Learning Personalized Retirement Policies from Observational Data using Offline Reinforcement Learning

• In Event: Advances in Computational Methods (Co-Sponsored by Mathematical Sociology Section)

Abstract

Determining an optimal retirement policy is a critical financial and personal decision, influenced by a multitude of factors including health, income, and demographic characteristics. This paper introduces offline reinforcement learning (RL) as a novel computational methodology for deriving personalized retirement policies from observational panel data, contributing to the growing toolkit of computational methods in sociology. Using the Panel Study of Income Dynamics (PSID), I model the retirement decision as a sequential decision-making problem and implement two state-of-the-art offline RL algorithms: Conservative Q-Learning (CQL) and Implicit Q-Learning (IQL). I experiment with three reward functions emphasizing different aspects of well-being: a balanced approach between income and health, income maximization, and health preservation. My findings demonstrate the potential of offline RL to derive data-driven, personalized retirement strategies and show that while average recommended ages move only modestly across reward designs, demographic gaps shift noticeably, which reveals how even small reward tweaks encode value choices that propagate into policy. This work highlights the importance of careful reward engineering and provides a methodological framework for in-silico policy experimentation in the social sciences.

Author

• Xinyue Wu, Washington State University