Paper: arXiv 2410.22346

Authors: Alexa Orton, Tim Gebbie

Abstract

We consider financial market regime detection from the perspective of deep representation learning of the causal information geometry underpinning traded asset systems using a hierarchical correlation structure to characterise market evolution. We assess the robustness of three toy models: SPDNet, SPD-NetBN and U-SPDNet whose architectures respect the underlying Riemannian manifold of input block hierarchical SPD correlation matrices. Market phase detection for each model is carried out using three data configurations: randomised JSE Top 60 data, synthetically-generated block hierarchical SPD matrices and block-resampled chronology-preserving JSE Top 60 data. We show that using a singular performance metric is misleading in our financial market investment use cases where deep learning models overfit in learning spatio-temporal correlation dynamics.

Complexity vs Empirical Score

  • Math Complexity: 9.2/10
  • Empirical Rigor: 4.5/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced deep learning on Riemannian manifolds (SPD matrices) and uses synthetic data for controlled experiments, showing high mathematical complexity but limited backtest-ready implementation details, aligning with the Lab Rats quadrant.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Deep Representation Learning<br>for Financial Regime Detection"] --> B["Data Configurations"]
  B --> C["JSE Top 60<br>Randomized"]
  B --> D["Synthetic Block<br>Hierarchical SPD"]
  B --> E["JSE Top 60<br>Chronology-Preserving"]
  
  C & D & E --> F["Computational Process:<br>Three SPD-Net Variants<br>SPDNet / SPD-NetBN / U-SPDNet"]
  
  F --> G["Key Findings"]
  G --> H["Deep models overfit<br>spatio-temporal dynamics"]
  G --> I["Singular performance<br>metrics are misleading<br>in investment use cases"]

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 using deep representation learning models (SPDNet, SPDNetBN, U‑SPDNet) on hierarchical SPD correlation matrices for financial market regime detection.

Source sentence
We consider financial market regime detection from the perspective of deep representation learning of the causal information geometry underpinning traded asset systems using a hierarchical correlation structure to characterise market evolution.

Assumes 2

  • The analysis assumes underlying data may violate Gaussianity and stationarity, requiring robustness. statistical
    Source sentence
    Novel approaches have been developed to address the key problems related to the assumptions of Guassianity, stationarity, and the need to address robustness and reflexivity
  • Backtesting assumes no trading costs are incurred. execution
    Source sentence
    Fig. 11: MVPO backtesting. We perform daily-rebalanced portfolio optimisation backtesting on the OOS JSE Top 60 data set from 01-2008 - 07-2012 with no trading costs.

Data

  • JSE Top 60 daily equity returns · Bloomberg market-prices commercial
    January 2000 - December 2023 · Top 60 high‑capitalised stocks traded on the JSE
    Source sentence
    To conduct a real‑world application of the selected manifold learning techniques for regime classification, the set of daily equity returns of the top 60 high‑capitalised stocks traded on the JSE over the period January 2000 - December 2023 is considered.
  • Synthetic block‑hierarchical SPD correlation matrices synthetic unknown
    Source sentence
    The synthetic construction uses techniques for nested factor timeseries generation as proposed by Yelibi and Gebbie (2021) to create 5 clusters with 3 hierarchies each, sampled from Student’s t‑distribution with v = 3 d.o.f.

Recognised sources: Bloomberg

Costs

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

Finds 3

  • Riemannian batch‑normalisation layers improve market state classification accuracy across models. predictive-accuracy
    Source sentence
    the introduction of Riemannian batchnorm layers improves the market state classification accuracy, irrespective of the dataset used for training
  • Increasing model depth beyond the baseline does not improve classification accuracy; performance converges around 71% (real‑world) and 66% (synthetic). predictive-accuracy
    Source sentence
    increased depth holds no relationship with improved classification accuracy since the performance measure convergence occurs within a narrow band around 71% and 66% for the real‑world and synthetic data scenarios
  • Mean‑Variance portfolio optimisation outperforms both regime‑dependent strategies after 2010, though U‑SPDNet regime‑dependent strategy outperforms before 2010 during stressed periods. performance
    Source sentence
    Before 2010 the U‑SPDNet regime‑dependent strategy outperforms both the benchmark mean‑variance and SPDNet regime‑dependent strategies during the stressed market periods

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 14 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

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

Cited References (82)

