This slides presents the development and capabilities of Chronos-2, an encoder-only transformer model designed to overcome the limitations of previous univariate time series foundation models.
"Chronos-2: From Univariate to Universal Forecasting" outlines the development and capabilities of Chronos-2, an encoder-only transformer model designed to overcome the limitations of previous univariate time series foundation models.
By introducing a novel architecture that alternates between time and group attention mechanisms, Chronos-2 enables zero-shot in-context learning for univariate, multivariate, and covariate-informed forecasting tasks without requiring architectural changes. The slides detail the model's tokenization process, its reliance on synthetic data generated via "multivariatizers" for two-stage training, and its state-of-the-art performance across comprehensive benchmarks like fev-bench and GIFT-Eval