NEWS

---

modelskill 0.1.1 (2026-09-17)

First CRAN release, published on 17 September 2026.

• Replace local README licence links with the full GitHub URL, resolving the invalid file URI reported by CRAN for the first submission.

modelskill 0.1.0

Submission preparation

• Avoid repeated plot rendering in solar and target help examples, and create coverage plots as assigned ggplot objects, to reduce example runtime.
• Accept predictive means and standard deviations (normal distributions), or equally weighted predictive samples, directly in interval and calibration diagnostics. Preserve explicit bounds and quantiles, use shared complete-row filtering and type-7 empirical quantiles, and support empirical PIT values.
• Document these input alternatives and test their numerical equivalence, validation, missing values, and compatibility with standard ggplot additions.
• Replace the verification-derived CRPS decomposition routine with a new implementation from the documented mathematical equations. Test the score and components against independent mathematical references, including ties, outliers, missing values and continuous and discrete predictive samples.
• Correct the build exclusion for locally rendered vignette HTML and update the bundled examples to work with the installed package and current API.
• Explicitly test expected undefined-metric warnings; retain the warnings in the public functions. Add methodological citations to DESCRIPTION.

Scientific-audit corrections

• Include observations tied with ensemble members in CRPS decomposition bins.
• Integrate accuracy-plot areas across identity-line crossings and return missing summaries with an informative warning when coverage is unavailable.
• Use a common complete-case sample across levels in coverage plots and areas.
• Make CCC symmetric for constant inputs: return zero for a positive denominator, and NA for identical constants with a zero denominator.
• Preserve matrix dimensions in normal QCP plots with one retained case.
• Scale extended-metric calculations to avoid intermediate overflow and underflow without changing their mathematical definitions.
• Document the RELI calculation, area interpolation and endpoint assumptions, finite-ensemble PIT limitations, and the descriptive status of median CRPS.

Author decisions

• Preserve the original sample-SD normalization.

• Constant inputs now return NA for r and r2. Constant-model coordinates remain usable; undefined correlations use grey in default colour scales.

• Alexandre M.J.-C. Wadoux is sole author, maintainer and copyright holder; the maintainer email is [email protected]. MIT grant confirmed.

• Author selected the current graphical defaults (A). Retain those conventions, with exterior legends, safer title margins, and reference-label obstacles for model-label repulsion to reduce overlap.

• Add exported diagram_stats() for independent access to diagram statistics.

• Accept numeric matrices and data frames (one model per column) consistently.

• Share pair filtering and numerical calculations across diagrams.

• Add na.rm at the end of each diagram interface; existing positional calls work.

• Reject invalid flags, infinite data, duplicate model names and invalid sizes.

• Preserve partial model names instead of discarding all supplied names.

• Fix missing-observation handling and reject degenerate paired plot subsets.

• Evaluate centred error directly and scale intermediate calculations to avoid cancellation and overflow. Regular complete-data definitions remain unchanged.

• Replace the earlier r = 0 convention for constant inputs with NA and an informative warning. CCC is zero when its denominator is positive and one input is constant, and NA when both inputs are identical constants.

• Keep original palettes, radii, sample SD normalization and axis-title placement.

• Replace repeated constant graphical layers with annotations, update legend syntax, expose Taylor point data, and remove redundant ggthemes dependency.

• Expand analytical, input, geometry, palette and real-device rendering tests.

• Make repelled labels reproducible without changing the user's RNG state.

• Add a three-platform GitHub Actions package-check workflow.

Compatibility: missing/nonfinite colour values, numeric logical flags, duplicate names and invalid dimensions now error. digits is limited to integers 0 through 22. Degenerate metric outputs are explicitly defined above. No existing arguments were renamed or reordered.

Prediction metric audit

• Breaking change: mape() and smape() now return percentages, matching mpe() and rrmse(). Previously, mape() and smape() returned fractions.

• model_metrics() now reports each statistic once. Duplicate ME, MAE, RMSE, r, nse/NSE/MEC, and rhoC columns are removed in favour of bias, mae, rmse, correlation, R2, and ccc. Standalone efficiency aliases remain available.

• Extended output includes KGE. Added independent numerical reference tests and a vignette table of equations, ranges, conventions, and references.