engleski

Deep self-normalizing networks for credit risk assessment

Credit risk assessment process includes evaluation of loan applications (approval of acceptable clients and rejection of clients that are likely to default) using application models, as well as monitoring behavior of existing clients using behavioral models. In this article we propose a deep self-normalizing neural network behavioral model trained on a large contract-level dataset. The proposed deep learning model outperformed conventional logistic regression based methods, with out-of-sample Somers’ D score of 84.08%. Moreover, when comparing accuracy scores with regard to actual month of default in the future, deep model once again exhibits higher predictive power.

self-normalizing nework ; credit risk assessment ; behavioral model

nije evidentirano