  1. Epistemic Limits of Empirical Finance: Causal Reductionism and Self-Reference — → 2311.16570
  2. Dissecting financial markets: sectors and states (2002) — link
  3. Agglomerative likelihood clustering (2021) — link
  4. Hierarchically nested factor model from multivariate data (2007) — link
  5. Correlation, hierarchies, and networks in financial markets (2010)
  6. A tool for filtering information in complex systems (2005) — link
  7. DYNAMIC HIERARCHICAL FACTOR MODELS (2013)
  8. Hierarchical structure in financial markets (1999)
and 74 more
  1. Information and hierarchical structure in financial markets (1999)
  2. On clustering financial time series: a need for distances between dependent random variables (2016) — link
  3. Statistically validated hierarchical clustering: Nested partitions in hierarchical trees (2022)
  4. Noise Thresholds for Spectral Clustering (2011) — link
  5. Efficient Active Algorithms for Hierarchical Clustering (2012)
  6. The development of hierarchical factor solutions (1957)
  7. Matrix Evolutions: Synthetic Correlations and Explainable Machine Learning for Constructing Robust Investment Portfolios (2021)
  8. Data clustering and noise undressing of correlation matrices (2001)
  9. Returns-Driven Macro Regimes and Characteristic Lead-Lag Behaviour between Asset Classes (2022) — link
  10. How Regimes Affect Asset Allocation (2004)
  11. Emergence of time-horizon invariant correlation structure in financial returns by subtraction of the market mode (2007) — link
  12. A hybrid learning approach to detecting regime switches in financial markets (2020) — link
  13. CorrGAN: Sampling Realistic Financial Correlation Matrices Using Generative Adversarial Networks (2020) — link
  14. Hierarchical Causality in Financial Economics (2014) — link
  15. Taxonomy of stock market indices (2000)
  16. Financial Applications of Random Matrix Theory: a short review (2011)
  17. A first course in random matrix theory : for physicists, engineers and data scientists (2021)
  18. Noise Dressing of Financial Correlation Matrices (1999) — link
  19. The Science of Can and Can’t: A Physicist’s Journey through the Land of Counterfactuals (2021)
  20. The Poverty of Historicism, I (1944)
  21. Riemannian batch normalization for SPD neural networks (2019) — link
  22. A Riemannian Network for SPD Matrix Learning (2016) — link
  23. Geometric means in a novel vector space structure on symmetric positive-definite matrices (2007)
  24. Positive definite matrices and the S-divergence (2013) — link
  25. Training Deep Networks with Structured Layers by Matrix Backpropagation (2016) — link
  26. SEMANTIC SEGMENTATION METHOD WITH SECOND-ORDER POOLING (2015)
  27. The Geometry of Algorithms with Orthogonality Constraints (2006) — link
  28. Jean-Louis Koszul and the Elementary Structures of Information Geometry (2019)
  29. On the Fisher-Rao Information Metric in the Space of Normal Distributions (2019) — link
  30. A Riemannian Framework for Tensor Computing (2006)
  31. Natural Langevin Dynamics for Neural Networks (2017)
  32. Rao’s Distance Measure (1981)
  33. Information Geometry and Its Applications: On the Occasion of Shun-Ichi Amari’s 80th Birthday, IGAIA IV Liblice, Czech Republic, June 2016 (2018)
  34. Averaging Symmetric Positive-Definite Matrices (2020)
  35. Riemannian Gaussian Distributions on the Space of Symmetric Positive Definite Matrices (2017)
  36. Geometric means (2004)
  37. Riemannian geometry and matrix geometric means (2006)
  38. Riemannian median, geometry of covariance matrices and radar target detection (2010)
  39. Riemannian median and its estimation (2010) — link
  40. The geometric median on Riemannian manifolds with application to robust atlas estimation (2009)
  41. Parallel Transport on the Cone Manifold of SPD Matrices for Domain Adaptation (2019) — link
  42. Riemannian metric and geometric mean for positive semidefinite matrices of fixed rank (2009)
  43. The Riemannian Barycentre as a Proxy for Global Optimisation (2019)
  44. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift (2015) — link
  45. Gradient-based learning applied to document recognition (1998)
  46. DeepKSPD: Learning Kernel-Matrix-Based SPD Representation For Fine-Grained Image Recognition (2018) — link
  47. Transfer Learning: A Riemannian Geometry Framework With Applications to Brain-Computer Interfaces (2018)
  48. Optimization Algorithms on Matrix Manifolds (2010)
  49. DeepKSPD: Learning Kernel-matrix-based SPD Representation for Fine-grained Image Recognition (2017) — link
  50. Classification of Correlation Matrices using SPDNet with Riemannian Batch Normalization (2021) — link
  51. Can we predict a market regime from correlation matrix features? (2020) — link
  52. Approximation by superpositions of a sigmoidal function (1989) — link
  53. DreamNet: A Deep Riemannian Network based on SPD Manifold Learning for Visual Classification (2022) — link
  54. U-SPDNet: An SPD manifold learning-based neural network for visual classification (2022) — link
  55. U-Net: Convolutional Networks for Biomedical Image Segmentation (2015) — link
  56. Geometry Regularized Autoencoders (2023) — link
  57. Tactical investment algorithms through Monte Carlo backtesting (2022)
  58. Statistical Consequences of Fat Tails: Real World Preasymptotics, Epistemology, and Applications (2022) — link
  59. timeseries_generator (2021) — link
  60. BH-SPD-CorrMat-Nets (2024) — link
  61. Representation Learning for Regime Detection in Block Hierarchical Financial Markets (2024)
  62. A Review of Two Decades of Correlations, Hierarchies, Networks and Clustering in Financial Markets (2021) — link
  63. Advances in financial machine learning (2018)
  64. Market state discovery (2022)
  65. The Impact of Bootstrap Methods on Time Series Analysis (2003)
  66. Portfolio optimization based on neural networks sensitivities from assets dynamics respect common drivers (2023)
  67. GNU Parallel 20230822 (‘Chandrayaan’) (2023) — link
  68. Dynamic portfolio optimization across hidden market regimes (2018) — link
  69. Detecting intraday financial market states using temporal clustering (2016) — link
  70. Modeling and pricing long memory in stock market volatility (1996)
  71. Fallibility, reflexivity, and the human uncertainty principle (2013)
  72. Empirical properties of asset returns: stylized facts and statistical issues (2000)
  73. Portfolio optimization for student t and skewed t returns (2010)
  74. Second-Order Networks in PyTorch (2019)

1 of 82 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.