Deploying Deep Learning Techniques to Estimate Probability of Default: Way to Data Driven Credit Risk Modeling

Abstract

Correct estimation of probability of default (PD) for credit loans is an essential task for BridgeFund, an online loan broker operating in the Dutch Small and Medium-Sized Enterprizes (SME) market . Advanced machine learning techniques are increasingly being explored to enhance prediction accuracy. Traditional models like logistic regression offer clear interpretability but often lack predictive power compared to more complex algorithms. Ensemble methods and deep learning techniques show potential for significant performance improvements in PD quantification. This study compares XGBoost, Random Forest, Feedforward Neural Networks (FNN) and Tabular Networks (TabNet) against logistic regression to determine their efficacy. The results show that XGBoost outperforms logistic regression and all other models, in all evaluation metrics for PD scoring. However, the "black box" nature of XGBoost raises concerns about model transparency and stakeholder trust, necessitating careful implementation. Developing techniques to demystify XGBoost’s decision-making process such as calculation of SHAP values will enhance the model’s interpretability and, therefore, applicability.