Experiments with different methods for class imbalance and bias in customers' credit risk assessment with little or no credit history

Customers defaulting on their loans payment present a big risk for credit providers. Therefore, credit issuers are not inclined to issue financial products or services to retail customers with little or no previous credit history due to the uncertainty about the customers creditworthiness. This results in limited access to financial resources and unfavorable terms for accessible resources. This research focus on improving the predictability of creditworthiness of these underserved groups with the help of ensemble models. Specifically, by applying various methods to handle the imbalance common in dataset used for credit risk assessment. Furthermore, this research attempts to improve fairness of ML model by limiting the group bias through the applications of disparate impact remover, a well known bias mitigation method. In applying three imbalance methods, the class weight was the most effective at handling class imbalance while still maintaining a reasonable performance. Additionally, disparate impact remover can have adverse effect on unprotected bias sensitive features.