high performance computing on graphics processing units: hgpu.org

We describe a high performance parallel implementation of a derivative pricing model, within which we introduce a new parallel method for the calibration of the industry standard SABR (stochastic-alpha beta rho) stochastic volatility model using three strike inputs. SABR calibration involves a non-linear three dimensional minimisation and parallelisation is achieved by incorporating several assumptions unique to the SABR class of models. Our calibration method is based on principles of surface intersection, guarantees convergence to a unique solution and operates by iteratively refining a two dimensional grid with local mesh refinement. As part of our pricing model we additionally present a fast parallel iterative algorithm for the creation of dynamically sized cumulative probability lookup tables that are able to cap maximum estimated linear interpolation error. We optimise performance for probability distributions that exhibit clustering of linear interpolation error. We also make an empirical assessment of error propagation through our pricing model as a result of changes in accuracy parameters within the pricing model’s multiple algorithmic steps. Algorithms are implemented on a GPU (graphics processing unit) using Nvidia’s Fermi architecture. The pricing model targets the evaluation of spread options using copula methods, however the presented algorithms can be applied to a wider class of financial instruments.