Automatic Time Series Forecasting: The forecast Package for R

This paper describes two automatic forecasting algorithms implemented in the forecast package for R, addressing the common need in business and other contexts to produce automatic forecasts for large numbers of univariate time series. The first algorithm is based on innovations state space models that underlie exponential smoothing methods, providing a principled framework for automatically choosing among exponential smoothing variants. The second is a step-wise algorithm for automatically identifying and fitting ARIMA models.

Both algorithms are applicable to seasonal as well as non-seasonal data, and the paper compares and illustrates them using four real time series, while also briefly describing other functionality available in the forecast package. By making robust, fully automatic forecasting methods accessible in a widely used open-source R package, the work became a standard tool for large-scale and reproducible time series forecasting.