Which Voices Move Markets? Speaker Identity and the Cross-Section of Post-Earnings Returns

Imagine you are sitting in a crowded room where a company is holding a press conference. There are three main groups of people talking: the Company Bosses (CEOs and CFOs), the Skeptical Reporters (Financial Analysts), and the Moderator (who just asks procedural questions).

For decades, investors and researchers treated this entire conversation as one giant, blurry blob of text. They would take the whole transcript, run it through a basic "good word/bad word" counter, and say, "This company sounds happy, so buy the stock."

This paper argues that this approach is like judging a movie by averaging the reviews of the director, the actors, and the audience together. It misses the most important part: who is speaking matters just as much as what they are saying.

Here is the breakdown of the paper's findings using simple analogies:

1. The "Who" Matters More Than the "What"

The researchers realized that not all voices in the room carry the same weight.

The Bosses (CEOs/CFOs): They are like a salesperson. They are paid to be optimistic. Even if things are going poorly, they might use fancy words to make it sound okay. Their "bullish" statements are often scripted and strategic.
The Reporters (Analysts): They are like investigative journalists. They have no incentive to lie; in fact, their job is to find the cracks in the story. When they ask a tough question or sound skeptical, it's usually because they see something the bosses are trying to hide.
The Finding: The paper found that the Analysts' voices are the most powerful predictor of how the stock will move. Even though bosses speak the most words, the market listens most closely to the reporters.

2. The "Smart Brain" vs. The "Dictionary"

To measure the "mood" of these conversations, the researchers used two different tools:

The Old Tool (Loughran-McDonald Dictionary): Imagine a robot that only knows a list of "good" words (like profit, growth) and "bad" words (like loss, risk). If a sentence says, "We hope to achieve growth," the robot sees "growth" and says "Good!" It doesn't understand that "hope" implies uncertainty. It's context-blind.
The New Tool (FinBERT): This is a super-smart AI trained specifically on financial language. It understands nuance, sarcasm, and context. If a CEO says, "We are hoping to grow," the AI knows that "hoping" is weak and the sentiment is actually shaky. It reads the whole sentence, not just the individual words.

The Result: The Smart AI (FinBERT) completely crushed the Old Dictionary. In a head-to-head race, the AI found the signal, while the dictionary was left looking at zero.

3. The "Weighted Scorecard"

The researchers didn't just use the AI; they built a smart scoring system.
Instead of giving every sentence an equal vote, they assigned "voting power" based on who spoke:

Analysts: 49% of the vote (The most important voice).
CFOs: 30% of the vote.
CEOs/Executives: 16% of the vote.
Others: 5% of the vote.

The Analogy: Imagine a classroom vote. If you let the whole class vote equally, the result is noisy. But if you give the expert teacher 50 votes, the assistant 30 votes, and the students 20 votes, the final decision is much more accurate. This "Section-Weighted" approach was the secret sauce.

4. The "Magic Money" Result

When they tested this new system against the stock market:

The Signal: They found that stocks with "positive analyst sentiment" went up, and those with "negative sentiment" went down.
The Profit: They created a strategy where they bought the "happy" stocks and sold the "sad" ones. This strategy made an extra 2% profit every month (which is huge in finance) that couldn't be explained by normal market risks.
The Surprise: Usually, when you test a strategy on new data (out-of-sample), it gets worse because the market changes. But here, the strategy got better in the new data (2023–2025). This suggests they didn't just get lucky; they found a real, structural truth about how the market works.

5. Why Does This Happen? (The "Slow Cooker" Effect)

The paper also looked at how fast the market reacts.

The Finding: The stock price doesn't jump instantly to the correct value. It "creeps" toward the right price over a few days or weeks.
The Analogy: Imagine dropping a heavy stone into a pond. The splash happens instantly, but the ripples take time to reach the other side. The market hears the "soft information" (the tone of the conversation) but takes a while to fully digest it. This delay creates a window for investors to make money.

Summary: The Big Takeaway

This paper teaches us that not all words are created equal.

Don't just read the transcript; listen to who is speaking.
The Analysts' skepticism is worth more than the Bosses' optimism.
Use smart AI that understands context, not just a simple word list.

By weighting the voices correctly and using advanced AI, investors can spot "soft information" (like confidence or fear) that traditional math misses, leading to better investment decisions.

1. Problem Statement

Existing literature on earnings call sentiment analysis predominantly treats transcripts as monolithic documents, aggregating sentiment scores by averaging all sentences equally. This approach ignores a critical dimension: speaker identity.

The Gap: A bullish statement by a CEO (often scripted and strategically managed) carries different informational weight than a skeptical question from a sell-side analyst (unscripted, real-time, and incentive-aligned with market truth).
The Limitation of Current Models: Traditional dictionary-based methods (e.g., Loughran–McDonald) are "context-blind," failing to capture negation, hedging, or comparative framing. Furthermore, they do not differentiate between the informational roles of different speakers.
Research Question: Does the identity of the speaker matter for predicting post-earnings stock returns, and can a speaker-weighted, deep-learning-based sentiment signal outperform traditional aggregate and dictionary-based approaches?

2. Methodology

A. Data Construction

Dataset: 6.5 million sentences extracted from 16,428 quarterly earnings call transcripts of S&P 500 firms (April 2015 – December 2025).
Speaker Classification: Sentences are categorized into four roles based on titles:
1. Analyst (Sell-side)
2. CFO
3. Executive (CEO, COO, CTO, etc.)
4. Other
Returns: Event-window returns are calculated for 1-day and 5-day horizons, distinguishing between Before Market Open (BMO) and After Market Close (AMC) calls.
Controls: Standardized Unexpected Earnings (SUE) and Fama-French five-factor data are used for regression controls.

