CRAXNet: Credit Rating via Advanced XGBoost and Neural Networks

Abstract

One of the most important criteria for evaluating corporate creditworthiness in the financial services sector is credit risk analysis. This paper presents a new two-stage model CRAXNet for corporate credit rating. CRAXNet combines the feature selection of XGBoost and the nonlinear pattern learning ability of Neural Networks (NN) to make high-accuracy credit score predictions. CRAXNet, unlike the studies in the literature, provides a unique architecture that provides the class probabilities generated by XGBoost as inputs for the classifier in the NN model. Thus, CRAXNet can successfully model relationships in complex financial data with linear and nonlinear patterns. Experimental results using two different public datasets confirm that CRAXNet outperforms five State of the Art (SOTA) baselines (KNN, FIKNN, AF, Doc2Vec, and CART) with up to 4.74% accuracy and 9.86% F1-Score performance improvement. The datasets and source code used in the paper are publicly available for future researchers.