Prepayment forecasting in a low-interest-rates environment: Evidence from the French market
TL;DR
This work model prepayment on French residential securitized mortgages observed quarterly from August 2020 through February 2022 and benchmark a single-hidden-layer neural network against logistic regression, deeper network architectures, Random Forest, and XGBoost, which proves the interaction and threshold effects that traditional models like logistic regression cannot represent.
Abstract
Prepayment risk creates cash flow uncertainties that complicate valuation, hedging, and risk management in mortgage lending. Prediction is indeed difficult: borrowers respond nonlinearly to economic drivers, prepayment events are rare, and shifting portfolio composition contaminates comparisons over time. We model prepayment on French residential securitized mortgages observed quarterly from August 2020 through February 2022 and benchmark a single-hidden-layer neural network against logistic regression, deeper network architectures, Random Forest, and XGBoost. Machine learning models consistently outperform the logistic baseline. The shallow network improves discriminative ability by roughly 20%, and XGBoost attains a mean out-of-sample AUC of 0.82, which evidences the interaction and threshold effects that traditional models like logistic regression cannot represent. SHAP and LIME attributions agree on the driver hierarchy. Under near-zero rates, refinancing incentives lose much of their predictive power, and prepayment is shaped by a broader set of conditions, real estate dynamics above all, followed by local unemployment and finally also by the financial incentive to refinance.