B. Sentiment Modeling

Core Model: FinBERT, a transformer-based model pre-trained on financial text, is used to generate sentence-level sentiment scores ( $S_i = P(positive) - P(negative)$ ).
Aggregation Strategies: The authors compare five methods:
1. Simple Mean (M1): Equal weight to all sentences (Baseline).
2. Confidence-Weighted (M2): Weighted by model confidence.
3. Extreme Fractions (M3): Focus on tails of the distribution.
4. Section-Weighted (M4): Key Innovation. Weights are derived empirically from the Information Coefficient (IC) of each speaker category on a training sample.
5. Analyst-Only (M5): Uses only analyst sentences.

C. Weight Derivation (Section-Weighted)

The weights for Method M4 are calculated using in-sample Information Coefficients (IC) (Spearman rank correlation between speaker sentiment and returns) from the training period (2015–2022):

Analyst: 49.0% weight (Highest predictive power).
CFO: 30.0% weight.
Executive: 16.2% weight.
Other: 4.7% weight.
Note: These weights are frozen and applied to the out-of-sample period (2023–2025) to prevent look-ahead bias.

D. Validation Framework

Temporal Split: Training (2015–2022) vs. Testing (2023–2025).
Statistical Tests:
- Information Coefficients (IC).
- Fama-MacBeth cross-sectional regressions (controlling for SUE).
- Fama-French five-factor time-series regressions (Alpha generation).
- Double-sort analysis (Sentiment vs. SUE).
- "Horse Race" regressions comparing FinBERT vs. Loughran-McDonald (LM) dictionary.

3. Key Contributions

Speaker-Level Decomposition: Moves beyond the binary "Presentation vs. Q&A" distinction to analyze individual speaker roles, proving that analyst sentiment is the most predictive voice despite comprising only ~20% of the text volume.
Data-Driven Weighting Scheme: Proposes a principled method to weight sentiment signals based on the empirical predictive power (IC) of specific speaker categories, rather than arbitrary or equal weighting.
Deep Learning Subsumption: Provides rigorous evidence that FinBERT completely subsumes the Loughran-McDonald dictionary. In joint regressions, FinBERT remains highly significant while the dictionary approach becomes statistically indistinguishable from zero.
Out-of-Sample Robustness: Demonstrates an unusual pattern where predictive power increases out-of-sample (IC rises from 0.115 to 0.142), suggesting the signal captures structural market inefficiencies rather than overfitting.

4. Empirical Results

A. Predictive Power (Information Coefficients)

Section-Weighted (M4) achieves a full-sample Spearman IC of 0.119, a 46% improvement over the Simple Mean baseline (0.081).
Analyst-Only (M5) yields the highest IC of 0.141.
Out-of-Sample: The Section-Weighted IC improves from 0.114 (in-sample) to 0.144 (out-of-sample), indicating the signal strengthens over time.

B. Economic Significance (Alpha Generation)

Long-Short Portfolios: A portfolio buying the top sentiment quintile (Q5) and selling the bottom (Q1) generates a monthly Fama-French 5-factor alpha of 2.03% ( $t=6.49$ ).
Annualized Alpha: ~24.3%.
Out-of-Sample Alpha: Increases to 2.77% per month ( $t=8.56$ ) in the 2023+ period.
Monotonicity: Returns are strictly monotonic across sentiment quintiles, with Q5 earning +1.30% and Q1 earning -0.96% on the day after the call.

C. Orthogonality to Earnings Surprises

In Fama-MacBeth regressions, both Section-Weighted Sentiment and SUE remain statistically significant when included together.
Double-Sort Analysis: Sentiment spreads remain significant across all SUE terciles (Low, Mid, High). The spread is largest for firms with negative earnings surprises (Low SUE), suggesting sentiment provides unique "soft information" (e.g., management confidence) that the market underweights during bad quarters.

D. FinBERT vs. Loughran-McDonald (LM)

Horse Race: In a joint regression, FinBERT's coefficient remains highly significant ( $t=5.90$ ), while the LM dictionary coefficient drops to insignificance ( $t=0.86$ ).
Interpretation: FinBERT captures context (negation, hedging, comparison) that the bag-of-words LM approach misses. For example, FinBERT correctly identifies "down 24% from strong performance" as negative, whereas LM may flag "strong" as positive.

E. Signal Decay

The signal has a half-life of 6–7 trading days. The IC decays from 0.119 (1-day) to 0.032 (21-day), consistent with the gradual assimilation of soft information (Post-Earnings Announcement Drift).

5. Significance and Implications

For Practitioners: Identifies a specific, implementable alpha source: tilting sentiment strategies toward analyst commentary rather than management remarks. The strategy is robust to transaction costs (implied by the magnitude of alpha) and works across market regimes (pre/post-COVID).
For Academics: Establishes that speaker identity is a fundamental variable in textual analysis. It challenges the "monolithic transcript" assumption and provides a framework for weighting multi-speaker corporate communications.
For Methodology: Validates context-aware deep learning (Transformers) as superior to dictionary-based methods in finance. The "subsumption" result suggests that future research should abandon simple lexicons in favor of transformer-based models for financial sentiment.
Market Microstructure: The findings support the theory that markets are sluggish in assimilating nuanced soft information, particularly the divergence between management optimism and analyst skepticism.

Conclusion

The paper conclusively demonstrates that not all voices in an earnings call are created equal. By leveraging FinBERT and an empirically derived weighting scheme that prioritizes analyst sentiment, the authors construct a sentiment signal that is more predictive, economically significant, and robust than traditional methods. This signal captures orthogonal information to earnings surprises and generates substantial risk-adjusted returns, fundamentally altering how textual data should be processed in asset pricing.