To prevent a similar crisis to 2007/2008, banks are regularly tested for their ability to respond to stress periods in the market. This stress-testing is performed at a regulatory level (e.g., CCAR / DFAST in the US, European-wide stress test exercises), and at bank-internal level. These include computationally expensive calculations such as XVA, which are tested under a large set of stress scenarios. As no algorithmic optimisations (e.g. sensitivity-based approach) are permitted, every scenario requires a full re-valuation. To keep the stress test execution time within reasonable bounds, it is therefore essential that the tested calculations run as fast as possible.
This paper gives an overview of financial stress-testing on different levels and highlights the complexities involved. It then details a range of techniques that can be used to optimise the performance of the calculations, modernising the code to run fast on recent multi-core CPUs. It further explores the application of accelerator processors to further optimise performance (GPU, many-core).