Added support for python 3.12

This minor release adds another parametric form of the CRPS and a new type of scoring rule: the interval-based Winkler score. Additionally, we start to widen the scope of the library by introducing the first graphical verification tool, the reliability diagram. In the future, we plan to add other graphical tools to diagnose calibration as they are important to complement scoring rules.

New features

• Closed-form CRPS for the exponential distribution in #27 by @sallen12 and @frazane
• CORP reliability diagram in #23 by @frazane
• Interval/Winkler Score in #24 by @simon-hirsch

New features

• Added quantile-based CRPS in #22 by @simon-hirsch

This release brings an important breaking change, improvements to the documentation and some preliminary work to include many new distributions for the analytical continuous ranked probability score.

Breaking changes 💥

• Changed the order of positional arguments for all functions in #18 by @frazane. Starting with this release, observations always come as the first positional argument.

New features

• Added special module and several statistical functions to the backends in #17 by @sallen12 and @frazane. This is preparatory work for #13.

Documentation

• General improvements in #12 by @frazane

What's changed

• Added Error Spread Score by @frazane in #10

This release brings a new framework-agnostic backend system to scoringrules, with added support for pytorch! This required a major internal refactoring to the codebase but there are no significant changes to the API except for some weighted scoring rules.

This release introduces weighted scores #4 by @sallen12 (👏 ). The following weighting schemes have been implemented:

• threshold weighting
• outcome weighting
• vertical rescaling

for the ensemble-based metrics: CRPS, Energy Score and Variogram Score.