Multi Scenario Financial Planning via Deep Reinforcement Learning AI

Abstract

Financial planning via deep reinforcement learning holds much promise. One implementation, AIPlanner, delivered near optimal financial results, but had a major shortcoming. It required training a separate neural network model for each financial scenario. This paper describes extending AIPlanner so that a small family of trained neural network models are capable of rapidly producing financial plans for a wide variety of financial scenarios. Additionally AIPlanner is extended to produce results over the lifecycle, both pre and post retirement, and for couples, as well as individuals. A reasonably realistic income tax model is incorporated. And finally, a more realistic stock model is used. Over the lifecycle, compared to the best discovered alternative strategy, reinforcement learning was found to effectively deliver 14% more retirement consumption.

Keywords: financial planning, asset allocation, consumption, reinforcement learning, deep learning