PT - JOURNAL ARTICLE
AU - Irlam, Gordon
TI - Machine Learning for Retirement Planning
AID - 10.3905/jor.2020.1.067
DP - 2020 Jul 31
TA - The Journal of Retirement
PG - 32--39
VI - 8
IP - 1
4099 - http://jor.pm-research.com/content/8/1/32.short
4100 - http://jor.pm-research.com/content/8/1/32.full
AB - Machine learning provides a new approach to solving problems in many fields. This article explores the use of machine learning to solve the retirement portfolio problem: deciding how much wealth to consume and how to allocate the remainder. After first reviewing existing approaches to the portfolio problem, this article looks in detail at the use of reinforcement learning. For simple financial scenarios where the optimal solution is known, reinforcement learning is found to deliver to within a few percent of the optimal solution. For more complicated financial scenarios, with no known optimal solution, reinforcement learning outperforms other common approaches. Reinforcement learning proves capable of optimizing highly complex financial models, including the effects of income taxes, mean-reverting asset classes, and time-varying bond yield curves, all of which other approaches cannot handle. Reinforcement learning appears to be the first fundamentally new approach to the portfolio problem in over 50 years.TOPICS: Retirement, big data/machine learningKey Findings• The field of machine learning is evolving rapidly. Within machine learning, the subfield of reinforcement learning appears to have the most relevance to retirement planning.• Reinforcement learning is capable of handling the real-world complexity of financial planning, including the effects of income taxes, mean-reverting asset classes, time-varying bond yield curves, and uncertain life expectancies.• For simple scenarios, machine learning delivers to within a few percent of the optimal solution. For more complex scenarios, whose optimal solution is unknown, machine learning is found to outperform other common approaches.