T0: A Time-Series Foundation Model for Forecasting with Context (arxiv.org)

Researchers have introduced $t_0$, a groundbreaking family of open-weight foundation models specifically designed for multivariate time-series forecasting. The initial models, $\texttt{t0-alpha}$ and $\texttt{t0-beta}$, feature 102 million and 256 million parameters, respectively, and are capable of generating probabilistic forecasts without the need for task-specific retraining. By conditioning their predictions on historical target data, past covariates, and known future covariates, these models effectively leverage transformer architectures that integrate attention mechanisms across both time and variable dimensions. The significance of $t_0$ lies in its advanced capability to enhance forecasting accuracy across various tasks. Notable achievements include $\texttt{t0-alpha}$ and $\texttt{t0-beta}$ demonstrating competitive performance on the GIFT-Eval benchmark with aggregate CRPS scores of 0.4941 and 0.4738, respectively. These models also show marked improvements when incorporating known future covariates, increasing forecasting skill by 6.3 percentage points. With promising results on benchmarks like the Victoria electricity-demand prediction and ERCOT price forecasting, $t_0$ stands to influence the AI/ML community by providing more robust, scalable, and accurate forecasting solutions tailored for complex datasets.