Credit Default Prediction from User-Generated Text in Peer-to-Peer Lending Using Deep Learning

Kriebel Johannes, Stitz Lennart

Abstract

Digital technologies produce vast amounts of unstructured data that can be stored and accessed by traditional banks and fintech companies. We employ deep learning and several other techniques to extract credit-relevant information from user-generated text on Lending Club. Our results show that even short pieces of user-generated text can improve credit default predictions significantly. The importance of text is further supported by an information fusion analysis. Compared with other approaches that use text, deep learning outperforms them in almost all cases. However, machine learning models combined with word frequencies or topic models also extract substantial credit-relevant information. A comparison of six deep neural network architectures, including state-of-the-art transformer models, finds that the architectures mostly provide similar performance. This means that simpler methods (such as average embedding neural networks) offer performance comparable to more complex methods (such as the transformer networks BERT and RoBERTa) in this credit scoring setting.

Keywords
OR in banking; Peer-to-peer lending; Deep learning; Textual data; Credit risk