Abstract:
Finance companies equally play a vital role in the Financial System in Sri Lanka. Loans and Advances represent the majority of the asset class of the financial position which accounted for 76.8% of the total assets in 2021. While financing through credit being a driving force of the economy, it has become the biggest risk for any financial institution. Non- Performing Loans (NPL) of the Non- Banking Finance sector have also been on the rise, reaching 13.9% and 11% in 2020 and 2021. As a solution to the problem statement, we have conducted a case study on employing machine learning models for credit risk modeling in a well reputed Finance Company in Sri Lanka. Five machine learning models namely, Logistic Regression, Decision trees, Random Forest, XGboost and Adaboost were used and evaluated the performance of each model. It was revealed that XGBoost model outperforms the other models with the highest model performance. (Accuracy (0.78), Weighted Avg Precision (0.87), Weighted Avg Recall (0.77), Weighted Avg F1-score (0.82) and AUC-ROC (0.60)). It was evident that feature selection and hyperparameter optimization will impact the performance of the model. Correlation coefficient heatmap, Chi- Squared test of independence, Select K best, Recursive Feature Elimination and Random Forest Feature importance were used as feature selection. The highest performances were shown with Random Forest feature importance's. It was observed that the importance of features namely, No of Rental (Term), Effective Rate, Rental, Age, Income, LTV and vehicle age are found to be significantly high. Addition to that, we discuss model interpretability using LIME method.
