Applying CRISP-DM Methodology in Developing Machine Learning Model for Credit Risk Prediction

Abstract

Banks and other financial institutions must assess credit risk when deciding on loan applications. Approving loans to ‘risky’ individuals is the largest source of financial loss. In other words, borrowers who default cause the largest amount of loss to the banks, and these institutions need to minimize this risk, before agreeing to approve loans. Quantitative modeling using machine learning techniques can now be used to get better insights from data, automate the process, reduce data management, and increase overall profitability. We used the CRoss Industry Standard Process for Data Mining (CRISP-DM) methodology in developing the machine learning solution for the given task of building a binary classifier that can predict loan applicants who are likely to default and who are not. Both linear and non-linear models were evaluated as baseline models to identify the potential candidate model for the final design solution to achieve the highest performance in terms of accuracy and recall values. Performance results are presented in the form of a 2-by-2 confusion matrix, classification reports, and ROC curve for easy understanding. The great benefit of the proposed design solution is that the bank can deploy a model that can automatically filter out potential customers who might default thus reducing data management requirements and the burden on underwriters.