LOAN DEFAULT PREDICTION IN MICROFINANCE GROUP LENDING WITH MACHINE LEARNING

• Main Author: Dwi Pradnyana, Kadek
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/71192
• Institution: Institut Teknologi Bandung

PT Gemilang, a microfinance fintech, enables the unbanked and underbanked communities to access credit by offering small, no collateral loans. The company focuses on group lending microfinance where the borrowers are women micro-entrepreneurs in Indonesia and are put together as a group. Each member of the group are jointly responsible for loan repayment. Due to a large number of borrowers failing to repay their loans, the company is currently facing an increase in non-performing loans. Although the company uses credit scoring methods to evaluate the risk of potential borrowers, these methods are now outdated and inaccurate, which means they are not able to filter out bad loans effectively. As a result, the company is struggling to manage its high level of non-performing loans. This research attempts to build a credit default prediction model for microfinance group lending using machine learning techniques. While credit scoring techniques for individual loans have been extensively studied, there is a lack of research on credit model for group loans. We examine six different machine learning methods, including XGBoost, logistic regression, linear discriminant analysis (LDA), decision trees, k-nearest neighbour (KNN) and random forest. The methods are then evaluated using accuracy and ROC AUC score. The XGBoost model performs the best during the first modeling phase. With an accuracy of 0.97 and an AUC score of 0.85, it performs better than other models. Decision tree and random forest give comparable outcomes, with AUCs of 0.81 and 0.80 and accuracies of 0.81, 0.95, and 0.97. In an effort to increase performance, class balancing is performed. it adjusts the class weights of the majority and minority classes during the model training phase. As a result, class weights will be in charge of allocating equal weights to both categories. The XGBoost model's performance was successfully enhanced, resulting in an increase in AUC from 0.85 to 0.89. Its accuracy stays the same as 0.97. False positive and false negative rates for this model are both low (2.05% and 1.38%, respectively). Consequently, the model has been effectively developed and is capable of differentiating between bad and good loans. This study demonstrates that it is feasible to develop a model with strong predictive ability for microfinance group financing. On top of that, it is advised that the model be incorporated into the PT Gemilang risk management system. It can assist the company identify bad loans and only book good loans. This will result in a higher repayment rate, a higher quality loan, and a low number of non-performing loans. The final model performance with XGBoost algorithm is expected to dramatically minimize the number of bad loans in the company. The proposed model has a recall of 0.79, in other words it can detect 79% out of total bad loans. If the number of bad loans in nature is 7% then, the reduction helped by this model will be 5.5%, leaving only 1.5% of bad loans in the company loan portfolio.