Measuring overconfidence from what executives say
CEO and managerial overconfidence appear on 98 recent papers, mostly measured from option-exercise behaviour or press coverage, as in the 2000s. Only 6 of them carry a text, language or speech keyword, and 4 of those are in AI journals. Can a language model score overconfidence from earnings calls, and does it predict the same decisions?
What prompted this
Since October 2025, 98 papers carry an overconfidence keyword (44 in finance journals, 24 in AI, 18 in economics): "overconfidence" on 51, "CEO overconfidence" on 22, "managerial overconfidence" on 15. Only 6 of the 98 also carry a text, language, machine-learning, speech, tone or linguistic keyword, and 4 of those 6 are in AI journals.
Questions
- Does a language model's reading of CEOs' answers on earnings calls (certainty, dismissal of risks, forecast ranges) agree with the option-based measure, and where do they disagree?
- Which measure better predicts overinvestment, overpriced acquisitions and later restatements?
- Can the text measure be computed for private firms and non-US firms, where option data do not exist?
Data that could help
Earnings call transcripts, option exercise data (ExecuComp), deal and investment data (Compustat, SDC).
Next step
Score 1,000 earnings calls with a language model for overconfidence and compare the scores with the option-based measure for the same CEOs.