Skip to content
BenchCouncilPublic

About

[Neurips 2026] Official implementation of "MSAR: Next-Scale Autoregressive Forecasting for Time Series"

Topics

Resources

Stars

0 stars

Watchers

1 watching

Forks

Latest commit

 

History

2 Commits

Folders and files

Repository files navigation

MSAR

Next-Scale Autoregressive Forecasting for Time Series via Modular Multi-Scale Decoupling

Paper · Forecasting framework and reproducible code

MSAR coarse-to-fine forecasting framework

MSAR is a modular, model-agnostic framework for time-series forecasting. It forecasts progressively from coarse to fine temporal scales: coarse forecasts guide finer-scale models, which refine the predictions at higher resolution.

This repository includes a shared training pipeline for single-scale baselines and MSAR-enhanced models.

Supported backbones

Backbone Baseline + MSAR
PatchTST ✓ ✓
DLinear ✓ ✓
TimeMixer ✓ ✓
iTransformer ✓ ✓
TimesNet ✓ ✓
SimpleTM ✓ ✓
DUET ✓ ✓

Usage

1. Environment setup

Use Python 3.9 or newer. Install a PyTorch build compatible with your CUDA environment, then install the remaining requirements:

pip install -r requirements.txt

2. Data preparation

The included loaders support ETTh1, ETTh2, ETTm1, ETTm2, and custom CSV files. For ETTh1, place the CSV here:

./dataset/ETT-small/ETTh1.csv

CSV files should have a date column followed by numeric time-series columns. For other ETT datasets, set --data and --data_path to the matching dataset; use --freq t for ETTm data. For custom data, pass --data custom together with --root_path, --data_path, --target, and --freq.

3. Training and evaluation

Run commands from the repository root. The default training loss is MAE and the default initial learning rate is 1e-3.

Baseline — single scale (interval_list=1):

python run.py \
  --model PatchTST \
  --data ETTh1 --root_path ./dataset/ETT-small --data_path ETTh1.csv \
  --seq_len 336 --label_len 0 --pred_len 96 \
  --interval_list 1 --seq_len_list 336 \
  --loss MAE

MSAR — coarse scale followed by full-resolution refinement:

bash scripts/run_msar.sh

The example uses interval_list=6,1, seq_len_list=56,336, and enables slim imputation (d_model=16, d_ff=32). Set MODEL to choose a different backbone, for example:

MODEL=TimeMixer bash scripts/run_msar.sh
MODEL=DLinear bash scripts/run_msar.sh
MODEL=SimpleTM bash scripts/run_msar.sh

DUET: slimming the imputation model is not recommended based on our width comparison. The script automatically leaves DUET full-width:

MODEL=DUET bash scripts/run_msar.sh

To train with MSE, pass --loss MSE to run.py or edit the example script. Use python run.py --help for model-specific parameters. The available backbones and MSAR training stages are implemented in models/ and exp/.

Outputs

  • Checkpoints: ./checkpoints/
  • Test metrics: appended to ./results.txt by default; change with --result_file
  • Dataset, checkpoint, and generated result files are excluded by .gitignore.

Citation

If you use MSAR, please cite:

@inproceedings{fan2026msar,
  title     = {{MSAR}: Next-Scale Autoregressive Forecasting for Time Series via Modular Multi-Scale Decoupling},
  author    = {Fanda Fan and Kuiye Ding and Liming Mao and Bingsong Wang and
               Yao Wang and Xiaorui Wang and Ruijie Jian and Zhipeng Liu and
               Luqi Gong and Zhenghua Lu and Chunjie Luo and Jianfeng Zhan},
  booktitle = {The Fortieth Annual Conference on Neural Information Processing Systems},
  year      = {2026},
  url       = {https://openreview.net/forum?id=ryJy1ZX3II}
}

Acknowledgements

This code is adapted from Time-Series-Library.

About

[Neurips 2026] Official implementation of "MSAR: Next-Scale Autoregressive Forecasting for Time Series"

Topics

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages