Credit Risk Scoring Analysis Based on Machine Learning Models

In the big data era, institutions can easily access a massive number of data describing different aspects of a user. Therefore, credit scoring models are now building from both the past credit records of the applicant, and other personal information including working years and characteristics of owned properties. A wide variety of usable information has required models to extract more expressive features from data and apply the effective models to fit the features. This paper reports our efforts in using feature engineering techniques and machine learning models for credit scoring modeling. Based on the Kaggle Home Credit Default Risk dataset, several current feature engineering techniques and machine learning models have been tested and compared in terms of the AUC score. The results have shown that the LightGBM model training on expert knowledge generated datasets can achieve the best result (About 78% AUC score).