EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Stability-Driven Approaches to Model Selection in Time Series Analysis

Rajasekhar Kalla, Hemanth Kumar Molapata, Hemasekhar Palla

Boletim da Sociedade Paranaense de Matemática · 19.10.2026 · 44 (8) · 1-12

DOI: 10.5269/bspm.84291 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Boletim da Sociedade Paranaense de Matemática, online erschienen am 19.10.2026. Verfasst von Rajasekhar Kalla und 2 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Model selection intricately intertwines with the assessment of models, presenting a formidable challenge when confronted with the task of evaluating multiple models, their respective outputs, or parameters within the context of a specific learning activity. This process assumes a pivotal role across a spectrum of data types, spanning both unstructured and structured domains, including supervised, semi-supervised, and unsupervised datasets. The complexity is accentuated in the realm of time series and forecasting analysis, where the inherent characteristics of temporal data introduce heightened challenges. In the landscape of time series and forecasting analysis, the abundance of data compounds the intricacies of model selection. With potentially thousands or even lakhs of time series, each manifesting distinct historical patterns ranging from truncated historical data and modest sales volumes to intermittent patterns and sustained trends, the task becomes notably daunting. Adding to this complexity, practitioners may employ varying forecast horizons, introducing an additional layer of intricacy. This dynamic setting renders the process of model fitting and selection particularly challenging, as models may necessitate adjustments when transitioning from one forecasting horizon to another. Traditionally, practitioners have relied on time series cross-validation techniques to appraise models. However, this paper introduces a pioneering model assessment methodology founded on diverse accuracy metrics, varying horizons, the utilization of multiple training datasets for a singular time series, and the inclusion of diverse models. Through experimentation, we conducted a comparative analysis of multiple models, evaluating their performance across distinct training periods, forecast horizons, and accurate metrics. The outcomes of our approach delineate the optimal model for each geographical product, surpassing the performance of alternative models subjected to the same training. This methodology not only alleviates a significant amount of manual effort but also yields a marked enhancement in the overall process.

Zitieren

APA

Kalla, R., Molapata, H. K., & Palla, H. (2026). Stability-Driven Approaches to Model Selection in Time Series Analysis. Boletim da Sociedade Paranaense de Matemática, 44(8), 1-12. https://doi.org/10.5269/bspm.84291

Vancouver

Kalla R, Molapata HK, Palla H. Stability-Driven Approaches to Model Selection in Time Series Analysis. Boletim da Sociedade Paranaense de Matemática. 2026;44(8):1-12. doi:10.5269/bspm.84291

BibTeX
@article{Kalla2026,
  author = {Kalla, Rajasekhar and Molapata, Hemanth Kumar and Palla, Hemasekhar},
  title = {Stability-Driven Approaches to Model Selection in Time Series Analysis},
  journal = {Boletim da Sociedade Paranaense de Matemática},
  year = {2026},
  volume = {44},
  number = {8},
  pages = {1--12},
  doi = {10.5269/bspm.84291}
}
Kostenlos starten →