Incorporating metadata: A novel variational neural topic model for bond default prediction

Abstract

Credit rating reports, especially the analyst rating opinions with positive and negative tendencies, offer key insights into the risks influencing bond issuers' ability to repay debt, making them vital for bond default prediction. However, meaningful topics are challenging to extract from these texts. Positive and negative rating opinions provide metadata that could guide topic modeling to capture both strengths and risks, but existing methods struggle to reflect these nuances. Additionally, the brevity of opinions leads to sparse data, which complicates topic extraction. To address these issues, MetaNTM is proposed as a metadata-guided variational neural topic model, which features a variational topic distribution module to manage sparsity and a metadata-guided attention mechanism to integrate rating tendencies. Experiments on Chinese corporate bond data show that MetaNTM outperforms benchmark topic models, achieving higher topic coherence. The topic features also enhance predictive accuracy across multiple bond default prediction models. Logistic regression achieves a 42.0% improvement in H-measure (from 0.355 to 0.504), random forest records a 12.9% increase in KS (from 0.605 to 0.683) and a 22.2% rise in H-measure (from 0.414 to 0.506), and XGBoost achieves a 7.2% increase in H-measure (from 0.581 to 0.623). These results underscore the predictive power of MetaNTM's topic features.