Uncertain drivers and potential biases of the BERT-based default risk score

Identify and characterize the primary factors that drive the BERT-based default risk score generated from Lending Club loan descriptions, and ascertain whether the score exhibits systematic biases that affect its predictions.

Background

The paper fine-tunes BERT on loan descriptions from the Lending Club dataset to produce a default risk score and shows that incorporating this score improves a conventional XGBoost granting model. Despite performance gains, the authors emphasize that the model’s opacity limits interpretability.

They explicitly state that the underlying determinants of the score and any potential biases are not presently known, highlighting the need for transparency to meet regulatory expectations and build user trust.