Changelog

All notable changes to tseda are documented here. This project follows Semantic Versioning.

0.1.3 (2026-06-22)

Fix incorrect PyPI install command in README and STL docstring (pip install tseda → pip install timeseries-eda).

0.1.2 (2026-06-22)

Bug fixes.

Correctness

  • core.TimeSeries._compute_is_regular — replaced DatetimeIndex.astype(np.int64) with unit-agnostic to_numpy() timedelta comparison; fixes TypeError on pandas 2.x with non-nanosecond resolution indexes.

  • forecastability.scorer._has_large_gaps — same timedelta fix.

  • statistics.StationarityTester (native ADF fallback) — rewrote regressor construction in _adf_native to use a fixed sample size (maxlag+1 start) for all lag models, making AIC comparisons valid and eliminating ValueError crashes for lag ≥ 1.

  • anomaly.AnomalyDetector.gesd — clamps max_outliers to n_finite // 2 − 1 before delegating to OutlierDetector; prevents crash on series with fewer than 22 observations.

  • changepoint.ChangepointDetector.cusum — replaced post-hoc reset logic with a single incremental accumulator so consecutive changepoints are detected correctly; scores now reflect the reset-adjusted CUSUM values rather than the pre-reset arrays.

  • changepoint.ChangepointDetector.variance_ratio — F-test degrees of freedom now assigned to the window with the larger variance (numerator), fixing an incorrect two-sided p-value.

  • quality.MissingValueAnalyzer.interpolate (method="linear") — leading and trailing NaN are now filled with the nearest boundary value when no limit is set, matching the documented “fill NaN values” contract.

Performance

  • anomaly.AnomalyDetector.rolling_iqr and rolling_z — replaced O(n) Python loops (series.iloc[i] per observation) with vectorized NumPy array operations.

0.1.1 (2026-06-21)

Documentation and CI improvements (no API changes).

0.1.0 (2026-06-21)

Initial release.

Modules