Introduced on day 27
Time-series foundation model
A model pretrained on a large corpus of many series (for example Chronos, TimesFM, Moirai, TimeGPT) and used to forecast new series with little or no training on them.
In detail
Families differ in architecture (tokenized values with a language-model backbone, patched decoders, masked encoders), in covariate support, in whether weights are released and under which licence, and in how their benchmark results were produced. Treat each version as a candidate to test, since vendor ranks are self-reported and leaderboards move.