// the find
facebookresearch/Kats
Kats, a kit to analyze time series data, a lightweight, easy-to-use, generalizable, and extendable framework to perform time series analysis, from understanding the key statistics and characteristics, detecting change points and anomalies, to forecasting future trends.
Kats is Facebook's time series toolkit covering forecasting, anomaly/changepoint detection, feature extraction, and model selection under one roof. It wraps Prophet, ARIMA, SARIMA, Holt-Winters, LSTM, and more behind a unified TimeSeriesData interface. Aimed at data scientists who want a breadth-first starting point rather than depth in any single method.
The detector collection is genuinely wide — CUSUM, BOCPD, ProphetTrendDetector, DTW-based detectors, and a meta-learning layer that recommends which detector to use for a given series. The TsFeatures extractor pulls 65+ statistical features in one call, which saves real work when you need to build downstream classifiers. The self-supervised hyperparameter tuning paper (arXiv:2102.05740) is backed by actual production use at Meta scale, not just a demo. The minimal install path (MINIMAL_KATS=1) is a thoughtful addition — you can pull in only what you need without dragging the full Prophet/neuralprophet/torch dependency chain.
The latest PyPI release is 0.2.0 from 2022 — the version string has not moved in four years even though commits continue, which makes it hard to know what you're actually getting from pip vs. main. Dependency management is a mess: the full install pulls Prophet, NeuralProphet, PyTorch, and statsmodels simultaneously, and version conflicts on Apple Silicon or newer Python environments are a recurring source of pain. The TimeSeriesData wrapper is an extra layer that doesn't buy you much over a plain pandas DataFrame and adds friction when integrating with other libraries. Documentation outside the Jupyter tutorials is sparse; the Sphinx site mostly mirrors docstrings, so understanding non-obvious model interactions requires reading source.