scores is a Python package containing mathematical functions for the verification, evaluation and optimisation of forecasts, predictions or models. It supports labelled n-dimensional (multidimensional) data, which is used in many scientific fields and in machine learning. At present, scores primarily supports the geoscience communities; in particular, the meteorological, climatological and oceanographic communities.

Documentation is hosted at scores.readthedocs.io
Source code is hosted at github.com/nci/scores
Journal article: scores: A Python package for verifying and evaluating models and predictions with xarray

Release Notes (What's New)

Version 2.4.0 (January 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

• Added an optional installation variant "fast" which introduces Numba as an optional dependency to support optimised implementations for some metrics. scores.probability.crps_cdf will now automatically switch to an optimised implementation if Numba is installed in the environment. The "fast" variant can be installed with pip install scores[fast] if wanted. See PR #931.

Bug Fixes

• Fixed a bug in threshold-weighed scoring methods that caused the code to fail if the first object in the tuple for interval_where_one was an xr.DataArray and the second was a float, e.g. np.inf. This method has now been corrected to allow a float, int, or xr.DataArray for the interval arguments. See PR #948.

Documentation

• Updated links to the new verification site https://jwgfvr.github.io/forecastverification (which will replace the prior site: https://www.cawcr.gov.au/projects/verification) in docs/included.md, tests/categorical/test_contingency.py and src/scores/continuous/standard_impl.py. See PR #933, PR #934 and PR #935.
• Updated the documentation and citation links for the scoringrules entry in "Related Works". See PR #937.
• Fixed rendering (removed an unintentional block quote), and thereby also resolved a sphinx build error, in the scores.continuous.nse docstring. See PR #936.

Internal Changes

• Sped up (improved the computational efficiency of) the continuous ranked probability score (CRPS) for ensembles, by sorting the ensemble members to compute the CRPS spread term. See PR #928.

Contributors to this Release

Belinda Trotta* (@btrotta-bom), Taylor Mandelbaum* (@aaTman), Tennessee Leeuwenburg (@tennlee), Nicholas Loveday (@nicholasloveday), Stephanie Chong (@Steph-Chong), Robert J. Taggart (@rob-taggart) and Nikeeth Ramanathan (@nikeethr).

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

We also acknowledge the developers of xskillscore and properscoring as we have adapted code from their repositories under a suitable compatible license. This acknowledgment has also been added to NOTICE.md as is best practice. The xarray wrapper function scores.probability.crps_numba.crps_cdf_exact_fast is based on the code for crps_ensemble from xskillscore (https://github.com/xarray-contrib/xskillscore/blob/main/xskillscore/core/probabilistic.py), released under the Apache-2.0 License with copyright attributed to xskillscore developers (as at 11 Dec 2025). The vectorisation of crps_at_point follows the example of _crps_ensemble_gufunc from properscoring (https://github.com/properscoring/properscoring/blob/master/properscoring/_gufuncs.py), released under the Apache-2.0 License with copyright attributed to The Climate Corporation (2015).