Enhancing accuracy and interpretability in corporate credit rating classification with the transformer-LSTM model

Abstract

Corporate credit rating classification is essential for assessing a corporation’s debt repayment ability. Previous research has demonstrated that neural network models exhibit high classification accuracy, particularly when incorporating time-series features. However, a significant challenge remains regarding their interpretability, often limited by nonlinear and intricate calculation processes. To address this trade-off between interpretability and the utilization of time-series features, we introduce a novel approach: the Transformer-Long Short-Term Memory (T-LSTM). Specifically, the attention matrix embedded within the T-LSTM architecture provides interpretability by revealing the temporal importance of features. Comparative experiments show that our T-LSTM model surpasses standard machine learning baselines in the long-history setting. Empirical results demonstrate that the proposed model, trained on a 10-year history of longitudinal financial ratios, yields an absolute accuracy improvement of approximately 1 percentage point when compared with a strong sequential baseline such as a Long Short-Term Memory (LSTM) model trained on the same 10-year history, and up to about 17 percentage points when compared with representative single-year “snapshot” baselines that use only the most recent year’s ratios. Furthermore, the attention matrix successfully visualizes specific time points where information is most critical for rating classification. Consequently, the proposed model offers a highly accurate and interpretable solution for credit rating classification in the financial industry.