Credit risk analysis using boosting methods

Abstract

The use of credit for various occasions has become a routine in our lives. In return, banking and financial institutions require to determine whether the loan demands from them contain any risk. Accordingly, these institutions have been increased their activities in determining whether credit rating models from past credit records of the person applying for the loan works properly. Machine learning-based technologies have opened a new era in this field. AI and machine learning based methods for credit scoring are currently implemented by banking or non-banking financial institutions. Employed models are to extract meaningful features from the required data in which wide variety of information available. In this study, credit risk assessment is conducted using boosting methods such as CatBoost, XGBoost and Light GBM. To this aim, Kaggle Home Credit Default Risk dataset is used and the effect of crediting tendency on the results is also considered. The results have shown that gradient boosting methods provide results that are close to each other, and crediting tendency produces better AUC score in CatBoost while it causes a small decrement in AUC score of XGBoost and LightGBM.

Enhancing credit scoring accuracy with a comprehensive evaluation of alternative data

“…The DeLong test [ 78 ] confirmed the statistical significance of these AUC differences (p-values < 0.05 for all comparisons). The LightGBM model using the full set of predictor variables achieved the highest AUC (0.79360), consistent with prior studies [ 11 , 13 , 14 ] and surpassing performance reported in previous research on this dataset. Models developed in this study also outperformed logistic regression benchmarks with AUC scores of 0.68031 [ 16 ] and 0.7574 [ 10 ].…”