Time series foundation model
Forecasts future values such as sales, traffic or demand from past numbers, often with no training on your data.
What it is
Time series foundation models are pre-trained on large collections of real and synthetic series, so they can forecast a new series zero-shot. They give strong baselines for demand planning, capacity and monitoring without building a custom model.
How it works
Past values are split into patches or quantized into tokens, much like text, and a transformer predicts the next values. Many output a range of likely values rather than a single guess, which helps with planning under uncertainty.
What it is good for
- Sales and demand forecasting
- Capacity and energy planning
- Spotting anomalies in metrics
Limits to know
- Cannot foresee events not reflected in the data
- Long horizons get uncertain fast
- Extra signals such as promotions need special handling
