A Deep Reinforcement Learning Approach to Asset-Liability Management

Asset-Liability Management (ALM) is a technique used to optimize investment portfolios, onsidering a future flow of liabilities. Its stochastic nature and multi-period decision structure favors its modeling as a Markov Decision Process (MDP). Reinforcement Learning is a state-of-the-art group of algorithms for MDP solving, and with its recent performance boost provided by deep neural networks, problems with long time horizons can be handled in just a few hours. In this paper, we address an ALM problem with a variation of Deep Deterministic Policy Gradient algorithm. Opposed to all other approaches in the literature, our model does not use scenario discretization, which is a major contribution to ALM study. Our experimental results show that the Reinforcement Learning framework is well fitted to solve this kind of problem, and has the additional benefit of using continuous state spaces.