Investigation To A Neural Network Approach To Optimal Dynamic Allocation Problem In Defined Contribution Pension Plans

Abstract

In this thesis, we propose a data-driven neural network (NN) optimization framework for solving a dynamic stochastic control problem under stochastic constraints. The objective function of the optimal control problem is based on expected wealth withdrawn (EW) and expected shortfall (ES) that directly targets left-tail risk. The optimal solution obtained from NN framework achieves high computational accuracy comparable to the Hamilton-Jacobi-Bellman (HJB) Partial Differential Equation (PDE) method. Additionally, the NN framework exhibits strong computational robustness, maintaining stable performance across different data distributions. Unlike traditional HJB PDE approaches, the NN framework can be extendable to high-dimensional multi-asset problems, overcoming the curse of dimensionality. To further enhance data diversity and improve generalization, we introduce TimeGAN and incorporate TimeGAN-generated data to generate historical financial time-series data, ensuring the robustness of model training.