Skip to content
ML Academy

Introduced on day 24

Global forecasting model

One model trained on many related series at once, using lags, calendar features and entity information, instead of one model per series.

In detail

Global models share information across series, keep complexity constant as the number of series grows and can forecast new or short series. They need series to be made comparable (scale by a pre-cutoff level, detrend) and should be benchmarked against local models by segment. Tree ensembles are common choices because they handle calendar and event features and mixed inputs.