This repository serves as a curated collection of research papers, code, and articles that critically evaluate the efficacy of Transformer and Large Language Model (LLM) architectures for time series forecasting. It aims to provide evidence and alternatives for practitioners and researchers who may be over-relying on these complex models, highlighting superior performance from simpler, non-Transformer State-of-the-Art (SOTA) methods.

How It Works

The repository compiles a comprehensive list of academic publications, many with accompanying code, that challenge the prevailing narrative around Transformers and LLMs in time series forecasting. It showcases research demonstrating that simpler models, such as MLPs, CNNs, and linear models, often achieve comparable or superior results with greater efficiency and interpretability, particularly for long-term forecasting tasks. The collection emphasizes frequency-domain analysis and decomposition techniques as key advantages.

Quick Start & Requirements

• Installation: Primarily involves cloning the repository and then installing dependencies for individual code samples (e.g., pip install -r requirements.txt).
• Prerequisites: Python 3.x, PyTorch/TensorFlow, and specific libraries for each model (e.g., NumPy, Pandas, Scikit-learn). Some code may require specific CUDA versions for GPU acceleration.
• Resources: Setup time varies per model; some may require downloading large datasets.
• Links:
  • Official GitHub Repository

Highlighted Details

• Extensive collection of papers with code, covering a wide range of non-Transformer SOTA models like SCINet, WINNET, TimesNet, and various MLP-based architectures.
• Critical analyses of LLMs and foundational models for time series, questioning their current practical utility.
• Benchmarking studies and discussions on the limitations of Transformers in handling time series data, including frequency-domain effectiveness and generalization issues.
• Inclusion of recent research (2023-2025) on alternative architectures and methodologies.

Maintenance & Community

• The repository is maintained by Valeriy Manokhin, with contributions indicated by "🔥🔥🔥🔥🔥" for highly relevant or impactful papers.
• Community engagement can be inferred from discussions on platforms like Hacker News related to TimeGPT.

Licensing & Compatibility

• The repository itself does not specify a license. Individual code samples within the repository will have their own licenses, which must be checked for compatibility with commercial or closed-source projects.

Limitations & Caveats

This repository is a curated list of research and does not provide a unified framework or single executable. Users must individually assess and implement the code for each paper. The "best" model is context-dependent, and direct comparisons across all listed papers may require significant effort.