Enhancing Credit Scoring in Peer-to-Peer Lending Platforms Through Ensemble Learning

Crediting scoring for P2P lending platforms can be defined as an essential field of improvement as the given work suggests the credit scoring system based on using ensemble learning approaches. Credit scoring systems, that analyze a lot of data, can face difficulties in processing it because of the typical shortcomings of the traditional credit scoring models, namely: data imbalance and an interdependence of multiple features. In order to address these challenges, the proposed methodology adopts a Hybrid Ensemble Learning technique that comprises of the following machine learning algorithms namely Random Forest, Gradient Boosting, XGBoost, and the Stacking Ensemble Model which harnesses the strengths of each of the algorithms. Some of the aspects of the methodology include enriched feature extraction that is used to increase the models ability to learn and SMOTE used when dealing with imbalanced data between defaulted and non-defaulted loans. Furthermore, the SHAP (Shapley Additive Explanations) importance is applied to the models to describe the distribution between the features and the decision about credits scoring. Such interpretability is important for a financial application especially because of the need to assure end-users about the operation and output of the system. Various classifications and activities of the models are then assessed by using important parameters such as accuracy, precision, recall, F1 measure, as well as AUC-ROC. Such findings indicate that the stacking ensemble model has a higher accuracy compared with the individual models at 92.4% and the Area Under the curve–Receiver Operating Characteristic (AUC–ROC) of 94.5%. As shown in the evaluation of the model, the incorporation of feature engineering as well as SMOTE highly enhanced the operation of the model, especially in the aspect of recall and precision with regard to loan defaulting.