Chronos-2: From univariate to universal forecasting (www.amazon.science)

Amazon announced Chronos-2, a time-series foundation model that extends prior univariate TSFMs to “universal” forecasting: it can zero-shot handle arbitrary numbers of target series (multivariate) and incorporate covariates (past-only, known-future, and categorical) using in-context learning. That means practitioners can jointly forecast coevolving signals (e.g., CPU, memory, I/O) or fold in scheduled promotions and weather without retraining. Chronos-2 also leverages cross-learning across series to improve univariate forecasts and help cold-start cases by borrowing patterns from related series. Technically, Chronos-2 pairs an architecture innovation—group attention that lets information flow within arbitrary-sized groups of series and between covariates—with a synthetic pretraining strategy that injects multivariate structure into base univariate generators to create the diverse training corpus needed for a universal model. Empirically it leads the field: Chronos-2 tops GIFT-Eval, outperforms existing TSFMs on fev-bench (biggest gains on covariate-informed tasks), and beats Chronos-Bolt in over 90% of head-to-head comparisons. The model is open source, positioned as a production-ready, general-purpose forecasting backbone that simplifies pipelines by delivering strong zero-shot multivariate and covariate-aware forecasts.