Reproduce “End-to-End Large Portfolio Optimization for Variance Minimization with Neural Networks through Covariance Cleaning” on OpenResearch

About this paper

Christian Bongiorno, Efstratios Manolakis, Rosario Nunzio Mantegna

Researchers developed an end-to-end rotation-invariant neural network for Global Minimum-Variance portfolio optimization that jointly processes historical returns and regularizes large equity covariance matrices through learnable components. The model achieves lower realized volatility and higher Sharpe ratios compared to state-of-the-art benchmarks in realistic, friction-inclusive backtesting across a universe of 1,000 equities, while generalizing from smaller training asset sets.

Key insights

• The learned lag-transformation module weights more recent data and applies soft-clipping to outliers, transitioning the correlation metric behavior from Spearman-like for short lags to Phi-like for long lags.
• The eigenvalue cleaning module (BiLSTM) compresses bulk eigenvalues to a narrow range, making them effectively insensitive to input data variability, while dynamically adjusting extreme eigenvalues, akin to an adaptive Average Oracle.
• The modular design allows for interpretability, demonstrating how the NN processes temporal dependencies, regularizes risk modes, and scales individual asset volatilities, moving beyond a black-box approach for financial models.

Summary by alphaXiv.