A reinforcement learning model for lending problems with limited budget and insufficient data

Abstract

Traditional lending policy requires sufficient data for making lending decisions, therefore, some small companies could not access to the fund. In this study, we propose a decision making model that can decide whether to accept or reject a sequence of unfamiliar loan applications while having a limited budget. Our model does not have any knowledge about the incoming loans, therefore, it can predict the default probability with low accuracy at the beginning. The model can learn by observing the outcomes of the accepted loans. The model's budget increases every time the model accepts a fully paid loan and decreases when the model accepts a defaulted loan. The objective of our model is to maximize the final budget. By using the reinforcement learning method, we propose a decision making model that takes the current budget and model accuracy into consideration when making decisions. Based on simulated data, the results show that our model yields a better performance compared to a traditional default prediction model. For the real data, our model performs well in some type of loans.