A functional approach to accelerating Monte Carlo based american option pricing

We study the feasibility and performance efficiency of expressing a complex financial numerical algorithm with high-level functional parallel constructs. The algorithm we investigate is a least-square regression-based Monte-Carlo simulation for pricing American options. We propose an accelerated parallel implementation in Futhark, a high-level functional data-parallel language. The Futhark language targets GPUs as the compute platform and we achieve a performance comparable to, and in particular cases up to 2.5X better than, an implementation optimised by NVIDIA CUDA engineers. In absolute terms, we can price a put option with 1 million simulation paths and 100 time steps in 17 ms on a NVIDIA Tesla V100 GPU. Furthermore, the high-level functional specification is much more accessible to the financial-domain experts than the low-level CUDA code, thus promoting code maintainability and facilitating algorithmic changes.