Maxence Prémont

Outputs

Reinforcement Learning Algorithms for a Dynamic Goal-Based Wealth Management Problem

This M.Sc. thesis explores the use of reinforcement learning (RL) algorithms to solve a dynamic goal based wealth management (GBWM) problem, where the objective is to maximize an investor’s probability of surpassing her terminal wealth goal. This asset allocation problem is usually solved with a dynamic programming algorithm. However, this thesis focuses on the implementation of a specific type of RL algorithm (Q-learning) specifically tailored for the GBWM problem examined here. Using RL, even if it can only approximate the solution, could be advantageous since, unlike dynamic programming, it can avoid the curse of dimensionality and easily integrate path-dependant constraints like transaction costs. Two contributions, which improve the algorithm’s performance, are then proposed. Unlike the previously presented algorithm, the improved version is able to correctly approximate the benchmark solution obtained with dynamic programming. Finally, another type of RL algorithm, Policy Gradient, is explored. While it successfully solved simplified problems, it was unfortunately not able to accurately approximate the benchmark solution for the full version of the GBWM problem