MIT's anomaly library runs a dozen algorithms against your time series

What it does Orion is a Python library that wraps multiple unsupervised anomaly detection pipelines for time series data. You load a signal with timestamps and values, pick a pipeline like aer or tadgan, fit it, then call detect() to get back intervals flagged as anomalous with severity scores. It ships with demo data and a scikit-learn-style API.

The interesting bit The project maintains a public leaderboard that scores every pipeline against 12 labeled datasets, measuring how often each beats a baseline ARIMA model. AER currently wins on all 12; Azure’s commercial detector wins on zero. That kind of baked-in benchmarking is unusual for open-source ML tooling.

Key highlights

• 13 verified pipelines including autoencoders, GANs, VAEs, transformers, and matrix profiles
• Built-in benchmark suite with ground-truth datasets and published comparison spreadsheets
• Pre-Alpha status (PyPI classifies it Development Status 2)
• Python 3.8 through 3.11 supported
• Part of the broader Sintel ecosystem from MIT’s Data to AI Lab

Caveats

• Pre-alpha status means APIs may shift; the README itself warns about harmless but noisy warnings during execution
• The leaderboard compares pipelines relative to ARIMA, not absolute accuracy metrics, so “wins” can obscure how small the margins are

Verdict Worth a look if you need to evaluate multiple anomaly detection approaches on time series without building the comparison harness yourself. Skip it if you want production-stable APIs or need real-time streaming detection—neither is claimed here.