TimeGPT is a forecasting model introduced by Azul Garza, Cristian Challu, and Max Mergenthaler-Canseco of Nixtla in the paper “TimeGPT-1,” posted to arXiv in October 2023. Its authors describe it as the first foundation model for time series, applying to forecasting the same recipe that made large language models successful.
The defining feature is zero-shot forecasting. A traditional forecasting model must be fitted to each dataset before it can predict. TimeGPT, by contrast, is pre-trained once on an enormous and varied collection of time series and can then produce forecasts for a brand-new series it has never seen, with no training step from the user. The paper reports that this pre-trained model outperforms established statistical and machine learning methods in zero-shot inference. Nixtla offers it as a hosted API, so a user can send a series and receive a forecast in seconds.
TimeGPT helped kick off the foundation-model era of forecasting, in which a single large pre-trained model replaces the per-dataset modeling that defined the field for decades.
Why business readers should care: TimeGPT promises usable forecasts without the cost and delay of building and training a custom model for every product, store, or metric.