Talking the Skeptic Down: 66% of Calls Leave the Doubter Reassured
Artul.ai's model scored 165,182 earnings calls from 1990 to 2026 on a battery of items, including whether the analyst-skeptic voice ended up reassured. In 109,699 calls (66.41%, 95% CI 66.18%–66.64%) the model answered YES to 'Skeptic Reassured.' These calls show higher candor (7.03 vs 6.86), specificity (7.81 vs 7.56), and confidence (7.55 vs 7.21), and lower evasion (2.44 vs 2.70) and stress (1.93 vs 2.43). Guidance was raised in 28.36% of cases vs 21.05% in the base, and lowered in 6.28% vs 11.56%. In the 18,030-call returns subsample, the median subsequent return was -5.71% vs -7.16% base, with 40.92% beating vs 39.47%.
- Skeptic Reassured calls make up 66.41% of the corpus (95% CI 66.18%-66.64%) out of 165,182 scored calls.
- Reassured calls show lower stress (1.93 vs 2.43) and evasion (2.44 vs 2.70) but higher confidence (7.55 vs 7.21).
- Guidance was lowered in only 6.28% of reassured calls versus 11.56% of the base, and raised in 28.36% versus 21.05%.
- The share of reassured calls peaked in 2021 at 72.96% before falling to 64.41% in 2022 and 64.32% in partial 2025.
1Introduction
Anyone who listens to earnings calls for a living knows the moment: an analyst pushes hard, management answers, and either the room settles or the tension lingers. Whether the skeptic on the other end of the line walks away convinced is one of the most human signals in the whole ritual, and it has never been systematically measured at scale. Artul.ai scored 165,182 earnings calls spanning 1990 through 2026 on a battery of tone and behavior items, including whether the model judged the skeptic reassured. This study describes the 109,699 calls where that answer was YES: how they differ in tone, guidance behavior, language patterns, and subsequent returns.
2Data & methodology
The corpus comprises 165,182 earnings-call transcripts published between 1990 and 2026, each scored independently by a large language model on an identical 37-field battery: seven categorical business verdicts, eight 0–9 behavioral meters, and twenty yes/no judgments. The study group is defined as calls where the model answered YES to the battery item "Skeptic Reassured" (n = 109,699; 66.4% of the reference set, 95% Wilson interval 66.2%–66.6%). Baseline figures use all scored calls. Market outcomes join a fixed sample of 22,449 calls with twelve-month total returns in excess of SPY, measured from the first close after each call; this sample skews toward liquid U.S. names and is reported as descriptive history only.
3Results
Reassured calls read differently. Candor runs 7.03 versus 6.86 in the base, specificity 7.81 versus 7.56, and confidence 7.55 versus 7.21, while evasion (2.44 vs 2.70) and stress (1.93 vs 2.43) sit lower. Guidance skews positive: 28.36% raised versus 21.05%, and 6.28% lowered versus 11.56%. Underrepresented language patterns are telling — 'Results Worse Than Direction' appears at 0.69x expected, 'The Question Left Hanging' at 0.54x, and 'Underused Fixed Costs' at 0.74x. The trend shows a 2020-2021 peak (71.76% and 72.96%) followed by a drop to 64.41% in 2022. In the returns subsample (n=18,030), the median subsequent return was -5.71% versus -7.16% base, and 40.92% beat versus 39.47%.
| Meter | Study group | Baseline | Δ |
|---|---|---|---|
| Candor | 7.03 | 6.86 | +0.16 |
| Evasion | 2.44 | 2.70 | -0.26 |
| Specificity | 7.81 | 7.56 | +0.25 |
| Stress | 1.93 | 2.43 | -0.49 |
| Promotion | 4.98 | 5.05 | -0.07 |
| Confidence | 7.55 | 7.21 | +0.34 |
| Action | Study group | Baseline |
|---|---|---|
| Raised | 28.4% | 21.1% |
| Maintained | 53.9% | 48.8% |
| Lowered | 6.3% | 11.6% |
| Withdrawn | 2.1% | 2.7% |
| Signal | Lift | In group | Baseline |
|---|---|---|---|
| Scale-Dependent Advantage Claims | 0.09× | 1.0% | 11.1% |
| The Question Left Hanging | 0.54× | 25.9% | 48.0% |
| Results Worse Than Direction | 0.69× | 35.4% | 51.1% |
| Underused Fixed Costs | 0.74× | 31.0% | 41.6% |
| Statistic | Study group | Returns sample |
|---|---|---|
| Median excess return | -5.7% | -7.2% |
| Interquartile range | -23.3% to +12.3% | — |
| Share beating SPY | 40.9% (95% CI 40%–42%) | 39.5% |
| Observations | 18,030 | 22,449 |
| Ticker | Quarter | Call date | Call grade |
|---|---|---|---|
| SBFG | Q2 2025 | 2025-07-25 | A |
| USCB | Q2 2025 | 2025-07-25 | B+ |
| HCA | Q2 2025 | 2025-07-25 | C |
| AON | Q2 2025 | 2025-07-25 | C |
| NWG | Q2 2025 | 2025-07-25 | B+ |
| BFH | Q2 2025 | 2025-07-25 | B |
| FFIC | Q2 2025 | 2025-07-25 | B+ |
| OMF | Q2 2025 | 2025-07-25 | A |
4Discussion
A careful reader should treat this as a description of association, not a recipe. Calls where the skeptic ends up reassured also feature more candor, more specificity, more raised guidance, and slightly better median subsequent returns — but none of that establishes that reassurance causes anything, or that it predicts returns. The beat-rate gap (40.92% vs 39.47%) is small and the medians overlap considerably. The pattern is best read as a coherent tone signature: conversations that resolve tend to be the more candid, more specific, less stressful ones. What the study cannot say is which direction the arrow runs.
5Limitations
The labels come from AI-scored fields and inherit their noise and possible bias; a YES on 'Skeptic Reassured' is a model judgment, not ground truth. The returns analysis covers 22,449 calls skewed toward liquid names, and our own forward tests falsified directional prediction from these features. LLMs partially remember famous stocks' histories, contaminating any backtest of this kind. The corpus spans 1990-2026 with uneven coverage, and 2025 is partial. Treat every number here as descriptive of this dataset and this model, not as a general law about markets. See the full methodology, including the C1 pattern’s forward-test failure and the LLM-memorization finding.