European derivatives pricing via constrained Gaussian process regressions

Arnaud Servais, Donatien Hainaut

Abstract

This article develops a Gaussian process method for pricing European derivatives when the underlyings, and the risk factors driving them, follow a multivariate jump-diffusion. A Gaussian process prior on the pricing function is conditioned on the Feynman-Kac partial integro-differential equation at a set of collocation points. The price is returned as a closed-form and differentiable function, allowing the Greeks to be obtained without simulation or finite differences. The jump terms are handled iteratively: the price is computed as the limit of a sequence of boundary value problems. Our main theoretical contribution is a convergence result for this iteration: using Weissinger's fixed-point theorem in the supremum norm, we show that it converges from any starting point, and provide a worst-case number of iterations for a given precision. The method is illustrated with derivatives of two underlyings under the Heston model with Kou double-exponential jumps, a five-dimensional problem, and benchmarked against Monte Carlo simulation.

Source: arxiv · PDF

Read the AI summary, key takeaways and discussion on WOBR Quant Research.


Open in the WOBR AI app → · WOBR.AI home