// the find
business-science/pytimetk
Time series easier, faster, more fun. Pytimetk.
pytimetk is a Python library that wraps pandas and Polars to make time series feature engineering, aggregation, visualization, and anomaly detection less painful. It's a port of the R timetk package, aimed at data scientists who want to go from raw time series data to model-ready features without writing boilerplate resampling and rolling-window code. Not a forecasting library — it stops at feature engineering.
Dual-engine design (pandas + Polars) is genuinely useful: you can prototype with pandas and flip `engine='polars'` for a speed boost without changing the rest of your code. The anomaly detection workflow (`anomalize` → `plot_anomalies` → `plot_anomalies_cleaned`) is a coherent pipeline that most libraries leave you to wire together yourself. `augment_timeseries_signature` generates 30+ calendar features in one call — day, week, quarter, fiscal period, holiday flags — which is real drudgery otherwise. The finance module (`augment_macd`, `augment_regime_detection` with HMM backends, drawdown, Hurst exponent) is wider than you'd expect from a generalist time series library.
Under 1,000 stars for a library that's been around several years is a warning sign about adoption — the R timetk has orders of magnitude more usage, and the Python version hasn't matched it. The Feature Store and GPU acceleration are both clearly marked beta with format-stability warnings, meaning you shouldn't build production pipelines on them yet. The dual-engine abstraction leaks: not every function supports both engines, and the docs don't always make it obvious which operations silently fall back to pandas. No built-in forecasting means you still need to stitch this onto Prophet, statsforecast, or sklearn pipelines yourself — the library gets you to the feature matrix but not across the finish line.