Forecasting model · Emerging · Advanced

Time series foundation models

Pretrained models that forecast unseen time series zero-shot, without fitting a model per series.

What Time series foundation models is

Trained across millions of diverse series, these models generalise patterns of seasonality and trend well enough to produce reasonable forecasts on new data immediately.

How it works

The series is tokenised into patches and processed by a transformer that predicts future values, optionally conditioned on covariates.

Why it matters

They shorten the path to a baseline forecast at scale, though strong classical and boosted-tree baselines remain competitive on well-understood series.

Common uses

→Forecasting thousands of SKUs
→Cold-start series with little history
→Rapid baselines

Strengths

✓Zero-shot forecasts
✓Handles many series at once

Watch for

✓Not always better than tuned classical models
✓Limited interpretability

Continue exploring

Time Series Forecasting Foundation Model Transformer