TempoPFN
A time-series foundation model method that pretrains linear RNNs on synthetic time series to achieve zero-shot time-series forecasting, and that has released an open-source library for synthetic data generation.
Timeline (last 90 days)
TempoPFN has been accepted by NeurIPS 2026.
The core contribution of TempoPFN is a library of synthetic time series generators and augmentation methods for training time series foundation models.
TempoPFN uses synthetic time series to pretrain a linear RNN, achieving zero-shot time series forecasting.
TempoPFN's synthetic data has been used by Toto-2 and TS-ICL.
The TempoPFN paper was published on arXiv under the number 2510.25502.
TempoPFN's code is hosted in the automl/TempoPFN repository on GitHub.
TempoPFN's model weights are hosted at AutoML-org/TempoPFN on Hugging Face.
TempoPFN is closer to a research infrastructure than an end-to-end product: the generator library is used to create data, the linear RNN is used to validate the effectiveness of synthetic pretraining, and zero-shot forecasting is the task form for testing the results.
The diversity boundary of synthetic data determines how wild a real series the model can generalize to, and the trade-offs linear RNNs make in expressive power may expose weaknesses on certain strongly nonlinear series patterns.
All Facts (9)
TempoPFN uses synthetic time series to pretrain a linear RNN, achieving zero-shot time series forecasting.
50%UnverifiedTempoPFN's synthetic data has been used by Toto-2 and TS-ICL.
50%UnverifiedTempoPFN is closer to a research infrastructure than an end-to-end product: the generator library is used to create data, the linear RNN is used to validate the effectiveness of synthetic pretraining, and zero-shot forecasting is the task form for testing the results.
50%UnverifiedThe core contribution of TempoPFN is a library of synthetic time series generators and augmentation methods for training time series foundation models.
50%UnverifiedTempoPFN has been accepted by NeurIPS 2026.
50%UnverifiedThe diversity boundary of synthetic data determines how wild a real series the model can generalize to, and the trade-offs linear RNNs make in expressive power may expose weaknesses on certain strongly nonlinear series patterns.
50%UnverifiedThe TempoPFN paper was published on arXiv under the number 2510.25502.
50%UnverifiedTempoPFN's code is hosted in the automl/TempoPFN repository on GitHub.
50%UnverifiedTempoPFN's model weights are hosted at AutoML-org/TempoPFN on Hugging Face.
50%