Neural Nonlinear Shrinkage of Covariance Matrices for Minimum Variance Portfolio Optimization

TL;DR

This work addresses covariance estimation for the global minimum variance portfolio by introducing a neural nonlinear eigenvalue shrinkage approach built on a Ledoit–Wolf decomposition. It directly learns a nonlinear mapping $\boldsymbol{\eta} = F_{\theta}(\boldsymbol{\lambda}, c)$ to tune the inverse covariance via $\hat{\mathbf{C}}_{\rm NN}^{-1} = \sum_i \eta_i {\bf u}_i{\bf u}_i^{\top}$, with a Transformer architecture conditioned on the sample-to-dimension ratio $c = N/n$. The network is trained end-to-end with portfolio risk as the objective, using out-of-sample returns to approximate risk. Empirical results on 50 S&P500 stocks show consistent out-of-sample risk reductions over benchmarks like SCM, LW, Chen, and DWE, across regimes where $n \le N$ and $n > N$, highlighting the value of combining structural statistical estimators with data-driven learning for scalable, risk-focused portfolio optimization.

Abstract

This paper introduces a neural network-based nonlinear shrinkage estimator of covariance matrices for the purpose of minimum variance portfolio optimization. It is a hybrid approach that integrates statistical estimation with machine learning. Starting from the Ledoit-Wolf (LW) shrinkage estimator, we decompose the LW covariance matrix into its eigenvalues and eigenvectors, and apply a lightweight transformer-based neural network to learn a nonlinear eigenvalue shrinkage function. Trained with portfolio risk as the loss function, the resulting precision matrix (the inverse covariance matrix) estimator directly targets portfolio risk minimization. By conditioning on the sample-to-dimension ratio, the approach remains scalable across different sample sizes and asset universes. Empirical results on stock daily returns from Standard & Poor's 500 Index (S&P500) demonstrate that the proposed method consistently achieves lower out-of-sample realized risk than benchmark approaches. This highlights the promise of integrating structural statistical models with data-driven learning.

Figures (4)

• Figure 1: Overall workflow of portfolio optimization with proposed neural network-based nonlinear precision matrix estimator. The "LW" represent Ledoit-Wolf covariance matrix estimation. The "ED" denotes eigenvalue decomposition.
• Figure 2: Lightweight transformer-based network architecture.
• Figure 3: Realized portfolio risks achieved out-of-sample over 400 days of S&P500 real market data by a GMVP implemented using different methods.
• Figure 4: Annualized rolling-window standard deviations of the most recent 40 out-of-sample log returns for the GMVP based on $\hat{\bf C}_{\rm NN}^{-1}$ and $\hat{\bf C}_{\rm LW}$ (N=50, n=100).