Global forecasting models
Many related series are shown. A local approach trains one model per series, so short series have little to learn from and the number of models grows with the number of series. A global approach trains a single model on all the series, which learns their shared weekly pattern and forecasts all of them. Four global approaches are then compared: independent (ForecasterRecursiveMultiSeries, each series forecast from its own past), dependent (ForecasterDirectMultiVariate, all series used to forecast one), foundation models (ForecasterFoundation, pre-trained, no training needed) and deep learning (ForecasterRnn, all series in and several forecast at once).
Global forecasting models
1
Many series
2
Local
3
Global
4
Approaches
skforecast.org
0.3 s