time series foundation model

A large neural network pretrained on a broad corpus of time series from many domains (retail demand, energy, traffic, weather, finance, web metrics), often supplemented with synthetic series, so that it can forecast unseen series zero-shot or after light fine-tuningfine-tuning, n.: Taking an AI model that has already learned a lot and training it a little more on a smaller, focused set of examples so it gets better at one job or behaves a certain way.. Most designs adapt the transformertransformer, n.: The design behind today's chatbots and much of modern AI. It handles text by letting every word look at every other word and decide which of them matter for understanding it.: each series is normalized and split into patches or quantized into tokens, and the model predicts future values or a distribution over them, either autoregressively with a decoder-only network (as in TimesFM and Lag-Llama) or with encoder-based or encoder–decoder designs (as in Moirai, and in Chronos, which reuses a language-model architecture over discretized values). The appeal is fast deployment without per-series training and strong accuracy on short or brand-new series. Open questions include the handling of external covariates and long horizons, robustness to domain shift, and whether gains over well-tuned statistical and gradient boostinggradient boosting, n.: A method that builds a strong predictor by adding small decision trees one at a time, each trained to correct the errors that remain after all the trees before it. baselines are consistent, since overlap between pretraining corpora and benchmark series (data contaminationdata contamination, n.: When the test questions used to grade an AI system were already in its training data, so a high score may reflect memorized answers rather than real ability.) can inflate reported results.

The foundation model gave usable zero-shot forecasts for 300 newly opened stores before we had any training data.