CONSUMER CREDIT RISK MODELLING USING HYBRID ENSEMBLE MODELS: EVIDENCE FROM ESTONIA

Abstract

The purpose of this research is to forecast consumer credit risk modelling and compare the performance of hybrid ensemble models and traditional ML models. The dataset used for this study is the Bondora data which spans from February 2009 to July 2021. The study’s main aim is to investigate the performance of hybrid ensemble models and whether they offer better predictive power compared to traditional models. The study utilises the Random Forest, Support Vector Machine, Extreme Gradient Boosting algorithm, Logistic regression and a hybrid ensemble model of these individual models. The performance metrics that were considered to assess the strength of the predictive model were accuracy, precision, recall and the f-measure. The study findings show that the hybrid ensemble model had the best performance with an f-measure score of 0.738 while the logistic regression was the worst individual learner. The random forest and extreme gradient boosting model also had good individual performances with the SVM being the weakest learner amongst the ML models. The study thus concluded that hybrid ensemble models have a good predictive accuracy and with enough data and good preprocessing techniques, default risk for client can be predicted more accurately through ensemble models.