Bayesian Modelling of Nonstationary Extreme Values Using a Nonparametric Hawkes Process
Bayesian Modelling of Nonstationary Extreme Values Using a Nonparametric Hawkes Process
Authors: Gordon J. Ross, Dean Markwick Date: 2026-05-05 Paper ID: openalex:2605.03331
Summary
This paper introduces a Bayesian nonparametric framework for modeling nonstationary extreme value processes by combining self-exciting Hawkes processes with hierarchical Extreme Value Theory. By utilizing a Dirichlet process to capture temporal excitation patterns and a Generalized Pareto Distribution with partial pooling for event magnitudes, the model effectively handles clustered and sparse extreme event data. Empirical results across four real-world datasets confirm the approach’s effectiveness in predictive performance for extreme events.
Key Contributions
- Proposes a Bayesian point process model that integrates a self-exciting Hawkes process with nonparametric Dirichlet process learning to capture nonstationary event rates.
- Integrates Extreme Value Theory via a Generalised Pareto Distribution with a hierarchical mark structure, enabling partial pooling for robust magnitude estimation in sparse clusters.
- Demonstrates superior predictive performance on four real-world datasets compared to existing extreme value modeling baselines.
Open Questions & Future Work
Key Concepts
- nonparametric-hawkes-extreme-value-modeling: A Bayesian nonparametric framework that models extreme event sequences as clustered processes with magnitudes governed by a hierarchically-pooled Generalized Pareto Distribution.
Archivist Review
The paper presents a coherent framework for handling nonstationary extremes using Bayesian point processes. I have approved the overarching model as a key concept, as it represents a significant, reusable paradigm in time series analysis for extremes. The open question was approved for its focus on addressing real-world complexities (covariates, nonstationarity, and multivariate dependencies) in extreme event forecasting. No datasets were approved as none were specifically named as persistent research benchmarks.
Approved Concepts
- Nonparametric Hawkes Extreme Value Modeling: Combines point process self-excitation with extreme value theory to handle nonstationarity in extreme event occurrence and magnitude, providing a robust Bayesian alternative to stationary extremes modeling.
Approved Open Questions
- Extending Nonstationary Extreme Models: These extensions are critical for moving from idealized, self-contained data analysis toward robust, real-world forecasting systems that can account for exogenous drivers and interconnected risk events.
Rejected Candidates
- [concept] Hierarchical Generalized Pareto Pooling (
nonparametric-hawkes-process-mark-pooling) - subcomponent_of_broader_mechanism: This is a specific subcomponent of the broader nonparametric Hawkes extreme value modeling framework which is being approved as a single concept.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.03331
- paper_id
- 2605.03331
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- nonparametric-hawkes-extreme-value-modeling
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-08T05:44:52Z