We present a generic solver for dynamical portfolio allocation problem when the market exhibits return predictability and price impact as well as partial observability. We assume that the prices modeling can be encoded into a linear state-space and show how the problem then falls into the LQG framework. We derive the optimal control policy and introduce tools to analyze it that preserve the intelligibility of the solution. Furthermore, we link theoretical assumptions for existence and uniqueness of the optimal controller to a dynamical non-arbitrage criterion. Finally, we illustrate our method on a synthetic portfolio allocation problem and provide intuition about the behavior of the controlled system.
↧