Margins Are Expanding, and So Is the Confidence: Reading 44% of Calls
We examine 73,217 earnings calls from a 165,182-call corpus spanning 1990-2026 where the margins signal was read as expanding, a 44.3% share (95% CI 44.1%-44.6%). Speakers on these calls score higher on confidence (7.57 vs 7.21), specificity (7.72 vs 7.56), and promotion (5.21 vs 5.05), and lower on evasion (2.55 vs 2.70) and stress (2.06 vs 2.43). Guidance behavior differs sharply: 32.1% raised guidance versus 21.1% in the base set, while 6.3% lowered it versus 11.6%. Overrepresented phrases include Pricing Recovering and Deferred Revenue Growing; underrepresented include Results Worse Than Direction. Median forward returns were -6.2% versus -7.2% in the base.
- Margin-expanding calls account for 44.3% of the 165,182-call corpus, with a 95% confidence interval of 44.1% to 44.6%.
- Confidence on these calls averages 7.57 versus 7.21 in the base set, while stress is lower at 2.06 versus 2.43.
- Guidance was raised on 32.1% of margin-expanding calls versus 21.1% of base calls, and lowered on 6.3% versus 11.6%.
- The phrase Pricing Recovering appears at a 1.26x lift on these calls, while Results Worse Than Direction appears at 0.66x.
- Median forward returns for the 11,186-call returns sample were -0.0618 versus -0.0716 in the base, with 40.7% beating versus 39.5%.
1Introduction
Margins are the line analysts quote first and executives defend hardest, so how management talks about them is a window into tone on the calls that matter. If nearly half of all earnings calls feature margins described as expanding, that framing is not a rarity but a default posture, and its linguistic fingerprints deserve scrutiny. For anyone who follows calls closely, the question is whether the confident, specific, promotion-heavy delivery that accompanies margin talk also travels with guidance behavior and market outcomes. This study profiles 73,217 calls where margins was read as expanding against the full 165,182-call corpus, comparing language, guidance actions, recurring phrases, and forward 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 margins was read as expanding (n = 73,217; 44.3% of the reference set, 95% Wilson interval 44.1%–44.6%). 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
Margin-expanding calls sound like it: confidence runs 7.57 versus 7.21, specificity 7.72 versus 7.56, and stress sits at 2.06 versus 2.43, with evasion lower at 2.55 versus 2.70. Guidance actions skew positive, with 32.1% raising guidance versus 21.1% in the base and only 6.3% lowering it versus 11.6%. The phrase table matches the mood: Pricing Recovering and Deferred Revenue Growing both show a 1.26x lift, while Results Worse Than Direction (0.66x) and Scale-Dependent Advantage Claims (0.67x) are underrepresented. The annual share is cyclical, dipping to 38.7% in 2020 and peaking at 50.2% in 2024. Median forward returns of -0.0618 versus -0.0716 in the base show only a modest gap, and the beat rate is 40.7% versus 39.5%.
| Meter | Study group | Baseline | Δ |
|---|---|---|---|
| Candor | 6.87 | 6.86 | +0.01 |
| Evasion | 2.55 | 2.70 | -0.15 |
| Specificity | 7.72 | 7.56 | +0.16 |
| Stress | 2.06 | 2.43 | -0.37 |
| Promotion | 5.21 | 5.05 | +0.16 |
| Confidence | 7.57 | 7.21 | +0.36 |
| Action | Study group | Baseline |
|---|---|---|
| Raised | 32.1% | 21.1% |
| Maintained | 48.7% | 48.8% |
| Lowered | 6.3% | 11.6% |
| Withdrawn | 1.5% | 2.7% |
| Signal | Lift | In group | Baseline |
|---|---|---|---|
| Pricing Recovering | 1.26× | 27.1% | 21.5% |
| Deferred Revenue Growing | 1.26× | 11.1% | 8.9% |
| Results Worse Than Direction | 0.66× | 33.7% | 51.1% |
| Scale-Dependent Advantage Claims | 0.67× | 7.5% | 11.1% |
| Statistic | Study group | Returns sample |
|---|---|---|
| Median excess return | -6.2% | -7.2% |
| Interquartile range | -24.9% to +12.8% | — |
| Share beating SPY | 40.7% (95% CI 40%–42%) | 39.5% |
| Observations | 11,186 | 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+ |
| FFIC | Q2 2025 | 2025-07-25 | B+ |
| OMF | Q2 2025 | 2025-07-25 | A |
| GBCI | Q2 2025 | 2025-07-25 | A |
4Discussion
A careful reader should conclude that calls with expanding margin language carry a distinctive, more confident tone and a guidance profile tilted toward raises. That is a description of co-occurrence, not a mechanism: nothing here shows that the language causes outcomes or that reading margins as expanding predicts returns. The returns gap is small and the beat-rate gap of 40.7% versus 39.5% is close to the base. The annual trend moves with market conditions rather than pointing anywhere forward. Treat the phrase lifts and language deltas as descriptive texture on how management frames good margin news, not as signals to trade on.
5Limitations
The margins reading and all language scores are AI-generated fields and are noisy, so measurement error is baked into every delta. The returns sample covers 11,186 calls from a base of 22,449 and is skewed toward liquid names, limiting generalizability. Our own forward tests falsified directional prediction, so no claimed edge exists here. Additionally, LLMs partially remember famous stocks' histories, which can contaminate any backtest built on these annotations. The confidence intervals are wide enough at the phrase level that small lifts should be read cautiously. See the full methodology, including the C1 pattern’s forward-test failure and the LLM-memorization finding.