NonAnswer score
Table view
| Reference group | Texts | Median | Your text is higher than |
|---|
Fewer than 100 words: the published measure is only computed for texts of at least 100 words, so treat this score as indicative.
Companion to Management Science 69(10), 2023
Score any text — an earnings-call answer, a hearing, an interview — with the non-answer measure of Barth, Mansouri & Wöbbeking (2023), alongside tone, uncertainty, jargon and forward-looking language, and see where it stands among hundreds of thousands of earnings calls.
| Reference group | Texts | Median | Your text is higher than |
|---|
Fewer than 100 words: the published measure is only computed for texts of at least 100 words, so treat this score as indicative.
Share of words and percentile among the selected earnings calls; the curves show how the calls are distributed (central 99%), the dashed line is your text. Definitions and sources in the Glossary.
Text after the measure’s cleaning (lowercase, no digits or punctuation, contractions expanded), with non-answer glossary phrases highlighted. Hover a phrase for its weight.
The measures on this site come from these papers. Please cite them when you use the scores or the data.
Management Science 2023 · 69(10), 6333–6348
Introduces the NonAnswer measure computed on this site.
Journal of Financial and Quantitative Analysis 2026 · 61(5), 2353–2391
Evasive answers in earnings calls are followed by less media coverage.
SSRN in revision
Context specific language or jargon helps, by definition, to efficiently and precisely transfer information. However, due to its complex nature, jargon might also be a tool to obfuscate information. This paper studies whether jargon is used in verbal firm disclosures to obfuscate or to efficiently transfer information. We observe that, within the Q&A of earnings conference calls, managers use less jargon in responses to tougher questions, and in calls after a quarter of bad success. Moreover, markets interpret the lack of precise information as a bad signal: we find lower cumulative abnormal returns and a higher implied volatility following earnings calls where managers use less jargon. These results support the argument that an excessive use of non-factual language is perceived as blathering that retards the reduction of information asymmetries.
SSRN working paper
This paper investigates the impact of precise language on market efficiency by examining the interplay between regulatory language simplification, as imposed by the Plain Writing Act of 2010, and information asymmetries in financial markets. We posit that financial jargon, while often criticized as complex, plays a pivotal role in facilitating efficient information dissemination within the market. Analyzing earnings conference call text data, we employ an instrumental variable event study framework to evaluate the consequences of the Plain Writing Act. Our findings suggest that the Act's promotion of plain language usage results in a reduction of financial jargon. Paradoxically, this reduction in jargon is associated with a decrease in market efficiency, implying that the Act may inadvertently hinder information flow and impair the market's price setting mechanism.
Call-level measures of management answers for Capital IQ and StreetEvents earnings calls since 2001,
one row per call, identified by date and Compustat gvkey. Choose the sample, years, measures and format.
Every file includes date, gvkey and the US flag is_us; CSV and Stata downloads come zipped with a
README describing the selection and the measures.
Transcript source. The data do not come from a single vendor: they combine S&P Capital IQ and StreetEvents transcripts, one transcript per call. The Management Science (2023) and JFQA (2026) papers used StreetEvents transcripts only, so for the same call the measures here can differ from the papers' published data. To replicate a paper, use its own dataset.
Definitions and sources of every measure are in the Glossary. If you use the data, please cite Barth, Mansouri & Wöbbeking (2023) and Boustanifar & Mansouri (2026).
How each measure is computed and where it comes from, and the non-answer phrases with their weights.
The non-answer measure of Barth, Mansouri & Wöbbeking (2023) counts glossary phrases in the cleaned text and scales by length:
NonAnswer = Σt countt / (words + 1) NonAnswerφ = Σt wt · countt / (words + 1)
Cleaning follows the paper exactly: lowercase; “don't”, “'ve”, “'d”, “can't” expanded; digits and punctuation removed; whitespace collapsed. Phrases are counted as non-overlapping substrings, as in the paper's code; this site reproduces the paper’s published scores to machine precision. Texts need at least 100 words.
Whole words of each word list, divided by words + 1; multi-word phrases are matched longest first, without overlap. The same conventions are used in the call-level data.
| Measure | Definition | Source |
|---|
The 1,364 three-word phrases (after cleaning) and their weights, learned from S&P 500 earnings calls with a multinomial inverse regression in the paper.
| Phrase | Weight |
|---|
Score many texts at once in the browser, or call the same measures from your own code.
Upload a CSV with a text column (and optionally an id column); you get back a CSV with all
measures for every row. Up to 10,000 texts or 10 MB per file. Download a template.
No key needed. Base URL:
/api/score
JSON {"text": "…"} (up to 300,000 characters) → all measures, percentiles in each reference group,
the closest presidential interview and the matched glossary phrases.
/api/batch
Multipart field file: a CSV with text and optional id columns → a CSV of
measures, one row per text.
/api/glossary
The non-answer phrases and their weights.
/api/meta
Definitions and sources of the measures.
/api/benchmarks
Percentiles and histograms of the reference groups (earnings calls, presidential interviews, sports press conferences).
/api/data/download
Call-level data. Parameters: sample (all, us), start, end,
measures (comma-separated; default all), format (csv, dta, parquet). Row counts:
/api/data/count; options: /api/data/meta.
Shell
curl -s -X POST /api/score \
-H 'Content-Type: application/json' \
-d '{"text": "We do not break that out, and I would rather not speculate."}'
curl -s -X POST /api/batch -F file=@answers.csv -o scores.csv
Python
import requests
API = ""
r = requests.post(f"{API}/api/score", json={"text": "We do not break that out."})
print(r.json()["NonAnswer"])
# US calls 2010-2020, all measures, Stata
d = requests.get(f"{API}/api/data/download",
params={"sample": "us", "start": 2010, "end": 2020, "format": "dta"})
open("calls.zip", "wb").write(d.content)