For a company that extends credit to its customers, in order to be profitable, the most important criterion for approving applications is whether they can pay back their debts. This is determined by a process called credit scoring that is based on the financial history and socio-economic information of the customer. Traditionally, for credit scoring, scorecards have been used, although in recent years, these simple models have given way to more sophisticated machine learning models. Scorecards are basically checklists of different items of information, each associated with points that are all added up in the end and compared to a pass mark.
We'll use a relatively small dataset of credit card applications; however, it can still give us some insights into how to do credit scoring with neural network models. We'll implement a model that includes a distribution of weights as well as a distribution over outputs. This is called epistemic, aleatoric uncertainty...
