Robustness of neural-network portfolio optimizers under varying cross-sectional dimensions

Investigate whether neural network–based portfolio optimization methods maintain stable and robust performance as the number of assets varies, by rigorously evaluating their stability under changes in cross-sectional dimension and determining the conditions under which performance remains robust when scaling from small or static universes to large universes.

Background

The paper surveys recent neural-network approaches to portfolio optimization and notes that many studies operate on low-dimensional or static asset universes and rely on heterogeneous benchmarks. Because these works seldom examine the sensitivity of performance to the number of assets, the stability of such methods when the cross-sectional dimension changes has not been thoroughly studied.

The authors emphasize that this gap leaves uncertainty about whether previously proposed neural-network portfolio optimizers retain their performance when applied to larger universes. Their own architecture is designed to be dimension-agnostic and is tested by training on a few hundred equities and applying to one thousand, but the general question for other neural methods remains unresolved.