Time Series Forecasting: Global Forecasting Model

published on 10 December 2021

The world of time series forecasting is currently undergoing a profound paradigm shift, moving from the traditional statistical approach to implementing what have been called Global Forecasting Models (GFM) [1], which, based on Machine Learning, fit a single model for the entire universe of series to be predicted.

When handling large volumes of data, the generalizability of GFMs clearly contrasts with the difficulty of scaling of local or statistical models, which fit an independent model for each series.

Thus, GFMs allow reducing the number of models required, reducing the need for human supervision, as well as reducing costs associated with putting into production, monitoring and maintenance. Along the same lines, the ability to generalize patterns in the data allows these models to predict series that were not present at the time of training.

Additionally, the GFM won all the last forecasting competitions carried out, such as M4 and M5 [2], and the others carried out in Kaggle [3], which has led to a great boom in the generation of new algorithms and libraries .

However, when the scale allows it and the gains derived from generalization are lower, the classical models will still often be the correct alternative, prevailing in terms of precision, interpretability and simplicity.

Therefore, the greatest contribution of those who are working with time series lies in the ability to understand the advantages and disadvantages of each of the available alternatives, and be able to select the one that fits the problem in question.

[1] Hansika Hewamalagea et al., Global Models for Time Series Forecasting: A Simulation Study.

[2] Makridakis et al., 2020a.

[3] Bojer and Meldgaard, 2020.