Earnings call sentiment analysis: what tone really tells you
Sentiment analysis on earnings calls has a strange reputation: quants have traded on it for two decades, while most private investors have never used it once. The idea is simple, the language of a call carries information beyond its literal claims, and it happens to be true, with published research behind it. What changed recently is access: tone used to require dictionaries and models of your own, and now it is a question you can ask in a chat. Worth understanding what the answer actually measures.
What sentiment on a call means
At its simplest, sentiment is the balance of positive against negative language, measured with finance-specific vocabularies, because ordinary sentiment dictionaries misread this domain: "liability", "gross", and "restructuring" are not emotional words here. The classic Loughran-McDonald word lists were built precisely because general-purpose sentiment tools called half of every annual report negative.
The more useful modern version is less about counting words and more about register: hedging density ("should", "hopefully", "assuming"), certainty markers, deflection patterns in the Q&A, and the gap between the prepared script's tone and the spontaneous answers. Prepared remarks are written by committees; the Q&A is where tone stops being a choice.
What the research supports
Findings that have replicated: negative-tone spikes relative to a company's own baseline precede analyst downgrades and weak follow-on quarters more often than chance; hedging language around guidance correlates with subsequent misses; and Q&A tone carries more signal than prepared-remarks tone, consistent with the committee-versus-spontaneous logic.
What does not hold: absolute tone as a cross-company signal. Some CEOs are permanently exuberant, some CFOs are professionally gloomy. Tone only means something against the same speaker's own history, which is why the delta question beats the level question every time.
Asking tone questions properly
- "How did management's tone change versus the previous quarter, with the passages that show it?" The baseline-relative form.
- "Where did management hedge in this call, and was the hedging new?" Density and novelty of caveats.
- "Compare the confidence of the prepared remarks with the Q&A answers." The committee-versus-spontaneous gap.
- "Which questions produced the most defensive answers?" Tone, located.
- "Across the sector, who sounded most and least confident about the same demand environment?" Tone as a cross-check.
Tone as evidence, not verdict
The right way to hold a tone finding is as a flag, not a conclusion. "Hedging around the margin outlook doubled versus last quarter" does not say sell; it says look at the margin line hard, and ask what changed in the business to change the language. Sometimes the answer is benign, a new CFO with a careful style. Sometimes it is the first visible edge of a problem that will have a name in two quarters.
This is also why tone answers must come with the underlying passages. An adjective like "cautious" is a model's judgment; the six quoted sentences behind it are something you can weigh yourself, and disagree with. earnings.chat returns the quotes on principle, because a sentiment read you cannot audit is just someone else's vibe.
Where it fits in a workflow
Tone is a second-pass instrument. First pass: figures, guidance, the Q&A topics. Second pass, on the names that matter: how was it said, and how does that compare to last quarter and the peer group. As a standing habit, tone deltas on your core positions each season, it costs a question per name and occasionally hands you the earliest warning you will get. As a standalone oracle, it will disappoint. Instruments beat oracles.
252,000+ earnings calls in the knowledge base, answers with verbatim quotes and sources, new calls within minutes.
Open earnings.chat