[AAAI 2025] Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation

Abstract

Deep Neural Networks have spearheaded remarkable advancements in time series forecasting (TSF), one of the major tasks in time series modeling. Nonetheless, the non-stationarity of time series undermines the reliability of pre-trained source time series forecasters in mission-critical deployment settings. In this study, we introduce a pioneering test-time adaptation framework tailored for TSF (TSF-TTA). TAFAS, the proposed approach to TSF-TTA, flexibly adapts source forecasters to continuously shifting test distributions while preserving the core semantic information learned during pre-training. The novel utilization of partially-observed ground truth and gated calibration module enables proactive, robust, and model-agnostic adaptation of source forecasters. Experiments on diverse benchmark datasets and cutting-edge architectures demonstrate the efficacy and generality of TAFAS, especially in long-term forecasting scenarios that suffer from significant distribution shifts.

The figure above illustrates the overview pipeline of our proposed TAFAS framework.

Prepare Datasets

The datasets can be downloaded from the Time-Series-Library.

Datasets

• ETTh1
• ETTm1
• ETTh2
• ETTm2
• exchange_rate
• weather

Place the downloaded datasets in data/{dataset}/{dataset}.csv.

Example: For ETTh1, place it in data/ETTh1/ETTh1.csv.

Model Checkpoints

Model checkpoints can be downloaded from Google Drive.

You need to place the checkpoints in the project folder as stored in the Google Drive.

Example: For model FreTS and dataset ETTh1, place checkpoint_best.pth in checkpoints/FreTS/ETTh1_720/.

Due to Google Drive storage limitations, we only provide checkpoints for pred_len = 720 for the following models and datasets:

Models

• PatchTST
• DLinear
• OLS
• FreTS

Datasets

• ETTh1
• ETTm1
• ETTh2
• ETTm2
• exchange_rate
• weather

Example Execution Code

The script files are located in the scripts/ directory. Run the following script to execute the model:

bash scripts/{model}/{dataset}_{pred_len}/run.sh

Example: For model FreTS and dataset ETTh1, run:

bash scripts/FreTS/ETTh1_720/run.sh

For settings where checkpoints are not provided, you can train the model by changing TRAIN.ENABLE from False to True in run.sh.

Models and Layers Packages

The implementation references the Time-Series-Library.

Citations

If you find our work useful, please cite our paper:

@article{kim2025battling,
    title={Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation},
    author={Kim, HyunGi and Kim, Siwon and Mok, Jisoo and Yoon, Sungroh},
    journal={arXiv preprint arXiv:2501.04970},
    year={2025}
}

License

This project is licensed under the MIT License. For commercial use, permission is required.

Acknowledgements

Please provide proper attribution if you use our codebase.
If you use our work, please cite our paper as mentioned in the Citations section.