Loss Given Default (LGD), a crucial metric in credit risk assessment, has been extensively examined in corporate lending and credit cards, yet its exploration in online microloans remains limited. In this study, we identify key macro variables for online microloan LGD modeling by screening a dataset containing 147 commonly used macro variables. The lasso method encounters challenges due to multicollinearity among variables, prompting the adoption of the latest statistical screening method, Group \(l_0\) . Our findings reveal that variables related to employment and business operations are significantly correlated with LGD. Then, we employ quantile regression and a neural network model to predict the LGD for each online microloan, enhancing predictive accuracy through the state-of-the-art statistical method, quantile conditioned moments (QCM). The theoretical analysis reveals that the convergence condition for estimates derived from QCM is weaker compared to using a benchmark quantile. This implies that our method performs more robustly than using the benchmark quantile as an estimate for LGD. Our approach, QRNN+QCM, demonstrates superior predictive power compared to the three machine learning models widely used in the literature, i.e, QRNN, XGBoost, and Random forest. Furthermore, we conduct a comprehensive analysis of covariate heterogeneity in the prediction model which provides valuable insights.
