INSIGHTS · AI × EQUITIES

How Language Models Read Earnings Calls and Filings

BY FOXITON RESEARCH·6 MIN READ

Every quarter, listed companies publish thousands of pages of transcripts, reports and filings. Most of what moves a stock is in there somewhere — the challenge is reading all of it, quickly and consistently. That is exactly where language models earn their place.

The problem: too much text, too little time

An earnings season produces far more material than any analyst team can read closely. A single company releases prepared remarks, a question-and-answer session with analysts, a results presentation and detailed filings. Multiply that across a sector and the important details — a softer tone on guidance, a new risk factor, an evasive answer — are easy to miss.

Language models change the economics of reading. They can process every document in full, apply the same criteria to each one, and flag what deserves a human’s attention.

What a model actually extracts

At Foxiton we treat a transcript or filing as data, not just text. A well-designed pipeline pulls out several layers of information:

  • Facts and figures — revenue, margins, guidance ranges and the numbers management chooses to highlight.
  • Tone and confidence — whether language around demand, pricing or costs has become more cautious or more assertive than last quarter.
  • Changes over time — new risk factors, removed sentences and shifts in emphasis between one filing and the next.
  • The Q&A — how directly management answers analysts’ questions, which is often more revealing than the prepared script.
The most useful signal is rarely a single word. It is a change in how a company talks about the same thing.

From text to signal

Raw model output is not a signal on its own. To be useful in equity research, each extracted feature has to be compared against the company’s own history, its peers and the market’s expectations. A cautious tone from a company that is always cautious means little; the same tone from a usually confident management team can mean a lot.

Only once those features are measured consistently can they be combined with price, volume and fundamental data — and tested to see whether they carried any information about what happened next.

Where language models fall short

Language models are powerful readers, but they are not infallible analysts. They can misread sarcasm or hedged language, over-weight words that sound negative but are routine, and occasionally state things the source text does not support. Financial language is also full of boilerplate that changes little from year to year.

That is why we keep the model’s job narrow and measurable: extract, compare and score — then validate every feature against real market outcomes before it is trusted.

Key takeaways

Language models make it possible to read every earnings call and filing in full, with consistent criteria.

The value lies in changes — in tone, emphasis and disclosure — not in individual words.

Extracted features only become signals after they are compared with history and tested against market outcomes.

THIS ARTICLE IS FOR INFORMATION AND EDUCATION ONLY. IT IS NOT INVESTMENT ADVICE OR A RECOMMENDATION TO BUY OR SELL ANY SECURITY.

NEXT ARTICLEMarket Sentiment from News: Signal or Noise?

GET OUR RESEARCH NOTES

A SHORT DIGEST OF OUR LATEST FINDINGS IN AI RESEARCH AND EQUITY MARKETS.

SUBSCRIBE
MORE INSIGHTS