Chronos is a family of time-series foundation models from Amazon, described in “Chronos: Learning the Language of Time Series,” posted to arXiv in March 2024. Its central idea is to treat forecasting as a language problem so that off-the-shelf language model architectures can be used almost unchanged.
To do this, Chronos converts a numeric time series into a sequence of discrete tokens by scaling the values and then quantizing them into a fixed vocabulary, much as words are tokens for a text model. It then trains standard transformer language models, ranging from 20 million to 710 million parameters, to predict the next token, which corresponds to the next value in the series. Training used a large collection of public datasets augmented with synthetic series generated from Gaussian processes. Evaluated across 42 datasets, Chronos performed strongly on data from its training domains and remained competitive in zero-shot forecasting on series it had never seen.
Chronos showed that the language-model toolchain, including its training tricks and scaling behavior, transfers surprisingly well to numeric forecasting.
Why business readers should care: Chronos lets organizations reuse the mature, well-understood machinery of language models for demand and metric forecasting, often without any per-dataset training.