Paper: arXiv 2510.19130

Authors: Andres Garcia-Medina

Abstract

Covariance matrices estimated from short, noisy, and non-Gaussian financial time series are notoriously unstable. Empirical evidence suggests that such covariance structures often exhibit power-law scaling, reflecting complex, hierarchical interactions among assets. Motivated by this observation, we introduce a power-law covariance model to characterize collective market dynamics and propose a hybrid estimator that integrates Random Matrix Theory (RMT) with deep Residual Neural Networks (ResNets). The RMT component regularizes the eigenvalue spectrum in high-dimensional noisy settings, while the ResNet learns data-driven corrections that recover latent structural dependencies encoded in the eigenvectors. Monte Carlo simulations show that the proposed ResNet-based estimators consistently minimize both Frobenius and minimum-variance losses across a range of population covariance models. Empirical experiments on 89 cryptocurrencies over the period 2020-2025, using a training window ending at the local Bitcoin peak in November 2021 and testing through the subsequent bear market, demonstrate that a two-step estimator combining hierarchical filtering with ResNet corrections produces the most profitable and well-balanced portfolios, remaining robust across market regime shifts. Beyond finance, the proposed hybrid framework applies broadly to high-dimensional systems described by low-rank deformations of Wishart ensembles, where incorporating eigenvector information enables the detection of multiscale and hierarchical structure that is inaccessible to purely eigenvalue-based methods.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)

Why this score: The paper employs advanced mathematics including Random Matrix Theory (RMT), Wishart ensembles, spiked covariance models, and deep ResNet architectures, indicating high mathematical complexity. It also demonstrates empirical rigor through Monte Carlo simulations and out-of-sample testing on real cryptocurrency data across bull and bear market regimes, evaluating portfolio performance with specific loss functions.

Research Flowchart

  flowchart TD
  A["<b>Research Goal</b><br>Improve Covariance Estimation for Short, Noisy Financial Time Series<br>(Power-law scaling & Hierarchical structure)"] --> B["<b>Data & Model</b><br>Data: 89 Cryptocurrencies (2020-2025)<br>Model: Power-law Covariance Generation"]

  B --> C["<b>Methodology Pipeline</b><br>Hybrid ResNet + RMT Estimator"]

  C --> D["<b>Step 1: RMT Regularization</b><br>Filter noise via eigenvalue spectrum regularization"]
  C --> E["<b>Step 2: Deep Learning Correction</b><br>ResNet learns eigenvector corrections"]

  D & E --> F["<b>Monte Carlo Simulation</b><br>Test on synthetic populations<br>Loss functions: Frobenius & Minimum-variance"]

  F --> G["<b>Key Findings & Outcomes</b><br>1. ResNet reduces estimation error<br>2. Two-step estimator yields robust portfolios<br>3. Detects multiscale hierarchical structure"]

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 hybrid covariance estimator that combines Random Matrix Theory regularization of eigenvalues with a deep Residual Neural Network that learns corrections to eigenvectors.

Source sentence
we introduce a power-law covariance model to characterize collective market dynamics and propose a hybrid estimator that integrates Random Matrix Theory (RMT) with deep Residual Neural Networks (ResNets). The RMT component regularizes the eigenvalue spectrum in high-dimensional noisy settings, while the ResNet learns d

Assumes 5

  • The training period ends at the Bitcoin peak on 2021-11-09 and the out‑of‑sample test period covers the subsequent bear market. market
    Source sentence
    We split our dataset into training and testing periods. The splitting point is set to 2021-11-09, corresponding to the maximum closing value of Bitcoin (BTC) during the pandemic turmoil.
  • Stablecoins are excluded because they do not reflect intrinsic cryptocurrency dynamics. data
    Source sentence
    Finally, we exclude all stablecoins (a total of 21), such as USDT or EURS, as their values are designed to remain pegged to fiat currencies or commodities (e.g., gold) and thus do not reflect the intrinsic dynamics of cryptocurrency markets.
  • Portfolio allocations are computed using the MVP+ investment strategy on the estimated covariance matrices. execution
    Source sentence
    we obtain the optimal portfolio allocations w by applying the MVP+ investment strategy, using as input each of the covariance estimators presented in Section 3
  • The ResNet models are retrained at each rebalancing point with a rolling window of approximately one additional preceding year. execution
    Source sentence
    ... the model is retrained at each rebalancing point using an extended training dataset that includes the in-sample period plus approximately one additional preceding year (282 days).
Show 1 more
  • Turnover is used as a proxy for transaction costs, but no explicit cost model is incorporated into performance metrics. other
    Source sentence
    Turnover measures the absolute difference between the asset weights at consecutive periods, normalized by the number of periods. A high value is associated with high transaction fees.

