Emotion Analysis of Financial Earnings Call Transcripts
12,521 earnings-call transcripts (2010–2020) analysed for CEO and analyst emotion, linked to stock returns
A finance-domain emotion analysis pipeline over more than 12,500 earnings-call transcripts, tracking emotional agreement and conflict between speakers and correlating the results against market variables.
- Python
- NLP
- Word embeddings
- Deep emotion extraction (Emotion AWARE)
- Document parsing
- Financial lexicons
- Statistical modelling
Problem
Earnings call transcripts carry a lot of signal about how a company’s leadership actually feels about its own results, signal that never shows up in the reported numbers. 12,521 transcripts of earnings conference calls from 2010 to 2020, involving the CEO, senior management, analysts and investors, sat as raw text with no systematic way to connect what was said to how the market subsequently moved.
Approach
The pipeline split into document handling and the emotion analysis itself:
- Document parsing and information extraction: parsing the source PDFs, segmenting each transcript into its presentation and Q&A sections, and mapping individual utterances to the participant who said them.
- Deep emotion analysis: per-participant, per-section emotion extraction using Plutchik’s wheel alongside Loughran–McDonald financial-domain lexicons, adapted for the finance domain. Plain positive/negative sentiment turned out to be too coarse: “there was limited growth in the economy” needs a magnitude-aware read, not a binary label, and a domain lexicon is what tells “significant growth” and “limited growth” apart. On top of that, the pipeline tracked emotion transitions across the meeting and quantified where CEO and analyst emotion agreed or conflicted, turn to turn.
- Comparison against market variables: cumulative abnormal returns and trading volume around each call, to test whether the emotional trajectory of a call carried information the raw text sentiment didn’t.
Result
Across regressions controlling for firm size, book-to-market ratio and volatility, a firm’s stock price moved positively following calls where the CEO and management expressed more positive emotion (joy, anticipation, trust), and negatively with more negative emotion (anger, sadness, fear, disgust). That relationship held after controlling for standard financial sentiment word lists. The work built on the Emotion AWARE framework published in the Journal of Big Data, and the market-returns analysis was published as Do Emotions Matter? The Role of Manager Emotions on Stock Returns (Langer, Gamage, Ranasinghe, De Silva, Mather, 2023).
What I’d do differently
Plain sentence-level positive/negative scoring was the first thing tried, and it degraded fast on hedged financial language: “limited growth” scored close to neutral instead of cautious. I’d start with magnitude-aware, domain-adapted lexicons from the outset rather than discovering the gap only after the simpler version was already built.