An approximate dynamic programming approach to risk sensitive control of execution costs

Abstract

We study the problem of optimal execution within a dynamic programming framework. Given an exponential objective function, system variables which are normally distributed, and linear market dynamics, we derive a closed form solution for optimal trading trajectories. We show that a trader lacking private information has trajectories which are static in nature, whilst a trader with private information requires real time observations to execute optimally. We further show that Bellman's equations become increasingly complex to solve if either the market dynamics are nonlinear, or if additional constraints are added to the problem. As such, we propose an approximate dynamic program using linear programming which achieves near-optimality. The algorithm approximates the exponential objective function within a class of linear architectures, and takes advantage of a probabilistic constraint sampling scheme in order to terminate. The performance of the algorithm relies on the quality of the approximation, and as such we propose a set of heuristics for its efficient implementation.