Applying Machine Learning for Troubleshooting Credit Exposure and xVA Profiles

Abstract

We explore the possibility of applying machine learning techniques to characterizing and troubleshooting problem in credit exposure profiles used in counterparty credit risk management and XVA. Machine learning models are trained to identify characteristics from the daily credit exposure profiles generated through Monte Carlo simulation and identify anomalies due to either counterparty behavioral changes or errors in inputs or calculation. In the paper, we have trained both K-means clustering and Convolutional Neural Network to a training data set containing more than 400,000 credit exposure profiles. The performance of models is satisfactory and we believe this is a promising area for further exploration. Identifying anomalies provides credit analysts early insight into emerging credit risk and errors in the data lineage.