CombinationTS

CombinationTS is a modular framework that decomposes, recombines, and diagnoses time-series forecasting models.

🔥News

• [2026-09] Our Nuwa: Evaluation-Grounded Agentic Construction of Time Series Forecasting Systems has been accepted to NeurIPS 2026! 🎉 Building on CombinationTS, Nuwa applies the principles of Evaluatology to automate the design of time-series forecasting systems through evaluation-grounded agentic construction.
• [2026-05-01] Our CombinationTS has been accepted to ICML 2026! 🎉

Overview

Figure 1. The conceptual framework of CombinationTS: modular decomposition and recombination of forecasting models (top), and distribution-based evaluation over a shared configuration space (bottom).

Installation

pip install -r requirements.txt

Requires Python 3.12+

Place datasets under ./dataset/ (ETT-small, weather, exchange_rate, traffic, electricity, illness).

Quick Start

# PatchTST
python run.py +model=patchtst +dataset=ETTh1 seq_len=96 pred_len=96

# iTransformer
python run.py +model=itransformer +dataset=Weather seq_len=96 pred_len=96

# DLinear
python run.py +model=dlinear +dataset=Exchange seq_len=336 pred_len=96

# TimesNet
python run.py +model=timesnet +dataset=ECL seq_len=96 pred_len=96

# FreTS
python run.py +model=frets +dataset=ETTh2 seq_len=96 pred_len=96

# TimeMixer
python run.py +model=timemixer +dataset=Traffic seq_len=96 pred_len=96

# Custom component combination
python run.py \
  +model/embedding=Patch16 \
  +model/encoder=Transformer \
  +model/decoder=Linear \
  +dataset=ETTh1 \
  model.use_norm=true \
  model.channel_independence=true \
  model.encoder.e_layers=1 \
  model.d_model=512

Citation

@inproceedings{
  wang2026combinationts,
  title={Combination{TS}: A Modular Framework for Understanding Time-Series Forecasting Models},
  author={Xiaorui Wang and Fanda Fan and Chenxi Wang and Yuxuan Yang and Rui Tang and Kuoyu Gao and simiao pang and Yuanfeng Shang and Zhipeng Liu and Wanling Gao and Lei Wang and Jianfeng Zhan},
  booktitle={Forty-third International Conference on Machine Learning},
  year={2026},
  url={https://openreview.net/forum?id=CwHRT46VmC}
}