Deep neural networks for probability of default modelling

In this paper we develop Deep Neural Networks for the approximation of the solution to Partial Integro-Differential Equations (PIDE) that arise in the calculation of Probability of Default functions. We consider a modelling framework in compliance with the spirit and regulations of the International Financial Reporting Standard 9 and use the resulting Deep Learning models to estimate default probabilities that can be used to solve credit risk problems. Detailed comparisons with standard numerical analysis schemes for the solutions to these PIDEs are also reported, enhancing the understanding and adding to the discussion regarding the applicability of the related Machine Learning methodologies.