Zhang, Tiantian (2025) Exploratory Analysis of Correlation between Earnings Call Acoustic Features and Credit Ratings: A FinBERT Validation Approach. Master thesis, Voice Technology (VT).
|
PDF
MScVT25Tiantian-Zhangs6055702executive-speech-credit-ratingvf.pdf Download (2MB) | Preview |
Abstract
This thesis introduces the first systematic investigation of correlation between earnings call acoustic features and companies’ subsequent credit rating outcomes by S&P Global, Moody’s, and Fitch. It develops an innovative framework that bridges speech technology and corporate financial communication analysis using NLP and machine learning approaches. While prior research has established relationships between speech sentiments and securities analysts’ ratings, no studies have examined correlations with international credit ratings - a critical gap given credit ratings’ role in global debt capital markets. This study addresses this intersection using the public Earnings-21 dataset, analyzing earnings calls of 44 US-listed companies, including 24 who received subsequent rating actions (21 affirmations, 2 downgrades, 1 upgrade). The dataset selection required extensive navigation of international legal frameworks, including GDPR compliance for cross-border speech data. Initial attempts to collect proprietary data were systematically explored under fair use doctrine. As explicit consent from data sources was required, the study chose to use publicly available datasets. This regulatory analysis process demonstrates proficiency in compliance requirements essential for research at the intersection of technology and regulated industries. Recognizing the data scarcity in this emerging field, the study employs percentile ranking, bootstrap confidence intervals, and MAD-based effect estimation, approaches particularly suitable for small and imbalanced sample. The study further uses the finance-domain NLP model FinBERT as a text sentiment validator. This innovative multimodal validation framework addresses the fundamental ambiguity that physiological arousal can stem from either financial optimism or distress. Acoustic features, particularly fundamental frequency, pause frequency, and jitter, are extracted and normalized using duration-weighted aggregation to address multi-speaker heterogeneity. The vali-dation framework successfully identifies convergent patterns (aligned acoustic-semantic stress) and divergent patterns (acoustic arousal with positive/neutral sentiment), providing interpretable insights despite limited sample size. The single upgrade case exhibits high acoustic variability coupled with notably negative semantic tone, suggesting complex relationships between speech sentiments and financial outcomes. The study provides an empirical method for integrating multimodal acoustic semantic analysis with financial outcome indicators, while openly acknowledging the limitations imposed by small sample and data imbalance. Future research is recommended to use larger datasets and multimodal fusion mechanisms. The findings highlight this as foundational work toward operational voice analytics for corporate disclosure analysis in credit assessments. Key innovations: novel application domain (earnings call speech-credit rating correlation), multi-modal validation framework without fusion, small-sample robust statistical methodology. Key words: earnings calls, speech sentiments, credit ratings, FinBERT, multimodality, speech technology
| Item Type: | Thesis (Master) |
|---|---|
| Name supervisor: | Schauble, J.K. |
| Date Deposited: | 05 Aug 2025 08:52 |
| Last Modified: | 05 Aug 2025 08:52 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/749 |
Actions (login required)
 |
View Item |