README.md

This is an official implementation of Time Series Forecasting via Direct Per-Step Probability Distribution Modeling.

Model Overview

We employ a base model based on trend-seasonality decomposition. The output head of the base model is modified to produce per-step discrete probability distribution over a predefined support set.

Interleaved dual branches: Two non-shared branches output probability distributions over interleaved support sets, with a consistency constraint applied between them.

Coarse-scale guidance: The interleaved dual-branch architecture is replicated at coarser temporal scales, serving as an additional self-supervised signal to guide long-term trend forecasting at regular scales.

Results

We provide below the results of multivariate time series forecasting on nine real-world datasets after hyperparameter search. Our method outperforms state-of-the-art approaches on most forecasting tasks.

Getting Started

1. Install conda environment: conda env create -f environment.yml

2. Download data. You can download the datasets from Google Driver, Baidu Driver or Kaggle Datasets. All datasets are pre-processed and can be used easily. Create a seperate folder ./dataset and put all the files in the directory.

3. Train the model. We provide the experiment scripts of all benchmarks under the folder ./scripts. The script for hyperparameter search is interPDN_search. Take ETTh1 dataset as example, you can reproduce the experiments by:

bash ./scripts/interPDN_search/ETTh1.sh

If you wish to reproduce the results on 4 ETT datasets when the base model is replaced with DLinear, you can run:

bash ./scripts/DLinear/ETTh1.sh

Acknowledgement

We appreciate the following github repos for their valuable code and effort:

• DLinear (https://github.com/cure-lab/LTSF-Linear)
• xPatch (https://github.com/stitsyuk/xPatch)
• Time-Series-Library (https://github.com/thuml/Time-Series-Library)
• RevIN (https://github.com/ts-kim/RevIN)

Contact

If you have any questions or concerns, please contact us at leonardokong486@gmail.com or submit an issue.

Citation

If you find this repository useful in your research, please consider citing our paper as follows:

@inproceedings{kong2026interPDN,
  title={Time Series Forecasting via Direct Per-Step Probability Distribution Modeling},
  author={Kong, Linghao and Hong, Xiaopeng},
  booktitle={Proceedings of the 40th Annual AAAI Conference on Artificial Intelligence},
  year={2026}
}