Replicating and Extending the M4 Study with sktime

This is a companion repository for our paper:

• Markus Löning, Franz Király: “Forecasting with sktime: Designing sktime's New Forecasting API and Applying It to Replicate and Extend the M4 Study”, 2020; arXiv:2005.08067

The repository contains the code for reproducing our results. Our code is heavily based on sktime.

• Makefile contains convenience commands for running the replication and extension from the command line,
• scripts/ contains our code for replicating and extending the M4 study,
• results/replicated/ contains our obtained predictive performance results,
• requirements.txt contains the list of required packages to run our scripts.

To reproduce our results, we recommend to download the original repository of the M4 study, particularly we use

• the M4 data set,
• the individual point forecast results for comparison,
• the summary results file to cross-check results.

In addition, you may have to adjust paths for the input and output directories in scripts/config.py.