Hypothesis Tested: Answers go deeper than the script

When an executive stops reading the script and actually answers the question, does the call get better? To find out, we had an AI read 144,500 earnings-call transcripts from the 165,000-call corpus and score each one on candor, specificity, and related measures. The 200 calls where answers went noticeably deeper than the prepared remarks form the group studied here.

These calls scored higher on specificity (7.76 vs 7.65) and confidence (7.57 vs 7.35), with slightly lower evasion (2.42 vs 2.66). They co-occur most with "Volume About to Step Up" at 1.55x the base rate (39.5% vs 25.5%), and less with rehearsed-sounding calls (0.72x). Guidance was raised in 33.5% of cases versus 24.4% in the base. Post-call returns were less negative: median -5.0% versus -6.9% overall.

These AI-read scores are noisy and the pattern is rare, so treat the profile differences cautiously. Return figures are descriptive history from a 22,449-call sample, not a trading signal; our own forward tests falsified directional prediction. Language models also partially remember famous stocks' histories, which contaminates any backtest of this kind.

Behavior profile

Guidance actions in this group

What travels with it

Share of calls by year

Recent examples

This hypothesis has its own page with every matching company: see the full question.