A foundation model for time series is a single large model, pre-trained once on an enormous and varied collection of time series, that can then forecast a brand-new series it has never seen without any additional training. The concept transplants to forecasting the recipe that reshaped language and vision: train one big general model on broad data, then reuse it everywhere.
This breaks with how forecasting worked for decades. Traditionally each dataset, each product, each metric, got its own model that had to be fitted before it could predict. A foundation model instead does zero-shot forecasting: the user supplies the recent history of a series and immediately gets a forecast, drawing on patterns the model absorbed during pre-training across millions of other series. Several such models appeared in close succession in 2023 and 2024, including Nixtla’s TimeGPT, which its authors billed as the first foundation model for time series, Google’s TimesFM, Amazon’s Chronos, and Salesforce’s Moirai. They differ in architecture and training data but share the pre-train-once, forecast-anything philosophy, and several are offered as simple APIs.
Whether these models reliably beat well-tuned classical methods on every problem is still being debated, but they have clearly changed how the field approaches forecasting.
Why business readers should care: foundation models promise good forecasts for any series on demand, removing the cost of building and maintaining a separate model for every product and metric.