How we source every claim

The numbers are not written by the AI

This is the part most people assume is impossible to get right, and the fix is unglamorous. Every price, every percentage move and every row of the portfolio table is rendered directly from market data by code. The language model never transcribes a number, because a model asked to copy a figure into a sentence will eventually copy it wrong.

What the model writes is the explanation: what happened, who reported it, and why it matters to a position you hold. That is the part worth having a writer for, and it is the only part it does.

Every claim carries its source

When a briefing says a company was downgraded, or that a short-seller report triggered a selloff, the sentence itself links to the article that reported it. Not a list of sources at the bottom. Not a vendor logo. The specific claim, linked to the specific story, so you can read it yourself in one click and decide whether you agree.

Go and check one: open a real briefing and click any link in the writing. Every dated briefing we have ever sent from the demo portfolio is public, going back months.

What stops a briefing from being sent

Before any briefing reaches an inbox it is checked against the data it was built from. These are not warnings that get logged somewhere. Any one of them blocks delivery:

• A link that was not in the source data

  Every URL in the briefing is checked against the articles the pipeline actually fetched. A plausible-looking link the model composed itself is not a citation, and it stops the send.

• A price that does not match the market data

  Any dollar figure in the prose is matched against the quote it refers to, within five cents or one tenth of one percent. Beyond that, the briefing does not go out.

• A ticker you do not hold

  If the writing attributes price data to a symbol that is not in your portfolio, your watchlist, or the source payload, it is treated as invented.

• An earnings date that is not on the calendar

  Claims that a company reports on a given day are checked against the earnings calendar. Unsupported ones block the send.

• A factual claim with no citation

  News-derived statements are checked sentence by sentence for a matching source link. Quoted phrases of six words or more must carry one.

• A holding that went uncovered

  If a position that was selected for write-up is missing from the finished briefing, the briefing is incomplete and is not sent.

• Anything that leaked from the machinery

  Raw URLs, unrendered citation markers, prompt scaffolding, and the model talking about itself ("I don't have data for...") are all treated as defects rather than shipped.

• A briefing that was cut off

  Output that ends mid-sentence or is missing its closing section is truncated, not finished, and is refused.

Where the data comes from

Prices and company fundamentals from Financial Modeling Prep and Polygon. Company news, earnings calendars and economic calendars from Finnhub. Treasury yields and policy rates from the Federal Reserve Bank of St. Louis. Company filings from SEC EDGAR.

Naming vendors is table stakes and nearly everyone does it. It tells you where the raw material came from, not whether the sentence you are reading is true. That is what everything above is for.

What we do not do

• No recommendations. We do not tell you what to buy or sell. Holdings Daily is not a registered investment adviser and none of this is financial advice. See the full disclaimer.
• No ads, no sponsors, no affiliate links. Ever. The product is entirely subscriber-funded, which means nobody is paying us to put a ticker in front of you.
• No price targets, no forecasts. The briefing reports what happened and what was scheduled. It does not predict.
• No selling your holdings data. What you own is the input to your briefing and nothing else. The public pages on this site come from a demo portfolio, never from a subscriber.