Reinforcement learning for Multiple Goals in Goals-Based Wealth Management

We formulate the goals-based wealth management (GBWM) problem as a Markov Decision Process (MDP) and assess deep reinforcement learning (RL) and dynamic programming (DP) algorithms for dynamic solutions. These algorithms seek to (i) optimally choose from multiple goals over time and (ii) optimally switch between several model investment portfolios over time. In RL, we implement a proximal policy optimization (PPO) algorithm with a customized environment in which the goal-taking choices and investment portfolio selections are implemented with a discrete two-dimensional action space using a deep neural network. We show this RL approach performs almost as well as exact DP solutions. By showing that RL works well for GBWM in a comparative setting with DP, we see that RL may be a suitable method in other GBWM models whose larger state spaces render DP methods computationally infeasible.