Branch-Level Mortgage Credit Risk Classification Using Loan Origination System Data for Managerial Decision Support

Abstract

This study addresses the challenge of early and accurate mortgage credit risk classification using operational data from the Loan Origination System (LOS) of Bank BJB. In this study, LOS consistently refers to the Loan Origination System used to record and manage mortgage loan operational data. The core research problem lies in the limited availability of granular borrower-level data in many banking environments, where credit risk information is commonly managed in aggregated branch-level form. To address this gap, this study develops a branch-level classification framework adapted to panel data by applying ratio-based, growth-based, volatility-based, and lag-based feature engineering techniques to enrich a limited set of original LOS variables. Two supervised classification models, Logistic Regression and Random Forest, were implemented and evaluated using mortgage loan data from 10 branch offices covering the period 2019–2024. The dataset consists of periodic branch-level observations constructed from three original LOS variables: outstanding loan amount, number of active accounts, and non-performing loan value. The binary risk label was defined using the 5% NPL ratio threshold, resulting in an imbalanced but operationally meaningful classification problem between high-risk and low-risk branches. Both models demonstrated strong discriminative performance, with AUC values of 0.91 for Logistic Regression and 0.90 for Random Forest. Logistic Regression produced better high-risk recall and a lower false negative rate, with recall of 0.79 and a false negative rate of 20.59%, while Random Forest achieved higher overall accuracy and high-risk precision but produced a higher false negative rate for high-risk cases. Feature importance analysis confirmed that lagged NPL ratio and lagged NPL value are the most influential predictors of future branch-level credit risk classification, followed by NPL volatility and average loan size. The selected classification model was integrated into a Decision Support System (DSS) dashboard that converts risk probabilities into ranked risk categories and management recommendations. Although the study is limited by its use of aggregated branch-level data from a single bank and one credit product, the findings show that meaningful branch-level risk classification signals can be extracted from routinely available LOS data without requiring individual borrower profiles.