Google Research has introduced TimesFM-3, a foundation model aimed at forecasting multiple time series at once. The model is described as zero-shot, which means it is intended to handle new forecasting tasks without requiring additional fine-tuning on each dataset. The announcement appears in Google Research's Data Management section.
Foundation models of this kind are notable because they can be reused across datasets rather than trained from scratch per use case. By focusing on multivariate forecasting, TimesFM-3 is aimed at settings where several related series need to be predicted together.
Since the source is a single announcement, there are no independent or conflicting assessments to compare. The claims here reflect Google Research's own description of the model.