| Summary: | Computing the price and risk of financial derivatives is a necessary activity for many financial market participants and is often undertaken by large and costly computing farms. This thesis seeks to explore the use of parallel computing, with particular focus on graphics processing units (GPUs), to improve the speed per cost ratio of such computation. This thesis addresses three distinct layers of high performance parallel financial derivatives computation: the first layer is related to the formulation of parallel algorithms that are generally used in the context of derivatives. The second layer is related to the optimum computation of pricing models, which consist of a series of computational steps or algorithms, where such pricing models are used to calculate the price and risk of individual derivatives. The third and final layer is related to deploying several pricing models within large scale infrastructures with particular focus on optimal scheduling approaches. Several contributions are made within this thesis: (i) with regard to the formulation of parallel algorithms, we introduce novel approaches for evaluating the normal cumulative distribution function (CDF), calculating option implied volatility, calibrating SABR (stochastic-αβρ) volatility models and generating CDF lookup tables. (ii) With regard to pricing models, we explore the computation of two dominant fixed income pricing models, namely non-callable bullet options and callable bond options. (iii) With regard to the computation of many such pricing models within large scale infrastructures, we devise and verify novel scheduling approaches that are able to optimally allocate tasks between a heterogeneous mix of CPU and GPU processors.
|
|---|
