Characterizing Differences Across Time Series Datasets

Characterize the specific differences in characteristics across various time series datasets.

Background

Time series datasets from different domains (e.g., energy, economics, weather, and health) often exhibit distinct properties such as sampling frequency, periodicity, noise characteristics, and non-stationarity. Understanding these differences is crucial to building models that generalize across domains or are tailored effectively to domain-specific traits.

The authors explicitly defer a systematic treatment of these dataset-specific differences, identifying it as a concrete item for future work.

References

We left the specific differences in characteristics across various time series datasets for future work.

— A Comprehensive Survey of Deep Learning for Time Series Forecasting: Architectural Diversity and Open Challenges  (2411.05793 - Kim et al., 2024) in Conclusion, Limitations and Future Work (Section 6)

One plausible source of the difference is traffic composition: the Danish corpus's ferry-dominated straits carry more strongly time-of-day and day-of-week structured routines than a port and waterway dominated by tug, tow and tanker transits (Section~\ref{sec:data}), so a chronological split has less regular short-term structure to leak across than a vessel-sharing one does on this corpus; we did not test this explanation directly.

— Protocol before progress: leakage-aware evaluation of AIS trajectory prediction  (2609.25827 - Raisi et al., 22 Sep 2026) in Section 6, subsection “Replication on the NOAA corpus”