Paper: arXiv 2601.07687

Authors: Efstratios Manolakis, Christian Bongiorno, Rosario Nunzio Mantegna

Abstract

A new wave of work on covariance cleaning and nonlinear shrinkage has delivered asymptotically optimal analytical solutions for large covariance matrices. Building on this progress, these ideas have been generalized to empirical cross-covariance matrices, whose singular-value shrinkage characterizes comovements between one set of assets and another. Existing analytical cross-covariance cleaners are derived under strong stationarity and large-sample assumptions, and they typically rely on mesoscopic regularity conditions such as bounded spectra; macroscopic common modes (e.g., a global market factor) violate these conditions. When applied to real equity returns, where dependence structures drift over time and global modes are prominent, we find that these theoretically optimal formulas do not translate into robust out-of-sample performance. We address this gap by designing a random-matrix-inspired neural architecture that operates in the empirical singular-vector basis and learns a nonlinear mapping from empirical singular values to their corresponding cleaned values. By construction, the network can recover the analytical solution as a special case, yet it remains flexible enough to adapt to non-stationary dynamics and mode-driven distortions. Trained on a long history of equity returns, the proposed method achieves a more favorable bias-variance trade-off than purely analytical cleaners and delivers systematically lower out-of-sample cross-covariance prediction errors. Our results demonstrate that combining random-matrix theory with machine learning makes asymptotic theories practically effective in realistic time-varying markets.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper is mathematically dense, featuring advanced random matrix theory derivations, singular value decomposition, and the formulation of a physics-informed neural network with symmetry constraints. Empirically, it proposes a specific neural architecture trained on historical equity returns and evaluates out-of-sample prediction errors, indicating a data-heavy, implementation-focused approach.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Address drift & macro modes<br>in cross-covariance cleaning"] --> B["Methodology<br>Physics-Informed Singular-Value Learning<br>(Neural Architecture)"]
  B --> C["Data Input<br>Long history of equity returns"]
  C --> D["Computational Process<br>Learn nonlinear mapping in<br>empirical singular-vector basis"]
  D --> E{"Outcome Analysis"}
  E -->|"Analytical Cleaner"| F["High Out-of-Sample Error<br>Non-stationary violation"]
  E -->|"Proposed Method"| G["Lower Prediction Error<br>Optimal Bias-Variance Trade-off"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a physics‑informed neural architecture that learns a nonlinear mapping from empirical singular values to cleaned cross‑covariance values, preserving random‑matrix theory symmetries.

Source sentence
we address this gap by designing a random-matrix-inspired neural architecture that operates in the empirical singular-vector basis and learns a nonlinear mapping from empirical singular values to their corresponding cleaned values.

Assumes 5

  • Analytical cross‑covariance cleaners are derived under strong stationarity and large‑sample assumptions. statistical
    Source sentence
    Existing analytical cross‑covariance cleaners are derived under strong stationarity and large-sample assumptions
  • The random‑matrix‑inspired network can recover the analytical solution as a special case. model
    Source sentence
    By construction, the network can recover the analytical solution as a special case
  • Zero‑mean random vectors X and Y are considered for cross‑covariance estimation. model
    Source sentence
    Consider two zero–mean random vectors, X ∈R^{n_x} and Y ∈R^{n_y}, observed over a…
  • Universe selection enforces stable short‑term liquidity and sufficiently long homogeneous trading histories. market
    Source sentence
    Our objective is to construct a time‑varying investable universe that enforces (i) stable short‑term liquidity and (ii) sufficiently long and homogeneous trading histories
Show 1 more
  • Stocks that will be delisted within the subsequent 240 trading days are excluded to avoid look‑ahead bias. execution
    Source sentence
    any stock that will be delisted within the subsequent

Data

  • NYSE and NASDAQ listed equities market-prices public
    1990-01-01 to 2024-12-31 · All NYSE‑ and NASDAQ‑listed equities (excluding ETFs, funds, non‑operating vehicles, ADRs)
    Source sentence
    The dataset comprises all NYSE- and NASDAQ-listed equities from January 1, 1990, to December 31, 2024, including ADRs, and excludes ETFs, funds, and non‑operating vehicles.
  • Synthetic spectral‑regime datasets synthetic unknown
    Source sentence
    To assess our Neural Network (NN) based estimator, we first benchmark it on synthetic datasets that reproduce the spectral regimes of finite‑rank signals and heavy‑tailed noise

