Initial Margin Simulation with Deep Learning

Abstract

Regulatory initial margin (IM) for non-cleared OTC derivatives is currently being implemented in the financial industry per BCBS-IOSCO requirements. To incorporate IM into counterparty credit risk (CCR) measurements and xVA calculations (especially MVA), it is necessary to simulate future IM requirements inside a CCR/xVA system. However, this is an extremely challenging task because the two readily available approaches – brute force simulation and regression-based approximation – are either prohibitively expensive in terms of computational cost, or extremely difficult for large and diverse portfolios due to the high dimensionality of the problem. In this paper, a practical deep learning approach to IM simulation is proposed, with a proof-of-concept implementation and test results demonstrating fast and accurate portfolio-level simulation of scenario-dependent IM through multiple time steps. Model training is shown to converge quickly, and model performance is robust under practical conditions. This approach separates offline training from online simulation, so that it can be implemented in production without significant system overhaul. Conceptually, since training data are generated by a deterministic function (in this case the sensitivity-based SIMM model), data noise is not a concern and overfitting can be avoided, assuming portfolio turnovers occur gradually during periods of time much longer than model training cycles. Model output can also be explained or validated by the underlying data-generating function for transparency. Other potential applications of this deep learning approach are also discussed, including collateral optimization.