Google Research released TimesFM-3, a forecasting model for time-based data. It predicts several related series at once, without task-specific training. It can also use features known only from the past, such as past foot traffic. It can use events already scheduled ahead, such as promotions, holidays or weather forecasts. Earlier versions forecast one series at a time and built the forecast step by step. TimesFM-3 produces the whole forecast horizon in one pass. Google reports top rank on three public forecasting leaderboards, against models including Chronos-2 and Toto 2.0. The model is on GitHub and Hugging Face now. A BigQuery integration is due in the coming weeks. Until then, the earlier TimesFM-2.5 can be used there for single-series tasks.
What changed
Earlier TimesFM models forecast only from a single series' own history.
What it unlocks
Forecasting several related series together, using known future events like promotions or holidays, without training on the task.
- 330 million parameters
- 1 trillion+ time points in training
- 32 time steps per patch
- 9 quantiles predicted per step
What you need to act on it
- download from GitHub or Hugging Face
- BigQuery integration not yet available
Sources