LLM-Extracted Earnings Sentiment as an Alpha Factor

Abstract

We evaluate frontier LLMs as extractors of forward-looking sentiment from earnings call transcripts. Semantic delta across consecutive calls achieves 3-day alpha of 1.8% with Sharpe 1.72 OOS across the S&P 500.

Research Question

Can frontier LLMs extract actionable forward-looking sentiment from earnings call transcripts, and does semantic delta between consecutive calls provide a robust alpha signal?

Data Sources

• Earnings call transcripts (2022–2026)
• S&P 500 universe
• Model: GPT-4, Claude 3.5 Sonnet
• Pricing data from CRSP

Sample Period

Q1 2022 – Q2 2026 (16 quarters)

Methodology

We prompt frontier LLMs to score forward-looking sentiment in earnings call transcripts on a scale from -5 (strongly negative) to +5 (strongly positive). The factor is constructed from the semantic delta between consecutive quarterly calls for each issuer. We use a 5-day holding period with decaying weights.

Baselines

• Traditional earnings call sentiment (BAG-of-words)
• Price momentum
• Equal-weighted S&P 500 benchmark

Results

The factor achieves a 3-day alpha of 1.8% with an out-of-sample Sharpe of 1.72 across the S&P 500, robust to transaction cost assumptions up to 15bps. The semantic delta signal outperforms traditional BAG-of-words sentiment extraction by 45%.

Transaction Cost Assumptions

• 10bps one-way average
• Higher for small-cap stocks
• Costs estimated from realistic execution slippage

Out-of-Sample Methodology

Time-series cross-validation with quarterly train-test splits, ensuring no future information leakage.

Limitations

• LLM output may be sensitive to prompt engineering
• Transcript quality varies by company
• Does not account for market-wide sentiment shifts
• Model inference costs may be prohibitive for real-time use

Reproducibility Information

Code, prompts, and data processing scripts will be published to the QuantHQ GitHub: https://github.com/quantheadquarters

Related QuantHQ Research

• Earnings Revision Momentum Decay in the Post-2023 Regime
• Yield Curve Inversion as a Regime Classifier for Equity Factor Rotation

Related Blog Posts

• Prompt Engineering for Financial AI: Best Practices — guidance on effective financial AI prompting
• Deflated Sharpe Ratios: How to Account for Multiple Testing — methodology for correcting selection bias
• Five Backtesting Pitfalls That Fake Your Sharpe — common backtesting errors

Keywords

LLM, earnings sentiment, NLP, semantic delta, financial AI, factor investing

@article{quanthq2026llm,
  title={LLM-Extracted Earnings Sentiment as an Alpha Factor},
  author={Rodriguez, M. and Kim, S.},
  year={2026}
}