CVA Hedging with Reinforcement Learning

This work considers the problem of a trader who must manage the Credit Valuation Adjustment (CVA) of a derivative, defined as the risk-neutral expectation of losses incurred if the counterparty of the derivative defaults. CVA can be regarded as a hybrid product, one of the most complex actively managed by a trading desk. Standard delta hedging based on sensitivities is not completely satisfactory for this product, because it ignores trading costs and jump-to-default risk while introducing unavoidable simplifications in the pricing model. In this paper we use risk-averse Reinforcement Learning to learn a superior hedging strategy compared to the standard delta hedging approach. Specifically, we generalize risk-averse Reinforcement Learning to stochastic horizons, to be compatible with counterparty defaults, and we introduce a realistic framework for the mechanics of the hedger's portfolio in which the data generating process of the underlying risk drivers can be inconsistent with the risk-neutral laws used to price the CVA and the hedging instruments. The potential of the proposed approach is investigated empirically by numerical examples on hedging the CVA of a forex forward.