Foundation models
A foundation model is pre-trained on a large collection of time series from many domains (seasonal, trending, intermittent, with level changes...), and its weights are then frozen. Three new stores, which the model has never seen, have different patterns: a weekly cycle, a growing trend and a peak every 3 days. Their recent values are given to the model as context (in skforecast, fit() only stores this context). The model forecasts the three stores in one call without any training (zero-shot), and each forecast continues the pattern of its own context: in-context learning.
Foundation models
Foundation models are pre-trained on large collections of series from many domains.
1
Pre-training
2
Context
3
Zero-shot forecast
energy
skforecast.org
0.3 s