Release Notes (What's New)

Version 2.5.0 (February 14, 2026)

For a list of all changes in this release, see the full changelog. Below are the changes we think users may wish to be aware of.

Features

• scores has introduced support for Python 3.14. See PR #989.
• Added probability integral transform (PIT) classes:
  • PIT for ensembles or cumulative distribution functions (CDFs): scores.probability.Pit
  • PIT for predictive CDFs evaluated at observations: scores.probability.PitFcstAtObs.
    See PR #919.
• Added a new function for generating data for rank histograms:
  • Rank histogram: scores.plotdata.rank_histogram (also available as scores.probability.rank_histogram). See PR #919 and PR #1012.

Deprecations

• Support for include_components will be removed from threshold-weighted continuous ranked probability score (twCRPS) functions in a future version of scores. The scores development team believe using include_components=True may lead to misleading results when used with twCRPS functions. As such, the following are now deprecated:
  • support for include_components in scores.probability.tw_crps_for_ensemble,
  • support for include_components in scores.probability.tail_tw_crps_for_ensemble and
  • support for include_components in scores.probability.interval_tw_crps_for_ensemble.
    See PR #991.

Bug Fixes

• Fixed an IndexError in receiver (relative) operating characteristic (ROC). As such, mutlidimensional input arrays are now supported when using automatic thresholds. See PR #963.
• Fixed a situation where receiver (relative) operating characteristic (ROC) calculations could trigger a NotImplementedError within Dask. scores.probability.roc_curve_data and scores.plotdata.roc have been updated so that if fcst or obs is an xarray object backed by a dask array, and check_args is True, the min and max of the arrays will be calcuated immediately, which triggers computation. This can be avoided by setting check_args=False. See PR #987.

Documentation

• Added two new tutorials:
  • "The Probability Integral Transform (PIT)". See PR #919.
  • "Rank Histogram". See PR #919.
• Updated documentation to say there are now over 75 metrics, statistical techniques and data processing tools contained in scores. See PR #1014.
• Updated "Acknowledging or Citing scores" to include citation details for both our Journal of Open Source Software paper and the Zenodo record for the version of scores being used. See PR #1003.
• Updated the "Contributing Guide" to include additional information about running pre-commit checks. See PR #977 and PR #1010.
• Corrected a function name in an example in the scores.stats.statistical_tests.diebold_mariano docstring. See PR #978.
• Pinned version of Sphinx to prior to version 8 (i.e. sphinx<8), due to a change in symlink handling. This will need to be resolved before scores can migrate to more recent versions of Sphinx. See commit f60ae0c.

Internal Changes

• Fixed Numba warnings in fast Continuous Ranked Probability Score (CRPS) implementation when NaNs are present in input data. See PR #957.
• Set join explicity to "outer" to be compatible with upcoming changes in Xarray. See PR #964.
• Replaced implementations of SciPy's legacy function interpolate.interp1d with a wrapper function. See PR #971.
• Removed the use of NetCDF data on disk from tests. Data is now created on the fly. See PR #966.
• Added Ruff. Ruff is now included in pre-commit and replaces Pylint, Black, Bandit and isort. See PR #967, PR #972, PR #979 and PR #990.
• Replaced mypy with ty. Added ty to pre-commit for type checking. PR #984 and PR #993.
• Updated CI/CD and pre-commit hooks to treat warnings as test failures. See commit a28042d.
• The directory name tests/probabilty/ was spelled incorrectly and has been renamed to tests/probability/. See PR #1021.

Contributors to this Release

Felix Esperson* (@fesperson), Jurian Beunk* (@jurianbeunk), Xiaoxi Wu* (@wuxx66), Robert J. Taggart (@rob-taggart), John Sharples (@John-Sharples), Belinda Trotta (@btrotta-bom), Tennessee Leeuwenburg (@tennlee), Nicholas Loveday (@nicholasloveday), Stephanie Chong (@Steph-Chong), Durga Shrestha (@durgals), Mohammadreza Khanarmuei (@reza-armuei) and Nikeeth Ramanathan (@nikeethr).

* indicates that this release contains their first contribution to scores.