How to Track Earnings Call Sentiment Across Companies

Earnings-call sentiment is useful as a change detector, not a self-contained investment signal. Score management's prepared remarks, management's Q&A answers, and analyst questions separately. Compare each section with the same company's history and an appropriate peer group. Preserve the passages behind large moves. A vendor label such as positive or negative is not interpretable until the corpus, model, speaker rules, and baseline are known.

This guide synthesizes public research and product documentation checked August 30, 2026. It does not report an AllMind sentiment test or rank platforms. We sell research software that can be configured for transcript analysis, so our conflict and the public-evidence limits are stated up front.

Decide what “sentiment” means for the research question

Do not combine these into one number at the start. Each is a different measurement with a different failure mode.

Finance-specific language is a real requirement

Loughran and McDonald's paper, “When Is a Liability Not a Liability?”, found that a general negative-word dictionary misclassified many words in 10-Ks and developed finance-specific lists. The lesson extends to transcripts: “liability,” “tax,” and “cost” are not automatically bearish without context.

FinBERT approaches the problem with financial-domain language models. The original FinBERT sentiment paper and a separate financial-communications model report improvements on labeled financial-text tasks. Those results do not establish performance on every earnings-call corpus. A production team still needs a transcript-specific validation set.

Published work also supports keeping speakers and sections separate. CFA Institute's summary of managerial tone changes describes an association between increased negativity in management language and future fundamentals after controls. Another CFA Institute discussion of managers and analysts on calls emphasizes differences between prepared management speech and the Q&A exchange. These findings motivate a measurement design. They do not justify a universal trading rule.

Build a stable transcript panel

The denominator matters. Define the universe, event types, history, language, transcript status, and exclusion rules before scoring. The following illustrative panel is specific enough to reproduce:

A platform that returns 47 companies from a 50-company list must show which three are missing. Silently averaging the 47 produces a clean chart with the wrong universe.

Score at passage level, then aggregate

Split each transcript by speaker and section. Preserve the raw passage, speaker, timestamp, topic, model version, and score. Then aggregate using a documented method.

One simple normalized method subtracts negative passage weight from positive passage weight and divides the difference by the number of scored passages. Compare that section result with the company's median over the prior eight calls. The peer-relative view then measures how far the company-level change sits from the peer median, using median absolute deviation rather than standard deviation to reduce the influence of outliers.

This is an illustrative method, not a recommended alpha model. The team should select the baseline after inspecting its data and document the treatment of neutral passages.

The output table should remain auditable:

The example values show the output structure only; they are not measured research results.

Validate classification before looking at returns

Create a labeled set from the intended corpus. Include financial negation, mixed statements, guidance ranges, sarcasm or rhetorical questions, analyst assertions, and answers that decline to quantify. Two domain reviewers should label each passage independently and resolve disagreements.

Report at least:

• class precision and recall, with sample counts;
• confusion matrix for positive, neutral, and negative;
• agreement between human reviewers;
• performance by prepared remarks, questions, and answers;
• performance by sector, language, and transcript status;
• stability across model or prompt versions;
• percentage of passages routed to manual review.

Do this before testing an association with fundamentals or returns. Otherwise the team may optimize a noisy label against a noisy outcome and mistake overfitting for information.

Freeze versions and monitor drift

Store the transcript provider and correction timestamp, segmentation code, dictionary or model identifier, prompt, threshold, and aggregation formula with every result. When any component changes, rerun an overlapping history and quantify the break. Never splice scores from two methods into one time series without a bridge analysis. Monitor the rate of missing calls, unclassified passages, changed speaker labels, and analyst overrides each quarter; a stable headline score can conceal a deteriorating input pipeline.

Use sentiment to route reading

A practical desk workflow ranks changes for review:

1. Score the current call with the frozen model version.
2. Compare each section with the company's history.
3. Normalize against peers reporting in the same season.
4. Retrieve the passages driving the largest moves.
5. Ask whether topic mix, speaker changes, or transcript corrections explain the move.
6. Link any remaining change to a thesis question or model assumption.
7. Record the analyst disposition: material, contextual, or false signal.

The analyst reads the evidence behind outliers. The sentiment number narrows the queue.

What a platform must expose

AlphaSense Company Topics documents topic mentions, relevance, and sentiment over calls. FactSet describes transcript sentiment visualization within Transcript Assistant. Quartr's Claude connector describes access to transcripts, speakers, and first-party IR materials for analysis. These official pages establish available surfaces, not a common sentiment methodology.

Evaluate products on whether they expose:

Our Grids can be evaluated as a way to run the same questions across a company list and retain cited cells, over a transcript layer we license rather than ask a buyer to supply: live earnings calls with real-time transcription, speaker-labeled processed transcripts, and investor presentations and event calendars beside them. We do not ship a proprietary sentiment score, benchmark, or default baseline. A buyer should therefore treat Grids as a configurable workflow surface and bring a validated method, not assume a ready-made signal.

What sentiment cannot establish

Text sentiment cannot show that management is truthful, that a forecast will be met, or that a price move will follow. Acoustic analysis adds another measurement layer and another set of confounders. A positive prepared script alongside pressured Q&A may be analytically interesting, but the interpretation requires the actual passages and company context.

Do not publish performance claims from a backtest without sample dates, universe, transaction costs, rebalancing rules, controls, and an untouched out-of-sample period. This article makes no return or risk forecast.

Research basis and limits

The methodology draws on the Loughran-McDonald paper, the two cited FinBERT papers, CFA Institute research summaries, official product documentation from AlphaSense, FactSet, and Quartr, and our own product pages. Sources were checked August 30, 2026. Competitor features are vendor-reported, and our own are first-party claims. No platform was run under common conditions, no transcript-specific classifier was validated for this article, and no sentiment score, return association, or investment result was independently verified.