Data

Backtest sample: 2021-11 – 2025-05 · cryptocurrency market · crypto · daily

  • Daily returns of 89 major (non‑stable) cryptocurrencies · yfinance API market-prices public
    2020-08-02 to 2025-07-31 · Top 89 cryptocurrencies by market capitalization after filtering
    Source sentence
    We analyze the daily returns of the p = 89 major (non‑stable) cryptocurrencies by market capitalization over a five‑year period, from 2020-08-02 to 2025-07-31, yielding a total of n = 1825 observations. The data are retrieved using the yfinance API.

Recognised sources: Yahoo Finance

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
Turnover measures the absolute difference between the asset weights at consecutive periods, normalized by the number of periods. A high value is associated with high transaction fees.

Finds 5

Reported performance (Ξ2S(CNN) two-step CNN-based estimator): Sharpe 1.34 (benchmark -0.50) · annual return 17.1% · total return 174.0% · max drawdown 63.6%

  • The two‑step CNN‑based estimator (Ξ2S(CNN)) achieves the highest cumulative portfolio return (17.14%) and the best Sharpe ratio among all estimators in the walk‑forward MVP+ analysis. performance
    Source sentence
    Ξ2S(CNN) 1.74 17.14% 46.60% 0.37 -63.63% 1.34
  • The hybrid two‑step estimator (Ξ2S(H)) attains the minimum MV loss for the block‑diagonal covariance model, while the single‑step ΞCNN minimizes Frobenius loss for that model. theoretical
    Source sentence
    We found that the two‑step estimator based on the ResNet architecture (Ξ2S(CNN)) minimizes the Frobenius loss function for the block‑diagonal model. For the same model, the two‑step hybrid estimator (Ξ2S(H)) attains the minimum value of the MV loss function.
  • When the covariance follows a power‑law structure, none of the state‑of‑the‑art estimators outperform the naive estimator in Frobenius loss, but the two‑step hybrid estimator yields the lowest MV loss. theoretical
    Source sentence
    in this case, none of the state‑of‑the‑art estimators outperform the naive one in terms of the Frobenius loss. However, regarding the MV loss, the two‑step hybrid estimator achieves the lowest noise level.
  • Empirical cryptocurrency data exhibit an eigenvalue spectrum with an approximate power‑law decay (α = 0.2), matching the stylized facts of both nested hierarchical and power‑law covariance models. statistical
    Source sentence
    the eigenvalue spectrum displays an approximate power‑law decay with coefficient α = 0.2.
  • Across all models, MV loss systematically decreases when using ResNet‑based estimators, indicating better risk control for portfolio optimization. statistical
    Source sentence
    a consistent pattern emerges MV loss systematically decreases across all models when using the ResNet‑based estimators.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 36 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

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

Cited References (87)

