Option pricing, or the valuation of an option, is constantly performed on a financial derivative market to determine the premium that the buyer pays to the seller of the option in a transaction. One of the common methods for option pricing is Monte-Carlo simulation, which generates a large number of random numbers to simulate the price paths over time steps, and computes the average and time-discounted price for the transaction day. This process is computation-intensive, and fast valuation speed as well as high numerical accuracy is crucial to the financial firms. Therefore, we explore to use Graphics Processing Unit (GPU), or graphics processors, to accelerate Monte-Carlo simulation in valuation of a large number of options.

There have been existing GPU-based Monte-Carlo softwa...[ Read more ]

Option pricing, or the valuation of an option, is constantly performed on a financial derivative market to determine the premium that the buyer pays to the seller of the option in a transaction. One of the common methods for option pricing is Monte-Carlo simulation, which generates a large number of random numbers to simulate the price paths over time steps, and computes the average and time-discounted price for the transaction day. This process is computation-intensive, and fast valuation speed as well as high numerical accuracy is crucial to the financial firms. Therefore, we explore to use Graphics Processing Unit (GPU), or graphics processors, to accelerate Monte-Carlo simulation in valuation of a large number of options.

There have been existing GPU-based Monte-Carlo software packages for option pricing, for example, one in NVIDIA’s CUDA SDK. However, we found that the repeated generation of a large number of random numbers on the GPU was the performance bottleneck. Furthermore, the numerical accuracy was problematic due to the differences in the random number generators on the CPU and the GPU. To address this problem, we proposed to generate the random numbers on the CPU once and transfer all of them to the GPU for subsequent computation. Additionally, we parallelized the computation on the GPU efficiently with data-parallel primitives. As a result, our GPU-based Monte-Carlo implementation outperformed the CUDA library by two orders of magnitude, and the result accuracy was improved by three times.