Reinforcement Learning for constrained Portfolio Allocation

Details

Supervisors

Ars, Pierre

Faculty

Faculté des sciences

Degree label

Master [120] en sciences actuarielles, à finalité spécialisée

Abstract

While machine learning has been employed in actuarial science for a long time, not all machine learning techniques have been explored by the field, including reinforcement learning. In contrast, the financial sector has successfully integrated reinforcement learning for tasks such as stock trading and portfolio management. This thesis focuses on constrained portfolio allocation driven by insurer liabilities. The aim is to improve decision-making and risk management by leveraging reinforcement learning capabilities to handle various market environments and risk aversions. The thesis details reinforcement learning algorithms and their underlying concepts. It then applies these methods to unconstrained and contrained portfolio optimization and proposes a reinforcement learning-based methodology for managing insurance products with target returns. This approach is tested in scenarios of interest rate increases to evaluate reinforcement learning's adaptive capabilities. This master thesis combines reinforcement learning with actuarial science and proposes new ways of exploring actuarial problems, ultimately contributing to the optimization and efficiency of insurance portfolio allocations.