Credit Exposure Modelling Using Differential Machine Learning

Abstract

Exposure modelling is a critical aspect of managing counterparty credit risk, and banks worldwide invest significant time and computational resources in this task. One approach to modelling exposure involves pricing trades with a counterparty in numerous potential future market scenarios. Suitable for this type of pricing is a framework presented in 2020 by Huge and Savine, which they call differential machine learning. It approximates the pricing function with a neural network that trains on Monte Carlo paths and the gradients along these paths. This thesis aims to demonstrate the application of differential machine learning in the context of exposure modelling. To better comply with this context, training is done on market variables, rather than some hidden model state. Simulated data is used from Heston type models to estimate the future exposure distribution of a portfolio consisting of European options. The conducted experiments reveal that training the machine learning model on market observables yields similar results to those obtained when training on hidden model states. Furthermore, the exposure modelling approach is subject to stress testing by evaluation of its performance under different levels of compatibility between the pricing model and future market scenarios in which the portfolio is priced. Results show that low compatibility leads to decreased accuracy of the predicted exposure distributions.