Optimal Asset Allocation with Quadratic Variation as a Risk Measure: A Neural Network Approach

Speaker: Chang Chen 
Affiliation: University of Queensland

Abstract

We determine the optimal decumulation strategy for defined contribution (DC) pension plan, with an Annually Recalculated Virtual Annuity (ARVA) spending rule. Our objective is to maximize expected total withdrawals and minimize quadratic variation, providing investors with better control over risk throughout the investment horizon. We propose a data-driven Neural Network approach to determine the optimal allocation, incorporating realistic constraints such as no leverage and discrete rebalancing, and verify its effectiveness using the PDE method. Compared to a constant weight strategy with the same expected withdrawals, the optimal strategy significantly reduces quadratic variation for most of the time, resulting in a more stable investment path. This conclusion holds under both a parametric market model based on historical data and a bootstrapped market simulation.