A Novel End-to-End Corporate Credit Rating Model Based on Self-Attention Mechanism

In traditional financial analysis discriminant analyses with statistic methods are considered as primary and widely applied corporate credit rating bases in which human impact plays an essential role. Machine learning approaches enormously assist credit rating in reducing human interference and improving rating efficiency. However, existing methods did not efficiently leverage the time series features of financial data that is beneficial for the classification tasks. We propose a novel end-to-end architecture, the SMAGRU, based on self multi-head attention with gated recurrent neural networks, which is capable of evaluating a corporate credit rating by capturing the time series features assigned the weights of market benchmarks. Experiments compared with statistic and machine learning baseline models on different datasets show that our model contains superiority in multi-class credit rating classification. The result of ablation experiments suggests that time series features are crucial in classification performance, and self multi-head attention can significantly enhance these features and thereby improve the accuracy and convergence speed. We show that the SMAGRU generalizes well, even on sparse data during K-fold cross validation experiments.