amazon-science/TSFM-Biases

amazon-science/TSFM-Biases

Description: Official Implementation of Understanding the Implicit Biases of Design Choices for Time Series Foundation Models

Language: Jupyter Notebook

License: Apache-2.0

Stars: 3

Forks: 0

Open issues: 0

Created: 2026-03-18T22:48:34Z

Pushed: 2026-03-18T23:15:07Z

Default branch: main

Fork: no

Archived: no

README:

Understanding the Implicit Biases of Design Choices for Time Series Foundation Models

This repository is associated with the paper "Understanding the Implicit Biases of Design Choices for Time Series Foundation Models" by Annan Yu, Danielle C. Maddix, Boran Han, Xiyuan Zhang, Abdul Fatir Ansari, Oleksandr Shchur, Christos Faloutsos, Andrew Gordon Wilson, Michael W. Mahoney, and Yuyang Wang. It contains codes that are used to investigate the biases one should be aware of when designing time-series foundation models. It is organized as follows.

Pretraining Checkpoints

• [hybrid-models](./hybrid-models/): containing the training scripts of two models formed by combining the design choices of Chronos and Chronos-Bolt. These models are used for ablation to better investigate the role of an isolated design choice. Please refer to the **chronos-forecasting** repository for additional training details.

Evaluating the Checkpoints

• [evaluation](./evaluation/): containing the code used to evaluate TSFMs' scale and offset biases on Chronos' evaluation corpus. These experiments are explained in Figure 5 and 6 in our paper.

Investigating Six Inductive Biases

• [notebooks](./notebooks/): containing the notebooks that we used to perform diagnostic investigations of biases induced by different design choices of time-series foundation models. These controlled experiments are explained in Figure 2-8 in our paper.

Source

This repository contains modified versions of the code found in the following repository:

**chronos-forecasting**: For training hybrid Chronos models that are used to establish controlled experiments to test the inductive biases of TSFMs with different design choices.

Citation

If you use this code, or our work, please cite:

@inproceedings{yu2026implicit,
title={Understanding the Implicit Biases of Design Choices for Time Series Foundation Models},
author={Yu, A. and Maddix, D.C. and Han, B. and Zhang, X. and Ansari, A.F. and Shchur, O. and Faloutsos, C. and Wilson, A.G. and Mahoney, M.W. and Wang, Y.},
booktitle={International Conference on Learning Representations},
year={2026}
}

License

This project is licensed under the Apache-2.0 License.