🕰️ TimeBase: The Power of Minimalism in Efficient Long-Term Time Series Forecasting

Welcome to the official repository for TimeBase, a Spotlight paper at ICML 2025. This repository provides all necessary code, datasets, and scripts to reproduce our results in a fully transparent and reproducible manner.

---

📦 Overview

This repository includes:

• 🔧 Scripts for training and evaluation
• 📊 Preprocessed datasets and download links
• 📉 Plug-and-play complexity reducer for PatchTST
• 📝 Detailed instructions for reproducibility

---

📁 Dataset Preparation

🔹 Benchmark Datasets (17 total)

We support the following public datasets:

ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, Traffic, Solar-Energy, Wind,
Exchange-Rate, METR-LA, ZafNoo, CzeLan, AQShunyi, AQWan, PM2.5, Temp

📥 Download from Google Drive, then unzip to the ./dataset/ directory.

🔹 Large-Scale Datasets

TimeBase is also evaluated on four large-scale real-world datasets:

• CA (4.52B records)
• GLA (2.02B)
• GBA (1.24B)
• SD (0.38B)

For these, please refer to the official preprocessing instructions from LargeST GitHub.

---

⚙️ Implementation Details

• Regularization coefficient: We perform grid search over λ_orth ∈ [0.00, 0.04, 0.08, 0.12, 0.16, 0.20].
• Learning rate: Searched in range [0.01, 0.5].
• Period settings: For datasets with a period length shorter than input length (e.g., ETTh1, ETTh2, Traffic, Electricity), we use P = 24 and #Basis R = 6.
• Loss function: Mean Squared Error (MSE)

---

🔌 Plug-and-Play Reducer for PatchTST

TimeBase can be used as a plug-in complexity reducer for patch-based models like PatchTST.

cd plug-and-play_for_patchtst
sh ./run_all.sh

---

🚀 Running TimeBase

To train and evaluate TimeBase on a given dataset:

sh ./scripts/$DATA_NAME.sh

Replace $DATA_NAME with one of the dataset names (e.g., ETTh1, Weather, etc.).

---

📄 Citation

If you find this project helpful, please cite us:

@inproceedings{huangtimebase,
  title={TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting},
  author={Huang, Qihe and Zhou, Zhengyang and Yang, Kuo and Yi, Zhongchao and Wang, Xu and Wang, Yang},
  booktitle={Forty-second International Conference on Machine Learning}
}

---

Acknowledgements

We would like to thank the authors of the following open-source projects for their valuable contributions, which provides significant help for our work:

• SparseTSF (ICML 2024)
• PatchTST (ICLR 2023)
• TFB (VLDB 2024)
• Time-Series-Library (THUML)

We gratefully acknowledge their contributions to the time series forecasting community.