Since Rob Hyndman & Stephan Kolassa wrote their Foresight article in 2010 on “Free Open-Source Forecasting Using R” much has happened. The forecast package for the R statistical language (Hyndman & Khandakar, 2008), abbreviated to “R Forecast package” in the following, was the main focus of the article then. Now, it is the reference implementation of many classical forecasting methods such as exponential smoothing and ARIMA. But many more R packages of high quality have appeared, such as Bayesian Structural Time Series (BSTS, Scott & Varian, 2014), a new package called Fable from Rob Hyndman and Mitchell O’Hara is setting out to replace the R Forecast package and other high-quality packages such as hts, tsintermittent, thief, smooth and tsutils are available. Outside of the R ecosystem, notable new packages include Prophet, Tensorflow STS, and Gluon Time Series.
An important difference to 2010 is that open-source seems to have reached the third stage of Arthur Schopenhauer’s truth-finding process:
“All truth passes through three stages. First, it is ridiculed. Second, it is violently opposed. Third, it is accepted as being self-evident.”
Much of the skepticism towards open-source has been overcome to the point that closed-source software is met with skepticism. This is particularly true in scientific circles, especially in the data science realm, and for the generation of students graduating from university.
Some things have stayed constant over the years however. Open-source forecasting software concerns very much the logic tier of Forecasting Support Systems (Petropoulos, 2015), to a limited degree the presentation layer and hardly ever the database layer. These layers are crucial for operating forecasting systems in practice.
In this article, we concentrate on the packages that are based on the Python programming language.
Research areas
