SimDiff: Simpler Yet Better Diffusion Model for Time Series Point Forecasting

📕 Official implementation for AAAI 2026 paper: SimDiff: Simpler Yet Better Diffusion Model for Time Series Point Forecasting

📖 Overview

Diffusion-based time series forecasters have recently shown strong probabilistic modeling ability, but they still struggle to match regression-style methods on point forecasts: they often lack sufficient contextual bias to track distribution shifts and face a hard trade-off between sample diversity and the stability required for low MSE/MAE. Existing approaches frequently rely on external pre-trained or jointly trained regressors to stabilize training, which complicates the pipeline and sacrifices the generative flexibility that diffusion models should provide.

SimDiff is a simple yet effective diffusion framework that closes this gap, with three core components:

• Unified End-to-End Diffusion Forecaster: A single Transformer backbone simultaneously serves as denoiser and predictor, removing the need for any external pre-trained or jointly trained regressors while still providing strong contextual bias from the history window.
• Normalization Independence for Drift-Robust Training: A diffusion-specific normalization scheme decouples the statistics of past and future segments, allowing the model to handle temporal distribution shifts without data leakage and significantly improving robustness on OOD-prone datasets.
• Median-of-Means Ensemble for Trustworthy Point Estimates: By ensembling diverse diffusion samples with a Median-of-Means estimator, SimDiff converts rich predictive distributions into stable point forecasts, reducing the influence of outliers and achieving consistently lower MSE.

Across nine multivariate benchmarks, SimDiff attains state-of-the-art point forecasting performance—achieving the best or second-best MSE on all datasets—while also matching leading probabilistic baselines in CRPS/CRPS-sum, and delivers over 90% faster single-sample inference than prior diffusion-based models.

Example of Prediction Results

Normalization Independence Alleviates OOD

🚀 Quickstart

📦 Installation

Create and activate the conda environment:

conda create -n simdiff python=3.10
pip install -r requirements.txt
conda activate simdiff

You may install other necessary packages depending on the runtime requirements.

▶️ Running the Training and Inference

# Usage:
# Drag the script from the script folder to the home directory, then run:
chmod +x xxx.sh
./xxx.sh

For datasets with large size(e.g., Traffic, Electricity), can be found in the official implementation of Autoformer:https://github.com/thuml/Autoformer?tab=readme-ov-file

🏆 Acknowledgements:

Special thanks to the codebase/repository and the developers of TimesNet and TimeDiff for providing invaluable resources and contributions which are instrumental in the progress and completion of this work.

✍️ Citation

If you use our work or are inspired by our work, please consider cite us:

@misc{ding2025simdiffsimplerbetterdiffusion,
      title={SimDiff: Simpler Yet Better Diffusion Model for Time Series Point Forecasting}, 
      author={Hang Ding and Xue Wang and Tian Zhou and Tao Yao},
      year={2025},
      eprint={2511.19256},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2511.19256}, 
}