Modelling Defaultable Loan Repayment Cashflows Using Neural Hawkes Processes

Abstract

We propose a novel framework for modelling defaultable loan repayment cashflows using Hawkes processes, where missed repayments are captured via a self-exciting point process and transition dynamics are learned through a recurrent neural network. Using a loan-level dataset of over 1,400 mortgages from 2014 to 2024, we demonstrate the model’s ability to accurately classify and simulate repayment behaviour across performing, missed-payment, bankruptcy, and restructuring states. The model achieves 94.8% accuracy in forecasting repayment states and passes the Kupiec Value-at-Risk test at 90%, 95%, and 99% confidence levels with consistently high pass rates across borrower credit categories, indicating strong calibration and value for risk assessment.