We compare of portfolios whose investment decisions are driven by neural networks (NN). We start by showing that often the portfolios evaluations yield optimistic results. With the aim of having more reliable measures of portfolio performance, we setup a methodology based on cross-validation that involves performance measurement across different holdout periods and varying portfolio compositions.

We compare a variety of NN strategies with classical extensions of the mean–variance model and the 1/N strategy. We prove that NN-based strategies, if set up correctly, systematically outperform the 1/N benchmark. Setting up the NN strategies requires advanced techniques such as weight-shrinkage, weight-regularization and the subtle tuning of regularization and model architectures.

Reference to the paper: Maggiolo, M. and Szehr, O. "Overfitting in portfolio optimization." Journal of Risk Model Validation, 17.3, 1-33, 2023.