Skip to content

MGAB (Mackey-Glass Anomaly Benchmark)

What is this data? Ten univariate series generated from the chaotic Mackey-Glass delay differential equation, each with 10 synthetically inserted anomalies that are invisible to the naked eye — the series looks identical before and after each anomaly. A stress test for methods that claim to model temporal dynamics rather than just detect visual outliers.

  • Source: https://github.com/MarkusThill/MGAB (Thill, Konen & Bäck, 2020)
  • Contents: 10 series × 100k points; columns value, is_anomaly, is_ignored
  • License: CC0 1.0 (public domain) — fully redistributable
  • Download: tsad-forge download mgab (or --subset 1,2)
  • Loader: load_mgab(series=1)

Known flaws / caveats

  • No official train/test split — our loader uses the longest anomaly-free prefix (capped at 30%) as train; a documented deviation.
  • Purely synthetic chaos: excellent for dynamics-modeling claims, but says nothing about sensor noise, drift or regime changes found in real data.
  • The is_ignored mask (transition segments the original benchmark excludes) is recorded in meta but not applied to metrics here.

Quick EDA (from local data)

Sample test channels with labeled anomalies shaded, plus the event-length distribution. Regenerate with tsad-forge viz after downloading the data.