Daily model calibration with AAD

How adjoint gradients with XAD took a tier-one bank's morning stochastic local volatility calibration across about 300 underlyings from three hours to 25 minutes, with risk through the calibration.

An equity derivatives desk at a tier-one bank recalibrates its stochastic local volatility models every morning, across about 300 underlyings. The optimiser needed the gradient of the fit with respect to every model parameter at each step, and by bumping, one full repricing per parameter, the calibration took three hours. Risk to the quoted volatilities was worse: each bumped quote meant another calibration.

Xcelerit built AAD into the desk's calibration library with XAD. The gradients now come from one adjoint pass per step, validated against bumping, and the implicit function theorem carries sensitivities through the calibration, so vega to each quote needs no recalibration. The morning calibration takes 25 minutes instead of three hours, on the same servers. The desk runs the library under an XAD Commercial subscription.

Challenge

Morning SLV calibration across about 300 underlyings; three hours with bumped gradients

What we did

AAD enablement: XAD in the desk's calibration library, with risk through the calibration

How it was checked

Same calibration, servers and convergence tolerance, bumped against adjoint gradients

Client-reported outcome. Exact figures are in the case study PDF, on request.

All resources