Reproduce “Same Company, Same Signal: The Role of Identity in Earnings Call Transcripts” on OpenResearch

About this paper

Ding Yu, Zhuo Liu, Hangfeng He

Researchers at the University of Rochester demonstrated that post-earnings volatility prediction is predominantly driven by historical, company-specific volatility patterns rather than semantic content from individual earnings call transcripts. The study introduced a new, dense dataset (DEC) and showed that simple, training-free historical baselines consistently outperform complex LLM-based models, as transcript embeddings primarily reflect ticker identity.

Key insights

• Each company exhibits a distinct and persistent 'Volatility Signature,' where its historical post-earnings volatility patterns are highly indicative of future volatility.
• The predictive power attributed to earnings call transcripts primarily stems from their ability to capture ticker identity and approximate prior post-earnings volatility distributions, rather than extracting novel, event-specific semantic insights.
• Simple, training-free historical baselines can achieve state-of-the-art performance, outperforming complex, resource-intensive NLP and LLM models for post-earnings volatility prediction.

Summary by alphaXiv.