TempoPFN: Synthetic Pre-Training of Linear RNNs for Zero-Shot Time Series Forecasting
TempoPFN introduced in TempoPFN: Synthetic Pre-Training of Linear RNNs for Zero-Shot Time Series Forecasting, is a univariate time series foundation model pretrained entirely on synthetic data. It delivers top-tier zero-shot forecasting accuracy while remaining fully reproducible and free from real-data leakage.
Built on a Linear RNN (GatedDeltaProduct) backbone, TempoPFN performs end-to-end forecasting without patching or windowing. Its design enables fully parallelizable training and inference while maintaining stable temporal state-tracking across long sequences. The GatedDeltaProduct architecture is based on DeltaProduct, extended with state-weaving for time series forecasting. For detailed information about the architecture and custom modifications, see src/models/gated_deltaproduct/README.md.
This repository includes the pretrained 38M parameter model (models/checkpoint_38M.pth), all training and inference code, and the complete synthetic data generation pipeline used for pretraining.
β¨ Why TempoPFN?
- High Performance, No Real Data: Achieves top-tier competitive results on GIFT-Eval, outperforming all existing synthetic-only approaches and surpassing the vast majority of models trained on real-world data. This ensures full reproducibility and eliminates benchmark leakage.
- Parallel and Efficient: The linear recurrence design enables full-sequence parallelization. This gives us the best of both worlds: the linear efficiency of an RNN, but with the training parallelism of a Transformer.
- Open and Reproducible: Includes the full synthetic data pipeline, configurations, and scripts to reproduce training from scratch.
- State-Tracking Stability: The GatedDeltaProduct recurrence and state-weaving mechanism preserve temporal continuity and information flow across long horizons, improving robustness without non-linear recurrence.
βοΈ Installation
Note on Model Weights: This repository uses Git LFS to store the model checkpoint (
.pthfile). You must have Git LFS installed to clone the repository correctly.# Install Git LFS (e.g., on Ubuntu) sudo apt-get install git-lfs git lfs install
- Clone the repository:
git clone https://huggingface.co/AutoML-org/TempoPFN
cd TempoPFN
- Set up the environment:
python -m venv venv && source venv/bin/activate
# 1. Install PyTorch version matching your CUDA version
# Example for CUDA 12.8:
pip install torch --index-url https://download.pytorch.org/whl/cu128
# 2. Clone the Hugging Face repository
git clone https://huggingface.co/AutoML-org/TempoPFN
cd TempoPFN
# 3. Set up the environment
python3.12 -m venv venv & source venv/bin/activate
export PYTHONPATH=$PWD
# 4. Install PyTorch version matching your CUDA version
pip install torch --index-url https://download.pytorch.org/whl/cu128
# 5. Install dependencies
pip install .
pip install .[dev]
# 4. Run the Quick Start Script
python examples/quick_start_tempo_pfn.py
# 5. Alternatively, you can run the Notebook version
jupyter notebook examples/quick_start_tempo_pfn.ipynb
Hardware & Performance Tips
GPU Required: Inference requires a CUDA-capable GPU with a matching PyTorch version installed. Tested on NVIDIA A100/H100.
First Run: The first inference for a new sequence length will be slow due to Triton kernel compilation. Subsequent runs will be fast.
Cache Tip: If using a network filesystem, prevent slowdowns by routing caches to a local directory (like /tmp) before running:
LOCAL_CACHE_BASE="${TMPDIR:-/tmp}/tsf-$(date +%s)"
mkdir -p "${LOCAL_CACHE_BASE}/triton" "${LOCAL_CACHE_BASE}/torchinductor"
export TRITON_CACHE_DIR="${LOCAL_CACHE_BASE}/triton"
export TORCHINDUCTOR_CACHE_DIR="${LOCAL_CACHE_BASE}/torchinductor"
python examples/quick_start_tempo_pfn.py
π Training
All training and model parameters are controlled via YAML files in configs/.
# Single-GPU (Debug)
torchrun --standalone --nproc_per_node=1 src/training/trainer_dist.py --config ./configs/train.yaml
# Multi-GPU (e.g., 8 GPUs)
torchrun --standalone --nproc_per_node=8 src/training/trainer_dist.py --config ./configs/train.yaml
πΎ Synthetic Data Generation
A core contribution of this work is our open-source synthetic data pipeline, located in src/synthetic_generation/. It combines diverse generators with a powerful augmentation cascade.
Generators Used:
- Adapted Priors: ForecastPFN, KernelSynth, GaussianProcess (GP), and CauKer (Structural Causal Models).
- Novel Priors: SDE (a flexible regime-switching Ornstein-Uhlenbeck process), Sawtooth, StepFunction, Anomaly, Spikes, SineWave, and Audio-Inspired generators (Stochastic Rhythms, Financial Volatility, Network Topology, Multi-Scale Fractals).
You can easily generate your own data by installing the development dependencies and instantiating a generator wrapper. See examples/generate_synthetic_data.py for a minimal script, or inspect the generator code in src/synthetic_generation/.
π€ License
This project is licensed under the Apache 2.0 License. See the LICENSE file for details. This permissive license allows for both academic and commercial use.
π Citation
If you find TempoPFN useful in your research, please consider citing our paper:
@misc{moroshan2025tempopfn,
title={TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-Shot Time Series Forecasting},
author={Vladyslav Moroshan and Julien Siems and Arber Zela and Timur Carstensen and Frank Hutter},
year={2025},
eprint={2510.25502},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
