About Subtext

Subtext is a research tool for fundamental investors. It applies natural-language processing to earnings-call transcripts, surfacing what management's language reveals beyond the numbers: tone shifts over time, divergence between how the CEO and CFO talk, and the gap between scripted prepared remarks and the unscripted Q&A.

Data

The transcripts come from the kurry/sp500_earnings_transcripts dataset — 33,000+ earnings calls for S&P 500 companies, 2005–2025, segmented speaker by speaker. Each call is split into utterances, and each utterance is tagged with a section (prepared remarks, Q&A question, Q&A response, operator) and a speaker role (CEO, CFO, COO, IR, analyst, operator).

Methodology

Sentiment is measured with the Loughran-McDonald Master Dictionary, a finance-specific word list. For every utterance we count positive, negative, uncertainty, litigious, and constraining words, and compute a net sentiment score of (positive − negative) ÷ total words. Aggregating these by call, by speaker role, and by section produces the charts on each company page.

Speaker roles and section boundaries are inferred with heuristics: roles from how people are introduced in the prepared remarks, and the Q&A boundary from the operator's hand-off to the first analyst. Classification is good but not perfect — treat role-level figures as strong signals rather than exact truth.

Related work

The same class of speaker-segmented S&P 500 transcript data underpins ECB Working Paper No. 3093 on central-bank and corporate communication. Subtext focuses on the investor's question: is this management team getting more cautious, and where?