Research › 20-Year Trends
Artul.ai Research LibraryStudy No. 820-Year TrendsUpdated 2026-08-28

38.6% of 165,182 earnings calls read rehearsed to our model, and the share keeps climbing

By Artul.ai Research Group · n = 63,830 earnings calls · First published 2026-08-28
Abstract

We asked a simple question across 165,182 earnings calls from 1990 to 2026: in how many calls did a language model answer YES to the prompt "Calls That Read Rehearsed"? The answer: 63,830 calls, or 38.6% (95% CI 38.4%–38.9%). The share has risen steadily, from 38.04% of calls in 2015 to 45.01% in 2025. Calls flagged as rehearsed show a distinct behavioral profile: promotion language runs 0.47 points higher and candor 0.34 points lower than the corpus baseline. They are also 1.7x more likely to feature scale-dependent advantage claims, and 0.68x as likely to discuss pricing recovering. Among 7,895 such calls with return data, the median next-day return was -0.09% versus -0.07% for the base sample.

Key findings
  • The model flagged 63,830 of 165,182 calls (38.6%) as reading rehearsed, with a 95% confidence interval of 38.4% to 38.9%.
  • The rehearsed share rose from 38.04% in 2015 to 45.01% in 2025, climbing in nearly every year of the trend window.
  • Rehearsed calls score 0.47 points higher on promotion and 0.34 points lower on candor than the overall corpus.
  • Among 7,895 rehearsed calls with next-day returns, the median was -0.09% versus -0.07% for the 22,449-call base sample, and 37.2% beat versus 39.5% for the base.

1Introduction

Earnings calls are the most scripted communications in public markets, yet investors routinely treat them as windows into management thinking. If a model can identify when a call reads rehearsed, the flag could describe a measurable style of communication rather than mere noise. Understanding how common that style is, how it has changed over a decade, and what it co-occurs with matters to anyone who parses management language for signal. This study examines 165,182 earnings calls from 1990 to 2026, counting the 63,830 calls where the model answered YES to "Calls That Read Rehearsed," and profiles those calls across behavioral dimensions, guidance actions, and next-day 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 "Calls That Read Rehearsed", tracked by year (n = 63,830; 38.6% of the reference set, 95% Wilson interval 38.4%–38.9%). 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

The headline pattern is the trend: the rehearsed share climbed from 38.04% in 2015 to 45.01% in 2025, with only 2020 (32.37%) dipping sharply below the surrounding years. Behaviorally, rehearsed calls tilt promotional: promotion runs 0.47 points above baseline while candor sits 0.34 points below and specificity 0.18 below. On individual markers, "Scale-Dependent Advantage Claims" appears 1.7x more often than in the base corpus, "The Finished-Story Tell" 1.37x, and "When the CFO Dominates" 1.34x; "Pricing Recovering" appears only 0.68x as often. Guidance-wise, 18.5% of rehearsed calls raised guidance versus 21.1% of the base, and 9.5% lowered versus 11.6%. The returns sample (n=7,895) shows a median next-day return of -0.09% versus -0.07% for the base, with 37.2% beating versus 39.5%.

Table 1. Mean behavioral scores (0–9 scale), study group versus baseline
MeterStudy groupBaselineΔ
Candor6.526.86-0.34
Evasion2.812.70+0.11
Specificity7.387.56-0.18
Stress2.492.43+0.06
Promotion5.525.05+0.47
Confidence7.267.21+0.04
Table 2. Guidance actions, study group versus baseline
ActionStudy groupBaseline
Raised18.5%21.1%
Maintained48.6%48.8%
Lowered9.5%11.6%
Withdrawn1.8%2.7%
Table 3. Co-occurring battery signals ranked by lift (group prevalence ÷ baseline prevalence)
SignalLiftIn groupBaseline
Scale-Dependent Advantage Claims1.70×18.8%11.1%
The Finished-Story Tell1.37×6.0%4.4%
When the CFO Dominates1.34×18.9%14.1%
A Tiny Fraction of the Market1.32×39.5%30.0%
Pricing Recovering0.68×14.7%21.5%
201538.04%
201635.06%
201734.85%
201837.15%
201938.69%
202032.37%
202139.37%
202241.11%
202342.18%
202443.53%
202545.01%
Figure 1. Share of all analyzed calls matching the study definition, by year.
Table 4. Twelve-month excess total returns versus SPY (descriptive history, not a signal)
StatisticStudy groupReturns sample
Median excess return-8.9%-7.2%
Interquartile range-28.3% to +10.7%
Share beating SPY37.2% (95% CI 36%–38%)39.5%
Observations7,89522,449
Table 5. Most recent calls matching the study definition
TickerQuarterCall dateCall grade
USCBQ2 20252025-07-25B+
HCAQ2 20252025-07-25C
AONQ2 20252025-07-25C
BFHQ2 20252025-07-25B
FFICQ2 20252025-07-25B+
OMFQ2 20252025-07-25A
GBCIQ2 20252025-07-25A
LARKQ2 20252025-07-25B

4Discussion

A careful reader should conclude that calls flagged as rehearsed are common, increasingly so, and carry a recognizable stylistic signature: more promotion, less candor, more grand claims about scale advantages. None of this establishes that rehearsal causes outcomes. The returns and beat-rate differences are small in magnitude and come from a subsample skewed toward liquid names, so they should be read as descriptive co-occurrence, not as a trading signal. The 2020 dip is an observation about a single unusual year, not evidence of what drives the long-run rise.

5Limitations

The YES/NO rehearsed flag is an AI-read judgment on noisy transcripts and may misclassify tone, especially for older calls with poorer audio and text quality. The returns analysis covers 7,895 flagged calls against a base of 22,449, skewed toward liquid large-cap names, so it is not representative of the full corpus. Our own forward tests falsified directional prediction from these flags, so no edge should be inferred. Finally, language models partially remember famous stocks' histories, which can contaminate any backtest of model-labeled text. See the full methodology, including the C1 pattern’s forward-test failure and the LLM-memorization finding.

Cite this study Artul.ai Research Group (2026). “38.6% of 165,182 earnings calls read rehearsed to our model, and the share keeps climbing.” Artul.ai Earnings-Call Research Library, Study No. 8. https://artul.ai/research/are-earnings-calls-getting-more-rehearsed

Related studies

87.8% of earnings calls contain critic ammunition - 66.4% of earnings calls left the model's skeptic reaWhen a Call Ends With a Question Left Hanging: 47.95When the model flags 'Volume About to Step Up' on 28Founder-led calls: 20.0% of the corpus, but only 41.When the CFO Dominates: 14.1% of 165,182 Calls, With
Not investment advice. Artul.ai publishes AI-generated earnings-call quality grades and expected-volatility estimates — never buy or sell recommendations. We tested over 1,600 predictive hypotheses against 165,000 transcripts; the honest result, including what failed, is documented in our methodology.