accuracylevel 0.1.0

Initial release. The package implements the accuracy-level evaluation metrics of Agustini, Fithriasari, and Prastyo (2026) doi:10.1016/j.dajour.2025.100661.

Features

• Four accuracy-level metrics on a 0-100% scale: Counted Squared Error (cse()), Counted Absolute Error (cae()), Counted Absolute Percentage Error (cape()), and Symmetric Counted Absolute Percentage Error (scape()), plus accuracy_level() for all four at once.
• Baseline threshold tools: calculate_threshold() and auto_threshold(), with quartiles computed from the inverse empirical CDF (type = 1) to match the paper.
• compare_models() implements the Figure 3 model-selection rule: Level-1 accuracy first, ties broken by the level’s mean error (lower is better) before advancing to the next level. The comparison table reports accuracy and mean error per level.
• Conventional and robust benchmarks: conventional_metrics() (R-squared, RMSE, NRMSE, MAE, MAPE, SMAPE), robust_metrics() (MedAE, trimmed MSE, Huber loss, quantile loss), and compare_all_metrics().
• Framework integration: caret_summary() / caret_summary_extended() / caret_single_metric() for caret; cse_l1(), cae_l1(), cape_l1(), scape_l1(), accuracy_level_metrics(), and al_metric_set() for tidymodels/yardstick; al_forecast_accuracy(), al_compare_forecasts(), al_extended_accuracy(), and al_tsCV() for forecast.
• vignette("replication") reproduces the simple-case (Table 4-6), regression-with-outlier, and time-series results, plus the caret/tidymodels/forecast integrations. The imputation case study is omitted because it relies on confidential firm-level microdata from BPS-Statistics Indonesia that cannot be redistributed.

Data and licensing

The package ships no datasets. The data used in the source article are referenced by link rather than redistributed: the simple-regression and candy-production series are public on Kaggle (the candy series originates from the public-domain FRED series IPG3113N), while the firm turnover microdata are confidential BPS-Statistics Indonesia survey microdata and are not redistributable. Examples and the vignette use small, reproducible simulated data generated inline.

Implementation notes

• Perfect predictions (actual == predicted) give a zero baseline threshold; a machine-epsilon boundary is used so that exact-zero errors are assigned to Level 1.
• Non-finite per-observation errors (for example absolute percentage error when an actual value is zero) are assigned to Level 4.
• When a pre-computed threshold object is supplied to accuracy_level(), all four per-error-type thresholds derive from the stored baseline quartiles, so every model is evaluated against the same baseline (Figure 2 of the paper).