v0.1.1

Enhancements:

• Support NaN and Null in compute_bias (PR #43)

Bug Fixes:

• Always output column "bias_weights" in compute_bias (PR #44)

v0.1.0

Some highlights:

• Confidence intervals for plot_reliability_diagram via arguments n_bootstrap and confidence_level (PR #32).
• New option diagram_type = "bias" for plot_reliability_diagram, which is roughly a 45 degree rotated plot (PR #35).
• Better visualisation of uncertainty/standard errors in plot_bias and distinction between numerical and categorical features (PR #37).
• Consistently sorted output, i.e. the different (model) columns of y_pred (PR #37).
• Number of effective (output) bins is now always at most n_bins in compute_bias and plot_bias (PR #37).
• Switch to ruff (PR #34)

v0.0.3

A new module scoring containing:

• Add strictly consistent, homogeneous scoring functions
  • HomogeneousExpectileScore for mean an expectiles
  • HomogeneousQuantileScore for quantiles
  • SquaredError, PoissonDeviance, GammaDeviance and PinballLoss for convenience
• Add LogLoss
• Add score decomposition decompose 🚀
  To my knowledge, this is the first time the score decomposition into miscalibration, discrimination (or resolution) is available in Python. R users can use the wonderful reliabilitydiag package of @aijordan for quite some time now.

v0.0.2

• Added support for case weights.
• Use of the fantastic https://www.pola.rs/ library (instead of pyarrow and pandas).

v0.0.1

First public release