Meta-learning how to forecast time series

Articles
Authors

Thiyanga S Talagala, Rob J Hyndman, George Athanasopoulos

Published

9 February 2023

Publication details

J Forecasting, 42(6), 1476-1501

Links

 

Features of time series are useful in identifying suitable models for forecasting. We present a general framework, labelled FFORMS (Feature-based FORecast Model Selection), which selects forecast models based on features calculated from each time series. The FFORMS framework builds a mapping that relates the features of a time series to the “best” forecast model using a classification algorithm such as a random forest. The framework is evaluated using time series from the M-forecasting competitions and is shown to yield forecasts that are almost as accurate as state-of-the-art methods, but are much faster to compute. We use model-agnostic machine learning interpretability methods to explore the results and to study what types of time series are best suited to each forecasting model.

Associated R package: seer

Shiny app: https://thiyangt.github.io/fformsviz/fforms.html