Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces an end-to-end neural shrinkage method for estimating positive definite covariance matrices from pairwise correlation data in small-cap-inclusive equity portfolios. By employing a rotation-invariant neural covariance estimator that processes signed spectra and utilizes a bidirectional gated recurrent unit, the model effectively addresses the challenges posed by indefinite correlation matrices. The approach significantly reduces annualized five-day volatility by approximately 20% and enhances the Sharpe ratio by around 40% compared to traditional covariance estimators, demonstrating robust performance across various risk metrics and execution frictions.
A neural estimator not only ensures positive definiteness of covariance matrices but also achieves a 40% increase in Sharpe ratio for small-cap equity portfolios.
Small-cap-inclusive equity universes contain recently listed and intermittently traded securities, so enforcing a common look-back discards a substantial fraction of the available information. Pairwise-complete estimation preserves the longest overlap for each asset pair, but the resulting correlation matrix can be indefinite because its entries are computed on different samples. This prevents direct use in Markowitz optimization and falls outside the assumptions of standard random-matrix shrinkage. We adapt a rotation-invariant neural covariance estimator to this setting. The model computes mask-aware marginal moments and a pairwise correlation matrix proxy, processes its signed spectrum, and uses a bidirectional gated recurrent unit conditioned on factor-aligned effective sample lengths derived from the overlap matrix and eigenvector loadings. It maps all eigenvalues, including negative ones, to a positive inverse spectrum. The reconstructed covariance is positive definite and is trained end-to-end to minimize five-day realized global-minimum-variance risk. We evaluate 26 expanding-window models from 2000 to 2025 on up to 1,500 U.S. equities in a closing-auction simulator with point-in-time selection, commissions, financing, corporate actions, and market impact. Across the 26-year out-of-sample period, the neural estimator reduces annualized five-day volatility by approximately 20\% and increases the Sharpe ratio by approximately 40\% relative to the next-best covariance estimator. These improvements are consistent across realized risk, risk-adjusted performance, and drawdown control, remain after the modeled execution frictions, and are supported by a 99.9\% Model Confidence Set that retains only the neural estimator.