Non-Performing Loan Prediction Using Machine Learning Techniques; Case Study for a Non-Banking Financial Institute in Sri Lanka

Abstract

In the realm of banking and financial institutions, the pivotal role of collecting public deposits and disbursing funds through various loan products underscores the importance of managing loan portfolios efficiently. Non-performing Loans (NPLs), characterized by continuous outstanding balances over four installments or 120 days, pose significant risks to financial stability. This study focuses exclusively on customer-related factors influencing NPLs within ABC Finance PLC, a leading non-banking financial institute in Sri Lanka, specifically analyzing vehicle loan data from the previous fiscal year. The objective of this research is to employ Machine Learning techniques to predict the likelihood of loans becoming NPLs based on customer, loan, vehicle, and account-related details. The study utilizes secondary data sourced from the lending database, employing a training dataset split in an 80:20 ratio for model development and testing. Key steps include preprocessing, exploratory data analysis, feature engineering, and model selection. Several Machine Learning models, including Naïve Bayes, K-Nearest Neighbors, Logistic Regression, Random Forest, Gradient Boosting Classifier, XG Boosting Classifier, and Decision Tree, are trained and evaluated using metrics such as Confusion Matrix, Accuracy, Precision, Recall, and F1 score. The findings from this analysis aim to empower responsible stakeholders, including recovery managers and branch heads, to proactively manage loan portfolios and mitigate NPL risks effectively. The analysis of machine learning models on the vehicle loan dataset from ABC Finance PLC yielded valuable insights into model performance and predictive capabilities. Among the evaluated models, XGBoost emerged as the top-performing algorithm, particularly after hyperparameter tuning. Post-tuning, XGBoost achieved the highest test accuracy of 0.7221 and an impressive Area Under the ROC Curve (AUC) of 0.8, demonstrating enhanced discriminatory power and predictive accuracy compared to other models. The significant improvement observed in XGBoost's performance highlights the effectiveness of hyperparameter tuning in optimizing model parameters and enhancing predictive capabilities. This finding underscores the importance of leveraging advanced machine learning techniques, such as XGBoost, to proactively identify and manage non-performing loans within financial institutions. In conclusion, the study's findings provide actionable insights for stakeholders at ABC Finance PLC to implement data-driven strategies for loan portfolio management and risk mitigation. By leveraging Machine Learning models like XGBoost, financial institutions can enhance decision making processes, optimize resource allocation, and minimize the impact of non-performing loans on financial stability. Future research directions may focus on incorporating additional data sources and advanced modeling techniques to further improve predictive performance and address evolving challenges in loan portfolio management.