Robust Accuracy-Level Metrics for Predictive Model Evaluation.

accuracylevel

Overview

The accuracylevel package implements novel accuracy-level metrics for evaluating continuous data prediction models, as described in Agustini, Fithriasari, and Prastyo (2026).

This is the initial release (version 0.1.0).

Key Features

• Robust: Gradual degradation under outlier presence, unlike conventional metrics
• Consistent: Same optimal model selection across all four metrics
• Objective: Multi-level assessment reveals distribution of prediction quality
• Comparable: Standardized 0-100% scale across all metrics and models

The Four Metrics

Each metric assigns observations to four accuracy levels:

• L1: error < T (highest accuracy)
• L2: T <= error < 2T
• L3: 2T <= error < 5T
• L4: error >= 5T (lowest accuracy)

where T is a threshold derived from the baseline model's error distribution.

Installation

The package is not yet on CRAN. Install the development version from GitHub:

# install.packages("devtools")
devtools::install_github("madsyair/accuracylevel")

Quick Start

library(accuracylevel)

# Sample data
actual <- c(10, 20, 30, 40, 50, 60, 70, 80, 90, 100)
predicted <- c(11, 19, 32, 38, 51, 58, 72, 78, 92, 98)

# Calculate all accuracy-level metrics
result <- accuracy_level(actual, predicted)
print(result)

# Get individual metrics
cse(actual, predicted, level = 1)  # CSE Level 1
cae(actual, predicted, level = 1)  # CAE Level 1

# Compare with conventional metrics
compare_all_metrics(actual, predicted)

Integration with Popular Frameworks

caret

library(caret)

ctrl <- trainControl(method = "cv", number = 5,
                     summaryFunction = caret_summary())

model <- train(y ~ ., data = training_data, method = "lm",
               trControl = ctrl, metric = "CAE_L1", maximize = TRUE)

tidymodels

library(tidymodels)

# Single metric
predictions |> cae_l1(truth = outcome, estimate = .pred)

# Or use the metric set
al_metrics <- al_metric_set()
predictions |> al_metrics(truth = outcome, estimate = .pred)

forecast

library(forecast)

fit <- auto.arima(train_ts)
fc  <- forecast(fit, h = 24)
al_forecast_accuracy(fc, test_ts)

Model Comparison

m1 <- list(actual = y, predicted = pred_a)
m2 <- list(actual = y, predicted = pred_b)

result <- compare_models(ModelA = m1, ModelB = m2, metric = "cae")
result$optimal_model
result$comparison

Data Used in the Article

This package does not bundle any datasets. The data used in the source article are obtained from their original providers:

• Simple linear regression and candy production time series: publicly available on Kaggle (<https://www.kaggle.com/code/goldens/candy-production-time-series-analysis). The candy production series originates from the U.S. Federal Reserve (FRED series IPG3113N), which is in the public domain.
• Firm turnover / imputation data: firm-level microdata from BPS-Statistics Indonesia. This is confidential survey microdata and is not redistributable; it must be requested directly from BPS-Statistics Indonesia. For this reason no imputation dataset is shipped with the package, and the package vignette does not reproduce the imputation case study.

The package vignette reproduces the article's results using small, reproducible simulated data generated inline (see vignette("replication", package = "accuracylevel")).

Why Accuracy-Level Metrics?

Citation

If you use this package in your research, please cite:

Agustini, M., Fithriasari, K., & Prastyo, D.D. (2026). An accuracy-level
method for robust evaluation in predictive analytics. Decision Analytics
Journal, 18, 100661. https://doi.org/10.1016/j.dajour.2025.100661

License

GPL-3 (c) The accuracylevel authors.