EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Volatility-informed representation for time-series clustering

Prakarn Taranodom, Thapana Boonchoo

Science, Engineering and Health Studies · 16.11.2026 · 26020004

DOI: 10.69598/sehs.20.26020004 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Science, Engineering and Health Studies, online erschienen am 16.11.2026. Verfasst von Prakarn Taranodom und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Time series clustering is a challenging task due to the presence of noise, non-stationarity, and volatility in real-world data. Traditional methods that rely on raw data often fail to capture meaningful temporal patterns. To address this issue, a time series clustering framework called enhanced clustering using conditional volatility (ECCV) is proposed, which systematically integrates conditional volatility (CV) as an alternative feature representation to improve clustering accuracy. In the feature extraction stage, an autoregressive integrated moving average - generalized autoregressive conditional heteroskedasticity (ARIMA-GARCH) model is applied to generate CV from the residuals of time series data, filtering out noise, preserving local structural patterns, and capturing temporal fluctuations that are often underutilized by conventional methods. The extracted CV is then employed as input for a wide range of clustering algorithms, including partition-based, hierarchical, shape-based, spectral, and neural network-based approaches. The main novelty of this work lies in the systematic evaluation and integration of CV for time series clustering, which, to the best of our knowledge, has not been comprehensively investigated before. Its effectiveness is demonstrated across 26 benchmark datasets through both quantitative metrics (the Rand Index) and statistical validation using Wilcoxon signed-rank tests, as well as visualizations and domain-specific analyses to show selective but meaningful improvements. Experiments reveal that CV-enhanced features significantly improve clustering performance in several algorithms, particularly those leveraging latent representations or graph-based embeddings, although their impact varies depending on the dataset type and algorithm. Overall, ECCV provides a scalable, interpretable, and statistically grounded framework for leveraging volatility information in time series clustering, offering a novel perspective on representation learning in dynamic and volatile environments.

Zitieren

APA

Taranodom, P., & Boonchoo, T. (2026). Volatility-informed representation for time-series clustering. Science, Engineering and Health Studies, 26020004. https://doi.org/10.69598/sehs.20.26020004

Vancouver

Taranodom P, Boonchoo T. Volatility-informed representation for time-series clustering. Science, Engineering and Health Studies. 2026:26020004. doi:10.69598/sehs.20.26020004

BibTeX
@article{Taranodom2026,
  author = {Taranodom, Prakarn and Boonchoo, Thapana},
  title = {Volatility-informed representation for time-series clustering},
  journal = {Science, Engineering and Health Studies},
  year = {2026},
  pages = {26020004},
  doi = {10.69598/sehs.20.26020004}
}
Kostenlos starten →