Resumo

Credit scoring is an efficient tool that allows financial institutions to distinguish their potential default borrowers. To this extent, researchers have developed a myriad of approaches, combining statistical and artificial intelligence techniques, to fulfill the task of credit scoring. Different studies have signaled that the combination of methods – ensembles -, which test different hypotheses to form a new hypothesis, typically outperform the other traditional credit scoring approaches. In this paper, we evaluated a machine learning algorithm called Extreme Gradient Boosting. The proposed

Credit scoring modeling

Basel II Accord requires financial institutions to disclose Probability of Default (PD), Exposure at Default (EAD), and Loss Given Default (LGD) in such a manner that accuracy and speed to analyze and predict data are paramount. Thus, an ever-increasing demand for smart and agile processes have exposed an opportunity for machine learning research in credit risk management. Basel II also requires companies to disclose risk management practices, which demands more reliable and accurate models to classify and quantify risk. (BIS - BANK FOR INTERNATIONAL SETTLEMENTS, 2006). For this reason, the ad

This present study tested Extreme Gradient Boosting (XGBoost), a state-of-the-art machine learning method, which is used for supervised learning problems (Chen and Guestrin, 2016), which the term Gradient Boosting was proposed by Friedman (2001). XGBoost is an enhancement and based on his original model. We have chosen this model because of demonstrated efficiency, accuracy and practicability of its algorithm (Chen and Guestrin, 2016). Besides that, its capacity to do parallel computation on a commonplace machine causes it to be alluring. It also has additional features for doing cross valida

In our study, we confirmed its performance and accuracy when compared to benchmark models (Logistic and Random Forest). XGboost had and accuracy rate of 82.29% against 82.01% from Random Forest and 81.93% from Logistic Regression.

From the results, we believe that credit-lending corporations can develop a cautionary system that would warn clients on behaviors that could affect their credit score. Adding to that, managers can avoid financial default by taking early appropriate action rather than waiting for the event to happen.