Deep Learning for Disentangling Liquidity-Constrained and Strategic Default

Abstract

We disentangle liquidity-constrained default and the incentives for strategic default using Deep Neural Network (DNN) methodology on a proprietary Trepp data set of commercial mortgages. Our results are robust (insensitive) to severe Financial Crisis (2008) and plausible economic catastrophe ensuing from COVID-19 pandemic (2020-2021). We demonstrate an identification strategy to retrieve the motive of default from observationally equivalent delinquency classes by bivariate analysis of default rate on Net operating income (NOI) and Loan-to-Value (LTV). NOI, appraisal reduction amount, prepayment penalty clause, balloon payment amongst others co-determine the delinquency class in highly nonlinear ways compared to more statistically significant variables such as LTV. Prediction accuracy for defaulted loans is higher when DNN is compared with other models, by increasing flexibility and relaxing the specification structure. These findings have significant implications for investors, rating agencies and policymakers.