Super-Linear: A Lightweight Pretrained Mixture of Linear Experts for Time Series Forecasting (TMLR 2026)

[GitHub] [Hugging Face] [Paper]

Super Linear provides a comprehensive foundation for time series forecasting achieveing competitive performance against more complex models with an efficient architecture combine mixture of frequencies linears experts.

Updates/News:

🚩 News (April 2026): Accepted to TMLR!

🚩 News (January 2026): Super Linear optimized code update! now inference is more than $\times 10$ faster! which makes it the fastest pre-trained forecaster!

🚩 News (September 2025): Super Linear v1.0.0 has been released!

Result Snapshot 🔬

Super-Linear delivers competitive, and in some cases surpassing , state-of-the-art Transformer models, all while using a simple linear backbone with just 2.5M parameters, and superior inference speed!

Getting Started 🚀

Installation

1. Clone the repository

pip install -r requirements.txt

Making Forecasts (Hugging Face) 📈

import torch
from transformers import AutoModelForCausalLM

model_path = "./model_files/" # local
model_path = "SequentialLearning/SuperLinear" # hugging_face
model = AutoModelForCausalLM.from_pretrained(model_path,trust_remote_code=True, force_download=True)

seq_len = 512
pred_len = 96

freq = 1/24
amp = 1
ph = 0

t = torch.arange(0, seq_len+ pred_len)
s = amp * torch.sin(2 * np.pi * freq * t + ph)
x = s[:seq_len].unsqueeze(0)  # Add batch dim
y = s[seq_len:].unsqueeze(0)  # Add batch dim

with torch.no_grad():
    # takes shapes (B, V, L) or (B, L)
    output = model(x, pred_len=pred_len, get_prob=True)
    preds = output.logits # Predicted values
    probs = output.attentions  # Expert probabilities stored here
expert_names = model.backbone.experts.keys()

Data Preparation

You can obtain all the benchmarks from Google Drive provided in Autoformer. All the datasets are well pre-processed and can be used easily.

mkdir dataset

Please put them in the ./dataset directory

Evaluation 📐

• [Example] Running the follow command to evaluate on ETTh1 with prediciton horizon 96

python run_eval.py -d dataset/ETT-small/ETTh1.csv -p 96

• [Example] Running the follow command to evaluate on ETTh1 with prediciton horizon 720

python run_eval.py -d dataset/ETT-small/ETTh1.csv -p 720

• [Example] Running the follow command to evaluate on weather with prediciton horizon 720

python run_eval.py -d dataset/weather/weather.csv -p 720

Full-Shot, Few-Shot 🔥

In progress

TODO ⏳

• Add Super-Linear to GIFT-Eval leaderboard. The results and notebook for GIFT-Eval can be found in the gift-eval directory.

Citation📕

If you use SuperLinear in your research, please cite:

@article{nochumsohn2025super,
  title={Super-Linear: A Lightweight Pretrained Mixture of Linear Experts for Time Series Forecasting},
  author={Nochumsohn, Liran and Marshanski, Raz and Zisling, Hedi and Azencot, Omri},
  journal={arXiv preprint arXiv:2509.15105},
  year={2025}
}

Resources 📄

• Are Transformers Effective for Time Series Forecasting? [paper] [GitHub Repo]
• Mixture-of-Linear-Experts for Long-term Time Series Forecasting, in AISTATS 2024. [paper] [GitHub Repo]
• A Multi-Task Learning Approach to Linear Multivariate Forecasting, in AISTATS 2025. [paper] [GitHub Repo]
• Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts, in ICLR 2025. [paper] [GitHub Repo]

Acknowledgments

We appreciate the following GitHub repos a lot for their valuable code and efforts.

• Time-Moe [repo]
• Time-Series-Library [repo]

Contact 📫

Liran Nochumsohn: lirannoc@post.bgu.ac.il

License

This project is licensed under the MIT License - see the LICENSE file for details.