Abstract
We investigate a machine learning approach to option Greeks approximationbased on Gaussian process (GP) surrogates. The method takes in noisily observedoption prices, fits a nonparametric input-output map and then analyticallydifferentiates the latter to obtain the various price sensitivities. Ourmotivation is to compute Greeks in cases where direct computation is expensive,such as in local volatility models, or can only ever be done approximately. Weprovide a detailed analysis of numerous aspects of GP surrogates, includingchoice of kernel family, simulation design, choice of trend function and impactof noise. We further discuss the application to Delta hedging, including a new Lemmathat relates quality of the Delta approximation to discrete-time hedging loss.Results are illustrated with two extensive case studies that considerestimation of Delta, Theta and Gamma and benchmark approximation quality anduncertainty quantification using a variety of statistical metrics. Among ourkey take-aways are the recommendation to use Matern kernels, the benefit ofincluding virtual training points to capture boundary conditions, and thesignificant loss of fidelity when training on stock-path-based datasets.