This repository contains the code accompanying our paper:

There are no Champions in Supervised Long-Term Time Series Forecasting
Transactions on Machine Learning Research (TMLR), January 2026
Lorenzo Brigato*, Rafael Morand*, Knut Strømmen*, Maria Panagiotou, Markus Schmidt, Stavroula Mougiakakou
University of Bern, Switzerland 🇨🇭
* Equal contribution

The repository provides the experimental framework, model implementations, datasets, and hyperparameter optimization workflows used to conduct the large-scale empirical study presented in the paper.

📊 Our work systematically evaluates supervised long-term time series forecasting models across diverse datasets and experimental settings, highlighting the lack of universally dominant architectures and offering recommendations for future research, including:

• Improving benchmarking practices
• Reducing unsubstantiated claims
• Increasing dataset diversity and revising guidelines for model selection

For additional details, please refer to our manuscript.

This codebase builds on top of the Time Series Library (TSLib) benchmark introduced at ICLR 2023:

• [Time Series Library (TSLib)]
• [ICLR 2023 paper]

We extend TSLib to support the experimental design of our study, including:

• additional baseline models
• additional datasets
• reproducible hyperparameter optimization pipelines

This repository contains all modifications and additions required to reproduce the results reported in our paper.

Beyond the existing DLinear, PatchTST, TimeMixer, iTransformer, and TimeXer implementations, we added the following models:

• ModernTCN - ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis [ICLR 2024] [Code]

• S-Mamba - Is Mamba Effective for Time Series Forecasting? [Neurocomputing 2025] [Code]

• xLSTMTime - xLSTMTime: Long-Term Time Series Forecasting with xLSTM [MDPI 2024] [Code]

• iPatch - Our hybrid transformed-based model introduced in this paper as a proof-of-concept model

In addition to the already available datasets from TSLib, the repository supports:

• UTSD (subset) - Timer: Generative Pre-trained Transformers Are Large Time Series Models [ICML 2024] [Hugging Face]
  • AustraliaRainfall
  • BeijingPM25Quality
  • KDD Cup 2018
  • Pedestrian Counts
  • TDBrain
  • BenzeneConcentration
  • MotorImagery

This repository includes utilities for large-scale and reproducible experimentation:

• Automated hyperparameter optimization using Optuna
• Reproducible search spaces aligned with the paper
• Scripts for running large-scale sweeps and model-efficiency analyses

You can obtain the well-preprocessed datasets included in the original TSLib from [Google Drive], [Baidu Drive] or [Hugging Face]. In addition, you can obtain the well-preprocessed UTSD datasets for our extensions from [Hugging Face]. Then place the downloaded data in the folder ./dataset.

1. Clone this repository.

   git clone https://github.com/AIHNlab/NoChamps
   cd NoChamps

2. Create a new Conda environment.

   conda create -n nochamps python=3.11
   conda activate nochamps

3. Install Core Dependencies.

   pip install -r requirements.txt

4. Install [PyTorch].

Run hyperparameter optimization for long-term forecasting using the provided shell script.

   bash scripts/long_term_forecast/hp_search/script_no_champions_in_ltsf.sh \
   96 \
   run.py \
   hp_results.txt

To run different configurations, adapt the shell script and call it using the following pattern.

   bash scripts/long_term_forecast/hp_search/<script>.sh \
   <prediction_length> \
   <training_script> \
   <results_file>

After completing HPO and generating the results file, you can evaluate efficiency metrics (e.g., FLOPs, training speed, etc.) for the same long-term forecasting models using the provided shell script.

   bash scripts/long_term_forecast/model_efficiency/script_no_champions_in_ltsf.sh \
   <results_file> \
   <task_name> \
   <path_to_dataset_config> \
   <batch_size> \
   <train_epochs>

Example arguments for the previous command, which corresponds to our default setup employed in the paper, are: hp_results.txt, long_term_forecast, ./scripts/long_term_forecast/model_efficiency/dataset_configs.json, 1, and 1000.

The script will generate a file containing the raw efficiency metrics per model and dataset by adding the pattern _efficiency to the <results_file>, e.g., hp_results_efficiency.txt. Additionally, the script produces another log file that aggregates all the efficiency results across datasets and further computes the efficiency-weighted error metric ξ. For more details, please refer to ./utils/exp_efficiency_analyser.py and our paper (Section 4.6).

If you find this repository useful for your research, please consider citing our paper:

@article{
brigato2026there,
title={There are no Champions in Supervised Long-Term Time Series Forecasting},
author={Lorenzo Brigato and Rafael Morand and Knut Joar Str{\o}mmen and Maria Panagiotou and Markus Schmidt and Stavroula Mougiakakou},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=yO1JuBpTBB},
note={}
}

We thank the authors and contributors of TSLib for providing a high-quality and extensible benchmark that enabled this work.