Tar by ByteDance

Tar by ByteDance is an open-source toolkit designed for efficient, scalable time series forecasting and analysis. Built around advanced Transformer-based architectures, it provides a unified framework for modeling complex temporal data across domains such as finance, energy, operations, and sensor analytics. The library supports both univariate and multivariate time series, enabling users to capture long-range dependencies, seasonality, and irregular patterns that traditional models often miss.

Key capabilities include modular model components, customizable encoder–decoder structures, and support for various attention mechanisms tailored to time series characteristics. Tar integrates data preprocessing utilities, including normalization, windowing, and feature engineering, as well as standardized training, validation, and evaluation pipelines. It offers ready-to-use baseline models and configuration templates, allowing practitioners to quickly benchmark and iterate on forecasting solutions.