Time-Series Foundation Model

A time-series foundation model (TSFM) is a transformer-based model pre-trained on a large corpus of time-series data, designed to perform forecasting on unseen time series in a zero-shot manner without task-specific fine-tuning. The paradigm emerged in 2023-2024 with the release of models like TimesFM (Google, 2024), Chronos (Amazon, 2024), and TimeGPT-1 (Nixtla, 2024), and has since become an active research area with major releases including TimesFM-2.5, TimesFM-3, Chronos-2, and others.

The training paradigm combines three ingredients: (1) a large pre-training corpus drawn from real-world and synthetic time-series sources (typically tens of millions to over a trillion time points); (2) a transformer architecture adapted for sequence modeling, often with patching (grouping contiguous time steps into patches of fixed size) and per-series normalization; and (3) a self-supervised training objective such as masked prediction or next-token prediction over the time-series tokens.

Time-series foundation models are evaluated on three families of public benchmarks as of 2026: Gift-Eval, FEV-Bench, and Time. Recent work has expanded the paradigm from strictly univariate forecasting (predicting future values of a single series) to multivariate forecasting with covariates — past covariates (features known historically), future covariates (known future events like promotions or holidays), and dynamic covariates that mix both. The 2026 generation of TSFMs (TimesFM-3, Chronos-2) is natively multivariate and produces both point forecasts and probabilistic forecasts (typically 9 quantiles).

Time-series foundation models are adjacent to foundation-model (the general pre-trained-transformer paradigm), zero-shot-capability (the broader transfer-learning property), and spacetime-patches (the tokenization strategy used for sequence modeling). They are distinct from statistical forecasting models (ARIMA, ETS) and from task-specific deep-learning forecasters (DeepAR, N-BEATS) in being pre-trained across domains rather than fit per series or per task.