// the find
sintel-dev/Orion
Unsupervised time series anomaly detection library
Orion is a Python library from MIT's DAI Lab that wraps a collection of unsupervised anomaly detection algorithms — LSTM autoencoders, GANs, VAEs, ARIMA, matrix profiles — behind a single fit/detect API. It's aimed at data scientists working with telemetry, sensor, or other univariate time series who want to benchmark multiple approaches without reimplementing each one. The pipeline abstraction is JSON-configured, which is either convenient or annoying depending on how much you like configuring neural nets in JSON.
The benchmarking infrastructure is genuinely useful — they run all pipelines against 12 labeled datasets every release and publish per-signal results, so you can actually see which methods fail on which signal types rather than trusting aggregate numbers. The AER pipeline (auto-encoder with regression) consistently beats ARIMA across all 12 datasets, which is a concrete and verifiable claim. Active maintenance with a push as recent as July 2026 and benchmarks versioned per release means regression tracking is real. The primitives-as-JSON pipeline design means you can swap components or tune hyperparameters without touching Python, which is useful for hyperparameter search workflows.
Still marked Pre-Alpha after years of development, and that badge isn't wrong — the API surface has shifted enough across versions that tutorials on older blog posts will mislead you. Multivariate time series support is basically absent; everything assumes a single value column, which rules out most real-world monitoring scenarios where you have correlated signals. The JSON pipeline configuration gets messy fast: non-trivial customizations require understanding the MLBlocks primitive spec, which is an additional abstraction layer with its own learning curve and sparse documentation. Dependency footprint is heavy — TensorFlow/Keras, PyTorch, and statsmodels all pulled in depending on which pipelines you use, which makes environment management painful.