End-to-End Neural Covariance Estimation for Portfolios with Illiquid Small-Cap Stocks
ORIGINAL / End-to-End Neural Shrinkage of Indefinite Pairwise Correlation Matrices for Small-Cap-Inclusive Portfolios
This paper introduces a neural covariance estimator that handles indefinite pairwise correlation matrices arising from incomplete data, improving portfolio optimization for small-cap-inclusive universes, with out-of-sample volatility reduction and Sharpe ratio enhancement.
01 ABSTRACT
The authors propose a novel neural covariance estimation method to address pairwise incomplete data, demonstrating superior out-of-sample performance in a US equity backtest, but conclusions rely on simulated trading and specific historical conditions.
02 KEY FINDINGS
- Pairwise-complete estimation preserves data but yields indefinite correlation matrices.
- A rotation-invariant neural estimator maps indefinite spectra to positive definite, end-to-end trained.
- Model leverages mask-aware moments, signed spectrum, and BiGRU conditioned on effective sample lengths.
- Out-of-sample tests on up to 1,500 US stocks show ~20% lower annualized 5-day volatility and ~40% higher Sharpe ratio.
- Results remain after execution frictions and pass model confidence set at 99.9%.
Return to the primary source
Read original ↗AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG · ARXIV Q-FIN