Software Issues and Performance of a Parallel Model for Stock Option Pricing

The finance industry is beginning to adopt parallel computing for numerical computation, and will soon be in a position to use parallel supercomputers. This paper examines software issues and performance of a stock option pricing model running on the Connection Machine-2 and DECmpp-12000. Pricing models incorporating stochastic volatility with American call (early exercise) are computationally intensive and require substantial communication. Three parallel versions of a stock option pricing model were developed which varied in data distribution, load balancing, and communication. The performance of this set of increasingly refined models ranged over no improvement, 10 times, and 100 times faster than a sequential model. A straightforward approach to this problem involves use of two-dimensional dynamic arrays. When asymmetric arrays are mapped on the DECmpp-12000, distribution of data to physical processors is inefficient and performance suffers. The regular communication patterns in the model can also be expressed in one dimensional arrays, improving data distribution. Performance of this version is similar on both parallel machines. Combining one dimensional parallel and sequential arrays achieves efficient data distribution, reduces interprocessor communication, and further improves performance (100 times faster than a sequential workstation model). The performance improvements possible on parallel supercomputers presents new opportunities for pricing entire portfolios, performing large scale model and market comparisons and using optimization techniques to improve model price estimates