Skip to content

Learning Time-Inhomogeneous Markov Dynamics in Financial Time Series via Neural Parameterization

Home / Papers / Learning Time-Inhomogeneous Markov Dynamics in Financial Time Series via Neural Parameterization

Learning Time-Inhomogeneous Markov Dynamics in Financial Time Series via Neural Parameterization

Authors: Jan Rovirosa, Jesse Schmolze Date: 2026-05-06 Paper ID: openalex:2605.04690

Summary

This paper addresses the statistical degeneracy of traditional Markov transition operators in high-resolution, noisy environments by parameterizing them with neural networks. By forcing the neural output into a formal stochastic operator structure, the model ensures mathematical transparency while capturing complex non-stationary dynamics. The approach is validated on financial time series, revealing that transition dynamics become more homogeneous during high-volatility regimes and providing a novel diagnostic for checking first-order memory violations.

Key Contributions

  • Introduces a framework that uses neural networks as parameterization engines to generate explicit, time-varying Markov transition matrices for non-stationary financial data.
  • Demonstrates that the learned stochastic operators capture complex regime shifts, with row entropy showing a strong correlation (r = -0.62) with realized volatility.
  • Proposes using the Chapman-Kolmogorov equations as a diagnostic tool for identifying temporal windows where first-order Markov assumptions fail in high-noise financial data.

Open Questions & Future Work

Key Concepts

  • neural-parameterized-markov-operator: A framework that uses neural networks as parameterization engines to output time-varying, valid stochastic matrices while maintaining structural interpretability.
  • chapman-kolmogorov-diagnostic: Using the Chapman-Kolmogorov equations as a local diagnostic tool to identify time windows where the first-order memory assumption of a Markov model fails.

Archivist Review

I approved two concepts that represent the paper’s core methodology—using NNs to parameterize structured operators and repurposing classical equations for diagnostic use—and two open questions regarding the scalability and temporal continuity of these structures. These items are highly reusable in the context of structured, interpretable time-series modeling. No datasets were approved as none were provided.

Approved Concepts

  • Neural-Parameterized Markov Operator: Provides a hybrid approach that bridges the gap between high-capacity deep learning and the interpretability of classical stochastic transition operators.
  • Chapman-Kolmogorov Diagnostic: Introduces a principled way to validate the Markovian assumption in high-noise temporal data, shifting the focus from enforcement to diagnostic utility.

Approved Open Questions

  • Multi-Asset Scaling Dynamics Bottleneck: Essential for the practical utility of the framework in systematic risk management and for overcoming signal-to-noise limitations inherent in single-asset modeling.
  • Continuous-Time Operator Modeling Bottleneck: Technically necessary to align structured operator modeling with the nature of continuous-time stochastic calculus and high-frequency trading data.

Metadata & Links

url
https://arxiv.org/abs/2605.04690
paper_id
2605.04690
paper_source
openalex
domain
time-series
tags
architectures
datasets
concept_slugs
neural-parameterized-markov-operatorchapman-kolmogorov-diagnostic
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-09T05:57:13Z