General-purpose forecasting models pre-trained on enormous collections of time series, which can then forecast a new series without being trained on it. The claim is that a model trained on millions of series learns patterns that transfer, and the results are competitive with, though not always better than, well-tuned classical methods.
Technical view
Trained on large heterogeneous corpora of series with tokenisation schemes over patches or quantised values, offering zero-shot forecasting with optional fine-tuning. Evaluated on standard forecasting benchmarks against statistical baselines and gradient boosting, where results are strong but not uniformly dominant. Advantages are cold-start forecasting for new series and reduced per-series engineering; open questions include covariate handling and calibrated prediction intervals.
Related
Part of AI Terminology Index, section: Modalities and Applications.