Abstract
Pretrained time series models have enabled inference-only forecasting systemsthat produce accurate predictions without task-specific training. However,existing approaches largely focus on univariate forecasting, limiting theirapplicability in real-world scenarios where multivariate data and covariatesplay a crucial role. We present Chronos-2, a pretrained model capable ofhandling univariate, multivariate, and covariate-informed forecasting tasks ina zero-shot manner. Chronos-2 employs a group attention mechanism thatfacilitates in-context learning (ICL) through efficient information sharingacross multiple time series within a group, which may represent sets of relatedseries, variates of a multivariate series, or targets and covariates in aforecasting task. These general capabilities are achieved through training onsynthetic datasets that impose diverse multivariate structures on univariateseries. Chronos-2 delivers state-of-the-art performance across threecomprehensive benchmarks: fev-bench, GIFT-Eval, and Chronos Benchmark II. Onfev-bench, which emphasizes multivariate and covariate-informed forecasting,Chronos-2's universal ICL capabilities lead to substantial improvements overexisting models. On tasks involving covariates, it consistently outperformsbaselines by a wide margin. Case studies in the energy and retail domainsfurther highlight its practical advantages. The in-context learningcapabilities of Chronos-2 establish it as a general-purpose forecasting modelthat can be used "as is" in real-world forecasting pipelines.