As Valuation Adjustments (XVAs) directly affect trade prices and capital charges, it is of central importance for banks to actively manage XVA through hedging. This requires computing XVA sensitivities to a large number of risk factors, typically hundreds to thousands. As XVA calculations involve a compute-intensive Monte-Carlo simulation, the hundreds to thousands of revaluations required for calculating sensitivities with the traditional bump-and-revalue approach (finite difference) is not practical. However, by applying Adjoint Algorithmic Differentiation (AAD), all sensitivities can be quickly and accurately computed.
This paper introduces the use of sensitivities for managing XVAs, details the complexities involved, and introduces AAD as a solution. It gives specific techniques to manage memory resources and improve performance for AAD on practical examples.