Fast American basket option pricing on a multi-GPU cluster

This article presents a multi GPU adaptation of a specific Monte Carlo and classification based method for pricing American basket options, due to Picazo [1]. The first part relates how to combine fine and coarse grained parallelization to price American basket options. In order to benefit from different GPU devices, a dynamic strategy of kernel calibration is proposed, and contributes to the dynamic split of GPU calculus. Our implementation achieves a realistic size option pricing in less than one hour against more than 7 for a multi CPU cluster-based solution. After an analysis of possible bottleneck effects, we distribute the sequential bottleneck due to the training phase. For this we rely upon Random Forests classification method which is suited to parallelization. We show through tests that the obtained parallel pricing algorithm is scalable.