Combination of Stacking with Genetic Algorithm Feature Selection to Improve Default Prediction in P2P Lending

Lending Club has become a popular P2P lending platform in America and has increased with a total loan of $13.4 billion in 2015. However, the development of the P2P market with online-based intermediaries allows for credit risk arising from defaults that can directly harm investors and threaten the development of P2P platforms. This study aims to determine a classification method that has the best accuracy for predicting credit risk, which reduce the emergence of risk in P2P lending practices. The method applied is data collection from the Lending Club platform, followed by preprocessing stage that applies the one-hot encoding method to handle categorical data and random oversampling to balance the data classes. The proposed classification method is a stacking ensemble combining three single classifiers, LightGBM, XGBoost, and Random Forest as a base learner. GAFS is also proposed as a feature selection method to get predictive attributes or features. The findings suggest that prior to implementing the GAFS (Genetic Algorithm Feature Selection) technique, the stacking ensemble learning achieved an accuracy of 98.05%. However, after incorporating GAFS, which identified 11 relevant features, the accuracy improved to 98.97%. From the accuracy results, it show an increase of 0.92%. Thus, the stacking method based on the result features of GAFS has demonstrated superior accuracy in predicting the risk of default on P2P loans.