PRADA: Prompt-guided Representation Alignment and Dynamic Adaption for Time Series Forecasting

Yinhao Liu¹, Zhenyu Kuang^2*, Hongyang Zhang³, Chen Li³, Feifei Li³, Xinghao Ding^3*

¹ Institute of Artificial Intelligence, Xiamen University ² School of Electronic and Information Engineering, Foshan University ³ School of Informatics, Xiamen University

---

🙋 Please let us know if you find out a mistake or have any suggestions!

🙏 If you find this resource helpful, please consider to star this repository and cite our research:

@article{liu2025prada,
  title={PRADA: Prompt-guided Representation Alignment and Dynamic Adaption for time series forecasting},
  author={Liu, Yinhao and Kuang, Zhenyu and Zhang, Hongyang and Li, Chen and Li, Feifei and Ding, Xinghao},
  journal={Knowledge-Based Systems},
  pages={113478},
  year={2025},
  publisher={Elsevier}
}

🏆 Updates/News:

🚩 News (April.1 2025): PRADA has been accpeted by Knowledge-Based Systems, 2025.

🚩 News (April.22 2025): The paper is availble at here.

🚩 News (April.27 2025): The code is now released.

📰 Introduction:

PRADA is a novel framework for time series forecasting based on LLMs. It decomposes time series into trend, seasonal, and residual terms, designs learnable textual prompts aligned with each component by the multi-view TSAA, which further mitigates the gap between natural language and time series. Additionally, a Time-Frequency Dual Constraint is applied to capture overlooked label autocorrelations.

🤗 Basic Preparation:

Datasets: You can access the well pre-processed datasets from [Google Drive] or [Baidu Drive], then place the downloaded contents under ./all_datasets.

Pretrained GPT-2: We use the pre-trained GPT-2 as our backbone network. You can download its weight from here and place it under ./gpt2.

🛠️ Requirements and Installation:

Ensure your environment meets the following requirements under python 3.11.0:

• einops==0.8.1
• joblib==1.4.2
• matplotlib==3.10.1
• numpy==1.23.5
• pandas==2.2.3
• peft==0.13.2
• scikit_learn==1.6.1
• statsmodels==0.14.4
• torch==2.1.0
• torchprofile==0.0.4
• tqdm==4.66.5
• transformers==4.45.2
• xlrd==2.0.1

For convenience, you can run:

conda create -n prada python=3.11.0 -y
conda activate prada
pip install -r requirements.txt

🚀 Getting Started:

Long-term forecasting: We provide all experimental scripts for 8 different benchmarks including ETTh1, ETTh2, ETTm1, ETTm2, Traffic, Weather, Electricity, and ILI for long-term forecasting task. The forecasting horizon is set to {24, 36, 48, 60} for ILI and {96, 192, 336, 720} for the others. For example, you can evaluate the model by:

# ETTh1
bash ./scripts/long_term_forecast/ETT_script/PRADA_ETTh1.sh
# Traffic
bash ./scripts/long_term_forecast/Traffic_script/PRADA.sh
# Weather
bash ./scripts/long_term_forecast/Weather_script/PRADA.sh
# Electricity
bash ./scripts/long_term_forecast/ECL_script/PRADA.sh
# ILI
bash ./scripts/long_term_forecast/ECL_script/ILI.sh

Short-term forecasting: We evaluate our model on M4 dataset for short-term forecasting task. The forecasting horizon is set to {6, 48} under different sampling intervals, including Yearly, Quarterly, Monthly, Weekly, Daily, and Hourly. You can run:

# M4
bash ./scripts/short_term_forecast/PRADA_M4.sh

Few-shot forecasting: You can set the parameter --percent to evaluate the model's few-shot forecasting performance. For example, for the few-shot task on 10% training data setting, you can set --percent 10 \ in your experimental script (only for long-term forecasting task).

Zero-shot forecasting: You can use the run_zsl.py file to evaluate the zero-shot performance of the model on ETT benchmark. For example, you can run the following code to evaluate the generalization in the ETTh1-ETTh2 setting:

python run_zsl.py --model_id ETTh1-ETTh2_512_96 --src_data ETTh1 --src_data_path ETTh1.csv --tgt_data ETTh2 --tgt_data_path ETTh2.csv --number_variable 7

📈 Train and Forecast:

• The model file is stored in ./models/PRADA.py, and the text prompt template is stored in ./models/prompt.py. You can imitate the content and modify it and register the model in ./exp/exp_basic.py.
• The domain description prompts for different benchmarks are stored in folder ./prompt_bank and support custom modification.

🌟 Acknowledgement:

Our implementation adapts Time-Series-Library, OFA (GPT4TS), and Time-LLM as the code base and have extensively modified it to our purposes. We thank the authors for sharing their implementations and related resources.

@inproceedings{wu2023timesnet,
  title={TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis},
  author={Haixu Wu and Tengge Hu and Yong Liu and Hang Zhou and Jianmin Wang and Mingsheng Long},
  booktitle={International Conference on Learning Representations},
  year={2023},
}

@article{wang2024tssurvey,
  title={Deep Time Series Models: A Comprehensive Survey and Benchmark},
  author={Yuxuan Wang and Haixu Wu and Jiaxiang Dong and Yong Liu and Mingsheng Long and Jianmin Wang},
  booktitle={arXiv preprint arXiv:2407.13278},
  year={2024},
}

@inproceedings{zhou2023onefitsall,
  title={{One Fits All}: Power General Time Series Analysis by Pretrained LM},
  author={Tian Zhou, Peisong Niu, Xue Wang, Liang Sun, Rong Jin},
  booktitle={NeurIPS},
  year={2023}
}

@inproceedings{jin2023time,
  title={{Time-LLM}: Time series forecasting by reprogramming large language models},
  author={Jin, Ming and Wang, Shiyu and Ma, Lintao and Chu, Zhixuan and Zhang, James Y and Shi, Xiaoming and Chen, Pin-Yu and Liang, Yuxuan and Li, Yuan-Fang and Pan, Shirui and Wen, Qingsong},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2024}
}