Transfer Learning and Loan Default Prediction

Predicting probability of default for potential loan customers is of the utmost importance to banks and other financial institutions. Nowadays, most financial institutions assess borrowers using machine learning algorithms. However, they may require large amounts of training data to make accurate predictions. Small financial institutions may not be able to collect large training datasets, and would benefit from the large datasets or pretrained models provided by larger financial institutions. This paper employs the use of transfer learning with neural networks to predict probability of default for new borrowers. We explore multiple architectures of deep neural networks trained on a large dataset and transfer the learned knowledge to a smaller dataset.