Case Study on Utilizing Machine Learning in Corporate Default Risk Prediction : A practical Implementation to Credit Risk Management Process

Abstract

The purpose of the case study is to create an in-house corporate default risk prediction model that outperforms the external corporate credit rating which the case company is currently using for this purpose. In addition, the study sets the framework for implementing the model into current system architecture and credit risk management process. The study consists of literature review and empirical analysis where the default prediction models are built and tested and the proposal for implementing the model into case company’s system architecture and processes is given. The data used in this study consists of historical financial figures & ratios, payment behaviour information and other background information of 2471 Finnish companies from period 2009-2017 of which 22,6% defaulted during this period. MissForest method was used in imputation of the missing values. The models used in this study are Multivariate Discriminant Analysis, Logistics Regression, Random Forest, CART, AdaBoost, Support Vector Machine and Neural Network. The dataset was split with 70/30 ratio to training and test set and 10-fold cross validation was used in training, feature selection and hyperparameter optimization for each model. Model performance was also tested over a two-year time horizon. The models’ performance was measured with ROC AUC & PR AUC and Brier Score. All the models overperformed the external credit rating with the selected metrics. The best performing model was the black box model Adaboost and the best performing white box model was the logistic regression with LASSO method used for the predictor variable selection.