Modelling Time Series of Counts with Hysteresis
Modelling Time Series of Counts with Hysteresis
Authors: Xuanye Ma, Dong Li, Howell Tong Date: 2026-04-13 Paper ID: openalex:2509.15508
Summary
This paper presents the Hysteretic Poisson Autoregressive (HPART) model, a novel approach for modelling count-valued time series that incorporates scientifically grounded hysteresis into a piecewise linear structure. By extending traditional Poisson autoregressive dynamics, the model improves upon the existing Buffered Poisson Autoregressive (BPART) model by offering more nuanced regime-switching capabilities. The authors provide a rigorous statistical foundation for estimation and testing and demonstrate superior interpretability and out-of-sample forecasting performance on real-world datasets.
Key Contributions
- Introduces the Hysteretic Poisson Autoregressive (HPART) model, a nonlinear extension of linear Poisson autoregressive models incorporating genuine hysteresis via a controlling factor.
- Establishes a unified framework for maximum likelihood estimation and asymptotic property derivation for both HPART and Buffered Poisson Autoregressive (BPART) models.
- Develops hypothesis tests for non-nested model comparison between BPART and HPART models to guide empirical selection.
Open Questions & Future Work
Archivist Review
I have approved the open question regarding intercept variance in count models because it identifies a concrete, persistent theoretical gap that transcends the specific HPART model introduced here. The proposed model itself (HPART) was rejected because it is a specific, specialized statistical architecture that does not meet the criteria for a high-level, reusable concept or paradigm.
Approved Open Questions
- Intercept parameter variance in count models: Understanding this persistent estimation issue is crucial for improving the robustness and interpretability of count time series models, which are widely used in financial and public health forecasting.
Rejected Candidates
- [concept] Hysteretic Poisson Autoregressive (HPART) model (
hpart-model) - not_reusable: This is a specific, narrow statistical model rather than a general-purpose concept likely to recur as a primary research topic across future domains.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2509.15508
- paper_id
- 2509.15508
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-16T05:46:15Z