Earnings call sentiment is mostly noise. Here is the part that is not.

Sentiment scoring on earnings calls is the most frequently cited and least useful piece of quantitative text analysis in equity research. It is easy to compute, it produces a number that goes up and down, and the number is mostly measuring something other than what people think it measures.

What you are actually scoring

Prepared remarks are a written document read aloud. They are drafted by investor relations, reviewed by legal, rehearsed, and in many cases informed by an explicit awareness that the transcript will be machine-read within minutes of the call ending. Running a sentiment model over them measures the competence and house style of a communications function.

This shows up in the data as stability. For most issuers, prepared-remarks sentiment is close to constant quarter over quarter, and the variance that does exist correlates with changes of CFO and changes of IR agency about as readily as it correlates with anything happening in the business.

The Q&A is a different document

The question-and-answer portion is semi-adversarial and largely unscripted. It is the only part of the call where somebody other than the company decides what gets discussed. And the informative content in it is structural rather than tonal.

Analyst persistence

The same question asked three times by three different analysts is a statement about what the sell side does not believe. It is a stronger signal than any tone measurement, because it is produced by people who are paid to know which answer was evasive and who have to file a note afterwards.

Answer typology

Classify each answer into one of four types and the ratio tells you more than the sentiment does:

• Quantified. A number, a range, or a rate. The company is willing to be measured against it next quarter.
• Directional. Better, worse, stable. Committal about sign, not magnitude.
• Deferred. Pushed to a future event — the analyst day, the next quarter, the close of a transaction.
• Non-answer. Restates the question as a strategic priority.

A company that answered six questions with numbers last quarter and two this quarter has changed its posture, regardless of how warm the adjectives were.

Absence as signal

The most useful thing a language model can do with four consecutive transcripts is not score them. It is notice what stopped being said.

Metrics that management used to volunteer unprompted and now only produce when asked. Targets that were reaffirmed every quarter for two years and were not reaffirmed this time — with nobody announcing that they had been withdrawn, because they have not been withdrawn, they have simply stopped being mentioned. Qualifiers that appear in a sentence that is otherwise word-for-word identical to last quarter's:

No headline is written about a subordinate clause. Nothing moves on the day. And a quarter later the qualifier turns out to have been the entire content of the sentence.

What we report, and what we do not

Call Notes report what management guided to, what changed from the prior quarter's language, which questions analysts pressed on repeatedly, and which ones did not get a straight answer. Sentiment appears as a diagnostic — a coarse indicator alongside the hedging-term count — and never as a score you are invited to trade on.

The honest caveat is that this is a qualitative read performed by a machine, and it is wrong sometimes too. Which is why every observation in a Call Note links to the timestamp in the transcript where it came from. Twelve seconds of listening resolves most disagreements.