The science behind the signal

Built on open research.

The analytical approach is grounded in two decades of peer-reviewed work on how corporate language reveals information. The phenomena Spoken Alpha detects — hedging, conditionality, ownership detachment, vocal affect — are public-domain academic constructs. Our contribution is the engineering: applying modern language models to score them, at scale, against every speaker's own historical baseline.

Foundational work

The four papers the field is built on.

If you only read four references in this space, read these. Each is peer-reviewed, widely cited, and publicly accessible.

Larcker, D. F. & Zakolyukina, A. A.(2012)
Detecting Deceptive Discussions in Conference Calls. Journal of Accounting Research, 50(2), 495–540.
The seminal empirical paper on linguistic deception markers in earnings calls. Establishes that word-category usage by CEOs and CFOs predicts subsequent restatements — the methodological template the field has built on for a decade.
Loughran, T. & McDonald, B.(2011)
When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks. Journal of Finance, 66(1), 35–65.
The financial-domain word lists (negative, uncertain, litigious, weak/strong modal, constraining) that became the standard lexicon for finance-text sentiment scoring. General-purpose dictionaries mis-score finance text; this is the corrective.
Pennebaker, J. W., Boyd, R. L., Jordan, K. & Blackburn, K.(2015)
The Development and Psychometric Properties of LIWC2015. Linguistic Inquiry and Word Count.
The foundational psycholinguistic text-analysis framework — tentativeness, certainty, cognitive-process, and affect dimensions used across the deception and disclosure literature.
Mayew, W. J. & Venkatachalam, M.(2012)
The Power of Voice: Managerial Affective States and Future Firm Performance. Journal of Finance, 67(1), 1–43.
Vocal affect on earnings calls carries information about firm fundamentals incremental to both reported numbers and the linguistic content of the call. Direct evidence that the speaker, not just the script, matters.

Supporting literature

Forensic linguistics and disclosure complexity.

How each individual linguistic phenomenon (temporal hedging, conditionality, detachment, forward-looking commitment) was empirically established before any AI tooling existed.

Li, F.(2010)
The Information Content of Forward-Looking Statements in Corporate Filings — A Naïve Bayesian Machine Learning Approach. Journal of Accounting Research, 48(5), 1049–1102.
Forward-looking statements carry information content distinct from realized results. Methodological anchor for treating speaker forward language as a signal in its own right.
Bushee, B. J., Gow, I. D. & Taylor, D. J.(2018)
Linguistic Complexity in Firm Disclosures: Obfuscation or Information?. Journal of Accounting Research, 56(1), 85–121.
How to score management acknowledgement of monitored conditions in disclosure text. Empirically distinguishes informative complexity from obfuscation — load-bearing distinction for any system reading executive language.
Hyland, K.(2005)
Metadiscourse: Exploring Interaction in Writing. Continuum.
Hedging and conditional-construction markers as commitment-modulation devices in expert discourse. Source for the linguistic theory behind why "if X then Y" and "provided we can" patterns are not interchangeable filler.
Vrij, A.(2008)
Detecting Lies and Deceit: Pitfalls and Opportunities (2nd ed.). Wiley.
The standard forensic-linguistics reference on temporal qualifiers, ownership detachment, and other commitment-reduction devices in spoken communication.
Newman, M. L., Pennebaker, J. W., Berry, D. S. & Richards, J. M.(2003)
Lying Words: Predicting Deception from Linguistic Styles. Personality and Social Psychology Bulletin, 29(5), 665–675.
Self-reference avoidance and third-person framing as deception indicators. Why pronoun usage on an earnings call is not stylistic noise.
Hancock, J. T., Curry, L. E., Goorha, S. & Woodworth, M.(2008)
On Lying and Being Lied To: A Linguistic Analysis of Deception in Computer-Mediated Communication. Discourse Processes, 45(1), 1–23.
Documents temporal-hedge usage patterns in deceptive vs. truthful exchanges. Complements Vrij for short-form spoken-style content.

