Simulation-Based Neural Policies for Portfolio Choice: Architecture, Training, and Interpretability

Abstract

Many economic decision problems, lifecycle consumption-saving and dynamic portfolio choice, are finite-horizon stochastic control problems with continuous states and actions. When the state is low-dimensional these problems are solved by dynamic programming on a grid. The grid cost grows exponentially in the state dimension, known as the curse of dimensionality, which motivates replacing the value-function grid with a neural policy optimized directly through simulation. Such policies are usually studied in the high-dimensional settings that motivate them, precisely where no reference solution exists. So the contribution of any single architectural or training choice cannot be isolated and diagnosed. We therefore take a step back and treat both the architecture and the solution method as the objects of study. To this end, we consider a lifecycle problem with a sufficiently low-dimensional normalized state space to admit an accurate dynamic programming solution, which is used for evaluation. We compare four architectures. The simplest consists of a single time-conditioned network. We then consider two networks concatenated across the regime switch, followed by one network per date trained backward against frozen downstream policies. Finally, we evaluate a constrained variant of the per-date architecture. Decoupling the policy across time gives each date a short, well-posed objective, which we pair with direction-dominant optimization that normalizes away gradient magnitude. Architectures that lead to similar realized utility objective can nevertheless differ in whether they respect the underlying problem's economics. We therefore evaluate each design jointly based on welfare, a solution-free Bellman residual, shape restrictions, and the resulting policy functions.