A map of the forecasting architectures that mattered, each in its own directory with a short README and, where it fits in a few hundred lines, a runnable PyTorch implementation.
We built this because the literature is confusing on purpose. Every paper since 2021 claims state of the art on the same eight datasets, the deltas are tiny, and an ICML 2026 position paper (arXiv 2512.22702) showed that normalization choices often matter more than the architecture being advertised. Reading the actual mechanisms side by side is the fastest way to form your own opinion.
| Directory | System | Year | The idea in one line |
|---|---|---|---|
| 01-linear-baselines | DLinear / NLinear | 2022 | one linear layer, embarrassingly hard to beat |
| 02-attention-era | Informer, Autoformer, FEDformer, Pyraformer | 2021–22 | make attention cheaper: sparsity, decomposition, Fourier bases, pyramids |
| 03-patchtst | PatchTST | 2023 | cut the series into patches and treat each patch as a token |
| 04-itransformer | iTransformer | 2024 | flip the axes: each variable's whole history becomes one token |
| 05-cyclenet | CycleNet | 2024 | learn the weekly/daily cycle explicitly, forecast the residual with a linear head |
| 06-mixers | TimeMixer, xPatch, WPMixer | 2024–25 | decompose (scales, EMA, wavelets), then mix with MLPs |
| 07-ssm-and-xlstm | S-Mamba, xLSTMTime, TiRex | 2024–25 | recurrence returns with linear scaling |
| 08-kan | TimeKAN | 2025 | learnable activation functions on frequency bands |
| 09-diffusion | TimeDiff, CCDM, ARMD | 2023–25 | sample forecast distributions; strong calibration, slow sampling |
| 10-foundation-models | Chronos-2, TimesFM 2.5, Moirai 2.0, Toto, Sundial | 2024–26 | pretrained zero-shot forecasters, the current frontier |
| 11-classical-baselines | AutoETS, AutoARIMA, seasonal naive | 1970s– | the control group; still wins on short seasonal series |
Directories 01, 03, 04, and 05 contain self-contained implementations that train on synthetic data in under a minute on a laptop. The rest are guided tours with pointers to the canonical code, because reimplementing them adds bulk without adding understanding. thuml/Time-Series-Library reimplements nearly all of them under one roof if you want to run the originals.
pip install torch numpy
python common/train.py --model dlinear
python common/train.py --model patchtst
python common/train.py --model itransformer
python common/train.py --model cyclenet
Each command trains on a synthetic multi-seasonal dataset and prints test MSE/MAE next to a seasonal-naive baseline. Drop ETTh1.csv into data/ to run on the standard benchmark instead (download).
Three things survived contact with our own experiments and the 2025–26 literature:
- Linear baselines first. DLinear is 50 lines and beats published transformers on several standard splits (arXiv 2512.10866). If your fancy model can't beat it, stop.
- Benchmark wins on ETT are mostly noise. Lookback length, normalization, and channel handling flip rankings. State your protocol before your architecture.
- The frontier moved to pretrained models. On GIFT-Eval, Chronos-2 led with a 79.8% win rate as of October 2025, ahead of TiRex and TimesFM 2.5. Zero-shot with covariates is the setting where 2026 claims get adjudicated, on GIFT-Eval and fev-bench. Leaderboards move; re-check before you quote them.
Dreamers Inc. We build forecasting and agent systems for clients and kept rewriting these models from papers to understand them; publishing the rewrites seemed more useful than hoarding them. Corrections welcome, issues especially.
If you are working on predictive alpha signals, or RL-based risk assessment and execution systems for live markets, write to info@thedreamers.us.
Apache-2.0.