timeseriesflattener

What timeseriesflattener is

timeseriesflattener is a Python package that turns irregular time-series observations into features for machine-learning models. It produces a dataframe with one row per prediction time and constructed feature columns, grouping observations by an ID. Users define prediction times and lookbehind windows for predictors or lookahead windows for outcomes. Documented aggregation choices include counts, earliest and latest values, maximums, means, minimums, slopes, sums, unique counts, and variances. The API also describes temporal predictors, outcomes, boolean outcomes, static features, and time-delta features. A fallback value can be specified when a feature window contains no data. The package accepts pandas or Polars dataframes for prediction-time and static frames, and a tutorial shows feature generation from pre-embedded text represented with IDs, timestamps, and embedding columns. Its introductory tutorial points to complicated, irregular time series in simple-model training as a particularly suitable use. Installation uses pip, and downloadable Jupyter notebook tutorials can be run locally. The package is free and self-hosted; bug reports and feature requests go to GitHub Issues, while usage discussion is directed to GitHub Discussions.