TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting Accepted at NeurIPS 2026

·2026.10.19 01:43

Key point

The synthetic data generators in TempoPFN have already been used in Toto-2 and TS-ICL.

TempoPFN is a library of synthetic time series generators and augmentations designed to train time series foundation models, specifically linear RNNs, for zero-shot forecasting.
The synthetic data approach has already been utilized in existing models Toto-2 and TS-ICL.
The paper was accepted at NeurIPS 2026.

Authors include Vladyslav Moroshan, Julien Siems, Arber Zela, Timur Carstensen, and Frank Hutter.
Code is available on GitHub under automl/TempoPFN and models on Hugging Face under AutoML-org/TempoPF.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.