Narrative forecasts

Abstract

We propose a novel methodology to identify managerial beliefs from earnings call transcripts, using lexicon-based and FinBERT sentiment analysis alongside machine-learning guided topic modeling. We provide a dual contribution to the literature. First, we find that managerial sentiment significantly predicts analyst forecast revisions, with presentation sentiment showing stronger associations than question and answer (Q&A) interactions. Second, we show that these sentiment-driven revisions lead to systematic forecast errors, suggesting that narrative content shapes analyst expectations beyond fundamental information. Our analysis offers a scalable alternative to traditional survey-based approaches for measuring economic beliefs, providing high-frequency and near-universal coverage across firms and time.