote can reach you. That is why every company card links to its record, with each figure’s date, and the record is the figure to trust.

Before an answer reaches you, these checks run over it:

  • A share price the model prints in a table’s Price column is compared with the price we hold for that company in the same answer. A price that disagrees, a Price cell holding something that is not a price (such as a market cap), or a market-cap cell holding the share price is replaced with “n/a”, and the answer states the price we hold and its date.
  • A technical reading (such as the distance from a 52-week high) that the model prints at 100 times the value we store — a unit slip — is corrected to the stored value in a table, and a sentence that repeats the slip about a named company is corrected or removed. Readings that are wrong in any other way are not caught by this check.
  • A sentence saying a company is not covered when it is in our catalog is removed.
  • A sentence saying a company’s dividend payout ratio is not available, when the lookup returned it, is removed, and the figure is added from the company’s filings.
  • A sentence that credits a company’s debt-to-equity or return on invested capital to its filings, when the figure we hold is market data, is removed, and a line stating each of those figures’ actual basis and date is added.
  • The figures the model is given carry their own as-of date and basis (which filing, or market data), so it can say when a number is from. This is what the model is handed, not a check on what it prints.
  • A stored figure that cannot be true — a margin above 100%, a ratio on negative equity, a return across an unadjusted stock split — is withheld with its reason instead of being shown, ranked or handed to the model.

These checks catch the specific errors they were built for and nothing else. They do not make the answer infallible.

How errors are corrected

  • Tell us. The feedback form on every page is the way to report a wrong figure or a broken page: . There is no e-mail address; the form is read.
  • Figures. When a source corrects a filing, or we repair a price series (for example across a stock split), the figure is recomputed and shows the date of the record it now comes from. A corrected figure is never back-dated.
  • Posts. A blog post that turns out to be wrong is retracted: its address then answers “This post has been retracted” instead of showing the old text.

How often each kind of data is refreshed

DataRefreshed
US prices (previous session’s close)Daily, from 01:00 UTC, once the session is published
UK prices (LSE close)Daily at 16:00 UTC; the London session usually arrives one to two days later
Technical indicators and technical screensDaily at 02:30 UTC, after prices
Top 10 listsDaily from 03:00 UTC
New SEC filingsDaily at 05:30 UTC
Financial statement facts (SEC XBRL)Weekly, Sunday 12:00 UTC
Insider transactions (SEC Form 3/4/5)Daily at 07:00 UTC, in rotation
U.S. House trading disclosuresDaily at 13:15 UTC
Short interest (FINRA, published twice a month)Checked daily at 14:30 UTC
Analyst consensus (market data)Daily at 10:00 UTC, in rotation
Dividends (market data)Monthly, on the 3rd
Index membership (S&P 500, Nasdaq-100, Russell)Quarterly

A value that has not been refreshed on schedule shows its older date, so a stale figure reads as stale.

Informational only — NOT financial advice. stocks-llm is a research aid, not a recommendation to buy or sell anything. Verify figures against the primary source before acting on them.

About stocks-llm · Privacy · Home