We first provide a mini-tutorial on Adjoint Algorithmic Differentiation (AAD) (also known as back-propagation in machine learning). We then illustrate how neural networks may be used to compute dynamic values and risks of trading books with applications to risk management of derivatives, valuation adjustments (XVA), counterpart credit risk, FRTB and SIMM margin valuation adjustments (MVA). We also describe new techniques to substantially improve deep learning on simulated data, and discuss how this is analogous to deriving approximate analytics in real time.
All the slides are available here. This material was also presented as a seminar for Oxford University, see here.