GPU-Accelerated American Option Pricing: The Case of the Longstaff-Schwartz Monte Carlo Model

In this article, the authors explore the implementation of graphics processing units (GPUs) for pricing American options using the Longstaff-Schwartz least squares approach. Unlike the sequential computation performed by traditional central processing units (CPUs), GPUs provide for parallel computing, which is better suited for financial modeling of computationally intensive tasks such as valuing American options. Their objective is to assess the potential of GPU computing, particularly its capacity for high-performance single instruction multiple data (SIMD) processing. While leveraging the performance improvements achieved through GPU-accelerated American option pricing with the Longstaff-Schwartz Monte Carlo (LSMC) approach, the authors also investigate the numerical instability in the LSMC results caused by the original choice of basis functions, which may produce nearly singular matrices. If not properly addressed, this instability can result in inaccurate results, negating the benefits of parallel computing efficiency. To mitigate this issue, the authors propose a suitability enhancement approach to the basis functions, which the authors find effectively stabilizes the results without compromising computational efficiency.