Master’s Research Paper Presentation • Scientific Computation • A Neural Network Approach to Optimal ARVA Decumulation

Please note: This master’s research paper presentation will take place in DC 2310.

Qinglan Cao, Master’s candidate
David R. Cheriton School of Computer Science

Supervisors: Professors Yuying Li, Peter Forsyth

We introduce a neural network (NN) based framework to optimize the Annually Recalculated Virtual Annuity (ARVA) spending rule based Defined Contribution (DC) pension decumulation strategy. We approximate the optimal allocation in stocks and bonds by a global-in-time NN, which is a function of time and current wealth. Our objective function balances maximizing total expected withdrawal and minimizing the downside withdrawal variability.

Experiments are conducted to compare our data-driven approach with controls obtained by solving (numerically) a Hamilton-Jacobi-Bellman (HJB) Partial Integro Differential Equation (PIDE) approach. The HJB-PIDE method provably converges to the optimal solution. Comparison to constant allocation strategies highlights the flexibility and robustness of our dynamic strategy for its derisking behaviour, which holds for both Monte Carlo simulated data using the market parametric model and bootstrapped data resampled from the historical path. The performance is validated on single historical paths, demonstrating the capability of our NN model in handling evolving market conditions.