TimesFM 3.0 (PyTorch)

TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.

This repository contains the official PyTorch weights and configurations for TimesFM 3.0.

License

This model is released under the TimesFM Non-Commercial License v1.0.

Model Details

• Architecture: Stacked Mixing Transformer with Variate Attention and CPM Iterative RevIN.
• Context Patch Length: 32
• Forecast Horizon Patch Length: 64
• Layers: 20 transformer layers (model dim: 1280, heads: 16)
• Quantiles: (median at index 4)

Data

timesfm-3.0 is pretrained using

• GiftEvalPretrain excluding the datasets that overlap with fev-bench
• Wikipedia Pageviews, cutoff Nov 2023 (see paper for details).
• Google Trends top queries, cutoff EoY 2022 (see paper for details).
• Synthetic and augmented data.

Citation

@article{das2023decoder, title={A decoder-only foundation model for time-series forecasting}, author={Das, Abhimanyu and Kong, Weihao and Sen, Rajat and Zhou, Yichen}, journal={arXiv preprint arXiv:2310.10688}, year={2023} }