Measuring Tail Dependence in Linear Processes: Theory and Empirics
Measuring Tail Dependence in Linear Processes: Theory and Empirics
Authors: Debanjana Datta, Diganta Mukherjee Date: 2026-05-11 Paper ID: openalex:2605.10303
Summary
This paper addresses the limitations of standard Gaussian models in capturing heavy-tailed distributions and extreme co-movements in financial time series. By leveraging copula theory and regular variation, the authors introduce a novel dependence measure designed to characterize joint extreme behavior. The framework is applied to high-frequency cryptocurrency datasets, with particular focus on the role of persistence properties in governing tail dependence.
Key Contributions
- Proposes a novel dependence measure for capturing joint extremes in linear processes that accounts for both identical and non-identical regularly varying distributions.
- Analyzes the impact of persistence properties on extreme co-movements within financial time series, specifically within the context of high-frequency cryptocurrency data.
- Validates the theoretical findings through extensive simulation studies, demonstrating the measure’s efficacy beyond standard Gaussian frameworks.
Open Questions & Future Work
Archivist Review
The submission proposes a mathematical dependence measure for extreme values in linear processes, which lacks the broader architectural or procedural reusability required for a permanent ML concept note. The open question regarding the robustness of dependence measures in the face of non-stationary or corrupted tail data remains a relevant and persistent challenge in time-series analysis, and is thus approved.
Approved Open Questions
- Robust Tail Dependence Measurement: Ensuring the robustness of dependence measures against data contamination is critical for reliable risk management and financial modeling, particularly in extreme scenarios where data quality can be compromised.
Rejected Candidates
- [concept] Novel Dependence Measure for Linear Processes (
novel-dependence-measure-for-linear-processes) - not_reusable: The proposal is framed as a mathematical tool for linear processes rather than a broadly reusable machine learning mechanism or inductive bias.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.10303
- paper_id
- 2605.10303
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-14T06:10:32Z