Per-speaker baselines

The stylometry behind “measure each speaker against themselves.”

Comparing a given speaker's current language to their own historical corpus is a standard pattern in authorship attribution and longitudinal linguistic analysis. These are the references behind that comparison.

Stamatatos, E.(2009)
A Survey of Modern Authorship Attribution Methods. Journal of the American Society for Information Science and Technology, 60(3), 538–556.
The methodological reference for comparing a given author's current text against their own prior corpus to detect deviation. The stylometric foundation behind per-speaker baselines.
Pennebaker, J. W. & King, L. A.(1999)
Linguistic styles: Language use as an individual difference. Journal of Personality and Social Psychology, 77(6), 1296–1312.
Empirical evidence that individuals have stable, measurable linguistic signatures over time — the precondition for treating a deviation from a speaker's own baseline as meaningful information.
Loughran, T. & McDonald, B.(2016)
Textual Analysis in Accounting and Finance: A Survey. Journal of Accounting Research, 54(4), 1187–1230.
Surveys per-firm and per-speaker textual-deviation methods across the financial-disclosure literature. The state-of-the-field paper if you only read one survey.

Active research

The field is openly studied, not a closed vendor space.

Recent open datasets and benchmarks specifically on earnings-call language. Cited as evidence that the underlying science continues to develop in public.

Ma, Y. et al.(2026)
EvasionBench: A Large-Scale Benchmark for Detecting Managerial Evasion in Earnings Call Q&A. arXiv:2601.09142.
A public, 84K-pair earnings-call Q&A dataset with a three-level evasion taxonomy and an open-weights classifier. Evidence the field is actively, openly researched — not a closed-vendor space.

What we add

The literature is the foundation. The engineering is the product.

Per-speaker longitudinal baselines

Earlier work scored language cross-sectionally — against industry averages or pooled corpora. We score each executive against their own prior calls, so a hedge-heavy speaker isn't flagged for being themselves and a normally crisp speaker is flagged the moment they aren't.

LLM scoring across the full universe

The 2011–2018 papers worked with hundreds or low thousands of calls and hand-curated dictionaries. We apply modern language models across every public earnings call, every quarter, with structured per-exchange scoring. The substrate changed; the questions the literature asks are the same.

Strict-prior dating

For a target call on date T, baselines are computed only from calls strictly prior to T. Anything closer leaks future information into the score. This is a data-engineering discipline, not a research claim — but it's the reason the signal is honestly investable rather than a backtest artifact.

Reading more

Want to see how the methodology shows up in product?

The demo walks through a worked example end-to-end — what we flag, why, and how the trade is structured.

See the demo Read the methodology notes

Loading…

Built on open research.

The four papers the field is built on.

If you only read four references in this space, read these. Each is peer-reviewed, widely cited, and publicly accessible.

Larcker, D. F. & Zakolyukina, A. A.(2012)

Detecting Deceptive Discussions in Conference Calls. Journal of Accounting Research, 50(2), 495–540.

The seminal empirical paper on linguistic deception markers in earnings calls. Establishes that word-category usage by CEOs and CFOs predicts subsequent restatements — the methodological template the field has built on for a decade.

Loughran, T. & McDonald, B.(2011)

When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks. Journal of Finance, 66(1), 35–65.

The financial-domain word lists (negative, uncertain, litigious, weak/strong modal, constraining) that became the standard lexicon for finance-text sentiment scoring. General-purpose dictionaries mis-score finance text; this is the corrective.

Pennebaker, J. W., Boyd, R. L., Jordan, K. & Blackburn, K.(2015)

The Development and Psychometric Properties of LIWC2015. Linguistic Inquiry and Word Count.

The foundational psycholinguistic text-analysis framework — tentativeness, certainty, cognitive-process, and affect dimensions used across the deception and disclosure literature.

Mayew, W. J. & Venkatachalam, M.(2012)

The Power of Voice: Managerial Affective States and Future Firm Performance. Journal of Finance, 67(1), 1–43.

