LTSF-Linear

Why we included this project

Forecasting far into the future is where transformer models tend to get heavy, and this repo offers a lighter argument. It ships three linear models: a plain Linear layer, NLinear for distribution shift, and DLinear, which separates trend from seasonality. The AAAI 2023 paper behind the code shows these one-layer models can beat popular transformer architectures on long-term forecasting benchmarks. To back that claim, the repo re-implements five well-known forecasting transformers, including Informer, Autoformer, and FEDformer, with scripts for univariate and multivariate forecasting, look-back window experiments, and weight visualization, plus a benchmark to support further research. Since each model is just a linear layer, training stays fast and light on memory, which makes this a practical baseline to check before reaching for heavier architectures.