  1. Hierarchically nested factor model from multivariate data (2007)
  2. Distribution of eigenvalues for some sets of random matrices (1967)
  3. The power of (non-) linear shrinking: A review and guide to covariance matrix estimation (2022)
  4. Shrinkage estimation of large covariance matrices: Keep it simple, statistician? (2021)
  5. Kullback-Leibler distance as a measure of the information filtered from multivariate data (2007)
  6. Noise dressing of financial correlation matrices (1999)
  7. Random matrix approach to cross correlations in financial data (2002)
  8. Cleaning large correlation matrices: tools from random matrix theory (2017)
and 79 more
  1. On the distribution of the largest eigenvalue in principal components analysis (2001)
  2. Estimation of a covariance matrix (1975)
  3. Eigenvectors of some large sample covariance matrix ensembles (2011)
  4. Numerical implementation of the QuEST function (2017)
  5. Analytical nonlinear shrinkage of large-dimensional covariance matrices (2020)
  6. Cleaning large-dimensional covariance matrices for correlated samples (2022)
  7. Optimal estimation of a large-dimensional covariance matrix under Stein’s loss (2018)
  8. The Student ensemble of correlation matrices: eigenvalue spectrum and Kullback-Leibler entropy (2007)
  9. Cross-validation based nonlinear shrinkage (2016) — link
  10. Shrinkage and spectral filtering of correlation matrices: a comparison via the Kullback-Leibler distance (2007)
  11. Free probability and random matrices (2017)
  12. A well-conditioned estimator for large-dimensional covariance matrices (2004)
  13. Cluster analysis for applications: probability and mathematical statistics: a series of monographs and textbooks (2014)
  14. Large dynamic covariance matrices (2019)
  15. No eigenvalues outside the support of the limiting spectral distribution of large-dimensional sample covariance matrices (1998)
  16. A First Course in Random Matrix Theory: For Physicists, Engineers and Data Scientists (2020)
  17. Spectral moments of correlated Wishart matrices (2005)
  18. Applying free random variables to random matrix analysis of financial data. Part I: The Gaussian case (2011)
  19. Machine learning for asset managers (2020)
  20. Overlaps between eigenvectors of correlated random matrices (2018)
  21. Matrix Iterative Analysis (2000)
  22. Localization of electron wave functions in disordered systems (1972)
  23. Applied multivariate statistical analysis (2002)
  24. New introduction to multiple time series analysis (2005)
  25. Optimal shrinkage of eigenvalues in the spiked covariance model (2018)
  26. Spectral density of the correlation matrix of factor models: A random matrix theory approach (2005)
  27. Cluster analysis for portfolio optimization (2008)
  28. Covariance matrix filtering with bootstrapped hierarchies (2021)
  29. Reactive global minimum variance portfolios with k-BAHC covariance cleaning (2022)
  30. When do improved covariance matrix estimators enhance portfolio optimization? An empirical comparative study of nine estimators (2011)
  31. Spectral properties of correlation matrices for hierarchically nested factor models (2007)
  32. Correlation, hierarchies, and networks in financial markets (2010)
  33. Two-step estimators of high-dimensional correlation matrices (2023)
  34. Criticality in a multisignal system using principal component analysis (2021)
  35. Introduction to risk parity and budgeting (2013)
  36. A robust statistics approach to minimum variance portfolio optimization (2015)
  37. Random matrix methods for machine learning (2022)
  38. Convex optimization (2004)
  39. Building diversified portfolios that outperform out-of-sample (2016)
  40. An Introduction to the Hierarchical Risk Parity Algorithm (2021) — link
  41. Capital allocation to business units and sub-portfolios: The Euler principle (2008)
  42. Market Beta is not dead: An approach from Random Matrix Theory (2023)
  43. Introduction to econophysics: correlations and complexity in finance (1999)
  44. The Architecture of Complexity (1962)
  45. More Is Different: Broken symmetry and the nature of the hierarchical structure of science (1972)
  46. The architecture of complexity (2007)
  47. Econophysics: Past and present (2017)
  48. Econophysics review: I. Empirical facts (2011)
  49. Free deterministic equivalents, rectangular random matrix models, and operator-valued free probability theory (2012)
  50. Random Matrix Theory and Nested Clustered Optimization on high-dimensional portfolios (2024)
  51. Portfolio selection (1952)
  52. Capital asset prices: A theory of market equilibrium under conditions of risk (1964)
  53. Fibonacci and Lucas numbers as tridiagonal matrix determinants (2004)
  54. Dynamic shrinkage estimation of the high-dimensional minimum-variance portfolio (2023)
  55. Two is better than one: Regularized shrinkage of large minimum variance portfolios (2024)
  56. High-dimensional covariance matrix estimators on simulated portfolios with complex structures — → 2412.08756
  57. End-to-End Large Portfolio Optimization for Variance Minimization with Neural Networks through Covariance Cleaning — → 2507.01918
  58. A spectral algorithm for seriation and the consecutive ones problem (1998)
  59. Deep residual learning for image recognition (2016)
  60. Multiscale characteristics of the emerging global cryptocurrency market (2021) — link
  61. Random matrices (2004)
  62. Characterization of chaotic quantum spectra and universality of level fluctuation laws (1984)
  63. Random-matrix theory of quantum transport (1997)
  64. Financial applications of random matrix theory: Old laces and new pieces (2005)
  65. Why resnet works? residuals generalize (2020)
  66. Avoiding overconfidence: Evidence from the M6 financial competition (2025) — link
  67. Shrinkage estimation with reinforcement learning of large variance matrices for portfolio selection (2023)
  68. Deep learning for time series forecasting: predict the future with MLPs, CNNs and LSTMs in Python (2018)
  69. Neural networks and deep learning (2015)
  70. The intelligent investor (2003)
  71. The generalised product moment distribution in samples from a normal multivariate population (1928)
  72. An 'a lisis de datos multivariantes. Primera edici 'o n, vol 24, Editorial McGraw-Hill (2002)
  73. The largest eigenvalues of sample covariance matrices for a spiked population: diagonal case (2009)
  74. Limits of spiked random matrices I (2013)
  75. Active antenna selection in multiuser MIMO communications (2007)
  76. Statistics of resonance poles, phase shifts and time delays in quantum chaotic scattering: Random matrix approach for systems with broken time-reversal invariance (1997)
  77. Spectrum of the QCD Dirac operator and chiral random matrix theory (1994)
  78. Exact distribution of the maximal height of p vicious walkers (2008)
  79. Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices (2005)

2 of 87 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.