Optimal Posting of Collateral with Recurrent Neural Networks

Abstract

In this paper, we consider the problem of optimal posting of collateral in two assets. The associated stochastic control problem is then solved by parameterizing the optimal control with a (recurrent) neural network. Then, the training of the neural network is achieved with a back-propagation algorithm, possibly complemented with a particle swarm optimization (or simulated annealing). Our algorithm is then illustrated with numerical examples.