functime

Production-ready time series models

functime is a machine learning library for time-series predictions that just works.

• Fully-featured: Powerful and easy-to-use API for forecasting and feature engineering (tsfresh, Catch22).
• Fast: Forecast 100,000 time series in seconds on your laptop
• Efficient: Extract 100s of time-series features in parallel using Polars
• Battle-tested: Algorithms that deliver real business impact and win competitions

Installation

Check out this guide to install functime. Requires Python 3.8+.

Supported Data Schemas

Panel Data

Forecasters, preprocessors, and splitters take a panel dataset where the first two columns represent entity (e.g. commodty name) and time (e.g. date). Subsequent columns represent observed values (e.g. price). The panel DataFrame must be sorted by entity, time.

>>> y_panel
shape: (47_583, 3)

commodity_type   time         price
------------------------------------
Aluminum         1960-01-01    511.47
                 1960-02-01    511.47
                 1960-03-01    511.47
                 1960-04-01    511.47
                 1960-05-01    511.47
...                     ...       ...
Zinc             2022-11-01   2938.92
                 2022-12-01   3129.48
                 2023-01-01   3309.81
                 2023-02-01   3133.84
                 2023-03-01   2967.46

Time Series

Feature extractors support both panel and time-series DataFrames. Time-series Dataframes represents the measurements for a single entity:

>>> y_time_series
shape: (756, 3)

time         price
-------------------
1960-01-01    511.47
1960-02-01    511.47
1960-03-01    511.47
...              ...
2022-11-01   2938.92
2022-12-01   3129.48
2023-01-01   3309.81

Features

Forecasting

Point and probabilistic forecasts using machine learning. Includes utilities to support the full forecasting lifecycle: preprocessing, feature extraction, time-series cross-validation / splitters, backtesting, automated hyperparameter tuning, and scoring.

• Every forecaster supports exogenous features
• Seasonality effects using calendar, Fourier, and holiday features
• Backtesting with expanding window and sliding window splitters
• Automated lags and hyperparameter tuning using FLAML
• Probabilistic forecasts via quantile regression and conformal prediction
• Forecast metrics (e.g. MASE, SMAPE, CRPS) for scoring in parallel
• Supports recursive and direct forecast strategies
• Censored model for zero-inflated forecasts

View the full walkthrough on forecasting with functime.

Feature Extraction

functime has over 100+ time-series feature extractors (e.g. binned_entropy, longest_streak_above_mean) available for any Polars Series. Approximately 85% of the implementations are optimized lazy queries and works on both polars.Series and polars.Expr.

• Over 100+ time-series features
• All features are registered under a custom ts Polars namespace
• ~85% optimized lazy queries and works on both polars.Series and polars.Expr
• 2x-200x speed-ups compared to tsfresh

• 200x speed-ups compared to tsfresh for group by operations

• Supports univariate feature extraction
• Supports feature extraction across many time-series (via group_by)
• Supports feature extraction across windows (via group_by_dynamic)

View the full walkthrough on forecasting with functime.

Preprocessing

View API reference for functime.preprocessing. Preprocessors take in a polars.DataFrame or polars.LazyFrame as input and always returns a polars.LazyFrame. No computation is run until the .collect() method is called on the LazyFrame. This allows Polars to optimize the whole query before execution.

from functime.preprocessing import boxcox, impute

# Use df.pipe to chain operations together
X_new: pl.LazyFrame = (
    X.pipe(boxcox(method="mle"))
    .pipe(detrend(method="linear"))
)
# Call .collect to execute query
X_new: pl.DataFrame = X_new.collect(streaming=True)

View quick examples of time-series preprocessing with functime.