Skip to contents

Changelog

Source: NEWS.md

alloscore2 0.1.0

  • Initial package setup: repository layout, GitHub Actions workflows for R CMD check, lint, format, test coverage and pkgdown, and the Air and lintr configuration.

  • Ported the allocation and scoring machinery from the original alloscore: allocate(), alloscore() and its methods, post_process(), the oracle functions, the gpl loss family, add_pdqr_funs(), slim(), the stdize_*_params() helpers and the zxh_tab2 and zxh_tab3 datasets.

  • Ported plot_components(), plot_components_slim(), plot_scores_slim() and plot_iterations().

  • Added as_alloscore_df(), allocate_model_out() and alloscore_model_out(), which take a hubverse model_out_tbl and oracle_output directly. Predictive quantiles are turned into cdfs and quantile functions with distfromq, so the conversion that callers used to hand-roll is no longer needed. Only the quantile output type is supported.

  • Added a README with runnable examples and two vignettes: an overview of scoring hubverse forecasts, and the validation against published newsvendor optima.

  • Results agree with the original package to within root-finding tolerance; see tests/testthat/test-legacy-equivalence.R.

Bug fixes relative to the original package

  • The O/U over/under-cost parameterization now works. gpl_loss_fun() and exp_gpl_loss_fun() previously errored on it and dexp_gpl_loss() silently returned NULL.

  • A supplied dg is now honored instead of being silently discarded and recomputed from g.

  • exp_gpl_loss_fun() now applies its offset argument, which was accepted and ignored.

  • new_gpl_df() can infer N from its arguments; the code path referred to an undefined variable.

  • alloscore() on a slim object now honors against_oracle, which was accepted and ignored.

  • allocate() returns a well-formed allocated object when no target has any marginal benefit. It previously returned a data frame with no xdf column, no class and no attributes, which nothing downstream could consume.

  • allocate() now fails with a clear message when g has an unbounded derivative at zero, as g = "log(x)" does. It previously emitted messages and returned its initial iterate unchanged. Use a shifted increment function such as g = "log(1 + x)".

  • Every dplyr and tidyr call is namespaced and declared. The original package only worked with the tidyverse attached.

  • weights.allocated() and alloscore.slim() now match the signatures of their generics.

  • plot_components() and plot_scores_slim() work. They previously called the nonexistent tidyr::any_of() and referred to an undefined variable respectively.

  • plot_iterations() no longer calls require() at runtime; patchwork is a declared dependency. It also accepts a single row of allocate_model_out() output, not only an allocated object.

  • Every ggplot2, forcats and scales call is namespaced and declared.