REAL TIME CREDIT RISK MONITORING WITH ML

Abstract

Credit risk is the possibility that a borrower will fail to repay a loan or meet other credit obligations on time. Banks, non-banking finance companies, and digital lenders assess this risk when a loan is sanctioned, but the financial condition of a borrower can change considerably during the life of the loan. Loss of income, rising debt from other lenders, or changes in spending behaviour can move a borrower from low risk to high risk within a few months. This paper presents a real-time credit risk monitoring system that uses machine learning to update the risk level of existing borrowers continuously as new repayment and transaction data arrives. The system uses a combination of static borrower attributes, such as age, income, employment type, loan amount, and tenure, and dynamic behavioural data, including monthly repayment status, days past due, credit utilisation, account balance trends, and the number of recent credit enquiries. Publicly available credit datasets are used for training, and a streaming simulator replays transaction and repayment events to reproduce the conditions of a live environment. The raw data is cleaned to handle missing values, inconsistent categories, and extreme outliers, and features describing recent behaviour are computed over rolling windows of one, three, and six months. Because defaulting borrowers form a small minority of the data, class imbalance is handled with resampling techniques and class weights. Several classifiers are trained and compared, including Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, XGBoost, and LightGBM. The models are evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and the Kolmogorov-Smirnov statistic, which is commonly used in credit scoring. The results show that gradient boosting methods provide the best discrimination between borrowers who will and will not default within the next three months, with days past due, utilisation, and balance trend being the strongest indicators. The trained model is deployed as a scoring service that receives new events and recalculates the probability of default for the affected borrower within seconds. Borrowers are placed into risk bands, and an alert is generated when a borrower moves to a higher band. A web dashboard shows the overall portfolio risk distribution, recent migrations between bands, individual borrower profiles with the factors behind their current score, and the performance of the model over time. The system is intended to help lenders identify deteriorating accounts early, so that they can contact borrowers, restructure loans, or adjust credit limits before the account becomes non-performing. It supports rather than replaces the judgement of credit officers, and every automated alert is reviewed before action is taken. Future work includes fairness assessment across borrower groups, integration of account aggregator data, and automatic detection of model drift.