Vocal affect on earnings calls carries information about firm fundamentals incremental to both reported numbers and the linguistic content of the call. Direct evidence that the speaker, not just the script, matters.

Forensic linguistics and disclosure complexity.

How each individual linguistic phenomenon (temporal hedging, conditionality, detachment, forward-looking commitment) was empirically established before any AI tooling existed.

Li, F.(2010)

The Information Content of Forward-Looking Statements in Corporate Filings — A Naïve Bayesian Machine Learning Approach. Journal of Accounting Research, 48(5), 1049–1102.

Forward-looking statements carry information content distinct from realized results. Methodological anchor for treating speaker forward language as a signal in its own right.

Bushee, B. J., Gow, I. D. & Taylor, D. J.(2018)

Linguistic Complexity in Firm Disclosures: Obfuscation or Information?. Journal of Accounting Research, 56(1), 85–121.

How to score management acknowledgement of monitored conditions in disclosure text. Empirically distinguishes informative complexity from obfuscation — load-bearing distinction for any system reading executive language.

Hyland, K.(2005)

Metadiscourse: Exploring Interaction in Writing. Continuum.

Hedging and conditional-construction markers as commitment-modulation devices in expert discourse. Source for the linguistic theory behind why "if X then Y" and "provided we can" patterns are not interchangeable filler.

Vrij, A.(2008)

Detecting Lies and Deceit: Pitfalls and Opportunities (2nd ed.). Wiley.

The standard forensic-linguistics reference on temporal qualifiers, ownership detachment, and other commitment-reduction devices in spoken communication.

Newman, M. L., Pennebaker, J. W., Berry, D. S. & Richards, J. M.(2003)

Lying Words: Predicting Deception from Linguistic Styles. Personality and Social Psychology Bulletin, 29(5), 665–675.

Self-reference avoidance and third-person framing as deception indicators. Why pronoun usage on an earnings call is not stylistic noise.

Hancock, J. T., Curry, L. E., Goorha, S. & Woodworth, M.(2008)

On Lying and Being Lied To: A Linguistic Analysis of Deception in Computer-Mediated Communication. Discourse Processes, 45(1), 1–23.

Documents temporal-hedge usage patterns in deceptive vs. truthful exchanges. Complements Vrij for short-form spoken-style content.

The stylometry behind “measure each speaker against themselves.”

Stamatatos, E.(2009)

A Survey of Modern Authorship Attribution Methods. Journal of the American Society for Information Science and Technology, 60(3), 538–556.

The methodological reference for comparing a given author's current text against their own prior corpus to detect deviation. The stylometric foundation behind per-speaker baselines.

Pennebaker, J. W. & King, L. A.(1999)

Linguistic styles: Language use as an individual difference. Journal of Personality and Social Psychology, 77(6), 1296–1312.

Empirical evidence that individuals have stable, measurable linguistic signatures over time — the precondition for treating a deviation from a speaker's own baseline as meaningful information.

Loughran, T. & McDonald, B.(2016)

Textual Analysis in Accounting and Finance: A Survey. Journal of Accounting Research, 54(4), 1187–1230.

Surveys per-firm and per-speaker textual-deviation methods across the financial-disclosure literature. The state-of-the-field paper if you only read one survey.

The field is openly studied, not a closed vendor space.

Recent open datasets and benchmarks specifically on earnings-call language. Cited as evidence that the underlying science continues to develop in public.

Ma, Y. et al.(2026)

EvasionBench: A Large-Scale Benchmark for Detecting Managerial Evasion in Earnings Call Q&A. arXiv:2601.09142.

A public, 84K-pair earnings-call Q&A dataset with a three-level evasion taxonomy and an open-weights classifier. Evidence the field is actively, openly researched — not a closed-vendor space.

The literature is the foundation. The engineering is the product.

Per-speaker longitudinal baselines

LLM scoring across the full universe

Strict-prior dating