Neural Networks for Credit Risk and xVA in a Front Office Pricing Environment

Abstract: We present a data-driven proof of concept model capable of reproducing expected counterparty credit exposures from market and trade data. The model has its greatest advantages in quick single-contract exposure evaluations that could be used in front office xVA solutions. The data was generated using short rates from the Hull-White One-Factor model. The best performance was obtained from a GRU neural network with two recurrent layers, which with adequate accuracy could reproduce the exposure profile for an interest rate swap contract. Errors were comparable to those expected from a Monte Carlo simulation with 5K paths. With regards to computational efficiency, the proposed model showed great potential in outperforming traditional numerical methods. Further development and calibration to actual market data is required for the model to be applicable in the industry. The proposed architectures may then prove useful, especially for contracts with high-rated counterparties, traded in a normal and liquid market.