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • The proposed neural method achieves a more favorable bias‑variance trade‑off than purely analytical cleaners and yields systematically lower out‑of‑sample cross‑covariance prediction errors. predictive-accuracy
    Source sentence
    the proposed method achieves a more favorable bias–variance trade-off than purely analytical cleaners and delivers systematically lower out-of-sample cross–covariance prediction errors
  • On synthetic benchmarks the neural estimator matches the BBP analytical performance in stationary regimes. performance
    Source sentence
    our estimator matches the BBP performance across these validation tests
  • In market‑data OOS tests (2017–2024) the NN consistently outperforms BBP and CV‑based shrinkage in MSE, except at n=1000 where BBP is comparable. performance
    Source sentence
    both BBP and the CV-based shrinkage remain consistently worse than the NN in OOS MSE. The only notable exception occurs at n = 1000, where BBP attains NN-level performance

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

https://github.com/bongiornoc/CrossRIEnet — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (53)

  1. Optimal denoising of rotationally invariant rectangular matrices (2022) — link
  2. Optimal cleaning for singular values of cross-covariance matrices (2023) — link
  3. Filtering time-dependent covariance matrices using time-independent eigenvalues (2023) — link
  4. End-to-end large portfolio optimization for variance minimization with neural networks through covariance cleaning — → 2507.01918
  5. Neural Network-Driven Volatility Drag Mitigation under Aggressive Leverage (2025) — link
  6. Optimal shrinkage of singular values (2017) — link
  7. Non-linear shrinkage of the price return covariance matrix is far from optimal for portfolio optimization (2023) — link
  8. Predictive regressions (1999) — link
and 45 more
  1. Cross-asset signals and time series momentum (2020) — link
  2. Asset allocation with factor-based covariance matrices (2025) — link
  3. Risk, jumps, and diversification (2008) — link
  4. Of shepherds, sheep, and the cross-autocorrelations in equity returns (1995) — link
  5. Improving hedging performance by using high–low range (2022) — link
  6. Time series momentum (2012) — link
  7. Do industries lead stock markets? (2007) — link
  8. Machine learning for US cross-industry return predictability under information uncertainty (2023) — link
  9. Estimating Covariance for Global Minimum Variance Portfolio: A Decision-Focused Learning Approach — → 2508.10776
  10. Cleaning large correlation matrices: tools from random matrix theory (2017) — link
  11. The effect of errors in means, variances, and covariances on optimal portfolio choice (1993) — link
  12. The Schur Complement and Symmetric Positive Semidefinite (and Definite) Matrices (2019) — link
  13. Aspects of multivariate statistical theory (2009)
  14. Schur complements and statistics (1981) — link
  15. A whitening approach to probabilistic canonical correlation analysis for omics data integration (2019) — link
  16. Canonical Correlation Analysis [preprint] (2024) — link
  17. On persistence in mutual fund performance (1997) — link
  18. Common risk factors in the returns on stocks and bonds (1993) — link
  19. A five-factor asset pricing model (2015) — link
  20. Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation (1982) — link
  21. Generalized autoregressive conditional heteroskedasticity (1986) — link
  22. The dynamics of exchange rate volatility: A multivariate latent factor ARCH model (1989) — link
  23. Asset pricing with a factor-ARCH covariance structure: Empirical estimates for Treasury bills (1990) — link
  24. Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models (2002) — link
  25. Comparing predictive accuracy (1995) — link
  26. Principal portfolios (2023) — link
  27. A benchmark-asset principal component factorization for index tracking on large investment universes (2025) — link
  28. A Mean/Variance Analysis of Tracking Error (1992) — link
  29. Nonparametric Eigenvalue-Regularized Precision or Covariance Matrix Estimator (2016) — link
  30. Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices (2005) — link
  31. Design-free estimation of variance matrices (2014) — link
  32. Quadratic shrinkage for large covariance matrices (2022) — link
  33. 17 CFR \S 240.12d2-2 – Removal from Listing and Registration (2024)
  34. CrossRIEnet [software] (2026)
  35. RIEnet: A Rotational Invariant Estimator Network for Global Minimum-Variance Optimisation [software] (2026) — link
  36. Optimal cleaning for singular values of cross-covariance matrices [software] (2021) — link
  37. Library Name [software] (unpublished results)
  38. High dimensional covariance matrix estimation using a factor model (2008) — link
  39. Data Library (2026) — link
  40. Bootstrap methods: Another look at the jackknife (1979) — link
  41. The stationary bootstrap (1994) — link
  42. Automatic block-length selection for the dependent bootstrap (2004) — link
  43. Correction to ``Automatic block-length selection for the dependent bootstrap’’ (2009) — link
  44. A simple sequentially rejective multiple test procedure (1979) — link
  45. CRIENet : CrossRIEnet for Rectangular Cross-Correlation Cleaning (2026)

2 of 53 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.