Extracting Alpha from Earnings Transcripts: Our NLP Methodology

This post details the methodology behind our EARNINGS_SENTIMENT_DIVERGENCE signal, which identifies short-term alpha opportunities by measuring the gap between management optimism and analyst skepticism in earnings call transcripts.

The Intuition

Earnings calls have two distinct sections: the prepared remarks (management's narrative) and the Q&A (analysts probing). Management has incentive to present positively. Analysts have incentive to probe weaknesses. When these two perspectives diverge significantly, it often signals information asymmetry — management knows something the consensus hasn't priced in yet.

The Pipeline

Our NLP pipeline processes transcripts in three stages:

Stage 1: Speaker Diarization & Section Classification We classify each utterance into "management prepared remarks," "management Q&A response," and "analyst question." This matters because management tone in prepared remarks (scripted, forward-looking) carries different information content than their tone in Q&A responses (spontaneous, defensive or confident).

Stage 2: Sentence-Level Sentiment Scoring Each sentence receives a sentiment score from a finance-specific transformer model fine-tuned on 50,000 manually-labeled financial statements. The model distinguishes between financial positive ("revenue growth accelerated") and generic positive ("we're excited about the future") — only the former is predictive.

Stage 3: Divergence Computation The divergence score = mean management sentiment (forward-looking statements only) minus mean analyst sentiment (Q&A questions only). We normalize by the historical standard deviation for each company to account for baseline differences in communication style.

Why It Works

Companies where management is significantly more optimistic than analysts (divergence > 2σ) tend to see positive price moves over the following 3-5 days. Our hypothesis: management has private information about pipeline, orders, or operational improvements that hasn't yet been reflected in consensus estimates. The Q&A questions reveal that analysts haven't yet incorporated this information.

The reverse also holds: when analysts are more optimistic than management (divergence < -2σ), it often signals management sandbagging ahead of guidance revisions or negative surprises.

Performance

• •Backtest period: Q4 2023 — Q3 2025 (8 quarters)
• •Universe: US mid-cap tech ($2B-$15B market cap)
• •Sharpe ratio: 1.4 (1.7 excluding Q2 2024 timing anomaly)
• •Win rate: 62%
• •Average holding period return: +1.8% (long), +0.9% (short)
• •Maximum drawdown: -4.2%

The signal is currently live in ARKRAFT at 0.3 weight, following our new-signal ramp-up protocol.