Skip to content

UCR Anomaly Archive (2021)

What is this data? 250 univariate series hand-curated by the UCR group from domains like ECG, EPG (insect probing), gait, and industrial sensors — each with exactly one confirmed anomaly. Built specifically to fix the flaws of older benchmarks; the reference set for univariate discord discovery.

  • Source: Wu & Keogh (TKDE 2021) — https://www.cs.ucr.edu/~eamonn/time_series_data_2018/
  • Contents: 250 univariate series. The filename encodes the metadata: NNN_UCR_Anomaly_<name>_<trainEnd>_<anomStart>_<anomEnd>.txt
  • License: free for research
  • Download: tsad-forge download ucr (zip, ~250 MB)
  • Loader: load_ucr(series=1) or load_ucr(series="001_UCR_Anomaly_...txt")

Design intent and caveats

  • The archive was built as a critique of flawed benchmarks (trivial anomalies, mislabeled data, excessive prior information) — with a strict one-anomaly-per-series convention.
  • Because of that single-anomaly convention, event-level metrics behave almost binarily on this archive. Keep that in mind when reading VUS-PR.
  • Files with DISTORTED/NOISE prefixes are variants of the same underlying series — the 250 series are not fully independent.