Skip to main content
TabPFN extends beyond tabular and regression tasks - it can perform zero-shot time series forecasting by reframing forecasting as a tabular regression problem.
Our method, TabPFN-TS, combines the TabPFN model with lightweight feature engineering to deliver both point and probabilistic forecasts instantly, without any retraining.

​
Key Capabilities

  • Zero-shot forecasting - Predicts future values immediately with no fine-tuning or pretraining on time-series data.
  • Point and probabilistic outputs - Produces accurate mean, median, and quantile forecasts directly from TabPFN’s predictive distribution.
  • Supports exogenous variables - Easily integrates external regressors or covariates as additional table columns.
  • Lightweight model - Achieves state-of-the-art results with just 11M parameters (40× smaller than some foundation models).
  • Fast, no-GPU inference - Works locally or via the tabpfn-client, requiring no dedicated hardware.

Google Colab Example

Check out our Google Colab for a detailed tutorial on using TabPFN for time-series forecasting.
⌘I