What is this?
A curated map of modern time-series forecasting architectures, from DLinear to foundation models like Chronos-2, each in its own directory with a short README and, for some, a runnable PyTorch implementation.
A curated map of modern time-series forecasting architectures, from DLinear to foundation models like Chronos-2, each in its own directory with a short README and, for some, a runnable PyTorch implementation.Is this a library I can import, or more like a reference collection?It is a reference collection: 11 numbered directories, each covering one family of forecasting models (linear baselines, attention-era transformers, PatchTST, iTransformer, mixers, SSMs, KAN, diffusion, foundation models, classical baselines), not an installable package.
It is a reference collection: 11 numbered directories, each covering one family of forecasting models (linear baselines, attention-era transformers, PatchTST, iTransformer, mixers, SSMs, KAN, diffusion, foundation models, classical baselines), not an installable package.How do I run one of these models?Each self-contained directory (01-linear-baselines, 03-patchtst, 04-itransformer, 05-cyclenet) trains on a synthetic multi-seasonal dataset and prints test MSE/MAE against a seasonal-naive baseline; drop ETTh1.csv into data/ to run the standard ETT benchmark instead.
Each self-contained directory (01-linear-baselines, 03-patchtst, 04-itransformer, 05-cyclenet) trains on a synthetic multi-seasonal dataset and prints test MSE/MAE against a seasonal-naive baseline; drop ETTh1.csv into data/ to run the standard ETT benchmark instead.Do I need anything beyond PyTorch?PyTorch is the core requirement for the self-contained implementations; the repo is pure Python with no Docker setup or other dependencies listed in the README.
PyTorch is the core requirement for the self-contained implementations; the repo is pure Python with no Docker setup or other dependencies listed in the README.Do all 11 directories have working code I can run?No, only directories 01, 03, 04, and 05 contain self-contained runnable implementations; the rest are guided tours pointing to canonical code, since reimplementing them 'adds bulk without adding understanding'.
No, only directories 01, 03, 04, and 05 contain self-contained runnable implementations; the rest are guided tours pointing to canonical code, since reimplementing them 'adds bulk without adding understanding'.Who is this map meant for?Anyone trying to understand the current forecasting landscape fast: researchers and engineers choosing a model, or students comparing mechanisms side by side instead of trusting leaderboard claims.
Anyone trying to understand the current forecasting landscape fast: researchers and engineers choosing a model, or students comparing mechanisms side by side instead of trusting leaderboard claims.Why did the authors build this instead of just reading the papers?They found the time-series literature 'confusing on purpose': every paper since 2021 claims state of the art on the same eight datasets with tiny deltas, and an ICML 2026 position paper showed normalization choices often matter more than architecture. Reading mechanisms side by side cuts through that noise.
They found the time-series literature 'confusing on purpose': every paper since 2021 claims state of the art on the same eight datasets with tiny deltas, and an ICML 2026 position paper showed normalization choices often matter more than architecture. Reading mechanisms side by side cuts through that noise.What's the big takeaway about linear baselines?DLinear, a roughly 50-line single linear layer, still beats published transformers on several standard splits, so the repo argues you should beat that baseline before trusting a fancier architecture.
DLinear, a roughly 50-line single linear layer, still beats published transformers on several standard splits, so the repo argues you should beat that baseline before trusting a fancier architecture.What's currently winning on real forecasting benchmarks?Pretrained foundation forecasters are the frontier: Chronos-2 led GIFT-Eval with a 79.8% win rate as of October 2025, ahead of TiRex and TimesFM 2.5, though the authors note leaderboards move fast and should be re-checked.
Pretrained foundation forecasters are the frontier: Chronos-2 led GIFT-Eval with a 79.8% win rate as of October 2025, ahead of TiRex and TimesFM 2.5, though the authors note leaderboards move fast and should be re-checked.Is there a place to run full reimplementations of all these models?Yes, thuml/Time-Series-Library reimplements nearly all of them under one roof; this repo links there for the models it only walks through rather than reimplements.
Yes, thuml/Time-Series-Library reimplements nearly all of them under one roof; this repo links there for the models it only walks through rather than reimplements.What are similar projects?timesfm, Time-Series-Library, Deep-Learning-Papers-Reading-Roadmap, SolarWM, Kronos, Attention-Residuals, Engram, autogluon
timesfm, Time-Series-Library, Deep-Learning-Papers-Reading-Roadmap, SolarWM, Kronos, Attention-Residuals, Engram, autogluonWhich topics is it in?forecasting, time-series, foundation-models, statistics, papers, pretrained-models, paper, neural-network, transformer, academic-research, data-science, pytorch
forecasting, time-series, foundation-models, statistics, papers, pretrained-models, paper, neural-network, transformer, academic-research, data-science, pytorch