Optimizing credit risk model classification using bert deep learning model and vader techniques

Loading...
Thumbnail Image

Date

Journal Title

Journal ISSN

Volume Title

Publisher

University of the Witwatersrand, Johannesburg

Abstract

Credit risk assessment remains a fundamental component of financial risk management, traditionally relying on structured quantitative indicators such as credit scores, finan- cial statements, and repayment histories. However, these approaches often overlook contextual and sentiment-driven signals embedded in unstructured textual data. This study investigates the optimization of credit risk classification through the integration of Natural Language Processing (NLP) techniques, specifically transformer-based models including BERT, RoBERTa, and LLaMA 3.2, alongside the rule-based sentiment analysis tool VADER. Financial news articles were collected using the NewsCatcher API over a one-year period and used as the primary source of unstructured textual data. Sentiment extracted from financial news is treated as an indirect indicator of credit risk–related signals rather than a direct borrower-level default prediction mechanism. Model per- formance was evaluated using Accuracy, Precision, Recall, and F1-score. Comparative analysis against traditional machine learning models revealed that XGBoost achieved the highest overall accuracy of 94%, followed by Logistic Regression at 92%. Among transformer-based models, BERT achieved 91% accuracy and RoBERTa 90%, while LLaMA 3.2 recorded 78%. Class-level performance further highlighted differences in contextual understanding. XGBoost achieved an F1-score of 0.97 for Positive sentiment and 0.88 for Neutral sentiment, demonstrating strong balance across classes. BERT achieved an F1-score of 0.79 for Negative sentiment but struggled with Neutral sen- timent (0.62), whereas RoBERTa achieved a more balanced Neutral F1-score of 0.67. Logistic Regression performed competitively with a Neutral F1-score of 0.87. LLaMA 3.2 exhibited significant bias toward Positive sentiment and failed to effectively classify Negative and Neutral categories. The findings demonstrate that while classical models achieved superior overall accuracy, transformer-based models provide valuable con- textual and semantic insights for sentiment-driven credit risk analysis. This research contributes to the advancement of NLP-enhanced financial risk modeling and estab- lishes a foundation for future work addressing class imbalance, model interpretability, and hybrid modeling approaches.

Description

A research report submitted in fulfillment of the requirements for the Master of Science, in the Faculty of Science, School of Computer Science and Applied Mathematics, University of the Witwatersrand, Johannesburg, 2025

Citation

Raliphada, Pfarelo. (2026). Optimizing credit risk model classification using bert deep learning model and vader techniques [Master’s dissertation, University of the Witwatersrand, Johannesburg]. WIReDSpace. https://hdl.handle.net/10539/50262

Endorsement

Review

Supplemented By

Referenced By