Textual Analysis of Earnings Calls for Predictive Risk Assessment: Evidence from Banking Sector

Abstract

This paper investigates the predictive relationship between linguistic patterns in earnings calls and subsequent risk events in banking institutions. Through comprehensive analysis of 2,480 transcripts from 62 banking institutions across North America and Europe spanning 2016-2023, we develop a multi-dimensional linguistic analysis framework that extracts and quantifies features including sentiment metrics, uncertainty markers, question evasiveness, and forward-looking statements. The research employs a combination of natural language processing techniques and machine learning models to establish correlations between textual features and risk materializations. Results demonstrate that linguistic features significantly enhance risk prediction capabilities beyond traditional financial indicators, with the integrated model achieving an AUC of 0.845 compared to 0.642 for financial metrics alone. Uncertainty-related language emerges as the strongest predictor across all bank types, with distinctive cross-sectional differences observed between global systemically important banks and regional institutions. Temporal analysis reveals progressive deterioration of linguistic indicators over multiple quarters preceding risk events, with uncertainty indices increasing from 0.014 to 0.045 and question evasiveness scores rising from 0.211 to 0.422 in quarters leading to significant risk materializations. The findings offer practical applications for regulatory oversight and market participants, enabling earlier identification of potential financial instability through systematic analysis of management communications approximately 60-90 days before risk events materialize in traditional metrics.