Neural Network Model Achieves Global Minimum-Variance Portfolio Optimization Through Covariance Cleaning

A new neural network that is invariant to rotation has been created to identify the global minimum-variance portfolio. It learns lag transformations of past returns and marginal volatilities while also regularizing the eigenvalues of extensive equity covariance matrices. This clear mathematical framework enhances the interpretability of each module's role, avoiding the perception of the model as a mere black box. The design mirrors the analytical structure of the global minimum-variance solution and is dimension-agnostic, enabling a single model trained on several hundred stocks to be effectively used on one thousand US equities without the need for retraining. Out-of-sample evaluations from January 2000 to December 2024 indicate that the estimator reliably produces lower variance.