Turning Earnings Call Transcripts into Quant Signals: Five Practical Features

Every quarter, thousands of listed companies produce the same artifact on a predictable schedule: an earnings call. The result is a remarkably well-behaved text corpus - recurring speakers, a stable format, comparable sections, and statements made on the record by the people who actually run the business. For a systematic team with even modest NLP tooling, that regularity is the whole appeal.

Below are five transcript features that appear repeatedly in the quant literature, each of which you can prototype against the EarningsAPI corpus - 250,000+ calls with 11.9M role-tagged speaker segments across 12,728 companies - in a single afternoon.

1. Theme counts across the cross-section

Start with the bluntest instrument: how many companies in your universe mentioned a given theme this quarter? Track that count over consecutive quarters. A count that keeps climbing is the classic sign of a narrative gaining institutional traction; a count that rolls over often flags a fading thesis before prices reflect it.

The mechanism is incentive alignment. Executives choose their prepared-remarks talking points based on what they believe investors want to hear. When many management teams independently start (or stop) raising the same topic, that is coordinated belief being expressed in text.

curl 'https://earningsapi.io/api/v1/search/by_ticker?q=agentic+AI&tickers=AAPL,MSFT,GOOGL,...&date_from=2026-01-01' \
  -H "X-API-Key: $KEY"

Mentions come back grouped per ticker, so a quarter-over-quarter diff gives you trend velocity directly.

2. Company-relative sentiment change

Raw sentiment scores are hard to compare across firms - every management team has its own baseline optimism. What separates cleanly is the delta: score the current call against the same company's trailing four-quarter average. Fetch the last five calls per name, run sentiment over the prepared remarks only, and use the deviation from baseline as your feature.

Restricting to prepared remarks matters. That section is fully scripted by management, which makes it a consistent measurement surface. The Q&A is shaped by whichever analysts showed up - interesting, but a different signal (next section).

3. Friction in the Q&A

Measure two rates per call. First, the fraction of analyst questions containing challenge markers (but, however, concern, clarify, walk us through, bridge that). Second, the fraction of management responses containing evasion markers (we'll come back to that, not commenting on, as we discussed last quarter, let me hand it to). Together they approximate how hard the questions were and how comfortable management was answering them.

Role-tagged speaker segments make this a filtered query rather than a parsing project:

GET /api/v1/search?q=clarify&type=speakers&speaker_type=analyst&ticker=NVDA

In backtests this feature tends to lead price by roughly one quarter - a factor-portfolio input, not an event-day trade.

4. Hedged guidance language

A sharper cut of the same idea: track how often executives wrap forward-looking statements in conditional phrasing (expect, anticipate, we believe, our view is, should, could). A rising rate relative to the company's own history is a soft caution flag, and it complements traditional estimate-revision factors nicely because it comes straight from management rather than from the sell side.

A phrase search scoped to executive segments does the job:

GET /api/v1/search?q=%22we+expect%22+OR+%22we+anticipate%22&type=speakers&speaker_type=executive&ticker=AMZN

5. Theme pairs, not single themes

Single-topic tracking throws away information. What a company discusses together is often more telling: tariff alongside pricing reads as a pricing-power claim; agentic AI alongside headcount reads as a cost story. Treat theme pairs as bigrams - sweep your universe, score every (theme A, theme B) co-occurrence per call, and rank the pairs.

Why build this on EarningsAPI

250,945 calls covering 12,728 companies - deep enough that cross-sectional ranks stay meaningful in narrow universes.
Every one of the 11.9M speaker segments carries a speaker_type tag (executive, analyst, and so on), which signals #2 through #4 depend on.
search/by_ticker aggregates counts for a whole ticker list in one request, so an index-wide sweep is cheap to backtest.
Plans start at $105 per quarter with self-serve signup - no data-vendor procurement cycle.

Endpoint reference and query syntax are in the API docs.