// the find
functime-org/functime
Time-series machine learning at scale. Built with Polars for embarrassingly parallel feature extraction and forecasts on panel data.
functime is a Python library for global forecasting and feature extraction on panel data — meaning many time series at once, not just one. It's built on Polars, so feature extraction that would take minutes in pandas/tsfresh runs in seconds. The target audience is data scientists doing demand forecasting, financial modeling, or any problem with hundreds to thousands of related time series.
The Polars-native design is the real differentiator: feature extraction runs as lazy expressions with query optimization, so you get parallelism and predicate pushdown for free without writing any multiprocessing code. The global forecasting approach (one model across all entities) is architecturally sound for panel data and often beats per-series models, and functime bakes this in rather than making you implement it yourself. The functional API — where a forecaster is a callable that returns predictions — is clean and composes well with the rest of a Polars pipeline. Including M4/M5 benchmark datasets in the repo means you can validate your setup against known results immediately.
Only 63 forks and 1179 stars for a library in this space suggests limited production adoption, which means sparse community knowledge and fewer real-world bug reports surfaced. The LLM forecast analyst feature feels bolted on — using an LLM to 'describe' forecasts is a gimmick that adds an API dependency without adding forecasting quality. Deep learning models (temporal fusion transformer, N-BEATS, etc.) are absent; if your problem needs them, you're back to PyTorch Forecasting or Darts. The Cargo.toml suggests some Rust extension code, but the README says nothing about this, which will confuse anyone who hits a build failure in a restricted environment.