Changelog • MetricsWeighted

Improvements

logLoss() and deviance_poisson() are now less picky about predictions of 0 (or 1).

Documentation

Fix Latex problem in MacOS help files.
Slight improvements in docu.

Documentation

Much more compact help files
Using Latex in help files

This release bumps the package to stable version 1.0.0.

Maintenance

Reorganization of code
Switch to package::function() notation

Maintenance

Use github workflows
Add github pages
Better README

Dependencies

Removed {dplyr} from suggested packages.

This is a maintainance update without any code change.

Fixed problematic unit tests.
Internal restructuring how package content is being generated.

Added unit tests
Added option reference_mean to r_squared() functions. This allows clean out-of-sample applications.
Added Murphy diagrams

Maintainance release. rmarkdown is now an explicit “Suggested” package.

New function

weighted_cor to calculate weighted correlation between actual and observed values.
prop_within to calculate weighted proportion of predictions within a tolerance around actual values.

Requirements

Reduced minimal R version from 3.5 to 3.1.

Documentation

Clarified that R-squared is calculated with respect to the null model calculated from the actual values.

New functions

Elementary scoring functions for expectiles and quantiles.
multi_metric: A way to create a named list of performance measures parametrized by a parameter.

New functions

Added the following convenience wrappers to r_squared.

r_squared_poisson
r_squared_gamma
r_squared_bernoulli

New function

Added function weighted_var to calculate variance weighted by sampling weights.

Improvement of documentation and examples.
Better handling of Tweedie special cases.
More strict error handling.
Added median absolute error (and weighted_median, weighted_quantile)

Initial release.