R interface to ForecastBaselines.jl

An R package providing access to 10 baseline forecasting models from the Julia ForecastBaselines.jl library, with uncertainty quantification and seamless integration with the scoringutils package and hubverse ecosystem.

• Baseline and benchmark models for forecast evaluation
• Multiple uncertainty quantification methods (bootstrap, analytical, conformal)
• Seamless integration with scoringutils for proper scoring rules
• Data transformations with automatic back-transformation

1. Julia (>= 1.9): Download from julialang.org
2. R (>= 3.5.0)
3. JuliaCall R package: install.packages("JuliaCall")

# Install from GitHub
remotes::install_github("epiforecasts/forecastbaselines")

After installation, initialize Julia and load ForecastBaselines.jl:

library(forecastbaselines)

# Initialize Julia and install/load ForecastBaselines.jl
setup_ForecastBaselines()

This only needs to be done once per R session.

library(scoringutils) # for scoring

# Your time series data
data <- c(1.2, 2.3, 3.1, 2.8, 3.5, 4.2, 3.9, 4.5, 4.1, 4.8)

# 1. Create and fit a model
model <- ARMAModel(p = 1, q = 1)
fitted <- fit_baseline(data, model)

# 2. Generate forecasts with prediction intervals
fc <- forecast(
  fitted,
  interval_method = EmpiricalInterval(n_trajectories = 1000),
  horizon = 1:5,
  levels = c(0.50, 0.95),
  model_name = "ARMA(1,1)"
)

# 3. Evaluate with true values
truth <- c(5.0, 5.2, 5.4, 5.1, 5.3)
fc_with_truth <- add_truth(fc, truth)

# 4. Score the forecast
fc_point <- as_forecast_point(fc_with_truth)
scores <- score(fc_point)
scores_summary <- summarise_scores(scores, by = "model")

scores_summary[, c("model", "ae_point", "se_point")]
#>        model  ae_point   se_point
#>       <char>     <num>      <num>
#> 1: ARMA(1,1) 0.2368795 0.07611188

• ConstantModel: Naive forecast using last observed value
• MarginalModel: Forecasts based on empirical marginal distribution
• KDEModel: Kernel density estimation

• LSDModel: Last Similar Dates method (seasonal patterns)
• OLSModel: Ordinary least squares with polynomial trends
• IDSModel: Increase-Decrease-Stable trend detection
• STLModel: Seasonal-Trend decomposition using Loess

• ARMAModel: Autoregressive Moving Average
• INARCHModel: Integer-valued ARCH for count data
• ETSModel: Error-Trend-Season exponential smoothing (all 30 variants)

For detailed guides and examples:

• vignette("forecastbaselines") - Complete introduction with examples
• vignette("forecast-models") - Detailed guide to all 10 models
• vignette("transformations") - Working with data transformations

Some of the models in {forecastbaselines} are also available through other R packages. For time series forecasting in R, consider:

• fable - A complete forecasting framework in the tidyverts ecosystem with ARIMA, ETS, and many other models
• forecast - The classic R forecasting package with auto.arima, ets, and many other methods
• prophet - Facebook’s forecasting package for time series with strong seasonal patterns
• modeltime - A tidymodels framework for time series forecasting

{forecastbaselines} focuses specifically on baseline models for forecast evaluation and integrates with the scoringutils/hubverse ecosystem.

If you use this package in your research, please cite the software and the associated preprint:

Software:

@software{forecastbaselinesr,
  title = {forecastbaselines: R Interface to ForecastBaselines.jl},
  author = {Stapper, Manuel and Funk, Sebastian},
  year = {2025},
  url = {https://github.com/epiforecasts/forecastbaselines}
}

Preprint:

@article{stapper2025baseline,
  title = {Mind the Baseline: The Hidden Impact of Reference Model Selection on Forecast Assessment},
  author = {Stapper, Manuel and Funk, Sebastian},
  year = {2025},
  doi = {10.1101/2025.08.01.25332807},
  url = {https://doi.org/10.1101/2025.08.01.25332807}
}

MIT License - see LICENSE file for details

For bugs and feature requests:

• R package issues: GitHub Issues
• Julia package issues: ForecastBaselines.jl Issues

All contributions to this project are gratefully acknowledged using the allcontributors package following the allcontributors specification. Contributions of any kind are welcome!

sbfnk